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_context *tls_ctx, 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 (tls_ctx->prot_info.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)
 135                                 return -EBADMSG;
 136                         err = skb_copy_bits(skb,
 137                                             rxm->offset + rxm->full_len - back,
 138                                             &content_type, 1);
 139                         if (content_type)
 140                                 break;
 141                         sub++;
 142                         back++;
 143                 }
 144                 ctx->control = content_type;
 145         }
 146         return sub;
 147 }
 148
 149 static void tls_decrypt_done(struct crypto_async_request *req, int err)
 150 {
 151         struct aead_request *aead_req = (struct aead_request *)req;
 152         struct scatterlist *sgout = aead_req->dst;
 153         struct scatterlist *sgin = aead_req->src;
 154         struct tls_sw_context_rx *ctx;
 155         struct tls_context *tls_ctx;
 156         struct tls_prot_info *prot;
 157         struct scatterlist *sg;
 158         struct sk_buff *skb;
 159         unsigned int pages;
 160         int pending;
 161
 162         skb = (struct sk_buff *)req->data;
 163         tls_ctx = tls_get_ctx(skb->sk);
 164         ctx = tls_sw_ctx_rx(tls_ctx);
 165         prot = &tls_ctx->prot_info;
 166
 167         /* Propagate if there was an err */
 168         if (err) {
 169                 ctx->async_wait.err = err;
 170                 tls_err_abort(skb->sk, err);
 171         } else {
 172                 struct strp_msg *rxm = strp_msg(skb);
 173                 rxm->full_len -= padding_length(ctx, tls_ctx, skb);
 174                 rxm->offset += prot->prepend_size;
 175                 rxm->full_len -= prot->overhead_size;
 176         }
 177
 178         /* After using skb->sk to propagate sk through crypto async callback
 179          * we need to NULL it again.
 180          */
 181         skb->sk = NULL;
 182
 183
 184         /* Free the destination pages if skb was not decrypted inplace */
 185         if (sgout != sgin) {
 186                 /* Skip the first S/G entry as it points to AAD */
 187                 for_each_sg(sg_next(sgout), sg, UINT_MAX, pages) {
 188                         if (!sg)
 189                                 break;
 190                         put_page(sg_page(sg));
 191                 }
 192         }
 193
 194         kfree(aead_req);
 195
 196         pending = atomic_dec_return(&ctx->decrypt_pending);
 197
 198         if (!pending && READ_ONCE(ctx->async_notify))
 199                 complete(&ctx->async_wait.completion);
 200 }
 201
 202 static int tls_do_decryption(struct sock *sk,
 203                              struct sk_buff *skb,
 204                              struct scatterlist *sgin,
 205                              struct scatterlist *sgout,
 206                              char *iv_recv,
 207                              size_t data_len,
 208                              struct aead_request *aead_req,
 209                              bool async)
 210 {
 211         struct tls_context *tls_ctx = tls_get_ctx(sk);
 212         struct tls_prot_info *prot = &tls_ctx->prot_info;
 213         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
 214         int ret;
 215
 216         aead_request_set_tfm(aead_req, ctx->aead_recv);
 217         aead_request_set_ad(aead_req, prot->aad_size);
 218         aead_request_set_crypt(aead_req, sgin, sgout,
 219                                data_len + prot->tag_size,
 220                                (u8 *)iv_recv);
 221
 222         if (async) {
 223                 /* Using skb->sk to push sk through to crypto async callback
 224                  * handler. This allows propagating errors up to the socket
 225                  * if needed. It _must_ be cleared in the async handler
 226                  * before consume_skb is called. We _know_ skb->sk is NULL
 227                  * because it is a clone from strparser.
 228                  */
 229                 skb->sk = sk;
 230                 aead_request_set_callback(aead_req,
 231                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 232                                           tls_decrypt_done, skb);
 233                 atomic_inc(&ctx->decrypt_pending);
 234         } else {
 235                 aead_request_set_callback(aead_req,
 236                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 237                                           crypto_req_done, &ctx->async_wait);
 238         }
 239
 240         ret = crypto_aead_decrypt(aead_req);
 241         if (ret == -EINPROGRESS) {
 242                 if (async)
 243                         return ret;
 244
 245                 ret = crypto_wait_req(ret, &ctx->async_wait);
 246         }
 247
 248         if (async)
 249                 atomic_dec(&ctx->decrypt_pending);
 250
 251         return ret;
 252 }
 253
 254 static void tls_trim_both_msgs(struct sock *sk, int target_size)
 255 {
 256         struct tls_context *tls_ctx = tls_get_ctx(sk);
 257         struct tls_prot_info *prot = &tls_ctx->prot_info;
 258         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 259         struct tls_rec *rec = ctx->open_rec;
 260
 261         sk_msg_trim(sk, &rec->msg_plaintext, target_size);
 262         if (target_size > 0)
 263                 target_size += prot->overhead_size;
 264         sk_msg_trim(sk, &rec->msg_encrypted, target_size);
 265 }
 266
 267 static int tls_alloc_encrypted_msg(struct sock *sk, int len)
 268 {
 269         struct tls_context *tls_ctx = tls_get_ctx(sk);
 270         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 271         struct tls_rec *rec = ctx->open_rec;
 272         struct sk_msg *msg_en = &rec->msg_encrypted;
 273
 274         return sk_msg_alloc(sk, msg_en, len, 0);
 275 }
 276
 277 static int tls_clone_plaintext_msg(struct sock *sk, int required)
 278 {
 279         struct tls_context *tls_ctx = tls_get_ctx(sk);
 280         struct tls_prot_info *prot = &tls_ctx->prot_info;
 281         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 282         struct tls_rec *rec = ctx->open_rec;
 283         struct sk_msg *msg_pl = &rec->msg_plaintext;
 284         struct sk_msg *msg_en = &rec->msg_encrypted;
 285         int skip, len;
 286
 287         /* We add page references worth len bytes from encrypted sg
 288          * at the end of plaintext sg. It is guaranteed that msg_en
 289          * has enough required room (ensured by caller).
