net/tls/tls_sw.c

   1 /*
   2  * Copyright (c) 2016-2017, Mellanox Technologies. All rights reserved.
   3  * Copyright (c) 2016-2017, Dave Watson <davejwatson@fb.com>. All rights reserved.
   4  * Copyright (c) 2016-2017, Lance Chao <lancerchao@fb.com>. All rights reserved.
   5  * Copyright (c) 2016, Fridolin Pokorny <fridolin.pokorny@gmail.com>. All rights reserved.
   6  * Copyright (c) 2016, Nikos Mavrogiannopoulos <nmav@gnutls.org>. All rights reserved.
   7  * Copyright (c) 2018, Covalent IO, Inc. http://covalent.io
   8  *
   9  * This software is available to you under a choice of one of two
  10  * licenses.  You may choose to be licensed under the terms of the GNU
  11  * General Public License (GPL) Version 2, available from the file
  12  * COPYING in the main directory of this source tree, or the
  13  * OpenIB.org BSD license below:
  14  *
  15  *     Redistribution and use in source and binary forms, with or
  16  *     without modification, are permitted provided that the following
  17  *     conditions are met:
  18  *
  19  *      - Redistributions of source code must retain the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer.
  22  *
  23  *      - Redistributions in binary form must reproduce the above
  24  *        copyright notice, this list of conditions and the following
  25  *        disclaimer in the documentation and/or other materials
  26  *        provided with the distribution.
  27  *
  28  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  29  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  30  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  31  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  32  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  33  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  34  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  35  * SOFTWARE.
  36  */
  37
  38 #include <linux/sched/signal.h>
  39 #include <linux/module.h>
  40 #include <crypto/aead.h>
  41
  42 #include <net/strparser.h>
  43 #include <net/tls.h>
  44
  45 static int __skb_nsg(struct sk_buff *skb, int offset, int len,
  46                      unsigned int recursion_level)
  47 {
  48         int start = skb_headlen(skb);
  49         int i, chunk = start - offset;
  50         struct sk_buff *frag_iter;
  51         int elt = 0;
  52
  53         if (unlikely(recursion_level >= 24))
  54                 return -EMSGSIZE;
  55
  56         if (chunk > 0) {
  57                 if (chunk > len)
  58                         chunk = len;
  59                 elt++;
  60                 len -= chunk;
  61                 if (len == 0)
  62                         return elt;
  63                 offset += chunk;
  64         }
  65
  66         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
  67                 int end;
  68
  69                 WARN_ON(start > offset + len);
  70
  71                 end = start + skb_frag_size(&skb_shinfo(skb)->frags[i]);
  72                 chunk = end - offset;
  73                 if (chunk > 0) {
  74                         if (chunk > len)
  75                                 chunk = len;
  76                         elt++;
  77                         len -= chunk;
  78                         if (len == 0)
  79                                 return elt;
  80                         offset += chunk;
  81                 }
  82                 start = end;
  83         }
  84
  85         if (unlikely(skb_has_frag_list(skb))) {
  86                 skb_walk_frags(skb, frag_iter) {
  87                         int end, ret;
  88
  89                         WARN_ON(start > offset + len);
  90
  91                         end = start + frag_iter->len;
  92                         chunk = end - offset;
  93                         if (chunk > 0) {
  94                                 if (chunk > len)
  95                                         chunk = len;
  96                                 ret = __skb_nsg(frag_iter, offset - start, chunk,
  97                                                 recursion_level + 1);
  98                                 if (unlikely(ret < 0))
  99                                         return ret;
 100                                 elt += ret;
 101                                 len -= chunk;
 102                                 if (len == 0)
 103                                         return elt;
 104                                 offset += chunk;
 105                         }
 106                         start = end;
 107                 }
 108         }
 109         BUG_ON(len);
 110         return elt;
 111 }
 112
 113 /* Return the number of scatterlist elements required to completely map the
 114  * skb, or -EMSGSIZE if the recursion depth is exceeded.
 115  */
 116 static int skb_nsg(struct sk_buff *skb, int offset, int len)
 117 {
 118         return __skb_nsg(skb, offset, len, 0);
 119 }
 120
 121 static int padding_length(struct tls_sw_context_rx *ctx,
 122                           struct tls_prot_info *prot, struct sk_buff *skb)
 123 {
 124         struct strp_msg *rxm = strp_msg(skb);
 125         int sub = 0;
 126
 127         /* Determine zero-padding length */
 128         if (prot->version == TLS_1_3_VERSION) {
 129                 char content_type = 0;
 130                 int err;
 131                 int back = 17;
 132
 133                 while (content_type == 0) {
 134                         if (back > rxm->full_len - prot->prepend_size)
 135                                 return -EBADMSG;
 136                         err = skb_copy_bits(skb,
 137                                             rxm->offset + rxm->full_len - back,
 138                                             &content_type, 1);
 139                         if (err)
 140                                 return err;
 141                         if (content_type)
 142                                 break;
 143                         sub++;
 144                         back++;
 145                 }
 146                 ctx->control = content_type;
 147         }
 148         return sub;
 149 }
 150
 151 static void tls_decrypt_done(struct crypto_async_request *req, int err)
 152 {
 153         struct aead_request *aead_req = (struct aead_request *)req;
 154         struct scatterlist *sgout = aead_req->dst;
 155         struct scatterlist *sgin = aead_req->src;
 156         struct tls_sw_context_rx *ctx;
 157         struct tls_context *tls_ctx;
 158         struct tls_prot_info *prot;
 159         struct scatterlist *sg;
 160         struct sk_buff *skb;
 161         unsigned int pages;
 162         int pending;
 163
 164         skb = (struct sk_buff *)req->data;
 165         tls_ctx = tls_get_ctx(skb->sk);
 166         ctx = tls_sw_ctx_rx(tls_ctx);
 167         prot = &tls_ctx->prot_info;
 168
 169         /* Propagate if there was an err */
 170         if (err) {
 171                 ctx->async_wait.err = err;
 172                 tls_err_abort(skb->sk, err);
 173         } else {
 174                 struct strp_msg *rxm = strp_msg(skb);
 175                 int pad;
 176
 177                 pad = padding_length(ctx, prot, skb);
 178                 if (pad < 0) {
 179                         ctx->async_wait.err = pad;
 180                         tls_err_abort(skb->sk, pad);
 181                 } else {
 182                         rxm->full_len -= pad;
 183                         rxm->offset += prot->prepend_size;
 184                         rxm->full_len -= prot->overhead_size;
 185                 }
 186         }
 187
 188         /* After using skb->sk to propagate sk through crypto async callback
 189          * we need to NULL it again.
 190          */
 191         skb->sk = NULL;
 192
 193
 194         /* Free the destination pages if skb was not decrypted inplace */
 195         if (sgout != sgin) {
 196                 /* Skip the first S/G entry as it points to AAD */
 197                 for_each_sg(sg_next(sgout), sg, UINT_MAX, pages) {
 198                         if (!sg)
 199                                 break;
 200                         put_page(sg_page(sg));
 201                 }
 202         }
 203
 204         kfree(aead_req);
 205
 206         pending = atomic_dec_return(&ctx->decrypt_pending);
 207
 208         if (!pending && READ_ONCE(ctx->async_notify))
 209                 complete(&ctx->async_wait.completion);
 210 }
 211
 212 static int tls_do_decryption(struct sock *sk,
 213                              struct sk_buff *skb,
 214                              struct scatterlist *sgin,
 215                              struct scatterlist *sgout,
 216                              char *iv_recv,
 217                              size_t data_len,
 218                              struct aead_request *aead_req,
 219                              bool async)
 220 {
 221         struct tls_context *tls_ctx = tls_get_ctx(sk);
 222         struct tls_prot_info *prot = &tls_ctx->prot_info;
 223         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
 224         int ret;
 225
 226         aead_request_set_tfm(aead_req, ctx->aead_recv);
 227         aead_request_set_ad(aead_req, prot->aad_size);
 228         aead_request_set_crypt(aead_req, sgin, sgout,
 229                                data_len + prot->tag_size,
 230                                (u8 *)iv_recv);
 231
 232         if (async) {
 233                 /* Using skb->sk to push sk through to crypto async callback
 234                  * handler. This allows propagating errors up to the socket
 235                  * if needed. It _must_ be cleared in the async handler
 236                  * before consume_skb is called. We _know_ skb->sk is NULL
 237                  * because it is a clone from strparser.
 238                  */
 239                 skb->sk = sk;
 240                 aead_request_set_callback(aead_req,
 241                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 242                                           tls_decrypt_done, skb);
 243                 atomic_inc(&ctx->decrypt_pending);
 244         } else {
 245                 aead_request_set_callback(aead_req,
 246                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 247                                           crypto_req_done, &ctx->async_wait);
 248         }
 249
 250         ret = crypto_aead_decrypt(aead_req);
 251         if (ret == -EINPROGRESS) {
 252                 if (async)
 253                         return ret;
 254
 255                 ret = crypto_wait_req(ret, &ctx->async_wait);
 256         }
 257
 258         if (async)
 259                 atomic_dec(&ctx->decrypt_pending);
 260
 261         return ret;
 262 }
 263
 264 static void tls_trim_both_msgs(struct sock *sk, int target_size)
 265 {
 266         struct tls_context *tls_ctx = tls_get_ctx(sk);
 267         struct tls_prot_info *prot = &tls_ctx->prot_info;
 268         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 269         struct tls_rec *rec = ctx->open_rec;
 270
 271         sk_msg_trim(sk, &rec->msg_plaintext, target_size);
 272         if (target_size > 0)
 273                 target_size += prot->overhead_size;
 274         sk_msg_trim(sk, &rec->msg_encrypted, target_size);
 275 }
 276
 277 static int tls_alloc_encrypted_msg(struct sock *sk, int len)
 278 {
 279         struct tls_context *tls_ctx = tls_get_ctx(sk);
 280         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 281         struct tls_rec *rec = ctx->open_rec;
 282         struct sk_msg *msg_en = &rec->msg_encrypted;
 283
 284         return sk_msg_alloc(sk, msg_en, len, 0);
 285 }
 286
 287 static int tls_clone_plaintext_msg(struct sock *sk, int required)
 288 {
 289         struct tls_context *tls_ctx = tls_get_ctx(sk);
 290         struct tls_prot_info *prot = &tls_ctx->prot_info;
 291         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 292         struct tls_rec *rec = ctx->open_rec;
 293         struct sk_msg *msg_pl = &rec->msg_plaintext;
 294         struct sk_msg *msg_en = &rec->msg_encrypted;
 295         int skip, len;
 296
 297         /* We add page references worth len bytes from encrypted sg
 298          * at the end of plaintext sg. It is guaranteed that msg_en
 299          * has enough required room (ensured by caller).
 300          */
 301         len = required - msg_pl->sg.size;
 302
 303         /* Skip initial bytes in msg_en's data to be able to use
 304          * same offset of both plain and encrypted data.
