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                                ssize_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 && sk->sk_err == EBADMSG) {
 800                         *copied -= sk_msg_free(sk, msg);
 801                         tls_free_open_rec(sk);
 802                         err = -sk->sk_err;
 803                 }
 804                 if (psock)
 805                         sk_psock_put(sk, psock);
 806                 return err;
 807         }
 808 more_data:
 809         enospc = sk_msg_full(msg);
 810         if (psock->eval == __SK_NONE) {
 811                 delta = msg->sg.size;
 812                 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
 813                 delta -= msg->sg.size;
 814         }
 815         if (msg->cork_bytes && msg->cork_bytes > msg->sg.size &&
 816             !enospc && !full_record) {
 817                 err = -ENOSPC;
 818                 goto out_err;
 819         }
 820         msg->cork_bytes = 0;
 821         send = msg->sg.size;
 822         if (msg->apply_bytes && msg->apply_bytes < send)
 823                 send = msg->apply_bytes;
 824
 825         switch (psock->eval) {
 826         case __SK_PASS:
 827                 err = tls_push_record(sk, flags, record_type);
 828                 if (err && sk->sk_err == EBADMSG) {
 829                         *copied -= sk_msg_free(sk, msg);
 830                         tls_free_open_rec(sk);
 831                         err = -sk->sk_err;
 832                         goto out_err;
 833                 }
 834                 break;
 835         case __SK_REDIRECT:
 836                 sk_redir = psock->sk_redir;
 837                 memcpy(&msg_redir, msg, sizeof(*msg));
 838                 if (msg->apply_bytes < send)
 839                         msg->apply_bytes = 0;
 840                 else
 841                         msg->apply_bytes -= send;
 842                 sk_msg_return_zero(sk, msg, send);
 843                 msg->sg.size -= send;
 844                 release_sock(sk);
 845                 err = tcp_bpf_sendmsg_redir(sk_redir, &msg_redir, send, flags);
 846                 lock_sock(sk);
 847                 if (err < 0) {
 848                         *copied -= sk_msg_free_nocharge(sk, &msg_redir);
 849                         msg->sg.size = 0;
 850                 }
 851                 if (msg->sg.size == 0)
 852                         tls_free_open_rec(sk);
 853                 break;
 854         case __SK_DROP:
 855         default:
 856                 sk_msg_free_partial(sk, msg, send);
 857                 if (msg->apply_bytes < send)
 858                         msg->apply_bytes = 0;
 859                 else
 860                         msg->apply_bytes -= send;
 861                 if (msg->sg.size == 0)
 862                         tls_free_open_rec(sk);
 863                 *copied -= (send + delta);
 864                 err = -EACCES;
 865         }
 866
 867         if (likely(!err)) {
 868                 bool reset_eval = !ctx->open_rec;
 869
 870                 rec = ctx->open_rec;
 871                 if (rec) {
 872                         msg = &rec->msg_plaintext;
 873                         if (!msg->apply_bytes)
 874                                 reset_eval = true;
 875                 }
 876                 if (reset_eval) {
 877                         psock->eval = __SK_NONE;
 878                         if (psock->sk_redir) {
 879                                 sock_put(psock->sk_redir);
 880                                 psock->sk_redir = NULL;
 881                         }
 882                 }
 883                 if (rec)
 884                         goto more_data;
 885         }
 886  out_err:
 887         sk_psock_put(sk, psock);
 888         return err;
 889 }
 890
 891 static int tls_sw_push_pending_record(struct sock *sk, int flags)
 892 {
 893         struct tls_context *tls_ctx = tls_get_ctx(sk);
 894         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 895         struct tls_rec *rec = ctx->open_rec;
 896         struct sk_msg *msg_pl;
 897         size_t copied;
 898
 899         if (!rec)
 900                 return 0;
 901
 902         msg_pl = &rec->msg_plaintext;
 903         copied = msg_pl->sg.size;
 904         if (!copied)
 905                 return 0;
 906
 907         return bpf_exec_tx_verdict(msg_pl, sk, true, TLS_RECORD_TYPE_DATA,
 908                                    &copied, flags);
 909 }
 910
 911 int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 912 {
 913         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 914         struct tls_context *tls_ctx = tls_get_ctx(sk);
 915         struct tls_prot_info *prot = &tls_ctx->prot_info;
 916         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 917         bool async_capable = ctx->async_capable;
 918         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 919         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
 920         bool eor = !(msg->msg_flags & MSG_MORE);
 921         size_t try_to_copy;
 922         ssize_t copied = 0;
 923         struct sk_msg *msg_pl, *msg_en;
 924         struct tls_rec *rec;
 925         int required_size;
 926         int num_async = 0;
 927         bool full_record;
 928         int record_room;
 929         int num_zc = 0;
 930         int orig_size;
 931         int ret = 0;
 932
 933         if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
 934                 return -EOPNOTSUPP;
 935
 936         mutex_lock(&tls_ctx->tx_lock);
 937         lock_sock(sk);
 938
 939         if (unlikely(msg->msg_controllen)) {
 940                 ret = tls_proccess_cmsg(sk, msg, &record_type);
 941                 if (ret) {
 942                         if (ret == -EINPROGRESS)
 943                                 num_async++;
 944                         else if (ret != -EAGAIN)
 945                                 goto send_end;
 946                 }
 947         }
 948
 949         while (msg_data_left(msg)) {
 950                 if (sk->sk_err) {
 951                         ret = -sk->sk_err;
 952                         goto send_end;
 953                 }
 954
 955                 if (ctx->open_rec)
 956                         rec = ctx->open_rec;
 957                 else
 958                         rec = ctx->open_rec = tls_get_rec(sk);
 959                 if (!rec) {
 960                         ret = -ENOMEM;
 961                         goto send_end;
 962                 }
 963
 964                 msg_pl = &rec->msg_plaintext;
 965                 msg_en = &rec->msg_encrypted;
 966
 967                 orig_size = msg_pl->sg.size;
 968                 full_record = false;
 969                 try_to_copy = msg_data_left(msg);
 970                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
 971                 if (try_to_copy >= record_room) {
 972                         try_to_copy = record_room;
 973                         full_record = true;
 974                 }
 975
 976                 required_size = msg_pl->sg.size + try_to_copy +
 977                                 prot->overhead_size;
 978
 979                 if (!sk_stream_memory_free(sk))
 980                         goto wait_for_sndbuf;
 981
 982 alloc_encrypted:
 983                 ret = tls_alloc_encrypted_msg(sk, required_size);
 984                 if (ret) {
 985                         if (ret != -ENOSPC)
 986                                 goto wait_for_memory;
 987
 988                         /* Adjust try_to_copy according to the amount that was
 989                          * actually allocated. The difference is due
 990                          * to max sg elements limit
 991                          */
 992                         try_to_copy -= required_size - msg_en->sg.size;
 993                         full_record = true;
 994                 }
 995
 996                 if (!is_kvec && (full_record || eor) && !async_capable) {
 997                         u32 first = msg_pl->sg.end;
 998
 999                         ret = sk_msg_zerocopy_from_iter(sk, &msg->msg_iter,
1000                                                         msg_pl, try_to_copy);
1001                         if (ret)
1002                                 goto fallback_to_reg_send;
1003
1004                         num_zc++;
1005                         copied += try_to_copy;
1006
1007                         sk_msg_sg_copy_set(msg_pl, first);
1008                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1009                                                   record_type, &copied,
1010                                                   msg->msg_flags);
1011                         if (ret) {
1012                                 if (ret == -EINPROGRESS)
1013                                         num_async++;
1014                                 else if (ret == -ENOMEM)
1015                                         goto wait_for_memory;
1016                                 else if (ctx->open_rec && ret == -ENOSPC)
1017                                         goto rollback_iter;
1018                                 else if (ret != -EAGAIN)
1019                                         goto send_end;
1020                         }
1021                         continue;
1022 rollback_iter:
1023                         copied -= try_to_copy;
1024                         sk_msg_sg_copy_clear(msg_pl, first);
1025                         iov_iter_revert(&msg->msg_iter,
1026                                         msg_pl->sg.size - orig_size);
1027 fallback_to_reg_send:
1028                         sk_msg_trim(sk, msg_pl, orig_size);
1029                 }
1030
1031                 required_size = msg_pl->sg.size + try_to_copy;
1032
1033                 ret = tls_clone_plaintext_msg(sk, required_size);
1034                 if (ret) {
1035                         if (ret != -ENOSPC)
1036                                 goto send_end;
1037
1038                         /* Adjust try_to_copy according to the amount that was
1039                          * actually allocated. The difference is due
1040                          * to max sg elements limit
1041                          */
1042                         try_to_copy -= required_size - msg_pl->sg.size;
1043                         full_record = true;
1044                         sk_msg_trim(sk, msg_en,
1045                                     msg_pl->sg.size + prot->overhead_size);
1046                 }
1047
1048                 if (try_to_copy) {
1049                         ret = sk_msg_memcopy_from_iter(sk, &msg->msg_iter,
1050                                                        msg_pl, try_to_copy);
1051                         if (ret < 0)
1052                                 goto trim_sgl;
1053                 }
1054
1055                 /* Open records defined only if successfully copied, otherwise
1056                  * we would trim the sg but not reset the open record frags.
