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 int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 935 {
 936         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 937         struct tls_context *tls_ctx = tls_get_ctx(sk);
 938         struct tls_prot_info *prot = &tls_ctx->prot_info;
 939         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 940         bool async_capable = ctx->async_capable;
 941         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 942         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
 943         bool eor = !(msg->msg_flags & MSG_MORE);
 944         size_t try_to_copy;
 945         ssize_t copied = 0;
 946         struct sk_msg *msg_pl, *msg_en;
 947         struct tls_rec *rec;
 948         int required_size;
 949         int num_async = 0;
 950         bool full_record;
 951         int record_room;
 952         int num_zc = 0;
 953         int orig_size;
 954         int ret = 0;
 955         int pending;
 956
 957         if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
 958                                MSG_CMSG_COMPAT | MSG_SPLICE_PAGES))
 959                 return -EOPNOTSUPP;
 960
 961         ret = mutex_lock_interruptible(&tls_ctx->tx_lock);
 962         if (ret)
 963                 return ret;
 964         lock_sock(sk);
 965
 966         if (unlikely(msg->msg_controllen)) {
 967                 ret = tls_process_cmsg(sk, msg, &record_type);
 968                 if (ret) {
 969                         if (ret == -EINPROGRESS)
 970                                 num_async++;
 971                         else if (ret != -EAGAIN)
 972                                 goto send_end;
 973                 }
 974         }
 975
 976         while (msg_data_left(msg)) {
 977                 if (sk->sk_err) {
 978                         ret = -sk->sk_err;
 979                         goto send_end;
 980                 }
 981
 982                 if (ctx->open_rec)
 983                         rec = ctx->open_rec;
 984                 else
 985                         rec = ctx->open_rec = tls_get_rec(sk);
 986                 if (!rec) {
 987                         ret = -ENOMEM;
 988                         goto send_end;
 989                 }
 990
 991                 msg_pl = &rec->msg_plaintext;
 992                 msg_en = &rec->msg_encrypted;
 993
 994                 orig_size = msg_pl->sg.size;
 995                 full_record = false;
 996                 try_to_copy = msg_data_left(msg);
 997                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
 998                 if (try_to_copy >= record_room) {
 999                         try_to_copy = record_room;
1000                         full_record = true;
1001                 }
1002
1003                 required_size = msg_pl->sg.size + try_to_copy +
1004                                 prot->overhead_size;
1005
1006                 if (!sk_stream_memory_free(sk))
1007                         goto wait_for_sndbuf;
1008
1009 alloc_encrypted:
1010                 ret = tls_alloc_encrypted_msg(sk, required_size);
1011                 if (ret) {
1012                         if (ret != -ENOSPC)
1013                                 goto wait_for_memory;
1014
1015                         /* Adjust try_to_copy according to the amount that was
1016                          * actually allocated. The difference is due
1017                          * to max sg elements limit
1018                          */
1019                         try_to_copy -= required_size - msg_en->sg.size;
1020                         full_record = true;
1021                 }
1022
1023                 if (!is_kvec && (full_record || eor) && !async_capable) {
1024                         u32 first = msg_pl->sg.end;
1025
1026                         ret = sk_msg_zerocopy_from_iter(sk, &msg->msg_iter,
1027                                                         msg_pl, try_to_copy);
1028                         if (ret)
1029                                 goto fallback_to_reg_send;
1030
1031                         num_zc++;
1032                         copied += try_to_copy;
1033
1034                         sk_msg_sg_copy_set(msg_pl, first);
1035                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1036                                                   record_type, &copied,
1037                                                   msg->msg_flags);
1038                         if (ret) {
1039                                 if (ret == -EINPROGRESS)
1040                                         num_async++;
1041                                 else if (ret == -ENOMEM)
1042                                         goto wait_for_memory;
1043                                 else if (ctx->open_rec && ret == -ENOSPC)
1044                                         goto rollback_iter;
1045                                 else if (ret != -EAGAIN)
1046                                         goto send_end;
1047                         }
1048                         continue;
1049 rollback_iter:
1050                         copied -= try_to_copy;
1051                         sk_msg_sg_copy_clear(msg_pl, first);
1052                         iov_iter_revert(&msg->msg_iter,
1053                                         msg_pl->sg.size - orig_size);
1054 fallback_to_reg_send:
1055                         sk_msg_trim(sk, msg_pl, orig_size);
1056                 }
1057
1058                 required_size = msg_pl->sg.size + try_to_copy;
1059
1060                 ret = tls_clone_plaintext_msg(sk, required_size);
1061                 if (ret) {
1062                         if (ret != -ENOSPC)
1063                                 goto send_end;
1064
1065                         /* Adjust try_to_copy according to the amount that was
1066                          * actually allocated. The difference is due
1067                          * to max sg elements limit
1068                          */
1069                         try_to_copy -= required_size - msg_pl->sg.size;
1070                         full_record = true;
1071                         sk_msg_trim(sk, msg_en,
1072                                     msg_pl->sg.size + prot->overhead_size);
1073                 }
1074
1075                 if (try_to_copy) {
1076                         ret = sk_msg_memcopy_from_iter(sk, &msg->msg_iter,
1077                                                        msg_pl, try_to_copy);
1078                         if (ret < 0)
1079                                 goto trim_sgl;
1080                 }
1081
1082                 /* Open records defined only if successfully copied, otherwise
1083                  * we would trim the sg but not reset the open record frags.
1084                  */
1085                 tls_ctx->pending_open_record_frags = true;
1086                 copied += try_to_copy;
1087                 if (full_record || eor) {
1088                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1089                                                   record_type, &copied,
1090                                                   msg->msg_flags);
1091                         if (ret) {
1092                                 if (ret == -EINPROGRESS)
1093                                         num_async++;
1094                                 else if (ret == -ENOMEM)
1095                                         goto wait_for_memory;
1096                                 else if (ret != -EAGAIN) {
1097                                         if (ret == -ENOSPC)
1098                                                 ret = 0;
1099                                         goto send_end;
1100                                 }
1101                         }
1102                 }
1103
1104                 continue;
1105
1106 wait_for_sndbuf:
1107                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1108 wait_for_memory:
1109                 ret = sk_stream_wait_memory(sk, &timeo);
1110                 if (ret) {
1111 trim_sgl:
1112                         if (ctx->open_rec)
1113                                 tls_trim_both_msgs(sk, orig_size);
1114                         goto send_end;
1115                 }
1116
1117                 if (ctx->open_rec && msg_en->sg.size < required_size)
1118                         goto alloc_encrypted;
1119         }
1120
1121         if (!num_async) {
1122                 goto send_end;
1123         } else if (num_zc) {
1124                 /* Wait for pending encryptions to get completed */
1125                 spin_lock_bh(&ctx->encrypt_compl_lock);
1126                 ctx->async_notify = true;
1127
1128                 pending = atomic_read(&ctx->encrypt_pending);
1129                 spin_unlock_bh(&ctx->encrypt_compl_lock);
1130                 if (pending)
1131                         crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1132                 else
1133                         reinit_completion(&ctx->async_wait.completion);
1134
1135                 /* There can be no concurrent accesses, since we have no
1136                  * pending encrypt operations
1137                  */
1138                 WRITE_ONCE(ctx->async_notify, false);
1139
1140                 if (ctx->async_wait.err) {
1141                         ret = ctx->async_wait.err;
1142                         copied = 0;
1143                 }
1144         }
1145
1146         /* Transmit if any encryptions have completed */
1147         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1148                 cancel_delayed_work(&ctx->tx_work.work);
1149                 tls_tx_records(sk, msg->msg_flags);
1150         }
1151
1152 send_end:
1153         ret = sk_stream_error(sk, msg->msg_flags, ret);
1154
1155         release_sock(sk);
1156         mutex_unlock(&tls_ctx->tx_lock);
1157         return copied > 0 ? copied : ret;
1158 }
1159
1160 static int tls_sw_do_sendpage(struct sock *sk, struct page *page,
1161                               int offset, size_t size, int flags)
1162 {
1163         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1164         struct tls_context *tls_ctx = tls_get_ctx(sk);
1165         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1166         struct tls_prot_info *prot = &tls_ctx->prot_info;
1167         unsigned char record_type = TLS_RECORD_TYPE_DATA;
1168         struct sk_msg *msg_pl;
1169         struct tls_rec *rec;
1170         int num_async = 0;
1171         ssize_t copied = 0;
1172         bool full_record;
1173         int record_room;
1174         int ret = 0;
1175         bool eor;
1176
1177         eor = !(flags & MSG_SENDPAGE_NOTLAST);
1178         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1179
1180         /* Call the sk_stream functions to manage the sndbuf mem. */
1181         while (size > 0) {
