net/tls/tls_sw.c

   1 /*
   2  * Copyright (c) 2016-2017, Mellanox Technologies. All rights reserved.
   3  * Copyright (c) 2016-2017, Dave Watson <davejwatson@fb.com>. All rights reserved.
   4  * Copyright (c) 2016-2017, Lance Chao <lancerchao@fb.com>. All rights reserved.
   5  * Copyright (c) 2016, Fridolin Pokorny <fridolin.pokorny@gmail.com>. All rights reserved.
   6  * Copyright (c) 2016, Nikos Mavrogiannopoulos <nmav@gnutls.org>. All rights reserved.
   7  * Copyright (c) 2018, Covalent IO, Inc. http://covalent.io
   8  *
   9  * This software is available to you under a choice of one of two
  10  * licenses.  You may choose to be licensed under the terms of the GNU
  11  * General Public License (GPL) Version 2, available from the file
  12  * COPYING in the main directory of this source tree, or the
  13  * OpenIB.org BSD license below:
  14  *
  15  *     Redistribution and use in source and binary forms, with or
  16  *     without modification, are permitted provided that the following
  17  *     conditions are met:
  18  *
  19  *      - Redistributions of source code must retain the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer.
  22  *
  23  *      - Redistributions in binary form must reproduce the above
  24  *        copyright notice, this list of conditions and the following
  25  *        disclaimer in the documentation and/or other materials
  26  *        provided with the distribution.
  27  *
  28  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  29  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  30  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  31  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  32  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  33  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  34  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  35  * SOFTWARE.
  36  */
  37
  38 #include <linux/sched/signal.h>
  39 #include <linux/module.h>
  40 #include <crypto/aead.h>
  41
  42 #include <net/strparser.h>
  43 #include <net/tls.h>
  44
  45 static int __skb_nsg(struct sk_buff *skb, int offset, int len,
  46                      unsigned int recursion_level)
  47 {
  48         int start = skb_headlen(skb);
  49         int i, chunk = start - offset;
  50         struct sk_buff *frag_iter;
  51         int elt = 0;
  52
  53         if (unlikely(recursion_level >= 24))
  54                 return -EMSGSIZE;
  55
  56         if (chunk > 0) {
  57                 if (chunk > len)
  58                         chunk = len;
  59                 elt++;
  60                 len -= chunk;
  61                 if (len == 0)
  62                         return elt;
  63                 offset += chunk;
  64         }
  65
  66         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
  67                 int end;
  68
  69                 WARN_ON(start > offset + len);
  70
  71                 end = start + skb_frag_size(&skb_shinfo(skb)->frags[i]);
  72                 chunk = end - offset;
  73                 if (chunk > 0) {
  74                         if (chunk > len)
  75                                 chunk = len;
  76                         elt++;
  77                         len -= chunk;
  78                         if (len == 0)
  79                                 return elt;
  80                         offset += chunk;
  81                 }
  82                 start = end;
  83         }
  84
  85         if (unlikely(skb_has_frag_list(skb))) {
  86                 skb_walk_frags(skb, frag_iter) {
  87                         int end, ret;
  88
  89                         WARN_ON(start > offset + len);
  90
  91                         end = start + frag_iter->len;
  92                         chunk = end - offset;
  93                         if (chunk > 0) {
  94                                 if (chunk > len)
  95                                         chunk = len;
  96                                 ret = __skb_nsg(frag_iter, offset - start, chunk,
  97                                                 recursion_level + 1);
  98                                 if (unlikely(ret < 0))
  99                                         return ret;
 100                                 elt += ret;
 101                                 len -= chunk;
 102                                 if (len == 0)
 103                                         return elt;
 104                                 offset += chunk;
 105                         }
 106                         start = end;
 107                 }
 108         }
 109         BUG_ON(len);
 110         return elt;
 111 }
 112
 113 /* Return the number of scatterlist elements required to completely map the
 114  * skb, or -EMSGSIZE if the recursion depth is exceeded.
 115  */
 116 static int skb_nsg(struct sk_buff *skb, int offset, int len)
 117 {
 118         return __skb_nsg(skb, offset, len, 0);
 119 }
 120
 121 static int padding_length(struct tls_sw_context_rx *ctx,
 122                           struct tls_prot_info *prot, struct sk_buff *skb)
 123 {
 124         struct strp_msg *rxm = strp_msg(skb);
 125         int sub = 0;
 126
 127         /* Determine zero-padding length */
 128         if (prot->version == TLS_1_3_VERSION) {
 129                 char content_type = 0;
 130                 int err;
 131                 int back = 17;
 132
 133                 while (content_type == 0) {
 134                         if (back > rxm->full_len - prot->prepend_size)
 135                                 return -EBADMSG;
 136                         err = skb_copy_bits(skb,
 137                                             rxm->offset + rxm->full_len - back,
 138                                             &content_type, 1);
 139                         if (err)
 140                                 return err;
 141                         if (content_type)
 142                                 break;
 143                         sub++;
 144                         back++;
 145                 }
 146                 ctx->control = content_type;
 147         }
 148         return sub;
 149 }
 150
 151 static void tls_decrypt_done(struct crypto_async_request *req, int err)
 152 {
 153         struct aead_request *aead_req = (struct aead_request *)req;
 154         struct scatterlist *sgout = aead_req->dst;
 155         struct scatterlist *sgin = aead_req->src;
 156         struct tls_sw_context_rx *ctx;
 157         struct tls_context *tls_ctx;
 158         struct tls_prot_info *prot;
 159         struct scatterlist *sg;
 160         struct sk_buff *skb;
 161         unsigned int pages;
 162         int pending;
 163
 164         skb = (struct sk_buff *)req->data;
 165         tls_ctx = tls_get_ctx(skb->sk);
 166         ctx = tls_sw_ctx_rx(tls_ctx);
 167         prot = &tls_ctx->prot_info;
 168
 169         /* Propagate if there was an err */
 170         if (err) {
 171                 if (err == -EBADMSG)
 172                         TLS_INC_STATS(sock_net(skb->sk),
 173                                       LINUX_MIB_TLSDECRYPTERROR);
 174                 ctx->async_wait.err = err;
 175                 tls_err_abort(skb->sk, err);
 176         } else {
 177                 struct strp_msg *rxm = strp_msg(skb);
 178                 int pad;
 179
 180                 pad = padding_length(ctx, prot, skb);
 181                 if (pad < 0) {
 182                         ctx->async_wait.err = pad;
 183                         tls_err_abort(skb->sk, pad);
 184                 } else {
 185                         rxm->full_len -= pad;
 186                         rxm->offset += prot->prepend_size;
 187                         rxm->full_len -= prot->overhead_size;
 188                 }
 189         }
 190
 191         /* After using skb->sk to propagate sk through crypto async callback
 192          * we need to NULL it again.
 193          */
 194         skb->sk = NULL;
 195
 196
 197         /* Free the destination pages if skb was not decrypted inplace */
 198         if (sgout != sgin) {
 199                 /* Skip the first S/G entry as it points to AAD */
 200                 for_each_sg(sg_next(sgout), sg, UINT_MAX, pages) {
 201                         if (!sg)
 202                                 break;
 203                         put_page(sg_page(sg));
 204                 }
 205         }
 206
 207         kfree(aead_req);
 208
 209         pending = atomic_dec_return(&ctx->decrypt_pending);
 210
 211         if (!pending && READ_ONCE(ctx->async_notify))
 212                 complete(&ctx->async_wait.completion);
 213 }
 214
 215 static int tls_do_decryption(struct sock *sk,
 216                              struct sk_buff *skb,
 217                              struct scatterlist *sgin,
 218                              struct scatterlist *sgout,
 219                              char *iv_recv,
 220                              size_t data_len,
 221                              struct aead_request *aead_req,
 222                              bool async)
 223 {
 224         struct tls_context *tls_ctx = tls_get_ctx(sk);
 225         struct tls_prot_info *prot = &tls_ctx->prot_info;
 226         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
 227         int ret;
 228
 229         aead_request_set_tfm(aead_req, ctx->aead_recv);
 230         aead_request_set_ad(aead_req, prot->aad_size);
 231         aead_request_set_crypt(aead_req, sgin, sgout,
 232                                data_len + prot->tag_size,
 233                                (u8 *)iv_recv);
 234
 235         if (async) {
 236                 /* Using skb->sk to push sk through to crypto async callback
 237                  * handler. This allows propagating errors up to the socket
 238                  * if needed. It _must_ be cleared in the async handler
 239                  * before consume_skb is called. We _know_ skb->sk is NULL
 240                  * because it is a clone from strparser.
 241                  */
 242                 skb->sk = sk;
 243                 aead_request_set_callback(aead_req,
 244                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 245                                           tls_decrypt_done, skb);
 246                 atomic_inc(&ctx->decrypt_pending);
 247         } else {
 248                 aead_request_set_callback(aead_req,
 249                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 250                                           crypto_req_done, &ctx->async_wait);
 251         }
 252
 253         ret = crypto_aead_decrypt(aead_req);
 254         if (ret == -EINPROGRESS) {
 255                 if (async)
 256                         return ret;
 257
 258                 ret = crypto_wait_req(ret, &ctx->async_wait);
 259         }
 260
 261         if (async)
 262                 atomic_dec(&ctx->decrypt_pending);
 263
 264         return ret;
 265 }
 266
 267 static void tls_trim_both_msgs(struct sock *sk, int target_size)
 268 {
 269         struct tls_context *tls_ctx = tls_get_ctx(sk);
 270         struct tls_prot_info *prot = &tls_ctx->prot_info;
 271         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 272         struct tls_rec *rec = ctx->open_rec;
 273
 274         sk_msg_trim(sk, &rec->msg_plaintext, target_size);
 275         if (target_size > 0)
 276                 target_size += prot->overhead_size;
 277         sk_msg_trim(sk, &rec->msg_encrypted, target_size);
 278 }
 279
 280 static int tls_alloc_encrypted_msg(struct sock *sk, int len)
 281 {
 282         struct tls_context *tls_ctx = tls_get_ctx(sk);
 283         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 284         struct tls_rec *rec = ctx->open_rec;
 285         struct sk_msg *msg_en = &rec->msg_encrypted;
 286
 287         return sk_msg_alloc(sk, msg_en, len, 0);
 288 }
 289
 290 static int tls_clone_plaintext_msg(struct sock *sk, int required)
 291 {
 292         struct tls_context *tls_ctx = tls_get_ctx(sk);
 293         struct tls_prot_info *prot = &tls_ctx->prot_info;
 294         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 295         struct tls_rec *rec = ctx->open_rec;
 296         struct sk_msg *msg_pl = &rec->msg_plaintext;
 297         struct sk_msg *msg_en = &rec->msg_encrypted;
 298         int skip, len;
 299
 300         /* We add page references worth len bytes from encrypted sg
 301          * at the end of plaintext sg. It is guaranteed that msg_en
 302          * has enough required room (ensured by caller).
 303          */
 304         len = required - msg_pl->sg.size;
 305
 306         /* Skip initial bytes in msg_en's data to be able to use
 307          * same offset of both plain and encrypted data.