 290          */
 291         len = required - msg_pl->sg.size;
 292
 293         /* Skip initial bytes in msg_en's data to be able to use
 294          * same offset of both plain and encrypted data.
 295          */
 296         skip = prot->prepend_size + msg_pl->sg.size;
 297
 298         return sk_msg_clone(sk, msg_pl, msg_en, skip, len);
 299 }
 300
 301 static struct tls_rec *tls_get_rec(struct sock *sk)
 302 {
 303         struct tls_context *tls_ctx = tls_get_ctx(sk);
 304         struct tls_prot_info *prot = &tls_ctx->prot_info;
 305         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 306         struct sk_msg *msg_pl, *msg_en;
 307         struct tls_rec *rec;
 308         int mem_size;
 309
 310         mem_size = sizeof(struct tls_rec) + crypto_aead_reqsize(ctx->aead_send);
 311
 312         rec = kzalloc(mem_size, sk->sk_allocation);
 313         if (!rec)
 314                 return NULL;
 315
 316         msg_pl = &rec->msg_plaintext;
 317         msg_en = &rec->msg_encrypted;
 318
 319         sk_msg_init(msg_pl);
 320         sk_msg_init(msg_en);
 321
 322         sg_init_table(rec->sg_aead_in, 2);
 323         sg_set_buf(&rec->sg_aead_in[0], rec->aad_space, prot->aad_size);
 324         sg_unmark_end(&rec->sg_aead_in[1]);
 325
 326         sg_init_table(rec->sg_aead_out, 2);
 327         sg_set_buf(&rec->sg_aead_out[0], rec->aad_space, prot->aad_size);
 328         sg_unmark_end(&rec->sg_aead_out[1]);
 329
 330         return rec;
 331 }
 332
 333 static void tls_free_rec(struct sock *sk, struct tls_rec *rec)
 334 {
 335         sk_msg_free(sk, &rec->msg_encrypted);
 336         sk_msg_free(sk, &rec->msg_plaintext);
 337         kfree(rec);
 338 }
 339
 340 static void tls_free_open_rec(struct sock *sk)
 341 {
 342         struct tls_context *tls_ctx = tls_get_ctx(sk);
 343         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 344         struct tls_rec *rec = ctx->open_rec;
 345
 346         if (rec) {
 347                 tls_free_rec(sk, rec);
 348                 ctx->open_rec = NULL;
 349         }
 350 }
 351
 352 int tls_tx_records(struct sock *sk, int flags)
 353 {
 354         struct tls_context *tls_ctx = tls_get_ctx(sk);
 355         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 356         struct tls_rec *rec, *tmp;
 357         struct sk_msg *msg_en;
 358         int tx_flags, rc = 0;
 359
 360         if (tls_is_partially_sent_record(tls_ctx)) {
 361                 rec = list_first_entry(&ctx->tx_list,
 362                                        struct tls_rec, list);
 363
 364                 if (flags == -1)
 365                         tx_flags = rec->tx_flags;
 366                 else
 367                         tx_flags = flags;
 368
 369                 rc = tls_push_partial_record(sk, tls_ctx, tx_flags);
 370                 if (rc)
 371                         goto tx_err;
 372
 373                 /* Full record has been transmitted.
 374                  * Remove the head of tx_list
 375                  */
 376                 list_del(&rec->list);
 377                 sk_msg_free(sk, &rec->msg_plaintext);
 378                 kfree(rec);
 379         }
 380
 381         /* Tx all ready records */
 382         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
 383                 if (READ_ONCE(rec->tx_ready)) {
 384                         if (flags == -1)
 385                                 tx_flags = rec->tx_flags;
 386                         else
 387                                 tx_flags = flags;
 388
 389                         msg_en = &rec->msg_encrypted;
 390                         rc = tls_push_sg(sk, tls_ctx,
 391                                          &msg_en->sg.data[msg_en->sg.curr],
 392                                          0, tx_flags);
 393                         if (rc)
 394                                 goto tx_err;
 395
 396                         list_del(&rec->list);
 397                         sk_msg_free(sk, &rec->msg_plaintext);
 398                         kfree(rec);
 399                 } else {
 400                         break;
 401                 }
 402         }
 403
 404 tx_err:
 405         if (rc < 0 && rc != -EAGAIN)
 406                 tls_err_abort(sk, EBADMSG);
 407
 408         return rc;
 409 }
 410
 411 static void tls_encrypt_done(struct crypto_async_request *req, int err)
 412 {
 413         struct aead_request *aead_req = (struct aead_request *)req;
 414         struct sock *sk = req->data;
 415         struct tls_context *tls_ctx = tls_get_ctx(sk);
 416         struct tls_prot_info *prot = &tls_ctx->prot_info;
 417         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 418         struct scatterlist *sge;
 419         struct sk_msg *msg_en;
 420         struct tls_rec *rec;
 421         bool ready = false;
 422         int pending;
 423
 424         rec = container_of(aead_req, struct tls_rec, aead_req);
 425         msg_en = &rec->msg_encrypted;
 426
 427         sge = sk_msg_elem(msg_en, msg_en->sg.curr);
 428         sge->offset -= prot->prepend_size;
 429         sge->length += prot->prepend_size;
 430
 431         /* Check if error is previously set on socket */
