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 #define MAX_IV_SIZE     TLS_CIPHER_AES_GCM_128_IV_SIZE
  46
  47 static int __skb_nsg(struct sk_buff *skb, int offset, int len,
  48                      unsigned int recursion_level)
  49 {
  50         int start = skb_headlen(skb);
  51         int i, chunk = start - offset;
  52         struct sk_buff *frag_iter;
  53         int elt = 0;
  54
  55         if (unlikely(recursion_level >= 24))
  56                 return -EMSGSIZE;
  57
  58         if (chunk > 0) {
  59                 if (chunk > len)
  60                         chunk = len;
  61                 elt++;
  62                 len -= chunk;
  63                 if (len == 0)
  64                         return elt;
  65                 offset += chunk;
  66         }
  67
  68         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
  69                 int end;
  70
  71                 WARN_ON(start > offset + len);
  72
  73                 end = start + skb_frag_size(&skb_shinfo(skb)->frags[i]);
  74                 chunk = end - offset;
  75                 if (chunk > 0) {
  76                         if (chunk > len)
  77                                 chunk = len;
  78                         elt++;
  79                         len -= chunk;
  80                         if (len == 0)
  81                                 return elt;
  82                         offset += chunk;
  83                 }
  84                 start = end;
  85         }
  86
  87         if (unlikely(skb_has_frag_list(skb))) {
  88                 skb_walk_frags(skb, frag_iter) {
  89                         int end, ret;
  90
  91                         WARN_ON(start > offset + len);
  92
  93                         end = start + frag_iter->len;
  94                         chunk = end - offset;
  95                         if (chunk > 0) {
  96                                 if (chunk > len)
  97                                         chunk = len;
  98                                 ret = __skb_nsg(frag_iter, offset - start, chunk,
  99                                                 recursion_level + 1);
 100                                 if (unlikely(ret < 0))
 101                                         return ret;
 102                                 elt += ret;
 103                                 len -= chunk;
 104                                 if (len == 0)
 105                                         return elt;
 106                                 offset += chunk;
 107                         }
 108                         start = end;
 109                 }
 110         }
 111         BUG_ON(len);
 112         return elt;
 113 }
 114
 115 /* Return the number of scatterlist elements required to completely map the
 116  * skb, or -EMSGSIZE if the recursion depth is exceeded.
 117  */
 118 static int skb_nsg(struct sk_buff *skb, int offset, int len)
 119 {
 120         return __skb_nsg(skb, offset, len, 0);
 121 }
 122
 123 static int padding_length(struct tls_sw_context_rx *ctx,
 124                           struct tls_context *tls_ctx, struct sk_buff *skb)
 125 {
 126         struct strp_msg *rxm = strp_msg(skb);
 127         int sub = 0;
 128
 129         /* Determine zero-padding length */
 130         if (tls_ctx->prot_info.version == TLS_1_3_VERSION) {
 131                 char content_type = 0;
 132                 int err;
 133                 int back = 17;
 134
 135                 while (content_type == 0) {
 136                         if (back > rxm->full_len)
 137                                 return -EBADMSG;
 138                         err = skb_copy_bits(skb,
 139                                             rxm->offset + rxm->full_len - back,
 140                                             &content_type, 1);
 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                 ctx->async_wait.err = err;
 172                 tls_err_abort(skb->sk, err);
 173         } else {
 174                 struct strp_msg *rxm = strp_msg(skb);
 175                 rxm->full_len -= padding_length(ctx, tls_ctx, skb);
 176                 rxm->offset += prot->prepend_size;
 177                 rxm->full_len -= prot->overhead_size;
 178         }
 179
 180         /* After using skb->sk to propagate sk through crypto async callback
 181          * we need to NULL it again.
 182          */
 183         skb->sk = NULL;
 184
 185
 186         /* Free the destination pages if skb was not decrypted inplace */
 187         if (sgout != sgin) {
 188                 /* Skip the first S/G entry as it points to AAD */
 189                 for_each_sg(sg_next(sgout), sg, UINT_MAX, pages) {
 190                         if (!sg)
 191                                 break;
 192                         put_page(sg_page(sg));
 193                 }
 194         }
 195
 196         kfree(aead_req);
 197
 198         pending = atomic_dec_return(&ctx->decrypt_pending);
 199
 200         if (!pending && READ_ONCE(ctx->async_notify))
 201                 complete(&ctx->async_wait.completion);
 202 }
 203
 204 static int tls_do_decryption(struct sock *sk,
 205                              struct sk_buff *skb,
 206                              struct scatterlist *sgin,
 207                              struct scatterlist *sgout,
 208                              char *iv_recv,
 209                              size_t data_len,
 210                              struct aead_request *aead_req,
 211                              bool async)
 212 {
 213         struct tls_context *tls_ctx = tls_get_ctx(sk);
 214         struct tls_prot_info *prot = &tls_ctx->prot_info;
 215         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
 216         int ret;
 217
 218         aead_request_set_tfm(aead_req, ctx->aead_recv);
 219         aead_request_set_ad(aead_req, prot->aad_size);
 220         aead_request_set_crypt(aead_req, sgin, sgout,
 221                                data_len + prot->tag_size,
 222                                (u8 *)iv_recv);
 223
 224         if (async) {
 225                 /* Using skb->sk to push sk through to crypto async callback
 226                  * handler. This allows propagating errors up to the socket
 227                  * if needed. It _must_ be cleared in the async handler
 228                  * before kfree_skb is called. We _know_ skb->sk is NULL
 229                  * because it is a clone from strparser.
 230                  */
 231                 skb->sk = sk;
 232                 aead_request_set_callback(aead_req,
 233                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 234                                           tls_decrypt_done, skb);
 235                 atomic_inc(&ctx->decrypt_pending);
 236         } else {
 237                 aead_request_set_callback(aead_req,
 238                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 239                                           crypto_req_done, &ctx->async_wait);
 240         }
 241
 242         ret = crypto_aead_decrypt(aead_req);
 243         if (ret == -EINPROGRESS) {
 244                 if (async)
 245                         return ret;
 246
 247                 ret = crypto_wait_req(ret, &ctx->async_wait);
 248         }
 249
 250         if (async)
 251                 atomic_dec(&ctx->decrypt_pending);
 252
 253         return ret;
 254 }
 255
 256 static void tls_trim_both_msgs(struct sock *sk, int target_size)
 257 {
 258         struct tls_context *tls_ctx = tls_get_ctx(sk);
 259         struct tls_prot_info *prot = &tls_ctx->prot_info;
 260         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 261         struct tls_rec *rec = ctx->open_rec;
 262
 263         sk_msg_trim(sk, &rec->msg_plaintext, target_size);
 264         if (target_size > 0)
 265                 target_size += prot->overhead_size;
 266         sk_msg_trim(sk, &rec->msg_encrypted, target_size);
 267 }
 268
 269 static int tls_alloc_encrypted_msg(struct sock *sk, int len)
 270 {
 271         struct tls_context *tls_ctx = tls_get_ctx(sk);
 272         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 273         struct tls_rec *rec = ctx->open_rec;
 274         struct sk_msg *msg_en = &rec->msg_encrypted;
 275
 276         return sk_msg_alloc(sk, msg_en, len, 0);
 277 }
 278
 279 static int tls_clone_plaintext_msg(struct sock *sk, int required)
 280 {
 281         struct tls_context *tls_ctx = tls_get_ctx(sk);
 282         struct tls_prot_info *prot = &tls_ctx->prot_info;
 283         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 284         struct tls_rec *rec = ctx->open_rec;
 285         struct sk_msg *msg_pl = &rec->msg_plaintext;
 286         struct sk_msg *msg_en = &rec->msg_encrypted;
 287         int skip, len;
 288
 289         /* We add page references worth len bytes from encrypted sg
 290          * at the end of plaintext sg. It is guaranteed that msg_en
 291          * has enough required room (ensured by caller).
 292          */
 293         len = required - msg_pl->sg.size;
 294
 295         /* Skip initial bytes in msg_en's data to be able to use
 296          * same offset of both plain and encrypted data.
 297          */
 298         skip = prot->prepend_size + msg_pl->sg.size;
 299
 300         return sk_msg_clone(sk, msg_pl, msg_en, skip, len);
 301 }
 302
 303 static struct tls_rec *tls_get_rec(struct sock *sk)
 304 {
 305         struct tls_context *tls_ctx = tls_get_ctx(sk);
 306         struct tls_prot_info *prot = &tls_ctx->prot_info;
 307         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 308         struct sk_msg *msg_pl, *msg_en;
 309         struct tls_rec *rec;
 310         int mem_size;
 311
 312         mem_size = sizeof(struct tls_rec) + crypto_aead_reqsize(ctx->aead_send);
 313
 314         rec = kzalloc(mem_size, sk->sk_allocation);
 315         if (!rec)
 316                 return NULL;
 317
 318         msg_pl = &rec->msg_plaintext;
 319         msg_en = &rec->msg_encrypted;
 320
 321         sk_msg_init(msg_pl);
 322         sk_msg_init(msg_en);
 323
 324         sg_init_table(rec->sg_aead_in, 2);
 325         sg_set_buf(&rec->sg_aead_in[0], rec->aad_space, prot->aad_size);
 326         sg_unmark_end(&rec->sg_aead_in[1]);
 327
 328         sg_init_table(rec->sg_aead_out, 2);
 329         sg_set_buf(&rec->sg_aead_out[0], rec->aad_space, prot->aad_size);
 330         sg_unmark_end(&rec->sg_aead_out[1]);
 331
 332         return rec;
 333 }
 334
 335 static void tls_free_rec(struct sock *sk, struct tls_rec *rec)
 336 {
 337         sk_msg_free(sk, &rec->msg_encrypted);
 338         sk_msg_free(sk, &rec->msg_plaintext);
 339         kfree(rec);
 340 }
 341
 342 static void tls_free_open_rec(struct sock *sk)
 343 {
 344         struct tls_context *tls_ctx = tls_get_ctx(sk);
 345         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 346         struct tls_rec *rec = ctx->open_rec;
 347
 348         if (rec) {
 349                 tls_free_rec(sk, rec);
 350                 ctx->open_rec = NULL;
 351         }
 352 }
 353
 354 int tls_tx_records(struct sock *sk, int flags)
 355 {
 356         struct tls_context *tls_ctx = tls_get_ctx(sk);
 357         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 358         struct tls_rec *rec, *tmp;
 359         struct sk_msg *msg_en;
 360         int tx_flags, rc = 0;
 361
 362         if (tls_is_partially_sent_record(tls_ctx)) {
 363                 rec = list_first_entry(&ctx->tx_list,
 364                                        struct tls_rec, list);
 365
 366                 if (flags == -1)
 367                         tx_flags = rec->tx_flags;
 368                 else
 369                         tx_flags = flags;
 370
 371                 rc = tls_push_partial_record(sk, tls_ctx, tx_flags);
 372                 if (rc)
 373                         goto tx_err;
 374
 375                 /* Full record has been transmitted.