 305          */
 306         skip = prot->prepend_size + msg_pl->sg.size;
 307
 308         return sk_msg_clone(sk, msg_pl, msg_en, skip, len);
 309 }
 310
 311 static struct tls_rec *tls_get_rec(struct sock *sk)
 312 {
 313         struct tls_context *tls_ctx = tls_get_ctx(sk);
 314         struct tls_prot_info *prot = &tls_ctx->prot_info;
 315         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 316         struct sk_msg *msg_pl, *msg_en;
 317         struct tls_rec *rec;
 318         int mem_size;
 319
 320         mem_size = sizeof(struct tls_rec) + crypto_aead_reqsize(ctx->aead_send);
 321
 322         rec = kzalloc(mem_size, sk->sk_allocation);
 323         if (!rec)
 324                 return NULL;
 325
 326         msg_pl = &rec->msg_plaintext;
 327         msg_en = &rec->msg_encrypted;
 328
 329         sk_msg_init(msg_pl);
 330         sk_msg_init(msg_en);
 331
 332         sg_init_table(rec->sg_aead_in, 2);
 333         sg_set_buf(&rec->sg_aead_in[0], rec->aad_space, prot->aad_size);
 334         sg_unmark_end(&rec->sg_aead_in[1]);
 335
 336         sg_init_table(rec->sg_aead_out, 2);
 337         sg_set_buf(&rec->sg_aead_out[0], rec->aad_space, prot->aad_size);
 338         sg_unmark_end(&rec->sg_aead_out[1]);
 339
 340         return rec;
 341 }
 342
 343 static void tls_free_rec(struct sock *sk, struct tls_rec *rec)
 344 {
 345         sk_msg_free(sk, &rec->msg_encrypted);
 346         sk_msg_free(sk, &rec->msg_plaintext);
 347         kfree(rec);
 348 }
 349
 350 static void tls_free_open_rec(struct sock *sk)
 351 {
 352         struct tls_context *tls_ctx = tls_get_ctx(sk);
 353         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 354         struct tls_rec *rec = ctx->open_rec;
 355
 356         if (rec) {
 357                 tls_free_rec(sk, rec);
 358                 ctx->open_rec = NULL;
 359         }
 360 }
 361
 362 int tls_tx_records(struct sock *sk, int flags)
 363 {
 364         struct tls_context *tls_ctx = tls_get_ctx(sk);
 365         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 366         struct tls_rec *rec, *tmp;
 367         struct sk_msg *msg_en;
 368         int tx_flags, rc = 0;
 369
 370         if (tls_is_partially_sent_record(tls_ctx)) {
 371                 rec = list_first_entry(&ctx->tx_list,
 372                                        struct tls_rec, list);
 373
 374                 if (flags == -1)
 375                         tx_flags = rec->tx_flags;
 376                 else
 377                         tx_flags = flags;
 378
 379                 rc = tls_push_partial_record(sk, tls_ctx, tx_flags);
 380                 if (rc)
 381                         goto tx_err;
 382
 383                 /* Full record has been transmitted.
 384                  * Remove the head of tx_list
 385                  */
 386                 list_del(&rec->list);
 387                 sk_msg_free(sk, &rec->msg_plaintext);
 388                 kfree(rec);
 389         }
 390
 391         /* Tx all ready records */
 392         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
 393                 if (READ_ONCE(rec->tx_ready)) {
 394                         if (flags == -1)
 395                                 tx_flags = rec->tx_flags;
 396                         else
 397                                 tx_flags = flags;
 398
 399                         msg_en = &rec->msg_encrypted;
 400                         rc = tls_push_sg(sk, tls_ctx,
 401                                          &msg_en->sg.data[msg_en->sg.curr],
 402                                          0, tx_flags);
 403                         if (rc)
 404                                 goto tx_err;
 405
 406                         list_del(&rec->list);
 407                         sk_msg_free(sk, &rec->msg_plaintext);
 408                         kfree(rec);
 409                 } else {
 410                         break;
 411                 }
 412         }
 413
 414 tx_err:
 415         if (rc < 0 && rc != -EAGAIN)
 416                 tls_err_abort(sk, EBADMSG);
 417
 418         return rc;
 419 }
 420
 421 static void tls_encrypt_done(struct crypto_async_request *req, int err)
 422 {
 423         struct aead_request *aead_req = (struct aead_request *)req;
 424         struct sock *sk = req->data;
 425         struct tls_context *tls_ctx = tls_get_ctx(sk);
 426         struct tls_prot_info *prot = &tls_ctx->prot_info;
 427         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 428         struct scatterlist *sge;
 429         struct sk_msg *msg_en;
 430         struct tls_rec *rec;
 431         bool ready = false;
 432         int pending;
 433
 434         rec = container_of(aead_req, struct tls_rec, aead_req);
 435         msg_en = &rec->msg_encrypted;
 436
 437         sge = sk_msg_elem(msg_en, msg_en->sg.curr);
 438         sge->offset -= prot->prepend_size;
 439         sge->length += prot->prepend_size;
 440
 441         /* Check if error is previously set on socket */
 442         if (err || sk->sk_err) {
 443                 rec = NULL;
 444
 445                 /* If err is already set on socket, return the same code */
 446                 if (sk->sk_err) {
 447                         ctx->async_wait.err = sk->sk_err;
 448                 } else {
 449                         ctx->async_wait.err = err;
 450                         tls_err_abort(sk, err);
 451                 }
 452         }
 453
 454         if (rec) {
 455                 struct tls_rec *first_rec;
 456
 457                 /* Mark the record as ready for transmission */
 458                 smp_store_mb(rec->tx_ready, true);
 459
 460                 /* If received record is at head of tx_list, schedule tx */
 461                 first_rec = list_first_entry(&ctx->tx_list,
 462                                              struct tls_rec, list);
 463                 if (rec == first_rec)
 464                         ready = true;
 465         }
 466
 467         pending = atomic_dec_return(&ctx->encrypt_pending);
 468
 469         if (!pending && READ_ONCE(ctx->async_notify))
 470                 complete(&ctx->async_wait.completion);
 471
 472         if (!ready)
 473                 return;
 474
 475         /* Schedule the transmission */
 476         if (!test_and_set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
 477                 schedule_delayed_work(&ctx->tx_work.work, 1);
 478 }
 479
 480 static int tls_do_encryption(struct sock *sk,
 481                              struct tls_context *tls_ctx,
 482                              struct tls_sw_context_tx *ctx,
 483                              struct aead_request *aead_req,
 484                              size_t data_len, u32 start)
 485 {
 486         struct tls_prot_info *prot = &tls_ctx->prot_info;
 487         struct tls_rec *rec = ctx->open_rec;
 488         struct sk_msg *msg_en = &rec->msg_encrypted;
 489         struct scatterlist *sge = sk_msg_elem(msg_en, start);
 490         int rc, iv_offset = 0;
 491
 492         /* For CCM based ciphers, first byte of IV is a constant */
 493         if (prot->cipher_type == TLS_CIPHER_AES_CCM_128) {
 494                 rec->iv_data[0] = TLS_AES_CCM_IV_B0_BYTE;
 495                 iv_offset = 1;
 496         }
 497
 498         memcpy(&rec->iv_data[iv_offset], tls_ctx->tx.iv,
 499                prot->iv_size + prot->salt_size);
 500
 501         xor_iv_with_seq(prot->version, rec->iv_data, tls_ctx->tx.rec_seq);
 502
 503         sge->offset += prot->prepend_size;
 504         sge->length -= prot->prepend_size;
 505
 506         msg_en->sg.curr = start;
 507
 508         aead_request_set_tfm(aead_req, ctx->aead_send);
 509         aead_request_set_ad(aead_req, prot->aad_size);
 510         aead_request_set_crypt(aead_req, rec->sg_aead_in,
 511                                rec->sg_aead_out,
 512                                data_len, rec->iv_data);
 513
 514         aead_request_set_callback(aead_req, CRYPTO_TFM_REQ_MAY_BACKLOG,
 515                                   tls_encrypt_done, sk);
 516
 517         /* Add the record in tx_list */
 518         list_add_tail((struct list_head *)&rec->list, &ctx->tx_list);
 519         atomic_inc(&ctx->encrypt_pending);
 520
 521         rc = crypto_aead_encrypt(aead_req);
 522         if (!rc || rc != -EINPROGRESS) {
 523                 atomic_dec(&ctx->encrypt_pending);
 524                 sge->offset -= prot->prepend_size;
 525                 sge->length += prot->prepend_size;
 526         }
 527
 528         if (!rc) {
 529                 WRITE_ONCE(rec->tx_ready, true);
 530         } else if (rc != -EINPROGRESS) {
 531                 list_del(&rec->list);
 532                 return rc;
 533         }
 534
 535         /* Unhook the record from context if encryption is not failure */
 536         ctx->open_rec = NULL;
 537         tls_advance_record_sn(sk, prot, &tls_ctx->tx);
 538         return rc;
 539 }
 540
 541 static int tls_split_open_record(struct sock *sk, struct tls_rec *from,
 542                                  struct tls_rec **to, struct sk_msg *msg_opl,
 543                                  struct sk_msg *msg_oen, u32 split_point,
 544                                  u32 tx_overhead_size, u32 *orig_end)
 545 {
 546         u32 i, j, bytes = 0, apply = msg_opl->apply_bytes;
 547         struct scatterlist *sge, *osge, *nsge;
 548         u32 orig_size = msg_opl->sg.size;
 549         struct scatterlist tmp = { };
 550         struct sk_msg *msg_npl;
 551         struct tls_rec *new;
 552         int ret;
 553
 554         new = tls_get_rec(sk);
 555         if (!new)
 556                 return -ENOMEM;
 557         ret = sk_msg_alloc(sk, &new->msg_encrypted, msg_opl->sg.size +
 558                            tx_overhead_size, 0);
 559         if (ret < 0) {
 560                 tls_free_rec(sk, new);
 561                 return ret;
 562         }
 563
 564         *orig_end = msg_opl->sg.end;
 565         i = msg_opl->sg.start;
 566         sge = sk_msg_elem(msg_opl, i);
 567         while (apply && sge->length) {
 568                 if (sge->length > apply) {
 569                         u32 len = sge->length - apply;
 570
 571                         get_page(sg_page(sge));
 572                         sg_set_page(&tmp, sg_page(sge), len,
 573                                     sge->offset + apply);
 574                         sge->length = apply;
 575                         bytes += apply;
 576                         apply = 0;
 577                 } else {
 578                         apply -= sge->length;
 579                         bytes += sge->length;
 580                 }
 581
 582                 sk_msg_iter_var_next(i);
 583                 if (i == msg_opl->sg.end)
 584                         break;
 585                 sge = sk_msg_elem(msg_opl, i);
 586         }
 587
 588         msg_opl->sg.end = i;
 589         msg_opl->sg.curr = i;
 590         msg_opl->sg.copybreak = 0;
 591         msg_opl->apply_bytes = 0;
 592         msg_opl->sg.size = bytes;
 593
 594         msg_npl = &new->msg_plaintext;
 595         msg_npl->apply_bytes = apply;
 596         msg_npl->sg.size = orig_size - bytes;
 597
 598         j = msg_npl->sg.start;
 599         nsge = sk_msg_elem(msg_npl, j);
 600         if (tmp.length) {
 601                 memcpy(nsge, &tmp, sizeof(*nsge));
 602                 sk_msg_iter_var_next(j);
 603                 nsge = sk_msg_elem(msg_npl, j);
 604         }
 605
 606         osge = sk_msg_elem(msg_opl, i);
 607         while (osge->length) {
 608                 memcpy(nsge, osge, sizeof(*nsge));
 609                 sg_unmark_end(nsge);
 610                 sk_msg_iter_var_next(i);
 611                 sk_msg_iter_var_next(j);
 612                 if (i == *orig_end)
 613                         break;
 614                 osge = sk_msg_elem(msg_opl, i);
 615                 nsge = sk_msg_elem(msg_npl, j);
 616         }
 617
 618         msg_npl->sg.end = j;
 619         msg_npl->sg.curr = j;
 620         msg_npl->sg.copybreak = 0;
 621
 622         *to = new;
 623         return 0;
 624 }
 625
 626 static void tls_merge_open_record(struct sock *sk, struct tls_rec *to,
 627                                   struct tls_rec *from, u32 orig_end)
 628 {
 629         struct sk_msg *msg_npl = &from->msg_plaintext;
 630         struct sk_msg *msg_opl = &to->msg_plaintext;
 631         struct scatterlist *osge, *nsge;
 632         u32 i, j;
 633
 634         i = msg_opl->sg.end;
 635         sk_msg_iter_var_prev(i);
 636         j = msg_npl->sg.start;
 637
 638         osge = sk_msg_elem(msg_opl, i);
 639         nsge = sk_msg_elem(msg_npl, j);
 640
 641         if (sg_page(osge) == sg_page(nsge) &&
