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         bool async_done;
  56         u8 tail;
  57         );
  58
  59         struct sk_buff *skb;
  60 };
  61
  62 struct tls_decrypt_ctx {
  63         struct sock *sk;
  64         u8 iv[TLS_MAX_IV_SIZE];
  65         u8 aad[TLS_MAX_AAD_SIZE];
  66         u8 tail;
  67         bool free_sgout;
  68         struct scatterlist sg[];
  69 };
  70
  71 noinline void tls_err_abort(struct sock *sk, int err)
  72 {
  73         WARN_ON_ONCE(err >= 0);
  74         /* sk->sk_err should contain a positive error code. */
  75         WRITE_ONCE(sk->sk_err, -err);
  76         /* Paired with smp_rmb() in tcp_poll() */
  77         smp_wmb();
  78         sk_error_report(sk);
  79 }
  80
  81 static int __skb_nsg(struct sk_buff *skb, int offset, int len,
  82                      unsigned int recursion_level)
  83 {
  84         int start = skb_headlen(skb);
  85         int i, chunk = start - offset;
  86         struct sk_buff *frag_iter;
  87         int elt = 0;
  88
  89         if (unlikely(recursion_level >= 24))
  90                 return -EMSGSIZE;
  91
  92         if (chunk > 0) {
  93                 if (chunk > len)
  94                         chunk = len;
  95                 elt++;
  96                 len -= chunk;
  97                 if (len == 0)
  98                         return elt;
  99                 offset += chunk;
 100         }
 101
 102         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
 103                 int end;
 104
 105                 WARN_ON(start > offset + len);
 106
 107                 end = start + skb_frag_size(&skb_shinfo(skb)->frags[i]);
 108                 chunk = end - offset;
 109                 if (chunk > 0) {
 110                         if (chunk > len)
 111                                 chunk = len;
 112                         elt++;
 113                         len -= chunk;
 114                         if (len == 0)
 115                                 return elt;
 116                         offset += chunk;
 117                 }
 118                 start = end;
 119         }
 120
 121         if (unlikely(skb_has_frag_list(skb))) {
 122                 skb_walk_frags(skb, frag_iter) {
 123                         int end, ret;
 124
 125                         WARN_ON(start > offset + len);
 126
 127                         end = start + frag_iter->len;
 128                         chunk = end - offset;
 129                         if (chunk > 0) {
 130                                 if (chunk > len)
 131                                         chunk = len;
 132                                 ret = __skb_nsg(frag_iter, offset - start, chunk,
 133                                                 recursion_level + 1);
 134                                 if (unlikely(ret < 0))
 135                                         return ret;
 136                                 elt += ret;
 137                                 len -= chunk;
 138                                 if (len == 0)
 139                                         return elt;
 140                                 offset += chunk;
 141                         }
 142                         start = end;
 143                 }
 144         }
 145         BUG_ON(len);
 146         return elt;
 147 }
 148
 149 /* Return the number of scatterlist elements required to completely map the
 150  * skb, or -EMSGSIZE if the recursion depth is exceeded.
 151  */
 152 static int skb_nsg(struct sk_buff *skb, int offset, int len)
 153 {
 154         return __skb_nsg(skb, offset, len, 0);
 155 }
 156
 157 static int tls_padding_length(struct tls_prot_info *prot, struct sk_buff *skb,
 158                               struct tls_decrypt_arg *darg)
 159 {
 160         struct strp_msg *rxm = strp_msg(skb);
 161         struct tls_msg *tlm = tls_msg(skb);
 162         int sub = 0;
 163
 164         /* Determine zero-padding length */
 165         if (prot->version == TLS_1_3_VERSION) {
 166                 int offset = rxm->full_len - TLS_TAG_SIZE - 1;
 167                 char content_type = darg->zc ? darg->tail : 0;
 168                 int err;
 169
 170                 while (content_type == 0) {
 171                         if (offset < prot->prepend_size)
 172                                 return -EBADMSG;
 173                         err = skb_copy_bits(skb, rxm->offset + offset,
 174                                             &content_type, 1);
 175                         if (err)
 176                                 return err;
 177                         if (content_type)
 178                                 break;
 179                         sub++;
 180                         offset--;
 181                 }
 182                 tlm->control = content_type;
 183         }
 184         return sub;
 185 }
 186
 187 static void tls_decrypt_done(void *data, int err)
 188 {
 189         struct aead_request *aead_req = data;
 190         struct crypto_aead *aead = crypto_aead_reqtfm(aead_req);
 191         struct scatterlist *sgout = aead_req->dst;
 192         struct tls_sw_context_rx *ctx;
 193         struct tls_decrypt_ctx *dctx;
 194         struct tls_context *tls_ctx;
 195         struct scatterlist *sg;
 196         unsigned int pages;
 197         struct sock *sk;
 198         int aead_size;
 199
 200         /* If requests get too backlogged crypto API returns -EBUSY and calls
 201          * ->complete(-EINPROGRESS) immediately followed by ->complete(0)
 202          * to make waiting for backlog to flush with crypto_wait_req() easier.
 203          * First wait converts -EBUSY -> -EINPROGRESS, and the second one
 204          * -EINPROGRESS -> 0.
 205          * We have a single struct crypto_async_request per direction, this
 206          * scheme doesn't help us, so just ignore the first ->complete().
 207          */
 208         if (err == -EINPROGRESS)
 209                 return;
 210
 211         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(aead);
 212         aead_size = ALIGN(aead_size, __alignof__(*dctx));
 213         dctx = (void *)((u8 *)aead_req + aead_size);
 214
 215         sk = dctx->sk;
 216         tls_ctx = tls_get_ctx(sk);
 217         ctx = tls_sw_ctx_rx(tls_ctx);
 218
 219         /* Propagate if there was an err */
 220         if (err) {
 221                 if (err == -EBADMSG)
 222                         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSDECRYPTERROR);
 223                 ctx->async_wait.err = err;
 224                 tls_err_abort(sk, err);
 225         }
 226
 227         /* Free the destination pages if skb was not decrypted inplace */
 228         if (dctx->free_sgout) {
 229                 /* Skip the first S/G entry as it points to AAD */
 230                 for_each_sg(sg_next(sgout), sg, UINT_MAX, pages) {
 231                         if (!sg)
 232                                 break;
 233                         put_page(sg_page(sg));
 234                 }
 235         }
 236
 237         kfree(aead_req);
 238
 239         if (atomic_dec_and_test(&ctx->decrypt_pending))
 240                 complete(&ctx->async_wait.completion);
 241 }
 242
 243 static int tls_decrypt_async_wait(struct tls_sw_context_rx *ctx)
 244 {
 245         if (!atomic_dec_and_test(&ctx->decrypt_pending))
 246                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
 247         atomic_inc(&ctx->decrypt_pending);
 248
 249         return ctx->async_wait.err;
 250 }
 251
 252 static int tls_do_decryption(struct sock *sk,
 253                              struct scatterlist *sgin,
 254                              struct scatterlist *sgout,
 255                              char *iv_recv,
 256                              size_t data_len,
 257                              struct aead_request *aead_req,
 258                              struct tls_decrypt_arg *darg)
 259 {
 260         struct tls_context *tls_ctx = tls_get_ctx(sk);
 261         struct tls_prot_info *prot = &tls_ctx->prot_info;
 262         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
 263         int ret;
 264
 265         aead_request_set_tfm(aead_req, ctx->aead_recv);
 266         aead_request_set_ad(aead_req, prot->aad_size);
 267         aead_request_set_crypt(aead_req, sgin, sgout,
 268                                data_len + prot->tag_size,
 269                                (u8 *)iv_recv);
 270
 271         if (darg->async) {
 272                 aead_request_set_callback(aead_req,
 273                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 274                                           tls_decrypt_done, aead_req);
 275                 DEBUG_NET_WARN_ON_ONCE(atomic_read(&ctx->decrypt_pending) < 1);
 276                 atomic_inc(&ctx->decrypt_pending);
 277         } else {
 278                 DECLARE_CRYPTO_WAIT(wait);
 279
 280                 aead_request_set_callback(aead_req,
 281                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 282                                           crypto_req_done, &wait);
 283                 ret = crypto_aead_decrypt(aead_req);
 284                 if (ret == -EINPROGRESS || ret == -EBUSY)
 285                         ret = crypto_wait_req(ret, &wait);
 286                 return ret;
 287         }
 288
 289         ret = crypto_aead_decrypt(aead_req);
 290         if (ret == -EINPROGRESS)
 291                 return 0;
 292
 293         if (ret == -EBUSY) {
 294                 ret = tls_decrypt_async_wait(ctx);
 295                 darg->async_done = true;
 296                 /* all completions have run, we're not doing async anymore */
 297                 darg->async = false;
 298                 return ret;
 299         }
 300
 301         atomic_dec(&ctx->decrypt_pending);
 302         darg->async = false;
 303
 304         return ret;
 305 }
 306
 307 static void tls_trim_both_msgs(struct sock *sk, int target_size)
 308 {
 309         struct tls_context *tls_ctx = tls_get_ctx(sk);
 310         struct tls_prot_info *prot = &tls_ctx->prot_info;
 311         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 312         struct tls_rec *rec = ctx->open_rec;
 313
 314         sk_msg_trim(sk, &rec->msg_plaintext, target_size);
 315         if (target_size > 0)
 316                 target_size += prot->overhead_size;
 317         sk_msg_trim(sk, &rec->msg_encrypted, target_size);
 318 }
 319
 320 static int tls_alloc_encrypted_msg(struct sock *sk, int len)
 321 {
 322         struct tls_context *tls_ctx = tls_get_ctx(sk);
 323         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 324         struct tls_rec *rec = ctx->open_rec;
 325         struct sk_msg *msg_en = &rec->msg_encrypted;
 326
 327         return sk_msg_alloc(sk, msg_en, len, 0);
 328 }
 329
 330 static int tls_clone_plaintext_msg(struct sock *sk, int required)
 331 {
 332         struct tls_context *tls_ctx = tls_get_ctx(sk);
 333         struct tls_prot_info *prot = &tls_ctx->prot_info;
 334         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 335         struct tls_rec *rec = ctx->open_rec;
 336         struct sk_msg *msg_pl = &rec->msg_plaintext;
 337         struct sk_msg *msg_en = &rec->msg_encrypted;
 338         int skip, len;
 339
 340         /* We add page references worth len bytes from encrypted sg
 341          * at the end of plaintext sg. It is guaranteed that msg_en
 342          * has enough required room (ensured by caller).
 343          */
 344         len = required - msg_pl->sg.size;
 345
 346         /* Skip initial bytes in msg_en's data to be able to use
 347          * same offset of both plain and encrypted data.
 348          */
 349         skip = prot->prepend_size + msg_pl->sg.size;
 350
 351         return sk_msg_clone(sk, msg_pl, msg_en, skip, len);
 352 }
 353
 354 static struct tls_rec *tls_get_rec(struct sock *sk)
 355 {
 356         struct tls_context *tls_ctx = tls_get_ctx(sk);
 357         struct tls_prot_info *prot = &tls_ctx->prot_info;
 358         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 359         struct sk_msg *msg_pl, *msg_en;
 360         struct tls_rec *rec;
 361         int mem_size;
 362
 363         mem_size = sizeof(struct tls_rec) + crypto_aead_reqsize(ctx->aead_send);
 364
 365         rec = kzalloc(mem_size, sk->sk_allocation);
 366         if (!rec)
 367                 return NULL;
 368
 369         msg_pl = &rec->msg_plaintext;
 370         msg_en = &rec->msg_encrypted;
 371
 372         sk_msg_init(msg_pl);
 373         sk_msg_init(msg_en);
 374
 375         sg_init_table(rec->sg_aead_in, 2);
 376         sg_set_buf(&rec->sg_aead_in[0], rec->aad_space, prot->aad_size);
 377         sg_unmark_end(&rec->sg_aead_in[1]);
 378
 379         sg_init_table(rec->sg_aead_out, 2);
 380         sg_set_buf(&rec->sg_aead_out[0], rec->aad_space, prot->aad_size);
 381         sg_unmark_end(&rec->sg_aead_out[1]);
 382
 383         rec->sk = sk;
 384
 385         return rec;
 386 }
 387
 388 static void tls_free_rec(struct sock *sk, struct tls_rec *rec)
 389 {
 390         sk_msg_free(sk, &rec->msg_encrypted);
 391         sk_msg_free(sk, &rec->msg_plaintext);
 392         kfree(rec);
 393 }
 394
 395 static void tls_free_open_rec(struct sock *sk)
 396 {
 397         struct tls_context *tls_ctx = tls_get_ctx(sk);
 398         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 399         struct tls_rec *rec = ctx->open_rec;
 400
 401         if (rec) {
 402                 tls_free_rec(sk, rec);
 403                 ctx->open_rec = NULL;
 404         }
 405 }
 406
 407 int tls_tx_records(struct sock *sk, int flags)
 408 {
 409         struct tls_context *tls_ctx = tls_get_ctx(sk);
 410         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 411         struct tls_rec *rec, *tmp;
 412         struct sk_msg *msg_en;
 413         int tx_flags, rc = 0;
 414
 415         if (tls_is_partially_sent_record(tls_ctx)) {
 416                 rec = list_first_entry(&ctx->tx_list,
 417                                        struct tls_rec, list);
 418
 419                 if (flags == -1)
 420                         tx_flags = rec->tx_flags;
 421                 else
 422                         tx_flags = flags;
 423
 424                 rc = tls_push_partial_record(sk, tls_ctx, tx_flags);
 425                 if (rc)
 426                         goto tx_err;
 427
 428                 /* Full record has been transmitted.