1182                 size_t copy, required_size;
1183
1184                 if (sk->sk_err) {
1185                         ret = -sk->sk_err;
1186                         goto sendpage_end;
1187                 }
1188
1189                 if (ctx->open_rec)
1190                         rec = ctx->open_rec;
1191                 else
1192                         rec = ctx->open_rec = tls_get_rec(sk);
1193                 if (!rec) {
1194                         ret = -ENOMEM;
1195                         goto sendpage_end;
1196                 }
1197
1198                 msg_pl = &rec->msg_plaintext;
1199
1200                 full_record = false;
1201                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
1202                 copy = size;
1203                 if (copy >= record_room) {
1204                         copy = record_room;
1205                         full_record = true;
1206                 }
1207
1208                 required_size = msg_pl->sg.size + copy + prot->overhead_size;
1209
1210                 if (!sk_stream_memory_free(sk))
1211                         goto wait_for_sndbuf;
1212 alloc_payload:
1213                 ret = tls_alloc_encrypted_msg(sk, required_size);
1214                 if (ret) {
1215                         if (ret != -ENOSPC)
1216                                 goto wait_for_memory;
1217
1218                         /* Adjust copy according to the amount that was
1219                          * actually allocated. The difference is due
1220                          * to max sg elements limit
1221                          */
1222                         copy -= required_size - msg_pl->sg.size;
1223                         full_record = true;
1224                 }
1225
1226                 sk_msg_page_add(msg_pl, page, copy, offset);
1227                 sk_mem_charge(sk, copy);
1228
1229                 offset += copy;
1230                 size -= copy;
1231                 copied += copy;
1232
1233                 tls_ctx->pending_open_record_frags = true;
1234                 if (full_record || eor || sk_msg_full(msg_pl)) {
1235                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1236                                                   record_type, &copied, flags);
1237                         if (ret) {
1238                                 if (ret == -EINPROGRESS)
1239                                         num_async++;
1240                                 else if (ret == -ENOMEM)
1241                                         goto wait_for_memory;
1242                                 else if (ret != -EAGAIN) {
1243                                         if (ret == -ENOSPC)
1244                                                 ret = 0;
1245                                         goto sendpage_end;
1246                                 }
1247                         }
1248                 }
1249                 continue;
1250 wait_for_sndbuf:
1251                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1252 wait_for_memory:
1253                 ret = sk_stream_wait_memory(sk, &timeo);
1254                 if (ret) {
1255                         if (ctx->open_rec)
1256                                 tls_trim_both_msgs(sk, msg_pl->sg.size);
1257                         goto sendpage_end;
1258                 }
1259
1260                 if (ctx->open_rec)
1261                         goto alloc_payload;
1262         }
1263
1264         if (num_async) {
1265                 /* Transmit if any encryptions have completed */
1266                 if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1267                         cancel_delayed_work(&ctx->tx_work.work);
1268                         tls_tx_records(sk, flags);
1269                 }
1270         }
1271 sendpage_end:
1272         ret = sk_stream_error(sk, flags, ret);
1273         return copied > 0 ? copied : ret;
1274 }
1275
1276 int tls_sw_sendpage_locked(struct sock *sk, struct page *page,
1277                            int offset, size_t size, int flags)
1278 {
1279         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1280                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY |
1281                       MSG_NO_SHARED_FRAGS))
1282                 return -EOPNOTSUPP;
1283
1284         return tls_sw_do_sendpage(sk, page, offset, size, flags);
1285 }
1286
1287 int tls_sw_sendpage(struct sock *sk, struct page *page,
1288                     int offset, size_t size, int flags)
1289 {
1290         struct tls_context *tls_ctx = tls_get_ctx(sk);
1291         int ret;
1292
1293         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1294                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY))
1295                 return -EOPNOTSUPP;
1296
1297         ret = mutex_lock_interruptible(&tls_ctx->tx_lock);
1298         if (ret)
1299                 return ret;
1300         lock_sock(sk);
1301         ret = tls_sw_do_sendpage(sk, page, offset, size, flags);
1302         release_sock(sk);
1303         mutex_unlock(&tls_ctx->tx_lock);
1304         return ret;
1305 }
1306
1307 static int
1308 tls_rx_rec_wait(struct sock *sk, struct sk_psock *psock, bool nonblock,
1309                 bool released)
1310 {
1311         struct tls_context *tls_ctx = tls_get_ctx(sk);
1312         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1313         DEFINE_WAIT_FUNC(wait, woken_wake_function);
1314         long timeo;
1315
1316         timeo = sock_rcvtimeo(sk, nonblock);
1317
1318         while (!tls_strp_msg_ready(ctx)) {
1319                 if (!sk_psock_queue_empty(psock))
1320                         return 0;
1321
1322                 if (sk->sk_err)
1323                         return sock_error(sk);
1324
1325                 if (!skb_queue_empty(&sk->sk_receive_queue)) {
1326                         tls_strp_check_rcv(&ctx->strp);
1327                         if (tls_strp_msg_ready(ctx))
1328                                 break;
1329                 }
1330
1331                 if (sk->sk_shutdown & RCV_SHUTDOWN)
1332                         return 0;
1333
1334                 if (sock_flag(sk, SOCK_DONE))
1335                         return 0;
1336
1337                 if (!timeo)
1338                         return -EAGAIN;
1339
1340                 released = true;
1341                 add_wait_queue(sk_sleep(sk), &wait);
1342                 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1343                 sk_wait_event(sk, &timeo,
1344                               tls_strp_msg_ready(ctx) ||
1345                               !sk_psock_queue_empty(psock),
1346                               &wait);
1347                 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1348                 remove_wait_queue(sk_sleep(sk), &wait);
1349
1350                 /* Handle signals */
1351                 if (signal_pending(current))
1352                         return sock_intr_errno(timeo);
1353         }
1354
1355         tls_strp_msg_load(&ctx->strp, released);
1356
1357         return 1;
1358 }
1359
1360 static int tls_setup_from_iter(struct iov_iter *from,
1361                                int length, int *pages_used,
1362                                struct scatterlist *to,
1363                                int to_max_pages)
1364 {
1365         int rc = 0, i = 0, num_elem = *pages_used, maxpages;
1366         struct page *pages[MAX_SKB_FRAGS];
1367         unsigned int size = 0;
1368         ssize_t copied, use;
1369         size_t offset;
1370
1371         while (length > 0) {
1372                 i = 0;
1373                 maxpages = to_max_pages - num_elem;
1374                 if (maxpages == 0) {
1375                         rc = -EFAULT;
1376                         goto out;
1377                 }
1378                 copied = iov_iter_get_pages2(from, pages,
1379                                             length,
1380                                             maxpages, &offset);
1381                 if (copied <= 0) {
1382                         rc = -EFAULT;
1383                         goto out;
1384                 }
1385
1386                 length -= copied;
1387                 size += copied;
1388                 while (copied) {
1389                         use = min_t(int, copied, PAGE_SIZE - offset);
1390
1391                         sg_set_page(&to[num_elem],
1392                                     pages[i], use, offset);
1393                         sg_unmark_end(&to[num_elem]);
1394                         /* We do not uncharge memory from this API */
1395
1396                         offset = 0;
1397                         copied -= use;
1398
1399                         i++;
1400                         num_elem++;
1401                 }
1402         }
1403         /* Mark the end in the last sg entry if newly added */
1404         if (num_elem > *pages_used)
1405                 sg_mark_end(&to[num_elem - 1]);
1406 out:
1407         if (rc)
1408                 iov_iter_revert(from, size);
1409         *pages_used = num_elem;
1410
1411         return rc;
1412 }
1413
1414 static struct sk_buff *
1415 tls_alloc_clrtxt_skb(struct sock *sk, struct sk_buff *skb,
1416                      unsigned int full_len)
1417 {
1418         struct strp_msg *clr_rxm;
1419         struct sk_buff *clr_skb;
1420         int err;
1421
1422         clr_skb = alloc_skb_with_frags(0, full_len, TLS_PAGE_ORDER,
1423                                        &err, sk->sk_allocation);
1424         if (!clr_skb)
1425                 return NULL;
1426
1427         skb_copy_header(clr_skb, skb);
1428         clr_skb->len = full_len;
1429         clr_skb->data_len = full_len;
1430
1431         clr_rxm = strp_msg(clr_skb);
1432         clr_rxm->offset = 0;
1433
1434         return clr_skb;
1435 }
1436
1437 /* Decrypt handlers
1438  *
1439  * tls_decrypt_sw() and tls_decrypt_device() are decrypt handlers.
1440  * They must transform the darg in/out argument are as follows:
1441  *       |          Input            |         Output
1442  * -------------------------------------------------------------------
1443  *    zc | Zero-copy decrypt allowed | Zero-copy performed
1444  * async | Async decrypt allowed     | Async crypto used / in progress
1445  *   skb |            *              | Output skb
1446  *
1447  * If ZC decryption was performed darg.skb will point to the input skb.
1448  */
1449
1450 /* This function decrypts the input skb into either out_iov or in out_sg
1451  * or in skb buffers itself. The input parameter 'darg->zc' indicates if
1452  * zero-copy mode needs to be tried or not. With zero-copy mode, either
1453  * out_iov or out_sg must be non-NULL. In case both out_iov and out_sg are
1454  * NULL, then the decryption happens inside skb buffers itself, i.e.
1455  * zero-copy gets disabled and 'darg->zc' is updated.