1057                  */
1058                 tls_ctx->pending_open_record_frags = true;
1059                 copied += try_to_copy;
1060                 if (full_record || eor) {
1061                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1062                                                   record_type, &copied,
1063                                                   msg->msg_flags);
1064                         if (ret) {
1065                                 if (ret == -EINPROGRESS)
1066                                         num_async++;
1067                                 else if (ret == -ENOMEM)
1068                                         goto wait_for_memory;
1069                                 else if (ret != -EAGAIN) {
1070                                         if (ret == -ENOSPC)
1071                                                 ret = 0;
1072                                         goto send_end;
1073                                 }
1074                         }
1075                 }
1076
1077                 continue;
1078
1079 wait_for_sndbuf:
1080                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1081 wait_for_memory:
1082                 ret = sk_stream_wait_memory(sk, &timeo);
1083                 if (ret) {
1084 trim_sgl:
1085                         if (ctx->open_rec)
1086                                 tls_trim_both_msgs(sk, orig_size);
1087                         goto send_end;
1088                 }
1089
1090                 if (ctx->open_rec && msg_en->sg.size < required_size)
1091                         goto alloc_encrypted;
1092         }
1093
1094         if (!num_async) {
1095                 goto send_end;
1096         } else if (num_zc) {
1097                 /* Wait for pending encryptions to get completed */
1098                 smp_store_mb(ctx->async_notify, true);
1099
1100                 if (atomic_read(&ctx->encrypt_pending))
1101                         crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1102                 else
1103                         reinit_completion(&ctx->async_wait.completion);
1104
1105                 WRITE_ONCE(ctx->async_notify, false);
1106
1107                 if (ctx->async_wait.err) {
1108                         ret = ctx->async_wait.err;
1109                         copied = 0;
1110                 }
1111         }
1112
1113         /* Transmit if any encryptions have completed */
1114         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1115                 cancel_delayed_work(&ctx->tx_work.work);
1116                 tls_tx_records(sk, msg->msg_flags);
1117         }
1118
1119 send_end:
1120         ret = sk_stream_error(sk, msg->msg_flags, ret);
1121
1122         release_sock(sk);
1123         mutex_unlock(&tls_ctx->tx_lock);
1124         return copied > 0 ? copied : ret;
1125 }
1126
1127 static int tls_sw_do_sendpage(struct sock *sk, struct page *page,
1128                               int offset, size_t size, int flags)
1129 {
1130         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1131         struct tls_context *tls_ctx = tls_get_ctx(sk);
1132         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1133         struct tls_prot_info *prot = &tls_ctx->prot_info;
1134         unsigned char record_type = TLS_RECORD_TYPE_DATA;
1135         struct sk_msg *msg_pl;
1136         struct tls_rec *rec;
1137         int num_async = 0;
1138         ssize_t copied = 0;
1139         bool full_record;
1140         int record_room;
1141         int ret = 0;
1142         bool eor;
1143
1144         eor = !(flags & (MSG_MORE | MSG_SENDPAGE_NOTLAST));
1145         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1146
1147         /* Call the sk_stream functions to manage the sndbuf mem. */
1148         while (size > 0) {
1149                 size_t copy, required_size;
1150
1151                 if (sk->sk_err) {
1152                         ret = -sk->sk_err;
1153                         goto sendpage_end;
1154                 }
1155
1156                 if (ctx->open_rec)
1157                         rec = ctx->open_rec;
1158                 else
1159                         rec = ctx->open_rec = tls_get_rec(sk);
1160                 if (!rec) {
1161                         ret = -ENOMEM;
1162                         goto sendpage_end;
1163                 }
1164
1165                 msg_pl = &rec->msg_plaintext;
1166
1167                 full_record = false;
1168                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
1169                 copy = size;
1170                 if (copy >= record_room) {
1171                         copy = record_room;
1172                         full_record = true;
1173                 }
1174
1175                 required_size = msg_pl->sg.size + copy + prot->overhead_size;
1176
1177                 if (!sk_stream_memory_free(sk))
1178                         goto wait_for_sndbuf;
1179 alloc_payload:
1180                 ret = tls_alloc_encrypted_msg(sk, required_size);
1181                 if (ret) {
1182                         if (ret != -ENOSPC)
1183                                 goto wait_for_memory;
1184
1185                         /* Adjust copy according to the amount that was
1186                          * actually allocated. The difference is due
1187                          * to max sg elements limit
1188                          */
1189                         copy -= required_size - msg_pl->sg.size;
1190                         full_record = true;
1191                 }
1192
1193                 sk_msg_page_add(msg_pl, page, copy, offset);
1194                 sk_mem_charge(sk, copy);
1195
1196                 offset += copy;
1197                 size -= copy;
1198                 copied += copy;
1199
1200                 tls_ctx->pending_open_record_frags = true;
1201                 if (full_record || eor || sk_msg_full(msg_pl)) {
1202                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1203                                                   record_type, &copied, flags);
1204                         if (ret) {
1205                                 if (ret == -EINPROGRESS)
1206                                         num_async++;
1207                                 else if (ret == -ENOMEM)
1208                                         goto wait_for_memory;
1209                                 else if (ret != -EAGAIN) {
1210                                         if (ret == -ENOSPC)
1211                                                 ret = 0;
1212                                         goto sendpage_end;
1213                                 }
1214                         }
1215                 }
1216                 continue;
1217 wait_for_sndbuf:
1218                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1219 wait_for_memory:
1220                 ret = sk_stream_wait_memory(sk, &timeo);
1221                 if (ret) {
1222                         if (ctx->open_rec)
1223                                 tls_trim_both_msgs(sk, msg_pl->sg.size);
1224                         goto sendpage_end;
1225                 }
1226
1227                 if (ctx->open_rec)
1228                         goto alloc_payload;
1229         }
1230
1231         if (num_async) {
1232                 /* Transmit if any encryptions have completed */
1233                 if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1234                         cancel_delayed_work(&ctx->tx_work.work);
1235                         tls_tx_records(sk, flags);
1236                 }
1237         }
1238 sendpage_end:
1239         ret = sk_stream_error(sk, flags, ret);
1240         return copied > 0 ? copied : ret;
1241 }
1242