 642             osge->offset + osge->length == nsge->offset) {
 643                 osge->length += nsge->length;
 644                 put_page(sg_page(nsge));
 645         }
 646
 647         msg_opl->sg.end = orig_end;
 648         msg_opl->sg.curr = orig_end;
 649         msg_opl->sg.copybreak = 0;
 650         msg_opl->apply_bytes = msg_opl->sg.size + msg_npl->sg.size;
 651         msg_opl->sg.size += msg_npl->sg.size;
 652
 653         sk_msg_free(sk, &to->msg_encrypted);
 654         sk_msg_xfer_full(&to->msg_encrypted, &from->msg_encrypted);
 655
 656         kfree(from);
 657 }
 658
 659 static int tls_push_record(struct sock *sk, int flags,
 660                            unsigned char record_type)
 661 {
 662         struct tls_context *tls_ctx = tls_get_ctx(sk);
 663         struct tls_prot_info *prot = &tls_ctx->prot_info;
 664         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 665         struct tls_rec *rec = ctx->open_rec, *tmp = NULL;
 666         u32 i, split_point, uninitialized_var(orig_end);
 667         struct sk_msg *msg_pl, *msg_en;
 668         struct aead_request *req;
 669         bool split;
 670         int rc;
 671
 672         if (!rec)
 673                 return 0;
 674
 675         msg_pl = &rec->msg_plaintext;
 676         msg_en = &rec->msg_encrypted;
 677
 678         split_point = msg_pl->apply_bytes;
 679         split = split_point && split_point < msg_pl->sg.size;
 680         if (split) {
 681                 rc = tls_split_open_record(sk, rec, &tmp, msg_pl, msg_en,
 682                                            split_point, prot->overhead_size,
 683                                            &orig_end);
 684                 if (rc < 0)
 685                         return rc;
 686                 sk_msg_trim(sk, msg_en, msg_pl->sg.size +
 687                             prot->overhead_size);
 688         }
 689
 690         rec->tx_flags = flags;
 691         req = &rec->aead_req;
 692
 693         i = msg_pl->sg.end;
 694         sk_msg_iter_var_prev(i);
 695
 696         rec->content_type = record_type;
 697         if (prot->version == TLS_1_3_VERSION) {
 698                 /* Add content type to end of message.  No padding added */
 699                 sg_set_buf(&rec->sg_content_type, &rec->content_type, 1);
 700                 sg_mark_end(&rec->sg_content_type);
 701                 sg_chain(msg_pl->sg.data, msg_pl->sg.end + 1,
 702                          &rec->sg_content_type);
 703         } else {
 704                 sg_mark_end(sk_msg_elem(msg_pl, i));
 705         }
 706
 707         i = msg_pl->sg.start;
 708         sg_chain(rec->sg_aead_in, 2, rec->inplace_crypto ?
 709                  &msg_en->sg.data[i] : &msg_pl->sg.data[i]);
 710
 711         i = msg_en->sg.end;
 712         sk_msg_iter_var_prev(i);
 713         sg_mark_end(sk_msg_elem(msg_en, i));
 714
 715         i = msg_en->sg.start;
 716         sg_chain(rec->sg_aead_out, 2, &msg_en->sg.data[i]);
 717
 718         tls_make_aad(rec->aad_space, msg_pl->sg.size + prot->tail_size,
 719                      tls_ctx->tx.rec_seq, prot->rec_seq_size,
 720                      record_type, prot->version);
 721
 722         tls_fill_prepend(tls_ctx,
 723                          page_address(sg_page(&msg_en->sg.data[i])) +
 724                          msg_en->sg.data[i].offset,
 725                          msg_pl->sg.size + prot->tail_size,
 726                          record_type, prot->version);
 727
 728         tls_ctx->pending_open_record_frags = false;
 729
 730         rc = tls_do_encryption(sk, tls_ctx, ctx, req,
 731                                msg_pl->sg.size + prot->tail_size, i);
 732         if (rc < 0) {
 733                 if (rc != -EINPROGRESS) {
 734                         tls_err_abort(sk, EBADMSG);
 735                         if (split) {
 736                                 tls_ctx->pending_open_record_frags = true;
 737                                 tls_merge_open_record(sk, rec, tmp, orig_end);
 738                         }
 739                 }
 740                 ctx->async_capable = 1;
 741                 return rc;
 742         } else if (split) {
 743                 msg_pl = &tmp->msg_plaintext;
 744                 msg_en = &tmp->msg_encrypted;
 745                 sk_msg_trim(sk, msg_en, msg_pl->sg.size + prot->overhead_size);
 746                 tls_ctx->pending_open_record_frags = true;
 747                 ctx->open_rec = tmp;
 748         }
 749
 750         return tls_tx_records(sk, flags);
 751 }
 752
 753 static int bpf_exec_tx_verdict(struct sk_msg *msg, struct sock *sk,
 754                                bool full_record, u8 record_type,
 755                                size_t *copied, int flags)
 756 {
 757         struct tls_context *tls_ctx = tls_get_ctx(sk);
 758         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 759         struct sk_msg msg_redir = { };
 760         struct sk_psock *psock;
 761         struct sock *sk_redir;
 762         struct tls_rec *rec;
 763         bool enospc, policy;
 764         int err = 0, send;
 765         u32 delta = 0;
 766
 767         policy = !(flags & MSG_SENDPAGE_NOPOLICY);
 768         psock = sk_psock_get(sk);
 769         if (!psock || !policy)
 770                 return tls_push_record(sk, flags, record_type);
 771 more_data:
 772         enospc = sk_msg_full(msg);
 773         if (psock->eval == __SK_NONE) {
 774                 delta = msg->sg.size;
 775                 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
 776                 if (delta < msg->sg.size)
 777                         delta -= msg->sg.size;
 778                 else
 779                         delta = 0;
 780         }
 781         if (msg->cork_bytes && msg->cork_bytes > msg->sg.size &&
 782             !enospc && !full_record) {
 783                 err = -ENOSPC;
 784                 goto out_err;
 785         }
 786         msg->cork_bytes = 0;
 787         send = msg->sg.size;
 788         if (msg->apply_bytes && msg->apply_bytes < send)
 789                 send = msg->apply_bytes;
 790
 791         switch (psock->eval) {
 792         case __SK_PASS:
 793                 err = tls_push_record(sk, flags, record_type);
 794                 if (err < 0) {
 795                         *copied -= sk_msg_free(sk, msg);
 796                         tls_free_open_rec(sk);
 797                         goto out_err;
 798                 }
 799                 break;
 800         case __SK_REDIRECT:
 801                 sk_redir = psock->sk_redir;
 802                 memcpy(&msg_redir, msg, sizeof(*msg));
 803                 if (msg->apply_bytes < send)
 804                         msg->apply_bytes = 0;
 805                 else
 806                         msg->apply_bytes -= send;
 807                 sk_msg_return_zero(sk, msg, send);
 808                 msg->sg.size -= send;
 809                 release_sock(sk);
 810                 err = tcp_bpf_sendmsg_redir(sk_redir, &msg_redir, send, flags);
 811                 lock_sock(sk);
 812                 if (err < 0) {
 813                         *copied -= sk_msg_free_nocharge(sk, &msg_redir);
 814                         msg->sg.size = 0;
 815                 }
 816                 if (msg->sg.size == 0)
 817                         tls_free_open_rec(sk);
 818                 break;
 819         case __SK_DROP:
 820         default:
 821                 sk_msg_free_partial(sk, msg, send);
 822                 if (msg->apply_bytes < send)
 823                         msg->apply_bytes = 0;
 824                 else
 825                         msg->apply_bytes -= send;
 826                 if (msg->sg.size == 0)
 827                         tls_free_open_rec(sk);
 828                 *copied -= (send + delta);
 829                 err = -EACCES;
 830         }
 831
 832         if (likely(!err)) {
 833                 bool reset_eval = !ctx->open_rec;
 834
 835                 rec = ctx->open_rec;
 836                 if (rec) {
 837                         msg = &rec->msg_plaintext;
 838                         if (!msg->apply_bytes)
 839                                 reset_eval = true;
 840                 }
 841                 if (reset_eval) {
 842                         psock->eval = __SK_NONE;
 843                         if (psock->sk_redir) {
 844                                 sock_put(psock->sk_redir);
 845                                 psock->sk_redir = NULL;
 846                         }
 847                 }
 848                 if (rec)
 849                         goto more_data;
 850         }
 851  out_err:
 852         sk_psock_put(sk, psock);
 853         return err;
 854 }
 855
 856 static int tls_sw_push_pending_record(struct sock *sk, int flags)
 857 {
 858         struct tls_context *tls_ctx = tls_get_ctx(sk);
 859         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 860         struct tls_rec *rec = ctx->open_rec;
 861         struct sk_msg *msg_pl;
 862         size_t copied;
 863
 864         if (!rec)
 865                 return 0;
 866
 867         msg_pl = &rec->msg_plaintext;
 868         copied = msg_pl->sg.size;
 869         if (!copied)
 870                 return 0;
 871
 872         return bpf_exec_tx_verdict(msg_pl, sk, true, TLS_RECORD_TYPE_DATA,
 873                                    &copied, flags);
 874 }
 875
 876 int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 877 {
 878         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 879         struct tls_context *tls_ctx = tls_get_ctx(sk);
 880         struct tls_prot_info *prot = &tls_ctx->prot_info;
 881         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 882         bool async_capable = ctx->async_capable;
 883         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 884         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
 885         bool eor = !(msg->msg_flags & MSG_MORE);
 886         size_t try_to_copy, copied = 0;
 887         struct sk_msg *msg_pl, *msg_en;
 888         struct tls_rec *rec;
 889         int required_size;
 890         int num_async = 0;
 891         bool full_record;
 892         int record_room;
 893         int num_zc = 0;
 894         int orig_size;
 895         int ret = 0;
 896
 897         if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
 898                 return -ENOTSUPP;
 899
 900         mutex_lock(&tls_ctx->tx_lock);
 901         lock_sock(sk);
 902
 903         if (unlikely(msg->msg_controllen)) {
 904                 ret = tls_proccess_cmsg(sk, msg, &record_type);
 905                 if (ret) {
 906                         if (ret == -EINPROGRESS)
 907                                 num_async++;
 908                         else if (ret != -EAGAIN)
 909                                 goto send_end;
 910                 }
 911         }
 912
 913         while (msg_data_left(msg)) {
 914                 if (sk->sk_err) {
 915                         ret = -sk->sk_err;
 916                         goto send_end;
 917                 }
 918
 919                 if (ctx->open_rec)
 920                         rec = ctx->open_rec;
 921                 else
 922                         rec = ctx->open_rec = tls_get_rec(sk);
 923                 if (!rec) {
 924                         ret = -ENOMEM;
 925                         goto send_end;
 926                 }
 927
 928                 msg_pl = &rec->msg_plaintext;
 929                 msg_en = &rec->msg_encrypted;
 930
 931                 orig_size = msg_pl->sg.size;
 932                 full_record = false;
 933                 try_to_copy = msg_data_left(msg);
 934                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
 935                 if (try_to_copy >= record_room) {
 936                         try_to_copy = record_room;
 937                         full_record = true;
 938                 }
 939
 940                 required_size = msg_pl->sg.size + try_to_copy +
 941                                 prot->overhead_size;
 942
 943                 if (!sk_stream_memory_free(sk))
 944                         goto wait_for_sndbuf;
 945
 946 alloc_encrypted:
 947                 ret = tls_alloc_encrypted_msg(sk, required_size);
 948                 if (ret) {
 949                         if (ret != -ENOSPC)
 950                                 goto wait_for_memory;
 951
 952                         /* Adjust try_to_copy according to the amount that was
 953                          * actually allocated. The difference is due
 954                          * to max sg elements limit
 955                          */
 956                         try_to_copy -= required_size - msg_en->sg.size;