1456  */
1457 static int tls_decrypt_sg(struct sock *sk, struct iov_iter *out_iov,
1458                           struct scatterlist *out_sg,
1459                           struct tls_decrypt_arg *darg)
1460 {
1461         struct tls_context *tls_ctx = tls_get_ctx(sk);
1462         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1463         struct tls_prot_info *prot = &tls_ctx->prot_info;
1464         int n_sgin, n_sgout, aead_size, err, pages = 0;
1465         struct sk_buff *skb = tls_strp_msg(ctx);
1466         const struct strp_msg *rxm = strp_msg(skb);
1467         const struct tls_msg *tlm = tls_msg(skb);
1468         struct aead_request *aead_req;
1469         struct scatterlist *sgin = NULL;
1470         struct scatterlist *sgout = NULL;
1471         const int data_len = rxm->full_len - prot->overhead_size;
1472         int tail_pages = !!prot->tail_size;
1473         struct tls_decrypt_ctx *dctx;
1474         struct sk_buff *clear_skb;
1475         int iv_offset = 0;
1476         u8 *mem;
1477
1478         n_sgin = skb_nsg(skb, rxm->offset + prot->prepend_size,
1479                          rxm->full_len - prot->prepend_size);
1480         if (n_sgin < 1)
1481                 return n_sgin ?: -EBADMSG;
1482
1483         if (darg->zc && (out_iov || out_sg)) {
1484                 clear_skb = NULL;
1485
1486                 if (out_iov)
1487                         n_sgout = 1 + tail_pages +
1488                                 iov_iter_npages_cap(out_iov, INT_MAX, data_len);
1489                 else
1490                         n_sgout = sg_nents(out_sg);
1491         } else {
1492                 darg->zc = false;
1493
1494                 clear_skb = tls_alloc_clrtxt_skb(sk, skb, rxm->full_len);
1495                 if (!clear_skb)
1496                         return -ENOMEM;
1497
1498                 n_sgout = 1 + skb_shinfo(clear_skb)->nr_frags;
1499         }
1500
1501         /* Increment to accommodate AAD */
1502         n_sgin = n_sgin + 1;
1503
1504         /* Allocate a single block of memory which contains
1505          *   aead_req || tls_decrypt_ctx.
1506          * Both structs are variable length.
1507          */
1508         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(ctx->aead_recv);
1509         aead_size = ALIGN(aead_size, __alignof__(*dctx));
1510         mem = kmalloc(aead_size + struct_size(dctx, sg, n_sgin + n_sgout),
1511                       sk->sk_allocation);
1512         if (!mem) {
1513                 err = -ENOMEM;
1514                 goto exit_free_skb;
1515         }
1516
1517         /* Segment the allocated memory */
1518         aead_req = (struct aead_request *)mem;
1519         dctx = (struct tls_decrypt_ctx *)(mem + aead_size);
1520         dctx->sk = sk;
1521         sgin = &dctx->sg[0];
1522         sgout = &dctx->sg[n_sgin];
1523
1524         /* For CCM based ciphers, first byte of nonce+iv is a constant */
1525         switch (prot->cipher_type) {
1526         case TLS_CIPHER_AES_CCM_128:
1527                 dctx->iv[0] = TLS_AES_CCM_IV_B0_BYTE;
1528                 iv_offset = 1;
1529                 break;
1530         case TLS_CIPHER_SM4_CCM:
1531                 dctx->iv[0] = TLS_SM4_CCM_IV_B0_BYTE;
1532                 iv_offset = 1;
1533                 break;
1534         }
1535
1536         /* Prepare IV */
1537         if (prot->version == TLS_1_3_VERSION ||
1538             prot->cipher_type == TLS_CIPHER_CHACHA20_POLY1305) {
1539                 memcpy(&dctx->iv[iv_offset], tls_ctx->rx.iv,
1540                        prot->iv_size + prot->salt_size);
1541         } else {
1542                 err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
1543                                     &dctx->iv[iv_offset] + prot->salt_size,
1544                                     prot->iv_size);
1545                 if (err < 0)
1546                         goto exit_free;
1547                 memcpy(&dctx->iv[iv_offset], tls_ctx->rx.iv, prot->salt_size);
1548         }
1549         tls_xor_iv_with_seq(prot, &dctx->iv[iv_offset], tls_ctx->rx.rec_seq);
1550
1551         /* Prepare AAD */
1552         tls_make_aad(dctx->aad, rxm->full_len - prot->overhead_size +
1553                      prot->tail_size,
1554                      tls_ctx->rx.rec_seq, tlm->control, prot);
1555
1556         /* Prepare sgin */
1557         sg_init_table(sgin, n_sgin);
1558         sg_set_buf(&sgin[0], dctx->aad, prot->aad_size);
1559         err = skb_to_sgvec(skb, &sgin[1],
1560                            rxm->offset + prot->prepend_size,
1561                            rxm->full_len - prot->prepend_size);
1562         if (err < 0)
1563                 goto exit_free;
1564
1565         if (clear_skb) {
1566                 sg_init_table(sgout, n_sgout);
1567                 sg_set_buf(&sgout[0], dctx->aad, prot->aad_size);
1568
1569                 err = skb_to_sgvec(clear_skb, &sgout[1], prot->prepend_size,
1570                                    data_len + prot->tail_size);
1571                 if (err < 0)
1572                         goto exit_free;
1573         } else if (out_iov) {
1574                 sg_init_table(sgout, n_sgout);
1575                 sg_set_buf(&sgout[0], dctx->aad, prot->aad_size);
1576
1577                 err = tls_setup_from_iter(out_iov, data_len, &pages, &sgout[1],
1578                                           (n_sgout - 1 - tail_pages));
1579                 if (err < 0)
1580                         goto exit_free_pages;
1581
1582                 if (prot->tail_size) {
1583                         sg_unmark_end(&sgout[pages]);
1584                         sg_set_buf(&sgout[pages + 1], &dctx->tail,
1585                                    prot->tail_size);
1586                         sg_mark_end(&sgout[pages + 1]);
1587                 }
1588         } else if (out_sg) {
1589                 memcpy(sgout, out_sg, n_sgout * sizeof(*sgout));
1590         }
1591
1592         /* Prepare and submit AEAD request */
1593         err = tls_do_decryption(sk, sgin, sgout, dctx->iv,
1594                                 data_len + prot->tail_size, aead_req, darg);
1595         if (err)
1596                 goto exit_free_pages;
1597
1598         darg->skb = clear_skb ?: tls_strp_msg(ctx);
1599         clear_skb = NULL;
1600
1601         if (unlikely(darg->async)) {
1602                 err = tls_strp_msg_hold(&ctx->strp, &ctx->async_hold);
1603                 if (err)
1604                         __skb_queue_tail(&ctx->async_hold, darg->skb);
1605                 return err;
1606         }
1607
1608         if (prot->tail_size)
1609                 darg->tail = dctx->tail;
1610
1611 exit_free_pages:
1612         /* Release the pages in case iov was mapped to pages */
1613         for (; pages > 0; pages--)
1614                 put_page(sg_page(&sgout[pages]));
1615 exit_free:
1616         kfree(mem);
1617 exit_free_skb:
1618         consume_skb(clear_skb);
1619         return err;
1620 }
1621
1622 static int
1623 tls_decrypt_sw(struct sock *sk, struct tls_context *tls_ctx,
1624                struct msghdr *msg, struct tls_decrypt_arg *darg)
1625 {
1626         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1627         struct tls_prot_info *prot = &tls_ctx->prot_info;
1628         struct strp_msg *rxm;
1629         int pad, err;
1630
1631         err = tls_decrypt_sg(sk, &msg->msg_iter, NULL, darg);
1632         if (err < 0) {
1633                 if (err == -EBADMSG)
1634                         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSDECRYPTERROR);
1635                 return err;
1636         }
1637         /* keep going even for ->async, the code below is TLS 1.3 */
1638
1639         /* If opportunistic TLS 1.3 ZC failed retry without ZC */
1640         if (unlikely(darg->zc && prot->version == TLS_1_3_VERSION &&
1641                      darg->tail != TLS_RECORD_TYPE_DATA)) {
1642                 darg->zc = false;
1643                 if (!darg->tail)
1644                         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSRXNOPADVIOL);
1645                 TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSDECRYPTRETRY);
1646                 return tls_decrypt_sw(sk, tls_ctx, msg, darg);
1647         }
1648
1649         pad = tls_padding_length(prot, darg->skb, darg);
1650         if (pad < 0) {
1651                 if (darg->skb != tls_strp_msg(ctx))
1652                         consume_skb(darg->skb);
1653                 return pad;
1654         }
1655
1656         rxm = strp_msg(darg->skb);
1657         rxm->full_len -= pad;
1658
1659         return 0;
1660 }
1661
1662 static int
1663 tls_decrypt_device(struct sock *sk, struct msghdr *msg,
1664                    struct tls_context *tls_ctx, struct tls_decrypt_arg *darg)
1665 {
1666         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1667         struct tls_prot_info *prot = &tls_ctx->prot_info;
1668         struct strp_msg *rxm;
1669         int pad, err;
1670
1671         if (tls_ctx->rx_conf != TLS_HW)
1672                 return 0;
1673
1674         err = tls_device_decrypted(sk, tls_ctx);
1675         if (err <= 0)
1676                 return err;
1677
1678         pad = tls_padding_length(prot, tls_strp_msg(ctx), darg);
1679         if (pad < 0)
1680                 return pad;
1681
1682         darg->async = false;
1683         darg->skb = tls_strp_msg(ctx);
1684         /* ->zc downgrade check, in case TLS 1.3 gets here */
1685         darg->zc &= !(prot->version == TLS_1_3_VERSION &&
1686                       tls_msg(darg->skb)->control != TLS_RECORD_TYPE_DATA);
1687
1688         rxm = strp_msg(darg->skb);
1689         rxm->full_len -= pad;
1690
1691         if (!darg->zc) {
1692                 /* Non-ZC case needs a real skb */
1693                 darg->skb = tls_strp_msg_detach(ctx);
1694                 if (!darg->skb)
1695                         return -ENOMEM;
1696         } else {
1697                 unsigned int off, len;
1698
1699                 /* In ZC case nobody cares about the output skb.