 308          */
 309         skip = prot->prepend_size + msg_pl->sg.size;
 310
 311         return sk_msg_clone(sk, msg_pl, msg_en, skip, len);
 312 }
 313
 314 static struct tls_rec *tls_get_rec(struct sock *sk)
 315 {
 316         struct tls_context *tls_ctx = tls_get_ctx(sk);
 317         struct tls_prot_info *prot = &tls_ctx->prot_info;
 318         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 319         struct sk_msg *msg_pl, *msg_en;
 320         struct tls_rec *rec;
 321         int mem_size;
 322
 323         mem_size = sizeof(struct tls_rec) + crypto_aead_reqsize(ctx->aead_send);
 324
 325         rec = kzalloc(mem_size, sk->sk_allocation);
 326         if (!rec)
 327                 return NULL;
 328
 329         msg_pl = &rec->msg_plaintext;
 330         msg_en = &rec->msg_encrypted;
 331
 332         sk_msg_init(msg_pl);
 333         sk_msg_init(msg_en);
 334
 335         sg_init_table(rec->sg_aead_in, 2);
 336         sg_set_buf(&rec->sg_aead_in[0], rec->aad_space, prot->aad_size);
 337         sg_unmark_end(&rec->sg_aead_in[1]);
 338
 339         sg_init_table(rec->sg_aead_out, 2);
 340         sg_set_buf(&rec->sg_aead_out[0], rec->aad_space, prot->aad_size);
 341         sg_unmark_end(&rec->sg_aead_out[1]);
 342
 343         return rec;
 344 }
 345
 346 static void tls_free_rec(struct sock *sk, struct tls_rec *rec)
 347 {
 348         sk_msg_free(sk, &rec->msg_encrypted);
 349         sk_msg_free(sk, &rec->msg_plaintext);
 350         kfree(rec);
 351 }
 352
 353 static void tls_free_open_rec(struct sock *sk)
 354 {
 355         struct tls_context *tls_ctx = tls_get_ctx(sk);
 356         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 357         struct tls_rec *rec = ctx->open_rec;
 358
 359         if (rec) {
 360                 tls_free_rec(sk, rec);
 361                 ctx->open_rec = NULL;
 362         }
 363 }
 364
 365 int tls_tx_records(struct sock *sk, int flags)
 366 {
 367         struct tls_context *tls_ctx = tls_get_ctx(sk);
 368         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 369         struct tls_rec *rec, *tmp;
 370         struct sk_msg *msg_en;
 371         int tx_flags, rc = 0;
 372
 373         if (tls_is_partially_sent_record(tls_ctx)) {
 374                 rec = list_first_entry(&ctx->tx_list,
 375                                        struct tls_rec, list);
 376
 377                 if (flags == -1)
 378                         tx_flags = rec->tx_flags;
 379                 else
 380                         tx_flags = flags;
 381
 382                 rc = tls_push_partial_record(sk, tls_ctx, tx_flags);
 383                 if (rc)
 384                         goto tx_err;
 385
 386                 /* Full record has been transmitted.
 387                  * Remove the head of tx_list
 388                  */
 389                 list_del(&rec->list);
 390                 sk_msg_free(sk, &rec->msg_plaintext);
 391                 kfree(rec);
 392         }
 393
 394         /* Tx all ready records */
 395         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
 396                 if (READ_ONCE(rec->tx_ready)) {
 397                         if (flags == -1)
 398                                 tx_flags = rec->tx_flags;
 399                         else
 400                                 tx_flags = flags;
 401
 402                         msg_en = &rec->msg_encrypted;
 403                         rc = tls_push_sg(sk, tls_ctx,
 404                                          &msg_en->sg.data[msg_en->sg.curr],
 405                                          0, tx_flags);
 406                         if (rc)
 407                                 goto tx_err;
 408
 409                         list_del(&rec->list);
 410                         sk_msg_free(sk, &rec->msg_plaintext);
 411                         kfree(rec);
 412                 } else {
 413                         break;
 414                 }
 415         }
 416
 417 tx_err:
 418         if (rc < 0 && rc != -EAGAIN)
 419                 tls_err_abort(sk, EBADMSG);
 420
 421         return rc;
 422 }
 423
 424 static void tls_encrypt_done(struct crypto_async_request *req, int err)
 425 {
 426         struct aead_request *aead_req = (struct aead_request *)req;
 427         struct sock *sk = req->data;
 428         struct tls_context *tls_ctx = tls_get_ctx(sk);
 429         struct tls_prot_info *prot = &tls_ctx->prot_info;
 430         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 431         struct scatterlist *sge;
 432         struct sk_msg *msg_en;
 433         struct tls_rec *rec;
 434         bool ready = false;
 435         int pending;
 436
 437         rec = container_of(aead_req, struct tls_rec, aead_req);
 438         msg_en = &rec->msg_encrypted;
 439
 440         sge = sk_msg_elem(msg_en, msg_en->sg.curr);
 441         sge->offset -= prot->prepend_size;
 442         sge->length += prot->prepend_size;
 443
 444         /* Check if error is previously set on socket */
 445         if (err || sk->sk_err) {
 446                 rec = NULL;
 447
 448                 /* If err is already set on socket, return the same code */
 449                 if (sk->sk_err) {
 450                         ctx->async_wait.err = sk->sk_err;
 451                 } else {
 452                         ctx->async_wait.err = err;
 453                         tls_err_abort(sk, err);
 454                 }
 455         }
 456
 457         if (rec) {
 458                 struct tls_rec *first_rec;
 459
 460                 /* Mark the record as ready for transmission */
 461                 smp_store_mb(rec->tx_ready, true);
 462
 463                 /* If received record is at head of tx_list, schedule tx */
 464                 first_rec = list_first_entry(&ctx->tx_list,
 465                                              struct tls_rec, list);
 466                 if (rec == first_rec)
 467                         ready = true;
 468         }
 469
 470         pending = atomic_dec_return(&ctx->encrypt_pending);
 471
 472         if (!pending && READ_ONCE(ctx->async_notify))
 473                 complete(&ctx->async_wait.completion);
 474
 475         if (!ready)
 476                 return;
 477
 478         /* Schedule the transmission */
 479         if (!test_and_set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
 480                 schedule_delayed_work(&ctx->tx_work.work, 1);
 481 }
 482
 483 static int tls_do_encryption(struct sock *sk,
 484                              struct tls_context *tls_ctx,
 485                              struct tls_sw_context_tx *ctx,
 486                              struct aead_request *aead_req,
 487                              size_t data_len, u32 start)
 488 {
 489         struct tls_prot_info *prot = &tls_ctx->prot_info;
 490         struct tls_rec *rec = ctx->open_rec;
 491         struct sk_msg *msg_en = &rec->msg_encrypted;
 492         struct scatterlist *sge = sk_msg_elem(msg_en, start);
 493         int rc, iv_offset = 0;
 494
 495         /* For CCM based ciphers, first byte of IV is a constant */
 496         if (prot->cipher_type == TLS_CIPHER_AES_CCM_128) {
 497                 rec->iv_data[0] = TLS_AES_CCM_IV_B0_BYTE;
 498                 iv_offset = 1;
 499         }
 500
 501         memcpy(&rec->iv_data[iv_offset], tls_ctx->tx.iv,
 502                prot->iv_size + prot->salt_size);
 503
 504         xor_iv_with_seq(prot->version, rec->iv_data, tls_ctx->tx.rec_seq);
 505
 506         sge->offset += prot->prepend_size;
 507         sge->length -= prot->prepend_size;
 508
 509         msg_en->sg.curr = start;
 510
 511         aead_request_set_tfm(aead_req, ctx->aead_send);
 512         aead_request_set_ad(aead_req, prot->aad_size);
 513         aead_request_set_crypt(aead_req, rec->sg_aead_in,
 514                                rec->sg_aead_out,
 515                                data_len, rec->iv_data);
 516
 517         aead_request_set_callback(aead_req, CRYPTO_TFM_REQ_MAY_BACKLOG,
 518                                   tls_encrypt_done, sk);
 519
 520         /* Add the record in tx_list */
 521         list_add_tail((struct list_head *)&rec->list, &ctx->tx_list);
 522         atomic_inc(&ctx->encrypt_pending);
 523
 524         rc = crypto_aead_encrypt(aead_req);
 525         if (!rc || rc != -EINPROGRESS) {
 526                 atomic_dec(&ctx->encrypt_pending);
 527                 sge->offset -= prot->prepend_size;
 528                 sge->length += prot->prepend_size;
 529         }
 530
 531         if (!rc) {
 532                 WRITE_ONCE(rec->tx_ready, true);
 533         } else if (rc != -EINPROGRESS) {
 534                 list_del(&rec->list);
 535                 return rc;
 536         }
 537
 538         /* Unhook the record from context if encryption is not failure */
 539         ctx->open_rec = NULL;
 540         tls_advance_record_sn(sk, prot, &tls_ctx->tx);
 541         return rc;
 542 }
 543
 544 static int tls_split_open_record(struct sock *sk, struct tls_rec *from,
 545                                  struct tls_rec **to, struct sk_msg *msg_opl,
 546                                  struct sk_msg *msg_oen, u32 split_point,
 547                                  u32 tx_overhead_size, u32 *orig_end)
 548 {
 549         u32 i, j, bytes = 0, apply = msg_opl->apply_bytes;
 550         struct scatterlist *sge, *osge, *nsge;
 551         u32 orig_size = msg_opl->sg.size;
 552         struct scatterlist tmp = { };
 553         struct sk_msg *msg_npl;
 554         struct tls_rec *new;
 555         int ret;
 556
 557         new = tls_get_rec(sk);
 558         if (!new)
 559                 return -ENOMEM;
 560         ret = sk_msg_alloc(sk, &new->msg_encrypted, msg_opl->sg.size +
 561                            tx_overhead_size, 0);
 562         if (ret < 0) {
 563                 tls_free_rec(sk, new);
 564                 return ret;
 565         }
 566
 567         *orig_end = msg_opl->sg.end;
 568         i = msg_opl->sg.start;
 569         sge = sk_msg_elem(msg_opl, i);
 570         while (apply && sge->length) {
 571                 if (sge->length > apply) {
 572                         u32 len = sge->length - apply;
 573
 574                         get_page(sg_page(sge));
 575                         sg_set_page(&tmp, sg_page(sge), len,
 576                                     sge->offset + apply);
 577                         sge->length = apply;
 578                         bytes += apply;
 579                         apply = 0;
 580                 } else {
 581                         apply -= sge->length;
 582                         bytes += sge->length;
 583                 }
 584
 585                 sk_msg_iter_var_next(i);
 586                 if (i == msg_opl->sg.end)
 587                         break;
 588                 sge = sk_msg_elem(msg_opl, i);
 589         }
 590
 591         msg_opl->sg.end = i;
 592         msg_opl->sg.curr = i;
 593         msg_opl->sg.copybreak = 0;
 594         msg_opl->apply_bytes = 0;
 595         msg_opl->sg.size = bytes;
 596
 597         msg_npl = &new->msg_plaintext;
 598         msg_npl->apply_bytes = apply;
 599         msg_npl->sg.size = orig_size - bytes;
 600
 601         j = msg_npl->sg.start;
 602         nsge = sk_msg_elem(msg_npl, j);
 603         if (tmp.length) {
 604                 memcpy(nsge, &tmp, sizeof(*nsge));
 605                 sk_msg_iter_var_next(j);
 606                 nsge = sk_msg_elem(msg_npl, j);
 607         }
 608
 609         osge = sk_msg_elem(msg_opl, i);
 610         while (osge->length) {
 611                 memcpy(nsge, osge, sizeof(*nsge));
 612                 sg_unmark_end(nsge);
 613                 sk_msg_iter_var_next(i);
 614                 sk_msg_iter_var_next(j);
 615                 if (i == *orig_end)
 616                         break;
 617                 osge = sk_msg_elem(msg_opl, i);
 618                 nsge = sk_msg_elem(msg_npl, j);
 619         }
 620
 621         msg_npl->sg.end = j;
 622         msg_npl->sg.curr = j;
 623         msg_npl->sg.copybreak = 0;
 624
 625         *to = new;
 626         return 0;
 627 }
 628
 629 static void tls_merge_open_record(struct sock *sk, struct tls_rec *to,
 630                                   struct tls_rec *from, u32 orig_end)
 631 {
 632         struct sk_msg *msg_npl = &from->msg_plaintext;
 633         struct sk_msg *msg_opl = &to->msg_plaintext;
 634         struct scatterlist *osge, *nsge;
 635         u32 i, j;
 636
 637         i = msg_opl->sg.end;
 638         sk_msg_iter_var_prev(i);
 639         j = msg_npl->sg.start;
 640
 641         osge = sk_msg_elem(msg_opl, i);
 642         nsge = sk_msg_elem(msg_npl, j);
 643
 644         if (sg_page(osge) == sg_page(nsge) &&
 645             osge->offset + osge->length == nsge->offset) {
 646                 osge->length += nsge->length;
 647                 put_page(sg_page(nsge));
 648         }
 649
 650         msg_opl->sg.end = orig_end;
 651         msg_opl->sg.curr = orig_end;
 652         msg_opl->sg.copybreak = 0;
 653         msg_opl->apply_bytes = msg_opl->sg.size + msg_npl->sg.size;
 654         msg_opl->sg.size += msg_npl->sg.size;