 429                  * Remove the head of tx_list
 430                  */
 431                 list_del(&rec->list);
 432                 sk_msg_free(sk, &rec->msg_plaintext);
 433                 kfree(rec);
 434         }
 435
 436         /* Tx all ready records */
 437         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
 438                 if (READ_ONCE(rec->tx_ready)) {
 439                         if (flags == -1)
 440                                 tx_flags = rec->tx_flags;
 441                         else
 442                                 tx_flags = flags;
 443
 444                         msg_en = &rec->msg_encrypted;
 445                         rc = tls_push_sg(sk, tls_ctx,
 446                                          &msg_en->sg.data[msg_en->sg.curr],
 447                                          0, tx_flags);
 448                         if (rc)
 449                                 goto tx_err;
 450
 451                         list_del(&rec->list);
 452                         sk_msg_free(sk, &rec->msg_plaintext);
 453                         kfree(rec);
 454                 } else {
 455                         break;
 456                 }
 457         }
 458
 459 tx_err:
 460         if (rc < 0 && rc != -EAGAIN)
 461                 tls_err_abort(sk, rc);
 462
 463         return rc;
 464 }
 465
 466 static void tls_encrypt_done(void *data, int err)
 467 {
 468         struct tls_sw_context_tx *ctx;
 469         struct tls_context *tls_ctx;
 470         struct tls_prot_info *prot;
 471         struct tls_rec *rec = data;
 472         struct scatterlist *sge;
 473         struct sk_msg *msg_en;
 474         struct sock *sk;
 475
 476         if (err == -EINPROGRESS) /* see the comment in tls_decrypt_done() */
 477                 return;
 478
 479         msg_en = &rec->msg_encrypted;
 480
 481         sk = rec->sk;
 482         tls_ctx = tls_get_ctx(sk);
 483         prot = &tls_ctx->prot_info;
 484         ctx = tls_sw_ctx_tx(tls_ctx);
 485
 486         sge = sk_msg_elem(msg_en, msg_en->sg.curr);
 487         sge->offset -= prot->prepend_size;
 488         sge->length += prot->prepend_size;
 489
 490         /* Check if error is previously set on socket */
 491         if (err || sk->sk_err) {
 492                 rec = NULL;
 493
 494                 /* If err is already set on socket, return the same code */
 495                 if (sk->sk_err) {
 496                         ctx->async_wait.err = -sk->sk_err;
 497                 } else {
 498                         ctx->async_wait.err = err;
 499                         tls_err_abort(sk, err);
 500                 }
 501         }
 502
 503         if (rec) {
 504                 struct tls_rec *first_rec;
 505
 506                 /* Mark the record as ready for transmission */
 507                 smp_store_mb(rec->tx_ready, true);
 508
 509                 /* If received record is at head of tx_list, schedule tx */
 510                 first_rec = list_first_entry(&ctx->tx_list,
 511                                              struct tls_rec, list);
 512                 if (rec == first_rec) {
 513                         /* Schedule the transmission */
 514                         if (!test_and_set_bit(BIT_TX_SCHEDULED,
 515                                               &ctx->tx_bitmask))
 516                                 schedule_delayed_work(&ctx->tx_work.work, 1);
 517                 }
 518         }
 519
 520         if (atomic_dec_and_test(&ctx->encrypt_pending))
 521                 complete(&ctx->async_wait.completion);
 522 }
 523
 524 static int tls_encrypt_async_wait(struct tls_sw_context_tx *ctx)
 525 {
 526         if (!atomic_dec_and_test(&ctx->encrypt_pending))
 527                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
 528         atomic_inc(&ctx->encrypt_pending);
 529
 530         return ctx->async_wait.err;
 531 }
 532
 533 static int tls_do_encryption(struct sock *sk,
 534                              struct tls_context *tls_ctx,
 535                              struct tls_sw_context_tx *ctx,
 536                              struct aead_request *aead_req,
 537                              size_t data_len, u32 start)
 538 {
 539         struct tls_prot_info *prot = &tls_ctx->prot_info;
 540         struct tls_rec *rec = ctx->open_rec;
 541         struct sk_msg *msg_en = &rec->msg_encrypted;
 542         struct scatterlist *sge = sk_msg_elem(msg_en, start);
 543         int rc, iv_offset = 0;
 544
 545         /* For CCM based ciphers, first byte of IV is a constant */
 546         switch (prot->cipher_type) {
 547         case TLS_CIPHER_AES_CCM_128:
 548                 rec->iv_data[0] = TLS_AES_CCM_IV_B0_BYTE;
 549                 iv_offset = 1;
 550                 break;
 551         case TLS_CIPHER_SM4_CCM:
 552                 rec->iv_data[0] = TLS_SM4_CCM_IV_B0_BYTE;
 553                 iv_offset = 1;
 554                 break;
 555         }
 556
 557         memcpy(&rec->iv_data[iv_offset], tls_ctx->tx.iv,
 558                prot->iv_size + prot->salt_size);
 559
 560         tls_xor_iv_with_seq(prot, rec->iv_data + iv_offset,
 561                             tls_ctx->tx.rec_seq);
 562
 563         sge->offset += prot->prepend_size;
 564         sge->length -= prot->prepend_size;
 565
 566         msg_en->sg.curr = start;
 567
 568         aead_request_set_tfm(aead_req, ctx->aead_send);
 569         aead_request_set_ad(aead_req, prot->aad_size);
 570         aead_request_set_crypt(aead_req, rec->sg_aead_in,
 571                                rec->sg_aead_out,
 572                                data_len, rec->iv_data);
 573
 574         aead_request_set_callback(aead_req, CRYPTO_TFM_REQ_MAY_BACKLOG,
 575                                   tls_encrypt_done, rec);
 576
 577         /* Add the record in tx_list */
 578         list_add_tail((struct list_head *)&rec->list, &ctx->tx_list);
 579         DEBUG_NET_WARN_ON_ONCE(atomic_read(&ctx->encrypt_pending) < 1);
 580         atomic_inc(&ctx->encrypt_pending);
 581
 582         rc = crypto_aead_encrypt(aead_req);
 583         if (rc == -EBUSY) {
 584                 rc = tls_encrypt_async_wait(ctx);
 585                 rc = rc ?: -EINPROGRESS;
 586         }
 587         if (!rc || rc != -EINPROGRESS) {
 588                 atomic_dec(&ctx->encrypt_pending);
 589                 sge->offset -= prot->prepend_size;
 590                 sge->length += prot->prepend_size;
 591         }
 592
 593         if (!rc) {
 594                 WRITE_ONCE(rec->tx_ready, true);
 595         } else if (rc != -EINPROGRESS) {
 596                 list_del(&rec->list);
 597                 return rc;
 598         }
 599
 600         /* Unhook the record from context if encryption is not failure */
 601         ctx->open_rec = NULL;
 602         tls_advance_record_sn(sk, prot, &tls_ctx->tx);
 603         return rc;
 604 }
 605
 606 static int tls_split_open_record(struct sock *sk, struct tls_rec *from,
 607                                  struct tls_rec **to, struct sk_msg *msg_opl,
 608                                  struct sk_msg *msg_oen, u32 split_point,
 609                                  u32 tx_overhead_size, u32 *orig_end)
 610 {
 611         u32 i, j, bytes = 0, apply = msg_opl->apply_bytes;
 612         struct scatterlist *sge, *osge, *nsge;
 613         u32 orig_size = msg_opl->sg.size;
 614         struct scatterlist tmp = { };
 615         struct sk_msg *msg_npl;
 616         struct tls_rec *new;
 617         int ret;
 618
 619         new = tls_get_rec(sk);
 620         if (!new)
 621                 return -ENOMEM;
 622         ret = sk_msg_alloc(sk, &new->msg_encrypted, msg_opl->sg.size +
 623                            tx_overhead_size, 0);
 624         if (ret < 0) {
 625                 tls_free_rec(sk, new);
 626                 return ret;
 627         }
 628
 629         *orig_end = msg_opl->sg.end;
 630         i = msg_opl->sg.start;
 631         sge = sk_msg_elem(msg_opl, i);
 632         while (apply && sge->length) {
 633                 if (sge->length > apply) {
 634                         u32 len = sge->length - apply;
 635
 636                         get_page(sg_page(sge));
 637                         sg_set_page(&tmp, sg_page(sge), len,
 638                                     sge->offset + apply);
 639                         sge->length = apply;
 640                         bytes += apply;
 641                         apply = 0;
 642                 } else {
 643                         apply -= sge->length;
 644                         bytes += sge->length;
 645                 }
 646
 647                 sk_msg_iter_var_next(i);
 648                 if (i == msg_opl->sg.end)
 649                         break;
 650                 sge = sk_msg_elem(msg_opl, i);
 651         }
 652
 653         msg_opl->sg.end = i;
 654         msg_opl->sg.curr = i;
 655         msg_opl->sg.copybreak = 0;
 656         msg_opl->apply_bytes = 0;
 657         msg_opl->sg.size = bytes;
 658
 659         msg_npl = &new->msg_plaintext;
 660         msg_npl->apply_bytes = apply;
 661         msg_npl->sg.size = orig_size - bytes;
 662
 663         j = msg_npl->sg.start;
 664         nsge = sk_msg_elem(msg_npl, j);
 665         if (tmp.length) {
 666                 memcpy(nsge, &tmp, sizeof(*nsge));
 667                 sk_msg_iter_var_next(j);
 668                 nsge = sk_msg_elem(msg_npl, j);
 669         }
 670
 671         osge = sk_msg_elem(msg_opl, i);
 672         while (osge->length) {
 673                 memcpy(nsge, osge, sizeof(*nsge));
 674                 sg_unmark_end(nsge);
 675                 sk_msg_iter_var_next(i);
 676                 sk_msg_iter_var_next(j);
 677                 if (i == *orig_end)
 678                         break;
 679                 osge = sk_msg_elem(msg_opl, i);
 680                 nsge = sk_msg_elem(msg_npl, j);
 681         }
 682
 683         msg_npl->sg.end = j;
 684         msg_npl->sg.curr = j;
 685         msg_npl->sg.copybreak = 0;
 686
 687         *to = new;
 688         return 0;
 689 }
 690
 691 static void tls_merge_open_record(struct sock *sk, struct tls_rec *to,
 692                                   struct tls_rec *from, u32 orig_end)
 693 {
 694         struct sk_msg *msg_npl = &from->msg_plaintext;
 695         struct sk_msg *msg_opl = &to->msg_plaintext;
 696         struct scatterlist *osge, *nsge;
 697         u32 i, j;
 698
 699         i = msg_opl->sg.end;
 700         sk_msg_iter_var_prev(i);
 701         j = msg_npl->sg.start;
 702
 703         osge = sk_msg_elem(msg_opl, i);
 704         nsge = sk_msg_elem(msg_npl, j);
 705
 706         if (sg_page(osge) == sg_page(nsge) &&
 707             osge->offset + osge->length == nsge->offset) {
 708                 osge->length += nsge->length;
 709                 put_page(sg_page(nsge));
 710         }
 711
 712         msg_opl->sg.end = orig_end;
 713         msg_opl->sg.curr = orig_end;
 714         msg_opl->sg.copybreak = 0;
 715         msg_opl->apply_bytes = msg_opl->sg.size + msg_npl->sg.size;
 716         msg_opl->sg.size += msg_npl->sg.size;
 717
 718         sk_msg_free(sk, &to->msg_encrypted);
 719         sk_msg_xfer_full(&to->msg_encrypted, &from->msg_encrypted);
 720
 721         kfree(from);
 722 }
 723
 724 static int tls_push_record(struct sock *sk, int flags,
 725                            unsigned char record_type)
 726 {
 727         struct tls_context *tls_ctx = tls_get_ctx(sk);
 728         struct tls_prot_info *prot = &tls_ctx->prot_info;
 729         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 730         struct tls_rec *rec = ctx->open_rec, *tmp = NULL;
 731         u32 i, split_point, orig_end;
 732         struct sk_msg *msg_pl, *msg_en;
 733         struct aead_request *req;
 734         bool split;
 735         int rc;
 736
 737         if (!rec)
 738                 return 0;
 739
 740         msg_pl = &rec->msg_plaintext;
 741         msg_en = &rec->msg_encrypted;
 742
 743         split_point = msg_pl->apply_bytes;
 744         split = split_point && split_point < msg_pl->sg.size;
 745         if (unlikely((!split &&
 746                       msg_pl->sg.size +
 747                       prot->overhead_size > msg_en->sg.size) ||
 748                      (split &&
 749                       split_point +
 750                       prot->overhead_size > msg_en->sg.size))) {
 751                 split = true;
 752                 split_point = msg_en->sg.size;
 753         }
 754         if (split) {
 755                 rc = tls_split_open_record(sk, rec, &tmp, msg_pl, msg_en,
 756                                            split_point, prot->overhead_size,
 757                                            &orig_end);
 758                 if (rc < 0)
 759                         return rc;
 760                 /* This can happen if above tls_split_open_record allocates
 761                  * a single large encryption buffer instead of two smaller
 762                  * ones. In this case adjust pointers and continue without
 763                  * split.