 432         if (err || sk->sk_err) {
 433                 rec = NULL;
 434
 435                 /* If err is already set on socket, return the same code */
 436                 if (sk->sk_err) {
 437                         ctx->async_wait.err = sk->sk_err;
 438                 } else {
 439                         ctx->async_wait.err = err;
 440                         tls_err_abort(sk, err);
 441                 }
 442         }
 443
 444         if (rec) {
 445                 struct tls_rec *first_rec;
 446
 447                 /* Mark the record as ready for transmission */
 448                 smp_store_mb(rec->tx_ready, true);
 449
 450                 /* If received record is at head of tx_list, schedule tx */
 451                 first_rec = list_first_entry(&ctx->tx_list,
 452                                              struct tls_rec, list);
 453                 if (rec == first_rec)
 454                         ready = true;
 455         }
 456
 457         pending = atomic_dec_return(&ctx->encrypt_pending);
 458
 459         if (!pending && READ_ONCE(ctx->async_notify))
 460                 complete(&ctx->async_wait.completion);
 461
 462         if (!ready)
 463                 return;
 464
 465         /* Schedule the transmission */
 466         if (!test_and_set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
 467                 schedule_delayed_work(&ctx->tx_work.work, 1);
 468 }
 469
 470 static int tls_do_encryption(struct sock *sk,
 471                              struct tls_context *tls_ctx,
 472                              struct tls_sw_context_tx *ctx,
 473                              struct aead_request *aead_req,
 474                              size_t data_len, u32 start)
 475 {
 476         struct tls_prot_info *prot = &tls_ctx->prot_info;
 477         struct tls_rec *rec = ctx->open_rec;
 478         struct sk_msg *msg_en = &rec->msg_encrypted;
 479         struct scatterlist *sge = sk_msg_elem(msg_en, start);
 480         int rc, iv_offset = 0;
 481
 482         /* For CCM based ciphers, first byte of IV is a constant */
 483         if (prot->cipher_type == TLS_CIPHER_AES_CCM_128) {
 484                 rec->iv_data[0] = TLS_AES_CCM_IV_B0_BYTE;
 485                 iv_offset = 1;
 486         }
 487
 488         memcpy(&rec->iv_data[iv_offset], tls_ctx->tx.iv,
 489                prot->iv_size + prot->salt_size);
 490
 491         xor_iv_with_seq(prot->version, rec->iv_data, tls_ctx->tx.rec_seq);
 492
 493         sge->offset += prot->prepend_size;
 494         sge->length -= prot->prepend_size;
 495
 496         msg_en->sg.curr = start;
 497
 498         aead_request_set_tfm(aead_req, ctx->aead_send);
 499         aead_request_set_ad(aead_req, prot->aad_size);
 500         aead_request_set_crypt(aead_req, rec->sg_aead_in,
 501                                rec->sg_aead_out,
 502                                data_len, rec->iv_data);
 503
 504         aead_request_set_callback(aead_req, CRYPTO_TFM_REQ_MAY_BACKLOG,
 505                                   tls_encrypt_done, sk);
 506
 507         /* Add the record in tx_list */
 508         list_add_tail((struct list_head *)&rec->list, &ctx->tx_list);
 509         atomic_inc(&ctx->encrypt_pending);
 510
 511         rc = crypto_aead_encrypt(aead_req);
 512         if (!rc || rc != -EINPROGRESS) {
 513                 atomic_dec(&ctx->encrypt_pending);
 514                 sge->offset -= prot->prepend_size;
 515                 sge->length += prot->prepend_size;
 516         }
 517
 518         if (!rc) {
 519                 WRITE_ONCE(rec->tx_ready, true);
 520         } else if (rc != -EINPROGRESS) {
 521                 list_del(&rec->list);
 522                 return rc;
 523         }
 524
 525         /* Unhook the record from context if encryption is not failure */
 526         ctx->open_rec = NULL;
 527         tls_advance_record_sn(sk, &tls_ctx->tx, prot->version);
 528         return rc;
 529 }
 530
 531 static int tls_split_open_record(struct sock *sk, struct tls_rec *from,
 532                                  struct tls_rec **to, struct sk_msg *msg_opl,
 533                                  struct sk_msg *msg_oen, u32 split_point,
 534                                  u32 tx_overhead_size, u32 *orig_end)
 535 {
 536         u32 i, j, bytes = 0, apply = msg_opl->apply_bytes;
 537         struct scatterlist *sge, *osge, *nsge;
 538         u32 orig_size = msg_opl->sg.size;
 539         struct scatterlist tmp = { };
 540         struct sk_msg *msg_npl;
 541         struct tls_rec *new;
 542         int ret;
 543
 544         new = tls_get_rec(sk);
 545         if (!new)
 546                 return -ENOMEM;
 547         ret = sk_msg_alloc(sk, &new->msg_encrypted, msg_opl->sg.size +
 548                            tx_overhead_size, 0);
 549         if (ret < 0) {
 550                 tls_free_rec(sk, new);
 551                 return ret;
 552         }
 553
 554         *orig_end = msg_opl->sg.end;
 555         i = msg_opl->sg.start;
 556         sge = sk_msg_elem(msg_opl, i);
 557         while (apply && sge->length) {
 558                 if (sge->length > apply) {
 559                         u32 len = sge->length - apply;
 560