1243 int tls_sw_sendpage_locked(struct sock *sk, struct page *page,
1244                            int offset, size_t size, int flags)
1245 {
1246         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1247                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY |
1248                       MSG_NO_SHARED_FRAGS))
1249                 return -EOPNOTSUPP;
1250
1251         return tls_sw_do_sendpage(sk, page, offset, size, flags);
1252 }
1253
1254 int tls_sw_sendpage(struct sock *sk, struct page *page,
1255                     int offset, size_t size, int flags)
1256 {
1257         struct tls_context *tls_ctx = tls_get_ctx(sk);
1258         int ret;
1259
1260         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1261                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY))
1262                 return -EOPNOTSUPP;
1263
1264         mutex_lock(&tls_ctx->tx_lock);
1265         lock_sock(sk);
1266         ret = tls_sw_do_sendpage(sk, page, offset, size, flags);
1267         release_sock(sk);
1268         mutex_unlock(&tls_ctx->tx_lock);
1269         return ret;
1270 }
1271
1272 static struct sk_buff *tls_wait_data(struct sock *sk, struct sk_psock *psock,
1273                                      int flags, long timeo, int *err)
1274 {
1275         struct tls_context *tls_ctx = tls_get_ctx(sk);
1276         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1277         struct sk_buff *skb;
1278         DEFINE_WAIT_FUNC(wait, woken_wake_function);
1279
1280         while (!(skb = ctx->recv_pkt) && sk_psock_queue_empty(psock)) {
1281                 if (sk->sk_err) {
1282                         *err = sock_error(sk);
1283                         return NULL;
1284                 }
1285
1286                 if (sk->sk_shutdown & RCV_SHUTDOWN)
1287                         return NULL;
1288
1289                 if (sock_flag(sk, SOCK_DONE))
1290                         return NULL;
1291
1292                 if ((flags & MSG_DONTWAIT) || !timeo) {
1293                         *err = -EAGAIN;
1294                         return NULL;
1295                 }
1296
1297                 add_wait_queue(sk_sleep(sk), &wait);
1298                 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1299                 sk_wait_event(sk, &timeo,
1300                               ctx->recv_pkt != skb ||
1301                               !sk_psock_queue_empty(psock),
1302                               &wait);
1303                 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1304                 remove_wait_queue(sk_sleep(sk), &wait);
1305
1306                 /* Handle signals */
1307                 if (signal_pending(current)) {
1308                         *err = sock_intr_errno(timeo);
1309                         return NULL;
1310                 }
1311         }
1312
1313         return skb;
1314 }
1315
1316 static int tls_setup_from_iter(struct sock *sk, struct iov_iter *from,
1317                                int length, int *pages_used,
1318                                unsigned int *size_used,
1319                                struct scatterlist *to,
1320                                int to_max_pages)
1321 {
1322         int rc = 0, i = 0, num_elem = *pages_used, maxpages;
1323         struct page *pages[MAX_SKB_FRAGS];
1324         unsigned int size = *size_used;
1325         ssize_t copied, use;
1326         size_t offset;
1327
1328         while (length > 0) {
1329                 i = 0;
1330                 maxpages = to_max_pages - num_elem;
1331                 if (maxpages == 0) {
1332                         rc = -EFAULT;
1333                         goto out;
1334                 }
1335                 copied = iov_iter_get_pages(from, pages,
1336                                             length,
1337                                             maxpages, &offset);
1338                 if (copied <= 0) {
1339                         rc = -EFAULT;
1340                         goto out;
1341                 }
1342
1343                 iov_iter_advance(from, copied);
1344
1345                 length -= copied;
1346                 size += copied;
1347                 while (copied) {
1348                         use = min_t(int, copied, PAGE_SIZE - offset);
1349
1350                         sg_set_page(&to[num_elem],
1351                                     pages[i], use, offset);
1352                         sg_unmark_end(&to[num_elem]);
1353                         /* We do not uncharge memory from this API */
1354
1355                         offset = 0;
1356                         copied -= use;
1357
1358                         i++;
1359                         num_elem++;
1360                 }
1361         }
1362         /* Mark the end in the last sg entry if newly added */
1363         if (num_elem > *pages_used)
1364                 sg_mark_end(&to[num_elem - 1]);
1365 out:
1366         if (rc)
1367                 iov_iter_revert(from, size - *size_used);
1368         *size_used = size;
1369         *pages_used = num_elem;
1370
1371         return rc;
1372 }
1373
1374 /* This function decrypts the input skb into either out_iov or in out_sg
1375  * or in skb buffers itself. The input parameter 'zc' indicates if
1376  * zero-copy mode needs to be tried or not. With zero-copy mode, either
1377  * out_iov or out_sg must be non-NULL. In case both out_iov and out_sg are
1378  * NULL, then the decryption happens inside skb buffers itself, i.e.
1379  * zero-copy gets disabled and 'zc' is updated.
1380  */
1381
1382 static int decrypt_internal(struct sock *sk, struct sk_buff *skb,
1383                             struct iov_iter *out_iov,
1384                             struct scatterlist *out_sg,
1385                             int *chunk, bool *zc, bool async)
1386 {
1387         struct tls_context *tls_ctx = tls_get_ctx(sk);
1388         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1389         struct tls_prot_info *prot = &tls_ctx->prot_info;
1390         struct strp_msg *rxm = strp_msg(skb);
1391         int n_sgin, n_sgout, nsg, mem_size, aead_size, err, pages = 0;
1392         struct aead_request *aead_req;
1393         struct sk_buff *unused;
1394         u8 *aad, *iv, *mem = NULL;
1395         struct scatterlist *sgin = NULL;
1396         struct scatterlist *sgout = NULL;
1397         const int data_len = rxm->full_len - prot->overhead_size +
1398                              prot->tail_size;
1399         int iv_offset = 0;
1400
1401         if (*zc && (out_iov || out_sg)) {
1402                 if (out_iov)
1403                         n_sgout = iov_iter_npages(out_iov, INT_MAX) + 1;
1404                 else
1405                         n_sgout = sg_nents(out_sg);
1406                 n_sgin = skb_nsg(skb, rxm->offset + prot->prepend_size,
1407                                  rxm->full_len - prot->prepend_size);
1408         } else {
1409                 n_sgout = 0;
1410                 *zc = false;
1411                 n_sgin = skb_cow_data(skb, 0, &unused);
1412         }
1413
1414         if (n_sgin < 1)
1415                 return -EBADMSG;
1416
1417         /* Increment to accommodate AAD */
1418         n_sgin = n_sgin + 1;
1419
1420         nsg = n_sgin + n_sgout;
1421
1422         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(ctx->aead_recv);
1423         mem_size = aead_size + (nsg * sizeof(struct scatterlist));
1424         mem_size = mem_size + prot->aad_size;
1425         mem_size = mem_size + crypto_aead_ivsize(ctx->aead_recv);
1426
1427         /* Allocate a single block of memory which contains
1428          * aead_req || sgin[] || sgout[] || aad || iv.
1429          * This order achieves correct alignment for aead_req, sgin, sgout.