 376                  * Remove the head of tx_list
 377                  */
 378                 list_del(&rec->list);
 379                 sk_msg_free(sk, &rec->msg_plaintext);
 380                 kfree(rec);
 381         }
 382
 383         /* Tx all ready records */
 384         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
 385                 if (READ_ONCE(rec->tx_ready)) {
 386                         if (flags == -1)
 387                                 tx_flags = rec->tx_flags;
 388                         else
 389                                 tx_flags = flags;
 390
 391                         msg_en = &rec->msg_encrypted;
 392                         rc = tls_push_sg(sk, tls_ctx,
 393                                          &msg_en->sg.data[msg_en->sg.curr],
 394                                          0, tx_flags);
 395                         if (rc)
 396                                 goto tx_err;
 397
 398                         list_del(&rec->list);
 399                         sk_msg_free(sk, &rec->msg_plaintext);
 400                         kfree(rec);
 401                 } else {
 402                         break;
 403                 }
 404         }
 405
 406 tx_err:
 407         if (rc < 0 && rc != -EAGAIN)
 408                 tls_err_abort(sk, EBADMSG);
 409
 410         return rc;
 411 }
 412
 413 static void tls_encrypt_done(struct crypto_async_request *req, int err)
 414 {
 415         struct aead_request *aead_req = (struct aead_request *)req;
 416         struct sock *sk = req->data;
 417         struct tls_context *tls_ctx = tls_get_ctx(sk);
 418         struct tls_prot_info *prot = &tls_ctx->prot_info;
 419         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 420         struct scatterlist *sge;
 421         struct sk_msg *msg_en;
 422         struct tls_rec *rec;
 423         bool ready = false;
 424         int pending;
 425
 426         rec = container_of(aead_req, struct tls_rec, aead_req);
 427         msg_en = &rec->msg_encrypted;
 428
 429         sge = sk_msg_elem(msg_en, msg_en->sg.curr);
 430         sge->offset -= prot->prepend_size;
 431         sge->length += prot->prepend_size;
 432
 433         /* Check if error is previously set on socket */
 434         if (err || sk->sk_err) {
 435                 rec = NULL;
 436
 437                 /* If err is already set on socket, return the same code */
 438                 if (sk->sk_err) {
 439                         ctx->async_wait.err = sk->sk_err;
 440                 } else {
 441                         ctx->async_wait.err = err;
 442                         tls_err_abort(sk, err);
 443                 }
 444         }
 445
 446         if (rec) {
 447                 struct tls_rec *first_rec;
 448
 449                 /* Mark the record as ready for transmission */
 450                 smp_store_mb(rec->tx_ready, true);
 451
 452                 /* If received record is at head of tx_list, schedule tx */
 453                 first_rec = list_first_entry(&ctx->tx_list,
 454                                              struct tls_rec, list);
 455                 if (rec == first_rec)
 456                         ready = true;
 457         }
 458
 459         pending = atomic_dec_return(&ctx->encrypt_pending);
 460
 461         if (!pending && READ_ONCE(ctx->async_notify))
 462                 complete(&ctx->async_wait.completion);
 463
 464         if (!ready)
 465                 return;
 466
 467         /* Schedule the transmission */
 468         if (!test_and_set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
 469                 schedule_delayed_work(&ctx->tx_work.work, 1);
 470 }
 471
 472 static int tls_do_encryption(struct sock *sk,
 473                              struct tls_context *tls_ctx,
 474                              struct tls_sw_context_tx *ctx,
 475                              struct aead_request *aead_req,
 476                              size_t data_len, u32 start)
 477 {
 478         struct tls_prot_info *prot = &tls_ctx->prot_info;
 479         struct tls_rec *rec = ctx->open_rec;
 480         struct sk_msg *msg_en = &rec->msg_encrypted;
 481         struct scatterlist *sge = sk_msg_elem(msg_en, start);
 482         int rc;
 483
 484         memcpy(rec->iv_data, tls_ctx->tx.iv, sizeof(rec->iv_data));
 485         xor_iv_with_seq(prot->version, rec->iv_data,
 486                         tls_ctx->tx.rec_seq);
 487
 488         sge->offset += prot->prepend_size;
 489         sge->length -= prot->prepend_size;
 490
 491         msg_en->sg.curr = start;
 492
 493         aead_request_set_tfm(aead_req, ctx->aead_send);
 494         aead_request_set_ad(aead_req, prot->aad_size);
 495         aead_request_set_crypt(aead_req, rec->sg_aead_in,
 496                                rec->sg_aead_out,
 497                                data_len, rec->iv_data);
 498
 499         aead_request_set_callback(aead_req, CRYPTO_TFM_REQ_MAY_BACKLOG,
 500                                   tls_encrypt_done, sk);
 501
 502         /* Add the record in tx_list */
 503         list_add_tail((struct list_head *)&rec->list, &ctx->tx_list);
 504         atomic_inc(&ctx->encrypt_pending);
 505
 506         rc = crypto_aead_encrypt(aead_req);
 507         if (!rc || rc != -EINPROGRESS) {
 508                 atomic_dec(&ctx->encrypt_pending);
 509                 sge->offset -= prot->prepend_size;
 510                 sge->length += prot->prepend_size;
 511         }
 512
 513         if (!rc) {
 514                 WRITE_ONCE(rec->tx_ready, true);
 515         } else if (rc != -EINPROGRESS) {
 516                 list_del(&rec->list);
 517                 return rc;
 518         }
 519
 520         /* Unhook the record from context if encryption is not failure */
 521         ctx->open_rec = NULL;
 522         tls_advance_record_sn(sk, &tls_ctx->tx, prot->version);
 523         return rc;
 524 }
 525
 526 static int tls_split_open_record(struct sock *sk, struct tls_rec *from,
 527                                  struct tls_rec **to, struct sk_msg *msg_opl,
 528                                  struct sk_msg *msg_oen, u32 split_point,
 529                                  u32 tx_overhead_size, u32 *orig_end)
 530 {
 531         u32 i, j, bytes = 0, apply = msg_opl->apply_bytes;
 532         struct scatterlist *sge, *osge, *nsge;
 533         u32 orig_size = msg_opl->sg.size;
 534         struct scatterlist tmp = { };
 535         struct sk_msg *msg_npl;
 536         struct tls_rec *new;
 537         int ret;
 538
 539         new = tls_get_rec(sk);
 540         if (!new)
 541                 return -ENOMEM;
 542         ret = sk_msg_alloc(sk, &new->msg_encrypted, msg_opl->sg.size +
 543                            tx_overhead_size, 0);
 544         if (ret < 0) {
 545                 tls_free_rec(sk, new);
 546                 return ret;
 547         }
 548
 549         *orig_end = msg_opl->sg.end;
 550         i = msg_opl->sg.start;
 551         sge = sk_msg_elem(msg_opl, i);
 552         while (apply && sge->length) {
 553                 if (sge->length > apply) {
 554                         u32 len = sge->length - apply;
 555
 556                         get_page(sg_page(sge));
 557                         sg_set_page(&tmp, sg_page(sge), len,
 558                                     sge->offset + apply);
 559                         sge->length = apply;
 560                         bytes += apply;
 561                         apply = 0;
 562                 } else {
 563                         apply -= sge->length;
 564                         bytes += sge->length;
 565                 }
 566
 567                 sk_msg_iter_var_next(i);
 568                 if (i == msg_opl->sg.end)
 569                         break;
 570                 sge = sk_msg_elem(msg_opl, i);
 571         }
 572
 573         msg_opl->sg.end = i;
 574         msg_opl->sg.curr = i;
 575         msg_opl->sg.copybreak = 0;
 576         msg_opl->apply_bytes = 0;
 577         msg_opl->sg.size = bytes;
 578
 579         msg_npl = &new->msg_plaintext;
 580         msg_npl->apply_bytes = apply;
 581         msg_npl->sg.size = orig_size - bytes;
 582
 583         j = msg_npl->sg.start;
 584         nsge = sk_msg_elem(msg_npl, j);
 585         if (tmp.length) {
 586                 memcpy(nsge, &tmp, sizeof(*nsge));
 587                 sk_msg_iter_var_next(j);
 588                 nsge = sk_msg_elem(msg_npl, j);
 589         }
 590
 591         osge = sk_msg_elem(msg_opl, i);
 592         while (osge->length) {
 593                 memcpy(nsge, osge, sizeof(*nsge));
 594                 sg_unmark_end(nsge);
 595                 sk_msg_iter_var_next(i);
 596                 sk_msg_iter_var_next(j);
 597                 if (i == *orig_end)
 598                         break;
 599                 osge = sk_msg_elem(msg_opl, i);
 600                 nsge = sk_msg_elem(msg_npl, j);
 601         }
 602
 603         msg_npl->sg.end = j;
 604         msg_npl->sg.curr = j;
 605         msg_npl->sg.copybreak = 0;
 606
 607         *to = new;
 608         return 0;
 609 }
 610
 611 static void tls_merge_open_record(struct sock *sk, struct tls_rec *to,
 612                                   struct tls_rec *from, u32 orig_end)
 613 {
 614         struct sk_msg *msg_npl = &from->msg_plaintext;
 615         struct sk_msg *msg_opl = &to->msg_plaintext;
 616         struct scatterlist *osge, *nsge;
 617         u32 i, j;
 618
 619         i = msg_opl->sg.end;
 620         sk_msg_iter_var_prev(i);
 621         j = msg_npl->sg.start;
 622
 623         osge = sk_msg_elem(msg_opl, i);
 624         nsge = sk_msg_elem(msg_npl, j);
 625
 626         if (sg_page(osge) == sg_page(nsge) &&
 627             osge->offset + osge->length == nsge->offset) {
 628                 osge->length += nsge->length;
 629                 put_page(sg_page(nsge));
 630         }
 631
 632         msg_opl->sg.end = orig_end;
 633         msg_opl->sg.curr = orig_end;
 634         msg_opl->sg.copybreak = 0;
 635         msg_opl->apply_bytes = msg_opl->sg.size + msg_npl->sg.size;
 636         msg_opl->sg.size += msg_npl->sg.size;
 637
 638         sk_msg_free(sk, &to->msg_encrypted);
 639         sk_msg_xfer_full(&to->msg_encrypted, &from->msg_encrypted);
 640
 641         kfree(from);
 642 }
 643
 644 static int tls_push_record(struct sock *sk, int flags,
 645                            unsigned char record_type)
 646 {
 647         struct tls_context *tls_ctx = tls_get_ctx(sk);
 648         struct tls_prot_info *prot = &tls_ctx->prot_info;
 649         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 650         struct tls_rec *rec = ctx->open_rec, *tmp = NULL;
 651         u32 i, split_point, uninitialized_var(orig_end);
 652         struct sk_msg *msg_pl, *msg_en;
 653         struct aead_request *req;
 654         bool split;
 655         int rc;
 656
 657         if (!rec)
 658                 return 0;
 659
 660         msg_pl = &rec->msg_plaintext;
 661         msg_en = &rec->msg_encrypted;
 662
 663         split_point = msg_pl->apply_bytes;
 664         split = split_point && split_point < msg_pl->sg.size;
 665         if (split) {
 666                 rc = tls_split_open_record(sk, rec, &tmp, msg_pl, msg_en,
 667                                            split_point, prot->overhead_size,
 668                                            &orig_end);
 669                 if (rc < 0)
 670                         return rc;
 671                 sk_msg_trim(sk, msg_en, msg_pl->sg.size +
 672                             prot->overhead_size);
 673         }
 674
 675         rec->tx_flags = flags;
 676         req = &rec->aead_req;
 677
 678         i = msg_pl->sg.end;
 679         sk_msg_iter_var_prev(i);
 680
 681         rec->content_type = record_type;
 682         if (prot->version == TLS_1_3_VERSION) {
 683                 /* Add content type to end of message.  No padding added */
 684                 sg_set_buf(&rec->sg_content_type, &rec->content_type, 1);
 685                 sg_mark_end(&rec->sg_content_type);
 686                 sg_chain(msg_pl->sg.data, msg_pl->sg.end + 1,
 687                          &rec->sg_content_type);
 688         } else {
 689                 sg_mark_end(sk_msg_elem(msg_pl, i));
 690         }
 691
 692         i = msg_pl->sg.start;
 693         sg_chain(rec->sg_aead_in, 2, rec->inplace_crypto ?