 561                         get_page(sg_page(sge));
 562                         sg_set_page(&tmp, sg_page(sge), len,
 563                                     sge->offset + apply);
 564                         sge->length = apply;
 565                         bytes += apply;
 566                         apply = 0;
 567                 } else {
 568                         apply -= sge->length;
 569                         bytes += sge->length;
 570                 }
 571
 572                 sk_msg_iter_var_next(i);
 573                 if (i == msg_opl->sg.end)
 574                         break;
 575                 sge = sk_msg_elem(msg_opl, i);
 576         }
 577
 578         msg_opl->sg.end = i;
 579         msg_opl->sg.curr = i;
 580         msg_opl->sg.copybreak = 0;
 581         msg_opl->apply_bytes = 0;
 582         msg_opl->sg.size = bytes;
 583
 584         msg_npl = &new->msg_plaintext;
 585         msg_npl->apply_bytes = apply;
 586         msg_npl->sg.size = orig_size - bytes;
 587
 588         j = msg_npl->sg.start;
 589         nsge = sk_msg_elem(msg_npl, j);
 590         if (tmp.length) {
 591                 memcpy(nsge, &tmp, sizeof(*nsge));
 592                 sk_msg_iter_var_next(j);
 593                 nsge = sk_msg_elem(msg_npl, j);
 594         }
 595
 596         osge = sk_msg_elem(msg_opl, i);
 597         while (osge->length) {
 598                 memcpy(nsge, osge, sizeof(*nsge));
 599                 sg_unmark_end(nsge);
 600                 sk_msg_iter_var_next(i);
 601                 sk_msg_iter_var_next(j);
 602                 if (i == *orig_end)
 603                         break;
 604                 osge = sk_msg_elem(msg_opl, i);
 605                 nsge = sk_msg_elem(msg_npl, j);
 606         }
 607
 608         msg_npl->sg.end = j;
 609         msg_npl->sg.curr = j;
 610         msg_npl->sg.copybreak = 0;
 611
 612         *to = new;
 613         return 0;
 614 }
 615
 616 static void tls_merge_open_record(struct sock *sk, struct tls_rec *to,
 617                                   struct tls_rec *from, u32 orig_end)
 618 {
 619         struct sk_msg *msg_npl = &from->msg_plaintext;
 620         struct sk_msg *msg_opl = &to->msg_plaintext;
 621         struct scatterlist *osge, *nsge;
 622         u32 i, j;
 623
 624         i = msg_opl->sg.end;
 625         sk_msg_iter_var_prev(i);
 626         j = msg_npl->sg.start;
 627
 628         osge = sk_msg_elem(msg_opl, i);
 629         nsge = sk_msg_elem(msg_npl, j);
 630
 631         if (sg_page(osge) == sg_page(nsge) &&
 632             osge->offset + osge->length == nsge->offset) {
 633                 osge->length += nsge->length;
 634                 put_page(sg_page(nsge));
 635         }
 636
 637         msg_opl->sg.end = orig_end;
 638         msg_opl->sg.curr = orig_end;
 639         msg_opl->sg.copybreak = 0;
 640         msg_opl->apply_bytes = msg_opl->sg.size + msg_npl->sg.size;
 641         msg_opl->sg.size += msg_npl->sg.size;
 642
 643         sk_msg_free(sk, &to->msg_encrypted);
 644         sk_msg_xfer_full(&to->msg_encrypted, &from->msg_encrypted);
 645
 646         kfree(from);
 647 }
 648
 649 static int tls_push_record(struct sock *sk, int flags,
 650                            unsigned char record_type)
 651 {
 652         struct tls_context *tls_ctx = tls_get_ctx(sk);
 653         struct tls_prot_info *prot = &tls_ctx->prot_info;
 654         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 655         struct tls_rec *rec = ctx->open_rec, *tmp = NULL;
 656         u32 i, split_point, uninitialized_var(orig_end);
 657         struct sk_msg *msg_pl, *msg_en;
 658         struct aead_request *req;
 659         bool split;
 660         int rc;
 661
 662         if (!rec)
 663                 return 0;
 664
 665         msg_pl = &rec->msg_plaintext;
 666         msg_en = &rec->msg_encrypted;
 667
 668         split_point = msg_pl->apply_bytes;
 669         split = split_point && split_point < msg_pl->sg.size;
 670         if (split) {
 671                 rc = tls_split_open_record(sk, rec, &tmp, msg_pl, msg_en,
 672                                            split_point, prot->overhead_size,
 673                                            &orig_end);
 674                 if (rc < 0)
 675                         return rc;
 676                 sk_msg_trim(sk, msg_en, msg_pl->sg.size +
 677                             prot->overhead_size);
 678         }
 679
 680         rec->tx_flags = flags;
 681         req = &rec->aead_req;
 682
 683         i = msg_pl->sg.end;
 684         sk_msg_iter_var_prev(i);
 685
 686         rec->content_type = record_type;
 687         if (prot->version == TLS_1_3_VERSION) {
 688                 /* Add content type to end of message.  No padding added */
 689                 sg_set_buf(&rec->sg_content_type, &rec->content_type, 1);
 690                 sg_mark_end(&rec->sg_content_type);
 691                 sg_chain(msg_pl->sg.data, msg_pl->sg.end + 1,
 692                          &rec->sg_content_type);
 693         } else {
 694                 sg_mark_end(sk_msg_elem(msg_pl, i));
 695         }
 696
 697         i = msg_pl->sg.start;
 698         sg_chain(rec->sg_aead_in, 2, rec->inplace_crypto ?