 655
 656         sk_msg_free(sk, &to->msg_encrypted);
 657         sk_msg_xfer_full(&to->msg_encrypted, &from->msg_encrypted);
 658
 659         kfree(from);
 660 }
 661
 662 static int tls_push_record(struct sock *sk, int flags,
 663                            unsigned char record_type)
 664 {
 665         struct tls_context *tls_ctx = tls_get_ctx(sk);
 666         struct tls_prot_info *prot = &tls_ctx->prot_info;
 667         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 668         struct tls_rec *rec = ctx->open_rec, *tmp = NULL;
 669         u32 i, split_point, uninitialized_var(orig_end);
 670         struct sk_msg *msg_pl, *msg_en;
 671         struct aead_request *req;
 672         bool split;
 673         int rc;
 674
 675         if (!rec)
 676                 return 0;
 677
 678         msg_pl = &rec->msg_plaintext;
 679         msg_en = &rec->msg_encrypted;
 680
 681         split_point = msg_pl->apply_bytes;
 682         split = split_point && split_point < msg_pl->sg.size;
 683         if (unlikely((!split &&
 684                       msg_pl->sg.size +
 685                       prot->overhead_size > msg_en->sg.size) ||
 686                      (split &&
 687                       split_point +
 688                       prot->overhead_size > msg_en->sg.size))) {
 689                 split = true;
 690                 split_point = msg_en->sg.size;
 691         }
 692         if (split) {
 693                 rc = tls_split_open_record(sk, rec, &tmp, msg_pl, msg_en,
 694                                            split_point, prot->overhead_size,
 695                                            &orig_end);
 696                 if (rc < 0)
 697                         return rc;
 698                 /* This can happen if above tls_split_open_record allocates
 699                  * a single large encryption buffer instead of two smaller
 700                  * ones. In this case adjust pointers and continue without
 701                  * split.
 702                  */
 703                 if (!msg_pl->sg.size) {
 704                         tls_merge_open_record(sk, rec, tmp, orig_end);
 705                         msg_pl = &rec->msg_plaintext;
 706                         msg_en = &rec->msg_encrypted;
 707                         split = false;
 708                 }
 709                 sk_msg_trim(sk, msg_en, msg_pl->sg.size +
 710                             prot->overhead_size);
 711         }
 712
 713         rec->tx_flags = flags;
 714         req = &rec->aead_req;
 715
 716         i = msg_pl->sg.end;
 717         sk_msg_iter_var_prev(i);
 718
 719         rec->content_type = record_type;
 720         if (prot->version == TLS_1_3_VERSION) {
 721                 /* Add content type to end of message.  No padding added */
 722                 sg_set_buf(&rec->sg_content_type, &rec->content_type, 1);
 723                 sg_mark_end(&rec->sg_content_type);
 724                 sg_chain(msg_pl->sg.data, msg_pl->sg.end + 1,
 725                          &rec->sg_content_type);
 726         } else {
 727                 sg_mark_end(sk_msg_elem(msg_pl, i));
 728         }
 729
 730         if (msg_pl->sg.end < msg_pl->sg.start) {
 731                 sg_chain(&msg_pl->sg.data[msg_pl->sg.start],
 732                          MAX_SKB_FRAGS - msg_pl->sg.start + 1,
 733                          msg_pl->sg.data);
 734         }
 735
 736         i = msg_pl->sg.start;
 737         sg_chain(rec->sg_aead_in, 2, &msg_pl->sg.data[i]);
 738
 739         i = msg_en->sg.end;
 740         sk_msg_iter_var_prev(i);
 741         sg_mark_end(sk_msg_elem(msg_en, i));
 742
 743         i = msg_en->sg.start;
 744         sg_chain(rec->sg_aead_out, 2, &msg_en->sg.data[i]);
 745
 746         tls_make_aad(rec->aad_space, msg_pl->sg.size + prot->tail_size,
 747                      tls_ctx->tx.rec_seq, prot->rec_seq_size,
 748                      record_type, prot->version);
 749
 750         tls_fill_prepend(tls_ctx,
 751                          page_address(sg_page(&msg_en->sg.data[i])) +
 752                          msg_en->sg.data[i].offset,
 753                          msg_pl->sg.size + prot->tail_size,
 754                          record_type, prot->version);
 755
 756         tls_ctx->pending_open_record_frags = false;
 757
 758         rc = tls_do_encryption(sk, tls_ctx, ctx, req,
 759                                msg_pl->sg.size + prot->tail_size, i);
 760         if (rc < 0) {
 761                 if (rc != -EINPROGRESS) {
 762                         tls_err_abort(sk, EBADMSG);
 763                         if (split) {
 764                                 tls_ctx->pending_open_record_frags = true;
 765                                 tls_merge_open_record(sk, rec, tmp, orig_end);
 766                         }
 767                 }
 768                 ctx->async_capable = 1;
 769                 return rc;
 770         } else if (split) {
 771                 msg_pl = &tmp->msg_plaintext;
 772                 msg_en = &tmp->msg_encrypted;
 773                 sk_msg_trim(sk, msg_en, msg_pl->sg.size + prot->overhead_size);
 774                 tls_ctx->pending_open_record_frags = true;
 775                 ctx->open_rec = tmp;
 776         }
 777
 778         return tls_tx_records(sk, flags);
 779 }
 780
 781 static int bpf_exec_tx_verdict(struct sk_msg *msg, struct sock *sk,
 782                                bool full_record, u8 record_type,
 783                                size_t *copied, int flags)
 784 {
 785         struct tls_context *tls_ctx = tls_get_ctx(sk);
 786         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 787         struct sk_msg msg_redir = { };
 788         struct sk_psock *psock;
 789         struct sock *sk_redir;
 790         struct tls_rec *rec;
 791         bool enospc, policy;
 792         int err = 0, send;
 793         u32 delta = 0;
 794
 795         policy = !(flags & MSG_SENDPAGE_NOPOLICY);
 796         psock = sk_psock_get(sk);
 797         if (!psock || !policy) {
 798                 err = tls_push_record(sk, flags, record_type);
 799                 if (err && err != -EINPROGRESS) {
 800                         *copied -= sk_msg_free(sk, msg);
 801                         tls_free_open_rec(sk);
 802                 }
 803                 if (psock)
 804                         sk_psock_put(sk, psock);
 805                 return err;
 806         }
 807 more_data:
 808         enospc = sk_msg_full(msg);
 809         if (psock->eval == __SK_NONE) {
 810                 delta = msg->sg.size;
 811                 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
 812                 delta -= msg->sg.size;
 813         }
 814         if (msg->cork_bytes && msg->cork_bytes > msg->sg.size &&
 815             !enospc && !full_record) {
 816                 err = -ENOSPC;
 817                 goto out_err;
 818         }
 819         msg->cork_bytes = 0;
 820         send = msg->sg.size;
 821         if (msg->apply_bytes && msg->apply_bytes < send)
 822                 send = msg->apply_bytes;
 823
 824         switch (psock->eval) {
 825         case __SK_PASS:
 826                 err = tls_push_record(sk, flags, record_type);
 827                 if (err && err != -EINPROGRESS) {
 828                         *copied -= sk_msg_free(sk, msg);
 829                         tls_free_open_rec(sk);
 830                         goto out_err;
 831                 }
 832                 break;
 833         case __SK_REDIRECT:
 834                 sk_redir = psock->sk_redir;
 835                 memcpy(&msg_redir, msg, sizeof(*msg));
 836                 if (msg->apply_bytes < send)
 837                         msg->apply_bytes = 0;
 838                 else
 839                         msg->apply_bytes -= send;
 840                 sk_msg_return_zero(sk, msg, send);
 841                 msg->sg.size -= send;
 842                 release_sock(sk);
 843                 err = tcp_bpf_sendmsg_redir(sk_redir, &msg_redir, send, flags);
 844                 lock_sock(sk);
 845                 if (err < 0) {
 846                         *copied -= sk_msg_free_nocharge(sk, &msg_redir);
 847                         msg->sg.size = 0;
 848                 }
 849                 if (msg->sg.size == 0)
 850                         tls_free_open_rec(sk);
 851                 break;
 852         case __SK_DROP:
 853         default:
 854                 sk_msg_free_partial(sk, msg, send);
 855                 if (msg->apply_bytes < send)
 856                         msg->apply_bytes = 0;
 857                 else
 858                         msg->apply_bytes -= send;
 859                 if (msg->sg.size == 0)
 860                         tls_free_open_rec(sk);
 861                 *copied -= (send + delta);
 862                 err = -EACCES;
 863         }
 864
 865         if (likely(!err)) {
 866                 bool reset_eval = !ctx->open_rec;
 867
 868                 rec = ctx->open_rec;
 869                 if (rec) {
 870                         msg = &rec->msg_plaintext;
 871                         if (!msg->apply_bytes)
 872                                 reset_eval = true;
 873                 }
 874                 if (reset_eval) {
 875                         psock->eval = __SK_NONE;
 876                         if (psock->sk_redir) {
 877                                 sock_put(psock->sk_redir);
 878                                 psock->sk_redir = NULL;
 879                         }
 880                 }
 881                 if (rec)
 882                         goto more_data;
 883         }
 884  out_err:
 885         sk_psock_put(sk, psock);
 886         return err;
 887 }
 888
 889 static int tls_sw_push_pending_record(struct sock *sk, int flags)
 890 {
 891         struct tls_context *tls_ctx = tls_get_ctx(sk);
 892         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 893         struct tls_rec *rec = ctx->open_rec;
 894         struct sk_msg *msg_pl;
 895         size_t copied;
 896
 897         if (!rec)
 898                 return 0;
 899
 900         msg_pl = &rec->msg_plaintext;
 901         copied = msg_pl->sg.size;
 902         if (!copied)
 903                 return 0;
 904
 905         return bpf_exec_tx_verdict(msg_pl, sk, true, TLS_RECORD_TYPE_DATA,
 906                                    &copied, flags);
 907 }
 908
 909 int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 910 {
 911         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 912         struct tls_context *tls_ctx = tls_get_ctx(sk);
 913         struct tls_prot_info *prot = &tls_ctx->prot_info;
 914         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 915         bool async_capable = ctx->async_capable;
 916         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 917         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
 918         bool eor = !(msg->msg_flags & MSG_MORE);
 919         size_t try_to_copy, copied = 0;
 920         struct sk_msg *msg_pl, *msg_en;
 921         struct tls_rec *rec;
 922         int required_size;
 923         int num_async = 0;
 924         bool full_record;
 925         int record_room;
 926         int num_zc = 0;
 927         int orig_size;
 928         int ret = 0;
 929
 930         if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
 931                 return -EOPNOTSUPP;
 932
 933         mutex_lock(&tls_ctx->tx_lock);
 934         lock_sock(sk);
 935
 936         if (unlikely(msg->msg_controllen)) {
 937                 ret = tls_proccess_cmsg(sk, msg, &record_type);
 938                 if (ret) {
 939                         if (ret == -EINPROGRESS)
 940                                 num_async++;
 941                         else if (ret != -EAGAIN)
 942                                 goto send_end;
 943                 }
 944         }
 945
 946         while (msg_data_left(msg)) {
 947                 if (sk->sk_err) {
 948                         ret = -sk->sk_err;
 949                         goto send_end;
 950                 }
 951
 952                 if (ctx->open_rec)
 953                         rec = ctx->open_rec;
 954                 else
 955                         rec = ctx->open_rec = tls_get_rec(sk);
 956                 if (!rec) {
 957                         ret = -ENOMEM;
 958                         goto send_end;
 959                 }
 960
 961                 msg_pl = &rec->msg_plaintext;
 962                 msg_en = &rec->msg_encrypted;
 963
 964                 orig_size = msg_pl->sg.size;
 965                 full_record = false;
 966                 try_to_copy = msg_data_left(msg);
 967                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
 968                 if (try_to_copy >= record_room) {
 969                         try_to_copy = record_room;
 970                         full_record = true;
 971                 }
 972
 973                 required_size = msg_pl->sg.size + try_to_copy +
 974                                 prot->overhead_size;
 975
 976                 if (!sk_stream_memory_free(sk))
 977                         goto wait_for_sndbuf;
 978
 979 alloc_encrypted:
