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, &msg_pl->sg.data[i]);
 714
 715         i = msg_en->sg.end;
 716         sk_msg_iter_var_prev(i);
 717         sg_mark_end(sk_msg_elem(msg_en, i));
 718
 719         i = msg_en->sg.start;
 720         sg_chain(rec->sg_aead_out, 2, &msg_en->sg.data[i]);
 721
 722         tls_make_aad(rec->aad_space, msg_pl->sg.size + prot->tail_size,
 723                      tls_ctx->tx.rec_seq, prot->rec_seq_size,
 724                      record_type, prot->version);
 725
 726         tls_fill_prepend(tls_ctx,
 727                          page_address(sg_page(&msg_en->sg.data[i])) +
 728                          msg_en->sg.data[i].offset,
 729                          msg_pl->sg.size + prot->tail_size,
 730                          record_type, prot->version);
 731
 732         tls_ctx->pending_open_record_frags = false;
 733
 734         rc = tls_do_encryption(sk, tls_ctx, ctx, req,
 735                                msg_pl->sg.size + prot->tail_size, i);
 736         if (rc < 0) {
 737                 if (rc != -EINPROGRESS) {
 738                         tls_err_abort(sk, EBADMSG);
 739                         if (split) {
 740                                 tls_ctx->pending_open_record_frags = true;
 741                                 tls_merge_open_record(sk, rec, tmp, orig_end);
 742                         }
 743                 }
 744                 ctx->async_capable = 1;
 745                 return rc;
 746         } else if (split) {
 747                 msg_pl = &tmp->msg_plaintext;
 748                 msg_en = &tmp->msg_encrypted;
 749                 sk_msg_trim(sk, msg_en, msg_pl->sg.size + prot->overhead_size);
 750                 tls_ctx->pending_open_record_frags = true;
 751                 ctx->open_rec = tmp;
 752         }
 753
 754         return tls_tx_records(sk, flags);
 755 }
 756
 757 static int bpf_exec_tx_verdict(struct sk_msg *msg, struct sock *sk,
 758                                bool full_record, u8 record_type,
 759                                size_t *copied, int flags)
 760 {
 761         struct tls_context *tls_ctx = tls_get_ctx(sk);
 762         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 763         struct sk_msg msg_redir = { };
 764         struct sk_psock *psock;
 765         struct sock *sk_redir;
 766         struct tls_rec *rec;
 767         bool enospc, policy;
 768         int err = 0, send;
 769         u32 delta = 0;
 770
 771         policy = !(flags & MSG_SENDPAGE_NOPOLICY);
 772         psock = sk_psock_get(sk);
 773         if (!psock || !policy) {
 774                 err = tls_push_record(sk, flags, record_type);
 775                 if (err) {
 776                         *copied -= sk_msg_free(sk, msg);
 777                         tls_free_open_rec(sk);
 778                 }
 779                 return err;
 780         }
 781 more_data:
 782         enospc = sk_msg_full(msg);
 783         if (psock->eval == __SK_NONE) {
 784                 delta = msg->sg.size;
 785                 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
 786                 if (delta < msg->sg.size)
 787                         delta -= msg->sg.size;
 788                 else
 789                         delta = 0;
 790         }
 791         if (msg->cork_bytes && msg->cork_bytes > msg->sg.size &&
 792             !enospc && !full_record) {
 793                 err = -ENOSPC;
 794                 goto out_err;
 795         }
 796         msg->cork_bytes = 0;
 797         send = msg->sg.size;
 798         if (msg->apply_bytes && msg->apply_bytes < send)
 799                 send = msg->apply_bytes;
 800
 801         switch (psock->eval) {
 802         case __SK_PASS:
 803                 err = tls_push_record(sk, flags, record_type);
 804                 if (err < 0) {
 805                         *copied -= sk_msg_free(sk, msg);
 806                         tls_free_open_rec(sk);
 807                         goto out_err;
 808                 }
 809                 break;
 810         case __SK_REDIRECT:
 811                 sk_redir = psock->sk_redir;
 812                 memcpy(&msg_redir, msg, sizeof(*msg));
 813                 if (msg->apply_bytes < send)
 814                         msg->apply_bytes = 0;
 815                 else
 816                         msg->apply_bytes -= send;
 817                 sk_msg_return_zero(sk, msg, send);
 818                 msg->sg.size -= send;
 819                 release_sock(sk);
 820                 err = tcp_bpf_sendmsg_redir(sk_redir, &msg_redir, send, flags);
 821                 lock_sock(sk);
 822                 if (err < 0) {
 823                         *copied -= sk_msg_free_nocharge(sk, &msg_redir);
 824                         msg->sg.size = 0;
 825                 }
 826                 if (msg->sg.size == 0)
 827                         tls_free_open_rec(sk);
 828                 break;
 829         case __SK_DROP:
 830         default:
 831                 sk_msg_free_partial(sk, msg, send);
 832                 if (msg->apply_bytes < send)
 833                         msg->apply_bytes = 0;
 834                 else
 835                         msg->apply_bytes -= send;
 836                 if (msg->sg.size == 0)
 837                         tls_free_open_rec(sk);
 838                 *copied -= (send + delta);
 839                 err = -EACCES;
 840         }
 841
 842         if (likely(!err)) {
 843                 bool reset_eval = !ctx->open_rec;
 844
 845                 rec = ctx->open_rec;
 846                 if (rec) {
 847                         msg = &rec->msg_plaintext;
 848                         if (!msg->apply_bytes)
 849                                 reset_eval = true;
 850                 }
 851                 if (reset_eval) {
 852                         psock->eval = __SK_NONE;
 853                         if (psock->sk_redir) {
 854                                 sock_put(psock->sk_redir);
 855                                 psock->sk_redir = NULL;
 856                         }
 857                 }
 858                 if (rec)
 859                         goto more_data;
 860         }
 861  out_err:
 862         sk_psock_put(sk, psock);
 863         return err;
 864 }
 865
 866 static int tls_sw_push_pending_record(struct sock *sk, int flags)
 867 {
 868         struct tls_context *tls_ctx = tls_get_ctx(sk);
 869         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 870         struct tls_rec *rec = ctx->open_rec;
 871         struct sk_msg *msg_pl;
 872         size_t copied;
 873
 874         if (!rec)
 875                 return 0;
 876
 877         msg_pl = &rec->msg_plaintext;
 878         copied = msg_pl->sg.size;
 879         if (!copied)
 880                 return 0;
 881
 882         return bpf_exec_tx_verdict(msg_pl, sk, true, TLS_RECORD_TYPE_DATA,
 883                                    &copied, flags);
 884 }
 885
 886 int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 887 {
 888         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 889         struct tls_context *tls_ctx = tls_get_ctx(sk);
 890         struct tls_prot_info *prot = &tls_ctx->prot_info;
 891         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 892         bool async_capable = ctx->async_capable;
 893         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 894         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
 895         bool eor = !(msg->msg_flags & MSG_MORE);
 896         size_t try_to_copy, copied = 0;
 897         struct sk_msg *msg_pl, *msg_en;
 898         struct tls_rec *rec;
 899         int required_size;
 900         int num_async = 0;
 901         bool full_record;
 902         int record_room;
 903         int num_zc = 0;
 904         int orig_size;
 905         int ret = 0;
 906
 907         if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
 908                 return -ENOTSUPP;
 909
 910         mutex_lock(&tls_ctx->tx_lock);
 911         lock_sock(sk);
 912
 913         if (unlikely(msg->msg_controllen)) {
 914                 ret = tls_proccess_cmsg(sk, msg, &record_type);
 915                 if (ret) {
 916                         if (ret == -EINPROGRESS)
 917                                 num_async++;
 918                         else if (ret != -EAGAIN)
 919                                 goto send_end;
 920                 }
 921         }
 922
 923         while (msg_data_left(msg)) {
 924                 if (sk->sk_err) {
 925                         ret = -sk->sk_err;
 926                         goto send_end;
 927                 }
 928
 929                 if (ctx->open_rec)
 930                         rec = ctx->open_rec;
 931                 else
 932                         rec = ctx->open_rec = tls_get_rec(sk);
 933                 if (!rec) {
 934                         ret = -ENOMEM;