1700                  * Just copy the data here. Note the skb is not fully trimmed.
1701                  */
1702                 off = rxm->offset + prot->prepend_size;
1703                 len = rxm->full_len - prot->overhead_size;
1704
1705                 err = skb_copy_datagram_msg(darg->skb, off, msg, len);
1706                 if (err)
1707                         return err;
1708         }
1709         return 1;
1710 }
1711
1712 static int tls_rx_one_record(struct sock *sk, struct msghdr *msg,
1713                              struct tls_decrypt_arg *darg)
1714 {
1715         struct tls_context *tls_ctx = tls_get_ctx(sk);
1716         struct tls_prot_info *prot = &tls_ctx->prot_info;
1717         struct strp_msg *rxm;
1718         int err;
1719
1720         err = tls_decrypt_device(sk, msg, tls_ctx, darg);
1721         if (!err)
1722                 err = tls_decrypt_sw(sk, tls_ctx, msg, darg);
1723         if (err < 0)
1724                 return err;
1725
1726         rxm = strp_msg(darg->skb);
1727         rxm->offset += prot->prepend_size;
1728         rxm->full_len -= prot->overhead_size;
1729         tls_advance_record_sn(sk, prot, &tls_ctx->rx);
1730
1731         return 0;
1732 }
1733
1734 int decrypt_skb(struct sock *sk, struct scatterlist *sgout)
1735 {
1736         struct tls_decrypt_arg darg = { .zc = true, };
1737
1738         return tls_decrypt_sg(sk, NULL, sgout, &darg);
1739 }
1740
1741 static int tls_record_content_type(struct msghdr *msg, struct tls_msg *tlm,
1742                                    u8 *control)
1743 {
1744         int err;
1745
1746         if (!*control) {
1747                 *control = tlm->control;
1748                 if (!*control)
1749                         return -EBADMSG;
1750
1751                 err = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1752                                sizeof(*control), control);
1753                 if (*control != TLS_RECORD_TYPE_DATA) {
1754                         if (err || msg->msg_flags & MSG_CTRUNC)
1755                                 return -EIO;
1756                 }
1757         } else if (*control != tlm->control) {
1758                 return 0;
1759         }
1760
1761         return 1;
1762 }
1763
1764 static void tls_rx_rec_done(struct tls_sw_context_rx *ctx)
1765 {
1766         tls_strp_msg_done(&ctx->strp);
1767 }
1768
1769 /* This function traverses the rx_list in tls receive context to copies the
1770  * decrypted records into the buffer provided by caller zero copy is not
1771  * true. Further, the records are removed from the rx_list if it is not a peek
1772  * case and the record has been consumed completely.
1773  */
1774 static int process_rx_list(struct tls_sw_context_rx *ctx,
1775                            struct msghdr *msg,
1776                            u8 *control,
1777                            size_t skip,
1778                            size_t len,
1779                            bool is_peek)
1780 {
1781         struct sk_buff *skb = skb_peek(&ctx->rx_list);
1782         struct tls_msg *tlm;
1783         ssize_t copied = 0;
1784         int err;
1785
1786         while (skip && skb) {
1787                 struct strp_msg *rxm = strp_msg(skb);
1788                 tlm = tls_msg(skb);
1789
1790                 err = tls_record_content_type(msg, tlm, control);
1791                 if (err <= 0)
1792                         goto out;
1793
1794                 if (skip < rxm->full_len)
1795                         break;
1796
1797                 skip = skip - rxm->full_len;
1798                 skb = skb_peek_next(skb, &ctx->rx_list);
1799         }
1800
1801         while (len && skb) {
1802                 struct sk_buff *next_skb;
1803                 struct strp_msg *rxm = strp_msg(skb);
1804                 int chunk = min_t(unsigned int, rxm->full_len - skip, len);
1805
1806                 tlm = tls_msg(skb);
1807
1808                 err = tls_record_content_type(msg, tlm, control);
1809                 if (err <= 0)
1810                         goto out;
1811
1812                 err = skb_copy_datagram_msg(skb, rxm->offset + skip,
1813                                             msg, chunk);
1814                 if (err < 0)
1815                         goto out;
1816
1817                 len = len - chunk;
1818                 copied = copied + chunk;
1819
1820                 /* Consume the data from record if it is non-peek case*/
1821                 if (!is_peek) {
1822                         rxm->offset = rxm->offset + chunk;
1823                         rxm->full_len = rxm->full_len - chunk;
1824
1825                         /* Return if there is unconsumed data in the record */
1826                         if (rxm->full_len - skip)
1827                                 break;
1828                 }
1829
1830                 /* The remaining skip-bytes must lie in 1st record in rx_list.
1831                  * So from the 2nd record, 'skip' should be 0.
1832                  */
1833                 skip = 0;
1834
1835                 if (msg)
1836                         msg->msg_flags |= MSG_EOR;
1837
1838                 next_skb = skb_peek_next(skb, &ctx->rx_list);
1839
1840                 if (!is_peek) {
1841                         __skb_unlink(skb, &ctx->rx_list);
1842                         consume_skb(skb);
1843                 }
1844
1845                 skb = next_skb;
1846         }
1847         err = 0;
1848
1849 out:
1850         return copied ? : err;
1851 }
1852
1853 static bool
1854 tls_read_flush_backlog(struct sock *sk, struct tls_prot_info *prot,
1855                        size_t len_left, size_t decrypted, ssize_t done,
1856                        size_t *flushed_at)
1857 {
1858         size_t max_rec;
1859
1860         if (len_left <= decrypted)
1861                 return false;
1862
1863         max_rec = prot->overhead_size - prot->tail_size + TLS_MAX_PAYLOAD_SIZE;
1864         if (done - *flushed_at < SZ_128K && tcp_inq(sk) > max_rec)
1865                 return false;
1866
1867         *flushed_at = done;
1868         return sk_flush_backlog(sk);
1869 }
1870
1871 static int tls_rx_reader_lock(struct sock *sk, struct tls_sw_context_rx *ctx,
1872                               bool nonblock)
1873 {
1874         long timeo;
1875         int err;
1876
1877         lock_sock(sk);
1878
1879         timeo = sock_rcvtimeo(sk, nonblock);
1880
1881         while (unlikely(ctx->reader_present)) {
1882                 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1883
1884                 ctx->reader_contended = 1;
1885
1886                 add_wait_queue(&ctx->wq, &wait);
1887                 sk_wait_event(sk, &timeo,
1888                               !READ_ONCE(ctx->reader_present), &wait);
1889                 remove_wait_queue(&ctx->wq, &wait);
1890
1891                 if (timeo <= 0) {
1892                         err = -EAGAIN;
1893                         goto err_unlock;
1894                 }
1895                 if (signal_pending(current)) {
1896                         err = sock_intr_errno(timeo);
1897                         goto err_unlock;
1898                 }
1899         }
1900
1901         WRITE_ONCE(ctx->reader_present, 1);
1902
1903         return 0;
1904
1905 err_unlock:
1906         release_sock(sk);
1907         return err;
1908 }
1909
1910 static void tls_rx_reader_unlock(struct sock *sk, struct tls_sw_context_rx *ctx)
1911 {
1912         if (unlikely(ctx->reader_contended)) {
1913                 if (wq_has_sleeper(&ctx->wq))
1914                         wake_up(&ctx->wq);
1915                 else
1916                         ctx->reader_contended = 0;
1917
1918                 WARN_ON_ONCE(!ctx->reader_present);
1919         }
1920
1921         WRITE_ONCE(ctx->reader_present, 0);
1922         release_sock(sk);
1923 }
1924
1925 int tls_sw_recvmsg(struct sock *sk,
1926                    struct msghdr *msg,
1927                    size_t len,
1928                    int flags,
1929                    int *addr_len)
1930 {
1931         struct tls_context *tls_ctx = tls_get_ctx(sk);
1932         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1933         struct tls_prot_info *prot = &tls_ctx->prot_info;
1934         ssize_t decrypted = 0, async_copy_bytes = 0;
1935         struct sk_psock *psock;
1936         unsigned char control = 0;
1937         size_t flushed_at = 0;
1938         struct strp_msg *rxm;
1939         struct tls_msg *tlm;
1940         ssize_t copied = 0;
1941         bool async = false;
1942         int target, err;
1943         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
1944         bool is_peek = flags & MSG_PEEK;
1945         bool released = true;
1946         bool bpf_strp_enabled;
1947         bool zc_capable;
1948
1949         if (unlikely(flags & MSG_ERRQUEUE))
1950                 return sock_recv_errqueue(sk, msg, len, SOL_IP, IP_RECVERR);
1951
1952         psock = sk_psock_get(sk);
1953         err = tls_rx_reader_lock(sk, ctx, flags & MSG_DONTWAIT);
1954         if (err < 0)
1955                 return err;
1956         bpf_strp_enabled = sk_psock_strp_enabled(psock);
1957
1958         /* If crypto failed the connection is broken */
1959         err = ctx->async_wait.err;
1960         if (err)
1961                 goto end;
1962
1963         /* Process pending decrypted records. It must be non-zero-copy */
1964         err = process_rx_list(ctx, msg, &control, 0, len, is_peek);
1965         if (err < 0)
1966                 goto end;
1967
1968         copied = err;
1969         if (len <= copied)
1970                 goto end;