 980                 ret = tls_alloc_encrypted_msg(sk, required_size);
 981                 if (ret) {
 982                         if (ret != -ENOSPC)
 983                                 goto wait_for_memory;
 984
 985                         /* Adjust try_to_copy according to the amount that was
 986                          * actually allocated. The difference is due
 987                          * to max sg elements limit
 988                          */
 989                         try_to_copy -= required_size - msg_en->sg.size;
 990                         full_record = true;
 991                 }
 992
 993                 if (!is_kvec && (full_record || eor) && !async_capable) {
 994                         u32 first = msg_pl->sg.end;
 995
 996                         ret = sk_msg_zerocopy_from_iter(sk, &msg->msg_iter,
 997                                                         msg_pl, try_to_copy);
 998                         if (ret)
 999                                 goto fallback_to_reg_send;
1000
1001                         num_zc++;
1002                         copied += try_to_copy;
1003
1004                         sk_msg_sg_copy_set(msg_pl, first);
1005                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1006                                                   record_type, &copied,
1007                                                   msg->msg_flags);
1008                         if (ret) {
1009                                 if (ret == -EINPROGRESS)
1010                                         num_async++;
1011                                 else if (ret == -ENOMEM)
1012                                         goto wait_for_memory;
1013                                 else if (ctx->open_rec && ret == -ENOSPC)
1014                                         goto rollback_iter;
1015                                 else if (ret != -EAGAIN)
1016                                         goto send_end;
1017                         }
1018                         continue;
1019 rollback_iter:
1020                         copied -= try_to_copy;
1021                         sk_msg_sg_copy_clear(msg_pl, first);
1022                         iov_iter_revert(&msg->msg_iter,
1023                                         msg_pl->sg.size - orig_size);
1024 fallback_to_reg_send:
1025                         sk_msg_trim(sk, msg_pl, orig_size);
1026                 }
1027
1028                 required_size = msg_pl->sg.size + try_to_copy;
1029
1030                 ret = tls_clone_plaintext_msg(sk, required_size);
1031                 if (ret) {
1032                         if (ret != -ENOSPC)
1033                                 goto send_end;
1034
1035                         /* Adjust try_to_copy according to the amount that was
1036                          * actually allocated. The difference is due
1037                          * to max sg elements limit
1038                          */
1039                         try_to_copy -= required_size - msg_pl->sg.size;
1040                         full_record = true;
1041                         sk_msg_trim(sk, msg_en,
1042                                     msg_pl->sg.size + prot->overhead_size);
1043                 }
1044
1045                 if (try_to_copy) {
1046                         ret = sk_msg_memcopy_from_iter(sk, &msg->msg_iter,
1047                                                        msg_pl, try_to_copy);
1048                         if (ret < 0)
1049                                 goto trim_sgl;
1050                 }
1051
1052                 /* Open records defined only if successfully copied, otherwise
1053                  * we would trim the sg but not reset the open record frags.
1054                  */
1055                 tls_ctx->pending_open_record_frags = true;
1056                 copied += try_to_copy;
1057                 if (full_record || eor) {
1058                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1059                                                   record_type, &copied,
1060                                                   msg->msg_flags);
1061                         if (ret) {
1062                                 if (ret == -EINPROGRESS)
1063                                         num_async++;
1064                                 else if (ret == -ENOMEM)
1065                                         goto wait_for_memory;
1066                                 else if (ret != -EAGAIN) {
1067                                         if (ret == -ENOSPC)
1068                                                 ret = 0;
1069                                         goto send_end;
1070                                 }
1071                         }
1072                 }
1073
1074                 continue;
1075
1076 wait_for_sndbuf:
1077                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1078 wait_for_memory:
1079                 ret = sk_stream_wait_memory(sk, &timeo);
1080                 if (ret) {
1081 trim_sgl:
1082                         if (ctx->open_rec)
1083                                 tls_trim_both_msgs(sk, orig_size);
1084                         goto send_end;
1085                 }
1086
1087                 if (ctx->open_rec && msg_en->sg.size < required_size)
1088                         goto alloc_encrypted;
1089         }
1090
1091         if (!num_async) {
1092                 goto send_end;
1093         } else if (num_zc) {
1094                 /* Wait for pending encryptions to get completed */
1095                 smp_store_mb(ctx->async_notify, true);
1096
1097                 if (atomic_read(&ctx->encrypt_pending))
1098                         crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1099                 else
1100                         reinit_completion(&ctx->async_wait.completion);
1101
1102                 WRITE_ONCE(ctx->async_notify, false);
1103
1104                 if (ctx->async_wait.err) {
1105                         ret = ctx->async_wait.err;
1106                         copied = 0;
1107                 }
1108         }
1109
1110         /* Transmit if any encryptions have completed */
1111         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1112                 cancel_delayed_work(&ctx->tx_work.work);
1113                 tls_tx_records(sk, msg->msg_flags);
1114         }
1115
1116 send_end:
1117         ret = sk_stream_error(sk, msg->msg_flags, ret);
1118
1119         release_sock(sk);
1120         mutex_unlock(&tls_ctx->tx_lock);
1121         return copied ? copied : ret;
1122 }
1123
1124 static int tls_sw_do_sendpage(struct sock *sk, struct page *page,
1125                               int offset, size_t size, int flags)
1126 {
1127         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1128         struct tls_context *tls_ctx = tls_get_ctx(sk);
1129         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1130         struct tls_prot_info *prot = &tls_ctx->prot_info;
1131         unsigned char record_type = TLS_RECORD_TYPE_DATA;
1132         struct sk_msg *msg_pl;
1133         struct tls_rec *rec;
1134         int num_async = 0;
1135         size_t copied = 0;
1136         bool full_record;
1137         int record_room;
1138         int ret = 0;
1139         bool eor;
1140
1141         eor = !(flags & (MSG_MORE | MSG_SENDPAGE_NOTLAST));
1142         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1143
1144         /* Call the sk_stream functions to manage the sndbuf mem. */
1145         while (size > 0) {
1146                 size_t copy, required_size;
1147
1148                 if (sk->sk_err) {
1149                         ret = -sk->sk_err;
1150                         goto sendpage_end;
1151                 }
1152
1153                 if (ctx->open_rec)
1154                         rec = ctx->open_rec;
1155                 else
1156                         rec = ctx->open_rec = tls_get_rec(sk);
1157                 if (!rec) {
1158                         ret = -ENOMEM;
1159                         goto sendpage_end;
1160                 }
1161
1162                 msg_pl = &rec->msg_plaintext;
1163
1164                 full_record = false;
1165                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
1166                 copy = size;
1167                 if (copy >= record_room) {
1168                         copy = record_room;
1169                         full_record = true;
1170                 }
1171
1172                 required_size = msg_pl->sg.size + copy + prot->overhead_size;
1173
1174                 if (!sk_stream_memory_free(sk))
1175                         goto wait_for_sndbuf;
1176 alloc_payload:
1177                 ret = tls_alloc_encrypted_msg(sk, required_size);
1178                 if (ret) {
1179                         if (ret != -ENOSPC)
1180                                 goto wait_for_memory;
1181
1182                         /* Adjust copy according to the amount that was
1183                          * actually allocated. The difference is due
1184                          * to max sg elements limit
1185                          */
1186                         copy -= required_size - msg_pl->sg.size;
1187                         full_record = true;
1188                 }
1189
1190                 sk_msg_page_add(msg_pl, page, copy, offset);
1191                 sk_mem_charge(sk, copy);
1192
1193                 offset += copy;
1194                 size -= copy;
1195                 copied += copy;
1196
1197                 tls_ctx->pending_open_record_frags = true;
1198                 if (full_record || eor || sk_msg_full(msg_pl)) {
1199                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1200                                                   record_type, &copied, flags);
1201                         if (ret) {
1202                                 if (ret == -EINPROGRESS)
1203                                         num_async++;
1204                                 else if (ret == -ENOMEM)
1205                                         goto wait_for_memory;
1206                                 else if (ret != -EAGAIN) {
1207                                         if (ret == -ENOSPC)
1208                                                 ret = 0;
1209                                         goto sendpage_end;
1210                                 }
1211                         }
1212                 }
1213                 continue;
1214 wait_for_sndbuf:
1215                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1216 wait_for_memory:
1217                 ret = sk_stream_wait_memory(sk, &timeo);
1218                 if (ret) {
1219                         if (ctx->open_rec)
1220                                 tls_trim_both_msgs(sk, msg_pl->sg.size);
1221                         goto sendpage_end;
1222                 }
1223
1224                 if (ctx->open_rec)
1225                         goto alloc_payload;
1226         }
1227
1228         if (num_async) {
1229                 /* Transmit if any encryptions have completed */
1230                 if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1231                         cancel_delayed_work(&ctx->tx_work.work);
1232                         tls_tx_records(sk, flags);
1233                 }
1234         }
1235 sendpage_end:
1236         ret = sk_stream_error(sk, flags, ret);
1237         return copied ? copied : ret;
1238 }
1239
1240 int tls_sw_sendpage_locked(struct sock *sk, struct page *page,
1241                            int offset, size_t size, int flags)
1242 {
1243         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1244                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY |
1245                       MSG_NO_SHARED_FRAGS))
1246                 return -EOPNOTSUPP;
1247
1248         return tls_sw_do_sendpage(sk, page, offset, size, flags);
1249 }
1250
1251 int tls_sw_sendpage(struct sock *sk, struct page *page,
1252                     int offset, size_t size, int flags)
1253 {
1254         struct tls_context *tls_ctx = tls_get_ctx(sk);
1255         int ret;
1256
1257         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1258                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY))
1259                 return -EOPNOTSUPP;
1260
1261         mutex_lock(&tls_ctx->tx_lock);
1262         lock_sock(sk);
1263         ret = tls_sw_do_sendpage(sk, page, offset, size, flags);
1264         release_sock(sk);
1265         mutex_unlock(&tls_ctx->tx_lock);
1266         return ret;
1267 }
1268
1269 static struct sk_buff *tls_wait_data(struct sock *sk, struct sk_psock *psock,
1270                                      int flags, long timeo, int *err)
1271 {
1272         struct tls_context *tls_ctx = tls_get_ctx(sk);
1273         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1274         struct sk_buff *skb;
1275         DEFINE_WAIT_FUNC(wait, woken_wake_function);
1276
1277         while (!(skb = ctx->recv_pkt) && sk_psock_queue_empty(psock)) {
1278                 if (sk->sk_err) {
1279                         *err = sock_error(sk);
1280                         return NULL;