 764                  */
 765                 if (!msg_pl->sg.size) {
 766                         tls_merge_open_record(sk, rec, tmp, orig_end);
 767                         msg_pl = &rec->msg_plaintext;
 768                         msg_en = &rec->msg_encrypted;
 769                         split = false;
 770                 }
 771                 sk_msg_trim(sk, msg_en, msg_pl->sg.size +
 772                             prot->overhead_size);
 773         }
 774
 775         rec->tx_flags = flags;
 776         req = &rec->aead_req;
 777
 778         i = msg_pl->sg.end;
 779         sk_msg_iter_var_prev(i);
 780
 781         rec->content_type = record_type;
 782         if (prot->version == TLS_1_3_VERSION) {
 783                 /* Add content type to end of message.  No padding added */
 784                 sg_set_buf(&rec->sg_content_type, &rec->content_type, 1);
 785                 sg_mark_end(&rec->sg_content_type);
 786                 sg_chain(msg_pl->sg.data, msg_pl->sg.end + 1,
 787                          &rec->sg_content_type);
 788         } else {
 789                 sg_mark_end(sk_msg_elem(msg_pl, i));
 790         }
 791
 792         if (msg_pl->sg.end < msg_pl->sg.start) {
 793                 sg_chain(&msg_pl->sg.data[msg_pl->sg.start],
 794                          MAX_SKB_FRAGS - msg_pl->sg.start + 1,
 795                          msg_pl->sg.data);
 796         }
 797
 798         i = msg_pl->sg.start;
 799         sg_chain(rec->sg_aead_in, 2, &msg_pl->sg.data[i]);
 800
 801         i = msg_en->sg.end;
 802         sk_msg_iter_var_prev(i);
 803         sg_mark_end(sk_msg_elem(msg_en, i));
 804
 805         i = msg_en->sg.start;
 806         sg_chain(rec->sg_aead_out, 2, &msg_en->sg.data[i]);
 807
 808         tls_make_aad(rec->aad_space, msg_pl->sg.size + prot->tail_size,
 809                      tls_ctx->tx.rec_seq, record_type, prot);
 810
 811         tls_fill_prepend(tls_ctx,
 812                          page_address(sg_page(&msg_en->sg.data[i])) +
 813                          msg_en->sg.data[i].offset,
 814                          msg_pl->sg.size + prot->tail_size,
 815                          record_type);
 816
 817         tls_ctx->pending_open_record_frags = false;
 818
 819         rc = tls_do_encryption(sk, tls_ctx, ctx, req,
 820                                msg_pl->sg.size + prot->tail_size, i);
 821         if (rc < 0) {
 822                 if (rc != -EINPROGRESS) {
 823                         tls_err_abort(sk, -EBADMSG);
 824                         if (split) {
 825                                 tls_ctx->pending_open_record_frags = true;
 826                                 tls_merge_open_record(sk, rec, tmp, orig_end);
 827                         }
 828                 }
 829                 ctx->async_capable = 1;
 830                 return rc;
 831         } else if (split) {
 832                 msg_pl = &tmp->msg_plaintext;
 833                 msg_en = &tmp->msg_encrypted;
 834                 sk_msg_trim(sk, msg_en, msg_pl->sg.size + prot->overhead_size);
 835                 tls_ctx->pending_open_record_frags = true;
 836                 ctx->open_rec = tmp;
 837         }
 838
 839         return tls_tx_records(sk, flags);
 840 }
 841
 842 static int bpf_exec_tx_verdict(struct sk_msg *msg, struct sock *sk,
 843                                bool full_record, u8 record_type,
 844                                ssize_t *copied, int flags)
 845 {
 846         struct tls_context *tls_ctx = tls_get_ctx(sk);
 847         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 848         struct sk_msg msg_redir = { };
 849         struct sk_psock *psock;
 850         struct sock *sk_redir;
 851         struct tls_rec *rec;
 852         bool enospc, policy, redir_ingress;
 853         int err = 0, send;
 854         u32 delta = 0;
 855
 856         policy = !(flags & MSG_SENDPAGE_NOPOLICY);
 857         psock = sk_psock_get(sk);
 858         if (!psock || !policy) {
 859                 err = tls_push_record(sk, flags, record_type);
 860                 if (err && err != -EINPROGRESS && sk->sk_err == EBADMSG) {
 861                         *copied -= sk_msg_free(sk, msg);
 862                         tls_free_open_rec(sk);
 863                         err = -sk->sk_err;
 864                 }
 865                 if (psock)
 866                         sk_psock_put(sk, psock);
 867                 return err;
 868         }
 869 more_data:
 870         enospc = sk_msg_full(msg);
 871         if (psock->eval == __SK_NONE) {
 872                 delta = msg->sg.size;
 873                 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
 874                 delta -= msg->sg.size;
 875         }
 876         if (msg->cork_bytes && msg->cork_bytes > msg->sg.size &&
 877             !enospc && !full_record) {
 878                 err = -ENOSPC;
 879                 goto out_err;
 880         }
 881         msg->cork_bytes = 0;
 882         send = msg->sg.size;
 883         if (msg->apply_bytes && msg->apply_bytes < send)
 884                 send = msg->apply_bytes;
 885
 886         switch (psock->eval) {
 887         case __SK_PASS:
 888                 err = tls_push_record(sk, flags, record_type);
 889                 if (err && err != -EINPROGRESS && sk->sk_err == EBADMSG) {
 890                         *copied -= sk_msg_free(sk, msg);
 891                         tls_free_open_rec(sk);
 892                         err = -sk->sk_err;
 893                         goto out_err;
 894                 }
 895                 break;
 896         case __SK_REDIRECT:
 897                 redir_ingress = psock->redir_ingress;
 898                 sk_redir = psock->sk_redir;
 899                 memcpy(&msg_redir, msg, sizeof(*msg));
 900                 if (msg->apply_bytes < send)
 901                         msg->apply_bytes = 0;
 902                 else
 903                         msg->apply_bytes -= send;
 904                 sk_msg_return_zero(sk, msg, send);
 905                 msg->sg.size -= send;
 906                 release_sock(sk);
 907                 err = tcp_bpf_sendmsg_redir(sk_redir, redir_ingress,
 908                                             &msg_redir, send, flags);
 909                 lock_sock(sk);
 910                 if (err < 0) {
 911                         /* Regardless of whether the data represented by
 912                          * msg_redir is sent successfully, we have already
 913                          * uncharged it via sk_msg_return_zero(). The
 914                          * msg->sg.size represents the remaining unprocessed
 915                          * data, which needs to be uncharged here.
 916                          */
 917                         sk_mem_uncharge(sk, msg->sg.size);
 918                         *copied -= sk_msg_free_nocharge(sk, &msg_redir);
 919                         msg->sg.size = 0;
 920                 }
 921                 if (msg->sg.size == 0)
 922                         tls_free_open_rec(sk);
 923                 break;
 924         case __SK_DROP:
 925         default:
 926                 sk_msg_free_partial(sk, msg, send);
 927                 if (msg->apply_bytes < send)
 928                         msg->apply_bytes = 0;
 929                 else
 930                         msg->apply_bytes -= send;
 931                 if (msg->sg.size == 0)
 932                         tls_free_open_rec(sk);
 933                 *copied -= (send + delta);
 934                 err = -EACCES;
 935         }
 936
 937         if (likely(!err)) {
 938                 bool reset_eval = !ctx->open_rec;
 939
 940                 rec = ctx->open_rec;
 941                 if (rec) {
 942                         msg = &rec->msg_plaintext;
 943                         if (!msg->apply_bytes)
 944                                 reset_eval = true;
 945                 }
 946                 if (reset_eval) {
 947                         psock->eval = __SK_NONE;
 948                         if (psock->sk_redir) {
 949                                 sock_put(psock->sk_redir);
 950                                 psock->sk_redir = NULL;
 951                         }
 952                 }
 953                 if (rec)
 954                         goto more_data;
 955         }
 956  out_err:
 957         sk_psock_put(sk, psock);
 958         return err;
 959 }
 960
 961 static int tls_sw_push_pending_record(struct sock *sk, int flags)
 962 {
 963         struct tls_context *tls_ctx = tls_get_ctx(sk);
 964         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 965         struct tls_rec *rec = ctx->open_rec;
 966         struct sk_msg *msg_pl;
 967         size_t copied;
 968
 969         if (!rec)
 970                 return 0;
 971
 972         msg_pl = &rec->msg_plaintext;
 973         copied = msg_pl->sg.size;
 974         if (!copied)
 975                 return 0;
 976
 977         return bpf_exec_tx_verdict(msg_pl, sk, true, TLS_RECORD_TYPE_DATA,
 978                                    &copied, flags);
 979 }
 980
 981 static int tls_sw_sendmsg_splice(struct sock *sk, struct msghdr *msg,
 982                                  struct sk_msg *msg_pl, size_t try_to_copy,
 983                                  ssize_t *copied)
 984 {
 985         struct page *page = NULL, **pages = &page;
 986
 987         do {
 988                 ssize_t part;
 989                 size_t off;
 990
 991                 part = iov_iter_extract_pages(&msg->msg_iter, &pages,
 992                                               try_to_copy, 1, 0, &off);
 993                 if (part <= 0)
 994                         return part ?: -EIO;
 995
 996                 if (WARN_ON_ONCE(!sendpage_ok(page))) {
 997                         iov_iter_revert(&msg->msg_iter, part);
 998                         return -EIO;
 999                 }
1000
1001                 sk_msg_page_add(msg_pl, page, part, off);
1002                 msg_pl->sg.copybreak = 0;
1003                 msg_pl->sg.curr = msg_pl->sg.end;
1004                 sk_mem_charge(sk, part);
1005                 *copied += part;
1006                 try_to_copy -= part;
1007         } while (try_to_copy && !sk_msg_full(msg_pl));
1008
1009         return 0;
1010 }
1011
1012 static int tls_sw_sendmsg_locked(struct sock *sk, struct msghdr *msg,
1013                                  size_t size)
1014 {
1015         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1016         struct tls_context *tls_ctx = tls_get_ctx(sk);
1017         struct tls_prot_info *prot = &tls_ctx->prot_info;
1018         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1019         bool async_capable = ctx->async_capable;
1020         unsigned char record_type = TLS_RECORD_TYPE_DATA;
1021         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
1022         bool eor = !(msg->msg_flags & MSG_MORE);
1023         size_t try_to_copy;
1024         ssize_t copied = 0;
1025         struct sk_msg *msg_pl, *msg_en;
1026         struct tls_rec *rec;
1027         int required_size;
1028         int num_async = 0;
1029         bool full_record;
1030         int record_room;
1031         int num_zc = 0;
1032         int orig_size;
1033         int ret = 0;
1034
1035         if (!eor && (msg->msg_flags & MSG_EOR))
1036                 return -EINVAL;
1037
1038         if (unlikely(msg->msg_controllen)) {
1039                 ret = tls_process_cmsg(sk, msg, &record_type);
1040                 if (ret) {
1041                         if (ret == -EINPROGRESS)
1042                                 num_async++;
1043                         else if (ret != -EAGAIN)
1044                                 goto send_end;
1045                 }
1046         }
1047
1048         while (msg_data_left(msg)) {
1049                 if (sk->sk_err) {
1050                         ret = -sk->sk_err;
1051                         goto send_end;
1052                 }
1053
1054                 if (ctx->open_rec)
1055                         rec = ctx->open_rec;
1056                 else
1057                         rec = ctx->open_rec = tls_get_rec(sk);
1058                 if (!rec) {
1059                         ret = -ENOMEM;
1060                         goto send_end;
1061                 }
1062
1063                 msg_pl = &rec->msg_plaintext;
1064                 msg_en = &rec->msg_encrypted;
1065
1066                 orig_size = msg_pl->sg.size;
1067                 full_record = false;
1068                 try_to_copy = msg_data_left(msg);
1069                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
1070                 if (try_to_copy >= record_room) {
1071                         try_to_copy = record_room;
1072                         full_record = true;
1073                 }
1074
1075                 required_size = msg_pl->sg.size + try_to_copy +
1076                                 prot->overhead_size;
1077
1078                 if (!sk_stream_memory_free(sk))
1079                         goto wait_for_sndbuf;
1080
1081 alloc_encrypted:
1082                 ret = tls_alloc_encrypted_msg(sk, required_size);
1083                 if (ret) {
1084                         if (ret != -ENOSPC)
1085                                 goto wait_for_memory;
1086
1087                         /* Adjust try_to_copy according to the amount that was
1088                          * actually allocated. The difference is due
1089                          * to max sg elements limit
1090                          */
1091                         try_to_copy -= required_size - msg_en->sg.size;
1092                         full_record = true;
1093                 }
1094
1095                 if (try_to_copy && (msg->msg_flags & MSG_SPLICE_PAGES)) {
1096                         ret = tls_sw_sendmsg_splice(sk, msg, msg_pl,
1097                                                     try_to_copy, &copied);
1098                         if (ret < 0)
1099                                 goto send_end;
1100                         tls_ctx->pending_open_record_frags = true;
1101
1102                         if (sk_msg_full(msg_pl))
1103                                 full_record = true;
1104
1105                         if (full_record || eor)
1106                                 goto copied;
1107                         continue;
1108                 }
1109
1110                 if (!is_kvec && (full_record || eor) && !async_capable) {
1111                         u32 first = msg_pl->sg.end;
1112
1113                         ret = sk_msg_zerocopy_from_iter(sk, &msg->msg_iter,
1114                                                         msg_pl, try_to_copy);
1115                         if (ret)
1116                                 goto fallback_to_reg_send;
1117
1118                         num_zc++;
1119                         copied += try_to_copy;
1120
1121                         sk_msg_sg_copy_set(msg_pl, first);
1122                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1123                                                   record_type, &copied,
1124                                                   msg->msg_flags);
1125                         if (ret) {
1126                                 if (ret == -EINPROGRESS)
1127                                         num_async++;
1128                                 else if (ret == -ENOMEM)
1129                                         goto wait_for_memory;
1130                                 else if (ctx->open_rec && ret == -ENOSPC) {
1131                                         if (msg_pl->cork_bytes) {
1132                                                 ret = 0;
1133                                                 goto send_end;
1134                                         }
1135                                         goto rollback_iter;
1136                                 } else if (ret != -EAGAIN)
1137                                         goto send_end;
1138                         }
1139                         continue;
1140 rollback_iter:
1141                         copied -= try_to_copy;
1142                         sk_msg_sg_copy_clear(msg_pl, first);
1143                         iov_iter_revert(&msg->msg_iter,
1144                                         msg_pl->sg.size - orig_size);
1145 fallback_to_reg_send:
1146                         sk_msg_trim(sk, msg_pl, orig_size);
1147                 }
1148
1149                 required_size = msg_pl->sg.size + try_to_copy;
1150
1151                 ret = tls_clone_plaintext_msg(sk, required_size);
1152                 if (ret) {
1153                         if (ret != -ENOSPC)
1154                                 goto send_end;
1155
1156                         /* Adjust try_to_copy according to the amount that was
1157                          * actually allocated. The difference is due
1158                          * to max sg elements limit
1159                          */
1160                         try_to_copy -= required_size - msg_pl->sg.size;
1161                         full_record = true;
1162                         sk_msg_trim(sk, msg_en,
1163                                     msg_pl->sg.size + prot->overhead_size);
1164                 }
1165
1166                 if (try_to_copy) {
1167                         ret = sk_msg_memcopy_from_iter(sk, &msg->msg_iter,
1168                                                        msg_pl, try_to_copy);
1169                         if (ret < 0)
1170                                 goto trim_sgl;
1171                 }
1172
1173                 /* Open records defined only if successfully copied, otherwise
1174                  * we would trim the sg but not reset the open record frags.