1430          */
1431         mem = kmalloc(mem_size, sk->sk_allocation);
1432         if (!mem)
1433                 return -ENOMEM;
1434
1435         /* Segment the allocated memory */
1436         aead_req = (struct aead_request *)mem;
1437         sgin = (struct scatterlist *)(mem + aead_size);
1438         sgout = sgin + n_sgin;
1439         aad = (u8 *)(sgout + n_sgout);
1440         iv = aad + prot->aad_size;
1441
1442         /* For CCM based ciphers, first byte of nonce+iv is always '2' */
1443         if (prot->cipher_type == TLS_CIPHER_AES_CCM_128) {
1444                 iv[0] = 2;
1445                 iv_offset = 1;
1446         }
1447
1448         /* Prepare IV */
1449         err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
1450                             iv + iv_offset + prot->salt_size,
1451                             prot->iv_size);
1452         if (err < 0) {
1453                 kfree(mem);
1454                 return err;
1455         }
1456         if (prot->version == TLS_1_3_VERSION)
1457                 memcpy(iv + iv_offset, tls_ctx->rx.iv,
1458                        crypto_aead_ivsize(ctx->aead_recv));
1459         else
1460                 memcpy(iv + iv_offset, tls_ctx->rx.iv, prot->salt_size);
1461
1462         xor_iv_with_seq(prot->version, iv, tls_ctx->rx.rec_seq);
1463
1464         /* Prepare AAD */
1465         tls_make_aad(aad, rxm->full_len - prot->overhead_size +
1466                      prot->tail_size,
1467                      tls_ctx->rx.rec_seq, prot->rec_seq_size,
1468                      ctx->control, prot->version);
1469
1470         /* Prepare sgin */
1471         sg_init_table(sgin, n_sgin);
1472         sg_set_buf(&sgin[0], aad, prot->aad_size);
1473         err = skb_to_sgvec(skb, &sgin[1],
1474                            rxm->offset + prot->prepend_size,
1475                            rxm->full_len - prot->prepend_size);
1476         if (err < 0) {
1477                 kfree(mem);
1478                 return err;
1479         }
1480
1481         if (n_sgout) {
1482                 if (out_iov) {
1483                         sg_init_table(sgout, n_sgout);
1484                         sg_set_buf(&sgout[0], aad, prot->aad_size);
1485
1486                         *chunk = 0;
1487                         err = tls_setup_from_iter(sk, out_iov, data_len,
1488                                                   &pages, chunk, &sgout[1],
1489                                                   (n_sgout - 1));
1490                         if (err < 0)
1491                                 goto fallback_to_reg_recv;
1492                 } else if (out_sg) {
1493                         memcpy(sgout, out_sg, n_sgout * sizeof(*sgout));
1494                 } else {
1495                         goto fallback_to_reg_recv;
1496                 }
1497         } else {
1498 fallback_to_reg_recv:
1499                 sgout = sgin;
1500                 pages = 0;
1501                 *chunk = data_len;
1502                 *zc = false;
1503         }
1504
1505         /* Prepare and submit AEAD request */
1506         err = tls_do_decryption(sk, skb, sgin, sgout, iv,
1507                                 data_len, aead_req, async);
1508         if (err == -EINPROGRESS)
1509                 return err;
1510
1511         /* Release the pages in case iov was mapped to pages */
1512         for (; pages > 0; pages--)
1513                 put_page(sg_page(&sgout[pages]));
1514
1515         kfree(mem);
1516         return err;
1517 }
1518
1519 static int decrypt_skb_update(struct sock *sk, struct sk_buff *skb,
1520                               struct iov_iter *dest, int *chunk, bool *zc,
1521                               bool async)
1522 {
1523         struct tls_context *tls_ctx = tls_get_ctx(sk);
1524         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1525         struct tls_prot_info *prot = &tls_ctx->prot_info;
1526         struct strp_msg *rxm = strp_msg(skb);
1527         int pad, err = 0;
1528
1529         if (!ctx->decrypted) {
1530                 if (tls_ctx->rx_conf == TLS_HW) {
1531                         err = tls_device_decrypted(sk, tls_ctx, skb, rxm);
1532                         if (err < 0)
1533                                 return err;
1534                 }
1535
1536                 /* Still not decrypted after tls_device */
1537                 if (!ctx->decrypted) {
1538                         err = decrypt_internal(sk, skb, dest, NULL, chunk, zc,
1539                                                async);
1540                         if (err < 0) {
1541                                 if (err == -EINPROGRESS)
1542                                         tls_advance_record_sn(sk, prot,
1543                                                               &tls_ctx->rx);
1544                                 else if (err == -EBADMSG)
1545                                         TLS_INC_STATS(sock_net(sk),
1546                                                       LINUX_MIB_TLSDECRYPTERROR);
1547                                 return err;
1548                         }
1549                 } else {
1550                         *zc = false;
1551                 }
1552
1553                 pad = padding_length(ctx, prot, skb);
1554                 if (pad < 0)
1555                         return pad;
1556
1557                 rxm->full_len -= pad;
1558                 rxm->offset += prot->prepend_size;
1559                 rxm->full_len -= prot->overhead_size;
1560                 tls_advance_record_sn(sk, prot, &tls_ctx->rx);
1561                 ctx->decrypted = 1;
1562                 ctx->saved_data_ready(sk);
1563         } else {
1564                 *zc = false;
1565         }
1566
1567         return err;
1568 }
1569
1570 int decrypt_skb(struct sock *sk, struct sk_buff *skb,
1571                 struct scatterlist *sgout)
1572 {
1573         bool zc = true;
1574         int chunk;
1575
1576         return decrypt_internal(sk, skb, NULL, sgout, &chunk, &zc, false);
1577 }
1578
1579 static bool tls_sw_advance_skb(struct sock *sk, struct sk_buff *skb,
1580                                unsigned int len)
1581 {
1582         struct tls_context *tls_ctx = tls_get_ctx(sk);
1583         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1584
1585         if (skb) {
1586                 struct strp_msg *rxm = strp_msg(skb);
1587
1588                 if (len < rxm->full_len) {
1589                         rxm->offset += len;
1590                         rxm->full_len -= len;
1591                         return false;
1592                 }
1593                 consume_skb(skb);
1594         }
1595
1596         /* Finished with message */
1597         ctx->recv_pkt = NULL;
1598         __strp_unpause(&ctx->strp);
1599
1600         return true;
1601 }
1602
1603 /* This function traverses the rx_list in tls receive context to copies the
1604  * decrypted records into the buffer provided by caller zero copy is not
1605  * true. Further, the records are removed from the rx_list if it is not a peek
1606  * case and the record has been consumed completely.
1607  */
1608 static int process_rx_list(struct tls_sw_context_rx *ctx,
1609                            struct msghdr *msg,
1610                            u8 *control,
1611                            bool *cmsg,
1612                            size_t skip,
1613                            size_t len,
1614                            bool zc,
1615                            bool is_peek)
1616 {
1617         struct sk_buff *skb = skb_peek(&ctx->rx_list);
1618         u8 ctrl = *control;
1619         u8 msgc = *cmsg;
1620         struct tls_msg *tlm;
1621         ssize_t copied = 0;
1622
1623         /* Set the record type in 'control' if caller didn't pass it */
1624         if (!ctrl && skb) {
1625                 tlm = tls_msg(skb);
1626                 ctrl = tlm->control;
1627         }
1628
1629         while (skip && skb) {
1630                 struct strp_msg *rxm = strp_msg(skb);
1631                 tlm = tls_msg(skb);
1632
1633                 /* Cannot process a record of different type */
1634                 if (ctrl != tlm->control)
1635                         return 0;
1636
1637                 if (skip < rxm->full_len)
1638                         break;
1639
1640                 skip = skip - rxm->full_len;
1641                 skb = skb_peek_next(skb, &ctx->rx_list);
1642         }
1643
1644         while (len && skb) {
1645                 struct sk_buff *next_skb;
1646                 struct strp_msg *rxm = strp_msg(skb);
1647                 int chunk = min_t(unsigned int, rxm->full_len - skip, len);
1648
1649                 tlm = tls_msg(skb);
1650
1651                 /* Cannot process a record of different type */
1652                 if (ctrl != tlm->control)
1653                         return 0;
1654
1655                 /* Set record type if not already done. For a non-data record,
1656                  * do not proceed if record type could not be copied.