1281                 }
1282
1283                 if (sk->sk_shutdown & RCV_SHUTDOWN)
1284                         return NULL;
1285
1286                 if (sock_flag(sk, SOCK_DONE))
1287                         return NULL;
1288
1289                 if ((flags & MSG_DONTWAIT) || !timeo) {
1290                         *err = -EAGAIN;
1291                         return NULL;
1292                 }
1293
1294                 add_wait_queue(sk_sleep(sk), &wait);
1295                 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1296                 sk_wait_event(sk, &timeo,
1297                               ctx->recv_pkt != skb ||
1298                               !sk_psock_queue_empty(psock),
1299                               &wait);
1300                 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1301                 remove_wait_queue(sk_sleep(sk), &wait);
1302
1303                 /* Handle signals */
1304                 if (signal_pending(current)) {
1305                         *err = sock_intr_errno(timeo);
1306                         return NULL;
1307                 }
1308         }
1309
1310         return skb;
1311 }
1312
1313 static int tls_setup_from_iter(struct sock *sk, struct iov_iter *from,
1314                                int length, int *pages_used,
1315                                unsigned int *size_used,
1316                                struct scatterlist *to,
1317                                int to_max_pages)
1318 {
1319         int rc = 0, i = 0, num_elem = *pages_used, maxpages;
1320         struct page *pages[MAX_SKB_FRAGS];
1321         unsigned int size = *size_used;
1322         ssize_t copied, use;
1323         size_t offset;
1324
1325         while (length > 0) {
1326                 i = 0;
1327                 maxpages = to_max_pages - num_elem;
1328                 if (maxpages == 0) {
1329                         rc = -EFAULT;
1330                         goto out;
1331                 }
1332                 copied = iov_iter_get_pages(from, pages,
1333                                             length,
1334                                             maxpages, &offset);
1335                 if (copied <= 0) {
1336                         rc = -EFAULT;
1337                         goto out;
1338                 }
1339
1340                 iov_iter_advance(from, copied);
1341
1342                 length -= copied;
1343                 size += copied;
1344                 while (copied) {
1345                         use = min_t(int, copied, PAGE_SIZE - offset);
1346
1347                         sg_set_page(&to[num_elem],
1348                                     pages[i], use, offset);
1349                         sg_unmark_end(&to[num_elem]);
1350                         /* We do not uncharge memory from this API */
1351
1352                         offset = 0;
1353                         copied -= use;
1354
1355                         i++;
1356                         num_elem++;
1357                 }
1358         }
1359         /* Mark the end in the last sg entry if newly added */
1360         if (num_elem > *pages_used)
1361                 sg_mark_end(&to[num_elem - 1]);
1362 out:
1363         if (rc)
1364                 iov_iter_revert(from, size - *size_used);
1365         *size_used = size;
1366         *pages_used = num_elem;
1367
1368         return rc;
1369 }
1370
1371 /* This function decrypts the input skb into either out_iov or in out_sg
1372  * or in skb buffers itself. The input parameter 'zc' indicates if
1373  * zero-copy mode needs to be tried or not. With zero-copy mode, either
1374  * out_iov or out_sg must be non-NULL. In case both out_iov and out_sg are
1375  * NULL, then the decryption happens inside skb buffers itself, i.e.
1376  * zero-copy gets disabled and 'zc' is updated.
1377  */
1378
1379 static int decrypt_internal(struct sock *sk, struct sk_buff *skb,
1380                             struct iov_iter *out_iov,
1381                             struct scatterlist *out_sg,
1382                             int *chunk, bool *zc, bool async)
1383 {
1384         struct tls_context *tls_ctx = tls_get_ctx(sk);
1385         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1386         struct tls_prot_info *prot = &tls_ctx->prot_info;
1387         struct strp_msg *rxm = strp_msg(skb);
1388         int n_sgin, n_sgout, nsg, mem_size, aead_size, err, pages = 0;
1389         struct aead_request *aead_req;
1390         struct sk_buff *unused;
1391         u8 *aad, *iv, *mem = NULL;
1392         struct scatterlist *sgin = NULL;
1393         struct scatterlist *sgout = NULL;
1394         const int data_len = rxm->full_len - prot->overhead_size +
1395                              prot->tail_size;
1396         int iv_offset = 0;
1397
1398         if (*zc && (out_iov || out_sg)) {
1399                 if (out_iov)
1400                         n_sgout = iov_iter_npages(out_iov, INT_MAX) + 1;
1401                 else
1402                         n_sgout = sg_nents(out_sg);
1403                 n_sgin = skb_nsg(skb, rxm->offset + prot->prepend_size,
1404                                  rxm->full_len - prot->prepend_size);
1405         } else {
1406                 n_sgout = 0;
1407                 *zc = false;
1408                 n_sgin = skb_cow_data(skb, 0, &unused);
1409         }
1410
1411         if (n_sgin < 1)
1412                 return -EBADMSG;
1413
1414         /* Increment to accommodate AAD */
1415         n_sgin = n_sgin + 1;
1416
1417         nsg = n_sgin + n_sgout;
1418
1419         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(ctx->aead_recv);
1420         mem_size = aead_size + (nsg * sizeof(struct scatterlist));
1421         mem_size = mem_size + prot->aad_size;
1422         mem_size = mem_size + crypto_aead_ivsize(ctx->aead_recv);
1423
1424         /* Allocate a single block of memory which contains
1425          * aead_req || sgin[] || sgout[] || aad || iv.
1426          * This order achieves correct alignment for aead_req, sgin, sgout.
1427          */
1428         mem = kmalloc(mem_size, sk->sk_allocation);
1429         if (!mem)
1430                 return -ENOMEM;
1431
1432         /* Segment the allocated memory */
1433         aead_req = (struct aead_request *)mem;
1434         sgin = (struct scatterlist *)(mem + aead_size);
1435         sgout = sgin + n_sgin;
1436         aad = (u8 *)(sgout + n_sgout);
1437         iv = aad + prot->aad_size;
1438
1439         /* For CCM based ciphers, first byte of nonce+iv is always '2' */
1440         if (prot->cipher_type == TLS_CIPHER_AES_CCM_128) {
1441                 iv[0] = 2;
1442                 iv_offset = 1;
1443         }
1444
1445         /* Prepare IV */
1446         err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
1447                             iv + iv_offset + prot->salt_size,
1448                             prot->iv_size);
1449         if (err < 0) {
1450                 kfree(mem);
1451                 return err;
1452         }
1453         if (prot->version == TLS_1_3_VERSION)
1454                 memcpy(iv + iv_offset, tls_ctx->rx.iv,
1455                        crypto_aead_ivsize(ctx->aead_recv));
1456         else
1457                 memcpy(iv + iv_offset, tls_ctx->rx.iv, prot->salt_size);
1458
1459         xor_iv_with_seq(prot->version, iv, tls_ctx->rx.rec_seq);
1460
1461         /* Prepare AAD */
1462         tls_make_aad(aad, rxm->full_len - prot->overhead_size +
1463                      prot->tail_size,
1464                      tls_ctx->rx.rec_seq, prot->rec_seq_size,
1465                      ctx->control, prot->version);
1466
1467         /* Prepare sgin */
1468         sg_init_table(sgin, n_sgin);
1469         sg_set_buf(&sgin[0], aad, prot->aad_size);
1470         err = skb_to_sgvec(skb, &sgin[1],
1471                            rxm->offset + prot->prepend_size,
1472                            rxm->full_len - prot->prepend_size);
1473         if (err < 0) {
1474                 kfree(mem);
1475                 return err;
1476         }
1477
1478         if (n_sgout) {
1479                 if (out_iov) {
1480                         sg_init_table(sgout, n_sgout);
1481                         sg_set_buf(&sgout[0], aad, prot->aad_size);
1482
1483                         *chunk = 0;
1484                         err = tls_setup_from_iter(sk, out_iov, data_len,
1485                                                   &pages, chunk, &sgout[1],
1486                                                   (n_sgout - 1));
1487                         if (err < 0)
1488                                 goto fallback_to_reg_recv;
1489                 } else if (out_sg) {
1490                         memcpy(sgout, out_sg, n_sgout * sizeof(*sgout));
1491                 } else {
1492                         goto fallback_to_reg_recv;
1493                 }
1494         } else {
1495 fallback_to_reg_recv:
1496                 sgout = sgin;
1497                 pages = 0;
1498                 *chunk = data_len;
1499                 *zc = false;
1500         }
1501
1502         /* Prepare and submit AEAD request */
1503         err = tls_do_decryption(sk, skb, sgin, sgout, iv,
1504                                 data_len, aead_req, async);
1505         if (err == -EINPROGRESS)
1506                 return err;
1507
1508         /* Release the pages in case iov was mapped to pages */
1509         for (; pages > 0; pages--)
1510                 put_page(sg_page(&sgout[pages]));
1511
1512         kfree(mem);
1513         return err;
1514 }
1515
1516 static int decrypt_skb_update(struct sock *sk, struct sk_buff *skb,
1517                               struct iov_iter *dest, int *chunk, bool *zc,
1518                               bool async)
1519 {
1520         struct tls_context *tls_ctx = tls_get_ctx(sk);
1521         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1522         struct tls_prot_info *prot = &tls_ctx->prot_info;
1523         struct strp_msg *rxm = strp_msg(skb);
1524         int pad, err = 0;
1525
1526         if (!ctx->decrypted) {
1527                 if (tls_ctx->rx_conf == TLS_HW) {
1528                         err = tls_device_decrypted(sk, tls_ctx, skb, rxm);
1529                         if (err < 0)
1530                                 return err;
1531                 }
1532
1533                 /* Still not decrypted after tls_device */
1534                 if (!ctx->decrypted) {
1535                         err = decrypt_internal(sk, skb, dest, NULL, chunk, zc,
1536                                                async);
1537                         if (err < 0) {
1538                                 if (err == -EINPROGRESS)
1539                                         tls_advance_record_sn(sk, prot,
1540                                                               &tls_ctx->rx);
1541                                 else if (err == -EBADMSG)
1542                                         TLS_INC_STATS(sock_net(sk),
1543                                                       LINUX_MIB_TLSDECRYPTERROR);
1544                                 return err;
1545                         }
1546                 } else {
1547                         *zc = false;
1548                 }
1549
1550                 pad = padding_length(ctx, prot, skb);
1551                 if (pad < 0)
1552                         return pad;
1553
1554                 rxm->full_len -= pad;
1555                 rxm->offset += prot->prepend_size;
1556                 rxm->full_len -= prot->overhead_size;
1557                 tls_advance_record_sn(sk, prot, &tls_ctx->rx);
1558                 ctx->decrypted = 1;
1559                 ctx->saved_data_ready(sk);
1560         } else {
1561                 *zc = false;
1562         }
1563
1564         return err;
1565 }
1566
1567 int decrypt_skb(struct sock *sk, struct sk_buff *skb,
1568                 struct scatterlist *sgout)
1569 {
1570         bool zc = true;
1571         int chunk;
1572
1573         return decrypt_internal(sk, skb, NULL, sgout, &chunk, &zc, false);
1574 }
1575
1576 static bool tls_sw_advance_skb(struct sock *sk, struct sk_buff *skb,
1577                                unsigned int len)
1578 {
1579         struct tls_context *tls_ctx = tls_get_ctx(sk);
1580         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1581
1582         if (skb) {
1583                 struct strp_msg *rxm = strp_msg(skb);
1584
1585                 if (len < rxm->full_len) {
1586                         rxm->offset += len;
1587                         rxm->full_len -= len;
1588                         return false;
1589                 }
1590                 consume_skb(skb);
1591         }
1592
1593         /* Finished with message */
1594         ctx->recv_pkt = NULL;
1595         __strp_unpause(&ctx->strp);
1596
1597         return true;
1598 }
1599
1600 /* This function traverses the rx_list in tls receive context to copies the
1601  * decrypted records into the buffer provided by caller zero copy is not
1602  * true. Further, the records are removed from the rx_list if it is not a peek
1603  * case and the record has been consumed completely.