 694                  &msg_en->sg.data[i] : &msg_pl->sg.data[i]);
 695
 696         i = msg_en->sg.end;
 697         sk_msg_iter_var_prev(i);
 698         sg_mark_end(sk_msg_elem(msg_en, i));
 699
 700         i = msg_en->sg.start;
 701         sg_chain(rec->sg_aead_out, 2, &msg_en->sg.data[i]);
 702
 703         tls_make_aad(rec->aad_space, msg_pl->sg.size + prot->tail_size,
 704                      tls_ctx->tx.rec_seq, prot->rec_seq_size,
 705                      record_type, prot->version);
 706
 707         tls_fill_prepend(tls_ctx,
 708                          page_address(sg_page(&msg_en->sg.data[i])) +
 709                          msg_en->sg.data[i].offset,
 710                          msg_pl->sg.size + prot->tail_size,
 711                          record_type, prot->version);
 712
 713         tls_ctx->pending_open_record_frags = false;
 714
 715         rc = tls_do_encryption(sk, tls_ctx, ctx, req,
 716                                msg_pl->sg.size + prot->tail_size, i);
 717         if (rc < 0) {
 718                 if (rc != -EINPROGRESS) {
 719                         tls_err_abort(sk, EBADMSG);
 720                         if (split) {
 721                                 tls_ctx->pending_open_record_frags = true;
 722                                 tls_merge_open_record(sk, rec, tmp, orig_end);
 723                         }
 724                 }
 725                 ctx->async_capable = 1;
 726                 return rc;
 727         } else if (split) {
 728                 msg_pl = &tmp->msg_plaintext;
 729                 msg_en = &tmp->msg_encrypted;
 730                 sk_msg_trim(sk, msg_en, msg_pl->sg.size + prot->overhead_size);
 731                 tls_ctx->pending_open_record_frags = true;
 732                 ctx->open_rec = tmp;
 733         }
 734
 735         return tls_tx_records(sk, flags);
 736 }
 737
 738 static int bpf_exec_tx_verdict(struct sk_msg *msg, struct sock *sk,
 739                                bool full_record, u8 record_type,
 740                                size_t *copied, int flags)
 741 {
 742         struct tls_context *tls_ctx = tls_get_ctx(sk);
 743         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 744         struct sk_msg msg_redir = { };
 745         struct sk_psock *psock;
 746         struct sock *sk_redir;
 747         struct tls_rec *rec;
 748         bool enospc, policy;
 749         int err = 0, send;
 750         u32 delta = 0;
 751
 752         policy = !(flags & MSG_SENDPAGE_NOPOLICY);
 753         psock = sk_psock_get(sk);
 754         if (!psock || !policy)
 755                 return tls_push_record(sk, flags, record_type);
 756 more_data:
 757         enospc = sk_msg_full(msg);
 758         if (psock->eval == __SK_NONE) {
 759                 delta = msg->sg.size;
 760                 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
 761                 if (delta < msg->sg.size)
 762                         delta -= msg->sg.size;
 763                 else
 764                         delta = 0;
 765         }
 766         if (msg->cork_bytes && msg->cork_bytes > msg->sg.size &&
 767             !enospc && !full_record) {
 768                 err = -ENOSPC;
 769                 goto out_err;
 770         }
 771         msg->cork_bytes = 0;
 772         send = msg->sg.size;
 773         if (msg->apply_bytes && msg->apply_bytes < send)
 774                 send = msg->apply_bytes;
 775
 776         switch (psock->eval) {
 777         case __SK_PASS:
 778                 err = tls_push_record(sk, flags, record_type);
 779                 if (err < 0) {
 780                         *copied -= sk_msg_free(sk, msg);
 781                         tls_free_open_rec(sk);
 782                         goto out_err;
 783                 }
 784                 break;
 785         case __SK_REDIRECT:
 786                 sk_redir = psock->sk_redir;
 787                 memcpy(&msg_redir, msg, sizeof(*msg));
 788                 if (msg->apply_bytes < send)
 789                         msg->apply_bytes = 0;
 790                 else
 791                         msg->apply_bytes -= send;
 792                 sk_msg_return_zero(sk, msg, send);
 793                 msg->sg.size -= send;
 794                 release_sock(sk);
 795                 err = tcp_bpf_sendmsg_redir(sk_redir, &msg_redir, send, flags);
 796                 lock_sock(sk);
 797                 if (err < 0) {
 798                         *copied -= sk_msg_free_nocharge(sk, &msg_redir);
 799                         msg->sg.size = 0;
 800                 }
 801                 if (msg->sg.size == 0)
 802                         tls_free_open_rec(sk);
 803                 break;
 804         case __SK_DROP:
 805         default:
 806                 sk_msg_free_partial(sk, msg, send);
 807                 if (msg->apply_bytes < send)
 808                         msg->apply_bytes = 0;
 809                 else
 810                         msg->apply_bytes -= send;
 811                 if (msg->sg.size == 0)
 812                         tls_free_open_rec(sk);
 813                 *copied -= (send + delta);
 814                 err = -EACCES;
 815         }
 816
 817         if (likely(!err)) {
 818                 bool reset_eval = !ctx->open_rec;
 819
 820                 rec = ctx->open_rec;
 821                 if (rec) {
 822                         msg = &rec->msg_plaintext;
 823                         if (!msg->apply_bytes)
 824                                 reset_eval = true;
 825                 }
 826                 if (reset_eval) {
 827                         psock->eval = __SK_NONE;
 828                         if (psock->sk_redir) {
 829                                 sock_put(psock->sk_redir);
 830                                 psock->sk_redir = NULL;
 831                         }
 832                 }
 833                 if (rec)
 834                         goto more_data;
 835         }
 836  out_err:
 837         sk_psock_put(sk, psock);
 838         return err;
 839 }
 840
 841 static int tls_sw_push_pending_record(struct sock *sk, int flags)
 842 {
 843         struct tls_context *tls_ctx = tls_get_ctx(sk);
 844         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 845         struct tls_rec *rec = ctx->open_rec;
 846         struct sk_msg *msg_pl;
 847         size_t copied;
 848
 849         if (!rec)
 850                 return 0;
 851
 852         msg_pl = &rec->msg_plaintext;
 853         copied = msg_pl->sg.size;
 854         if (!copied)
 855                 return 0;
 856
 857         return bpf_exec_tx_verdict(msg_pl, sk, true, TLS_RECORD_TYPE_DATA,
 858                                    &copied, flags);
 859 }
 860
 861 int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 862 {
 863         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 864         struct tls_context *tls_ctx = tls_get_ctx(sk);
 865         struct tls_prot_info *prot = &tls_ctx->prot_info;
 866         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 867         bool async_capable = ctx->async_capable;
 868         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 869         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
 870         bool eor = !(msg->msg_flags & MSG_MORE);
 871         size_t try_to_copy, copied = 0;
 872         struct sk_msg *msg_pl, *msg_en;
 873         struct tls_rec *rec;
 874         int required_size;
 875         int num_async = 0;
 876         bool full_record;
 877         int record_room;
 878         int num_zc = 0;
 879         int orig_size;
 880         int ret = 0;
 881
 882         if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
 883                 return -ENOTSUPP;
 884
 885         lock_sock(sk);
 886
 887         /* Wait till there is any pending write on socket */
 888         if (unlikely(sk->sk_write_pending)) {
 889                 ret = wait_on_pending_writer(sk, &timeo);
 890                 if (unlikely(ret))
 891                         goto send_end;
 892         }
 893
 894         if (unlikely(msg->msg_controllen)) {
 895                 ret = tls_proccess_cmsg(sk, msg, &record_type);
 896                 if (ret) {
 897                         if (ret == -EINPROGRESS)
 898                                 num_async++;
 899                         else if (ret != -EAGAIN)
 900                                 goto send_end;
 901                 }
 902         }
 903
 904         while (msg_data_left(msg)) {
 905                 if (sk->sk_err) {
 906                         ret = -sk->sk_err;
 907                         goto send_end;
 908                 }
 909
 910                 if (ctx->open_rec)
 911                         rec = ctx->open_rec;
 912                 else
 913                         rec = ctx->open_rec = tls_get_rec(sk);
 914                 if (!rec) {
 915                         ret = -ENOMEM;
 916                         goto send_end;
 917                 }
 918
 919                 msg_pl = &rec->msg_plaintext;
 920                 msg_en = &rec->msg_encrypted;
 921
 922                 orig_size = msg_pl->sg.size;
 923                 full_record = false;
 924                 try_to_copy = msg_data_left(msg);
 925                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
 926                 if (try_to_copy >= record_room) {
 927                         try_to_copy = record_room;
 928                         full_record = true;
 929                 }
 930
 931                 required_size = msg_pl->sg.size + try_to_copy +
 932                                 prot->overhead_size;
 933
 934                 if (!sk_stream_memory_free(sk))
 935                         goto wait_for_sndbuf;
 936
 937 alloc_encrypted:
 938                 ret = tls_alloc_encrypted_msg(sk, required_size);
 939                 if (ret) {
 940                         if (ret != -ENOSPC)
 941                                 goto wait_for_memory;
 942
 943                         /* Adjust try_to_copy according to the amount that was
 944                          * actually allocated. The difference is due
 945                          * to max sg elements limit
 946                          */
 947                         try_to_copy -= required_size - msg_en->sg.size;
 948                         full_record = true;
 949                 }
 950
 951                 if (!is_kvec && (full_record || eor) && !async_capable) {
 952                         u32 first = msg_pl->sg.end;
 953
 954                         ret = sk_msg_zerocopy_from_iter(sk, &msg->msg_iter,
 955                                                         msg_pl, try_to_copy);
 956                         if (ret)
 957                                 goto fallback_to_reg_send;
 958
 959                         rec->inplace_crypto = 0;
 960
 961                         num_zc++;
 962                         copied += try_to_copy;
 963
 964                         sk_msg_sg_copy_set(msg_pl, first);
 965                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
 966                                                   record_type, &copied,
 967                                                   msg->msg_flags);
 968                         if (ret) {
 969                                 if (ret == -EINPROGRESS)
 970                                         num_async++;
 971                                 else if (ret == -ENOMEM)
 972                                         goto wait_for_memory;
 973                                 else if (ret == -ENOSPC)
 974                                         goto rollback_iter;
 975                                 else if (ret != -EAGAIN)