1657                  */
1658                 if (!msgc) {
1659                         int cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1660                                             sizeof(ctrl), &ctrl);
1661                         msgc = true;
1662                         if (ctrl != TLS_RECORD_TYPE_DATA) {
1663                                 if (cerr || msg->msg_flags & MSG_CTRUNC)
1664                                         return -EIO;
1665
1666                                 *cmsg = msgc;
1667                         }
1668                 }
1669
1670                 if (!zc || (rxm->full_len - skip) > len) {
1671                         int err = skb_copy_datagram_msg(skb, rxm->offset + skip,
1672                                                     msg, chunk);
1673                         if (err < 0)
1674                                 return err;
1675                 }
1676
1677                 len = len - chunk;
1678                 copied = copied + chunk;
1679
1680                 /* Consume the data from record if it is non-peek case*/
1681                 if (!is_peek) {
1682                         rxm->offset = rxm->offset + chunk;
1683                         rxm->full_len = rxm->full_len - chunk;
1684
1685                         /* Return if there is unconsumed data in the record */
1686                         if (rxm->full_len - skip)
1687                                 break;
1688                 }
1689
1690                 /* The remaining skip-bytes must lie in 1st record in rx_list.
1691                  * So from the 2nd record, 'skip' should be 0.
1692                  */
1693                 skip = 0;
1694
1695                 if (msg)
1696                         msg->msg_flags |= MSG_EOR;
1697
1698                 next_skb = skb_peek_next(skb, &ctx->rx_list);
1699
1700                 if (!is_peek) {
1701                         skb_unlink(skb, &ctx->rx_list);
1702                         consume_skb(skb);
1703                 }
1704
1705                 skb = next_skb;
1706         }
1707
1708         *control = ctrl;
1709         return copied;
1710 }
1711
1712 int tls_sw_recvmsg(struct sock *sk,
1713                    struct msghdr *msg,
1714                    size_t len,
1715                    int nonblock,
1716                    int flags,
1717                    int *addr_len)
1718 {
1719         struct tls_context *tls_ctx = tls_get_ctx(sk);
1720         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1721         struct tls_prot_info *prot = &tls_ctx->prot_info;
1722         struct sk_psock *psock;
1723         unsigned char control = 0;
1724         ssize_t decrypted = 0;
1725         struct strp_msg *rxm;
1726         struct tls_msg *tlm;
1727         struct sk_buff *skb;
1728         ssize_t copied = 0;
1729         bool cmsg = false;
1730         int target, err = 0;
1731         long timeo;
1732         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
1733         bool is_peek = flags & MSG_PEEK;
1734         int num_async = 0;
1735
1736         flags |= nonblock;
1737
1738         if (unlikely(flags & MSG_ERRQUEUE))
1739                 return sock_recv_errqueue(sk, msg, len, SOL_IP, IP_RECVERR);
1740
1741         psock = sk_psock_get(sk);
1742         lock_sock(sk);
1743
1744         /* Process pending decrypted records. It must be non-zero-copy */
1745         err = process_rx_list(ctx, msg, &control, &cmsg, 0, len, false,
1746                               is_peek);
1747         if (err < 0) {
1748                 tls_err_abort(sk, err);
1749                 goto end;
1750         } else {
1751                 copied = err;
1752         }
1753
1754         if (len <= copied)
1755                 goto recv_end;
1756
1757         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1758         len = len - copied;
1759         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1760
1761         while (len && (decrypted + copied < target || ctx->recv_pkt)) {
1762                 bool retain_skb = false;
1763                 bool zc = false;
1764                 int to_decrypt;
1765                 int chunk = 0;
1766                 bool async_capable;
1767                 bool async = false;
1768
1769                 skb = tls_wait_data(sk, psock, flags, timeo, &err);
1770                 if (!skb) {
1771                         if (psock) {
1772                                 int ret = __tcp_bpf_recvmsg(sk, psock,
1773                                                             msg, len, flags);
1774
1775                                 if (ret > 0) {
1776                                         decrypted += ret;
1777                                         len -= ret;
1778                                         continue;
1779                                 }
1780                         }
1781                         goto recv_end;
1782                 } else {
1783                         tlm = tls_msg(skb);
1784                         if (prot->version == TLS_1_3_VERSION)
1785                                 tlm->control = 0;
1786                         else
1787                                 tlm->control = ctx->control;
1788                 }
1789
1790                 rxm = strp_msg(skb);
1791
1792                 to_decrypt = rxm->full_len - prot->overhead_size;
1793
1794                 if (to_decrypt <= len && !is_kvec && !is_peek &&
1795                     ctx->control == TLS_RECORD_TYPE_DATA &&
1796                     prot->version != TLS_1_3_VERSION)
1797                         zc = true;
1798
1799                 /* Do not use async mode if record is non-data */
1800                 if (ctx->control == TLS_RECORD_TYPE_DATA)
1801                         async_capable = ctx->async_capable;
1802                 else
1803                         async_capable = false;
1804
1805                 err = decrypt_skb_update(sk, skb, &msg->msg_iter,
1806                                          &chunk, &zc, async_capable);
1807                 if (err < 0 && err != -EINPROGRESS) {
1808                         tls_err_abort(sk, EBADMSG);
1809                         goto recv_end;
1810                 }
1811
1812                 if (err == -EINPROGRESS) {
1813                         async = true;
1814                         num_async++;
1815                 } else if (prot->version == TLS_1_3_VERSION) {
1816                         tlm->control = ctx->control;
1817                 }
1818
1819                 /* If the type of records being processed is not known yet,
1820                  * set it to record type just dequeued. If it is already known,
1821                  * but does not match the record type just dequeued, go to end.
1822                  * We always get record type here since for tls1.2, record type
1823                  * is known just after record is dequeued from stream parser.
1824                  * For tls1.3, we disable async.