1604  */
1605 static int process_rx_list(struct tls_sw_context_rx *ctx,
1606                            struct msghdr *msg,
1607                            u8 *control,
1608                            bool *cmsg,
1609                            size_t skip,
1610                            size_t len,
1611                            bool zc,
1612                            bool is_peek)
1613 {
1614         struct sk_buff *skb = skb_peek(&ctx->rx_list);
1615         u8 ctrl = *control;
1616         u8 msgc = *cmsg;
1617         struct tls_msg *tlm;
1618         ssize_t copied = 0;
1619
1620         /* Set the record type in 'control' if caller didn't pass it */
1621         if (!ctrl && skb) {
1622                 tlm = tls_msg(skb);
1623                 ctrl = tlm->control;
1624         }
1625
1626         while (skip && skb) {
1627                 struct strp_msg *rxm = strp_msg(skb);
1628                 tlm = tls_msg(skb);
1629
1630                 /* Cannot process a record of different type */
1631                 if (ctrl != tlm->control)
1632                         return 0;
1633
1634                 if (skip < rxm->full_len)
1635                         break;
1636
1637                 skip = skip - rxm->full_len;
1638                 skb = skb_peek_next(skb, &ctx->rx_list);
1639         }
1640
1641         while (len && skb) {
1642                 struct sk_buff *next_skb;
1643                 struct strp_msg *rxm = strp_msg(skb);
1644                 int chunk = min_t(unsigned int, rxm->full_len - skip, len);
1645
1646                 tlm = tls_msg(skb);
1647
1648                 /* Cannot process a record of different type */
1649                 if (ctrl != tlm->control)
1650                         return 0;
1651
1652                 /* Set record type if not already done. For a non-data record,
1653                  * do not proceed if record type could not be copied.
1654                  */
1655                 if (!msgc) {
1656                         int cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1657                                             sizeof(ctrl), &ctrl);
1658                         msgc = true;
1659                         if (ctrl != TLS_RECORD_TYPE_DATA) {
1660                                 if (cerr || msg->msg_flags & MSG_CTRUNC)
1661                                         return -EIO;
1662
1663                                 *cmsg = msgc;
1664                         }
1665                 }
1666
1667                 if (!zc || (rxm->full_len - skip) > len) {
1668                         int err = skb_copy_datagram_msg(skb, rxm->offset + skip,
1669                                                     msg, chunk);
1670                         if (err < 0)
1671                                 return err;
1672                 }
1673
1674                 len = len - chunk;
1675                 copied = copied + chunk;
1676
1677                 /* Consume the data from record if it is non-peek case*/
1678                 if (!is_peek) {
1679                         rxm->offset = rxm->offset + chunk;
1680                         rxm->full_len = rxm->full_len - chunk;
1681
1682                         /* Return if there is unconsumed data in the record */
1683                         if (rxm->full_len - skip)
1684                                 break;
1685                 }
1686
1687                 /* The remaining skip-bytes must lie in 1st record in rx_list.
1688                  * So from the 2nd record, 'skip' should be 0.
1689                  */
1690                 skip = 0;
1691
1692                 if (msg)
1693                         msg->msg_flags |= MSG_EOR;
1694
1695                 next_skb = skb_peek_next(skb, &ctx->rx_list);
1696
1697                 if (!is_peek) {
1698                         skb_unlink(skb, &ctx->rx_list);
1699                         consume_skb(skb);
1700                 }
1701
1702                 skb = next_skb;
1703         }
1704
1705         *control = ctrl;
1706         return copied;
1707 }
1708
1709 int tls_sw_recvmsg(struct sock *sk,
1710                    struct msghdr *msg,
1711                    size_t len,
1712                    int nonblock,
1713                    int flags,
1714                    int *addr_len)
1715 {
1716         struct tls_context *tls_ctx = tls_get_ctx(sk);
1717         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1718         struct tls_prot_info *prot = &tls_ctx->prot_info;
1719         struct sk_psock *psock;
1720         unsigned char control = 0;
1721         ssize_t decrypted = 0;
1722         struct strp_msg *rxm;
1723         struct tls_msg *tlm;
1724         struct sk_buff *skb;
1725         ssize_t copied = 0;
1726         bool cmsg = false;
1727         int target, err = 0;
1728         long timeo;
1729         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
1730         bool is_peek = flags & MSG_PEEK;
1731         int num_async = 0;
1732
1733         flags |= nonblock;
1734
1735         if (unlikely(flags & MSG_ERRQUEUE))
1736                 return sock_recv_errqueue(sk, msg, len, SOL_IP, IP_RECVERR);
1737
1738         psock = sk_psock_get(sk);
1739         lock_sock(sk);
1740
1741         /* Process pending decrypted records. It must be non-zero-copy */
1742         err = process_rx_list(ctx, msg, &control, &cmsg, 0, len, false,
1743                               is_peek);
1744         if (err < 0) {
1745                 tls_err_abort(sk, err);
1746                 goto end;
1747         } else {
1748                 copied = err;
1749         }
1750
1751         if (len <= copied)
1752                 goto recv_end;
1753
1754         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1755         len = len - copied;
1756         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1757
1758         while (len && (decrypted + copied < target || ctx->recv_pkt)) {
1759                 bool retain_skb = false;
1760                 bool zc = false;
1761                 int to_decrypt;
1762                 int chunk = 0;
1763                 bool async_capable;
1764                 bool async = false;
1765
1766                 skb = tls_wait_data(sk, psock, flags, timeo, &err);
1767                 if (!skb) {
1768                         if (psock) {
1769                                 int ret = __tcp_bpf_recvmsg(sk, psock,
1770                                                             msg, len, flags);
1771
1772                                 if (ret > 0) {
1773                                         decrypted += ret;
1774                                         len -= ret;
1775                                         continue;
1776                                 }
1777                         }
1778                         goto recv_end;
1779                 } else {
1780                         tlm = tls_msg(skb);
1781                         if (prot->version == TLS_1_3_VERSION)
1782                                 tlm->control = 0;
1783                         else
1784                                 tlm->control = ctx->control;
1785                 }
1786
1787                 rxm = strp_msg(skb);
1788
1789                 to_decrypt = rxm->full_len - prot->overhead_size;
1790
1791                 if (to_decrypt <= len && !is_kvec && !is_peek &&
1792                     ctx->control == TLS_RECORD_TYPE_DATA &&
1793                     prot->version != TLS_1_3_VERSION)
1794                         zc = true;
1795
1796                 /* Do not use async mode if record is non-data */
1797                 if (ctx->control == TLS_RECORD_TYPE_DATA)
1798                         async_capable = ctx->async_capable;
1799                 else
1800                         async_capable = false;
1801
1802                 err = decrypt_skb_update(sk, skb, &msg->msg_iter,
1803                                          &chunk, &zc, async_capable);
1804                 if (err < 0 && err != -EINPROGRESS) {
1805                         tls_err_abort(sk, EBADMSG);
1806                         goto recv_end;
1807                 }
1808
1809                 if (err == -EINPROGRESS) {
1810                         async = true;
1811                         num_async++;
1812                 } else if (prot->version == TLS_1_3_VERSION) {
1813                         tlm->control = ctx->control;
1814                 }
1815
1816                 /* If the type of records being processed is not known yet,
1817                  * set it to record type just dequeued. If it is already known,
1818                  * but does not match the record type just dequeued, go to end.
1819                  * We always get record type here since for tls1.2, record type
1820                  * is known just after record is dequeued from stream parser.
1821                  * For tls1.3, we disable async.