 935                         goto send_end;
 936                 }
 937
 938                 msg_pl = &rec->msg_plaintext;
 939                 msg_en = &rec->msg_encrypted;
 940
 941                 orig_size = msg_pl->sg.size;
 942                 full_record = false;
 943                 try_to_copy = msg_data_left(msg);
 944                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
 945                 if (try_to_copy >= record_room) {
 946                         try_to_copy = record_room;
 947                         full_record = true;
 948                 }
 949
 950                 required_size = msg_pl->sg.size + try_to_copy +
 951                                 prot->overhead_size;
 952
 953                 if (!sk_stream_memory_free(sk))
 954                         goto wait_for_sndbuf;
 955
 956 alloc_encrypted:
 957                 ret = tls_alloc_encrypted_msg(sk, required_size);
 958                 if (ret) {
 959                         if (ret != -ENOSPC)
 960                                 goto wait_for_memory;
 961
 962                         /* Adjust try_to_copy according to the amount that was
 963                          * actually allocated. The difference is due
 964                          * to max sg elements limit
 965                          */
 966                         try_to_copy -= required_size - msg_en->sg.size;
 967                         full_record = true;
 968                 }
 969
 970                 if (!is_kvec && (full_record || eor) && !async_capable) {
 971                         u32 first = msg_pl->sg.end;
 972
 973                         ret = sk_msg_zerocopy_from_iter(sk, &msg->msg_iter,
 974                                                         msg_pl, try_to_copy);
 975                         if (ret)
 976                                 goto fallback_to_reg_send;
 977
 978                         num_zc++;
 979                         copied += try_to_copy;
 980
 981                         sk_msg_sg_copy_set(msg_pl, first);
 982                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
 983                                                   record_type, &copied,
 984                                                   msg->msg_flags);
 985                         if (ret) {
 986                                 if (ret == -EINPROGRESS)
 987                                         num_async++;
 988                                 else if (ret == -ENOMEM)
 989                                         goto wait_for_memory;
 990                                 else if (ctx->open_rec && ret == -ENOSPC)
 991                                         goto rollback_iter;
 992                                 else if (ret != -EAGAIN)
 993                                         goto send_end;
 994                         }
 995                         continue;
 996 rollback_iter:
 997                         copied -= try_to_copy;
 998                         sk_msg_sg_copy_clear(msg_pl, first);
 999                         iov_iter_revert(&msg->msg_iter,
1000                                         msg_pl->sg.size - orig_size);
1001 fallback_to_reg_send:
1002                         sk_msg_trim(sk, msg_pl, orig_size);
1003                 }
1004
1005                 required_size = msg_pl->sg.size + try_to_copy;
1006
1007                 ret = tls_clone_plaintext_msg(sk, required_size);
1008                 if (ret) {
1009                         if (ret != -ENOSPC)
1010                                 goto send_end;
1011
1012                         /* Adjust try_to_copy according to the amount that was
1013                          * actually allocated. The difference is due
1014                          * to max sg elements limit
1015                          */
1016                         try_to_copy -= required_size - msg_pl->sg.size;
1017                         full_record = true;
1018                         sk_msg_trim(sk, msg_en,
1019                                     msg_pl->sg.size + prot->overhead_size);
1020                 }
1021
1022                 if (try_to_copy) {
1023                         ret = sk_msg_memcopy_from_iter(sk, &msg->msg_iter,
1024                                                        msg_pl, try_to_copy);
1025                         if (ret < 0)
1026                                 goto trim_sgl;
1027                 }
1028
1029                 /* Open records defined only if successfully copied, otherwise
1030                  * we would trim the sg but not reset the open record frags.
1031                  */
1032                 tls_ctx->pending_open_record_frags = true;
1033                 copied += try_to_copy;
1034                 if (full_record || eor) {
1035                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1036                                                   record_type, &copied,
1037                                                   msg->msg_flags);
1038                         if (ret) {
1039                                 if (ret == -EINPROGRESS)
1040                                         num_async++;
1041                                 else if (ret == -ENOMEM)
1042                                         goto wait_for_memory;
1043                                 else if (ret != -EAGAIN) {
1044                                         if (ret == -ENOSPC)
1045                                                 ret = 0;
1046                                         goto send_end;
1047                                 }
1048                         }
1049                 }
1050
1051                 continue;
1052
1053 wait_for_sndbuf:
1054                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1055 wait_for_memory:
1056                 ret = sk_stream_wait_memory(sk, &timeo);
1057                 if (ret) {
1058 trim_sgl:
1059                         if (ctx->open_rec)
1060                                 tls_trim_both_msgs(sk, orig_size);
1061                         goto send_end;
1062                 }
1063
1064                 if (ctx->open_rec && msg_en->sg.size < required_size)
1065                         goto alloc_encrypted;
1066         }
1067
1068         if (!num_async) {
1069                 goto send_end;
1070         } else if (num_zc) {
1071                 /* Wait for pending encryptions to get completed */
1072                 smp_store_mb(ctx->async_notify, true);
1073
1074                 if (atomic_read(&ctx->encrypt_pending))
1075                         crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1076                 else
1077                         reinit_completion(&ctx->async_wait.completion);
1078
1079                 WRITE_ONCE(ctx->async_notify, false);
1080
1081                 if (ctx->async_wait.err) {
1082                         ret = ctx->async_wait.err;
1083                         copied = 0;
1084                 }
1085         }
1086
1087         /* Transmit if any encryptions have completed */
1088         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1089                 cancel_delayed_work(&ctx->tx_work.work);
1090                 tls_tx_records(sk, msg->msg_flags);
1091         }
1092
1093 send_end:
1094         ret = sk_stream_error(sk, msg->msg_flags, ret);
1095
1096         release_sock(sk);
1097         mutex_unlock(&tls_ctx->tx_lock);
1098         return copied ? copied : ret;
1099 }
1100
1101 static int tls_sw_do_sendpage(struct sock *sk, struct page *page,
1102                               int offset, size_t size, int flags)
1103 {
1104         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1105         struct tls_context *tls_ctx = tls_get_ctx(sk);
1106         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1107         struct tls_prot_info *prot = &tls_ctx->prot_info;
1108         unsigned char record_type = TLS_RECORD_TYPE_DATA;
1109         struct sk_msg *msg_pl;
1110         struct tls_rec *rec;
1111         int num_async = 0;
1112         size_t copied = 0;
1113         bool full_record;
1114         int record_room;
1115         int ret = 0;
1116         bool eor;
1117
1118         eor = !(flags & (MSG_MORE | MSG_SENDPAGE_NOTLAST));
1119         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1120
1121         /* Call the sk_stream functions to manage the sndbuf mem. */
1122         while (size > 0) {
1123                 size_t copy, required_size;
1124
1125                 if (sk->sk_err) {
1126                         ret = -sk->sk_err;
1127                         goto sendpage_end;
1128                 }
1129
1130                 if (ctx->open_rec)
1131                         rec = ctx->open_rec;
1132                 else
1133                         rec = ctx->open_rec = tls_get_rec(sk);
1134                 if (!rec) {
1135                         ret = -ENOMEM;
1136                         goto sendpage_end;
1137                 }