 976                                         goto send_end;
 977                         }
 978                         continue;
 979 rollback_iter:
 980                         copied -= try_to_copy;
 981                         sk_msg_sg_copy_clear(msg_pl, first);
 982                         iov_iter_revert(&msg->msg_iter,
 983                                         msg_pl->sg.size - orig_size);
 984 fallback_to_reg_send:
 985                         sk_msg_trim(sk, msg_pl, orig_size);
 986                 }
 987
 988                 required_size = msg_pl->sg.size + try_to_copy;
 989
 990                 ret = tls_clone_plaintext_msg(sk, required_size);
 991                 if (ret) {
 992                         if (ret != -ENOSPC)
 993                                 goto send_end;
 994
 995                         /* Adjust try_to_copy according to the amount that was
 996                          * actually allocated. The difference is due
 997                          * to max sg elements limit
 998                          */
 999                         try_to_copy -= required_size - msg_pl->sg.size;
1000                         full_record = true;
1001                         sk_msg_trim(sk, msg_en,
1002                                     msg_pl->sg.size + prot->overhead_size);
1003                 }
1004
1005                 if (try_to_copy) {
1006                         ret = sk_msg_memcopy_from_iter(sk, &msg->msg_iter,
1007                                                        msg_pl, try_to_copy);
1008                         if (ret < 0)
1009                                 goto trim_sgl;
1010                 }
1011
1012                 /* Open records defined only if successfully copied, otherwise
1013                  * we would trim the sg but not reset the open record frags.
1014                  */
1015                 tls_ctx->pending_open_record_frags = true;
1016                 copied += try_to_copy;
1017                 if (full_record || eor) {
1018                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1019                                                   record_type, &copied,
1020                                                   msg->msg_flags);
1021                         if (ret) {
1022                                 if (ret == -EINPROGRESS)
1023                                         num_async++;
1024                                 else if (ret == -ENOMEM)
1025                                         goto wait_for_memory;
1026                                 else if (ret != -EAGAIN) {
1027                                         if (ret == -ENOSPC)
1028                                                 ret = 0;
1029                                         goto send_end;
1030                                 }
1031                         }
1032                 }
1033
1034                 continue;
1035
1036 wait_for_sndbuf:
1037                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1038 wait_for_memory:
1039                 ret = sk_stream_wait_memory(sk, &timeo);
1040                 if (ret) {
1041 trim_sgl:
1042                         tls_trim_both_msgs(sk, orig_size);
1043                         goto send_end;
1044                 }
1045
1046                 if (msg_en->sg.size < required_size)
1047                         goto alloc_encrypted;
1048         }
1049
1050         if (!num_async) {
1051                 goto send_end;
1052         } else if (num_zc) {
1053                 /* Wait for pending encryptions to get completed */
1054                 smp_store_mb(ctx->async_notify, true);
1055
1056                 if (atomic_read(&ctx->encrypt_pending))
1057                         crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1058                 else
1059                         reinit_completion(&ctx->async_wait.completion);
1060
1061                 WRITE_ONCE(ctx->async_notify, false);
1062
1063                 if (ctx->async_wait.err) {
1064                         ret = ctx->async_wait.err;
1065                         copied = 0;
1066                 }
1067         }
1068
1069         /* Transmit if any encryptions have completed */
1070         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1071                 cancel_delayed_work(&ctx->tx_work.work);
1072                 tls_tx_records(sk, msg->msg_flags);
1073         }
1074
1075 send_end:
1076         ret = sk_stream_error(sk, msg->msg_flags, ret);
1077
1078         release_sock(sk);
1079         return copied ? copied : ret;
1080 }
1081
1082 static int tls_sw_do_sendpage(struct sock *sk, struct page *page,
1083                               int offset, size_t size, int flags)
1084 {
1085         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1086         struct tls_context *tls_ctx = tls_get_ctx(sk);
1087         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1088         struct tls_prot_info *prot = &tls_ctx->prot_info;
1089         unsigned char record_type = TLS_RECORD_TYPE_DATA;
1090         struct sk_msg *msg_pl;
1091         struct tls_rec *rec;
1092         int num_async = 0;
1093         size_t copied = 0;
1094         bool full_record;
1095         int record_room;
1096         int ret = 0;
1097         bool eor;
1098
1099         eor = !(flags & (MSG_MORE | MSG_SENDPAGE_NOTLAST));
1100         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1101
1102         /* Wait till there is any pending write on socket */
1103         if (unlikely(sk->sk_write_pending)) {
1104                 ret = wait_on_pending_writer(sk, &timeo);
1105                 if (unlikely(ret))
1106                         goto sendpage_end;
1107         }
1108
1109         /* Call the sk_stream functions to manage the sndbuf mem. */
1110         while (size > 0) {
1111                 size_t copy, required_size;
1112
1113                 if (sk->sk_err) {
1114                         ret = -sk->sk_err;
1115                         goto sendpage_end;
1116                 }
1117
1118                 if (ctx->open_rec)
1119                         rec = ctx->open_rec;
1120                 else
1121                         rec = ctx->open_rec = tls_get_rec(sk);
1122                 if (!rec) {
1123                         ret = -ENOMEM;
1124                         goto sendpage_end;
1125                 }
1126
1127                 msg_pl = &rec->msg_plaintext;
1128
1129                 full_record = false;
1130                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
1131                 copied = 0;
1132                 copy = size;
1133                 if (copy >= record_room) {
1134                         copy = record_room;
1135                         full_record = true;
1136                 }
1137
1138                 required_size = msg_pl->sg.size + copy + prot->overhead_size;
1139
1140                 if (!sk_stream_memory_free(sk))
1141                         goto wait_for_sndbuf;
1142 alloc_payload:
1143                 ret = tls_alloc_encrypted_msg(sk, required_size);
1144                 if (ret) {
1145                         if (ret != -ENOSPC)
1146                                 goto wait_for_memory;
1147
1148                         /* Adjust copy according to the amount that was
1149                          * actually allocated. The difference is due
1150                          * to max sg elements limit
1151                          */
1152                         copy -= required_size - msg_pl->sg.size;
1153                         full_record = true;
1154                 }
1155
1156                 sk_msg_page_add(msg_pl, page, copy, offset);
1157                 sk_mem_charge(sk, copy);
1158
1159                 offset += copy;
1160                 size -= copy;
1161                 copied += copy;
1162
1163                 tls_ctx->pending_open_record_frags = true;
1164                 if (full_record || eor || sk_msg_full(msg_pl)) {
1165                         rec->inplace_crypto = 0;
1166                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1167                                                   record_type, &copied, flags);
1168                         if (ret) {
1169                                 if (ret == -EINPROGRESS)
1170                                         num_async++;
1171                                 else if (ret == -ENOMEM)
1172                                         goto wait_for_memory;
1173                                 else if (ret != -EAGAIN) {
1174                                         if (ret == -ENOSPC)
1175                                                 ret = 0;
1176                                         goto sendpage_end;
1177                                 }
1178                         }
1179                 }
1180                 continue;
1181 wait_for_sndbuf:
1182                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1183 wait_for_memory:
1184                 ret = sk_stream_wait_memory(sk, &timeo);
1185                 if (ret) {
1186                         tls_trim_both_msgs(sk, msg_pl->sg.size);
1187                         goto sendpage_end;
1188                 }
1189
1190                 goto alloc_payload;
1191         }
1192
1193         if (num_async) {
1194                 /* Transmit if any encryptions have completed */
1195                 if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1196                         cancel_delayed_work(&ctx->tx_work.work);
1197                         tls_tx_records(sk, flags);
1198                 }
1199         }
1200 sendpage_end:
1201         ret = sk_stream_error(sk, flags, ret);
1202         return copied ? copied : ret;
1203 }
1204
1205 int tls_sw_sendpage(struct sock *sk, struct page *page,
1206                     int offset, size_t size, int flags)
1207 {
1208         int ret;
1209
1210         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1211                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY))
1212                 return -ENOTSUPP;
1213
1214         lock_sock(sk);
1215         ret = tls_sw_do_sendpage(sk, page, offset, size, flags);
1216         release_sock(sk);
1217         return ret;
1218 }
1219
1220 static struct sk_buff *tls_wait_data(struct sock *sk, struct sk_psock *psock,
1221                                      int flags, long timeo, int *err)
1222 {
1223         struct tls_context *tls_ctx = tls_get_ctx(sk);
1224         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1225         struct sk_buff *skb;
1226         DEFINE_WAIT_FUNC(wait, woken_wake_function);
1227
1228         while (!(skb = ctx->recv_pkt) && sk_psock_queue_empty(psock)) {
1229                 if (sk->sk_err) {
1230                         *err = sock_error(sk);
1231                         return NULL;
1232                 }
1233
1234                 if (sk->sk_shutdown & RCV_SHUTDOWN)
1235                         return NULL;
1236
1237                 if (sock_flag(sk, SOCK_DONE))
1238                         return NULL;
1239
1240                 if ((flags & MSG_DONTWAIT) || !timeo) {
1241                         *err = -EAGAIN;
1242                         return NULL;
1243                 }
1244
1245                 add_wait_queue(sk_sleep(sk), &wait);
1246                 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1247                 sk_wait_event(sk, &timeo,
1248                               ctx->recv_pkt != skb ||
1249                               !sk_psock_queue_empty(psock),
1250                               &wait);
1251                 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1252                 remove_wait_queue(sk_sleep(sk), &wait);