1825                  */
1826
1827                 if (!control)
1828                         control = tlm->control;
1829                 else if (control != tlm->control)
1830                         goto recv_end;
1831
1832                 if (!cmsg) {
1833                         int cerr;
1834
1835                         cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1836                                         sizeof(control), &control);
1837                         cmsg = true;
1838                         if (control != TLS_RECORD_TYPE_DATA) {
1839                                 if (cerr || msg->msg_flags & MSG_CTRUNC) {
1840                                         err = -EIO;
1841                                         goto recv_end;
1842                                 }
1843                         }
1844                 }
1845
1846                 if (async)
1847                         goto pick_next_record;
1848
1849                 if (!zc) {
1850                         if (rxm->full_len > len) {
1851                                 retain_skb = true;
1852                                 chunk = len;
1853                         } else {
1854                                 chunk = rxm->full_len;
1855                         }
1856
1857                         err = skb_copy_datagram_msg(skb, rxm->offset,
1858                                                     msg, chunk);
1859                         if (err < 0)
1860                                 goto recv_end;
1861
1862                         if (!is_peek) {
1863                                 rxm->offset = rxm->offset + chunk;
1864                                 rxm->full_len = rxm->full_len - chunk;
1865                         }
1866                 }
1867
1868 pick_next_record:
1869                 if (chunk > len)
1870                         chunk = len;
1871
1872                 decrypted += chunk;
1873                 len -= chunk;
1874
1875                 /* For async or peek case, queue the current skb */
1876                 if (async || is_peek || retain_skb) {
1877                         skb_queue_tail(&ctx->rx_list, skb);
1878                         skb = NULL;
1879                 }
1880
1881                 if (tls_sw_advance_skb(sk, skb, chunk)) {
1882                         /* Return full control message to
1883                          * userspace before trying to parse
1884                          * another message type
1885                          */
1886                         msg->msg_flags |= MSG_EOR;
1887                         if (ctx->control != TLS_RECORD_TYPE_DATA)
1888                                 goto recv_end;
1889                 } else {
1890                         break;
1891                 }
1892         }
1893
1894 recv_end:
1895         if (num_async) {
1896                 /* Wait for all previously submitted records to be decrypted */
1897                 smp_store_mb(ctx->async_notify, true);
1898                 if (atomic_read(&ctx->decrypt_pending)) {
1899                         err = crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1900                         if (err) {
1901                                 /* one of async decrypt failed */
1902                                 tls_err_abort(sk, err);
1903                                 copied = 0;
1904                                 decrypted = 0;
1905                                 goto end;
1906                         }
1907                 } else {
1908                         reinit_completion(&ctx->async_wait.completion);
1909                 }
1910                 WRITE_ONCE(ctx->async_notify, false);
1911
1912                 /* Drain records from the rx_list & copy if required */
1913                 if (is_peek || is_kvec)
1914                         err = process_rx_list(ctx, msg, &control, &cmsg, copied,
1915                                               decrypted, false, is_peek);
1916                 else
1917                         err = process_rx_list(ctx, msg, &control, &cmsg, 0,
1918                                               decrypted, true, is_peek);
1919                 if (err < 0) {
1920                         tls_err_abort(sk, err);
1921                         copied = 0;
1922                         goto end;
1923                 }
1924         }
1925
1926         copied += decrypted;
1927
1928 end:
1929         release_sock(sk);
1930         if (psock)
1931                 sk_psock_put(sk, psock);
1932         return copied ? : err;
1933 }
1934
1935 ssize_t tls_sw_splice_read(struct socket *sock,  loff_t *ppos,
1936                            struct pipe_inode_info *pipe,
1937                            size_t len, unsigned int flags)
1938 {
1939         struct tls_context *tls_ctx = tls_get_ctx(sock->sk);
1940         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1941         struct strp_msg *rxm = NULL;
1942         struct sock *sk = sock->sk;
1943         struct sk_buff *skb;
1944         ssize_t copied = 0;
1945         int err = 0;
1946         long timeo;
1947         int chunk;
1948         bool zc = false;
1949
1950         lock_sock(sk);
1951
1952         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1953
1954         skb = tls_wait_data(sk, NULL, flags, timeo, &err);
1955         if (!skb)
1956                 goto splice_read_end;
1957
1958         if (!ctx->decrypted) {
1959                 err = decrypt_skb_update(sk, skb, NULL, &chunk, &zc, false);
1960
1961                 /* splice does not support reading control messages */
1962                 if (ctx->control != TLS_RECORD_TYPE_DATA) {
1963                         err = -EINVAL;
1964                         goto splice_read_end;
1965                 }
1966
1967                 if (err < 0) {
1968                         tls_err_abort(sk, EBADMSG);
1969                         goto splice_read_end;
1970                 }
1971                 ctx->decrypted = 1;
1972         }
1973         rxm = strp_msg(skb);
1974
1975         chunk = min_t(unsigned int, rxm->full_len, len);
1976         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, chunk, flags);
1977         if (copied < 0)
1978                 goto splice_read_end;
1979
1980         if (likely(!(flags & MSG_PEEK)))
1981                 tls_sw_advance_skb(sk, skb, copied);
1982
1983 splice_read_end:
1984         release_sock(sk);
1985         return copied ? : err;
1986 }
1987
1988 bool tls_sw_stream_read(const struct sock *sk)
1989 {
1990         struct tls_context *tls_ctx = tls_get_ctx(sk);
1991         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1992         bool ingress_empty = true;
1993         struct sk_psock *psock;
1994
1995         rcu_read_lock();
1996         psock = sk_psock(sk);
1997         if (psock)
1998                 ingress_empty = list_empty(&psock->ingress_msg);
1999         rcu_read_unlock();
2000
2001         return !ingress_empty || ctx->recv_pkt ||
2002                 !skb_queue_empty(&ctx->rx_list);
2003 }
2004
2005 static int tls_read_size(struct strparser *strp, struct sk_buff *skb)
2006 {
2007         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2008         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2009         struct tls_prot_info *prot = &tls_ctx->prot_info;
2010         char header[TLS_HEADER_SIZE + MAX_IV_SIZE];
2011         struct strp_msg *rxm = strp_msg(skb);
2012         size_t cipher_overhead;
2013         size_t data_len = 0;
2014         int ret;
2015
2016         /* Verify that we have a full TLS header, or wait for more data */
2017         if (rxm->offset + prot->prepend_size > skb->len)
2018                 return 0;
2019
2020         /* Sanity-check size of on-stack buffer. */
2021         if (WARN_ON(prot->prepend_size > sizeof(header))) {
2022                 ret = -EINVAL;
2023                 goto read_failure;
2024         }
2025
2026         /* Linearize header to local buffer */
2027         ret = skb_copy_bits(skb, rxm->offset, header, prot->prepend_size);
2028
2029         if (ret < 0)
2030                 goto read_failure;
2031
2032         ctx->control = header[0];
2033
2034         data_len = ((header[4] & 0xFF) | (header[3] << 8));
2035
2036         cipher_overhead = prot->tag_size;
2037         if (prot->version != TLS_1_3_VERSION)
2038                 cipher_overhead += prot->iv_size;
2039
2040         if (data_len > TLS_MAX_PAYLOAD_SIZE + cipher_overhead +
2041             prot->tail_size) {
2042                 ret = -EMSGSIZE;
2043                 goto read_failure;
2044         }
2045         if (data_len < cipher_overhead) {
2046                 ret = -EBADMSG;
2047                 goto read_failure;
2048         }
2049
2050         /* Note that both TLS1.3 and TLS1.2 use TLS_1_2 version here */
2051         if (header[1] != TLS_1_2_VERSION_MINOR ||
2052             header[2] != TLS_1_2_VERSION_MAJOR) {
2053                 ret = -EINVAL;
2054                 goto read_failure;
2055         }
2056
2057         tls_device_rx_resync_new_rec(strp->sk, data_len + TLS_HEADER_SIZE,
2058                                      TCP_SKB_CB(skb)->seq + rxm->offset);
2059         return data_len + TLS_HEADER_SIZE;
2060
2061 read_failure:
2062         tls_err_abort(strp->sk, ret);
2063
2064         return ret;
2065 }
2066
2067 static void tls_queue(struct strparser *strp, struct sk_buff *skb)
2068 {
2069         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2070         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2071
2072         ctx->decrypted = 0;
2073
2074         ctx->recv_pkt = skb;
2075         strp_pause(strp);
2076
2077         ctx->saved_data_ready(strp->sk);
2078 }
2079
2080 static void tls_data_ready(struct sock *sk)
2081 {
2082         struct tls_context *tls_ctx = tls_get_ctx(sk);
2083         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2084         struct sk_psock *psock;
2085
2086         strp_data_ready(&ctx->strp);
2087
2088         psock = sk_psock_get(sk);
2089         if (psock) {
2090                 if (!list_empty(&psock->ingress_msg))
2091                         ctx->saved_data_ready(sk);
2092                 sk_psock_put(sk, psock);
2093         }
2094 }
2095
2096 void tls_sw_cancel_work_tx(struct tls_context *tls_ctx)
2097 {
2098         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2099
2100         set_bit(BIT_TX_CLOSING, &ctx->tx_bitmask);
2101         set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask);
2102         cancel_delayed_work_sync(&ctx->tx_work.work);
2103 }
2104
2105 void tls_sw_release_resources_tx(struct sock *sk)
2106 {
2107         struct tls_context *tls_ctx = tls_get_ctx(sk);
2108         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2109         struct tls_rec *rec, *tmp;
2110
2111         /* Wait for any pending async encryptions to complete */
2112         smp_store_mb(ctx->async_notify, true);
2113         if (atomic_read(&ctx->encrypt_pending))
2114                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
2115
2116         tls_tx_records(sk, -1);
2117
2118         /* Free up un-sent records in tx_list. First, free
2119          * the partially sent record if any at head of tx_list.