1138
1139                 msg_pl = &rec->msg_plaintext;
1140
1141                 full_record = false;
1142                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
1143                 copy = size;
1144                 if (copy >= record_room) {
1145                         copy = record_room;
1146                         full_record = true;
1147                 }
1148
1149                 required_size = msg_pl->sg.size + copy + prot->overhead_size;
1150
1151                 if (!sk_stream_memory_free(sk))
1152                         goto wait_for_sndbuf;
1153 alloc_payload:
1154                 ret = tls_alloc_encrypted_msg(sk, required_size);
1155                 if (ret) {
1156                         if (ret != -ENOSPC)
1157                                 goto wait_for_memory;
1158
1159                         /* Adjust copy according to the amount that was
1160                          * actually allocated. The difference is due
1161                          * to max sg elements limit
1162                          */
1163                         copy -= required_size - msg_pl->sg.size;
1164                         full_record = true;
1165                 }
1166
1167                 sk_msg_page_add(msg_pl, page, copy, offset);
1168                 sk_mem_charge(sk, copy);
1169
1170                 offset += copy;
1171                 size -= copy;
1172                 copied += copy;
1173
1174                 tls_ctx->pending_open_record_frags = true;
1175                 if (full_record || eor || sk_msg_full(msg_pl)) {
1176                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1177                                                   record_type, &copied, flags);
1178                         if (ret) {
1179                                 if (ret == -EINPROGRESS)
1180                                         num_async++;
1181                                 else if (ret == -ENOMEM)
1182                                         goto wait_for_memory;
1183                                 else if (ret != -EAGAIN) {
1184                                         if (ret == -ENOSPC)
1185                                                 ret = 0;
1186                                         goto sendpage_end;
1187                                 }
1188                         }
1189                 }
1190                 continue;
1191 wait_for_sndbuf:
1192                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1193 wait_for_memory:
1194                 ret = sk_stream_wait_memory(sk, &timeo);
1195                 if (ret) {
1196                         if (ctx->open_rec)
1197                                 tls_trim_both_msgs(sk, msg_pl->sg.size);
1198                         goto sendpage_end;
1199                 }
1200
1201                 if (ctx->open_rec)
1202                         goto alloc_payload;
1203         }
1204
1205         if (num_async) {
1206                 /* Transmit if any encryptions have completed */
1207                 if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1208                         cancel_delayed_work(&ctx->tx_work.work);
1209                         tls_tx_records(sk, flags);
1210                 }
1211         }
1212 sendpage_end:
1213         ret = sk_stream_error(sk, flags, ret);
1214         return copied ? copied : ret;
1215 }
1216
1217 int tls_sw_sendpage_locked(struct sock *sk, struct page *page,
1218                            int offset, size_t size, int flags)
1219 {
1220         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1221                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY |
1222                       MSG_NO_SHARED_FRAGS))
1223                 return -ENOTSUPP;
1224
1225         return tls_sw_do_sendpage(sk, page, offset, size, flags);
1226 }
1227
1228 int tls_sw_sendpage(struct sock *sk, struct page *page,
1229                     int offset, size_t size, int flags)
1230 {
1231         struct tls_context *tls_ctx = tls_get_ctx(sk);
1232         int ret;
1233
1234         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1235                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY))
1236                 return -ENOTSUPP;
1237
1238         mutex_lock(&tls_ctx->tx_lock);
1239         lock_sock(sk);
1240         ret = tls_sw_do_sendpage(sk, page, offset, size, flags);
1241         release_sock(sk);
1242         mutex_unlock(&tls_ctx->tx_lock);
1243         return ret;
1244 }
1245
1246 static struct sk_buff *tls_wait_data(struct sock *sk, struct sk_psock *psock,
1247                                      int flags, long timeo, int *err)
1248 {
1249         struct tls_context *tls_ctx = tls_get_ctx(sk);
1250         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1251         struct sk_buff *skb;
1252         DEFINE_WAIT_FUNC(wait, woken_wake_function);
1253
1254         while (!(skb = ctx->recv_pkt) && sk_psock_queue_empty(psock)) {
1255                 if (sk->sk_err) {
1256                         *err = sock_error(sk);
1257                         return NULL;
1258                 }
1259
1260                 if (sk->sk_shutdown & RCV_SHUTDOWN)
1261                         return NULL;
1262
1263                 if (sock_flag(sk, SOCK_DONE))
1264                         return NULL;
1265
1266                 if ((flags & MSG_DONTWAIT) || !timeo) {
1267                         *err = -EAGAIN;
1268                         return NULL;
1269                 }
1270
1271                 add_wait_queue(sk_sleep(sk), &wait);
1272                 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1273                 sk_wait_event(sk, &timeo,
1274                               ctx->recv_pkt != skb ||
1275                               !sk_psock_queue_empty(psock),
1276                               &wait);
1277                 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1278                 remove_wait_queue(sk_sleep(sk), &wait);
1279
1280                 /* Handle signals */
1281                 if (signal_pending(current)) {
1282                         *err = sock_intr_errno(timeo);
1283                         return NULL;
1284                 }
1285         }
1286
1287         return skb;
1288 }
1289
1290 static int tls_setup_from_iter(struct sock *sk, struct iov_iter *from,
1291                                int length, int *pages_used,
1292                                unsigned int *size_used,
1293                                struct scatterlist *to,
1294                                int to_max_pages)
1295 {
1296         int rc = 0, i = 0, num_elem = *pages_used, maxpages;
1297         struct page *pages[MAX_SKB_FRAGS];
1298         unsigned int size = *size_used;
1299         ssize_t copied, use;
1300         size_t offset;
1301
1302         while (length > 0) {
1303                 i = 0;
1304                 maxpages = to_max_pages - num_elem;
1305                 if (maxpages == 0) {
1306                         rc = -EFAULT;
1307                         goto out;
1308                 }
1309                 copied = iov_iter_get_pages(from, pages,
1310                                             length,
1311                                             maxpages, &offset);
1312                 if (copied <= 0) {
1313                         rc = -EFAULT;
1314                         goto out;
1315                 }
1316
1317                 iov_iter_advance(from, copied);
1318
1319                 length -= copied;
1320                 size += copied;
1321                 while (copied) {
1322                         use = min_t(int, copied, PAGE_SIZE - offset);
1323
1324                         sg_set_page(&to[num_elem],
1325                                     pages[i], use, offset);
1326                         sg_unmark_end(&to[num_elem]);
1327                         /* We do not uncharge memory from this API */
1328
1329                         offset = 0;
1330                         copied -= use;
1331
1332                         i++;
1333                         num_elem++;
1334                 }
1335         }
1336         /* Mark the end in the last sg entry if newly added */
1337         if (num_elem > *pages_used)
1338                 sg_mark_end(&to[num_elem - 1]);
1339 out:
1340         if (rc)
1341                 iov_iter_revert(from, size - *size_used);
1342         *size_used = size;
1343         *pages_used = num_elem;
1344
1345         return rc;
1346 }
1347
1348 /* This function decrypts the input skb into either out_iov or in out_sg
1349  * or in skb buffers itself. The input parameter 'zc' indicates if
1350  * zero-copy mode needs to be tried or not. With zero-copy mode, either
1351  * out_iov or out_sg must be non-NULL. In case both out_iov and out_sg are
1352  * NULL, then the decryption happens inside skb buffers itself, i.e.
1353  * zero-copy gets disabled and 'zc' is updated.