1253
1254                 /* Handle signals */
1255                 if (signal_pending(current)) {
1256                         *err = sock_intr_errno(timeo);
1257                         return NULL;
1258                 }
1259         }
1260
1261         return skb;
1262 }
1263
1264 static int tls_setup_from_iter(struct sock *sk, struct iov_iter *from,
1265                                int length, int *pages_used,
1266                                unsigned int *size_used,
1267                                struct scatterlist *to,
1268                                int to_max_pages)
1269 {
1270         int rc = 0, i = 0, num_elem = *pages_used, maxpages;
1271         struct page *pages[MAX_SKB_FRAGS];
1272         unsigned int size = *size_used;
1273         ssize_t copied, use;
1274         size_t offset;
1275
1276         while (length > 0) {
1277                 i = 0;
1278                 maxpages = to_max_pages - num_elem;
1279                 if (maxpages == 0) {
1280                         rc = -EFAULT;
1281                         goto out;
1282                 }
1283                 copied = iov_iter_get_pages(from, pages,
1284                                             length,
1285                                             maxpages, &offset);
1286                 if (copied <= 0) {
1287                         rc = -EFAULT;
1288                         goto out;
1289                 }
1290
1291                 iov_iter_advance(from, copied);
1292
1293                 length -= copied;
1294                 size += copied;
1295                 while (copied) {
1296                         use = min_t(int, copied, PAGE_SIZE - offset);
1297
1298                         sg_set_page(&to[num_elem],
1299                                     pages[i], use, offset);
1300                         sg_unmark_end(&to[num_elem]);
1301                         /* We do not uncharge memory from this API */
1302
1303                         offset = 0;
1304                         copied -= use;
1305
1306                         i++;
1307                         num_elem++;
1308                 }
1309         }
1310         /* Mark the end in the last sg entry if newly added */
1311         if (num_elem > *pages_used)
1312                 sg_mark_end(&to[num_elem - 1]);
1313 out:
1314         if (rc)
1315                 iov_iter_revert(from, size - *size_used);
1316         *size_used = size;
1317         *pages_used = num_elem;
1318
1319         return rc;
1320 }
1321
1322 /* This function decrypts the input skb into either out_iov or in out_sg
1323  * or in skb buffers itself. The input parameter 'zc' indicates if
1324  * zero-copy mode needs to be tried or not. With zero-copy mode, either
1325  * out_iov or out_sg must be non-NULL. In case both out_iov and out_sg are
1326  * NULL, then the decryption happens inside skb buffers itself, i.e.
1327  * zero-copy gets disabled and 'zc' is updated.
1328  */
1329
1330 static int decrypt_internal(struct sock *sk, struct sk_buff *skb,
1331                             struct iov_iter *out_iov,
1332                             struct scatterlist *out_sg,
1333                             int *chunk, bool *zc, bool async)
1334 {
1335         struct tls_context *tls_ctx = tls_get_ctx(sk);
1336         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1337         struct tls_prot_info *prot = &tls_ctx->prot_info;
1338         struct strp_msg *rxm = strp_msg(skb);
1339         int n_sgin, n_sgout, nsg, mem_size, aead_size, err, pages = 0;
1340         struct aead_request *aead_req;
1341         struct sk_buff *unused;
1342         u8 *aad, *iv, *mem = NULL;
1343         struct scatterlist *sgin = NULL;
1344         struct scatterlist *sgout = NULL;
1345         const int data_len = rxm->full_len - prot->overhead_size +
1346                              prot->tail_size;
1347
1348         if (*zc && (out_iov || out_sg)) {
1349                 if (out_iov)
1350                         n_sgout = iov_iter_npages(out_iov, INT_MAX) + 1;
1351                 else
1352                         n_sgout = sg_nents(out_sg);
1353                 n_sgin = skb_nsg(skb, rxm->offset + prot->prepend_size,
1354                                  rxm->full_len - prot->prepend_size);
1355         } else {
1356                 n_sgout = 0;
1357                 *zc = false;
1358                 n_sgin = skb_cow_data(skb, 0, &unused);
1359         }
1360
1361         if (n_sgin < 1)
1362                 return -EBADMSG;
1363
1364         /* Increment to accommodate AAD */
1365         n_sgin = n_sgin + 1;
1366
1367         nsg = n_sgin + n_sgout;
1368
1369         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(ctx->aead_recv);
1370         mem_size = aead_size + (nsg * sizeof(struct scatterlist));
1371         mem_size = mem_size + prot->aad_size;
1372         mem_size = mem_size + crypto_aead_ivsize(ctx->aead_recv);
1373
1374         /* Allocate a single block of memory which contains
1375          * aead_req || sgin[] || sgout[] || aad || iv.
1376          * This order achieves correct alignment for aead_req, sgin, sgout.
1377          */
1378         mem = kmalloc(mem_size, sk->sk_allocation);
1379         if (!mem)
1380                 return -ENOMEM;
1381
1382         /* Segment the allocated memory */
1383         aead_req = (struct aead_request *)mem;
1384         sgin = (struct scatterlist *)(mem + aead_size);
1385         sgout = sgin + n_sgin;
1386         aad = (u8 *)(sgout + n_sgout);
1387         iv = aad + prot->aad_size;
1388
1389         /* Prepare IV */
1390         err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
1391                             iv + TLS_CIPHER_AES_GCM_128_SALT_SIZE,
1392                             prot->iv_size);
1393         if (err < 0) {
1394                 kfree(mem);
1395                 return err;
1396         }
1397         if (prot->version == TLS_1_3_VERSION)
1398                 memcpy(iv, tls_ctx->rx.iv, crypto_aead_ivsize(ctx->aead_recv));
1399         else
1400                 memcpy(iv, tls_ctx->rx.iv, TLS_CIPHER_AES_GCM_128_SALT_SIZE);
1401
1402         xor_iv_with_seq(prot->version, iv, tls_ctx->rx.rec_seq);
1403
1404         /* Prepare AAD */
1405         tls_make_aad(aad, rxm->full_len - prot->overhead_size +
1406                      prot->tail_size,
1407                      tls_ctx->rx.rec_seq, prot->rec_seq_size,
1408                      ctx->control, prot->version);
1409
1410         /* Prepare sgin */
1411         sg_init_table(sgin, n_sgin);
1412         sg_set_buf(&sgin[0], aad, prot->aad_size);
1413         err = skb_to_sgvec(skb, &sgin[1],
1414                            rxm->offset + prot->prepend_size,
1415                            rxm->full_len - prot->prepend_size);
1416         if (err < 0) {
1417                 kfree(mem);
1418                 return err;
1419         }
1420
1421         if (n_sgout) {
1422                 if (out_iov) {
1423                         sg_init_table(sgout, n_sgout);
1424                         sg_set_buf(&sgout[0], aad, prot->aad_size);
1425
1426                         *chunk = 0;
1427                         err = tls_setup_from_iter(sk, out_iov, data_len,
1428                                                   &pages, chunk, &sgout[1],
1429                                                   (n_sgout - 1));
1430                         if (err < 0)
1431                                 goto fallback_to_reg_recv;
1432                 } else if (out_sg) {
1433                         memcpy(sgout, out_sg, n_sgout * sizeof(*sgout));
1434                 } else {
1435                         goto fallback_to_reg_recv;
1436                 }
1437         } else {
1438 fallback_to_reg_recv:
1439                 sgout = sgin;
1440                 pages = 0;
1441                 *chunk = data_len;
1442                 *zc = false;
1443         }
1444
1445         /* Prepare and submit AEAD request */
1446         err = tls_do_decryption(sk, skb, sgin, sgout, iv,
1447                                 data_len, aead_req, async);
1448         if (err == -EINPROGRESS)
1449                 return err;
1450
1451         /* Release the pages in case iov was mapped to pages */
1452         for (; pages > 0; pages--)
1453                 put_page(sg_page(&sgout[pages]));
1454
1455         kfree(mem);
1456         return err;
1457 }
1458
1459 static int decrypt_skb_update(struct sock *sk, struct sk_buff *skb,
1460                               struct iov_iter *dest, int *chunk, bool *zc,
1461                               bool async)
1462 {
1463         struct tls_context *tls_ctx = tls_get_ctx(sk);
1464         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1465         struct tls_prot_info *prot = &tls_ctx->prot_info;
1466         int version = prot->version;
1467         struct strp_msg *rxm = strp_msg(skb);
1468         int err = 0;
1469
1470         if (!ctx->decrypted) {
1471 #ifdef CONFIG_TLS_DEVICE
1472                 err = tls_device_decrypted(sk, skb);
1473                 if (err < 0)
1474                         return err;
1475 #endif
1476                 /* Still not decrypted after tls_device */
1477                 if (!ctx->decrypted) {
1478                         err = decrypt_internal(sk, skb, dest, NULL, chunk, zc,
1479                                                async);
1480                         if (err < 0) {
1481                                 if (err == -EINPROGRESS)
1482                                         tls_advance_record_sn(sk, &tls_ctx->rx,
1483                                                               version);
1484
1485                                 return err;
1486                         }
1487                 } else {
1488                         *zc = false;
1489                 }
1490
1491                 rxm->full_len -= padding_length(ctx, tls_ctx, skb);
1492                 rxm->offset += prot->prepend_size;
1493                 rxm->full_len -= prot->overhead_size;
1494                 tls_advance_record_sn(sk, &tls_ctx->rx, version);
1495                 ctx->decrypted = true;
1496                 ctx->saved_data_ready(sk);
1497         } else {
1498                 *zc = false;
1499         }
1500
1501         return err;
1502 }
1503
1504 int decrypt_skb(struct sock *sk, struct sk_buff *skb,
1505                 struct scatterlist *sgout)
1506 {
1507         bool zc = true;
1508         int chunk;
1509
1510         return decrypt_internal(sk, skb, NULL, sgout, &chunk, &zc, false);
1511 }
1512
1513 static bool tls_sw_advance_skb(struct sock *sk, struct sk_buff *skb,
1514                                unsigned int len)
1515 {
1516         struct tls_context *tls_ctx = tls_get_ctx(sk);
1517         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1518
1519         if (skb) {
1520                 struct strp_msg *rxm = strp_msg(skb);
1521
1522                 if (len < rxm->full_len) {
1523                         rxm->offset += len;
1524                         rxm->full_len -= len;
1525                         return false;
1526                 }
1527                 kfree_skb(skb);
1528         }
1529
1530         /* Finished with message */
1531         ctx->recv_pkt = NULL;
1532         __strp_unpause(&ctx->strp);
1533
1534         return true;
1535 }
1536
1537 /* This function traverses the rx_list in tls receive context to copies the
1538  * decrypted records into the buffer provided by caller zero copy is not
1539  * true. Further, the records are removed from the rx_list if it is not a peek
1540  * case and the record has been consumed completely.