2120          */
2121         if (tls_ctx->partially_sent_record) {
2122                 tls_free_partial_record(sk, tls_ctx);
2123                 rec = list_first_entry(&ctx->tx_list,
2124                                        struct tls_rec, list);
2125                 list_del(&rec->list);
2126                 sk_msg_free(sk, &rec->msg_plaintext);
2127                 kfree(rec);
2128         }
2129
2130         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
2131                 list_del(&rec->list);
2132                 sk_msg_free(sk, &rec->msg_encrypted);
2133                 sk_msg_free(sk, &rec->msg_plaintext);
2134                 kfree(rec);
2135         }
2136
2137         crypto_free_aead(ctx->aead_send);
2138         tls_free_open_rec(sk);
2139 }
2140
2141 void tls_sw_free_ctx_tx(struct tls_context *tls_ctx)
2142 {
2143         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2144
2145         kfree(ctx);
2146 }
2147
2148 void tls_sw_release_resources_rx(struct sock *sk)
2149 {
2150         struct tls_context *tls_ctx = tls_get_ctx(sk);
2151         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2152
2153         kfree(tls_ctx->rx.rec_seq);
2154         kfree(tls_ctx->rx.iv);
2155
2156         if (ctx->aead_recv) {
2157                 kfree_skb(ctx->recv_pkt);
2158                 ctx->recv_pkt = NULL;
2159                 skb_queue_purge(&ctx->rx_list);
2160                 crypto_free_aead(ctx->aead_recv);
2161                 strp_stop(&ctx->strp);
2162                 /* If tls_sw_strparser_arm() was not called (cleanup paths)
2163                  * we still want to strp_stop(), but sk->sk_data_ready was
2164                  * never swapped.
2165                  */
2166                 if (ctx->saved_data_ready) {
2167                         write_lock_bh(&sk->sk_callback_lock);
2168                         sk->sk_data_ready = ctx->saved_data_ready;
2169                         write_unlock_bh(&sk->sk_callback_lock);
2170                 }
2171         }
2172 }
2173
2174 void tls_sw_strparser_done(struct tls_context *tls_ctx)
2175 {
2176         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2177
2178         strp_done(&ctx->strp);
2179 }
2180
2181 void tls_sw_free_ctx_rx(struct tls_context *tls_ctx)
2182 {
2183         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2184
2185         kfree(ctx);
2186 }
2187
2188 void tls_sw_free_resources_rx(struct sock *sk)
2189 {
2190         struct tls_context *tls_ctx = tls_get_ctx(sk);
2191
2192         tls_sw_release_resources_rx(sk);
2193         tls_sw_free_ctx_rx(tls_ctx);
2194 }
2195
2196 /* The work handler to transmitt the encrypted records in tx_list */
2197 static void tx_work_handler(struct work_struct *work)
2198 {
2199         struct delayed_work *delayed_work = to_delayed_work(work);
2200         struct tx_work *tx_work = container_of(delayed_work,
2201                                                struct tx_work, work);
2202         struct sock *sk = tx_work->sk;
2203         struct tls_context *tls_ctx = tls_get_ctx(sk);
2204         struct tls_sw_context_tx *ctx;
2205
2206         if (unlikely(!tls_ctx))
2207                 return;
2208
2209         ctx = tls_sw_ctx_tx(tls_ctx);
2210         if (test_bit(BIT_TX_CLOSING, &ctx->tx_bitmask))
2211                 return;
2212
2213         if (!test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
2214                 return;
2215         mutex_lock(&tls_ctx->tx_lock);
2216         lock_sock(sk);
2217         tls_tx_records(sk, -1);
2218         release_sock(sk);
2219         mutex_unlock(&tls_ctx->tx_lock);
2220 }
2221
2222 void tls_sw_write_space(struct sock *sk, struct tls_context *ctx)
2223 {
2224         struct tls_sw_context_tx *tx_ctx = tls_sw_ctx_tx(ctx);
2225
2226         /* Schedule the transmission if tx list is ready */
2227         if (is_tx_ready(tx_ctx) &&
2228             !test_and_set_bit(BIT_TX_SCHEDULED, &tx_ctx->tx_bitmask))
2229                 schedule_delayed_work(&tx_ctx->tx_work.work, 0);
2230 }
2231
2232 void tls_sw_strparser_arm(struct sock *sk, struct tls_context *tls_ctx)
2233 {
2234         struct tls_sw_context_rx *rx_ctx = tls_sw_ctx_rx(tls_ctx);
2235
2236         write_lock_bh(&sk->sk_callback_lock);
2237         rx_ctx->saved_data_ready = sk->sk_data_ready;
2238         sk->sk_data_ready = tls_data_ready;
2239         write_unlock_bh(&sk->sk_callback_lock);
2240
2241         strp_check_rcv(&rx_ctx->strp);
2242 }
2243
2244 int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
2245 {
2246         struct tls_context *tls_ctx = tls_get_ctx(sk);
2247         struct tls_prot_info *prot = &tls_ctx->prot_info;
2248         struct tls_crypto_info *crypto_info;
2249         struct tls12_crypto_info_aes_gcm_128 *gcm_128_info;
2250         struct tls12_crypto_info_aes_gcm_256 *gcm_256_info;
2251         struct tls12_crypto_info_aes_ccm_128 *ccm_128_info;
2252         struct tls_sw_context_tx *sw_ctx_tx = NULL;
2253         struct tls_sw_context_rx *sw_ctx_rx = NULL;
2254         struct cipher_context *cctx;
2255         struct crypto_aead **aead;
2256         struct strp_callbacks cb;
2257         u16 nonce_size, tag_size, iv_size, rec_seq_size, salt_size;
2258         struct crypto_tfm *tfm;
2259         char *iv, *rec_seq, *key, *salt, *cipher_name;
2260         size_t keysize;
2261         int rc = 0;
2262
2263         if (!ctx) {
2264                 rc = -EINVAL;
2265                 goto out;
2266         }
2267
2268         if (tx) {
2269                 if (!ctx->priv_ctx_tx) {
2270                         sw_ctx_tx = kzalloc(sizeof(*sw_ctx_tx), GFP_KERNEL);
2271                         if (!sw_ctx_tx) {
2272                                 rc = -ENOMEM;
2273                                 goto out;
2274                         }
2275                         ctx->priv_ctx_tx = sw_ctx_tx;
2276                 } else {
2277                         sw_ctx_tx =
2278                                 (struct tls_sw_context_tx *)ctx->priv_ctx_tx;
2279                 }
2280         } else {
2281                 if (!ctx->priv_ctx_rx) {
2282                         sw_ctx_rx = kzalloc(sizeof(*sw_ctx_rx), GFP_KERNEL);
2283                         if (!sw_ctx_rx) {
2284                                 rc = -ENOMEM;
2285                                 goto out;
2286                         }
2287                         ctx->priv_ctx_rx = sw_ctx_rx;
2288                 } else {
2289                         sw_ctx_rx =
2290                                 (struct tls_sw_context_rx *)ctx->priv_ctx_rx;
2291                 }
2292         }
2293
2294         if (tx) {
2295                 crypto_init_wait(&sw_ctx_tx->async_wait);
2296                 crypto_info = &ctx->crypto_send.info;
2297                 cctx = &ctx->tx;
2298                 aead = &sw_ctx_tx->aead_send;
2299                 INIT_LIST_HEAD(&sw_ctx_tx->tx_list);
2300                 INIT_DELAYED_WORK(&sw_ctx_tx->tx_work.work, tx_work_handler);