1971
1972         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1973         len = len - copied;
1974
1975         zc_capable = !bpf_strp_enabled && !is_kvec && !is_peek &&
1976                 ctx->zc_capable;
1977         decrypted = 0;
1978         while (len && (decrypted + copied < target || tls_strp_msg_ready(ctx))) {
1979                 struct tls_decrypt_arg darg;
1980                 int to_decrypt, chunk;
1981
1982                 err = tls_rx_rec_wait(sk, psock, flags & MSG_DONTWAIT,
1983                                       released);
1984                 if (err <= 0) {
1985                         if (psock) {
1986                                 chunk = sk_msg_recvmsg(sk, psock, msg, len,
1987                                                        flags);
1988                                 if (chunk > 0) {
1989                                         decrypted += chunk;
1990                                         len -= chunk;
1991                                         continue;
1992                                 }
1993                         }
1994                         goto recv_end;
1995                 }
1996
1997                 memset(&darg.inargs, 0, sizeof(darg.inargs));
1998
1999                 rxm = strp_msg(tls_strp_msg(ctx));
2000                 tlm = tls_msg(tls_strp_msg(ctx));
2001
2002                 to_decrypt = rxm->full_len - prot->overhead_size;
2003
2004                 if (zc_capable && to_decrypt <= len &&
2005                     tlm->control == TLS_RECORD_TYPE_DATA)
2006                         darg.zc = true;
2007
2008                 /* Do not use async mode if record is non-data */
2009                 if (tlm->control == TLS_RECORD_TYPE_DATA && !bpf_strp_enabled)
2010                         darg.async = ctx->async_capable;
2011                 else
2012                         darg.async = false;
2013
2014                 err = tls_rx_one_record(sk, msg, &darg);
2015                 if (err < 0) {
2016                         tls_err_abort(sk, -EBADMSG);
2017                         goto recv_end;
2018                 }
2019
2020                 async |= darg.async;
2021
2022                 /* If the type of records being processed is not known yet,
2023                  * set it to record type just dequeued. If it is already known,
2024                  * but does not match the record type just dequeued, go to end.
2025                  * We always get record type here since for tls1.2, record type
2026                  * is known just after record is dequeued from stream parser.
2027                  * For tls1.3, we disable async.
2028                  */
2029                 err = tls_record_content_type(msg, tls_msg(darg.skb), &control);
2030                 if (err <= 0) {
2031                         DEBUG_NET_WARN_ON_ONCE(darg.zc);
2032                         tls_rx_rec_done(ctx);
2033 put_on_rx_list_err:
2034                         __skb_queue_tail(&ctx->rx_list, darg.skb);
2035                         goto recv_end;
2036                 }
2037
2038                 /* periodically flush backlog, and feed strparser */
2039                 released = tls_read_flush_backlog(sk, prot, len, to_decrypt,
2040                                                   decrypted + copied,
2041                                                   &flushed_at);
2042
2043                 /* TLS 1.3 may have updated the length by more than overhead */
2044                 rxm = strp_msg(darg.skb);
2045                 chunk = rxm->full_len;
2046                 tls_rx_rec_done(ctx);
2047
2048                 if (!darg.zc) {
2049                         bool partially_consumed = chunk > len;
2050                         struct sk_buff *skb = darg.skb;
2051
2052                         DEBUG_NET_WARN_ON_ONCE(darg.skb == ctx->strp.anchor);
2053
2054                         if (async) {
2055                                 /* TLS 1.2-only, to_decrypt must be text len */
2056                                 chunk = min_t(int, to_decrypt, len);
2057                                 async_copy_bytes += chunk;
2058 put_on_rx_list:
2059                                 decrypted += chunk;
2060                                 len -= chunk;
2061                                 __skb_queue_tail(&ctx->rx_list, skb);
2062                                 continue;
2063                         }
2064
2065                         if (bpf_strp_enabled) {
2066                                 released = true;
2067                                 err = sk_psock_tls_strp_read(psock, skb);
2068                                 if (err != __SK_PASS) {
2069                                         rxm->offset = rxm->offset + rxm->full_len;
2070                                         rxm->full_len = 0;
2071                                         if (err == __SK_DROP)
2072                                                 consume_skb(skb);
2073                                         continue;
2074                                 }
2075                         }
2076
2077                         if (partially_consumed)
2078                                 chunk = len;
2079
2080                         err = skb_copy_datagram_msg(skb, rxm->offset,
2081                                                     msg, chunk);
2082                         if (err < 0)
2083                                 goto put_on_rx_list_err;
2084
2085                         if (is_peek)
2086                                 goto put_on_rx_list;
2087
2088                         if (partially_consumed) {
2089                                 rxm->offset += chunk;
2090                                 rxm->full_len -= chunk;
2091                                 goto put_on_rx_list;
2092                         }
2093
2094                         consume_skb(skb);
2095                 }
2096
2097                 decrypted += chunk;
2098                 len -= chunk;
2099
2100                 /* Return full control message to userspace before trying
2101                  * to parse another message type
2102                  */
2103                 msg->msg_flags |= MSG_EOR;
2104                 if (control != TLS_RECORD_TYPE_DATA)
2105                         break;
2106         }
2107
2108 recv_end:
2109         if (async) {
2110                 int ret, pending;
2111
2112                 /* Wait for all previously submitted records to be decrypted */
2113                 spin_lock_bh(&ctx->decrypt_compl_lock);
2114                 reinit_completion(&ctx->async_wait.completion);
2115                 pending = atomic_read(&ctx->decrypt_pending);
2116                 spin_unlock_bh(&ctx->decrypt_compl_lock);
2117                 ret = 0;
2118                 if (pending)
2119                         ret = crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
2120                 __skb_queue_purge(&ctx->async_hold);
2121
2122                 if (ret) {
2123                         if (err >= 0 || err == -EINPROGRESS)
2124                                 err = ret;
2125                         decrypted = 0;
2126                         goto end;
2127                 }
2128
2129                 /* Drain records from the rx_list & copy if required */
2130                 if (is_peek || is_kvec)
2131                         err = process_rx_list(ctx, msg, &control, copied,
2132                                               decrypted, is_peek);
2133                 else
2134                         err = process_rx_list(ctx, msg, &control, 0,
2135                                               async_copy_bytes, is_peek);
2136                 decrypted += max(err, 0);
2137         }
2138
2139         copied += decrypted;
2140
2141 end:
2142         tls_rx_reader_unlock(sk, ctx);
2143         if (psock)
2144                 sk_psock_put(sk, psock);
2145         return copied ? : err;
2146 }
2147
2148 ssize_t tls_sw_splice_read(struct socket *sock,  loff_t *ppos,
2149                            struct pipe_inode_info *pipe,
2150                            size_t len, unsigned int flags)
2151 {
2152         struct tls_context *tls_ctx = tls_get_ctx(sock->sk);
2153         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2154         struct strp_msg *rxm = NULL;
2155         struct sock *sk = sock->sk;
2156         struct tls_msg *tlm;
2157         struct sk_buff *skb;
2158         ssize_t copied = 0;
2159         int chunk;
2160         int err;
2161
2162         err = tls_rx_reader_lock(sk, ctx, flags & SPLICE_F_NONBLOCK);
2163         if (err < 0)
2164                 return err;
2165
2166         if (!skb_queue_empty(&ctx->rx_list)) {
2167                 skb = __skb_dequeue(&ctx->rx_list);
2168         } else {
2169                 struct tls_decrypt_arg darg;
2170
2171                 err = tls_rx_rec_wait(sk, NULL, flags & SPLICE_F_NONBLOCK,
2172                                       true);
2173                 if (err <= 0)
2174                         goto splice_read_end;
2175
2176                 memset(&darg.inargs, 0, sizeof(darg.inargs));
2177
2178                 err = tls_rx_one_record(sk, NULL, &darg);
2179                 if (err < 0) {
2180                         tls_err_abort(sk, -EBADMSG);
2181                         goto splice_read_end;
2182                 }
2183
2184                 tls_rx_rec_done(ctx);
2185                 skb = darg.skb;
2186         }
2187
2188         rxm = strp_msg(skb);
2189         tlm = tls_msg(skb);
2190
2191         /* splice does not support reading control messages */
2192         if (tlm->control != TLS_RECORD_TYPE_DATA) {
