net/tls/tls_sw.c

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