1354  */
1355
1356 static int decrypt_internal(struct sock *sk, struct sk_buff *skb,
1357                             struct iov_iter *out_iov,
1358                             struct scatterlist *out_sg,
1359                             int *chunk, bool *zc, bool async)
1360 {
1361         struct tls_context *tls_ctx = tls_get_ctx(sk);
1362         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1363         struct tls_prot_info *prot = &tls_ctx->prot_info;
1364         struct strp_msg *rxm = strp_msg(skb);
1365         int n_sgin, n_sgout, nsg, mem_size, aead_size, err, pages = 0;
1366         struct aead_request *aead_req;
1367         struct sk_buff *unused;
1368         u8 *aad, *iv, *mem = NULL;
1369         struct scatterlist *sgin = NULL;
1370         struct scatterlist *sgout = NULL;
1371         const int data_len = rxm->full_len - prot->overhead_size +
1372                              prot->tail_size;
1373         int iv_offset = 0;
1374
1375         if (*zc && (out_iov || out_sg)) {
1376                 if (out_iov)
1377                         n_sgout = iov_iter_npages(out_iov, INT_MAX) + 1;
1378                 else
1379                         n_sgout = sg_nents(out_sg);
1380                 n_sgin = skb_nsg(skb, rxm->offset + prot->prepend_size,
1381                                  rxm->full_len - prot->prepend_size);
1382         } else {
1383                 n_sgout = 0;
1384                 *zc = false;
1385                 n_sgin = skb_cow_data(skb, 0, &unused);
1386         }
1387
1388         if (n_sgin < 1)
1389                 return -EBADMSG;
1390
1391         /* Increment to accommodate AAD */
1392         n_sgin = n_sgin + 1;
1393
1394         nsg = n_sgin + n_sgout;
1395
1396         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(ctx->aead_recv);
1397         mem_size = aead_size + (nsg * sizeof(struct scatterlist));
1398         mem_size = mem_size + prot->aad_size;
1399         mem_size = mem_size + crypto_aead_ivsize(ctx->aead_recv);
1400
1401         /* Allocate a single block of memory which contains
1402          * aead_req || sgin[] || sgout[] || aad || iv.
1403          * This order achieves correct alignment for aead_req, sgin, sgout.
1404          */
1405         mem = kmalloc(mem_size, sk->sk_allocation);
1406         if (!mem)
1407                 return -ENOMEM;
1408
1409         /* Segment the allocated memory */
1410         aead_req = (struct aead_request *)mem;
1411         sgin = (struct scatterlist *)(mem + aead_size);
1412         sgout = sgin + n_sgin;
1413         aad = (u8 *)(sgout + n_sgout);
1414         iv = aad + prot->aad_size;
1415
1416         /* For CCM based ciphers, first byte of nonce+iv is always '2' */
1417         if (prot->cipher_type == TLS_CIPHER_AES_CCM_128) {
1418                 iv[0] = 2;
1419                 iv_offset = 1;
1420         }
1421
1422         /* Prepare IV */
1423         err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
1424                             iv + iv_offset + prot->salt_size,
1425                             prot->iv_size);
1426         if (err < 0) {
1427                 kfree(mem);
1428                 return err;
1429         }
1430         if (prot->version == TLS_1_3_VERSION)
1431                 memcpy(iv + iv_offset, tls_ctx->rx.iv,
1432                        crypto_aead_ivsize(ctx->aead_recv));
1433         else
1434                 memcpy(iv + iv_offset, tls_ctx->rx.iv, prot->salt_size);
1435
1436         xor_iv_with_seq(prot->version, iv, tls_ctx->rx.rec_seq);
1437
1438         /* Prepare AAD */
1439         tls_make_aad(aad, rxm->full_len - prot->overhead_size +
1440                      prot->tail_size,
1441                      tls_ctx->rx.rec_seq, prot->rec_seq_size,
1442                      ctx->control, prot->version);
1443
1444         /* Prepare sgin */
1445         sg_init_table(sgin, n_sgin);
1446         sg_set_buf(&sgin[0], aad, prot->aad_size);
1447         err = skb_to_sgvec(skb, &sgin[1],
1448                            rxm->offset + prot->prepend_size,
1449                            rxm->full_len - prot->prepend_size);
1450         if (err < 0) {
1451                 kfree(mem);
1452                 return err;
1453         }
1454
1455         if (n_sgout) {
1456                 if (out_iov) {
1457                         sg_init_table(sgout, n_sgout);
1458                         sg_set_buf(&sgout[0], aad, prot->aad_size);
1459
1460                         *chunk = 0;
1461                         err = tls_setup_from_iter(sk, out_iov, data_len,
1462                                                   &pages, chunk, &sgout[1],
1463                                                   (n_sgout - 1));
1464                         if (err < 0)
1465                                 goto fallback_to_reg_recv;
1466                 } else if (out_sg) {
1467                         memcpy(sgout, out_sg, n_sgout * sizeof(*sgout));
1468                 } else {
1469                         goto fallback_to_reg_recv;
1470                 }
1471         } else {
1472 fallback_to_reg_recv:
1473                 sgout = sgin;
1474                 pages = 0;
1475                 *chunk = data_len;
1476                 *zc = false;
1477         }
1478
1479         /* Prepare and submit AEAD request */
1480         err = tls_do_decryption(sk, skb, sgin, sgout, iv,
1481                                 data_len, aead_req, async);
1482         if (err == -EINPROGRESS)
1483                 return err;
1484
1485         /* Release the pages in case iov was mapped to pages */
1486         for (; pages > 0; pages--)
1487                 put_page(sg_page(&sgout[pages]));
1488
1489         kfree(mem);
1490         return err;
1491 }
1492
1493 static int decrypt_skb_update(struct sock *sk, struct sk_buff *skb,
1494                               struct iov_iter *dest, int *chunk, bool *zc,
1495                               bool async)
1496 {
1497         struct tls_context *tls_ctx = tls_get_ctx(sk);
1498         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1499         struct tls_prot_info *prot = &tls_ctx->prot_info;
1500         struct strp_msg *rxm = strp_msg(skb);
1501         int pad, err = 0;
1502
1503         if (!ctx->decrypted) {
1504                 if (tls_ctx->rx_conf == TLS_HW) {
1505                         err = tls_device_decrypted(sk, tls_ctx, skb, rxm);
1506                         if (err < 0)
1507                                 return err;
1508                 }
1509
1510                 /* Still not decrypted after tls_device */
1511                 if (!ctx->decrypted) {
1512                         err = decrypt_internal(sk, skb, dest, NULL, chunk, zc,
1513                                                async);
1514                         if (err < 0) {
1515                                 if (err == -EINPROGRESS)
1516                                         tls_advance_record_sn(sk, prot,
1517                                                               &tls_ctx->rx);
1518
1519                                 return err;
1520                         }
1521                 } else {
1522                         *zc = false;
1523                 }
1524
1525                 pad = padding_length(ctx, prot, skb);
1526                 if (pad < 0)
1527                         return pad;
1528
1529                 rxm->full_len -= pad;
1530                 rxm->offset += prot->prepend_size;
1531                 rxm->full_len -= prot->overhead_size;
1532                 tls_advance_record_sn(sk, prot, &tls_ctx->rx);
1533                 ctx->decrypted = 1;
1534                 ctx->saved_data_ready(sk);
1535         } else {
1536                 *zc = false;
1537         }
1538
1539         return err;
1540 }
1541
1542 int decrypt_skb(struct sock *sk, struct sk_buff *skb,
1543                 struct scatterlist *sgout)
1544 {
1545         bool zc = true;
1546         int chunk;
1547
1548         return decrypt_internal(sk, skb, NULL, sgout, &chunk, &zc, false);
1549 }
1550
1551 static bool tls_sw_advance_skb(struct sock *sk, struct sk_buff *skb,
1552                                unsigned int len)
1553 {
1554         struct tls_context *tls_ctx = tls_get_ctx(sk);
1555         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1556
1557         if (skb) {
1558                 struct strp_msg *rxm = strp_msg(skb);
1559
1560                 if (len < rxm->full_len) {
1561                         rxm->offset += len;
1562                         rxm->full_len -= len;
1563                         return false;
1564                 }
1565                 consume_skb(skb);
1566         }
1567
1568         /* Finished with message */
1569         ctx->recv_pkt = NULL;
1570         __strp_unpause(&ctx->strp);
1571
1572         return true;
1573 }
1574
1575 /* This function traverses the rx_list in tls receive context to copies the
1576  * decrypted records into the buffer provided by caller zero copy is not
1577  * true. Further, the records are removed from the rx_list if it is not a peek
1578  * case and the record has been consumed completely.