1822                  */
1823
1824                 if (!control)
1825                         control = tlm->control;
1826                 else if (control != tlm->control)
1827                         goto recv_end;
1828
1829                 if (!cmsg) {
1830                         int cerr;
1831
1832                         cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1833                                         sizeof(control), &control);
1834                         cmsg = true;
1835                         if (control != TLS_RECORD_TYPE_DATA) {
1836                                 if (cerr || msg->msg_flags & MSG_CTRUNC) {
1837                                         err = -EIO;
1838                                         goto recv_end;
1839                                 }
1840                         }
1841                 }
1842
1843                 if (async)
1844                         goto pick_next_record;
1845
1846                 if (!zc) {
1847                         if (rxm->full_len > len) {
1848                                 retain_skb = true;
1849                                 chunk = len;
1850                         } else {
1851                                 chunk = rxm->full_len;
1852                         }
1853
1854                         err = skb_copy_datagram_msg(skb, rxm->offset,
1855                                                     msg, chunk);
1856                         if (err < 0)
1857                                 goto recv_end;
1858
1859                         if (!is_peek) {
1860                                 rxm->offset = rxm->offset + chunk;
1861                                 rxm->full_len = rxm->full_len - chunk;
1862                         }
1863                 }
1864
1865 pick_next_record:
1866                 if (chunk > len)
1867                         chunk = len;
1868
1869                 decrypted += chunk;
1870                 len -= chunk;
1871
1872                 /* For async or peek case, queue the current skb */
1873                 if (async || is_peek || retain_skb) {
1874                         skb_queue_tail(&ctx->rx_list, skb);
1875                         skb = NULL;
1876                 }
1877
1878                 if (tls_sw_advance_skb(sk, skb, chunk)) {
1879                         /* Return full control message to
1880                          * userspace before trying to parse
1881                          * another message type
1882                          */
1883                         msg->msg_flags |= MSG_EOR;
1884                         if (ctx->control != TLS_RECORD_TYPE_DATA)
1885                                 goto recv_end;
1886                 } else {
1887                         break;
1888                 }
1889         }
1890
1891 recv_end:
1892         if (num_async) {
1893                 /* Wait for all previously submitted records to be decrypted */
1894                 smp_store_mb(ctx->async_notify, true);
1895                 if (atomic_read(&ctx->decrypt_pending)) {
1896                         err = crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1897                         if (err) {
1898                                 /* one of async decrypt failed */
1899                                 tls_err_abort(sk, err);
1900                                 copied = 0;
1901                                 decrypted = 0;
1902                                 goto end;
1903                         }
1904                 } else {
1905                         reinit_completion(&ctx->async_wait.completion);
1906                 }
1907                 WRITE_ONCE(ctx->async_notify, false);
1908
1909                 /* Drain records from the rx_list & copy if required */
1910                 if (is_peek || is_kvec)
1911                         err = process_rx_list(ctx, msg, &control, &cmsg, copied,
1912                                               decrypted, false, is_peek);
1913                 else
1914                         err = process_rx_list(ctx, msg, &control, &cmsg, 0,
1915                                               decrypted, true, is_peek);
1916                 if (err < 0) {
1917                         tls_err_abort(sk, err);
1918                         copied = 0;
1919                         goto end;
1920                 }
1921         }
1922
1923         copied += decrypted;
1924
1925 end:
1926         release_sock(sk);
1927         if (psock)
1928                 sk_psock_put(sk, psock);
1929         return copied ? : err;
1930 }
1931
1932 ssize_t tls_sw_splice_read(struct socket *sock,  loff_t *ppos,
1933                            struct pipe_inode_info *pipe,
1934                            size_t len, unsigned int flags)
1935 {
1936         struct tls_context *tls_ctx = tls_get_ctx(sock->sk);
1937         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1938         struct strp_msg *rxm = NULL;
1939         struct sock *sk = sock->sk;
1940         struct sk_buff *skb;
1941         ssize_t copied = 0;
1942         int err = 0;
1943         long timeo;
1944         int chunk;
1945         bool zc = false;
1946
1947         lock_sock(sk);
1948
1949         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1950
1951         skb = tls_wait_data(sk, NULL, flags, timeo, &err);
1952         if (!skb)
1953                 goto splice_read_end;
1954
1955         if (!ctx->decrypted) {
1956                 err = decrypt_skb_update(sk, skb, NULL, &chunk, &zc, false);
1957
1958                 /* splice does not support reading control messages */
1959                 if (ctx->control != TLS_RECORD_TYPE_DATA) {
1960                         err = -EINVAL;
1961                         goto splice_read_end;
1962                 }
1963
1964                 if (err < 0) {
1965                         tls_err_abort(sk, EBADMSG);
1966                         goto splice_read_end;
1967                 }
1968                 ctx->decrypted = 1;
1969         }
1970         rxm = strp_msg(skb);
1971
1972         chunk = min_t(unsigned int, rxm->full_len, len);
1973         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, chunk, flags);
1974         if (copied < 0)
1975                 goto splice_read_end;
1976
1977         if (likely(!(flags & MSG_PEEK)))
1978                 tls_sw_advance_skb(sk, skb, copied);
1979
1980 splice_read_end:
1981         release_sock(sk);
1982         return copied ? : err;
1983 }
1984
1985 bool tls_sw_stream_read(const struct sock *sk)
1986 {
1987         struct tls_context *tls_ctx = tls_get_ctx(sk);
1988         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1989         bool ingress_empty = true;
1990         struct sk_psock *psock;
1991
1992         rcu_read_lock();
1993         psock = sk_psock(sk);
1994         if (psock)
1995                 ingress_empty = list_empty(&psock->ingress_msg);
1996         rcu_read_unlock();
1997
1998         return !ingress_empty || ctx->recv_pkt ||
1999                 !skb_queue_empty(&ctx->rx_list);
2000 }
2001
2002 static int tls_read_size(struct strparser *strp, struct sk_buff *skb)
2003 {
2004         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2005         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2006         struct tls_prot_info *prot = &tls_ctx->prot_info;
2007         char header[TLS_HEADER_SIZE + MAX_IV_SIZE];
2008         struct strp_msg *rxm = strp_msg(skb);
2009         size_t cipher_overhead;
2010         size_t data_len = 0;
2011         int ret;
2012
2013         /* Verify that we have a full TLS header, or wait for more data */
2014         if (rxm->offset + prot->prepend_size > skb->len)
2015                 return 0;
2016
2017         /* Sanity-check size of on-stack buffer. */
2018         if (WARN_ON(prot->prepend_size > sizeof(header))) {
2019                 ret = -EINVAL;
2020                 goto read_failure;
2021         }
2022
2023         /* Linearize header to local buffer */
2024         ret = skb_copy_bits(skb, rxm->offset, header, prot->prepend_size);
2025
2026         if (ret < 0)
2027                 goto read_failure;
2028
2029         ctx->control = header[0];
2030
2031         data_len = ((header[4] & 0xFF) | (header[3] << 8));
2032
2033         cipher_overhead = prot->tag_size;
2034         if (prot->version != TLS_1_3_VERSION)
2035                 cipher_overhead += prot->iv_size;
2036
2037         if (data_len > TLS_MAX_PAYLOAD_SIZE + cipher_overhead +
2038             prot->tail_size) {
2039                 ret = -EMSGSIZE;
2040                 goto read_failure;
2041         }
2042         if (data_len < cipher_overhead) {
2043                 ret = -EBADMSG;
2044                 goto read_failure;
2045         }
2046
2047         /* Note that both TLS1.3 and TLS1.2 use TLS_1_2 version here */
2048         if (header[1] != TLS_1_2_VERSION_MINOR ||
2049             header[2] != TLS_1_2_VERSION_MAJOR) {
2050                 ret = -EINVAL;
2051                 goto read_failure;
2052         }
2053
2054         tls_device_rx_resync_new_rec(strp->sk, data_len + TLS_HEADER_SIZE,
2055                                      TCP_SKB_CB(skb)->seq + rxm->offset);
2056         return data_len + TLS_HEADER_SIZE;
2057
2058 read_failure:
2059         tls_err_abort(strp->sk, ret);
2060
2061         return ret;
2062 }
2063
2064 static void tls_queue(struct strparser *strp, struct sk_buff *skb)
2065 {
2066         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2067         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2068
2069         ctx->decrypted = 0;
2070
2071         ctx->recv_pkt = skb;
2072         strp_pause(strp);
2073
2074         ctx->saved_data_ready(strp->sk);
2075 }
2076
2077 static void tls_data_ready(struct sock *sk)
2078 {
2079         struct tls_context *tls_ctx = tls_get_ctx(sk);
2080         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2081         struct sk_psock *psock;
2082
2083         strp_data_ready(&ctx->strp);
2084
2085         psock = sk_psock_get(sk);
2086         if (psock) {
2087                 if (!list_empty(&psock->ingress_msg))
2088                         ctx->saved_data_ready(sk);
2089                 sk_psock_put(sk, psock);
2090         }
2091 }
2092
2093 void tls_sw_cancel_work_tx(struct tls_context *tls_ctx)
2094 {
2095         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2096
2097         set_bit(BIT_TX_CLOSING, &ctx->tx_bitmask);
2098         set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask);
2099         cancel_delayed_work_sync(&ctx->tx_work.work);
2100 }
2101
2102 void tls_sw_release_resources_tx(struct sock *sk)
2103 {
2104         struct tls_context *tls_ctx = tls_get_ctx(sk);
2105         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2106         struct tls_rec *rec, *tmp;
2107
2108         /* Wait for any pending async encryptions to complete */
2109         smp_store_mb(ctx->async_notify, true);
2110         if (atomic_read(&ctx->encrypt_pending))
2111                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
2112
2113         tls_tx_records(sk, -1);
2114
2115         /* Free up un-sent records in tx_list. First, free
2116          * the partially sent record if any at head of tx_list.
2117          */
2118         if (tls_ctx->partially_sent_record) {
2119                 tls_free_partial_record(sk, tls_ctx);
2120                 rec = list_first_entry(&ctx->tx_list,
2121                                        struct tls_rec, list);
2122                 list_del(&rec->list);
2123                 sk_msg_free(sk, &rec->msg_plaintext);
2124                 kfree(rec);
2125         }
2126
2127         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
2128                 list_del(&rec->list);
2129                 sk_msg_free(sk, &rec->msg_encrypted);
2130                 sk_msg_free(sk, &rec->msg_plaintext);
2131                 kfree(rec);
2132         }
2133
2134         crypto_free_aead(ctx->aead_send);
2135         tls_free_open_rec(sk);
2136 }
2137
2138 void tls_sw_free_ctx_tx(struct tls_context *tls_ctx)
2139 {
2140         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2141
2142         kfree(ctx);
2143 }
2144
2145 void tls_sw_release_resources_rx(struct sock *sk)
2146 {
2147         struct tls_context *tls_ctx = tls_get_ctx(sk);
2148         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2149
2150         kfree(tls_ctx->rx.rec_seq);
2151         kfree(tls_ctx->rx.iv);
2152
2153         if (ctx->aead_recv) {
2154                 kfree_skb(ctx->recv_pkt);
2155                 ctx->recv_pkt = NULL;
2156                 skb_queue_purge(&ctx->rx_list);
2157                 crypto_free_aead(ctx->aead_recv);
2158                 strp_stop(&ctx->strp);
2159                 /* If tls_sw_strparser_arm() was not called (cleanup paths)
2160                  * we still want to strp_stop(), but sk->sk_data_ready was
2161                  * never swapped.