1541  */
1542 static int process_rx_list(struct tls_sw_context_rx *ctx,
1543                            struct msghdr *msg,
1544                            u8 *control,
1545                            bool *cmsg,
1546                            size_t skip,
1547                            size_t len,
1548                            bool zc,
1549                            bool is_peek)
1550 {
1551         struct sk_buff *skb = skb_peek(&ctx->rx_list);
1552         u8 ctrl = *control;
1553         u8 msgc = *cmsg;
1554         struct tls_msg *tlm;
1555         ssize_t copied = 0;
1556
1557         /* Set the record type in 'control' if caller didn't pass it */
1558         if (!ctrl && skb) {
1559                 tlm = tls_msg(skb);
1560                 ctrl = tlm->control;
1561         }
1562
1563         while (skip && skb) {
1564                 struct strp_msg *rxm = strp_msg(skb);
1565                 tlm = tls_msg(skb);
1566
1567                 /* Cannot process a record of different type */
1568                 if (ctrl != tlm->control)
1569                         return 0;
1570
1571                 if (skip < rxm->full_len)
1572                         break;
1573
1574                 skip = skip - rxm->full_len;
1575                 skb = skb_peek_next(skb, &ctx->rx_list);
1576         }
1577
1578         while (len && skb) {
1579                 struct sk_buff *next_skb;
1580                 struct strp_msg *rxm = strp_msg(skb);
1581                 int chunk = min_t(unsigned int, rxm->full_len - skip, len);
1582
1583                 tlm = tls_msg(skb);
1584
1585                 /* Cannot process a record of different type */
1586                 if (ctrl != tlm->control)
1587                         return 0;
1588
1589                 /* Set record type if not already done. For a non-data record,
1590                  * do not proceed if record type could not be copied.
1591                  */
1592                 if (!msgc) {
1593                         int cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1594                                             sizeof(ctrl), &ctrl);
1595                         msgc = true;
1596                         if (ctrl != TLS_RECORD_TYPE_DATA) {
1597                                 if (cerr || msg->msg_flags & MSG_CTRUNC)
1598                                         return -EIO;
1599
1600                                 *cmsg = msgc;
1601                         }
1602                 }
1603
1604                 if (!zc || (rxm->full_len - skip) > len) {
1605                         int err = skb_copy_datagram_msg(skb, rxm->offset + skip,
1606                                                     msg, chunk);
1607                         if (err < 0)
1608                                 return err;
1609                 }
1610
1611                 len = len - chunk;
1612                 copied = copied + chunk;
1613
1614                 /* Consume the data from record if it is non-peek case*/
1615                 if (!is_peek) {
1616                         rxm->offset = rxm->offset + chunk;
1617                         rxm->full_len = rxm->full_len - chunk;
1618
1619                         /* Return if there is unconsumed data in the record */
1620                         if (rxm->full_len - skip)
1621                                 break;
1622                 }
1623
1624                 /* The remaining skip-bytes must lie in 1st record in rx_list.
1625                  * So from the 2nd record, 'skip' should be 0.
1626                  */
1627                 skip = 0;
1628
1629                 if (msg)
1630                         msg->msg_flags |= MSG_EOR;
1631
1632                 next_skb = skb_peek_next(skb, &ctx->rx_list);
1633
1634                 if (!is_peek) {
1635                         skb_unlink(skb, &ctx->rx_list);
1636                         kfree_skb(skb);
1637                 }
1638
1639                 skb = next_skb;
1640         }
1641
1642         *control = ctrl;
1643         return copied;
1644 }
1645
1646 int tls_sw_recvmsg(struct sock *sk,
1647                    struct msghdr *msg,
1648                    size_t len,
1649                    int nonblock,
1650                    int flags,
1651                    int *addr_len)
1652 {
1653         struct tls_context *tls_ctx = tls_get_ctx(sk);
1654         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1655         struct tls_prot_info *prot = &tls_ctx->prot_info;
1656         struct sk_psock *psock;
1657         unsigned char control = 0;
1658         ssize_t decrypted = 0;
1659         struct strp_msg *rxm;
1660         struct tls_msg *tlm;
1661         struct sk_buff *skb;
1662         ssize_t copied = 0;
1663         bool cmsg = false;
1664         int target, err = 0;
1665         long timeo;
1666         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
1667         bool is_peek = flags & MSG_PEEK;
1668         int num_async = 0;
1669
1670         flags |= nonblock;
1671
1672         if (unlikely(flags & MSG_ERRQUEUE))
1673                 return sock_recv_errqueue(sk, msg, len, SOL_IP, IP_RECVERR);
1674
1675         psock = sk_psock_get(sk);
1676         lock_sock(sk);
1677
1678         /* Process pending decrypted records. It must be non-zero-copy */
1679         err = process_rx_list(ctx, msg, &control, &cmsg, 0, len, false,
1680                               is_peek);
1681         if (err < 0) {
1682                 tls_err_abort(sk, err);
1683                 goto end;
1684         } else {
1685                 copied = err;
1686         }
1687
1688         len = len - copied;
1689         if (len) {
1690                 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1691                 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1692         } else {
1693                 goto recv_end;
1694         }
1695
1696         do {
1697                 bool retain_skb = false;
1698                 bool zc = false;
1699                 int to_decrypt;
1700                 int chunk = 0;
1701                 bool async_capable;
1702                 bool async = false;
1703
1704                 skb = tls_wait_data(sk, psock, flags, timeo, &err);
1705                 if (!skb) {
1706                         if (psock) {
1707                                 int ret = __tcp_bpf_recvmsg(sk, psock,
1708                                                             msg, len, flags);
1709
1710                                 if (ret > 0) {
1711                                         decrypted += ret;
1712                                         len -= ret;
1713                                         continue;
1714                                 }
1715                         }
1716                         goto recv_end;
1717                 } else {
1718                         tlm = tls_msg(skb);
1719                         if (prot->version == TLS_1_3_VERSION)
1720                                 tlm->control = 0;
1721                         else
1722                                 tlm->control = ctx->control;
1723                 }
1724
1725                 rxm = strp_msg(skb);
1726
1727                 to_decrypt = rxm->full_len - prot->overhead_size;
1728
1729                 if (to_decrypt <= len && !is_kvec && !is_peek &&
1730                     ctx->control == TLS_RECORD_TYPE_DATA &&
1731                     prot->version != TLS_1_3_VERSION)
1732                         zc = true;
1733
1734                 /* Do not use async mode if record is non-data */
1735                 if (ctx->control == TLS_RECORD_TYPE_DATA)
1736                         async_capable = ctx->async_capable;
1737                 else
1738                         async_capable = false;
1739
1740                 err = decrypt_skb_update(sk, skb, &msg->msg_iter,
1741                                          &chunk, &zc, async_capable);
1742                 if (err < 0 && err != -EINPROGRESS) {
1743                         tls_err_abort(sk, EBADMSG);
1744                         goto recv_end;
1745                 }
1746
1747                 if (err == -EINPROGRESS) {
1748                         async = true;
1749                         num_async++;
1750                 } else if (prot->version == TLS_1_3_VERSION) {
1751                         tlm->control = ctx->control;
1752                 }
1753
1754                 /* If the type of records being processed is not known yet,
1755                  * set it to record type just dequeued. If it is already known,
1756                  * but does not match the record type just dequeued, go to end.
1757                  * We always get record type here since for tls1.2, record type
1758                  * is known just after record is dequeued from stream parser.
1759                  * For tls1.3, we disable async.