1579  */
1580 static int process_rx_list(struct tls_sw_context_rx *ctx,
1581                            struct msghdr *msg,
1582                            u8 *control,
1583                            bool *cmsg,
1584                            size_t skip,
1585                            size_t len,
1586                            bool zc,
1587                            bool is_peek)
1588 {
1589         struct sk_buff *skb = skb_peek(&ctx->rx_list);
1590         u8 ctrl = *control;
1591         u8 msgc = *cmsg;
1592         struct tls_msg *tlm;
1593         ssize_t copied = 0;
1594
1595         /* Set the record type in 'control' if caller didn't pass it */
1596         if (!ctrl && skb) {
1597                 tlm = tls_msg(skb);
1598                 ctrl = tlm->control;
1599         }
1600
1601         while (skip && skb) {
1602                 struct strp_msg *rxm = strp_msg(skb);
1603                 tlm = tls_msg(skb);
1604
1605                 /* Cannot process a record of different type */
1606                 if (ctrl != tlm->control)
1607                         return 0;
1608
1609                 if (skip < rxm->full_len)
1610                         break;
1611
1612                 skip = skip - rxm->full_len;
1613                 skb = skb_peek_next(skb, &ctx->rx_list);
1614         }
1615
1616         while (len && skb) {
1617                 struct sk_buff *next_skb;
1618                 struct strp_msg *rxm = strp_msg(skb);
1619                 int chunk = min_t(unsigned int, rxm->full_len - skip, len);
1620
1621                 tlm = tls_msg(skb);
1622
1623                 /* Cannot process a record of different type */
1624                 if (ctrl != tlm->control)
1625                         return 0;
1626
1627                 /* Set record type if not already done. For a non-data record,
1628                  * do not proceed if record type could not be copied.
1629                  */
1630                 if (!msgc) {
1631                         int cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1632                                             sizeof(ctrl), &ctrl);
1633                         msgc = true;
1634                         if (ctrl != TLS_RECORD_TYPE_DATA) {
1635                                 if (cerr || msg->msg_flags & MSG_CTRUNC)
1636                                         return -EIO;
1637
1638                                 *cmsg = msgc;
1639                         }
1640                 }
1641
1642                 if (!zc || (rxm->full_len - skip) > len) {
1643                         int err = skb_copy_datagram_msg(skb, rxm->offset + skip,
1644                                                     msg, chunk);
1645                         if (err < 0)
1646                                 return err;
1647                 }
1648
1649                 len = len - chunk;
1650                 copied = copied + chunk;
1651
1652                 /* Consume the data from record if it is non-peek case*/
1653                 if (!is_peek) {
1654                         rxm->offset = rxm->offset + chunk;
1655                         rxm->full_len = rxm->full_len - chunk;
1656
1657                         /* Return if there is unconsumed data in the record */
1658                         if (rxm->full_len - skip)
1659                                 break;
1660                 }
1661
1662                 /* The remaining skip-bytes must lie in 1st record in rx_list.
1663                  * So from the 2nd record, 'skip' should be 0.
1664                  */
1665                 skip = 0;
1666
1667                 if (msg)
1668                         msg->msg_flags |= MSG_EOR;
1669
1670                 next_skb = skb_peek_next(skb, &ctx->rx_list);
1671
1672                 if (!is_peek) {
1673                         skb_unlink(skb, &ctx->rx_list);
1674                         consume_skb(skb);
1675                 }
1676
1677                 skb = next_skb;
1678         }
1679
1680         *control = ctrl;
1681         return copied;
1682 }
1683
1684 int tls_sw_recvmsg(struct sock *sk,
1685                    struct msghdr *msg,
1686                    size_t len,
1687                    int nonblock,
1688                    int flags,
1689                    int *addr_len)
1690 {
1691         struct tls_context *tls_ctx = tls_get_ctx(sk);
1692         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1693         struct tls_prot_info *prot = &tls_ctx->prot_info;
1694         struct sk_psock *psock;
1695         unsigned char control = 0;
1696         ssize_t decrypted = 0;
1697         struct strp_msg *rxm;
1698         struct tls_msg *tlm;
1699         struct sk_buff *skb;
1700         ssize_t copied = 0;
1701         bool cmsg = false;
1702         int target, err = 0;
1703         long timeo;
1704         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
1705         bool is_peek = flags & MSG_PEEK;
1706         int num_async = 0;
1707
1708         flags |= nonblock;
1709
1710         if (unlikely(flags & MSG_ERRQUEUE))
1711                 return sock_recv_errqueue(sk, msg, len, SOL_IP, IP_RECVERR);
1712
1713         psock = sk_psock_get(sk);
1714         lock_sock(sk);
1715
1716         /* Process pending decrypted records. It must be non-zero-copy */
1717         err = process_rx_list(ctx, msg, &control, &cmsg, 0, len, false,
1718                               is_peek);
1719         if (err < 0) {
1720                 tls_err_abort(sk, err);
1721                 goto end;
1722         } else {
1723                 copied = err;
1724         }
1725
1726         if (len <= copied)
1727                 goto recv_end;
1728
1729         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1730         len = len - copied;
1731         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1732
1733         while (len && (decrypted + copied < target || ctx->recv_pkt)) {
1734                 bool retain_skb = false;
1735                 bool zc = false;
1736                 int to_decrypt;
1737                 int chunk = 0;
1738                 bool async_capable;
1739                 bool async = false;
1740
1741                 skb = tls_wait_data(sk, psock, flags, timeo, &err);
1742                 if (!skb) {
1743                         if (psock) {
1744                                 int ret = __tcp_bpf_recvmsg(sk, psock,
1745                                                             msg, len, flags);
1746
1747                                 if (ret > 0) {
1748                                         decrypted += ret;
1749                                         len -= ret;
1750                                         continue;
1751                                 }
1752                         }
1753                         goto recv_end;
1754                 } else {
1755                         tlm = tls_msg(skb);
1756                         if (prot->version == TLS_1_3_VERSION)
1757                                 tlm->control = 0;
1758                         else
1759                                 tlm->control = ctx->control;
1760                 }
1761
1762                 rxm = strp_msg(skb);
1763
1764                 to_decrypt = rxm->full_len - prot->overhead_size;
1765
1766                 if (to_decrypt <= len && !is_kvec && !is_peek &&
1767                     ctx->control == TLS_RECORD_TYPE_DATA &&
1768                     prot->version != TLS_1_3_VERSION)
1769                         zc = true;
1770
1771                 /* Do not use async mode if record is non-data */
1772                 if (ctx->control == TLS_RECORD_TYPE_DATA)
1773                         async_capable = ctx->async_capable;
1774                 else
1775                         async_capable = false;
1776
1777                 err = decrypt_skb_update(sk, skb, &msg->msg_iter,
1778                                          &chunk, &zc, async_capable);
1779                 if (err < 0 && err != -EINPROGRESS) {
1780                         tls_err_abort(sk, EBADMSG);
1781                         goto recv_end;
1782                 }
1783
1784                 if (err == -EINPROGRESS) {
1785                         async = true;
1786                         num_async++;
1787                 } else if (prot->version == TLS_1_3_VERSION) {
1788                         tlm->control = ctx->control;
1789                 }
1790
1791                 /* If the type of records being processed is not known yet,
1792                  * set it to record type just dequeued. If it is already known,
1793                  * but does not match the record type just dequeued, go to end.
1794                  * We always get record type here since for tls1.2, record type
1795                  * is known just after record is dequeued from stream parser.
1796                  * For tls1.3, we disable async.