 957                         full_record = true;
 958                 }
 959
 960                 if (!is_kvec && (full_record || eor) && !async_capable) {
 961                         u32 first = msg_pl->sg.end;
 962
 963                         ret = sk_msg_zerocopy_from_iter(sk, &msg->msg_iter,
 964                                                         msg_pl, try_to_copy);
 965                         if (ret)
 966                                 goto fallback_to_reg_send;
 967
 968                         rec->inplace_crypto = 0;
 969
 970                         num_zc++;
 971                         copied += try_to_copy;
 972
 973                         sk_msg_sg_copy_set(msg_pl, first);
 974                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
 975                                                   record_type, &copied,
 976                                                   msg->msg_flags);
 977                         if (ret) {
 978                                 if (ret == -EINPROGRESS)
 979                                         num_async++;
 980                                 else if (ret == -ENOMEM)
 981                                         goto wait_for_memory;
 982                                 else if (ret == -ENOSPC)
 983                                         goto rollback_iter;
 984                                 else if (ret != -EAGAIN)
 985                                         goto send_end;
 986                         }
 987                         continue;
 988 rollback_iter:
 989                         copied -= try_to_copy;
 990                         sk_msg_sg_copy_clear(msg_pl, first);
 991                         iov_iter_revert(&msg->msg_iter,
 992                                         msg_pl->sg.size - orig_size);
 993 fallback_to_reg_send:
 994                         sk_msg_trim(sk, msg_pl, orig_size);
 995                 }
 996
 997                 required_size = msg_pl->sg.size + try_to_copy;
 998
 999                 ret = tls_clone_plaintext_msg(sk, required_size);
1000                 if (ret) {
1001                         if (ret != -ENOSPC)
1002                                 goto send_end;
1003
1004                         /* Adjust try_to_copy according to the amount that was
1005                          * actually allocated. The difference is due
1006                          * to max sg elements limit
1007                          */
1008                         try_to_copy -= required_size - msg_pl->sg.size;
1009                         full_record = true;
1010                         sk_msg_trim(sk, msg_en,
1011                                     msg_pl->sg.size + prot->overhead_size);
1012                 }
1013
1014                 if (try_to_copy) {
1015                         ret = sk_msg_memcopy_from_iter(sk, &msg->msg_iter,
1016                                                        msg_pl, try_to_copy);
1017                         if (ret < 0)
1018                                 goto trim_sgl;
1019                 }
1020
1021                 /* Open records defined only if successfully copied, otherwise
1022                  * we would trim the sg but not reset the open record frags.
1023                  */
1024                 tls_ctx->pending_open_record_frags = true;
1025                 copied += try_to_copy;
1026                 if (full_record || eor) {
1027                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1028                                                   record_type, &copied,
1029                                                   msg->msg_flags);
1030                         if (ret) {
1031                                 if (ret == -EINPROGRESS)
1032                                         num_async++;
1033                                 else if (ret == -ENOMEM)
1034                                         goto wait_for_memory;
1035                                 else if (ret != -EAGAIN) {
1036                                         if (ret == -ENOSPC)
1037                                                 ret = 0;
1038                                         goto send_end;
1039                                 }
1040                         }
1041                 }
1042
1043                 continue;
1044
1045 wait_for_sndbuf:
1046                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1047 wait_for_memory:
1048                 ret = sk_stream_wait_memory(sk, &timeo);
1049                 if (ret) {
1050 trim_sgl:
1051                         tls_trim_both_msgs(sk, orig_size);
1052                         goto send_end;
1053                 }
1054
1055                 if (msg_en->sg.size < required_size)
1056                         goto alloc_encrypted;
1057         }
1058
1059         if (!num_async) {
1060                 goto send_end;
1061         } else if (num_zc) {
1062                 /* Wait for pending encryptions to get completed */
1063                 smp_store_mb(ctx->async_notify, true);
1064
1065                 if (atomic_read(&ctx->encrypt_pending))
1066                         crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1067                 else
1068                         reinit_completion(&ctx->async_wait.completion);
1069
1070                 WRITE_ONCE(ctx->async_notify, false);
1071
1072                 if (ctx->async_wait.err) {
1073                         ret = ctx->async_wait.err;
1074                         copied = 0;
1075                 }
1076         }
1077
1078         /* Transmit if any encryptions have completed */
1079         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1080                 cancel_delayed_work(&ctx->tx_work.work);
1081                 tls_tx_records(sk, msg->msg_flags);
1082         }
1083
1084 send_end:
1085         ret = sk_stream_error(sk, msg->msg_flags, ret);
1086
1087         release_sock(sk);
1088         mutex_unlock(&tls_ctx->tx_lock);
1089         return copied ? copied : ret;
1090 }
1091
1092 static int tls_sw_do_sendpage(struct sock *sk, struct page *page,
1093                               int offset, size_t size, int flags)
1094 {
1095         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1096         struct tls_context *tls_ctx = tls_get_ctx(sk);
1097         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1098         struct tls_prot_info *prot = &tls_ctx->prot_info;
1099         unsigned char record_type = TLS_RECORD_TYPE_DATA;
1100         struct sk_msg *msg_pl;
1101         struct tls_rec *rec;
1102         int num_async = 0;
1103         size_t copied = 0;
1104         bool full_record;
1105         int record_room;
1106         int ret = 0;
1107         bool eor;
1108
1109         eor = !(flags & (MSG_MORE | MSG_SENDPAGE_NOTLAST));
1110         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1111
1112         /* Call the sk_stream functions to manage the sndbuf mem. */
1113         while (size > 0) {
1114                 size_t copy, required_size;
1115
1116                 if (sk->sk_err) {
1117                         ret = -sk->sk_err;
1118                         goto sendpage_end;
1119                 }
1120
1121                 if (ctx->open_rec)
1122                         rec = ctx->open_rec;
1123                 else
1124                         rec = ctx->open_rec = tls_get_rec(sk);
1125                 if (!rec) {
1126                         ret = -ENOMEM;
1127                         goto sendpage_end;
1128                 }
1129
1130                 msg_pl = &rec->msg_plaintext;
1131
1132                 full_record = false;
1133                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
1134                 copy = size;
1135                 if (copy >= record_room) {
1136                         copy = record_room;
1137                         full_record = true;
1138                 }
1139
1140                 required_size = msg_pl->sg.size + copy + prot->overhead_size;
1141
1142                 if (!sk_stream_memory_free(sk))
1143                         goto wait_for_sndbuf;
1144 alloc_payload:
1145                 ret = tls_alloc_encrypted_msg(sk, required_size);
1146                 if (ret) {
1147                         if (ret != -ENOSPC)
1148                                 goto wait_for_memory;
1149
1150                         /* Adjust copy according to the amount that was
1151                          * actually allocated. The difference is due
1152                          * to max sg elements limit
1153                          */
1154                         copy -= required_size - msg_pl->sg.size;
1155                         full_record = true;
1156                 }
1157
1158                 sk_msg_page_add(msg_pl, page, copy, offset);
1159                 sk_mem_charge(sk, copy);
1160
1161                 offset += copy;
1162                 size -= copy;
1163                 copied += copy;
1164
1165                 tls_ctx->pending_open_record_frags = true;
1166                 if (full_record || eor || sk_msg_full(msg_pl)) {
1167                         rec->inplace_crypto = 0;
1168                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1169                                                   record_type, &copied, flags);
1170                         if (ret) {
1171                                 if (ret == -EINPROGRESS)
1172                                         num_async++;
1173                                 else if (ret == -ENOMEM)
1174                                         goto wait_for_memory;
1175                                 else if (ret != -EAGAIN) {
1176                                         if (ret == -ENOSPC)
1177                                                 ret = 0;
1178                                         goto sendpage_end;
1179                                 }
1180                         }
1181                 }
1182                 continue;
1183 wait_for_sndbuf:
1184                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1185 wait_for_memory:
1186                 ret = sk_stream_wait_memory(sk, &timeo);
1187                 if (ret) {
1188                         tls_trim_both_msgs(sk, msg_pl->sg.size);
1189                         goto sendpage_end;
1190                 }
1191
1192                 goto alloc_payload;
1193         }
1194
1195         if (num_async) {
1196                 /* Transmit if any encryptions have completed */
1197                 if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1198                         cancel_delayed_work(&ctx->tx_work.work);
1199                         tls_tx_records(sk, flags);
1200                 }
1201         }
1202 sendpage_end:
1203         ret = sk_stream_error(sk, flags, ret);
1204         return copied ? copied : ret;
1205 }
1206
1207 int tls_sw_sendpage_locked(struct sock *sk, struct page *page,
1208                            int offset, size_t size, int flags)
1209 {
1210         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1211                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY |
1212                       MSG_NO_SHARED_FRAGS))
1213                 return -ENOTSUPP;
1214
1215         return tls_sw_do_sendpage(sk, page, offset, size, flags);
1216 }
1217
1218 int tls_sw_sendpage(struct sock *sk, struct page *page,
1219                     int offset, size_t size, int flags)
1220 {
1221         struct tls_context *tls_ctx = tls_get_ctx(sk);
1222         int ret;
1223
1224         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1225                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY))
1226                 return -ENOTSUPP;
1227
1228         mutex_lock(&tls_ctx->tx_lock);
1229         lock_sock(sk);
1230         ret = tls_sw_do_sendpage(sk, page, offset, size, flags);
1231         release_sock(sk);
1232         mutex_unlock(&tls_ctx->tx_lock);
1233         return ret;
1234 }
1235
1236 static struct sk_buff *tls_wait_data(struct sock *sk, struct sk_psock *psock,
1237                                      int flags, long timeo, int *err)
1238 {
1239         struct tls_context *tls_ctx = tls_get_ctx(sk);
1240         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1241         struct sk_buff *skb;
1242         DEFINE_WAIT_FUNC(wait, woken_wake_function);
1243
1244         while (!(skb = ctx->recv_pkt) && sk_psock_queue_empty(psock)) {
1245                 if (sk->sk_err) {
1246                         *err = sock_error(sk);
1247                         return NULL;
1248                 }
1249
1250                 if (sk->sk_shutdown & RCV_SHUTDOWN)
1251                         return NULL;
1252
1253                 if (sock_flag(sk, SOCK_DONE))