 699                  &msg_en->sg.data[i] : &msg_pl->sg.data[i]);
 700
 701         i = msg_en->sg.end;
 702         sk_msg_iter_var_prev(i);
 703         sg_mark_end(sk_msg_elem(msg_en, i));
 704
 705         i = msg_en->sg.start;
 706         sg_chain(rec->sg_aead_out, 2, &msg_en->sg.data[i]);
 707
 708         tls_make_aad(rec->aad_space, msg_pl->sg.size + prot->tail_size,
 709                      tls_ctx->tx.rec_seq, prot->rec_seq_size,
 710                      record_type, prot->version);
 711
 712         tls_fill_prepend(tls_ctx,
 713                          page_address(sg_page(&msg_en->sg.data[i])) +
 714                          msg_en->sg.data[i].offset,
 715                          msg_pl->sg.size + prot->tail_size,
 716                          record_type, prot->version);
 717
 718         tls_ctx->pending_open_record_frags = false;
 719
 720         rc = tls_do_encryption(sk, tls_ctx, ctx, req,
 721                                msg_pl->sg.size + prot->tail_size, i);
 722         if (rc < 0) {
 723                 if (rc != -EINPROGRESS) {
 724                         tls_err_abort(sk, EBADMSG);
 725                         if (split) {
 726                                 tls_ctx->pending_open_record_frags = true;
 727                                 tls_merge_open_record(sk, rec, tmp, orig_end);
 728                         }
 729                 }
 730                 ctx->async_capable = 1;
 731                 return rc;
 732         } else if (split) {
 733                 msg_pl = &tmp->msg_plaintext;
 734                 msg_en = &tmp->msg_encrypted;
 735                 sk_msg_trim(sk, msg_en, msg_pl->sg.size + prot->overhead_size);
 736                 tls_ctx->pending_open_record_frags = true;
 737                 ctx->open_rec = tmp;
 738         }
 739
 740         return tls_tx_records(sk, flags);
 741 }
 742
 743 static int bpf_exec_tx_verdict(struct sk_msg *msg, struct sock *sk,
 744                                bool full_record, u8 record_type,
 745                                size_t *copied, int flags)
 746 {
 747         struct tls_context *tls_ctx = tls_get_ctx(sk);
 748         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 749         struct sk_msg msg_redir = { };
 750         struct sk_psock *psock;
 751         struct sock *sk_redir;
 752         struct tls_rec *rec;
 753         bool enospc, policy;
 754         int err = 0, send;
 755         u32 delta = 0;
 756
 757         policy = !(flags & MSG_SENDPAGE_NOPOLICY);
 758         psock = sk_psock_get(sk);
 759         if (!psock || !policy)
 760                 return tls_push_record(sk, flags, record_type);
 761 more_data:
 762         enospc = sk_msg_full(msg);
 763         if (psock->eval == __SK_NONE) {
 764                 delta = msg->sg.size;
 765                 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
 766                 if (delta < msg->sg.size)
 767                         delta -= msg->sg.size;
 768                 else
 769                         delta = 0;
 770         }
 771         if (msg->cork_bytes && msg->cork_bytes > msg->sg.size &&
 772             !enospc && !full_record) {
 773                 err = -ENOSPC;
 774                 goto out_err;
 775         }
 776         msg->cork_bytes = 0;
 777         send = msg->sg.size;
 778         if (msg->apply_bytes && msg->apply_bytes < send)
 779                 send = msg->apply_bytes;
 780
 781         switch (psock->eval) {
 782         case __SK_PASS:
 783                 err = tls_push_record(sk, flags, record_type);
 784                 if (err < 0) {
 785                         *copied -= sk_msg_free(sk, msg);
 786                         tls_free_open_rec(sk);
 787                         goto out_err;
 788                 }
 789                 break;
 790         case __SK_REDIRECT:
 791                 sk_redir = psock->sk_redir;
 792                 memcpy(&msg_redir, msg, sizeof(*msg));
 793                 if (msg->apply_bytes < send)
 794                         msg->apply_bytes = 0;
 795                 else
 796                         msg->apply_bytes -= send;
 797                 sk_msg_return_zero(sk, msg, send);
 798                 msg->sg.size -= send;
 799                 release_sock(sk);
 800                 err = tcp_bpf_sendmsg_redir(sk_redir, &msg_redir, send, flags);
 801                 lock_sock(sk);
 802                 if (err < 0) {
 803                         *copied -= sk_msg_free_nocharge(sk, &msg_redir);
 804                         msg->sg.size = 0;
 805                 }
 806                 if (msg->sg.size == 0)
 807                         tls_free_open_rec(sk);
 808                 break;
 809         case __SK_DROP:
 810         default:
 811                 sk_msg_free_partial(sk, msg, send);
 812                 if (msg->apply_bytes < send)
 813                         msg->apply_bytes = 0;
 814                 else
 815                         msg->apply_bytes -= send;
 816                 if (msg->sg.size == 0)
 817                         tls_free_open_rec(sk);
 818                 *copied -= (send + delta);
 819                 err = -EACCES;
 820         }
 821
 822         if (likely(!err)) {
 823                 bool reset_eval = !ctx->open_rec;
 824
 825                 rec = ctx->open_rec;
 826                 if (rec) {
 827                         msg = &rec->msg_plaintext;
 828                         if (!msg->apply_bytes)
 829                                 reset_eval = true;
 830                 }
 831                 if (reset_eval) {
 832                         psock->eval = __SK_NONE;
 833                         if (psock->sk_redir) {
 834                                 sock_put(psock->sk_redir);
 835                                 psock->sk_redir = NULL;
 836                         }
 837                 }
 838                 if (rec)
 839                         goto more_data;
 840         }
 841  out_err:
 842         sk_psock_put(sk, psock);
 843         return err;
 844 }
 845
 846 static int tls_sw_push_pending_record(struct sock *sk, int flags)
 847 {
 848         struct tls_context *tls_ctx = tls_get_ctx(sk);
 849         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 850         struct tls_rec *rec = ctx->open_rec;
 851         struct sk_msg *msg_pl;
 852         size_t copied;
 853
 854         if (!rec)
 855                 return 0;
 856
 857         msg_pl = &rec->msg_plaintext;
 858         copied = msg_pl->sg.size;
 859         if (!copied)
 860                 return 0;
 861
 862         return bpf_exec_tx_verdict(msg_pl, sk, true, TLS_RECORD_TYPE_DATA,
 863                                    &copied, flags);
 864 }
 865
 866 int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 867 {