1797                  */
1798
1799                 if (!control)
1800                         control = tlm->control;
1801                 else if (control != tlm->control)
1802                         goto recv_end;
1803
1804                 if (!cmsg) {
1805                         int cerr;
1806
1807                         cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1808                                         sizeof(control), &control);
1809                         cmsg = true;
1810                         if (control != TLS_RECORD_TYPE_DATA) {
1811                                 if (cerr || msg->msg_flags & MSG_CTRUNC) {
1812                                         err = -EIO;
1813                                         goto recv_end;
1814                                 }
1815                         }
1816                 }
1817
1818                 if (async)
1819                         goto pick_next_record;
1820
1821                 if (!zc) {
1822                         if (rxm->full_len > len) {
1823                                 retain_skb = true;
1824                                 chunk = len;
1825                         } else {
1826                                 chunk = rxm->full_len;
1827                         }
1828
1829                         err = skb_copy_datagram_msg(skb, rxm->offset,
1830                                                     msg, chunk);
1831                         if (err < 0)
1832                                 goto recv_end;
1833
1834                         if (!is_peek) {
1835                                 rxm->offset = rxm->offset + chunk;
1836                                 rxm->full_len = rxm->full_len - chunk;
1837                         }
1838                 }
1839
1840 pick_next_record:
1841                 if (chunk > len)
1842                         chunk = len;
1843
1844                 decrypted += chunk;
1845                 len -= chunk;
1846
1847                 /* For async or peek case, queue the current skb */
1848                 if (async || is_peek || retain_skb) {
1849                         skb_queue_tail(&ctx->rx_list, skb);
1850                         skb = NULL;
1851                 }
1852
1853                 if (tls_sw_advance_skb(sk, skb, chunk)) {
1854                         /* Return full control message to
1855                          * userspace before trying to parse
1856                          * another message type
1857                          */
1858                         msg->msg_flags |= MSG_EOR;
1859                         if (ctx->control != TLS_RECORD_TYPE_DATA)
1860                                 goto recv_end;
1861                 } else {
1862                         break;
1863                 }
1864         }
1865
1866 recv_end:
1867         if (num_async) {
1868                 /* Wait for all previously submitted records to be decrypted */
1869                 smp_store_mb(ctx->async_notify, true);
1870                 if (atomic_read(&ctx->decrypt_pending)) {
1871                         err = crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1872                         if (err) {
1873                                 /* one of async decrypt failed */
1874                                 tls_err_abort(sk, err);
1875                                 copied = 0;
1876                                 decrypted = 0;
1877                                 goto end;
1878                         }
1879                 } else {
1880                         reinit_completion(&ctx->async_wait.completion);
1881                 }
1882                 WRITE_ONCE(ctx->async_notify, false);
1883
1884                 /* Drain records from the rx_list & copy if required */
1885                 if (is_peek || is_kvec)
1886                         err = process_rx_list(ctx, msg, &control, &cmsg, copied,
1887                                               decrypted, false, is_peek);
1888                 else
1889                         err = process_rx_list(ctx, msg, &control, &cmsg, 0,
1890                                               decrypted, true, is_peek);
1891                 if (err < 0) {
1892                         tls_err_abort(sk, err);
1893                         copied = 0;
1894                         goto end;
1895                 }
1896         }
1897
1898         copied += decrypted;
1899
1900 end:
1901         release_sock(sk);
1902         if (psock)
1903                 sk_psock_put(sk, psock);
1904         return copied ? : err;
1905 }
1906
1907 ssize_t tls_sw_splice_read(struct socket *sock,  loff_t *ppos,
1908                            struct pipe_inode_info *pipe,
1909                            size_t len, unsigned int flags)
1910 {
1911         struct tls_context *tls_ctx = tls_get_ctx(sock->sk);
1912         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1913         struct strp_msg *rxm = NULL;
1914         struct sock *sk = sock->sk;
1915         struct sk_buff *skb;
1916         ssize_t copied = 0;
1917         int err = 0;
1918         long timeo;
1919         int chunk;
1920         bool zc = false;
1921
1922         lock_sock(sk);
1923
1924         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1925
1926         skb = tls_wait_data(sk, NULL, flags, timeo, &err);
1927         if (!skb)
1928                 goto splice_read_end;
1929
1930         if (!ctx->decrypted) {
1931                 err = decrypt_skb_update(sk, skb, NULL, &chunk, &zc, false);
1932
1933                 /* splice does not support reading control messages */
1934                 if (ctx->control != TLS_RECORD_TYPE_DATA) {
1935                         err = -ENOTSUPP;
1936                         goto splice_read_end;
1937                 }
1938
1939                 if (err < 0) {
1940                         tls_err_abort(sk, EBADMSG);
1941                         goto splice_read_end;
1942                 }
1943                 ctx->decrypted = 1;
1944         }
1945         rxm = strp_msg(skb);
1946
1947         chunk = min_t(unsigned int, rxm->full_len, len);
1948         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, chunk, flags);
1949         if (copied < 0)
1950                 goto splice_read_end;
1951
1952         if (likely(!(flags & MSG_PEEK)))
1953                 tls_sw_advance_skb(sk, skb, copied);
1954
1955 splice_read_end:
1956         release_sock(sk);
1957         return copied ? : err;
1958 }
1959
1960 bool tls_sw_stream_read(const struct sock *sk)
1961 {
1962         struct tls_context *tls_ctx = tls_get_ctx(sk);
1963         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1964         bool ingress_empty = true;
1965         struct sk_psock *psock;
1966
1967         rcu_read_lock();
1968         psock = sk_psock(sk);
1969         if (psock)
1970                 ingress_empty = list_empty(&psock->ingress_msg);
1971         rcu_read_unlock();
1972
1973         return !ingress_empty || ctx->recv_pkt ||
1974                 !skb_queue_empty(&ctx->rx_list);
1975 }
1976
1977 static int tls_read_size(struct strparser *strp, struct sk_buff *skb)
1978 {
1979         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
1980         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1981         struct tls_prot_info *prot = &tls_ctx->prot_info;
1982         char header[TLS_HEADER_SIZE + MAX_IV_SIZE];
1983         struct strp_msg *rxm = strp_msg(skb);
1984         size_t cipher_overhead;
1985         size_t data_len = 0;
1986         int ret;
1987
1988         /* Verify that we have a full TLS header, or wait for more data */
1989         if (rxm->offset + prot->prepend_size > skb->len)
1990                 return 0;
1991
1992         /* Sanity-check size of on-stack buffer. */
1993         if (WARN_ON(prot->prepend_size > sizeof(header))) {
1994                 ret = -EINVAL;
1995                 goto read_failure;
1996         }
1997
1998         /* Linearize header to local buffer */
1999         ret = skb_copy_bits(skb, rxm->offset, header, prot->prepend_size);
2000
2001         if (ret < 0)
2002                 goto read_failure;
2003
2004         ctx->control = header[0];
2005
2006         data_len = ((header[4] & 0xFF) | (header[3] << 8));
2007
2008         cipher_overhead = prot->tag_size;
2009         if (prot->version != TLS_1_3_VERSION)
2010                 cipher_overhead += prot->iv_size;
2011
2012         if (data_len > TLS_MAX_PAYLOAD_SIZE + cipher_overhead +
2013             prot->tail_size) {
2014                 ret = -EMSGSIZE;
2015                 goto read_failure;
2016         }
2017         if (data_len < cipher_overhead) {
2018                 ret = -EBADMSG;
2019                 goto read_failure;
2020         }
2021
2022         /* Note that both TLS1.3 and TLS1.2 use TLS_1_2 version here */
2023         if (header[1] != TLS_1_2_VERSION_MINOR ||
2024             header[2] != TLS_1_2_VERSION_MAJOR) {
2025                 ret = -EINVAL;
2026                 goto read_failure;
2027         }
2028
2029         tls_device_rx_resync_new_rec(strp->sk, data_len + TLS_HEADER_SIZE,
2030                                      TCP_SKB_CB(skb)->seq + rxm->offset);
2031         return data_len + TLS_HEADER_SIZE;
2032
2033 read_failure:
2034         tls_err_abort(strp->sk, ret);
2035
2036         return ret;
2037 }
2038
2039 static void tls_queue(struct strparser *strp, struct sk_buff *skb)
2040 {
2041         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2042         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2043
2044         ctx->decrypted = 0;
2045
2046         ctx->recv_pkt = skb;
2047         strp_pause(strp);
2048
2049         ctx->saved_data_ready(strp->sk);
2050 }
2051
2052 static void tls_data_ready(struct sock *sk)
2053 {
2054         struct tls_context *tls_ctx = tls_get_ctx(sk);
2055         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2056         struct sk_psock *psock;
2057
2058         strp_data_ready(&ctx->strp);
2059
2060         psock = sk_psock_get(sk);
2061         if (psock && !list_empty(&psock->ingress_msg)) {
2062                 ctx->saved_data_ready(sk);
2063                 sk_psock_put(sk, psock);
2064         }
2065 }
2066
2067 void tls_sw_cancel_work_tx(struct tls_context *tls_ctx)
2068 {
2069         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2070
2071         set_bit(BIT_TX_CLOSING, &ctx->tx_bitmask);
2072         set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask);
2073         cancel_delayed_work_sync(&ctx->tx_work.work);
2074 }
2075
2076 void tls_sw_release_resources_tx(struct sock *sk)
2077 {
2078         struct tls_context *tls_ctx = tls_get_ctx(sk);
2079         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2080         struct tls_rec *rec, *tmp;
2081
2082         /* Wait for any pending async encryptions to complete */
2083         smp_store_mb(ctx->async_notify, true);
2084         if (atomic_read(&ctx->encrypt_pending))
2085                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
2086
2087         tls_tx_records(sk, -1);
2088
2089         /* Free up un-sent records in tx_list. First, free
2090          * the partially sent record if any at head of tx_list.