1175                  */
1176                 tls_ctx->pending_open_record_frags = true;
1177                 copied += try_to_copy;
1178 copied:
1179                 if (full_record || eor) {
1180                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1181                                                   record_type, &copied,
1182                                                   msg->msg_flags);
1183                         if (ret) {
1184                                 if (ret == -EINPROGRESS)
1185                                         num_async++;
1186                                 else if (ret == -ENOMEM)
1187                                         goto wait_for_memory;
1188                                 else if (ret != -EAGAIN) {
1189                                         if (ret == -ENOSPC)
1190                                                 ret = 0;
1191                                         goto send_end;
1192                                 }
1193                         }
1194                 }
1195
1196                 continue;
1197
1198 wait_for_sndbuf:
1199                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1200 wait_for_memory:
1201                 ret = sk_stream_wait_memory(sk, &timeo);
1202                 if (ret) {
1203 trim_sgl:
1204                         if (ctx->open_rec)
1205                                 tls_trim_both_msgs(sk, orig_size);
1206                         goto send_end;
1207                 }
1208
1209                 if (ctx->open_rec && msg_en->sg.size < required_size)
1210                         goto alloc_encrypted;
1211         }
1212
1213         if (!num_async) {
1214                 goto send_end;
1215         } else if (num_zc || eor) {
1216                 int err;
1217
1218                 /* Wait for pending encryptions to get completed */
1219                 err = tls_encrypt_async_wait(ctx);
1220                 if (err) {
1221                         ret = err;
1222                         copied = 0;
1223                 }
1224         }
1225
1226         /* Transmit if any encryptions have completed */
1227         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1228                 cancel_delayed_work(&ctx->tx_work.work);
1229                 tls_tx_records(sk, msg->msg_flags);
1230         }
1231
1232 send_end:
1233         ret = sk_stream_error(sk, msg->msg_flags, ret);
1234         return copied > 0 ? copied : ret;
1235 }
1236
1237 int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1238 {
1239         struct tls_context *tls_ctx = tls_get_ctx(sk);
1240         int ret;
1241
1242         if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1243                                MSG_CMSG_COMPAT | MSG_SPLICE_PAGES | MSG_EOR |
1244                                MSG_SENDPAGE_NOPOLICY))
1245                 return -EOPNOTSUPP;
1246
1247         ret = mutex_lock_interruptible(&tls_ctx->tx_lock);
1248         if (ret)
1249                 return ret;
1250         lock_sock(sk);
1251         ret = tls_sw_sendmsg_locked(sk, msg, size);
1252         release_sock(sk);
1253         mutex_unlock(&tls_ctx->tx_lock);
1254         return ret;
1255 }
1256
1257 /*
1258  * Handle unexpected EOF during splice without SPLICE_F_MORE set.
1259  */
1260 void tls_sw_splice_eof(struct socket *sock)
1261 {
1262         struct sock *sk = sock->sk;
1263         struct tls_context *tls_ctx = tls_get_ctx(sk);
1264         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1265         struct tls_rec *rec;
1266         struct sk_msg *msg_pl;
1267         ssize_t copied = 0;
1268         bool retrying = false;
1269         int ret = 0;
1270
1271         if (!ctx->open_rec)
1272                 return;
1273
1274         mutex_lock(&tls_ctx->tx_lock);
1275         lock_sock(sk);
1276
1277 retry:
1278         /* same checks as in tls_sw_push_pending_record() */
1279         rec = ctx->open_rec;
1280         if (!rec)
1281                 goto unlock;
1282
1283         msg_pl = &rec->msg_plaintext;
1284         if (msg_pl->sg.size == 0)
1285                 goto unlock;
1286
1287         /* Check the BPF advisor and perform transmission. */
1288         ret = bpf_exec_tx_verdict(msg_pl, sk, false, TLS_RECORD_TYPE_DATA,
1289                                   &copied, 0);
1290         switch (ret) {
1291         case 0:
1292         case -EAGAIN:
1293                 if (retrying)
1294                         goto unlock;
1295                 retrying = true;
1296                 goto retry;
1297         case -EINPROGRESS:
1298                 break;
1299         default:
1300                 goto unlock;
1301         }
1302
1303         /* Wait for pending encryptions to get completed */
1304         if (tls_encrypt_async_wait(ctx))
1305                 goto unlock;
1306
1307         /* Transmit if any encryptions have completed */
1308         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1309                 cancel_delayed_work(&ctx->tx_work.work);
1310                 tls_tx_records(sk, 0);
1311         }
1312
1313 unlock:
1314         release_sock(sk);
1315         mutex_unlock(&tls_ctx->tx_lock);
1316 }
1317
1318 static int
1319 tls_rx_rec_wait(struct sock *sk, struct sk_psock *psock, bool nonblock,
1320                 bool released)
1321 {
1322         struct tls_context *tls_ctx = tls_get_ctx(sk);
1323         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1324         DEFINE_WAIT_FUNC(wait, woken_wake_function);
1325         int ret = 0;
1326         long timeo;
1327
1328         /* a rekey is pending, let userspace deal with it */
1329         if (unlikely(ctx->key_update_pending))
1330                 return -EKEYEXPIRED;
1331
1332         timeo = sock_rcvtimeo(sk, nonblock);
1333
1334         while (!tls_strp_msg_ready(ctx)) {
1335                 if (!sk_psock_queue_empty(psock))
1336                         return 0;
1337
1338                 if (sk->sk_err)
1339                         return sock_error(sk);
1340
1341                 if (ret < 0)
1342                         return ret;
1343
1344                 if (!skb_queue_empty(&sk->sk_receive_queue)) {
1345                         tls_strp_check_rcv(&ctx->strp);
1346                         if (tls_strp_msg_ready(ctx))
1347                                 break;
1348                 }
1349
1350                 if (sk->sk_shutdown & RCV_SHUTDOWN)
1351                         return 0;
1352
1353                 if (sock_flag(sk, SOCK_DONE))
1354                         return 0;
1355
1356                 if (!timeo)
1357                         return -EAGAIN;
1358
1359                 released = true;
1360                 add_wait_queue(sk_sleep(sk), &wait);
1361                 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1362                 ret = sk_wait_event(sk, &timeo,
1363                                     tls_strp_msg_ready(ctx) ||
1364                                     !sk_psock_queue_empty(psock),
1365                                     &wait);
1366                 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1367                 remove_wait_queue(sk_sleep(sk), &wait);
1368
1369                 /* Handle signals */
1370                 if (signal_pending(current))
1371                         return sock_intr_errno(timeo);
1372         }
1373
1374         tls_strp_msg_load(&ctx->strp, released);
1375
1376         return 1;
1377 }
1378
1379 static int tls_setup_from_iter(struct iov_iter *from,
1380                                int length, int *pages_used,
1381                                struct scatterlist *to,
1382                                int to_max_pages)
1383 {
1384         int rc = 0, i = 0, num_elem = *pages_used, maxpages;
1385         struct page *pages[MAX_SKB_FRAGS];
1386         unsigned int size = 0;
1387         ssize_t copied, use;
1388         size_t offset;
1389
1390         while (length > 0) {
1391                 i = 0;
1392                 maxpages = to_max_pages - num_elem;
1393                 if (maxpages == 0) {
1394                         rc = -EFAULT;
1395                         goto out;
1396                 }
1397                 copied = iov_iter_get_pages2(from, pages,
1398                                             length,
1399                                             maxpages, &offset);
1400                 if (copied <= 0) {
1401                         rc = -EFAULT;
1402                         goto out;
1403                 }
1404
1405                 length -= copied;
1406                 size += copied;
1407                 while (copied) {
1408                         use = min_t(int, copied, PAGE_SIZE - offset);
1409
1410                         sg_set_page(&to[num_elem],
1411                                     pages[i], use, offset);
1412                         sg_unmark_end(&to[num_elem]);
1413                         /* We do not uncharge memory from this API */
1414
1415                         offset = 0;
1416                         copied -= use;
1417
1418                         i++;
1419                         num_elem++;
1420                 }
1421         }
1422         /* Mark the end in the last sg entry if newly added */
1423         if (num_elem > *pages_used)
1424                 sg_mark_end(&to[num_elem - 1]);
1425 out:
1426         if (rc)
1427                 iov_iter_revert(from, size);
1428         *pages_used = num_elem;
1429
1430         return rc;
1431 }
1432
1433 static struct sk_buff *
1434 tls_alloc_clrtxt_skb(struct sock *sk, struct sk_buff *skb,
1435                      unsigned int full_len)
1436 {
1437         struct strp_msg *clr_rxm;
1438         struct sk_buff *clr_skb;
1439         int err;
1440
1441         clr_skb = alloc_skb_with_frags(0, full_len, TLS_PAGE_ORDER,
1442                                        &err, sk->sk_allocation);
1443         if (!clr_skb)
1444                 return NULL;
1445
1446         skb_copy_header(clr_skb, skb);
1447         clr_skb->len = full_len;
1448         clr_skb->data_len = full_len;
1449
1450         clr_rxm = strp_msg(clr_skb);
1451         clr_rxm->offset = 0;
1452
1453         return clr_skb;
1454 }
1455
1456 /* Decrypt handlers
1457  *
1458  * tls_decrypt_sw() and tls_decrypt_device() are decrypt handlers.
1459  * They must transform the darg in/out argument are as follows:
1460  *       |          Input            |         Output
1461  * -------------------------------------------------------------------
1462  *    zc | Zero-copy decrypt allowed | Zero-copy performed
1463  * async | Async decrypt allowed     | Async crypto used / in progress
1464  *   skb |            *              | Output skb
1465  *
1466  * If ZC decryption was performed darg.skb will point to the input skb.
1467  */
1468
1469 /* This function decrypts the input skb into either out_iov or in out_sg
1470  * or in skb buffers itself. The input parameter 'darg->zc' indicates if
1471  * zero-copy mode needs to be tried or not. With zero-copy mode, either
1472  * out_iov or out_sg must be non-NULL. In case both out_iov and out_sg are
1473  * NULL, then the decryption happens inside skb buffers itself, i.e.
1474  * zero-copy gets disabled and 'darg->zc' is updated.