 868         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 869         struct tls_context *tls_ctx = tls_get_ctx(sk);
 870         struct tls_prot_info *prot = &tls_ctx->prot_info;
 871         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 872         bool async_capable = ctx->async_capable;
 873         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 874         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
 875         bool eor = !(msg->msg_flags & MSG_MORE);
 876         size_t try_to_copy, copied = 0;
 877         struct sk_msg *msg_pl, *msg_en;
 878         struct tls_rec *rec;
 879         int required_size;
 880         int num_async = 0;
 881         bool full_record;
 882         int record_room;
 883         int num_zc = 0;
 884         int orig_size;
 885         int ret = 0;
 886
 887         if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
 888                 return -ENOTSUPP;
 889
 890         lock_sock(sk);
 891
 892         /* Wait till there is any pending write on socket */
 893         if (unlikely(sk->sk_write_pending)) {
 894                 ret = wait_on_pending_writer(sk, &timeo);
 895                 if (unlikely(ret))
 896                         goto send_end;
 897         }
 898
 899         if (unlikely(msg->msg_controllen)) {
 900                 ret = tls_proccess_cmsg(sk, msg, &record_type);
 901                 if (ret) {
 902                         if (ret == -EINPROGRESS)
 903                                 num_async++;
 904                         else if (ret != -EAGAIN)
 905                                 goto send_end;
 906                 }
 907         }
 908
 909         while (msg_data_left(msg)) {
 910                 if (sk->sk_err) {
 911                         ret = -sk->sk_err;
 912                         goto send_end;
 913                 }
 914
 915                 if (ctx->open_rec)
 916                         rec = ctx->open_rec;
 917                 else
 918                         rec = ctx->open_rec = tls_get_rec(sk);
 919                 if (!rec) {
 920                         ret = -ENOMEM;
 921                         goto send_end;
 922                 }
 923
 924                 msg_pl = &rec->msg_plaintext;
 925                 msg_en = &rec->msg_encrypted;
 926
 927                 orig_size = msg_pl->sg.size;
 928                 full_record = false;
 929                 try_to_copy = msg_data_left(msg);
 930                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
 931                 if (try_to_copy >= record_room) {
 932                         try_to_copy = record_room;
 933                         full_record = true;
 934                 }
 935
 936                 required_size = msg_pl->sg.size + try_to_copy +
 937                                 prot->overhead_size;
 938
 939                 if (!sk_stream_memory_free(sk))
 940                         goto wait_for_sndbuf;
 941
 942 alloc_encrypted:
 943                 ret = tls_alloc_encrypted_msg(sk, required_size);
 944                 if (ret) {
 945                         if (ret != -ENOSPC)
 946                                 goto wait_for_memory;
 947
 948                         /* Adjust try_to_copy according to the amount that was
 949                          * actually allocated. The difference is due
 950                          * to max sg elements limit
 951                          */
 952                         try_to_copy -= required_size - msg_en->sg.size;
 953                         full_record = true;
 954                 }
 955
 956                 if (!is_kvec && (full_record || eor) && !async_capable) {
 957                         u32 first = msg_pl->sg.end;
 958
 959                         ret = sk_msg_zerocopy_from_iter(sk, &msg->msg_iter,
 960                                                         msg_pl, try_to_copy);
 961                         if (ret)
 962                                 goto fallback_to_reg_send;
 963
 964                         rec->inplace_crypto = 0;
 965
 966                         num_zc++;
 967                         copied += try_to_copy;
 968
 969                         sk_msg_sg_copy_set(msg_pl, first);
 970                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
 971                                                   record_type, &copied,
 972                                                   msg->msg_flags);
 973                         if (ret) {
 974                                 if (ret == -EINPROGRESS)
 975                                         num_async++;
 976                                 else if (ret == -ENOMEM)
 977                                         goto wait_for_memory;
 978                                 else if (ret == -ENOSPC)
 979                                         goto rollback_iter;
 980                                 else if (ret != -EAGAIN)
 981                                         goto send_end;
 982                         }
 983                         continue;
 984 rollback_iter:
 985                         copied -= try_to_copy;
 986                         sk_msg_sg_copy_clear(msg_pl, first);
 987                         iov_iter_revert(&msg->msg_iter,
 988                                         msg_pl->sg.size - orig_size);
 989 fallback_to_reg_send:
 990                         sk_msg_trim(sk, msg_pl, orig_size);
 991                 }
 992
 993                 required_size = msg_pl->sg.size + try_to_copy;
 994
 995                 ret = tls_clone_plaintext_msg(sk, required_size);
 996                 if (ret) {
 997                         if (ret != -ENOSPC)
 998                                 goto send_end;
 999
1000                         /* Adjust try_to_copy according to the amount that was
1001                          * actually allocated. The difference is due
1002                          * to max sg elements limit
1003                          */
1004                         try_to_copy -= required_size - msg_pl->sg.size;
1005                         full_record = true;
1006                         sk_msg_trim(sk, msg_en,
1007                                     msg_pl->sg.size + prot->overhead_size);
1008                 }
1009
1010                 if (try_to_copy) {
1011                         ret = sk_msg_memcopy_from_iter(sk, &msg->msg_iter,
1012                                                        msg_pl, try_to_copy);
1013                         if (ret < 0)
1014                                 goto trim_sgl;
1015                 }
1016
1017                 /* Open records defined only if successfully copied, otherwise
1018                  * we would trim the sg but not reset the open record frags.