2091          */
2092         if (tls_ctx->partially_sent_record) {
2093                 tls_free_partial_record(sk, tls_ctx);
2094                 rec = list_first_entry(&ctx->tx_list,
2095                                        struct tls_rec, list);
2096                 list_del(&rec->list);
2097                 sk_msg_free(sk, &rec->msg_plaintext);
2098                 kfree(rec);
2099         }
2100
2101         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
2102                 list_del(&rec->list);
2103                 sk_msg_free(sk, &rec->msg_encrypted);
2104                 sk_msg_free(sk, &rec->msg_plaintext);
2105                 kfree(rec);
2106         }
2107
2108         crypto_free_aead(ctx->aead_send);
2109         tls_free_open_rec(sk);
2110 }
2111
2112 void tls_sw_free_ctx_tx(struct tls_context *tls_ctx)
2113 {
2114         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2115
2116         kfree(ctx);
2117 }
2118
2119 void tls_sw_release_resources_rx(struct sock *sk)
2120 {
2121         struct tls_context *tls_ctx = tls_get_ctx(sk);
2122         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2123
2124         kfree(tls_ctx->rx.rec_seq);
2125         kfree(tls_ctx->rx.iv);
2126
2127         if (ctx->aead_recv) {
2128                 kfree_skb(ctx->recv_pkt);
2129                 ctx->recv_pkt = NULL;
2130                 skb_queue_purge(&ctx->rx_list);
2131                 crypto_free_aead(ctx->aead_recv);
2132                 strp_stop(&ctx->strp);
2133                 /* If tls_sw_strparser_arm() was not called (cleanup paths)
2134                  * we still want to strp_stop(), but sk->sk_data_ready was
2135                  * never swapped.
2136                  */
2137                 if (ctx->saved_data_ready) {
2138                         write_lock_bh(&sk->sk_callback_lock);
2139                         sk->sk_data_ready = ctx->saved_data_ready;
2140                         write_unlock_bh(&sk->sk_callback_lock);
2141                 }
2142         }
2143 }
2144
2145 void tls_sw_strparser_done(struct tls_context *tls_ctx)
2146 {
2147         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2148
2149         strp_done(&ctx->strp);
2150 }
2151
2152 void tls_sw_free_ctx_rx(struct tls_context *tls_ctx)
2153 {
2154         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2155
2156         kfree(ctx);
2157 }
2158
2159 void tls_sw_free_resources_rx(struct sock *sk)
2160 {
2161         struct tls_context *tls_ctx = tls_get_ctx(sk);
2162
2163         tls_sw_release_resources_rx(sk);
2164         tls_sw_free_ctx_rx(tls_ctx);
2165 }
2166
2167 /* The work handler to transmitt the encrypted records in tx_list */
2168 static void tx_work_handler(struct work_struct *work)
2169 {
2170         struct delayed_work *delayed_work = to_delayed_work(work);
2171         struct tx_work *tx_work = container_of(delayed_work,
2172                                                struct tx_work, work);
2173         struct sock *sk = tx_work->sk;
2174         struct tls_context *tls_ctx = tls_get_ctx(sk);
2175         struct tls_sw_context_tx *ctx;
2176
2177         if (unlikely(!tls_ctx))
2178                 return;
2179
2180         ctx = tls_sw_ctx_tx(tls_ctx);
2181         if (test_bit(BIT_TX_CLOSING, &ctx->tx_bitmask))
2182                 return;
2183
2184         if (!test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
2185                 return;
2186         mutex_lock(&tls_ctx->tx_lock);
2187         lock_sock(sk);
2188         tls_tx_records(sk, -1);
2189         release_sock(sk);
2190         mutex_unlock(&tls_ctx->tx_lock);
2191 }
2192
2193 void tls_sw_write_space(struct sock *sk, struct tls_context *ctx)
2194 {
2195         struct tls_sw_context_tx *tx_ctx = tls_sw_ctx_tx(ctx);
2196
2197         /* Schedule the transmission if tx list is ready */
2198         if (is_tx_ready(tx_ctx) &&
2199             !test_and_set_bit(BIT_TX_SCHEDULED, &tx_ctx->tx_bitmask))
2200                 schedule_delayed_work(&tx_ctx->tx_work.work, 0);
2201 }
2202
2203 void tls_sw_strparser_arm(struct sock *sk, struct tls_context *tls_ctx)
2204 {
2205         struct tls_sw_context_rx *rx_ctx = tls_sw_ctx_rx(tls_ctx);
2206
2207         write_lock_bh(&sk->sk_callback_lock);
2208         rx_ctx->saved_data_ready = sk->sk_data_ready;
2209         sk->sk_data_ready = tls_data_ready;
2210         write_unlock_bh(&sk->sk_callback_lock);
2211
2212         strp_check_rcv(&rx_ctx->strp);
2213 }
2214
2215 int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
2216 {
2217         struct tls_context *tls_ctx = tls_get_ctx(sk);
2218         struct tls_prot_info *prot = &tls_ctx->prot_info;
2219         struct tls_crypto_info *crypto_info;
2220         struct tls12_crypto_info_aes_gcm_128 *gcm_128_info;
2221         struct tls12_crypto_info_aes_gcm_256 *gcm_256_info;
2222         struct tls12_crypto_info_aes_ccm_128 *ccm_128_info;
2223         struct tls_sw_context_tx *sw_ctx_tx = NULL;
2224         struct tls_sw_context_rx *sw_ctx_rx = NULL;
2225         struct cipher_context *cctx;
2226         struct crypto_aead **aead;
2227         struct strp_callbacks cb;
2228         u16 nonce_size, tag_size, iv_size, rec_seq_size, salt_size;
2229         struct crypto_tfm *tfm;
2230         char *iv, *rec_seq, *key, *salt, *cipher_name;
2231         size_t keysize;
2232         int rc = 0;
2233
2234         if (!ctx) {
2235                 rc = -EINVAL;
2236                 goto out;
2237         }
2238
2239         if (tx) {
2240                 if (!ctx->priv_ctx_tx) {
2241                         sw_ctx_tx = kzalloc(sizeof(*sw_ctx_tx), GFP_KERNEL);
2242                         if (!sw_ctx_tx) {
2243                                 rc = -ENOMEM;
2244                                 goto out;
2245                         }
2246                         ctx->priv_ctx_tx = sw_ctx_tx;
2247                 } else {
2248                         sw_ctx_tx =
2249                                 (struct tls_sw_context_tx *)ctx->priv_ctx_tx;
2250                 }
2251         } else {
2252                 if (!ctx->priv_ctx_rx) {
2253                         sw_ctx_rx = kzalloc(sizeof(*sw_ctx_rx), GFP_KERNEL);
2254                         if (!sw_ctx_rx) {
2255                                 rc = -ENOMEM;
2256                                 goto out;
2257                         }
2258                         ctx->priv_ctx_rx = sw_ctx_rx;
2259                 } else {
2260                         sw_ctx_rx =
2261                                 (struct tls_sw_context_rx *)ctx->priv_ctx_rx;
2262                 }
2263         }
2264
2265         if (tx) {
2266                 crypto_init_wait(&sw_ctx_tx->async_wait);
2267                 crypto_info = &ctx->crypto_send.info;
2268                 cctx = &ctx->tx;
2269                 aead = &sw_ctx_tx->aead_send;
2270                 INIT_LIST_HEAD(&sw_ctx_tx->tx_list);
2271                 INIT_DELAYED_WORK(&sw_ctx_tx->tx_work.work, tx_work_handler);