2193                 err = -EINVAL;
2194                 goto splice_requeue;
2195         }
2196
2197         chunk = min_t(unsigned int, rxm->full_len, len);
2198         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, chunk, flags);
2199         if (copied < 0)
2200                 goto splice_requeue;
2201
2202         if (chunk < rxm->full_len) {
2203                 rxm->offset += len;
2204                 rxm->full_len -= len;
2205                 goto splice_requeue;
2206         }
2207
2208         consume_skb(skb);
2209
2210 splice_read_end:
2211         tls_rx_reader_unlock(sk, ctx);
2212         return copied ? : err;
2213
2214 splice_requeue:
2215         __skb_queue_head(&ctx->rx_list, skb);
2216         goto splice_read_end;
2217 }
2218
2219 bool tls_sw_sock_is_readable(struct sock *sk)
2220 {
2221         struct tls_context *tls_ctx = tls_get_ctx(sk);
2222         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2223         bool ingress_empty = true;
2224         struct sk_psock *psock;
2225
2226         rcu_read_lock();
2227         psock = sk_psock(sk);
2228         if (psock)
2229                 ingress_empty = list_empty(&psock->ingress_msg);
2230         rcu_read_unlock();
2231
2232         return !ingress_empty || tls_strp_msg_ready(ctx) ||
2233                 !skb_queue_empty(&ctx->rx_list);
2234 }
2235
2236 int tls_rx_msg_size(struct tls_strparser *strp, struct sk_buff *skb)
2237 {
2238         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2239         struct tls_prot_info *prot = &tls_ctx->prot_info;
2240         char header[TLS_HEADER_SIZE + MAX_IV_SIZE];
2241         size_t cipher_overhead;
2242         size_t data_len = 0;
2243         int ret;
2244
2245         /* Verify that we have a full TLS header, or wait for more data */
2246         if (strp->stm.offset + prot->prepend_size > skb->len)
2247                 return 0;
2248
2249         /* Sanity-check size of on-stack buffer. */
2250         if (WARN_ON(prot->prepend_size > sizeof(header))) {
2251                 ret = -EINVAL;
2252                 goto read_failure;
2253         }
2254
2255         /* Linearize header to local buffer */
2256         ret = skb_copy_bits(skb, strp->stm.offset, header, prot->prepend_size);
2257         if (ret < 0)
2258                 goto read_failure;
2259
2260         strp->mark = header[0];
2261
2262         data_len = ((header[4] & 0xFF) | (header[3] << 8));
2263
2264         cipher_overhead = prot->tag_size;
2265         if (prot->version != TLS_1_3_VERSION &&
2266             prot->cipher_type != TLS_CIPHER_CHACHA20_POLY1305)
2267                 cipher_overhead += prot->iv_size;
2268
2269         if (data_len > TLS_MAX_PAYLOAD_SIZE + cipher_overhead +
2270             prot->tail_size) {
2271                 ret = -EMSGSIZE;
2272                 goto read_failure;
2273         }
2274         if (data_len < cipher_overhead) {
2275                 ret = -EBADMSG;
2276                 goto read_failure;
2277         }
2278
2279         /* Note that both TLS1.3 and TLS1.2 use TLS_1_2 version here */
2280         if (header[1] != TLS_1_2_VERSION_MINOR ||
2281             header[2] != TLS_1_2_VERSION_MAJOR) {
2282                 ret = -EINVAL;
2283                 goto read_failure;
2284         }
2285
2286         tls_device_rx_resync_new_rec(strp->sk, data_len + TLS_HEADER_SIZE,
2287                                      TCP_SKB_CB(skb)->seq + strp->stm.offset);
2288         return data_len + TLS_HEADER_SIZE;
2289
2290 read_failure:
2291         tls_err_abort(strp->sk, ret);
2292
2293         return ret;
2294 }
2295
2296 void tls_rx_msg_ready(struct tls_strparser *strp)
2297 {
2298         struct tls_sw_context_rx *ctx;
2299
2300         ctx = container_of(strp, struct tls_sw_context_rx, strp);
2301         ctx->saved_data_ready(strp->sk);
2302 }
2303
2304 static void tls_data_ready(struct sock *sk)
2305 {
2306         struct tls_context *tls_ctx = tls_get_ctx(sk);
2307         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2308         struct sk_psock *psock;
2309         gfp_t alloc_save;
2310
2311         trace_sk_data_ready(sk);
2312
2313         alloc_save = sk->sk_allocation;
2314         sk->sk_allocation = GFP_ATOMIC;
2315         tls_strp_data_ready(&ctx->strp);
2316         sk->sk_allocation = alloc_save;
2317
2318         psock = sk_psock_get(sk);
2319         if (psock) {
2320                 if (!list_empty(&psock->ingress_msg))
2321                         ctx->saved_data_ready(sk);
2322                 sk_psock_put(sk, psock);
2323         }
2324 }
2325
2326 void tls_sw_cancel_work_tx(struct tls_context *tls_ctx)
2327 {
2328         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2329
2330         set_bit(BIT_TX_CLOSING, &ctx->tx_bitmask);
2331         set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask);
2332         cancel_delayed_work_sync(&ctx->tx_work.work);
2333 }
2334
2335 void tls_sw_release_resources_tx(struct sock *sk)
2336 {
2337         struct tls_context *tls_ctx = tls_get_ctx(sk);
2338         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2339         struct tls_rec *rec, *tmp;
2340         int pending;
2341
2342         /* Wait for any pending async encryptions to complete */
2343         spin_lock_bh(&ctx->encrypt_compl_lock);
2344         ctx->async_notify = true;
2345         pending = atomic_read(&ctx->encrypt_pending);
2346         spin_unlock_bh(&ctx->encrypt_compl_lock);
2347
2348         if (pending)
2349                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
2350
2351         tls_tx_records(sk, -1);
2352
2353         /* Free up un-sent records in tx_list. First, free
2354          * the partially sent record if any at head of tx_list.
2355          */
2356         if (tls_ctx->partially_sent_record) {
2357                 tls_free_partial_record(sk, tls_ctx);
2358                 rec = list_first_entry(&ctx->tx_list,
2359                                        struct tls_rec, list);
2360                 list_del(&rec->list);
2361                 sk_msg_free(sk, &rec->msg_plaintext);
2362                 kfree(rec);
2363         }
2364
2365         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
2366                 list_del(&rec->list);
2367                 sk_msg_free(sk, &rec->msg_encrypted);
2368                 sk_msg_free(sk, &rec->msg_plaintext);
2369                 kfree(rec);
2370         }
2371
2372         crypto_free_aead(ctx->aead_send);
2373         tls_free_open_rec(sk);
2374 }
2375
2376 void tls_sw_free_ctx_tx(struct tls_context *tls_ctx)
2377 {
2378         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2379
2380         kfree(ctx);
2381 }
2382
2383 void tls_sw_release_resources_rx(struct sock *sk)
2384 {
2385         struct tls_context *tls_ctx = tls_get_ctx(sk);
2386         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2387
2388         kfree(tls_ctx->rx.rec_seq);
2389         kfree(tls_ctx->rx.iv);
2390
2391         if (ctx->aead_recv) {
2392                 __skb_queue_purge(&ctx->rx_list);
2393                 crypto_free_aead(ctx->aead_recv);
2394                 tls_strp_stop(&ctx->strp);
2395                 /* If tls_sw_strparser_arm() was not called (cleanup paths)
2396                  * we still want to tls_strp_stop(), but sk->sk_data_ready was
2397                  * never swapped.
2398                  */
2399                 if (ctx->saved_data_ready) {
2400                         write_lock_bh(&sk->sk_callback_lock);
2401                         sk->sk_data_ready = ctx->saved_data_ready;
2402                         write_unlock_bh(&sk->sk_callback_lock);
2403                 }
2404         }
2405 }
2406
2407 void tls_sw_strparser_done(struct tls_context *tls_ctx)
2408 {
2409         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2410
2411         tls_strp_done(&ctx->strp);
2412 }
2413
2414 void tls_sw_free_ctx_rx(struct tls_context *tls_ctx)
2415 {
2416         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2417
2418         kfree(ctx);
2419 }
2420
2421 void tls_sw_free_resources_rx(struct sock *sk)
2422 {
2423         struct tls_context *tls_ctx = tls_get_ctx(sk);
2424
2425         tls_sw_release_resources_rx(sk);
2426         tls_sw_free_ctx_rx(tls_ctx);
2427 }
2428
2429 /* The work handler to transmitt the encrypted records in tx_list */
2430 static void tx_work_handler(struct work_struct *work)
2431 {
2432         struct delayed_work *delayed_work = to_delayed_work(work);
2433         struct tx_work *tx_work = container_of(delayed_work,
2434                                                struct tx_work, work);
2435         struct sock *sk = tx_work->sk;
2436         struct tls_context *tls_ctx = tls_get_ctx(sk);
2437         struct tls_sw_context_tx *ctx;
2438
2439         if (unlikely(!tls_ctx))
2440                 return;
2441
2442         ctx = tls_sw_ctx_tx(tls_ctx);
2443         if (test_bit(BIT_TX_CLOSING, &ctx->tx_bitmask))
2444                 return;
2445
2446         if (!test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
2447                 return;
2448
2449         if (mutex_trylock(&tls_ctx->tx_lock)) {
2450                 lock_sock(sk);
2451                 tls_tx_records(sk, -1);
2452                 release_sock(sk);
2453                 mutex_unlock(&tls_ctx->tx_lock);
2454         } else if (!test_and_set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
2455                 /* Someone is holding the tx_lock, they will likely run Tx
2456                  * and cancel the work on their way out of the lock section.