1760                  */
1761
1762                 if (!control)
1763                         control = tlm->control;
1764                 else if (control != tlm->control)
1765                         goto recv_end;
1766
1767                 if (!cmsg) {
1768                         int cerr;
1769
1770                         cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1771                                         sizeof(control), &control);
1772                         cmsg = true;
1773                         if (control != TLS_RECORD_TYPE_DATA) {
1774                                 if (cerr || msg->msg_flags & MSG_CTRUNC) {
1775                                         err = -EIO;
1776                                         goto recv_end;
1777                                 }
1778                         }
1779                 }
1780
1781                 if (async)
1782                         goto pick_next_record;
1783
1784                 if (!zc) {
1785                         if (rxm->full_len > len) {
1786                                 retain_skb = true;
1787                                 chunk = len;
1788                         } else {
1789                                 chunk = rxm->full_len;
1790                         }
1791
1792                         err = skb_copy_datagram_msg(skb, rxm->offset,
1793                                                     msg, chunk);
1794                         if (err < 0)
1795                                 goto recv_end;
1796
1797                         if (!is_peek) {
1798                                 rxm->offset = rxm->offset + chunk;
1799                                 rxm->full_len = rxm->full_len - chunk;
1800                         }
1801                 }
1802
1803 pick_next_record:
1804                 if (chunk > len)
1805                         chunk = len;
1806
1807                 decrypted += chunk;
1808                 len -= chunk;
1809
1810                 /* For async or peek case, queue the current skb */
1811                 if (async || is_peek || retain_skb) {
1812                         skb_queue_tail(&ctx->rx_list, skb);
1813                         skb = NULL;
1814                 }
1815
1816                 if (tls_sw_advance_skb(sk, skb, chunk)) {
1817                         /* Return full control message to
1818                          * userspace before trying to parse
1819                          * another message type
1820                          */
1821                         msg->msg_flags |= MSG_EOR;
1822                         if (ctx->control != TLS_RECORD_TYPE_DATA)
1823                                 goto recv_end;
1824                 } else {
1825                         break;
1826                 }
1827
1828                 /* If we have a new message from strparser, continue now. */
1829                 if (decrypted >= target && !ctx->recv_pkt)
1830                         break;
1831         } while (len);
1832
1833 recv_end:
1834         if (num_async) {
1835                 /* Wait for all previously submitted records to be decrypted */
1836                 smp_store_mb(ctx->async_notify, true);
1837                 if (atomic_read(&ctx->decrypt_pending)) {
1838                         err = crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1839                         if (err) {
1840                                 /* one of async decrypt failed */
1841                                 tls_err_abort(sk, err);
1842                                 copied = 0;
1843                                 decrypted = 0;
1844                                 goto end;
1845                         }
1846                 } else {
1847                         reinit_completion(&ctx->async_wait.completion);
1848                 }
1849                 WRITE_ONCE(ctx->async_notify, false);
1850
1851                 /* Drain records from the rx_list & copy if required */
1852                 if (is_peek || is_kvec)
1853                         err = process_rx_list(ctx, msg, &control, &cmsg, copied,
1854                                               decrypted, false, is_peek);
1855                 else
1856                         err = process_rx_list(ctx, msg, &control, &cmsg, 0,
1857                                               decrypted, true, is_peek);
1858                 if (err < 0) {
1859                         tls_err_abort(sk, err);
1860                         copied = 0;
1861                         goto end;
1862                 }
1863         }
1864
1865         copied += decrypted;
1866
1867 end:
1868         release_sock(sk);
1869         if (psock)
1870                 sk_psock_put(sk, psock);
1871         return copied ? : err;
1872 }
1873
1874 ssize_t tls_sw_splice_read(struct socket *sock,  loff_t *ppos,
1875                            struct pipe_inode_info *pipe,
1876                            size_t len, unsigned int flags)
1877 {
1878         struct tls_context *tls_ctx = tls_get_ctx(sock->sk);
1879         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1880         struct strp_msg *rxm = NULL;
1881         struct sock *sk = sock->sk;
1882         struct sk_buff *skb;
1883         ssize_t copied = 0;
1884         int err = 0;
1885         long timeo;
1886         int chunk;
1887         bool zc = false;
1888
1889         lock_sock(sk);
1890
1891         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1892
1893         skb = tls_wait_data(sk, NULL, flags, timeo, &err);
1894         if (!skb)
1895                 goto splice_read_end;
1896
1897         if (!ctx->decrypted) {
1898                 err = decrypt_skb_update(sk, skb, NULL, &chunk, &zc, false);
1899
1900                 /* splice does not support reading control messages */
1901                 if (ctx->control != TLS_RECORD_TYPE_DATA) {
1902                         err = -ENOTSUPP;
1903                         goto splice_read_end;
1904                 }
1905
1906                 if (err < 0) {
1907                         tls_err_abort(sk, EBADMSG);
1908                         goto splice_read_end;
1909                 }
1910                 ctx->decrypted = true;
1911         }
1912         rxm = strp_msg(skb);
1913
1914         chunk = min_t(unsigned int, rxm->full_len, len);
1915         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, chunk, flags);
1916         if (copied < 0)
1917                 goto splice_read_end;
1918
1919         if (likely(!(flags & MSG_PEEK)))
1920                 tls_sw_advance_skb(sk, skb, copied);
1921
1922 splice_read_end:
1923         release_sock(sk);
1924         return copied ? : err;
1925 }
1926
1927 bool tls_sw_stream_read(const struct sock *sk)
1928 {
1929         struct tls_context *tls_ctx = tls_get_ctx(sk);
1930         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1931         bool ingress_empty = true;
1932         struct sk_psock *psock;
1933
1934         rcu_read_lock();
1935         psock = sk_psock(sk);
1936         if (psock)
1937                 ingress_empty = list_empty(&psock->ingress_msg);
1938         rcu_read_unlock();
1939
1940         return !ingress_empty || ctx->recv_pkt;
1941 }
1942
1943 static int tls_read_size(struct strparser *strp, struct sk_buff *skb)
1944 {
1945         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
1946         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1947         struct tls_prot_info *prot = &tls_ctx->prot_info;
1948         char header[TLS_HEADER_SIZE + MAX_IV_SIZE];
1949         struct strp_msg *rxm = strp_msg(skb);
1950         size_t cipher_overhead;
1951         size_t data_len = 0;
1952         int ret;
1953
1954         /* Verify that we have a full TLS header, or wait for more data */
1955         if (rxm->offset + prot->prepend_size > skb->len)
1956                 return 0;
1957
1958         /* Sanity-check size of on-stack buffer. */
1959         if (WARN_ON(prot->prepend_size > sizeof(header))) {
1960                 ret = -EINVAL;
1961                 goto read_failure;
1962         }
1963
1964         /* Linearize header to local buffer */
1965         ret = skb_copy_bits(skb, rxm->offset, header, prot->prepend_size);
1966
1967         if (ret < 0)
1968                 goto read_failure;
1969
1970         ctx->control = header[0];
1971
1972         data_len = ((header[4] & 0xFF) | (header[3] << 8));
1973
1974         cipher_overhead = prot->tag_size;
1975         if (prot->version != TLS_1_3_VERSION)
1976                 cipher_overhead += prot->iv_size;
1977
1978         if (data_len > TLS_MAX_PAYLOAD_SIZE + cipher_overhead +
1979             prot->tail_size) {
1980                 ret = -EMSGSIZE;
1981                 goto read_failure;
1982         }
1983         if (data_len < cipher_overhead) {
1984                 ret = -EBADMSG;
1985                 goto read_failure;
1986         }
1987
1988         /* Note that both TLS1.3 and TLS1.2 use TLS_1_2 version here */
1989         if (header[1] != TLS_1_2_VERSION_MINOR ||
1990             header[2] != TLS_1_2_VERSION_MAJOR) {
1991                 ret = -EINVAL;
1992                 goto read_failure;
1993         }
1994 #ifdef CONFIG_TLS_DEVICE
1995         handle_device_resync(strp->sk, TCP_SKB_CB(skb)->seq + rxm->offset,
1996                              *(u64*)tls_ctx->rx.rec_seq);
1997 #endif
1998         return data_len + TLS_HEADER_SIZE;
1999
2000 read_failure:
2001         tls_err_abort(strp->sk, ret);
2002
2003         return ret;
2004 }
2005
2006 static void tls_queue(struct strparser *strp, struct sk_buff *skb)
2007 {
2008         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2009         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2010
2011         ctx->decrypted = false;
2012
2013         ctx->recv_pkt = skb;
2014         strp_pause(strp);
2015
2016         ctx->saved_data_ready(strp->sk);
2017 }
2018
2019 static void tls_data_ready(struct sock *sk)
2020 {
2021         struct tls_context *tls_ctx = tls_get_ctx(sk);
2022         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2023         struct sk_psock *psock;
2024
2025         strp_data_ready(&ctx->strp);
2026
2027         psock = sk_psock_get(sk);
2028         if (psock && !list_empty(&psock->ingress_msg)) {
2029                 ctx->saved_data_ready(sk);
2030                 sk_psock_put(sk, psock);
2031         }
2032 }
2033
2034 void tls_sw_free_resources_tx(struct sock *sk)
2035 {
2036         struct tls_context *tls_ctx = tls_get_ctx(sk);
2037         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2038         struct tls_rec *rec, *tmp;
2039
2040         /* Wait for any pending async encryptions to complete */
2041         smp_store_mb(ctx->async_notify, true);
2042         if (atomic_read(&ctx->encrypt_pending))
2043                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
2044
2045         release_sock(sk);
2046         cancel_delayed_work_sync(&ctx->tx_work.work);
2047         lock_sock(sk);
2048
2049         /* Tx whatever records we can transmit and abandon the rest */
2050         tls_tx_records(sk, -1);
2051
2052         /* Free up un-sent records in tx_list. First, free
2053          * the partially sent record if any at head of tx_list.