1254                         return NULL;
1255
1256                 if ((flags & MSG_DONTWAIT) || !timeo) {
1257                         *err = -EAGAIN;
1258                         return NULL;
1259                 }
1260
1261                 add_wait_queue(sk_sleep(sk), &wait);
1262                 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1263                 sk_wait_event(sk, &timeo,
1264                               ctx->recv_pkt != skb ||
1265                               !sk_psock_queue_empty(psock),
1266                               &wait);
1267                 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1268                 remove_wait_queue(sk_sleep(sk), &wait);
1269
1270                 /* Handle signals */
1271                 if (signal_pending(current)) {
1272                         *err = sock_intr_errno(timeo);
1273                         return NULL;
1274                 }
1275         }
1276
1277         return skb;
1278 }
1279
1280 static int tls_setup_from_iter(struct sock *sk, struct iov_iter *from,
1281                                int length, int *pages_used,
1282                                unsigned int *size_used,
1283                                struct scatterlist *to,
1284                                int to_max_pages)
1285 {
1286         int rc = 0, i = 0, num_elem = *pages_used, maxpages;
1287         struct page *pages[MAX_SKB_FRAGS];
1288         unsigned int size = *size_used;
1289         ssize_t copied, use;
1290         size_t offset;
1291
1292         while (length > 0) {
1293                 i = 0;
1294                 maxpages = to_max_pages - num_elem;
1295                 if (maxpages == 0) {
1296                         rc = -EFAULT;
1297                         goto out;
1298                 }
1299                 copied = iov_iter_get_pages(from, pages,
1300                                             length,
1301                                             maxpages, &offset);
1302                 if (copied <= 0) {
1303                         rc = -EFAULT;
1304                         goto out;
1305                 }
1306
1307                 iov_iter_advance(from, copied);
1308
1309                 length -= copied;
1310                 size += copied;
1311                 while (copied) {
1312                         use = min_t(int, copied, PAGE_SIZE - offset);
1313
1314                         sg_set_page(&to[num_elem],
1315                                     pages[i], use, offset);
1316                         sg_unmark_end(&to[num_elem]);
1317                         /* We do not uncharge memory from this API */
1318
1319                         offset = 0;
1320                         copied -= use;
1321
1322                         i++;
1323                         num_elem++;
1324                 }
1325         }
1326         /* Mark the end in the last sg entry if newly added */
1327         if (num_elem > *pages_used)
1328                 sg_mark_end(&to[num_elem - 1]);
1329 out:
1330         if (rc)
1331                 iov_iter_revert(from, size - *size_used);
1332         *size_used = size;
1333         *pages_used = num_elem;
1334
1335         return rc;
1336 }
1337
1338 /* This function decrypts the input skb into either out_iov or in out_sg
1339  * or in skb buffers itself. The input parameter 'zc' indicates if
1340  * zero-copy mode needs to be tried or not. With zero-copy mode, either
1341  * out_iov or out_sg must be non-NULL. In case both out_iov and out_sg are
1342  * NULL, then the decryption happens inside skb buffers itself, i.e.
1343  * zero-copy gets disabled and 'zc' is updated.
1344  */
1345
1346 static int decrypt_internal(struct sock *sk, struct sk_buff *skb,
1347                             struct iov_iter *out_iov,
1348                             struct scatterlist *out_sg,
1349                             int *chunk, bool *zc, bool async)
1350 {
1351         struct tls_context *tls_ctx = tls_get_ctx(sk);
1352         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1353         struct tls_prot_info *prot = &tls_ctx->prot_info;
1354         struct strp_msg *rxm = strp_msg(skb);
1355         int n_sgin, n_sgout, nsg, mem_size, aead_size, err, pages = 0;
1356         struct aead_request *aead_req;
1357         struct sk_buff *unused;
1358         u8 *aad, *iv, *mem = NULL;
1359         struct scatterlist *sgin = NULL;
1360         struct scatterlist *sgout = NULL;
1361         const int data_len = rxm->full_len - prot->overhead_size +
1362                              prot->tail_size;
1363         int iv_offset = 0;
1364
1365         if (*zc && (out_iov || out_sg)) {
1366                 if (out_iov)
1367                         n_sgout = iov_iter_npages(out_iov, INT_MAX) + 1;
1368                 else
1369                         n_sgout = sg_nents(out_sg);
1370                 n_sgin = skb_nsg(skb, rxm->offset + prot->prepend_size,
1371                                  rxm->full_len - prot->prepend_size);
1372         } else {
1373                 n_sgout = 0;
1374                 *zc = false;
1375                 n_sgin = skb_cow_data(skb, 0, &unused);
1376         }
1377
1378         if (n_sgin < 1)
1379                 return -EBADMSG;
1380
1381         /* Increment to accommodate AAD */
1382         n_sgin = n_sgin + 1;
1383
1384         nsg = n_sgin + n_sgout;
1385
1386         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(ctx->aead_recv);
1387         mem_size = aead_size + (nsg * sizeof(struct scatterlist));
1388         mem_size = mem_size + prot->aad_size;
1389         mem_size = mem_size + crypto_aead_ivsize(ctx->aead_recv);
1390
1391         /* Allocate a single block of memory which contains
1392          * aead_req || sgin[] || sgout[] || aad || iv.
1393          * This order achieves correct alignment for aead_req, sgin, sgout.
1394          */
1395         mem = kmalloc(mem_size, sk->sk_allocation);
1396         if (!mem)
1397                 return -ENOMEM;
1398
1399         /* Segment the allocated memory */
1400         aead_req = (struct aead_request *)mem;
1401         sgin = (struct scatterlist *)(mem + aead_size);
1402         sgout = sgin + n_sgin;
1403         aad = (u8 *)(sgout + n_sgout);
1404         iv = aad + prot->aad_size;
1405
1406         /* For CCM based ciphers, first byte of nonce+iv is always '2' */
1407         if (prot->cipher_type == TLS_CIPHER_AES_CCM_128) {
1408                 iv[0] = 2;
1409                 iv_offset = 1;
1410         }
1411
1412         /* Prepare IV */
1413         err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
1414                             iv + iv_offset + prot->salt_size,
1415                             prot->iv_size);
1416         if (err < 0) {
1417                 kfree(mem);
1418                 return err;
1419         }
1420         if (prot->version == TLS_1_3_VERSION)
1421                 memcpy(iv + iv_offset, tls_ctx->rx.iv,
1422                        crypto_aead_ivsize(ctx->aead_recv));
1423         else
1424                 memcpy(iv + iv_offset, tls_ctx->rx.iv, prot->salt_size);
1425
1426         xor_iv_with_seq(prot->version, iv, tls_ctx->rx.rec_seq);
1427
1428         /* Prepare AAD */
1429         tls_make_aad(aad, rxm->full_len - prot->overhead_size +
1430                      prot->tail_size,
1431                      tls_ctx->rx.rec_seq, prot->rec_seq_size,
1432                      ctx->control, prot->version);
1433
1434         /* Prepare sgin */
1435         sg_init_table(sgin, n_sgin);
1436         sg_set_buf(&sgin[0], aad, prot->aad_size);
1437         err = skb_to_sgvec(skb, &sgin[1],
1438                            rxm->offset + prot->prepend_size,
1439                            rxm->full_len - prot->prepend_size);
1440         if (err < 0) {
1441                 kfree(mem);
1442                 return err;
1443         }
1444
1445         if (n_sgout) {
1446                 if (out_iov) {
1447                         sg_init_table(sgout, n_sgout);
1448                         sg_set_buf(&sgout[0], aad, prot->aad_size);
1449
1450                         *chunk = 0;
1451                         err = tls_setup_from_iter(sk, out_iov, data_len,
1452                                                   &pages, chunk, &sgout[1],
1453                                                   (n_sgout - 1));
1454                         if (err < 0)
1455                                 goto fallback_to_reg_recv;
1456                 } else if (out_sg) {
1457                         memcpy(sgout, out_sg, n_sgout * sizeof(*sgout));
1458                 } else {
1459                         goto fallback_to_reg_recv;
1460                 }
1461         } else {
1462 fallback_to_reg_recv:
1463                 sgout = sgin;
1464                 pages = 0;
1465                 *chunk = data_len;
1466                 *zc = false;
1467         }
1468
1469         /* Prepare and submit AEAD request */
1470         err = tls_do_decryption(sk, skb, sgin, sgout, iv,
1471                                 data_len, aead_req, async);
1472         if (err == -EINPROGRESS)
1473                 return err;
1474
1475         /* Release the pages in case iov was mapped to pages */
1476         for (; pages > 0; pages--)
1477                 put_page(sg_page(&sgout[pages]));
1478
1479         kfree(mem);
1480         return err;
1481 }
1482
1483 static int decrypt_skb_update(struct sock *sk, struct sk_buff *skb,
1484                               struct iov_iter *dest, int *chunk, bool *zc,
1485                               bool async)
1486 {
1487         struct tls_context *tls_ctx = tls_get_ctx(sk);
1488         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1489         struct tls_prot_info *prot = &tls_ctx->prot_info;
1490         struct strp_msg *rxm = strp_msg(skb);
1491         int pad, err = 0;
1492
1493         if (!ctx->decrypted) {
1494                 if (tls_ctx->rx_conf == TLS_HW) {
1495                         err = tls_device_decrypted(sk, skb);
1496                         if (err < 0)
1497                                 return err;
1498                 }
1499
1500                 /* Still not decrypted after tls_device */
1501                 if (!ctx->decrypted) {
1502                         err = decrypt_internal(sk, skb, dest, NULL, chunk, zc,
1503                                                async);
1504                         if (err < 0) {
1505                                 if (err == -EINPROGRESS)
1506                                         tls_advance_record_sn(sk, prot,
1507                                                               &tls_ctx->rx);
1508
1509                                 return err;
1510                         }
1511                 } else {
1512                         *zc = false;
1513                 }
1514
1515                 pad = padding_length(ctx, prot, skb);
1516                 if (pad < 0)
1517                         return pad;
1518
1519                 rxm->full_len -= pad;
1520                 rxm->offset += prot->prepend_size;
1521                 rxm->full_len -= prot->overhead_size;
1522                 tls_advance_record_sn(sk, prot, &tls_ctx->rx);
1523                 ctx->decrypted = true;
1524                 ctx->saved_data_ready(sk);
1525         } else {
1526                 *zc = false;
1527         }
1528
1529         return err;
1530 }
1531
1532 int decrypt_skb(struct sock *sk, struct sk_buff *skb,
1533                 struct scatterlist *sgout)
1534 {
1535         bool zc = true;
1536         int chunk;
1537
1538         return decrypt_internal(sk, skb, NULL, sgout, &chunk, &zc, false);
1539 }
1540
1541 static bool tls_sw_advance_skb(struct sock *sk, struct sk_buff *skb,
1542                                unsigned int len)
1543 {
1544         struct tls_context *tls_ctx = tls_get_ctx(sk);
1545         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1546
1547         if (skb) {
1548                 struct strp_msg *rxm = strp_msg(skb);
1549
1550                 if (len < rxm->full_len) {
1551                         rxm->offset += len;
1552                         rxm->full_len -= len;
1553                         return false;
1554                 }
1555                 consume_skb(skb);
1556         }
1557
1558         /* Finished with message */
1559         ctx->recv_pkt = NULL;
1560         __strp_unpause(&ctx->strp);
1561
1562         return true;
1563 }
1564
1565 /* This function traverses the rx_list in tls receive context to copies the
1566  * decrypted records into the buffer provided by caller zero copy is not
1567  * true. Further, the records are removed from the rx_list if it is not a peek
1568  * case and the record has been consumed completely.