2162                  */
2163                 if (ctx->saved_data_ready) {
2164                         write_lock_bh(&sk->sk_callback_lock);
2165                         sk->sk_data_ready = ctx->saved_data_ready;
2166                         write_unlock_bh(&sk->sk_callback_lock);
2167                 }
2168         }
2169 }
2170
2171 void tls_sw_strparser_done(struct tls_context *tls_ctx)
2172 {
2173         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2174
2175         strp_done(&ctx->strp);
2176 }
2177
2178 void tls_sw_free_ctx_rx(struct tls_context *tls_ctx)
2179 {
2180         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2181
2182         kfree(ctx);
2183 }
2184
2185 void tls_sw_free_resources_rx(struct sock *sk)
2186 {
2187         struct tls_context *tls_ctx = tls_get_ctx(sk);
2188
2189         tls_sw_release_resources_rx(sk);
2190         tls_sw_free_ctx_rx(tls_ctx);
2191 }
2192
2193 /* The work handler to transmitt the encrypted records in tx_list */
2194 static void tx_work_handler(struct work_struct *work)
2195 {
2196         struct delayed_work *delayed_work = to_delayed_work(work);
2197         struct tx_work *tx_work = container_of(delayed_work,
2198                                                struct tx_work, work);
2199         struct sock *sk = tx_work->sk;
2200         struct tls_context *tls_ctx = tls_get_ctx(sk);
2201         struct tls_sw_context_tx *ctx;
2202
2203         if (unlikely(!tls_ctx))
2204                 return;
2205
2206         ctx = tls_sw_ctx_tx(tls_ctx);
2207         if (test_bit(BIT_TX_CLOSING, &ctx->tx_bitmask))
2208                 return;
2209
2210         if (!test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
2211                 return;
2212         mutex_lock(&tls_ctx->tx_lock);
2213         lock_sock(sk);
2214         tls_tx_records(sk, -1);
2215         release_sock(sk);
2216         mutex_unlock(&tls_ctx->tx_lock);
2217 }
2218
2219 void tls_sw_write_space(struct sock *sk, struct tls_context *ctx)
2220 {
2221         struct tls_sw_context_tx *tx_ctx = tls_sw_ctx_tx(ctx);
2222
2223         /* Schedule the transmission if tx list is ready */
2224         if (is_tx_ready(tx_ctx) &&
2225             !test_and_set_bit(BIT_TX_SCHEDULED, &tx_ctx->tx_bitmask))
2226                 schedule_delayed_work(&tx_ctx->tx_work.work, 0);
2227 }
2228
2229 void tls_sw_strparser_arm(struct sock *sk, struct tls_context *tls_ctx)
2230 {
2231         struct tls_sw_context_rx *rx_ctx = tls_sw_ctx_rx(tls_ctx);
2232
2233         write_lock_bh(&sk->sk_callback_lock);
2234         rx_ctx->saved_data_ready = sk->sk_data_ready;
2235         sk->sk_data_ready = tls_data_ready;
2236         write_unlock_bh(&sk->sk_callback_lock);
2237
2238         strp_check_rcv(&rx_ctx->strp);
2239 }
2240
2241 int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
2242 {
2243         struct tls_context *tls_ctx = tls_get_ctx(sk);
2244         struct tls_prot_info *prot = &tls_ctx->prot_info;
2245         struct tls_crypto_info *crypto_info;
2246         struct tls12_crypto_info_aes_gcm_128 *gcm_128_info;
2247         struct tls12_crypto_info_aes_gcm_256 *gcm_256_info;
2248         struct tls12_crypto_info_aes_ccm_128 *ccm_128_info;
2249         struct tls_sw_context_tx *sw_ctx_tx = NULL;
2250         struct tls_sw_context_rx *sw_ctx_rx = NULL;
2251         struct cipher_context *cctx;
2252         struct crypto_aead **aead;
2253         struct strp_callbacks cb;
2254         u16 nonce_size, tag_size, iv_size, rec_seq_size, salt_size;
2255         struct crypto_tfm *tfm;
2256         char *iv, *rec_seq, *key, *salt, *cipher_name;
2257         size_t keysize;
2258         int rc = 0;
2259
2260         if (!ctx) {
2261                 rc = -EINVAL;
2262                 goto out;
2263         }
2264
2265         if (tx) {
2266                 if (!ctx->priv_ctx_tx) {
2267                         sw_ctx_tx = kzalloc(sizeof(*sw_ctx_tx), GFP_KERNEL);
2268                         if (!sw_ctx_tx) {
2269                                 rc = -ENOMEM;
2270                                 goto out;
2271                         }
2272                         ctx->priv_ctx_tx = sw_ctx_tx;
2273                 } else {
2274                         sw_ctx_tx =
2275                                 (struct tls_sw_context_tx *)ctx->priv_ctx_tx;
2276                 }
2277         } else {
2278                 if (!ctx->priv_ctx_rx) {
2279                         sw_ctx_rx = kzalloc(sizeof(*sw_ctx_rx), GFP_KERNEL);
2280                         if (!sw_ctx_rx) {
2281                                 rc = -ENOMEM;
2282                                 goto out;
2283                         }
2284                         ctx->priv_ctx_rx = sw_ctx_rx;
2285                 } else {
2286                         sw_ctx_rx =
2287                                 (struct tls_sw_context_rx *)ctx->priv_ctx_rx;
2288                 }
2289         }
2290
2291         if (tx) {
2292                 crypto_init_wait(&sw_ctx_tx->async_wait);
2293                 crypto_info = &ctx->crypto_send.info;
2294                 cctx = &ctx->tx;
2295                 aead = &sw_ctx_tx->aead_send;
2296                 INIT_LIST_HEAD(&sw_ctx_tx->tx_list);
2297                 INIT_DELAYED_WORK(&sw_ctx_tx->tx_work.work, tx_work_handler);
2298                 sw_ctx_tx->tx_work.sk = sk;
2299         } else {
2300                 crypto_init_wait(&sw_ctx_rx->async_wait);
2301                 crypto_info = &ctx->crypto_recv.info;
2302                 cctx = &ctx->rx;
2303                 skb_queue_head_init(&sw_ctx_rx->rx_list);
2304                 aead = &sw_ctx_rx->aead_recv;
2305         }
2306
2307         switch (crypto_info->cipher_type) {
2308         case TLS_CIPHER_AES_GCM_128: {
2309                 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2310                 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
2311                 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2312                 iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
2313                 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
2314                 rec_seq =
2315                  ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
2316                 gcm_128_info =
2317                         (struct tls12_crypto_info_aes_gcm_128 *)crypto_info;
2318                 keysize = TLS_CIPHER_AES_GCM_128_KEY_SIZE;
2319                 key = gcm_128_info->key;
2320                 salt = gcm_128_info->salt;
2321                 salt_size = TLS_CIPHER_AES_GCM_128_SALT_SIZE;
2322                 cipher_name = "gcm(aes)";
2323                 break;
2324         }
2325         case TLS_CIPHER_AES_GCM_256: {
2326                 nonce_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2327                 tag_size = TLS_CIPHER_AES_GCM_256_TAG_SIZE;
2328                 iv_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2329                 iv = ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->iv;
2330                 rec_seq_size = TLS_CIPHER_AES_GCM_256_REC_SEQ_SIZE;
2331                 rec_seq =
2332                  ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->rec_seq;
2333                 gcm_256_info =
2334                         (struct tls12_crypto_info_aes_gcm_256 *)crypto_info;
2335                 keysize = TLS_CIPHER_AES_GCM_256_KEY_SIZE;
2336                 key = gcm_256_info->key;
2337                 salt = gcm_256_info->salt;
2338                 salt_size = TLS_CIPHER_AES_GCM_256_SALT_SIZE;
2339                 cipher_name = "gcm(aes)";
2340                 break;
2341         }
2342         case TLS_CIPHER_AES_CCM_128: {
2343                 nonce_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2344                 tag_size = TLS_CIPHER_AES_CCM_128_TAG_SIZE;
2345                 iv_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2346                 iv = ((struct tls12_crypto_info_aes_ccm_128 *)crypto_info)->iv;
2347                 rec_seq_size = TLS_CIPHER_AES_CCM_128_REC_SEQ_SIZE;
2348                 rec_seq =
2349                 ((struct tls12_crypto_info_aes_ccm_128 *)crypto_info)->rec_seq;
2350                 ccm_128_info =
2351                 (struct tls12_crypto_info_aes_ccm_128 *)crypto_info;
2352                 keysize = TLS_CIPHER_AES_CCM_128_KEY_SIZE;
2353                 key = ccm_128_info->key;
2354                 salt = ccm_128_info->salt;
2355                 salt_size = TLS_CIPHER_AES_CCM_128_SALT_SIZE;
2356                 cipher_name = "ccm(aes)";
2357                 break;
2358         }
2359         default:
2360                 rc = -EINVAL;
2361                 goto free_priv;
2362         }
2363
2364         /* Sanity-check the sizes for stack allocations. */
2365         if (iv_size > MAX_IV_SIZE || nonce_size > MAX_IV_SIZE ||
2366             rec_seq_size > TLS_MAX_REC_SEQ_SIZE) {
2367                 rc = -EINVAL;
2368                 goto free_priv;
2369         }
2370
2371         if (crypto_info->version == TLS_1_3_VERSION) {
2372                 nonce_size = 0;
2373                 prot->aad_size = TLS_HEADER_SIZE;
2374                 prot->tail_size = 1;
2375         } else {
2376                 prot->aad_size = TLS_AAD_SPACE_SIZE;
2377                 prot->tail_size = 0;
2378         }
2379
2380         prot->version = crypto_info->version;
2381         prot->cipher_type = crypto_info->cipher_type;
2382         prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
2383         prot->tag_size = tag_size;
2384         prot->overhead_size = prot->prepend_size +
2385                               prot->tag_size + prot->tail_size;
2386         prot->iv_size = iv_size;
2387         prot->salt_size = salt_size;
2388         cctx->iv = kmalloc(iv_size + salt_size, GFP_KERNEL);
2389         if (!cctx->iv) {
2390                 rc = -ENOMEM;
2391                 goto free_priv;
2392         }
2393         /* Note: 128 & 256 bit salt are the same size */
2394         prot->rec_seq_size = rec_seq_size;
2395         memcpy(cctx->iv, salt, salt_size);
2396         memcpy(cctx->iv + salt_size, iv, iv_size);
2397         cctx->rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
2398         if (!cctx->rec_seq) {
2399                 rc = -ENOMEM;
2400                 goto free_iv;
2401         }
2402
2403         if (!*aead) {
2404                 *aead = crypto_alloc_aead(cipher_name, 0, 0);
2405                 if (IS_ERR(*aead)) {
2406                         rc = PTR_ERR(*aead);
2407                         *aead = NULL;
2408                         goto free_rec_seq;
2409                 }
2410         }
2411
2412         ctx->push_pending_record = tls_sw_push_pending_record;
2413
2414         rc = crypto_aead_setkey(*aead, key, keysize);
2415
2416         if (rc)
2417                 goto free_aead;
2418
2419         rc = crypto_aead_setauthsize(*aead, prot->tag_size);
2420         if (rc)
2421                 goto free_aead;
2422
2423         if (sw_ctx_rx) {
2424                 tfm = crypto_aead_tfm(sw_ctx_rx->aead_recv);
2425
2426                 if (crypto_info->version == TLS_1_3_VERSION)
2427                         sw_ctx_rx->async_capable = 0;
2428                 else
2429                         sw_ctx_rx->async_capable =
2430                                 !!(tfm->__crt_alg->cra_flags &
2431                                    CRYPTO_ALG_ASYNC);
2432
2433                 /* Set up strparser */
2434                 memset(&cb, 0, sizeof(cb));
2435                 cb.rcv_msg = tls_queue;
2436                 cb.parse_msg = tls_read_size;
2437
2438                 strp_init(&sw_ctx_rx->strp, sk, &cb);
2439         }
2440
2441         goto out;
2442
2443 free_aead:
2444         crypto_free_aead(*aead);
2445         *aead = NULL;
2446 free_rec_seq:
2447         kfree(cctx->rec_seq);
2448         cctx->rec_seq = NULL;
2449 free_iv:
2450         kfree(cctx->iv);
2451         cctx->iv = NULL;
2452 free_priv:
2453         if (tx) {
2454                 kfree(ctx->priv_ctx_tx);
2455                 ctx->priv_ctx_tx = NULL;
2456         } else {
2457                 kfree(ctx->priv_ctx_rx);
2458                 ctx->priv_ctx_rx = NULL;
2459         }
2460 out:
2461         return rc;
2462 }