1019                  */
1020                 tls_ctx->pending_open_record_frags = true;
1021                 copied += try_to_copy;
1022                 if (full_record || eor) {
1023                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1024                                                   record_type, &copied,
1025                                                   msg->msg_flags);
1026                         if (ret) {
1027                                 if (ret == -EINPROGRESS)
1028                                         num_async++;
1029                                 else if (ret == -ENOMEM)
1030                                         goto wait_for_memory;
1031                                 else if (ret != -EAGAIN) {
1032                                         if (ret == -ENOSPC)
1033                                                 ret = 0;
1034                                         goto send_end;
1035                                 }
1036                         }
1037                 }
1038
1039                 continue;
1040
1041 wait_for_sndbuf:
1042                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1043 wait_for_memory:
1044                 ret = sk_stream_wait_memory(sk, &timeo);
1045                 if (ret) {
1046 trim_sgl:
1047                         tls_trim_both_msgs(sk, orig_size);
1048                         goto send_end;
1049                 }
1050
1051                 if (msg_en->sg.size < required_size)
1052                         goto alloc_encrypted;
1053         }
1054
1055         if (!num_async) {
1056                 goto send_end;
1057         } else if (num_zc) {
1058                 /* Wait for pending encryptions to get completed */
1059                 smp_store_mb(ctx->async_notify, true);
1060
1061                 if (atomic_read(&ctx->encrypt_pending))
1062                         crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1063                 else
1064                         reinit_completion(&ctx->async_wait.completion);
1065
1066                 WRITE_ONCE(ctx->async_notify, false);
1067
1068                 if (ctx->async_wait.err) {
1069                         ret = ctx->async_wait.err;
1070                         copied = 0;
1071                 }
1072         }
1073
1074         /* Transmit if any encryptions have completed */
1075         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1076                 cancel_delayed_work(&ctx->tx_work.work);
1077                 tls_tx_records(sk, msg->msg_flags);
1078         }
1079
1080 send_end:
1081         ret = sk_stream_error(sk, msg->msg_flags, ret);
1082
1083         release_sock(sk);
1084         return copied ? copied : ret;
1085 }
1086
1087 static int tls_sw_do_sendpage(struct sock *sk, struct page *page,
1088                               int offset, size_t size, int flags)
1089 {
1090         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1091         struct tls_context *tls_ctx = tls_get_ctx(sk);
1092         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1093         struct tls_prot_info *prot = &tls_ctx->prot_info;
1094         unsigned char record_type = TLS_RECORD_TYPE_DATA;
1095         struct sk_msg *msg_pl;
1096         struct tls_rec *rec;
1097         int num_async = 0;
1098         size_t copied = 0;
1099         bool full_record;
1100         int record_room;
1101         int ret = 0;
1102         bool eor;
1103
1104         eor = !(flags & (MSG_MORE | MSG_SENDPAGE_NOTLAST));
1105         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1106
1107         /* Wait till there is any pending write on socket */
1108         if (unlikely(sk->sk_write_pending)) {
1109                 ret = wait_on_pending_writer(sk, &timeo);
1110                 if (unlikely(ret))
1111                         goto sendpage_end;
1112         }
1113
1114         /* Call the sk_stream functions to manage the sndbuf mem. */
1115         while (size > 0) {
1116                 size_t copy, required_size;
1117
1118                 if (sk->sk_err) {
1119                         ret = -sk->sk_err;
1120                         goto sendpage_end;
1121                 }
1122
1123                 if (ctx->open_rec)
1124                         rec = ctx->open_rec;
1125                 else
1126                         rec = ctx->open_rec = tls_get_rec(sk);
1127                 if (!rec) {
1128                         ret = -ENOMEM;
1129                         goto sendpage_end;
1130                 }
1131
1132                 msg_pl = &rec->msg_plaintext;
1133
1134                 full_record = false;
1135                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
1136                 copied = 0;
1137                 copy = size;
1138                 if (copy >= record_room) {
1139                         copy = record_room;
1140                         full_record = true;
1141                 }
1142
1143                 required_size = msg_pl->sg.size + copy + prot->overhead_size;
1144
1145                 if (!sk_stream_memory_free(sk))
1146                         goto wait_for_sndbuf;
1147 alloc_payload:
1148                 ret = tls_alloc_encrypted_msg(sk, required_size);
1149                 if (ret) {
1150                         if (ret != -ENOSPC)
1151                                 goto wait_for_memory;
1152
1153                         /* Adjust copy according to the amount that was
1154                          * actually allocated. The difference is due
1155                          * to max sg elements limit
1156                          */
1157                         copy -= required_size - msg_pl->sg.size;
1158                         full_record = true;
1159                 }
1160
1161                 sk_msg_page_add(msg_pl, page, copy, offset);
1162                 sk_mem_charge(sk, copy);
1163
1164                 offset += copy;
1165                 size -= copy;
1166                 copied += copy;
1167
1168                 tls_ctx->pending_open_record_frags = true;
1169                 if (full_record || eor || sk_msg_full(msg_pl)) {
1170                         rec->inplace_crypto = 0;
1171                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1172                                                   record_type, &copied, flags);