1475  */
1476 static int tls_decrypt_sg(struct sock *sk, struct iov_iter *out_iov,
1477                           struct scatterlist *out_sg,
1478                           struct tls_decrypt_arg *darg)
1479 {
1480         struct tls_context *tls_ctx = tls_get_ctx(sk);
1481         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1482         struct tls_prot_info *prot = &tls_ctx->prot_info;
1483         int n_sgin, n_sgout, aead_size, err, pages = 0;
1484         struct sk_buff *skb = tls_strp_msg(ctx);
1485         const struct strp_msg *rxm = strp_msg(skb);
1486         const struct tls_msg *tlm = tls_msg(skb);
1487         struct aead_request *aead_req;
1488         struct scatterlist *sgin = NULL;
1489         struct scatterlist *sgout = NULL;
1490         const int data_len = rxm->full_len - prot->overhead_size;
1491         int tail_pages = !!prot->tail_size;
1492         struct tls_decrypt_ctx *dctx;
1493         struct sk_buff *clear_skb;
1494         int iv_offset = 0;
1495         u8 *mem;
1496
1497         n_sgin = skb_nsg(skb, rxm->offset + prot->prepend_size,
1498                          rxm->full_len - prot->prepend_size);
1499         if (n_sgin < 1)
1500                 return n_sgin ?: -EBADMSG;
1501
1502         if (darg->zc && (out_iov || out_sg)) {
1503                 clear_skb = NULL;
1504
1505                 if (out_iov)
1506                         n_sgout = 1 + tail_pages +
1507                                 iov_iter_npages_cap(out_iov, INT_MAX, data_len);
1508                 else
1509                         n_sgout = sg_nents(out_sg);
1510         } else {
1511                 darg->zc = false;
1512
1513                 clear_skb = tls_alloc_clrtxt_skb(sk, skb, rxm->full_len);
1514                 if (!clear_skb)
1515                         return -ENOMEM;
1516
1517                 n_sgout = 1 + skb_shinfo(clear_skb)->nr_frags;
1518         }
1519
1520         /* Increment to accommodate AAD */
1521         n_sgin = n_sgin + 1;
1522
1523         /* Allocate a single block of memory which contains
1524          *   aead_req || tls_decrypt_ctx.
1525          * Both structs are variable length.
1526          */
1527         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(ctx->aead_recv);
1528         aead_size = ALIGN(aead_size, __alignof__(*dctx));
1529         mem = kmalloc(aead_size + struct_size(dctx, sg, size_add(n_sgin, n_sgout)),
1530                       sk->sk_allocation);
1531         if (!mem) {
1532                 err = -ENOMEM;
1533                 goto exit_free_skb;
1534         }
1535
1536         /* Segment the allocated memory */
1537         aead_req = (struct aead_request *)mem;
1538         dctx = (struct tls_decrypt_ctx *)(mem + aead_size);
1539         dctx->sk = sk;
1540         sgin = &dctx->sg[0];
1541         sgout = &dctx->sg[n_sgin];
1542
1543         /* For CCM based ciphers, first byte of nonce+iv is a constant */
1544         switch (prot->cipher_type) {
1545         case TLS_CIPHER_AES_CCM_128:
1546                 dctx->iv[0] = TLS_AES_CCM_IV_B0_BYTE;
1547                 iv_offset = 1;
1548                 break;
1549         case TLS_CIPHER_SM4_CCM:
1550                 dctx->iv[0] = TLS_SM4_CCM_IV_B0_BYTE;
1551                 iv_offset = 1;
1552                 break;
1553         }
1554
1555         /* Prepare IV */
1556         if (prot->version == TLS_1_3_VERSION ||
1557             prot->cipher_type == TLS_CIPHER_CHACHA20_POLY1305) {
1558                 memcpy(&dctx->iv[iv_offset], tls_ctx->rx.iv,
1559                        prot->iv_size + prot->salt_size);
1560         } else {
1561                 err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
1562                                     &dctx->iv[iv_offset] + prot->salt_size,
1563                                     prot->iv_size);
1564                 if (err < 0)
1565                         goto exit_free;
1566                 memcpy(&dctx->iv[iv_offset], tls_ctx->rx.iv, prot->salt_size);
1567         }
1568         tls_xor_iv_with_seq(prot, &dctx->iv[iv_offset], tls_ctx->rx.rec_seq);
1569
1570         /* Prepare AAD */
1571         tls_make_aad(dctx->aad, rxm->full_len - prot->overhead_size +
1572                      prot->tail_size,
1573                      tls_ctx->rx.rec_seq, tlm->control, prot);
1574
1575         /* Prepare sgin */
1576         sg_init_table(sgin, n_sgin);
1577         sg_set_buf(&sgin[0], dctx->aad, prot->aad_size);
1578         err = skb_to_sgvec(skb, &sgin[1],
1579                            rxm->offset + prot->prepend_size,
1580                            rxm->full_len - prot->prepend_size);
1581         if (err < 0)
1582                 goto exit_free;
1583
1584         if (clear_skb) {
1585                 sg_init_table(sgout, n_sgout);
1586                 sg_set_buf(&sgout[0], dctx->aad, prot->aad_size);
1587
1588                 err = skb_to_sgvec(clear_skb, &sgout[1], prot->prepend_size,
1589                                    data_len + prot->tail_size);
1590                 if (err < 0)
1591                         goto exit_free;
1592         } else if (out_iov) {
1593                 sg_init_table(sgout, n_sgout);
1594                 sg_set_buf(&sgout[0], dctx->aad, prot->aad_size);
1595
1596                 err = tls_setup_from_iter(out_iov, data_len, &pages, &sgout[1],
1597                                           (n_sgout - 1 - tail_pages));
1598                 if (err < 0)
1599                         goto exit_free_pages;
1600
1601                 if (prot->tail_size) {
1602                         sg_unmark_end(&sgout[pages]);
1603                         sg_set_buf(&sgout[pages + 1], &dctx->tail,
1604                                    prot->tail_size);
1605                         sg_mark_end(&sgout[pages + 1]);
1606                 }
1607         } else if (out_sg) {
1608                 memcpy(sgout, out_sg, n_sgout * sizeof(*sgout));
1609         }
1610         dctx->free_sgout = !!pages;
1611
1612         /* Prepare and submit AEAD request */
1613         err = tls_do_decryption(sk, sgin, sgout, dctx->iv,
1614                                 data_len + prot->tail_size, aead_req, darg);
1615         if (err) {
1616                 if (darg->async_done)
1617                         goto exit_free_skb;
1618                 goto exit_free_pages;
1619         }
1620
1621         darg->skb = clear_skb ?: tls_strp_msg(ctx);
1622         clear_skb = NULL;
1623
1624         if (unlikely(darg->async)) {
1625                 err = tls_strp_msg_hold(&ctx->strp, &ctx->async_hold);
1626                 if (err)
1627                         __skb_queue_tail(&ctx->async_hold, darg->skb);
1628                 return err;
1629         }
1630
1631         if (unlikely(darg->async_done))
1632                 return 0;
1633
1634         if (prot->tail_size)
1635                 darg->tail = dctx->tail;
1636
1637 exit_free_pages:
1638         /* Release the pages in case iov was mapped to pages */
1639         for (; pages > 0; pages--)
1640                 put_page(sg_page(&sgout[pages]));
1641 exit_free:
1642         kfree(mem);
1643 exit_free_skb:
1644         consume_skb(clear_skb);
1645         return err;
1646 }
1647
1648 static int
1649 tls_decrypt_sw(struct sock *sk, struct tls_context *tls_ctx,
1650                struct msghdr *msg, struct tls_decrypt_arg *darg)
1651 {
1652         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1653         struct tls_prot_info *prot = &tls_ctx->prot_info;
1654         struct strp_msg *rxm;
1655         int pad, err;
1656
1657         err = tls_decrypt_sg(sk, &msg->msg_iter, NULL, darg);
1658         if (err < 0) {
1659                 if (err == -EBADMSG)
1660                         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSDECRYPTERROR);
1661                 return err;
1662         }
1663         /* keep going even for ->async, the code below is TLS 1.3 */
1664
1665         /* If opportunistic TLS 1.3 ZC failed retry without ZC */
1666         if (unlikely(darg->zc && prot->version == TLS_1_3_VERSION &&
1667                      darg->tail != TLS_RECORD_TYPE_DATA)) {
1668                 darg->zc = false;
1669                 if (!darg->tail)
1670                         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSRXNOPADVIOL);
1671                 TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSDECRYPTRETRY);
1672                 return tls_decrypt_sw(sk, tls_ctx, msg, darg);
1673         }
1674
1675         pad = tls_padding_length(prot, darg->skb, darg);
1676         if (pad < 0) {
1677                 if (darg->skb != tls_strp_msg(ctx))
1678                         consume_skb(darg->skb);
1679                 return pad;
1680         }
1681
1682         rxm = strp_msg(darg->skb);
1683         rxm->full_len -= pad;
1684
1685         return 0;
1686 }
1687
1688 static int
1689 tls_decrypt_device(struct sock *sk, struct msghdr *msg,
1690                    struct tls_context *tls_ctx, struct tls_decrypt_arg *darg)
1691 {
1692         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1693         struct tls_prot_info *prot = &tls_ctx->prot_info;
1694         struct strp_msg *rxm;
1695         int pad, err;
1696
1697         if (tls_ctx->rx_conf != TLS_HW)
1698                 return 0;
1699
1700         err = tls_device_decrypted(sk, tls_ctx);
1701         if (err <= 0)
1702                 return err;
1703
1704         pad = tls_padding_length(prot, tls_strp_msg(ctx), darg);
1705         if (pad < 0)
1706                 return pad;
1707
1708         darg->async = false;
1709         darg->skb = tls_strp_msg(ctx);
1710         /* ->zc downgrade check, in case TLS 1.3 gets here */
1711         darg->zc &= !(prot->version == TLS_1_3_VERSION &&
1712                       tls_msg(darg->skb)->control != TLS_RECORD_TYPE_DATA);
1713
1714         rxm = strp_msg(darg->skb);
1715         rxm->full_len -= pad;
1716
1717         if (!darg->zc) {
1718                 /* Non-ZC case needs a real skb */
1719                 darg->skb = tls_strp_msg_detach(ctx);
1720                 if (!darg->skb)
1721                         return -ENOMEM;
1722         } else {
1723                 unsigned int off, len;
1724
1725                 /* In ZC case nobody cares about the output skb.
1726                  * Just copy the data here. Note the skb is not fully trimmed.
1727                  */
1728                 off = rxm->offset + prot->prepend_size;
1729                 len = rxm->full_len - prot->overhead_size;
1730
1731                 err = skb_copy_datagram_msg(darg->skb, off, msg, len);
1732                 if (err)
1733                         return err;
1734         }
1735         return 1;
1736 }
1737
1738 static int tls_check_pending_rekey(struct sock *sk, struct tls_context *ctx,
1739                                    struct sk_buff *skb)
1740 {
1741         const struct strp_msg *rxm = strp_msg(skb);
1742         const struct tls_msg *tlm = tls_msg(skb);
1743         char hs_type;
1744         int err;
1745
1746         if (likely(tlm->control != TLS_RECORD_TYPE_HANDSHAKE))
1747                 return 0;
1748
1749         if (rxm->full_len < 1)
1750                 return 0;
1751
1752         err = skb_copy_bits(skb, rxm->offset, &hs_type, 1);
1753         if (err < 0) {
1754                 DEBUG_NET_WARN_ON_ONCE(1);
1755                 return err;
1756         }
1757
1758         if (hs_type == TLS_HANDSHAKE_KEYUPDATE) {
1759                 struct tls_sw_context_rx *rx_ctx = ctx->priv_ctx_rx;
1760
1761                 WRITE_ONCE(rx_ctx->key_update_pending, true);
1762                 TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSRXREKEYRECEIVED);
1763         }
1764
1765         return 0;
1766 }
1767
1768 static int tls_rx_one_record(struct sock *sk, struct msghdr *msg,
1769                              struct tls_decrypt_arg *darg)
1770 {
1771         struct tls_context *tls_ctx = tls_get_ctx(sk);
1772         struct tls_prot_info *prot = &tls_ctx->prot_info;
1773         struct strp_msg *rxm;
1774         int err;
1775
1776         err = tls_decrypt_device(sk, msg, tls_ctx, darg);
1777         if (!err)
1778                 err = tls_decrypt_sw(sk, tls_ctx, msg, darg);
1779         if (err < 0)
1780                 return err;
1781
1782         rxm = strp_msg(darg->skb);
1783         rxm->offset += prot->prepend_size;
1784         rxm->full_len -= prot->overhead_size;
1785         tls_advance_record_sn(sk, prot, &tls_ctx->rx);
1786
1787         return tls_check_pending_rekey(sk, tls_ctx, darg->skb);
1788 }
1789
1790 int decrypt_skb(struct sock *sk, struct scatterlist *sgout)
1791 {
1792         struct tls_decrypt_arg darg = { .zc = true, };
1793
1794         return tls_decrypt_sg(sk, NULL, sgout, &darg);
1795 }
1796
1797 static int tls_record_content_type(struct msghdr *msg, struct tls_msg *tlm,
1798                                    u8 *control)
1799 {
1800         int err;
1801
1802         if (!*control) {
1803                 *control = tlm->control;
1804                 if (!*control)
1805                         return -EBADMSG;
1806
1807                 err = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1808                                sizeof(*control), control);
1809                 if (*control != TLS_RECORD_TYPE_DATA) {
1810                         if (err || msg->msg_flags & MSG_CTRUNC)
1811                                 return -EIO;
1812                 }
1813         } else if (*control != tlm->control) {
1814                 return 0;
1815         }
1816
1817         return 1;
1818 }
1819
1820 static void tls_rx_rec_done(struct tls_sw_context_rx *ctx)
1821 {
1822         tls_strp_msg_done(&ctx->strp);
1823 }
1824
1825 /* This function traverses the rx_list in tls receive context to copies the
1826  * decrypted records into the buffer provided by caller zero copy is not
1827  * true. Further, the records are removed from the rx_list if it is not a peek
1828  * case and the record has been consumed completely.