1569  */
1570 static int process_rx_list(struct tls_sw_context_rx *ctx,
1571                            struct msghdr *msg,
1572                            u8 *control,
1573                            bool *cmsg,
1574                            size_t skip,
1575                            size_t len,
1576                            bool zc,
1577                            bool is_peek)
1578 {
1579         struct sk_buff *skb = skb_peek(&ctx->rx_list);
1580         u8 ctrl = *control;
1581         u8 msgc = *cmsg;
1582         struct tls_msg *tlm;
1583         ssize_t copied = 0;
1584
1585         /* Set the record type in 'control' if caller didn't pass it */
1586         if (!ctrl && skb) {
1587                 tlm = tls_msg(skb);
1588                 ctrl = tlm->control;
1589         }
1590
1591         while (skip && skb) {
1592                 struct strp_msg *rxm = strp_msg(skb);
1593                 tlm = tls_msg(skb);
1594
1595                 /* Cannot process a record of different type */
1596                 if (ctrl != tlm->control)
1597                         return 0;
1598
1599                 if (skip < rxm->full_len)
1600                         break;
1601
1602                 skip = skip - rxm->full_len;
1603                 skb = skb_peek_next(skb, &ctx->rx_list);
1604         }
1605
1606         while (len && skb) {
1607                 struct sk_buff *next_skb;
1608                 struct strp_msg *rxm = strp_msg(skb);
1609                 int chunk = min_t(unsigned int, rxm->full_len - skip, len);
1610
1611                 tlm = tls_msg(skb);
1612
1613                 /* Cannot process a record of different type */
1614                 if (ctrl != tlm->control)
1615                         return 0;
1616
1617                 /* Set record type if not already done. For a non-data record,
1618                  * do not proceed if record type could not be copied.
1619                  */
1620                 if (!msgc) {
1621                         int cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1622                                             sizeof(ctrl), &ctrl);
1623                         msgc = true;
1624                         if (ctrl != TLS_RECORD_TYPE_DATA) {
1625                                 if (cerr || msg->msg_flags & MSG_CTRUNC)
1626                                         return -EIO;
1627
1628                                 *cmsg = msgc;
1629                         }
1630                 }
1631
1632                 if (!zc || (rxm->full_len - skip) > len) {
1633                         int err = skb_copy_datagram_msg(skb, rxm->offset + skip,
1634                                                     msg, chunk);
1635                         if (err < 0)
1636                                 return err;
1637                 }
1638
1639                 len = len - chunk;
1640                 copied = copied + chunk;
1641
1642                 /* Consume the data from record if it is non-peek case*/
1643                 if (!is_peek) {
1644                         rxm->offset = rxm->offset + chunk;
1645                         rxm->full_len = rxm->full_len - chunk;
1646
1647                         /* Return if there is unconsumed data in the record */
1648                         if (rxm->full_len - skip)
1649                                 break;
1650                 }
1651
1652                 /* The remaining skip-bytes must lie in 1st record in rx_list.
1653                  * So from the 2nd record, 'skip' should be 0.
1654                  */
1655                 skip = 0;
1656
1657                 if (msg)
1658                         msg->msg_flags |= MSG_EOR;
1659
1660                 next_skb = skb_peek_next(skb, &ctx->rx_list);
1661
1662                 if (!is_peek) {
1663                         skb_unlink(skb, &ctx->rx_list);
1664                         consume_skb(skb);
1665                 }
1666
1667                 skb = next_skb;
1668         }
1669
1670         *control = ctrl;
1671         return copied;
1672 }
1673
1674 int tls_sw_recvmsg(struct sock *sk,
1675                    struct msghdr *msg,
1676                    size_t len,
1677                    int nonblock,
1678                    int flags,
1679                    int *addr_len)
1680 {
1681         struct tls_context *tls_ctx = tls_get_ctx(sk);
1682         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1683         struct tls_prot_info *prot = &tls_ctx->prot_info;
1684         struct sk_psock *psock;
1685         unsigned char control = 0;
1686         ssize_t decrypted = 0;
1687         struct strp_msg *rxm;
1688         struct tls_msg *tlm;
1689         struct sk_buff *skb;
1690         ssize_t copied = 0;
1691         bool cmsg = false;
1692         int target, err = 0;
1693         long timeo;
1694         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
1695         bool is_peek = flags & MSG_PEEK;
1696         int num_async = 0;
1697
1698         flags |= nonblock;
1699
1700         if (unlikely(flags & MSG_ERRQUEUE))
1701                 return sock_recv_errqueue(sk, msg, len, SOL_IP, IP_RECVERR);
1702
1703         psock = sk_psock_get(sk);
1704         lock_sock(sk);
1705
1706         /* Process pending decrypted records. It must be non-zero-copy */
1707         err = process_rx_list(ctx, msg, &control, &cmsg, 0, len, false,
1708                               is_peek);
1709         if (err < 0) {
1710                 tls_err_abort(sk, err);
1711                 goto end;
1712         } else {
1713                 copied = err;
1714         }
1715
1716         if (len <= copied)
1717                 goto recv_end;
1718
1719         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1720         len = len - copied;
1721         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1722
1723         while (len && (decrypted + copied < target || ctx->recv_pkt)) {
1724                 bool retain_skb = false;
1725                 bool zc = false;
1726                 int to_decrypt;
1727                 int chunk = 0;
1728                 bool async_capable;
1729                 bool async = false;
1730
1731                 skb = tls_wait_data(sk, psock, flags, timeo, &err);
1732                 if (!skb) {
1733                         if (psock) {
1734                                 int ret = __tcp_bpf_recvmsg(sk, psock,
1735                                                             msg, len, flags);
1736
1737                                 if (ret > 0) {
1738                                         decrypted += ret;
1739                                         len -= ret;
1740                                         continue;
1741                                 }
1742                         }
1743                         goto recv_end;
1744                 } else {
1745                         tlm = tls_msg(skb);
1746                         if (prot->version == TLS_1_3_VERSION)
1747                                 tlm->control = 0;
1748                         else
1749                                 tlm->control = ctx->control;
1750                 }
1751
1752                 rxm = strp_msg(skb);
1753
1754                 to_decrypt = rxm->full_len - prot->overhead_size;
1755
1756                 if (to_decrypt <= len && !is_kvec && !is_peek &&
1757                     ctx->control == TLS_RECORD_TYPE_DATA &&
1758                     prot->version != TLS_1_3_VERSION)
1759                         zc = true;
1760
1761                 /* Do not use async mode if record is non-data */
1762                 if (ctx->control == TLS_RECORD_TYPE_DATA)
1763                         async_capable = ctx->async_capable;
1764                 else
1765                         async_capable = false;
1766
1767                 err = decrypt_skb_update(sk, skb, &msg->msg_iter,
1768                                          &chunk, &zc, async_capable);
1769                 if (err < 0 && err != -EINPROGRESS) {
1770                         tls_err_abort(sk, EBADMSG);
1771                         goto recv_end;
1772                 }
1773
1774                 if (err == -EINPROGRESS) {
1775                         async = true;
1776                         num_async++;
1777                 } else if (prot->version == TLS_1_3_VERSION) {
1778                         tlm->control = ctx->control;
1779                 }
1780
1781                 /* If the type of records being processed is not known yet,
1782                  * set it to record type just dequeued. If it is already known,
1783                  * but does not match the record type just dequeued, go to end.
1784                  * We always get record type here since for tls1.2, record type
1785                  * is known just after record is dequeued from stream parser.
1786                  * For tls1.3, we disable async.