2301                 sw_ctx_tx->tx_work.sk = sk;
2302         } else {
2303                 crypto_init_wait(&sw_ctx_rx->async_wait);
2304                 crypto_info = &ctx->crypto_recv.info;
2305                 cctx = &ctx->rx;
2306                 skb_queue_head_init(&sw_ctx_rx->rx_list);
2307                 aead = &sw_ctx_rx->aead_recv;
2308         }
2309
2310         switch (crypto_info->cipher_type) {
2311         case TLS_CIPHER_AES_GCM_128: {
2312                 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2313                 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
2314                 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2315                 iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
2316                 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
2317                 rec_seq =
2318                  ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
2319                 gcm_128_info =
2320                         (struct tls12_crypto_info_aes_gcm_128 *)crypto_info;
2321                 keysize = TLS_CIPHER_AES_GCM_128_KEY_SIZE;
2322                 key = gcm_128_info->key;
2323                 salt = gcm_128_info->salt;
2324                 salt_size = TLS_CIPHER_AES_GCM_128_SALT_SIZE;
2325                 cipher_name = "gcm(aes)";
2326                 break;
2327         }
2328         case TLS_CIPHER_AES_GCM_256: {
2329                 nonce_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2330                 tag_size = TLS_CIPHER_AES_GCM_256_TAG_SIZE;
2331                 iv_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2332                 iv = ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->iv;
2333                 rec_seq_size = TLS_CIPHER_AES_GCM_256_REC_SEQ_SIZE;
2334                 rec_seq =
2335                  ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->rec_seq;
2336                 gcm_256_info =
2337                         (struct tls12_crypto_info_aes_gcm_256 *)crypto_info;
2338                 keysize = TLS_CIPHER_AES_GCM_256_KEY_SIZE;
2339                 key = gcm_256_info->key;
2340                 salt = gcm_256_info->salt;
2341                 salt_size = TLS_CIPHER_AES_GCM_256_SALT_SIZE;
2342                 cipher_name = "gcm(aes)";
2343                 break;
2344         }
2345         case TLS_CIPHER_AES_CCM_128: {
2346                 nonce_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2347                 tag_size = TLS_CIPHER_AES_CCM_128_TAG_SIZE;
2348                 iv_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2349                 iv = ((struct tls12_crypto_info_aes_ccm_128 *)crypto_info)->iv;
2350                 rec_seq_size = TLS_CIPHER_AES_CCM_128_REC_SEQ_SIZE;
2351                 rec_seq =
2352                 ((struct tls12_crypto_info_aes_ccm_128 *)crypto_info)->rec_seq;
2353                 ccm_128_info =
2354                 (struct tls12_crypto_info_aes_ccm_128 *)crypto_info;
2355                 keysize = TLS_CIPHER_AES_CCM_128_KEY_SIZE;
2356                 key = ccm_128_info->key;
2357                 salt = ccm_128_info->salt;
2358                 salt_size = TLS_CIPHER_AES_CCM_128_SALT_SIZE;
2359                 cipher_name = "ccm(aes)";
2360                 break;
2361         }
2362         default:
2363                 rc = -EINVAL;
2364                 goto free_priv;
2365         }
2366
2367         /* Sanity-check the sizes for stack allocations. */
2368         if (iv_size > MAX_IV_SIZE || nonce_size > MAX_IV_SIZE ||
2369             rec_seq_size > TLS_MAX_REC_SEQ_SIZE) {
2370                 rc = -EINVAL;
2371                 goto free_priv;
2372         }
2373
2374         if (crypto_info->version == TLS_1_3_VERSION) {
2375                 nonce_size = 0;
2376                 prot->aad_size = TLS_HEADER_SIZE;
2377                 prot->tail_size = 1;
2378         } else {
2379                 prot->aad_size = TLS_AAD_SPACE_SIZE;
2380                 prot->tail_size = 0;
2381         }
2382
2383         prot->version = crypto_info->version;
2384         prot->cipher_type = crypto_info->cipher_type;
2385         prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
2386         prot->tag_size = tag_size;
2387         prot->overhead_size = prot->prepend_size +
2388                               prot->tag_size + prot->tail_size;
2389         prot->iv_size = iv_size;
2390         prot->salt_size = salt_size;
2391         cctx->iv = kmalloc(iv_size + salt_size, GFP_KERNEL);
2392         if (!cctx->iv) {
2393                 rc = -ENOMEM;
2394                 goto free_priv;
2395         }
2396         /* Note: 128 & 256 bit salt are the same size */
2397         prot->rec_seq_size = rec_seq_size;
2398         memcpy(cctx->iv, salt, salt_size);
2399         memcpy(cctx->iv + salt_size, iv, iv_size);
2400         cctx->rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
2401         if (!cctx->rec_seq) {
2402                 rc = -ENOMEM;
2403                 goto free_iv;
2404         }
2405
2406         if (!*aead) {
2407                 *aead = crypto_alloc_aead(cipher_name, 0, 0);
2408                 if (IS_ERR(*aead)) {
2409                         rc = PTR_ERR(*aead);
2410                         *aead = NULL;
2411                         goto free_rec_seq;
2412                 }
2413         }
2414
2415         ctx->push_pending_record = tls_sw_push_pending_record;
2416
2417         rc = crypto_aead_setkey(*aead, key, keysize);
2418
2419         if (rc)
2420                 goto free_aead;
2421
2422         rc = crypto_aead_setauthsize(*aead, prot->tag_size);
2423         if (rc)
2424                 goto free_aead;
2425
2426         if (sw_ctx_rx) {
2427                 tfm = crypto_aead_tfm(sw_ctx_rx->aead_recv);
2428
2429                 if (crypto_info->version == TLS_1_3_VERSION)
2430                         sw_ctx_rx->async_capable = 0;
2431                 else
2432                         sw_ctx_rx->async_capable =
2433                                 !!(tfm->__crt_alg->cra_flags &
2434                                    CRYPTO_ALG_ASYNC);
2435
2436                 /* Set up strparser */
2437                 memset(&cb, 0, sizeof(cb));
2438                 cb.rcv_msg = tls_queue;
2439                 cb.parse_msg = tls_read_size;
2440
2441                 strp_init(&sw_ctx_rx->strp, sk, &cb);
2442         }
2443
2444         goto out;
2445
2446 free_aead:
2447         crypto_free_aead(*aead);
2448         *aead = NULL;
2449 free_rec_seq:
2450         kfree(cctx->rec_seq);
2451         cctx->rec_seq = NULL;
2452 free_iv:
2453         kfree(cctx->iv);
2454         cctx->iv = NULL;
2455 free_priv:
2456         if (tx) {
2457                 kfree(ctx->priv_ctx_tx);
2458                 ctx->priv_ctx_tx = NULL;
2459         } else {
2460                 kfree(ctx->priv_ctx_rx);
2461                 ctx->priv_ctx_rx = NULL;
2462         }
2463 out:
2464         return rc;
2465 }