1829  */
1830 static int process_rx_list(struct tls_sw_context_rx *ctx,
1831                            struct msghdr *msg,
1832                            u8 *control,
1833                            size_t skip,
1834                            size_t len,
1835                            bool is_peek,
1836                            bool *more)
1837 {
1838         struct sk_buff *skb = skb_peek(&ctx->rx_list);
1839         struct tls_msg *tlm;
1840         ssize_t copied = 0;
1841         int err;
1842
1843         while (skip && skb) {
1844                 struct strp_msg *rxm = strp_msg(skb);
1845                 tlm = tls_msg(skb);
1846
1847                 err = tls_record_content_type(msg, tlm, control);
1848                 if (err <= 0)
1849                         goto more;
1850
1851                 if (skip < rxm->full_len)
1852                         break;
1853
1854                 skip = skip - rxm->full_len;
1855                 skb = skb_peek_next(skb, &ctx->rx_list);
1856         }
1857
1858         while (len && skb) {
1859                 struct sk_buff *next_skb;
1860                 struct strp_msg *rxm = strp_msg(skb);
1861                 int chunk = min_t(unsigned int, rxm->full_len - skip, len);
1862
1863                 tlm = tls_msg(skb);
1864
1865                 err = tls_record_content_type(msg, tlm, control);
1866                 if (err <= 0)
1867                         goto more;
1868
1869                 err = skb_copy_datagram_msg(skb, rxm->offset + skip,
1870                                             msg, chunk);
1871                 if (err < 0)
1872                         goto more;
1873
1874                 len = len - chunk;
1875                 copied = copied + chunk;
1876
1877                 /* Consume the data from record if it is non-peek case*/
1878                 if (!is_peek) {
1879                         rxm->offset = rxm->offset + chunk;
1880                         rxm->full_len = rxm->full_len - chunk;
1881
1882                         /* Return if there is unconsumed data in the record */
1883                         if (rxm->full_len - skip)
1884                                 break;
1885                 }
1886
1887                 /* The remaining skip-bytes must lie in 1st record in rx_list.
1888                  * So from the 2nd record, 'skip' should be 0.
1889                  */
1890                 skip = 0;
1891
1892                 if (msg)
1893                         msg->msg_flags |= MSG_EOR;
1894
1895                 next_skb = skb_peek_next(skb, &ctx->rx_list);
1896
1897                 if (!is_peek) {
1898                         __skb_unlink(skb, &ctx->rx_list);
1899                         consume_skb(skb);
1900                 }
1901
1902                 skb = next_skb;
1903         }
1904         err = 0;
1905
1906 out:
1907         return copied ? : err;
1908 more:
1909         if (more)
1910                 *more = true;
1911         goto out;
1912 }
1913
1914 static bool
1915 tls_read_flush_backlog(struct sock *sk, struct tls_prot_info *prot,
1916                        size_t len_left, size_t decrypted, ssize_t done,
1917                        size_t *flushed_at)
1918 {
1919         size_t max_rec;
1920
1921         if (len_left <= decrypted)
1922                 return false;
1923
1924         max_rec = prot->overhead_size - prot->tail_size + TLS_MAX_PAYLOAD_SIZE;
1925         if (done - *flushed_at < SZ_128K && tcp_inq(sk) > max_rec)
1926                 return false;
1927
1928         *flushed_at = done;
1929         return sk_flush_backlog(sk);
1930 }
1931
1932 static int tls_rx_reader_acquire(struct sock *sk, struct tls_sw_context_rx *ctx,
1933                                  bool nonblock)
1934 {
1935         long timeo;
1936         int ret;
1937
1938         timeo = sock_rcvtimeo(sk, nonblock);
1939
1940         while (unlikely(ctx->reader_present)) {
1941                 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1942
1943                 ctx->reader_contended = 1;
1944
1945                 add_wait_queue(&ctx->wq, &wait);
1946                 ret = sk_wait_event(sk, &timeo,
1947                                     !READ_ONCE(ctx->reader_present), &wait);
1948                 remove_wait_queue(&ctx->wq, &wait);
1949
1950                 if (timeo <= 0)
1951                         return -EAGAIN;
1952                 if (signal_pending(current))
1953                         return sock_intr_errno(timeo);
1954                 if (ret < 0)
1955                         return ret;
1956         }
1957
1958         WRITE_ONCE(ctx->reader_present, 1);
1959
1960         return 0;
1961 }
1962
1963 static int tls_rx_reader_lock(struct sock *sk, struct tls_sw_context_rx *ctx,
1964                               bool nonblock)
1965 {
1966         int err;
1967
1968         lock_sock(sk);
1969         err = tls_rx_reader_acquire(sk, ctx, nonblock);
1970         if (err)
1971                 release_sock(sk);
1972         return err;
1973 }
1974
1975 static void tls_rx_reader_release(struct sock *sk, struct tls_sw_context_rx *ctx)
1976 {
1977         if (unlikely(ctx->reader_contended)) {
1978                 if (wq_has_sleeper(&ctx->wq))
1979                         wake_up(&ctx->wq);
1980                 else
1981                         ctx->reader_contended = 0;
1982
1983                 WARN_ON_ONCE(!ctx->reader_present);
1984         }
1985
1986         WRITE_ONCE(ctx->reader_present, 0);
1987 }
1988
1989 static void tls_rx_reader_unlock(struct sock *sk, struct tls_sw_context_rx *ctx)
1990 {
1991         tls_rx_reader_release(sk, ctx);
1992         release_sock(sk);
1993 }
1994
1995 int tls_sw_recvmsg(struct sock *sk,
1996                    struct msghdr *msg,
1997                    size_t len,
1998                    int flags,
1999                    int *addr_len)
2000 {
2001         struct tls_context *tls_ctx = tls_get_ctx(sk);
2002         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2003         struct tls_prot_info *prot = &tls_ctx->prot_info;
2004         ssize_t decrypted = 0, async_copy_bytes = 0;
2005         struct sk_psock *psock;
2006         unsigned char control = 0;
2007         size_t flushed_at = 0;
2008         struct strp_msg *rxm;
2009         struct tls_msg *tlm;
2010         ssize_t copied = 0;
2011         ssize_t peeked = 0;
2012         bool async = false;
2013         int target, err;
2014         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
2015         bool is_peek = flags & MSG_PEEK;
2016         bool rx_more = false;
2017         bool released = true;
2018         bool bpf_strp_enabled;
2019         bool zc_capable;
2020
2021         if (unlikely(flags & MSG_ERRQUEUE))
2022                 return sock_recv_errqueue(sk, msg, len, SOL_IP, IP_RECVERR);
2023
2024         err = tls_rx_reader_lock(sk, ctx, flags & MSG_DONTWAIT);
2025         if (err < 0)
2026                 return err;
2027         psock = sk_psock_get(sk);
2028         bpf_strp_enabled = sk_psock_strp_enabled(psock);
2029
2030         /* If crypto failed the connection is broken */
2031         err = ctx->async_wait.err;
2032         if (err)
2033                 goto end;
2034
2035         /* Process pending decrypted records. It must be non-zero-copy */
2036         err = process_rx_list(ctx, msg, &control, 0, len, is_peek, &rx_more);
2037         if (err < 0)
2038                 goto end;
2039
2040         copied = err;
2041         if (len <= copied || (copied && control != TLS_RECORD_TYPE_DATA) || rx_more)
2042                 goto end;
2043
2044         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2045         len = len - copied;
2046
2047         zc_capable = !bpf_strp_enabled && !is_kvec && !is_peek &&
2048                 ctx->zc_capable;
2049         decrypted = 0;
2050         while (len && (decrypted + copied < target || tls_strp_msg_ready(ctx))) {
2051                 struct tls_decrypt_arg darg;
2052                 int to_decrypt, chunk;
2053
2054                 err = tls_rx_rec_wait(sk, psock, flags & MSG_DONTWAIT,
2055                                       released);
2056                 if (err <= 0) {
2057                         if (psock) {
2058                                 chunk = sk_msg_recvmsg(sk, psock, msg, len,
2059                                                        flags);
2060                                 if (chunk > 0) {
2061                                         decrypted += chunk;
2062                                         len -= chunk;
2063                                         continue;
2064                                 }
2065                         }
2066                         goto recv_end;
2067                 }
2068
2069                 memset(&darg.inargs, 0, sizeof(darg.inargs));
2070
2071                 rxm = strp_msg(tls_strp_msg(ctx));
2072                 tlm = tls_msg(tls_strp_msg(ctx));
2073
2074                 to_decrypt = rxm->full_len - prot->overhead_size;
2075
2076                 if (zc_capable && to_decrypt <= len &&
2077                     tlm->control == TLS_RECORD_TYPE_DATA)
2078                         darg.zc = true;
2079
2080                 /* Do not use async mode if record is non-data */
2081                 if (tlm->control == TLS_RECORD_TYPE_DATA && !bpf_strp_enabled)
2082                         darg.async = ctx->async_capable;
2083                 else
2084                         darg.async = false;
2085
2086                 err = tls_rx_one_record(sk, msg, &darg);
2087                 if (err < 0) {
2088                         tls_err_abort(sk, -EBADMSG);
2089                         goto recv_end;
2090                 }
2091
2092                 async |= darg.async;
2093
2094                 /* If the type of records being processed is not known yet,
2095                  * set it to record type just dequeued. If it is already known,
2096                  * but does not match the record type just dequeued, go to end.
2097                  * We always get record type here since for tls1.2, record type
2098                  * is known just after record is dequeued from stream parser.
2099                  * For tls1.3, we disable async.