1787                  */
1788
1789                 if (!control)
1790                         control = tlm->control;
1791                 else if (control != tlm->control)
1792                         goto recv_end;
1793
1794                 if (!cmsg) {
1795                         int cerr;
1796
1797                         cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1798                                         sizeof(control), &control);
1799                         cmsg = true;
1800                         if (control != TLS_RECORD_TYPE_DATA) {
1801                                 if (cerr || msg->msg_flags & MSG_CTRUNC) {
1802                                         err = -EIO;
1803                                         goto recv_end;
1804                                 }
1805                         }
1806                 }
1807
1808                 if (async)
1809                         goto pick_next_record;
1810
1811                 if (!zc) {
1812                         if (rxm->full_len > len) {
1813                                 retain_skb = true;
1814                                 chunk = len;
1815                         } else {
1816                                 chunk = rxm->full_len;
1817                         }
1818
1819                         err = skb_copy_datagram_msg(skb, rxm->offset,
1820                                                     msg, chunk);
1821                         if (err < 0)
1822                                 goto recv_end;
1823
1824                         if (!is_peek) {
1825                                 rxm->offset = rxm->offset + chunk;
1826                                 rxm->full_len = rxm->full_len - chunk;
1827                         }
1828                 }
1829
1830 pick_next_record:
1831                 if (chunk > len)
1832                         chunk = len;
1833
1834                 decrypted += chunk;
1835                 len -= chunk;
1836
1837                 /* For async or peek case, queue the current skb */
1838                 if (async || is_peek || retain_skb) {
1839                         skb_queue_tail(&ctx->rx_list, skb);
1840                         skb = NULL;
1841                 }
1842
1843                 if (tls_sw_advance_skb(sk, skb, chunk)) {
1844                         /* Return full control message to
1845                          * userspace before trying to parse
1846                          * another message type
1847                          */
1848                         msg->msg_flags |= MSG_EOR;
1849                         if (ctx->control != TLS_RECORD_TYPE_DATA)
1850                                 goto recv_end;
1851                 } else {
1852                         break;
1853                 }
1854         }
1855
1856 recv_end:
1857         if (num_async) {
1858                 /* Wait for all previously submitted records to be decrypted */
1859                 smp_store_mb(ctx->async_notify, true);
1860                 if (atomic_read(&ctx->decrypt_pending)) {
1861                         err = crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1862                         if (err) {
1863                                 /* one of async decrypt failed */
1864                                 tls_err_abort(sk, err);
1865                                 copied = 0;
1866                                 decrypted = 0;
1867                                 goto end;
1868                         }
1869                 } else {
1870                         reinit_completion(&ctx->async_wait.completion);
1871                 }
1872                 WRITE_ONCE(ctx->async_notify, false);
1873
1874                 /* Drain records from the rx_list & copy if required */
1875                 if (is_peek || is_kvec)
1876                         err = process_rx_list(ctx, msg, &control, &cmsg, copied,
1877                                               decrypted, false, is_peek);
1878                 else
1879                         err = process_rx_list(ctx, msg, &control, &cmsg, 0,
1880                                               decrypted, true, is_peek);
1881                 if (err < 0) {
1882                         tls_err_abort(sk, err);
1883                         copied = 0;
1884                         goto end;
1885                 }
1886         }
1887
1888         copied += decrypted;
1889
1890 end:
1891         release_sock(sk);
1892         if (psock)
1893                 sk_psock_put(sk, psock);
1894         return copied ? : err;
1895 }
1896
1897 ssize_t tls_sw_splice_read(struct socket *sock,  loff_t *ppos,
1898                            struct pipe_inode_info *pipe,
1899                            size_t len, unsigned int flags)
1900 {
1901         struct tls_context *tls_ctx = tls_get_ctx(sock->sk);
1902         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1903         struct strp_msg *rxm = NULL;
1904         struct sock *sk = sock->sk;
1905         struct sk_buff *skb;
1906         ssize_t copied = 0;
1907         int err = 0;
1908         long timeo;
1909         int chunk;
1910         bool zc = false;
1911
1912         lock_sock(sk);
1913
1914         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1915
1916         skb = tls_wait_data(sk, NULL, flags, timeo, &err);
1917         if (!skb)
1918                 goto splice_read_end;
1919
1920         if (!ctx->decrypted) {
1921                 err = decrypt_skb_update(sk, skb, NULL, &chunk, &zc, false);
1922
1923                 /* splice does not support reading control messages */
1924                 if (ctx->control != TLS_RECORD_TYPE_DATA) {
1925                         err = -ENOTSUPP;
1926                         goto splice_read_end;
1927                 }
1928
1929                 if (err < 0) {
1930                         tls_err_abort(sk, EBADMSG);
1931                         goto splice_read_end;
1932                 }
1933                 ctx->decrypted = true;
1934         }
1935         rxm = strp_msg(skb);
1936
1937         chunk = min_t(unsigned int, rxm->full_len, len);
1938         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, chunk, flags);
1939         if (copied < 0)
1940                 goto splice_read_end;
1941
1942         if (likely(!(flags & MSG_PEEK)))
1943                 tls_sw_advance_skb(sk, skb, copied);
1944
1945 splice_read_end:
1946         release_sock(sk);
1947         return copied ? : err;
1948 }
1949
1950 bool tls_sw_stream_read(const struct sock *sk)
1951 {
1952         struct tls_context *tls_ctx = tls_get_ctx(sk);
1953         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1954         bool ingress_empty = true;
1955         struct sk_psock *psock;
1956
1957         rcu_read_lock();
1958         psock = sk_psock(sk);
1959         if (psock)
1960                 ingress_empty = list_empty(&psock->ingress_msg);
1961         rcu_read_unlock();
1962
1963         return !ingress_empty || ctx->recv_pkt ||
1964                 !skb_queue_empty(&ctx->rx_list);
1965 }
1966
1967 static int tls_read_size(struct strparser *strp, struct sk_buff *skb)
1968 {
1969         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
1970         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1971         struct tls_prot_info *prot = &tls_ctx->prot_info;
1972         char header[TLS_HEADER_SIZE + MAX_IV_SIZE];
1973         struct strp_msg *rxm = strp_msg(skb);
1974         size_t cipher_overhead;
1975         size_t data_len = 0;
1976         int ret;
1977
1978         /* Verify that we have a full TLS header, or wait for more data */
1979         if (rxm->offset + prot->prepend_size > skb->len)
1980                 return 0;
1981
1982         /* Sanity-check size of on-stack buffer. */
1983         if (WARN_ON(prot->prepend_size > sizeof(header))) {
1984                 ret = -EINVAL;
1985                 goto read_failure;
1986         }
1987
1988         /* Linearize header to local buffer */
1989         ret = skb_copy_bits(skb, rxm->offset, header, prot->prepend_size);
1990
1991         if (ret < 0)
1992                 goto read_failure;
1993
1994         ctx->control = header[0];
1995
1996         data_len = ((header[4] & 0xFF) | (header[3] << 8));
1997
1998         cipher_overhead = prot->tag_size;
1999         if (prot->version != TLS_1_3_VERSION)
2000                 cipher_overhead += prot->iv_size;
2001
2002         if (data_len > TLS_MAX_PAYLOAD_SIZE + cipher_overhead +
2003             prot->tail_size) {
2004                 ret = -EMSGSIZE;
2005                 goto read_failure;
2006         }
2007         if (data_len < cipher_overhead) {
2008                 ret = -EBADMSG;
2009                 goto read_failure;
2010         }
2011
2012         /* Note that both TLS1.3 and TLS1.2 use TLS_1_2 version here */
2013         if (header[1] != TLS_1_2_VERSION_MINOR ||
2014             header[2] != TLS_1_2_VERSION_MAJOR) {
2015                 ret = -EINVAL;
2016                 goto read_failure;
2017         }
2018
2019         tls_device_rx_resync_new_rec(strp->sk, data_len + TLS_HEADER_SIZE,
2020                                      TCP_SKB_CB(skb)->seq + rxm->offset);
2021         return data_len + TLS_HEADER_SIZE;
2022
2023 read_failure:
2024         tls_err_abort(strp->sk, ret);
2025
2026         return ret;
2027 }
2028
2029 static void tls_queue(struct strparser *strp, struct sk_buff *skb)
2030 {
2031         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2032         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2033
2034         ctx->decrypted = false;
2035
2036         ctx->recv_pkt = skb;
2037         strp_pause(strp);
2038
2039         ctx->saved_data_ready(strp->sk);
2040 }
2041
2042 static void tls_data_ready(struct sock *sk)
2043 {
2044         struct tls_context *tls_ctx = tls_get_ctx(sk);
2045         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2046         struct sk_psock *psock;
2047
2048         strp_data_ready(&ctx->strp);
2049
2050         psock = sk_psock_get(sk);
2051         if (psock && !list_empty(&psock->ingress_msg)) {
2052                 ctx->saved_data_ready(sk);
2053                 sk_psock_put(sk, psock);
2054         }
2055 }
2056
2057 void tls_sw_cancel_work_tx(struct tls_context *tls_ctx)
2058 {
2059         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2060
2061         set_bit(BIT_TX_CLOSING, &ctx->tx_bitmask);
2062         set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask);
2063         cancel_delayed_work_sync(&ctx->tx_work.work);
2064 }
2065
2066 void tls_sw_release_resources_tx(struct sock *sk)
2067 {
2068         struct tls_context *tls_ctx = tls_get_ctx(sk);
2069         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2070         struct tls_rec *rec, *tmp;
2071
2072         /* Wait for any pending async encryptions to complete */
2073         smp_store_mb(ctx->async_notify, true);
2074         if (atomic_read(&ctx->encrypt_pending))
2075                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
2076
2077         tls_tx_records(sk, -1);
2078
2079         /* Free up un-sent records in tx_list. First, free
2080          * the partially sent record if any at head of tx_list.
2081          */
2082         if (tls_free_partial_record(sk, tls_ctx)) {
2083                 rec = list_first_entry(&ctx->tx_list,
2084                                        struct tls_rec, list);
2085                 list_del(&rec->list);
2086                 sk_msg_free(sk, &rec->msg_plaintext);
2087                 kfree(rec);
2088         }
2089
2090         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
2091                 list_del(&rec->list);
2092                 sk_msg_free(sk, &rec->msg_encrypted);
2093                 sk_msg_free(sk, &rec->msg_plaintext);
2094                 kfree(rec);
2095         }
2096
2097         crypto_free_aead(ctx->aead_send);
2098         tls_free_open_rec(sk);
2099 }
2100
2101 void tls_sw_free_ctx_tx(struct tls_context *tls_ctx)
2102 {
2103         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2104
2105         kfree(ctx);
2106 }
2107
2108 void tls_sw_release_resources_rx(struct sock *sk)
2109 {
2110         struct tls_context *tls_ctx = tls_get_ctx(sk);
2111         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2112
2113         kfree(tls_ctx->rx.rec_seq);
2114         kfree(tls_ctx->rx.iv);
2115
2116         if (ctx->aead_recv) {
2117                 kfree_skb(ctx->recv_pkt);
2118                 ctx->recv_pkt = NULL;
2119                 skb_queue_purge(&ctx->rx_list);
2120                 crypto_free_aead(ctx->aead_recv);
2121                 strp_stop(&ctx->strp);
2122                 /* If tls_sw_strparser_arm() was not called (cleanup paths)
2123                  * we still want to strp_stop(), but sk->sk_data_ready was
2124                  * never swapped.
