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