2100                  */
2101                 err = tls_record_content_type(msg, tls_msg(darg.skb), &control);
2102                 if (err <= 0) {
2103                         DEBUG_NET_WARN_ON_ONCE(darg.zc);
2104                         tls_rx_rec_done(ctx);
2105 put_on_rx_list_err:
2106                         __skb_queue_tail(&ctx->rx_list, darg.skb);
2107                         goto recv_end;
2108                 }
2109
2110                 /* periodically flush backlog, and feed strparser */
2111                 released = tls_read_flush_backlog(sk, prot, len, to_decrypt,
2112                                                   decrypted + copied,
2113                                                   &flushed_at);
2114
2115                 /* TLS 1.3 may have updated the length by more than overhead */
2116                 rxm = strp_msg(darg.skb);
2117                 chunk = rxm->full_len;
2118                 tls_rx_rec_done(ctx);
2119
2120                 if (!darg.zc) {
2121                         bool partially_consumed = chunk > len;
2122                         struct sk_buff *skb = darg.skb;
2123
2124                         DEBUG_NET_WARN_ON_ONCE(darg.skb == ctx->strp.anchor);
2125
2126                         if (async) {
2127                                 /* TLS 1.2-only, to_decrypt must be text len */
2128                                 chunk = min_t(int, to_decrypt, len);
2129                                 async_copy_bytes += chunk;
2130 put_on_rx_list:
2131                                 decrypted += chunk;
2132                                 len -= chunk;
2133                                 __skb_queue_tail(&ctx->rx_list, skb);
2134                                 if (unlikely(control != TLS_RECORD_TYPE_DATA))
2135                                         break;
2136                                 continue;
2137                         }
2138
2139                         if (bpf_strp_enabled) {
2140                                 released = true;
2141                                 err = sk_psock_tls_strp_read(psock, skb);
2142                                 if (err != __SK_PASS) {
2143                                         rxm->offset = rxm->offset + rxm->full_len;
2144                                         rxm->full_len = 0;
2145                                         if (err == __SK_DROP)
2146                                                 consume_skb(skb);
2147                                         continue;
2148                                 }
2149                         }
2150
2151                         if (partially_consumed)
2152                                 chunk = len;
2153
2154                         err = skb_copy_datagram_msg(skb, rxm->offset,
2155                                                     msg, chunk);
2156                         if (err < 0)
2157                                 goto put_on_rx_list_err;
2158
2159                         if (is_peek) {
2160                                 peeked += chunk;
2161                                 goto put_on_rx_list;
2162                         }
2163
2164                         if (partially_consumed) {
2165                                 rxm->offset += chunk;
2166                                 rxm->full_len -= chunk;
2167                                 goto put_on_rx_list;
2168                         }
2169
2170                         consume_skb(skb);
2171                 }
2172
2173                 decrypted += chunk;
2174                 len -= chunk;
2175
2176                 /* Return full control message to userspace before trying
2177                  * to parse another message type
2178                  */
2179                 msg->msg_flags |= MSG_EOR;
2180                 if (control != TLS_RECORD_TYPE_DATA)
2181                         break;
2182         }
2183
2184 recv_end:
2185         if (async) {
2186                 int ret;
2187
2188                 /* Wait for all previously submitted records to be decrypted */
2189                 ret = tls_decrypt_async_wait(ctx);
2190                 __skb_queue_purge(&ctx->async_hold);
2191
2192                 if (ret) {
2193                         if (err >= 0 || err == -EINPROGRESS)
2194                                 err = ret;
2195                         goto end;
2196                 }
2197
2198                 /* Drain records from the rx_list & copy if required */
2199                 if (is_peek)
2200                         err = process_rx_list(ctx, msg, &control, copied + peeked,
2201                                               decrypted - peeked, is_peek, NULL);
2202                 else
2203                         err = process_rx_list(ctx, msg, &control, 0,
2204                                               async_copy_bytes, is_peek, NULL);
2205
2206                 /* we could have copied less than we wanted, and possibly nothing */
2207                 decrypted += max(err, 0) - async_copy_bytes;
2208         }
2209
2210         copied += decrypted;
2211
2212 end:
2213         tls_rx_reader_unlock(sk, ctx);
2214         if (psock)
2215                 sk_psock_put(sk, psock);
2216         return copied ? : err;
2217 }
2218
2219 ssize_t tls_sw_splice_read(struct socket *sock,  loff_t *ppos,
2220                            struct pipe_inode_info *pipe,
2221                            size_t len, unsigned int flags)
2222 {
2223         struct tls_context *tls_ctx = tls_get_ctx(sock->sk);
2224         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2225         struct strp_msg *rxm = NULL;
2226         struct sock *sk = sock->sk;
2227         struct tls_msg *tlm;
2228         struct sk_buff *skb;
2229         ssize_t copied = 0;
2230         int chunk;
2231         int err;
2232
2233         err = tls_rx_reader_lock(sk, ctx, flags & SPLICE_F_NONBLOCK);
2234         if (err < 0)
2235                 return err;
2236
2237         if (!skb_queue_empty(&ctx->rx_list)) {
2238                 skb = __skb_dequeue(&ctx->rx_list);
2239         } else {
2240                 struct tls_decrypt_arg darg;
2241
2242                 err = tls_rx_rec_wait(sk, NULL, flags & SPLICE_F_NONBLOCK,
2243                                       true);
2244                 if (err <= 0)
2245                         goto splice_read_end;
2246
2247                 memset(&darg.inargs, 0, sizeof(darg.inargs));
2248
2249                 err = tls_rx_one_record(sk, NULL, &darg);
2250                 if (err < 0) {
2251                         tls_err_abort(sk, -EBADMSG);
2252                         goto splice_read_end;
2253                 }
2254
2255                 tls_rx_rec_done(ctx);
2256                 skb = darg.skb;
2257         }
2258
2259         rxm = strp_msg(skb);
2260         tlm = tls_msg(skb);
2261
2262         /* splice does not support reading control messages */
2263         if (tlm->control != TLS_RECORD_TYPE_DATA) {
2264                 err = -EINVAL;
2265                 goto splice_requeue;
2266         }
2267
2268         chunk = min_t(unsigned int, rxm->full_len, len);
2269         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, chunk, flags);
2270         if (copied < 0)
2271                 goto splice_requeue;
2272
2273         if (chunk < rxm->full_len) {
2274                 rxm->offset += len;
2275                 rxm->full_len -= len;
2276                 goto splice_requeue;
2277         }
2278
2279         consume_skb(skb);
2280
2281 splice_read_end:
2282         tls_rx_reader_unlock(sk, ctx);
2283         return copied ? : err;
2284
2285 splice_requeue:
2286         __skb_queue_head(&ctx->rx_list, skb);
2287         goto splice_read_end;
2288 }
2289
2290 int tls_sw_read_sock(struct sock *sk, read_descriptor_t *desc,
2291                      sk_read_actor_t read_actor)
2292 {
2293         struct tls_context *tls_ctx = tls_get_ctx(sk);
2294         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2295         struct tls_prot_info *prot = &tls_ctx->prot_info;
2296         struct strp_msg *rxm = NULL;
2297         struct sk_buff *skb = NULL;
2298         struct sk_psock *psock;
2299         size_t flushed_at = 0;
2300         bool released = true;
2301         struct tls_msg *tlm;
2302         ssize_t copied = 0;
2303         ssize_t decrypted;
2304         int err, used;
2305
2306         psock = sk_psock_get(sk);
2307         if (psock) {
2308                 sk_psock_put(sk, psock);
2309                 return -EINVAL;
2310         }
2311         err = tls_rx_reader_acquire(sk, ctx, true);
2312         if (err < 0)
2313                 return err;
2314
2315         /* If crypto failed the connection is broken */
2316         err = ctx->async_wait.err;
2317         if (err)
2318                 goto read_sock_end;
2319
2320         decrypted = 0;
2321         do {
2322                 if (!skb_queue_empty(&ctx->rx_list)) {
2323                         skb = __skb_dequeue(&ctx->rx_list);
2324                         rxm = strp_msg(skb);
2325                         tlm = tls_msg(skb);
2326                 } else {
2327                         struct tls_decrypt_arg darg;
2328
2329                         err = tls_rx_rec_wait(sk, NULL, true, released);
2330                         if (err <= 0)
2331                                 goto read_sock_end;
2332
2333                         memset(&darg.inargs, 0, sizeof(darg.inargs));
2334
2335                         err = tls_rx_one_record(sk, NULL, &darg);
2336                         if (err < 0) {
2337                                 tls_err_abort(sk, -EBADMSG);
2338                                 goto read_sock_end;
2339                         }
2340
2341                         released = tls_read_flush_backlog(sk, prot, INT_MAX,
2342                                                           0, decrypted,
2343                                                           &flushed_at);
2344                         skb = darg.skb;
2345                         rxm = strp_msg(skb);
2346                         tlm = tls_msg(skb);
2347                         decrypted += rxm->full_len;
2348
2349                         tls_rx_rec_done(ctx);
2350                 }
2351
2352                 /* read_sock does not support reading control messages */
2353                 if (tlm->control != TLS_RECORD_TYPE_DATA) {
2354                         err = -EINVAL;
2355                         goto read_sock_requeue;
2356                 }
2357
2358                 used = read_actor(desc, skb, rxm->offset, rxm->full_len);
2359                 if (used <= 0) {
2360                         if (!copied)
2361                                 err = used;
2362                         goto read_sock_requeue;
2363                 }
2364                 copied += used;
2365                 if (used < rxm->full_len) {
2366                         rxm->offset += used;
2367                         rxm->full_len -= used;
2368                         if (!desc->count)
2369                                 goto read_sock_requeue;
2370                 } else {
2371                         consume_skb(skb);
2372                         if (!desc->count)
2373                                 skb = NULL;
2374                 }
2375         } while (skb);
2376
2377 read_sock_end:
2378         tls_rx_reader_release(sk, ctx);
2379         return copied ? : err;
2380
2381 read_sock_requeue:
2382         __skb_queue_head(&ctx->rx_list, skb);
2383         goto read_sock_end;
2384 }
2385
2386 bool tls_sw_sock_is_readable(struct sock *sk)
2387 {
2388         struct tls_context *tls_ctx = tls_get_ctx(sk);
2389         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2390         bool ingress_empty = true;
2391         struct sk_psock *psock;
2392
2393         rcu_read_lock();
2394         psock = sk_psock(sk);
2395         if (psock)
2396                 ingress_empty = list_empty(&psock->ingress_msg);
2397         rcu_read_unlock();
2398
2399         return !ingress_empty || tls_strp_msg_ready(ctx) ||
2400                 !skb_queue_empty(&ctx->rx_list);
2401 }
2402
2403 int tls_rx_msg_size(struct tls_strparser *strp, struct sk_buff *skb)
2404 {
2405         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2406         struct tls_prot_info *prot = &tls_ctx->prot_info;
2407         char header[TLS_HEADER_SIZE + TLS_MAX_IV_SIZE];
2408         size_t cipher_overhead;
2409         size_t data_len = 0;
2410         int ret;
2411
2412         /* Verify that we have a full TLS header, or wait for more data */
2413         if (strp->stm.offset + prot->prepend_size > skb->len)
2414                 return 0;
2415
2416         /* Sanity-check size of on-stack buffer. */
2417         if (WARN_ON(prot->prepend_size > sizeof(header))) {
2418                 ret = -EINVAL;
2419                 goto read_failure;
2420         }
2421
2422         /* Linearize header to local buffer */
2423         ret = skb_copy_bits(skb, strp->stm.offset, header, prot->prepend_size);
2424         if (ret < 0)
2425                 goto read_failure;
2426
2427         strp->mark = header[0];
2428
2429         data_len = ((header[4] & 0xFF) | (header[3] << 8));
2430
2431         cipher_overhead = prot->tag_size;
2432         if (prot->version != TLS_1_3_VERSION &&
2433             prot->cipher_type != TLS_CIPHER_CHACHA20_POLY1305)
2434                 cipher_overhead += prot->iv_size;
2435
2436         if (data_len > TLS_MAX_PAYLOAD_SIZE + cipher_overhead +
2437             prot->tail_size) {
2438                 ret = -EMSGSIZE;
2439                 goto read_failure;
2440         }
2441         if (data_len < cipher_overhead) {
2442                 ret = -EBADMSG;
2443                 goto read_failure;
2444         }
2445
2446         /* Note that both TLS1.3 and TLS1.2 use TLS_1_2 version here */
2447         if (header[1] != TLS_1_2_VERSION_MINOR ||
2448             header[2] != TLS_1_2_VERSION_MAJOR) {
2449                 ret = -EINVAL;
2450                 goto read_failure;
2451         }
2452
2453         tls_device_rx_resync_new_rec(strp->sk, data_len + TLS_HEADER_SIZE,
2454                                      TCP_SKB_CB(skb)->seq + strp->stm.offset);
2455         return data_len + TLS_HEADER_SIZE;
2456
2457 read_failure:
2458         tls_err_abort(strp->sk, ret);
2459
2460         return ret;
2461 }
2462
2463 void tls_rx_msg_ready(struct tls_strparser *strp)
2464 {
2465         struct tls_sw_context_rx *ctx;
2466
2467         ctx = container_of(strp, struct tls_sw_context_rx, strp);
2468         ctx->saved_data_ready(strp->sk);
2469 }
2470
2471 static void tls_data_ready(struct sock *sk)
2472 {
2473         struct tls_context *tls_ctx = tls_get_ctx(sk);
2474         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2475         struct sk_psock *psock;
2476         gfp_t alloc_save;
2477
2478         trace_sk_data_ready(sk);
2479
2480         alloc_save = sk->sk_allocation;
2481         sk->sk_allocation = GFP_ATOMIC;
2482         tls_strp_data_ready(&ctx->strp);
2483         sk->sk_allocation = alloc_save;
2484
2485         psock = sk_psock_get(sk);
2486         if (psock) {
2487                 if (!list_empty(&psock->ingress_msg))
2488                         ctx->saved_data_ready(sk);
2489                 sk_psock_put(sk, psock);
2490         }
2491 }
2492
2493 void tls_sw_cancel_work_tx(struct tls_context *tls_ctx)
2494 {
2495         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2496
2497         set_bit(BIT_TX_CLOSING, &ctx->tx_bitmask);
2498         set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask);
2499         cancel_delayed_work_sync(&ctx->tx_work.work);
2500 }
2501
2502 void tls_sw_release_resources_tx(struct sock *sk)
2503 {
2504         struct tls_context *tls_ctx = tls_get_ctx(sk);
2505         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2506         struct tls_rec *rec, *tmp;
2507
2508         /* Wait for any pending async encryptions to complete */
2509         tls_encrypt_async_wait(ctx);
2510
2511         tls_tx_records(sk, -1);
2512
2513         /* Free up un-sent records in tx_list. First, free
2514          * the partially sent record if any at head of tx_list.
2515          */
2516         if (tls_ctx->partially_sent_record) {
2517                 tls_free_partial_record(sk, tls_ctx);
2518                 rec = list_first_entry(&ctx->tx_list,
2519                                        struct tls_rec, list);
2520                 list_del(&rec->list);
2521                 sk_msg_free(sk, &rec->msg_plaintext);
2522                 kfree(rec);
2523         }
2524
2525         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
2526                 list_del(&rec->list);
2527                 sk_msg_free(sk, &rec->msg_encrypted);
2528                 sk_msg_free(sk, &rec->msg_plaintext);
2529                 kfree(rec);
2530         }
2531
2532         crypto_free_aead(ctx->aead_send);
2533         tls_free_open_rec(sk);
2534 }
2535
2536 void tls_sw_free_ctx_tx(struct tls_context *tls_ctx)
2537 {
2538         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2539
2540         kfree(ctx);
2541 }
2542
2543 void tls_sw_release_resources_rx(struct sock *sk)
2544 {
2545         struct tls_context *tls_ctx = tls_get_ctx(sk);
2546         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2547
2548         if (ctx->aead_recv) {
2549                 __skb_queue_purge(&ctx->rx_list);
2550                 crypto_free_aead(ctx->aead_recv);
2551                 tls_strp_stop(&ctx->strp);
2552                 /* If tls_sw_strparser_arm() was not called (cleanup paths)
2553                  * we still want to tls_strp_stop(), but sk->sk_data_ready was
2554                  * never swapped.