2125                  */
2126                 if (ctx->saved_data_ready) {
2127                         write_lock_bh(&sk->sk_callback_lock);
2128                         sk->sk_data_ready = ctx->saved_data_ready;
2129                         write_unlock_bh(&sk->sk_callback_lock);
2130                 }
2131         }
2132 }
2133
2134 void tls_sw_strparser_done(struct tls_context *tls_ctx)
2135 {
2136         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2137
2138         strp_done(&ctx->strp);
2139 }
2140
2141 void tls_sw_free_ctx_rx(struct tls_context *tls_ctx)
2142 {
2143         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2144
2145         kfree(ctx);
2146 }
2147
2148 void tls_sw_free_resources_rx(struct sock *sk)
2149 {
2150         struct tls_context *tls_ctx = tls_get_ctx(sk);
2151
2152         tls_sw_release_resources_rx(sk);
2153         tls_sw_free_ctx_rx(tls_ctx);
2154 }
2155
2156 /* The work handler to transmitt the encrypted records in tx_list */
2157 static void tx_work_handler(struct work_struct *work)
2158 {
2159         struct delayed_work *delayed_work = to_delayed_work(work);
2160         struct tx_work *tx_work = container_of(delayed_work,
2161                                                struct tx_work, work);
2162         struct sock *sk = tx_work->sk;
2163         struct tls_context *tls_ctx = tls_get_ctx(sk);
2164         struct tls_sw_context_tx *ctx;
2165
2166         if (unlikely(!tls_ctx))
2167                 return;
2168
2169         ctx = tls_sw_ctx_tx(tls_ctx);
2170         if (test_bit(BIT_TX_CLOSING, &ctx->tx_bitmask))
2171                 return;
2172
2173         if (!test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
2174                 return;
2175         mutex_lock(&tls_ctx->tx_lock);
2176         lock_sock(sk);
2177         tls_tx_records(sk, -1);
2178         release_sock(sk);
2179         mutex_unlock(&tls_ctx->tx_lock);
2180 }
2181
2182 void tls_sw_write_space(struct sock *sk, struct tls_context *ctx)
2183 {
2184         struct tls_sw_context_tx *tx_ctx = tls_sw_ctx_tx(ctx);
2185
2186         /* Schedule the transmission if tx list is ready */
2187         if (is_tx_ready(tx_ctx) &&
2188             !test_and_set_bit(BIT_TX_SCHEDULED, &tx_ctx->tx_bitmask))
2189                 schedule_delayed_work(&tx_ctx->tx_work.work, 0);
2190 }
2191
2192 void tls_sw_strparser_arm(struct sock *sk, struct tls_context *tls_ctx)
2193 {
2194         struct tls_sw_context_rx *rx_ctx = tls_sw_ctx_rx(tls_ctx);
2195
2196         write_lock_bh(&sk->sk_callback_lock);
2197         rx_ctx->saved_data_ready = sk->sk_data_ready;
2198         sk->sk_data_ready = tls_data_ready;
2199         write_unlock_bh(&sk->sk_callback_lock);
2200
2201         strp_check_rcv(&rx_ctx->strp);
2202 }
2203
2204 int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
2205 {
2206         struct tls_context *tls_ctx = tls_get_ctx(sk);
2207         struct tls_prot_info *prot = &tls_ctx->prot_info;
2208         struct tls_crypto_info *crypto_info;
2209         struct tls12_crypto_info_aes_gcm_128 *gcm_128_info;
2210         struct tls12_crypto_info_aes_gcm_256 *gcm_256_info;
2211         struct tls12_crypto_info_aes_ccm_128 *ccm_128_info;
2212         struct tls_sw_context_tx *sw_ctx_tx = NULL;
2213         struct tls_sw_context_rx *sw_ctx_rx = NULL;
2214         struct cipher_context *cctx;
2215         struct crypto_aead **aead;
2216         struct strp_callbacks cb;
2217         u16 nonce_size, tag_size, iv_size, rec_seq_size, salt_size;
2218         struct crypto_tfm *tfm;
2219         char *iv, *rec_seq, *key, *salt, *cipher_name;
2220         size_t keysize;
2221         int rc = 0;
2222
2223         if (!ctx) {
2224                 rc = -EINVAL;
2225                 goto out;
2226         }
2227
2228         if (tx) {
2229                 if (!ctx->priv_ctx_tx) {
2230                         sw_ctx_tx = kzalloc(sizeof(*sw_ctx_tx), GFP_KERNEL);
2231                         if (!sw_ctx_tx) {
2232                                 rc = -ENOMEM;
2233                                 goto out;
2234                         }
2235                         ctx->priv_ctx_tx = sw_ctx_tx;
2236                 } else {
2237                         sw_ctx_tx =
2238                                 (struct tls_sw_context_tx *)ctx->priv_ctx_tx;
2239                 }
2240         } else {
2241                 if (!ctx->priv_ctx_rx) {
2242                         sw_ctx_rx = kzalloc(sizeof(*sw_ctx_rx), GFP_KERNEL);
2243                         if (!sw_ctx_rx) {
2244                                 rc = -ENOMEM;
2245                                 goto out;
2246                         }
2247                         ctx->priv_ctx_rx = sw_ctx_rx;
2248                 } else {
2249                         sw_ctx_rx =
2250                                 (struct tls_sw_context_rx *)ctx->priv_ctx_rx;
2251                 }
2252         }
2253
2254         if (tx) {
2255                 crypto_init_wait(&sw_ctx_tx->async_wait);
2256                 crypto_info = &ctx->crypto_send.info;
2257                 cctx = &ctx->tx;
2258                 aead = &sw_ctx_tx->aead_send;
2259                 INIT_LIST_HEAD(&sw_ctx_tx->tx_list);
2260                 INIT_DELAYED_WORK(&sw_ctx_tx->tx_work.work, tx_work_handler);
2261                 sw_ctx_tx->tx_work.sk = sk;
2262         } else {
2263                 crypto_init_wait(&sw_ctx_rx->async_wait);
2264                 crypto_info = &ctx->crypto_recv.info;
2265                 cctx = &ctx->rx;
2266                 skb_queue_head_init(&sw_ctx_rx->rx_list);
2267                 aead = &sw_ctx_rx->aead_recv;
2268         }
2269
2270         switch (crypto_info->cipher_type) {
2271         case TLS_CIPHER_AES_GCM_128: {
2272                 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2273                 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
2274                 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2275                 iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
2276                 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
2277                 rec_seq =
2278                  ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
2279                 gcm_128_info =
2280                         (struct tls12_crypto_info_aes_gcm_128 *)crypto_info;
2281                 keysize = TLS_CIPHER_AES_GCM_128_KEY_SIZE;
2282                 key = gcm_128_info->key;
2283                 salt = gcm_128_info->salt;
2284                 salt_size = TLS_CIPHER_AES_GCM_128_SALT_SIZE;
2285                 cipher_name = "gcm(aes)";
2286                 break;
2287         }
2288         case TLS_CIPHER_AES_GCM_256: {
2289                 nonce_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2290                 tag_size = TLS_CIPHER_AES_GCM_256_TAG_SIZE;
2291                 iv_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2292                 iv = ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->iv;
2293                 rec_seq_size = TLS_CIPHER_AES_GCM_256_REC_SEQ_SIZE;
2294                 rec_seq =
2295                  ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->rec_seq;
2296                 gcm_256_info =
2297                         (struct tls12_crypto_info_aes_gcm_256 *)crypto_info;
2298                 keysize = TLS_CIPHER_AES_GCM_256_KEY_SIZE;
2299                 key = gcm_256_info->key;
2300                 salt = gcm_256_info->salt;
2301                 salt_size = TLS_CIPHER_AES_GCM_256_SALT_SIZE;
2302                 cipher_name = "gcm(aes)";
2303                 break;
2304         }
2305         case TLS_CIPHER_AES_CCM_128: {
2306                 nonce_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2307                 tag_size = TLS_CIPHER_AES_CCM_128_TAG_SIZE;
2308                 iv_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2309                 iv = ((struct tls12_crypto_info_aes_ccm_128 *)crypto_info)->iv;
2310                 rec_seq_size = TLS_CIPHER_AES_CCM_128_REC_SEQ_SIZE;
2311                 rec_seq =
2312                 ((struct tls12_crypto_info_aes_ccm_128 *)crypto_info)->rec_seq;
2313                 ccm_128_info =
2314                 (struct tls12_crypto_info_aes_ccm_128 *)crypto_info;
2315                 keysize = TLS_CIPHER_AES_CCM_128_KEY_SIZE;
2316                 key = ccm_128_info->key;
2317                 salt = ccm_128_info->salt;
2318                 salt_size = TLS_CIPHER_AES_CCM_128_SALT_SIZE;
2319                 cipher_name = "ccm(aes)";
2320                 break;
2321         }
2322         default:
2323                 rc = -EINVAL;
2324                 goto free_priv;
2325         }
2326
2327         /* Sanity-check the sizes for stack allocations. */
2328         if (iv_size > MAX_IV_SIZE || nonce_size > MAX_IV_SIZE ||
2329             rec_seq_size > TLS_MAX_REC_SEQ_SIZE) {
2330                 rc = -EINVAL;
2331                 goto free_priv;
2332         }
2333
2334         if (crypto_info->version == TLS_1_3_VERSION) {
2335                 nonce_size = 0;
2336                 prot->aad_size = TLS_HEADER_SIZE;
2337                 prot->tail_size = 1;
2338         } else {
2339                 prot->aad_size = TLS_AAD_SPACE_SIZE;
2340                 prot->tail_size = 0;
2341         }
2342
2343         prot->version = crypto_info->version;
2344         prot->cipher_type = crypto_info->cipher_type;
2345         prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
2346         prot->tag_size = tag_size;
2347         prot->overhead_size = prot->prepend_size +
2348                               prot->tag_size + prot->tail_size;
2349         prot->iv_size = iv_size;
2350         prot->salt_size = salt_size;
2351         cctx->iv = kmalloc(iv_size + salt_size, GFP_KERNEL);
2352         if (!cctx->iv) {
2353                 rc = -ENOMEM;
2354                 goto free_priv;
2355         }
2356         /* Note: 128 & 256 bit salt are the same size */
2357         prot->rec_seq_size = rec_seq_size;
2358         memcpy(cctx->iv, salt, salt_size);
2359         memcpy(cctx->iv + salt_size, iv, iv_size);
2360         cctx->rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
2361         if (!cctx->rec_seq) {
2362                 rc = -ENOMEM;
2363                 goto free_iv;
2364         }
2365
2366         if (!*aead) {
2367                 *aead = crypto_alloc_aead(cipher_name, 0, 0);
2368                 if (IS_ERR(*aead)) {
2369                         rc = PTR_ERR(*aead);
2370                         *aead = NULL;
2371                         goto free_rec_seq;
2372                 }
2373         }
2374
2375         ctx->push_pending_record = tls_sw_push_pending_record;
2376
2377         rc = crypto_aead_setkey(*aead, key, keysize);
2378
2379         if (rc)
2380                 goto free_aead;
2381
2382         rc = crypto_aead_setauthsize(*aead, prot->tag_size);
2383         if (rc)
2384                 goto free_aead;
2385
2386         if (sw_ctx_rx) {
2387                 tfm = crypto_aead_tfm(sw_ctx_rx->aead_recv);
2388
2389                 if (crypto_info->version == TLS_1_3_VERSION)
2390                         sw_ctx_rx->async_capable = false;
2391                 else
2392                         sw_ctx_rx->async_capable =
2393                                 tfm->__crt_alg->cra_flags & CRYPTO_ALG_ASYNC;
2394
2395                 /* Set up strparser */
2396                 memset(&cb, 0, sizeof(cb));
2397                 cb.rcv_msg = tls_queue;
2398                 cb.parse_msg = tls_read_size;
2399
2400                 strp_init(&sw_ctx_rx->strp, sk, &cb);
2401         }
2402
2403         goto out;
2404
2405 free_aead:
2406         crypto_free_aead(*aead);
2407         *aead = NULL;
2408 free_rec_seq:
2409         kfree(cctx->rec_seq);
2410         cctx->rec_seq = NULL;
2411 free_iv:
2412         kfree(cctx->iv);
2413         cctx->iv = NULL;
2414 free_priv:
2415         if (tx) {
2416                 kfree(ctx->priv_ctx_tx);
2417                 ctx->priv_ctx_tx = NULL;
2418         } else {
2419                 kfree(ctx->priv_ctx_rx);
2420                 ctx->priv_ctx_rx = NULL;
2421         }
2422 out:
2423         return rc;
2424 }