2272                 sw_ctx_tx->tx_work.sk = sk;
2273         } else {
2274                 crypto_init_wait(&sw_ctx_rx->async_wait);
2275                 crypto_info = &ctx->crypto_recv.info;
2276                 cctx = &ctx->rx;
2277                 skb_queue_head_init(&sw_ctx_rx->rx_list);
2278                 aead = &sw_ctx_rx->aead_recv;
2279         }
2280
2281         switch (crypto_info->cipher_type) {
2282         case TLS_CIPHER_AES_GCM_128: {
2283                 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2284                 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
2285                 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2286                 iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
2287                 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
2288                 rec_seq =
2289                  ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
2290                 gcm_128_info =
2291                         (struct tls12_crypto_info_aes_gcm_128 *)crypto_info;
2292                 keysize = TLS_CIPHER_AES_GCM_128_KEY_SIZE;
2293                 key = gcm_128_info->key;
2294                 salt = gcm_128_info->salt;
2295                 salt_size = TLS_CIPHER_AES_GCM_128_SALT_SIZE;
2296                 cipher_name = "gcm(aes)";
2297                 break;
2298         }
2299         case TLS_CIPHER_AES_GCM_256: {
2300                 nonce_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2301                 tag_size = TLS_CIPHER_AES_GCM_256_TAG_SIZE;
2302                 iv_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2303                 iv = ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->iv;
2304                 rec_seq_size = TLS_CIPHER_AES_GCM_256_REC_SEQ_SIZE;
2305                 rec_seq =
2306                  ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->rec_seq;
2307                 gcm_256_info =
2308                         (struct tls12_crypto_info_aes_gcm_256 *)crypto_info;
2309                 keysize = TLS_CIPHER_AES_GCM_256_KEY_SIZE;
2310                 key = gcm_256_info->key;
2311                 salt = gcm_256_info->salt;
2312                 salt_size = TLS_CIPHER_AES_GCM_256_SALT_SIZE;
2313                 cipher_name = "gcm(aes)";
2314                 break;
2315         }
2316         case TLS_CIPHER_AES_CCM_128: {
2317                 nonce_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2318                 tag_size = TLS_CIPHER_AES_CCM_128_TAG_SIZE;
2319                 iv_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2320                 iv = ((struct tls12_crypto_info_aes_ccm_128 *)crypto_info)->iv;
2321                 rec_seq_size = TLS_CIPHER_AES_CCM_128_REC_SEQ_SIZE;
2322                 rec_seq =
2323                 ((struct tls12_crypto_info_aes_ccm_128 *)crypto_info)->rec_seq;
2324                 ccm_128_info =
2325                 (struct tls12_crypto_info_aes_ccm_128 *)crypto_info;
2326                 keysize = TLS_CIPHER_AES_CCM_128_KEY_SIZE;
2327                 key = ccm_128_info->key;
2328                 salt = ccm_128_info->salt;
2329                 salt_size = TLS_CIPHER_AES_CCM_128_SALT_SIZE;
2330                 cipher_name = "ccm(aes)";
2331                 break;
2332         }
2333         default:
2334                 rc = -EINVAL;
2335                 goto free_priv;
2336         }
2337
2338         /* Sanity-check the sizes for stack allocations. */
2339         if (iv_size > MAX_IV_SIZE || nonce_size > MAX_IV_SIZE ||
2340             rec_seq_size > TLS_MAX_REC_SEQ_SIZE) {
2341                 rc = -EINVAL;
2342                 goto free_priv;
2343         }
2344
2345         if (crypto_info->version == TLS_1_3_VERSION) {
2346                 nonce_size = 0;
2347                 prot->aad_size = TLS_HEADER_SIZE;
2348                 prot->tail_size = 1;
2349         } else {
2350                 prot->aad_size = TLS_AAD_SPACE_SIZE;
2351                 prot->tail_size = 0;
2352         }
2353
2354         prot->version = crypto_info->version;
2355         prot->cipher_type = crypto_info->cipher_type;
2356         prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
2357         prot->tag_size = tag_size;
2358         prot->overhead_size = prot->prepend_size +
2359                               prot->tag_size + prot->tail_size;
2360         prot->iv_size = iv_size;
2361         prot->salt_size = salt_size;
2362         cctx->iv = kmalloc(iv_size + salt_size, GFP_KERNEL);
2363         if (!cctx->iv) {
2364                 rc = -ENOMEM;
2365                 goto free_priv;
2366         }
2367         /* Note: 128 & 256 bit salt are the same size */
2368         prot->rec_seq_size = rec_seq_size;
2369         memcpy(cctx->iv, salt, salt_size);
2370         memcpy(cctx->iv + salt_size, iv, iv_size);
2371         cctx->rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
2372         if (!cctx->rec_seq) {
2373                 rc = -ENOMEM;
2374                 goto free_iv;
2375         }
2376
2377         if (!*aead) {
2378                 *aead = crypto_alloc_aead(cipher_name, 0, 0);
2379                 if (IS_ERR(*aead)) {
2380                         rc = PTR_ERR(*aead);
2381                         *aead = NULL;
2382                         goto free_rec_seq;
2383                 }
2384         }
2385
2386         ctx->push_pending_record = tls_sw_push_pending_record;
2387
2388         rc = crypto_aead_setkey(*aead, key, keysize);
2389
2390         if (rc)
2391                 goto free_aead;
2392
2393         rc = crypto_aead_setauthsize(*aead, prot->tag_size);
2394         if (rc)
2395                 goto free_aead;
2396
2397         if (sw_ctx_rx) {
2398                 tfm = crypto_aead_tfm(sw_ctx_rx->aead_recv);
2399
2400                 if (crypto_info->version == TLS_1_3_VERSION)
2401                         sw_ctx_rx->async_capable = 0;
2402                 else
2403                         sw_ctx_rx->async_capable =
2404                                 !!(tfm->__crt_alg->cra_flags &
2405                                    CRYPTO_ALG_ASYNC);
2406
2407                 /* Set up strparser */
2408                 memset(&cb, 0, sizeof(cb));
2409                 cb.rcv_msg = tls_queue;
2410                 cb.parse_msg = tls_read_size;
2411
2412                 strp_init(&sw_ctx_rx->strp, sk, &cb);
2413         }
2414
2415         goto out;
2416
2417 free_aead:
2418         crypto_free_aead(*aead);
2419         *aead = NULL;
2420 free_rec_seq:
2421         kfree(cctx->rec_seq);
2422         cctx->rec_seq = NULL;
2423 free_iv:
2424         kfree(cctx->iv);
2425         cctx->iv = NULL;
2426 free_priv:
2427         if (tx) {
2428                 kfree(ctx->priv_ctx_tx);
2429                 ctx->priv_ctx_tx = NULL;
2430         } else {
2431                 kfree(ctx->priv_ctx_rx);
2432                 ctx->priv_ctx_rx = NULL;
2433         }
2434 out:
2435         return rc;
2436 }