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