2457                  * Schedule a long delay just in case.
2458                  */
2459                 schedule_delayed_work(&ctx->tx_work.work, msecs_to_jiffies(10));
2460         }
2461 }
2462
2463 static bool tls_is_tx_ready(struct tls_sw_context_tx *ctx)
2464 {
2465         struct tls_rec *rec;
2466
2467         rec = list_first_entry_or_null(&ctx->tx_list, struct tls_rec, list);
2468         if (!rec)
2469                 return false;
2470
2471         return READ_ONCE(rec->tx_ready);
2472 }
2473
2474 void tls_sw_write_space(struct sock *sk, struct tls_context *ctx)
2475 {
2476         struct tls_sw_context_tx *tx_ctx = tls_sw_ctx_tx(ctx);
2477
2478         /* Schedule the transmission if tx list is ready */
2479         if (tls_is_tx_ready(tx_ctx) &&
2480             !test_and_set_bit(BIT_TX_SCHEDULED, &tx_ctx->tx_bitmask))
2481                 schedule_delayed_work(&tx_ctx->tx_work.work, 0);
2482 }
2483
2484 void tls_sw_strparser_arm(struct sock *sk, struct tls_context *tls_ctx)
2485 {
2486         struct tls_sw_context_rx *rx_ctx = tls_sw_ctx_rx(tls_ctx);
2487
2488         write_lock_bh(&sk->sk_callback_lock);
2489         rx_ctx->saved_data_ready = sk->sk_data_ready;
2490         sk->sk_data_ready = tls_data_ready;
2491         write_unlock_bh(&sk->sk_callback_lock);
2492 }
2493
2494 void tls_update_rx_zc_capable(struct tls_context *tls_ctx)
2495 {
2496         struct tls_sw_context_rx *rx_ctx = tls_sw_ctx_rx(tls_ctx);
2497
2498         rx_ctx->zc_capable = tls_ctx->rx_no_pad ||
2499                 tls_ctx->prot_info.version != TLS_1_3_VERSION;
2500 }
2501
2502 int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
2503 {
2504         struct tls_context *tls_ctx = tls_get_ctx(sk);
2505         struct tls_prot_info *prot = &tls_ctx->prot_info;
2506         struct tls_crypto_info *crypto_info;
2507         struct tls_sw_context_tx *sw_ctx_tx = NULL;
2508         struct tls_sw_context_rx *sw_ctx_rx = NULL;
2509         struct cipher_context *cctx;
2510         struct crypto_aead **aead;
2511         u16 nonce_size, tag_size, iv_size, rec_seq_size, salt_size;
2512         struct crypto_tfm *tfm;
2513         char *iv, *rec_seq, *key, *salt, *cipher_name;
2514         size_t keysize;
2515         int rc = 0;
2516
2517         if (!ctx) {
2518                 rc = -EINVAL;
2519                 goto out;
2520         }
2521
2522         if (tx) {
2523                 if (!ctx->priv_ctx_tx) {
2524                         sw_ctx_tx = kzalloc(sizeof(*sw_ctx_tx), GFP_KERNEL);
2525                         if (!sw_ctx_tx) {
2526                                 rc = -ENOMEM;
2527                                 goto out;
2528                         }
2529                         ctx->priv_ctx_tx = sw_ctx_tx;
2530                 } else {
2531                         sw_ctx_tx =
2532                                 (struct tls_sw_context_tx *)ctx->priv_ctx_tx;
2533                 }
2534         } else {
2535                 if (!ctx->priv_ctx_rx) {
2536                         sw_ctx_rx = kzalloc(sizeof(*sw_ctx_rx), GFP_KERNEL);
2537                         if (!sw_ctx_rx) {
2538                                 rc = -ENOMEM;
2539                                 goto out;
2540                         }
2541                         ctx->priv_ctx_rx = sw_ctx_rx;
2542                 } else {
2543                         sw_ctx_rx =
2544                                 (struct tls_sw_context_rx *)ctx->priv_ctx_rx;
2545                 }
2546         }
2547
2548         if (tx) {
2549                 crypto_init_wait(&sw_ctx_tx->async_wait);
2550                 spin_lock_init(&sw_ctx_tx->encrypt_compl_lock);
2551                 crypto_info = &ctx->crypto_send.info;
2552                 cctx = &ctx->tx;
2553                 aead = &sw_ctx_tx->aead_send;
2554                 INIT_LIST_HEAD(&sw_ctx_tx->tx_list);
2555                 INIT_DELAYED_WORK(&sw_ctx_tx->tx_work.work, tx_work_handler);
2556                 sw_ctx_tx->tx_work.sk = sk;
2557         } else {
2558                 crypto_init_wait(&sw_ctx_rx->async_wait);
2559                 spin_lock_init(&sw_ctx_rx->decrypt_compl_lock);
2560                 init_waitqueue_head(&sw_ctx_rx->wq);
2561                 crypto_info = &ctx->crypto_recv.info;
2562                 cctx = &ctx->rx;
2563                 skb_queue_head_init(&sw_ctx_rx->rx_list);
2564                 skb_queue_head_init(&sw_ctx_rx->async_hold);
2565                 aead = &sw_ctx_rx->aead_recv;
2566         }
2567
2568         switch (crypto_info->cipher_type) {
2569         case TLS_CIPHER_AES_GCM_128: {
2570                 struct tls12_crypto_info_aes_gcm_128 *gcm_128_info;
2571
2572                 gcm_128_info = (void *)crypto_info;
2573                 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2574                 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
2575                 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2576                 iv = gcm_128_info->iv;
2577                 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
2578                 rec_seq = gcm_128_info->rec_seq;
2579                 keysize = TLS_CIPHER_AES_GCM_128_KEY_SIZE;
2580                 key = gcm_128_info->key;
2581                 salt = gcm_128_info->salt;
2582                 salt_size = TLS_CIPHER_AES_GCM_128_SALT_SIZE;
2583                 cipher_name = "gcm(aes)";
2584                 break;
2585         }
2586         case TLS_CIPHER_AES_GCM_256: {
2587                 struct tls12_crypto_info_aes_gcm_256 *gcm_256_info;
2588
2589                 gcm_256_info = (void *)crypto_info;
2590                 nonce_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2591                 tag_size = TLS_CIPHER_AES_GCM_256_TAG_SIZE;
2592                 iv_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2593                 iv = gcm_256_info->iv;
2594                 rec_seq_size = TLS_CIPHER_AES_GCM_256_REC_SEQ_SIZE;
2595                 rec_seq = gcm_256_info->rec_seq;
2596                 keysize = TLS_CIPHER_AES_GCM_256_KEY_SIZE;
2597                 key = gcm_256_info->key;
2598                 salt = gcm_256_info->salt;
2599                 salt_size = TLS_CIPHER_AES_GCM_256_SALT_SIZE;
2600                 cipher_name = "gcm(aes)";
2601                 break;
2602         }
2603         case TLS_CIPHER_AES_CCM_128: {
2604                 struct tls12_crypto_info_aes_ccm_128 *ccm_128_info;
2605
2606                 ccm_128_info = (void *)crypto_info;
2607                 nonce_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2608                 tag_size = TLS_CIPHER_AES_CCM_128_TAG_SIZE;
2609                 iv_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2610                 iv = ccm_128_info->iv;
2611                 rec_seq_size = TLS_CIPHER_AES_CCM_128_REC_SEQ_SIZE;
2612                 rec_seq = ccm_128_info->rec_seq;
2613                 keysize = TLS_CIPHER_AES_CCM_128_KEY_SIZE;
2614                 key = ccm_128_info->key;
2615                 salt = ccm_128_info->salt;
2616                 salt_size = TLS_CIPHER_AES_CCM_128_SALT_SIZE;
2617                 cipher_name = "ccm(aes)";
2618                 break;
2619         }
2620         case TLS_CIPHER_CHACHA20_POLY1305: {
2621                 struct tls12_crypto_info_chacha20_poly1305 *chacha20_poly1305_info;
2622
2623                 chacha20_poly1305_info = (void *)crypto_info;
2624                 nonce_size = 0;
2625                 tag_size = TLS_CIPHER_CHACHA20_POLY1305_TAG_SIZE;