2555                  */
2556                 if (ctx->saved_data_ready) {
2557                         write_lock_bh(&sk->sk_callback_lock);
2558                         sk->sk_data_ready = ctx->saved_data_ready;
2559                         write_unlock_bh(&sk->sk_callback_lock);
2560                 }
2561         }
2562 }
2563
2564 void tls_sw_strparser_done(struct tls_context *tls_ctx)
2565 {
2566         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2567
2568         tls_strp_done(&ctx->strp);
2569 }
2570
2571 void tls_sw_free_ctx_rx(struct tls_context *tls_ctx)
2572 {
2573         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2574
2575         kfree(ctx);
2576 }
2577
2578 void tls_sw_free_resources_rx(struct sock *sk)
2579 {
2580         struct tls_context *tls_ctx = tls_get_ctx(sk);
2581
2582         tls_sw_release_resources_rx(sk);
2583         tls_sw_free_ctx_rx(tls_ctx);
2584 }
2585
2586 /* The work handler to transmitt the encrypted records in tx_list */
2587 static void tx_work_handler(struct work_struct *work)
2588 {
2589         struct delayed_work *delayed_work = to_delayed_work(work);
2590         struct tx_work *tx_work = container_of(delayed_work,
2591                                                struct tx_work, work);
2592         struct sock *sk = tx_work->sk;
2593         struct tls_context *tls_ctx = tls_get_ctx(sk);
2594         struct tls_sw_context_tx *ctx;
2595
2596         if (unlikely(!tls_ctx))
2597                 return;
2598
2599         ctx = tls_sw_ctx_tx(tls_ctx);
2600         if (test_bit(BIT_TX_CLOSING, &ctx->tx_bitmask))
2601                 return;
2602
2603         if (!test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
2604                 return;
2605
2606         if (mutex_trylock(&tls_ctx->tx_lock)) {
2607                 lock_sock(sk);
2608                 tls_tx_records(sk, -1);
2609                 release_sock(sk);
2610                 mutex_unlock(&tls_ctx->tx_lock);
2611         } else if (!test_and_set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
2612                 /* Someone is holding the tx_lock, they will likely run Tx
2613                  * and cancel the work on their way out of the lock section.
2614                  * Schedule a long delay just in case.
2615                  */
2616                 schedule_delayed_work(&ctx->tx_work.work, msecs_to_jiffies(10));
2617         }
2618 }
2619
2620 static bool tls_is_tx_ready(struct tls_sw_context_tx *ctx)
2621 {
2622         struct tls_rec *rec;
2623
2624         rec = list_first_entry_or_null(&ctx->tx_list, struct tls_rec, list);
2625         if (!rec)
2626                 return false;
2627
2628         return READ_ONCE(rec->tx_ready);
2629 }
2630
2631 void tls_sw_write_space(struct sock *sk, struct tls_context *ctx)
2632 {
2633         struct tls_sw_context_tx *tx_ctx = tls_sw_ctx_tx(ctx);
2634
2635         /* Schedule the transmission if tx list is ready */
2636         if (tls_is_tx_ready(tx_ctx) &&
2637             !test_and_set_bit(BIT_TX_SCHEDULED, &tx_ctx->tx_bitmask))
2638                 schedule_delayed_work(&tx_ctx->tx_work.work, 0);
2639 }
2640
2641 void tls_sw_strparser_arm(struct sock *sk, struct tls_context *tls_ctx)
2642 {
2643         struct tls_sw_context_rx *rx_ctx = tls_sw_ctx_rx(tls_ctx);
2644
2645         write_lock_bh(&sk->sk_callback_lock);
2646         rx_ctx->saved_data_ready = sk->sk_data_ready;
2647         sk->sk_data_ready = tls_data_ready;
2648         write_unlock_bh(&sk->sk_callback_lock);
2649 }
2650
2651 void tls_update_rx_zc_capable(struct tls_context *tls_ctx)
2652 {
2653         struct tls_sw_context_rx *rx_ctx = tls_sw_ctx_rx(tls_ctx);
2654
2655         rx_ctx->zc_capable = tls_ctx->rx_no_pad ||
2656                 tls_ctx->prot_info.version != TLS_1_3_VERSION;
2657 }
2658
2659 static struct tls_sw_context_tx *init_ctx_tx(struct tls_context *ctx, struct sock *sk)
2660 {
2661         struct tls_sw_context_tx *sw_ctx_tx;
2662
2663         if (!ctx->priv_ctx_tx) {
2664                 sw_ctx_tx = kzalloc(sizeof(*sw_ctx_tx), GFP_KERNEL);
2665                 if (!sw_ctx_tx)
2666                         return NULL;
2667         } else {
2668                 sw_ctx_tx = ctx->priv_ctx_tx;
2669         }
2670
2671         crypto_init_wait(&sw_ctx_tx->async_wait);
2672         atomic_set(&sw_ctx_tx->encrypt_pending, 1);
2673         INIT_LIST_HEAD(&sw_ctx_tx->tx_list);
2674         INIT_DELAYED_WORK(&sw_ctx_tx->tx_work.work, tx_work_handler);
2675         sw_ctx_tx->tx_work.sk = sk;
2676
2677         return sw_ctx_tx;
2678 }
2679
2680 static struct tls_sw_context_rx *init_ctx_rx(struct tls_context *ctx)
2681 {
2682         struct tls_sw_context_rx *sw_ctx_rx;
2683
2684         if (!ctx->priv_ctx_rx) {
2685                 sw_ctx_rx = kzalloc(sizeof(*sw_ctx_rx), GFP_KERNEL);
2686                 if (!sw_ctx_rx)
2687                         return NULL;
2688         } else {
2689                 sw_ctx_rx = ctx->priv_ctx_rx;
2690         }
2691
2692         crypto_init_wait(&sw_ctx_rx->async_wait);
2693         atomic_set(&sw_ctx_rx->decrypt_pending, 1);
2694         init_waitqueue_head(&sw_ctx_rx->wq);
2695         skb_queue_head_init(&sw_ctx_rx->rx_list);
2696         skb_queue_head_init(&sw_ctx_rx->async_hold);
2697
2698         return sw_ctx_rx;
2699 }
2700
2701 int init_prot_info(struct tls_prot_info *prot,
2702                    const struct tls_crypto_info *crypto_info,
2703                    const struct tls_cipher_desc *cipher_desc)
2704 {
2705         u16 nonce_size = cipher_desc->nonce;
2706
2707         if (crypto_info->version == TLS_1_3_VERSION) {
2708                 nonce_size = 0;
2709                 prot->aad_size = TLS_HEADER_SIZE;
2710                 prot->tail_size = 1;
2711         } else {
2712                 prot->aad_size = TLS_AAD_SPACE_SIZE;
2713                 prot->tail_size = 0;
2714         }
2715
2716         /* Sanity-check the sizes for stack allocations. */
2717         if (nonce_size > TLS_MAX_IV_SIZE || prot->aad_size > TLS_MAX_AAD_SIZE)
2718                 return -EINVAL;
2719
2720         prot->version = crypto_info->version;
2721         prot->cipher_type = crypto_info->cipher_type;
2722         prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
2723         prot->tag_size = cipher_desc->tag;
2724         prot->overhead_size = prot->prepend_size + prot->tag_size + prot->tail_size;
2725         prot->iv_size = cipher_desc->iv;
2726         prot->salt_size = cipher_desc->salt;
2727         prot->rec_seq_size = cipher_desc->rec_seq;
2728
2729         return 0;
2730 }
2731
2732 static void tls_finish_key_update(struct sock *sk, struct tls_context *tls_ctx)
2733 {
2734         struct tls_sw_context_rx *ctx = tls_ctx->priv_ctx_rx;
2735
2736         WRITE_ONCE(ctx->key_update_pending, false);
2737         /* wake-up pre-existing poll() */
2738         ctx->saved_data_ready(sk);
2739 }
2740
2741 int tls_set_sw_offload(struct sock *sk, int tx,
2742                        struct tls_crypto_info *new_crypto_info)
2743 {
2744         struct tls_crypto_info *crypto_info, *src_crypto_info;
2745         struct tls_sw_context_tx *sw_ctx_tx = NULL;
2746         struct tls_sw_context_rx *sw_ctx_rx = NULL;
2747         const struct tls_cipher_desc *cipher_desc;
2748         char *iv, *rec_seq, *key, *salt;
2749         struct cipher_context *cctx;
2750         struct tls_prot_info *prot;
2751         struct crypto_aead **aead;
2752         struct tls_context *ctx;
2753         struct crypto_tfm *tfm;
2754         int rc = 0;
2755
2756         ctx = tls_get_ctx(sk);
2757         prot = &ctx->prot_info;
2758
2759         /* new_crypto_info != NULL means rekey */
2760         if (!new_crypto_info) {
2761                 if (tx) {
2762                         ctx->priv_ctx_tx = init_ctx_tx(ctx, sk);
2763                         if (!ctx->priv_ctx_tx)
2764                                 return -ENOMEM;
2765                 } else {
2766                         ctx->priv_ctx_rx = init_ctx_rx(ctx);
2767                         if (!ctx->priv_ctx_rx)
2768                                 return -ENOMEM;
2769                 }
2770         }
2771
2772         if (tx) {
2773                 sw_ctx_tx = ctx->priv_ctx_tx;
2774                 crypto_info = &ctx->crypto_send.info;
2775                 cctx = &ctx->tx;
2776                 aead = &sw_ctx_tx->aead_send;
2777         } else {
2778                 sw_ctx_rx = ctx->priv_ctx_rx;
2779                 crypto_info = &ctx->crypto_recv.info;
2780                 cctx = &ctx->rx;
2781                 aead = &sw_ctx_rx->aead_recv;
2782         }
2783
2784         src_crypto_info = new_crypto_info ?: crypto_info;
2785
2786         cipher_desc = get_cipher_desc(src_crypto_info->cipher_type);
2787         if (!cipher_desc) {
2788                 rc = -EINVAL;
2789                 goto free_priv;
2790         }
2791
2792         rc = init_prot_info(prot, src_crypto_info, cipher_desc);
2793         if (rc)
2794                 goto free_priv;
2795
2796         iv = crypto_info_iv(src_crypto_info, cipher_desc);
2797         key = crypto_info_key(src_crypto_info, cipher_desc);
2798         salt = crypto_info_salt(src_crypto_info, cipher_desc);
2799         rec_seq = crypto_info_rec_seq(src_crypto_info, cipher_desc);
2800
2801         if (!*aead) {
2802                 *aead = crypto_alloc_aead(cipher_desc->cipher_name, 0, 0);
2803                 if (IS_ERR(*aead)) {
2804                         rc = PTR_ERR(*aead);
2805                         *aead = NULL;
2806                         goto free_priv;
2807                 }
2808         }
2809
2810         ctx->push_pending_record = tls_sw_push_pending_record;
2811
2812         /* setkey is the last operation that could fail during a
2813          * rekey. if it succeeds, we can start modifying the
2814          * context.
2815          */
2816         rc = crypto_aead_setkey(*aead, key, cipher_desc->key);
2817         if (rc) {
2818                 if (new_crypto_info)
2819                         goto out;
2820                 else
2821                         goto free_aead;
2822         }
2823
2824         if (!new_crypto_info) {
2825                 rc = crypto_aead_setauthsize(*aead, prot->tag_size);
2826                 if (rc)
2827                         goto free_aead;
2828         }
2829
2830         if (!tx && !new_crypto_info) {
2831                 tfm = crypto_aead_tfm(sw_ctx_rx->aead_recv);
2832
2833                 tls_update_rx_zc_capable(ctx);
2834                 sw_ctx_rx->async_capable =
2835                         src_crypto_info->version != TLS_1_3_VERSION &&
2836                         !!(tfm->__crt_alg->cra_flags & CRYPTO_ALG_ASYNC);
2837
2838                 rc = tls_strp_init(&sw_ctx_rx->strp, sk);
2839                 if (rc)
2840                         goto free_aead;
2841         }
2842
2843         memcpy(cctx->iv, salt, cipher_desc->salt);
2844         memcpy(cctx->iv + cipher_desc->salt, iv, cipher_desc->iv);
2845         memcpy(cctx->rec_seq, rec_seq, cipher_desc->rec_seq);
2846
2847         if (new_crypto_info) {
2848                 unsafe_memcpy(crypto_info, new_crypto_info,
2849                               cipher_desc->crypto_info,
2850                               /* size was checked in do_tls_setsockopt_conf */);
2851                 memzero_explicit(new_crypto_info, cipher_desc->crypto_info);
2852                 if (!tx)
2853                         tls_finish_key_update(sk, ctx);
2854         }
2855
2856         goto out;
2857
2858 free_aead:
2859         crypto_free_aead(*aead);
2860         *aead = NULL;
2861 free_priv:
2862         if (!new_crypto_info) {
2863                 if (tx) {
2864                         kfree(ctx->priv_ctx_tx);
2865                         ctx->priv_ctx_tx = NULL;
2866                 } else {
2867                         kfree(ctx->priv_ctx_rx);
2868                         ctx->priv_ctx_rx = NULL;
2869                 }
2870         }
2871 out:
2872         return rc;
2873 }