2054          */
2055         if (tls_ctx->partially_sent_record) {
2056                 struct scatterlist *sg = tls_ctx->partially_sent_record;
2057
2058                 while (1) {
2059                         put_page(sg_page(sg));
2060                         sk_mem_uncharge(sk, sg->length);
2061
2062                         if (sg_is_last(sg))
2063                                 break;
2064                         sg++;
2065                 }
2066
2067                 tls_ctx->partially_sent_record = NULL;
2068
2069                 rec = list_first_entry(&ctx->tx_list,
2070                                        struct tls_rec, list);
2071                 list_del(&rec->list);
2072                 sk_msg_free(sk, &rec->msg_plaintext);
2073                 kfree(rec);
2074         }
2075
2076         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
2077                 list_del(&rec->list);
2078                 sk_msg_free(sk, &rec->msg_encrypted);
2079                 sk_msg_free(sk, &rec->msg_plaintext);
2080                 kfree(rec);
2081         }
2082
2083         crypto_free_aead(ctx->aead_send);
2084         tls_free_open_rec(sk);
2085
2086         kfree(ctx);
2087 }
2088
2089 void tls_sw_release_resources_rx(struct sock *sk)
2090 {
2091         struct tls_context *tls_ctx = tls_get_ctx(sk);
2092         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2093
2094         if (ctx->aead_recv) {
2095                 kfree_skb(ctx->recv_pkt);
2096                 ctx->recv_pkt = NULL;
2097                 skb_queue_purge(&ctx->rx_list);
2098                 crypto_free_aead(ctx->aead_recv);
2099                 strp_stop(&ctx->strp);
2100                 write_lock_bh(&sk->sk_callback_lock);
2101                 sk->sk_data_ready = ctx->saved_data_ready;
2102                 write_unlock_bh(&sk->sk_callback_lock);
2103                 release_sock(sk);
2104                 strp_done(&ctx->strp);
2105                 lock_sock(sk);
2106         }
2107 }
2108
2109 void tls_sw_free_resources_rx(struct sock *sk)
2110 {
2111         struct tls_context *tls_ctx = tls_get_ctx(sk);
2112         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2113
2114         tls_sw_release_resources_rx(sk);
2115
2116         kfree(ctx);
2117 }
2118
2119 /* The work handler to transmitt the encrypted records in tx_list */
2120 static void tx_work_handler(struct work_struct *work)
2121 {
2122         struct delayed_work *delayed_work = to_delayed_work(work);
2123         struct tx_work *tx_work = container_of(delayed_work,
2124                                                struct tx_work, work);
2125         struct sock *sk = tx_work->sk;
2126         struct tls_context *tls_ctx = tls_get_ctx(sk);
2127         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2128
2129         if (!test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
2130                 return;
2131
2132         lock_sock(sk);
2133         tls_tx_records(sk, -1);
2134         release_sock(sk);
2135 }
2136
2137 void tls_sw_write_space(struct sock *sk, struct tls_context *ctx)
2138 {
2139         struct tls_sw_context_tx *tx_ctx = tls_sw_ctx_tx(ctx);
2140
2141         /* Schedule the transmission if tx list is ready */
2142         if (is_tx_ready(tx_ctx) && !sk->sk_write_pending) {
2143                 /* Schedule the transmission */
2144                 if (!test_and_set_bit(BIT_TX_SCHEDULED,
2145                                       &tx_ctx->tx_bitmask))
2146                         schedule_delayed_work(&tx_ctx->tx_work.work, 0);
2147         }
2148 }
2149
2150 int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
2151 {
2152         struct tls_context *tls_ctx = tls_get_ctx(sk);
2153         struct tls_prot_info *prot = &tls_ctx->prot_info;
2154         struct tls_crypto_info *crypto_info;
2155         struct tls12_crypto_info_aes_gcm_128 *gcm_128_info;
2156         struct tls12_crypto_info_aes_gcm_256 *gcm_256_info;
2157         struct tls_sw_context_tx *sw_ctx_tx = NULL;
2158         struct tls_sw_context_rx *sw_ctx_rx = NULL;
2159         struct cipher_context *cctx;
2160         struct crypto_aead **aead;
2161         struct strp_callbacks cb;
2162         u16 nonce_size, tag_size, iv_size, rec_seq_size;
2163         struct crypto_tfm *tfm;
2164         char *iv, *rec_seq, *key, *salt;
2165         size_t keysize;
2166         int rc = 0;
2167
2168         if (!ctx) {
2169                 rc = -EINVAL;
2170                 goto out;
2171         }
2172
2173         if (tx) {
2174                 if (!ctx->priv_ctx_tx) {
2175                         sw_ctx_tx = kzalloc(sizeof(*sw_ctx_tx), GFP_KERNEL);
2176                         if (!sw_ctx_tx) {
2177                                 rc = -ENOMEM;
2178                                 goto out;
2179                         }
2180                         ctx->priv_ctx_tx = sw_ctx_tx;
2181                 } else {
2182                         sw_ctx_tx =
2183                                 (struct tls_sw_context_tx *)ctx->priv_ctx_tx;
2184                 }
2185         } else {
2186                 if (!ctx->priv_ctx_rx) {
2187                         sw_ctx_rx = kzalloc(sizeof(*sw_ctx_rx), GFP_KERNEL);
2188                         if (!sw_ctx_rx) {
2189                                 rc = -ENOMEM;
2190                                 goto out;
2191                         }
2192                         ctx->priv_ctx_rx = sw_ctx_rx;
2193                 } else {
2194                         sw_ctx_rx =
2195                                 (struct tls_sw_context_rx *)ctx->priv_ctx_rx;
2196                 }
2197         }
2198
2199         if (tx) {
2200                 crypto_init_wait(&sw_ctx_tx->async_wait);
2201                 crypto_info = &ctx->crypto_send.info;
2202                 cctx = &ctx->tx;
2203                 aead = &sw_ctx_tx->aead_send;
2204                 INIT_LIST_HEAD(&sw_ctx_tx->tx_list);
2205                 INIT_DELAYED_WORK(&sw_ctx_tx->tx_work.work, tx_work_handler);
2206                 sw_ctx_tx->tx_work.sk = sk;
2207         } else {
2208                 crypto_init_wait(&sw_ctx_rx->async_wait);
2209                 crypto_info = &ctx->crypto_recv.info;
2210                 cctx = &ctx->rx;
2211                 skb_queue_head_init(&sw_ctx_rx->rx_list);
2212                 aead = &sw_ctx_rx->aead_recv;
2213         }
2214
2215         switch (crypto_info->cipher_type) {
2216         case TLS_CIPHER_AES_GCM_128: {
2217                 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2218                 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
2219                 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2220                 iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
2221                 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
2222                 rec_seq =
2223                  ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
2224                 gcm_128_info =
2225                         (struct tls12_crypto_info_aes_gcm_128 *)crypto_info;
2226                 keysize = TLS_CIPHER_AES_GCM_128_KEY_SIZE;
2227                 key = gcm_128_info->key;
2228                 salt = gcm_128_info->salt;
2229                 break;
2230         }
2231         case TLS_CIPHER_AES_GCM_256: {
2232                 nonce_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2233                 tag_size = TLS_CIPHER_AES_GCM_256_TAG_SIZE;
2234                 iv_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2235                 iv = ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->iv;
2236                 rec_seq_size = TLS_CIPHER_AES_GCM_256_REC_SEQ_SIZE;
2237                 rec_seq =
2238                  ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->rec_seq;
2239                 gcm_256_info =
2240                         (struct tls12_crypto_info_aes_gcm_256 *)crypto_info;
2241                 keysize = TLS_CIPHER_AES_GCM_256_KEY_SIZE;
2242                 key = gcm_256_info->key;
2243                 salt = gcm_256_info->salt;
2244                 break;
2245         }
2246         default:
2247                 rc = -EINVAL;
2248                 goto free_priv;
2249         }
2250
2251         /* Sanity-check the IV size for stack allocations. */
2252         if (iv_size > MAX_IV_SIZE || nonce_size > MAX_IV_SIZE) {
2253                 rc = -EINVAL;
2254                 goto free_priv;
2255         }
2256
2257         if (crypto_info->version == TLS_1_3_VERSION) {
2258                 nonce_size = 0;
2259                 prot->aad_size = TLS_HEADER_SIZE;
2260                 prot->tail_size = 1;
2261         } else {
2262                 prot->aad_size = TLS_AAD_SPACE_SIZE;
2263                 prot->tail_size = 0;
2264         }
2265
2266         prot->version = crypto_info->version;
2267         prot->cipher_type = crypto_info->cipher_type;
2268         prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
2269         prot->tag_size = tag_size;
2270         prot->overhead_size = prot->prepend_size +
2271                               prot->tag_size + prot->tail_size;
2272         prot->iv_size = iv_size;
2273         cctx->iv = kmalloc(iv_size + TLS_CIPHER_AES_GCM_128_SALT_SIZE,
2274                            GFP_KERNEL);
2275         if (!cctx->iv) {
2276                 rc = -ENOMEM;
2277                 goto free_priv;
2278         }
2279         /* Note: 128 & 256 bit salt are the same size */
2280         memcpy(cctx->iv, salt, TLS_CIPHER_AES_GCM_128_SALT_SIZE);
2281         memcpy(cctx->iv + TLS_CIPHER_AES_GCM_128_SALT_SIZE, iv, iv_size);
2282         prot->rec_seq_size = rec_seq_size;
2283         cctx->rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
2284         if (!cctx->rec_seq) {
2285                 rc = -ENOMEM;
2286                 goto free_iv;
2287         }
2288
2289         if (!*aead) {
2290                 *aead = crypto_alloc_aead("gcm(aes)", 0, 0);
2291                 if (IS_ERR(*aead)) {
2292                         rc = PTR_ERR(*aead);
2293                         *aead = NULL;
2294                         goto free_rec_seq;
2295                 }
2296         }
2297
2298         ctx->push_pending_record = tls_sw_push_pending_record;
2299
2300         rc = crypto_aead_setkey(*aead, key, keysize);
2301
2302         if (rc)
2303                 goto free_aead;
2304
2305         rc = crypto_aead_setauthsize(*aead, prot->tag_size);
2306         if (rc)
2307                 goto free_aead;
2308
2309         if (sw_ctx_rx) {
2310                 tfm = crypto_aead_tfm(sw_ctx_rx->aead_recv);
2311
2312                 if (crypto_info->version == TLS_1_3_VERSION)
2313                         sw_ctx_rx->async_capable = false;
2314                 else
2315                         sw_ctx_rx->async_capable =
2316                                 tfm->__crt_alg->cra_flags & CRYPTO_ALG_ASYNC;
2317
2318                 /* Set up strparser */
2319                 memset(&cb, 0, sizeof(cb));
2320                 cb.rcv_msg = tls_queue;
2321                 cb.parse_msg = tls_read_size;
2322
2323                 strp_init(&sw_ctx_rx->strp, sk, &cb);
2324
2325                 write_lock_bh(&sk->sk_callback_lock);
2326                 sw_ctx_rx->saved_data_ready = sk->sk_data_ready;
2327                 sk->sk_data_ready = tls_data_ready;
2328                 write_unlock_bh(&sk->sk_callback_lock);
2329
2330                 strp_check_rcv(&sw_ctx_rx->strp);
2331         }
2332
2333         goto out;
2334
2335 free_aead:
2336         crypto_free_aead(*aead);
2337         *aead = NULL;
2338 free_rec_seq:
2339         kfree(cctx->rec_seq);
2340         cctx->rec_seq = NULL;
2341 free_iv:
2342         kfree(cctx->iv);
2343         cctx->iv = NULL;
2344 free_priv:
2345         if (tx) {
2346                 kfree(ctx->priv_ctx_tx);
2347                 ctx->priv_ctx_tx = NULL;
2348         } else {
2349                 kfree(ctx->priv_ctx_rx);
2350                 ctx->priv_ctx_rx = NULL;
2351         }
2352 out:
2353         return rc;
2354 }