2626                 iv_size = TLS_CIPHER_CHACHA20_POLY1305_IV_SIZE;
2627                 iv = chacha20_poly1305_info->iv;
2628                 rec_seq_size = TLS_CIPHER_CHACHA20_POLY1305_REC_SEQ_SIZE;
2629                 rec_seq = chacha20_poly1305_info->rec_seq;
2630                 keysize = TLS_CIPHER_CHACHA20_POLY1305_KEY_SIZE;
2631                 key = chacha20_poly1305_info->key;
2632                 salt = chacha20_poly1305_info->salt;
2633                 salt_size = TLS_CIPHER_CHACHA20_POLY1305_SALT_SIZE;
2634                 cipher_name = "rfc7539(chacha20,poly1305)";
2635                 break;
2636         }
2637         case TLS_CIPHER_SM4_GCM: {
2638                 struct tls12_crypto_info_sm4_gcm *sm4_gcm_info;
2639
2640                 sm4_gcm_info = (void *)crypto_info;
2641                 nonce_size = TLS_CIPHER_SM4_GCM_IV_SIZE;
2642                 tag_size = TLS_CIPHER_SM4_GCM_TAG_SIZE;
2643                 iv_size = TLS_CIPHER_SM4_GCM_IV_SIZE;
2644                 iv = sm4_gcm_info->iv;
2645                 rec_seq_size = TLS_CIPHER_SM4_GCM_REC_SEQ_SIZE;
2646                 rec_seq = sm4_gcm_info->rec_seq;
2647                 keysize = TLS_CIPHER_SM4_GCM_KEY_SIZE;
2648                 key = sm4_gcm_info->key;
2649                 salt = sm4_gcm_info->salt;
2650                 salt_size = TLS_CIPHER_SM4_GCM_SALT_SIZE;
2651                 cipher_name = "gcm(sm4)";
2652                 break;
2653         }
2654         case TLS_CIPHER_SM4_CCM: {
2655                 struct tls12_crypto_info_sm4_ccm *sm4_ccm_info;
2656
2657                 sm4_ccm_info = (void *)crypto_info;
2658                 nonce_size = TLS_CIPHER_SM4_CCM_IV_SIZE;
2659                 tag_size = TLS_CIPHER_SM4_CCM_TAG_SIZE;
2660                 iv_size = TLS_CIPHER_SM4_CCM_IV_SIZE;
2661                 iv = sm4_ccm_info->iv;
2662                 rec_seq_size = TLS_CIPHER_SM4_CCM_REC_SEQ_SIZE;
2663                 rec_seq = sm4_ccm_info->rec_seq;
2664                 keysize = TLS_CIPHER_SM4_CCM_KEY_SIZE;
2665                 key = sm4_ccm_info->key;
2666                 salt = sm4_ccm_info->salt;
2667                 salt_size = TLS_CIPHER_SM4_CCM_SALT_SIZE;
2668                 cipher_name = "ccm(sm4)";
2669                 break;
2670         }
2671         case TLS_CIPHER_ARIA_GCM_128: {
2672                 struct tls12_crypto_info_aria_gcm_128 *aria_gcm_128_info;
2673
2674                 aria_gcm_128_info = (void *)crypto_info;
2675                 nonce_size = TLS_CIPHER_ARIA_GCM_128_IV_SIZE;
2676                 tag_size = TLS_CIPHER_ARIA_GCM_128_TAG_SIZE;
2677                 iv_size = TLS_CIPHER_ARIA_GCM_128_IV_SIZE;
2678                 iv = aria_gcm_128_info->iv;
2679                 rec_seq_size = TLS_CIPHER_ARIA_GCM_128_REC_SEQ_SIZE;
2680                 rec_seq = aria_gcm_128_info->rec_seq;
2681                 keysize = TLS_CIPHER_ARIA_GCM_128_KEY_SIZE;
2682                 key = aria_gcm_128_info->key;
2683                 salt = aria_gcm_128_info->salt;
2684                 salt_size = TLS_CIPHER_ARIA_GCM_128_SALT_SIZE;
2685                 cipher_name = "gcm(aria)";
2686                 break;
2687         }
2688         case TLS_CIPHER_ARIA_GCM_256: {
2689                 struct tls12_crypto_info_aria_gcm_256 *gcm_256_info;
2690
2691                 gcm_256_info = (void *)crypto_info;
2692                 nonce_size = TLS_CIPHER_ARIA_GCM_256_IV_SIZE;
2693                 tag_size = TLS_CIPHER_ARIA_GCM_256_TAG_SIZE;
2694                 iv_size = TLS_CIPHER_ARIA_GCM_256_IV_SIZE;
2695                 iv = gcm_256_info->iv;
2696                 rec_seq_size = TLS_CIPHER_ARIA_GCM_256_REC_SEQ_SIZE;
2697                 rec_seq = gcm_256_info->rec_seq;
2698                 keysize = TLS_CIPHER_ARIA_GCM_256_KEY_SIZE;
2699                 key = gcm_256_info->key;
2700                 salt = gcm_256_info->salt;
2701                 salt_size = TLS_CIPHER_ARIA_GCM_256_SALT_SIZE;
2702                 cipher_name = "gcm(aria)";
2703                 break;
2704         }
2705         default:
2706                 rc = -EINVAL;
2707                 goto free_priv;
2708         }
2709
2710         if (crypto_info->version == TLS_1_3_VERSION) {
2711                 nonce_size = 0;
2712                 prot->aad_size = TLS_HEADER_SIZE;
2713                 prot->tail_size = 1;
2714         } else {
2715                 prot->aad_size = TLS_AAD_SPACE_SIZE;
2716                 prot->tail_size = 0;
2717         }
2718
2719         /* Sanity-check the sizes for stack allocations. */
2720         if (iv_size > MAX_IV_SIZE || nonce_size > MAX_IV_SIZE ||
2721             rec_seq_size > TLS_MAX_REC_SEQ_SIZE || tag_size != TLS_TAG_SIZE ||
2722             prot->aad_size > TLS_MAX_AAD_SIZE) {
2723                 rc = -EINVAL;
2724                 goto free_priv;
2725         }
2726
2727         prot->version = crypto_info->version;
2728         prot->cipher_type = crypto_info->cipher_type;
2729         prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
2730         prot->tag_size = tag_size;
2731         prot->overhead_size = prot->prepend_size +
2732                               prot->tag_size + prot->tail_size;
2733         prot->iv_size = iv_size;
2734         prot->salt_size = salt_size;
2735         cctx->iv = kmalloc(iv_size + salt_size, GFP_KERNEL);
2736         if (!cctx->iv) {
2737                 rc = -ENOMEM;
2738                 goto free_priv;
2739         }
2740         /* Note: 128 & 256 bit salt are the same size */
2741         prot->rec_seq_size = rec_seq_size;
2742         memcpy(cctx->iv, salt, salt_size);
2743         memcpy(cctx->iv + salt_size, iv, iv_size);
2744         cctx->rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
2745         if (!cctx->rec_seq) {
2746                 rc = -ENOMEM;
2747                 goto free_iv;
2748         }
2749
2750         if (!*aead) {
2751                 *aead = crypto_alloc_aead(cipher_name, 0, 0);
2752                 if (IS_ERR(*aead)) {
2753                         rc = PTR_ERR(*aead);
2754                         *aead = NULL;
2755                         goto free_rec_seq;
2756                 }
2757         }
2758
2759         ctx->push_pending_record = tls_sw_push_pending_record;
2760
2761         rc = crypto_aead_setkey(*aead, key, keysize);
2762
2763         if (rc)
2764                 goto free_aead;
2765
2766         rc = crypto_aead_setauthsize(*aead, prot->tag_size);
2767         if (rc)
2768                 goto free_aead;
2769
2770         if (sw_ctx_rx) {
2771                 tfm = crypto_aead_tfm(sw_ctx_rx->aead_recv);
2772
2773                 tls_update_rx_zc_capable(ctx);
2774                 sw_ctx_rx->async_capable =
2775                         crypto_info->version != TLS_1_3_VERSION &&
2776                         !!(tfm->__crt_alg->cra_flags & CRYPTO_ALG_ASYNC);
2777
2778                 rc = tls_strp_init(&sw_ctx_rx->strp, sk);
2779                 if (rc)
2780                         goto free_aead;
2781         }
2782
2783         goto out;
2784
2785 free_aead:
2786         crypto_free_aead(*aead);
2787         *aead = NULL;
2788 free_rec_seq:
2789         kfree(cctx->rec_seq);
2790         cctx->rec_seq = NULL;
2791 free_iv:
2792         kfree(cctx->iv);
2793         cctx->iv = NULL;
2794 free_priv:
2795         if (tx) {
2796                 kfree(ctx->priv_ctx_tx);
2797                 ctx->priv_ctx_tx = NULL;
2798         } else {
2799                 kfree(ctx->priv_ctx_rx);
2800                 ctx->priv_ctx_rx = NULL;
2801         }
2802 out:
2803         return rc;
2804 }