1173                         if (ret) {
1174                                 if (ret == -EINPROGRESS)
1175                                         num_async++;
1176                                 else if (ret == -ENOMEM)
1177                                         goto wait_for_memory;
1178                                 else if (ret != -EAGAIN) {
1179                                         if (ret == -ENOSPC)
1180                                                 ret = 0;
1181                                         goto sendpage_end;
1182                                 }
1183                         }
1184                 }
1185                 continue;
1186 wait_for_sndbuf:
1187                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1188 wait_for_memory:
1189                 ret = sk_stream_wait_memory(sk, &timeo);
1190                 if (ret) {
1191                         tls_trim_both_msgs(sk, msg_pl->sg.size);
1192                         goto sendpage_end;
1193                 }
1194
1195                 goto alloc_payload;
1196         }
1197
1198         if (num_async) {
1199                 /* Transmit if any encryptions have completed */
1200                 if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1201                         cancel_delayed_work(&ctx->tx_work.work);
1202                         tls_tx_records(sk, flags);
1203                 }
1204         }
1205 sendpage_end:
1206         ret = sk_stream_error(sk, flags, ret);
1207         return copied ? copied : ret;
1208 }
1209
1210 int tls_sw_sendpage(struct sock *sk, struct page *page,
1211                     int offset, size_t size, int flags)
1212 {
1213         int ret;
1214
1215         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1216                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY))
1217                 return -ENOTSUPP;
1218
1219         lock_sock(sk);
1220         ret = tls_sw_do_sendpage(sk, page, offset, size, flags);
1221         release_sock(sk);
1222         return ret;
1223 }
1224
1225 static struct sk_buff *tls_wait_data(struct sock *sk, struct sk_psock *psock,
1226                                      int flags, long timeo, int *err)
1227 {
1228         struct tls_context *tls_ctx = tls_get_ctx(sk);
1229         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1230         struct sk_buff *skb;
1231         DEFINE_WAIT_FUNC(wait, woken_wake_function);
1232
1233         while (!(skb = ctx->recv_pkt) && sk_psock_queue_empty(psock)) {
1234                 if (sk->sk_err) {
1235                         *err = sock_error(sk);
1236                         return NULL;
1237                 }
1238
1239                 if (sk->sk_shutdown & RCV_SHUTDOWN)
1240                         return NULL;
1241
1242                 if (sock_flag(sk, SOCK_DONE))
1243                         return NULL;
1244
1245                 if ((flags & MSG_DONTWAIT) || !timeo) {
1246                         *err = -EAGAIN;
1247                         return NULL;
1248                 }
1249
1250                 add_wait_queue(sk_sleep(sk), &wait);
1251                 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1252                 sk_wait_event(sk, &timeo,
1253                               ctx->recv_pkt != skb ||
1254                               !sk_psock_queue_empty(psock),
1255                               &wait);
1256                 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1257                 remove_wait_queue(sk_sleep(sk), &wait);
1258
1259                 /* Handle signals */
1260                 if (signal_pending(current)) {
1261                         *err = sock_intr_errno(timeo);
1262                         return NULL;
1263                 }
1264         }
1265
1266         return skb;
1267 }
1268
1269 static int tls_setup_from_iter(struct sock *sk, struct iov_iter *from,
1270                                int length, int *pages_used,
1271                                unsigned int *size_used,
1272                                struct scatterlist *to,
1273                                int to_max_pages)
1274 {
1275         int rc = 0, i = 0, num_elem = *pages_used, maxpages;
1276         struct page *pages[MAX_SKB_FRAGS];
1277         unsigned int size = *size_used;
1278         ssize_t copied, use;
1279         size_t offset;
1280
1281         while (length > 0) {
1282                 i = 0;
1283                 maxpages = to_max_pages - num_elem;
1284                 if (maxpages == 0) {
1285                         rc = -EFAULT;
1286                         goto out;
1287                 }
1288                 copied = iov_iter_get_pages(from, pages,
1289                                             length,
1290                                             maxpages, &offset);
1291                 if (copied <= 0) {
1292                         rc = -EFAULT;
1293                         goto out;
1294                 }
1295
1296                 iov_iter_advance(from, copied);
1297
1298                 length -= copied;
1299                 size += copied;
1300                 while (copied) {
1301                         use = min_t(int, copied, PAGE_SIZE - offset);
1302
1303                         sg_set_page(&to[num_elem],
1304                                     pages[i], use, offset);
1305                         sg_unmark_end(&to[num_elem]);
1306                         /* We do not uncharge memory from this API */
1307
1308                         offset = 0;
1309                         copied -= use;
1310
1311                         i++;
1312                         num_elem++;
1313                 }
1314         }
1315         /* Mark the end in the last sg entry if newly added */
1316         if (num_elem > *pages_used)
1317                 sg_mark_end(&to[num_elem - 1]);
1318 out:
1319         if (rc)
1320                 iov_iter_revert(from, size - *size_used);
1321         *size_used = size;
1322         *pages_used = num_elem;
1323
1324         return rc;
1325 }
1326
1327 /* This function decrypts the input skb into either out_iov or in out_sg
1328  * or in skb buffers itself. The input parameter 'zc' indicates if
1329  * zero-copy mode needs to be tried or not. With zero-copy mode, either
1330  * out_iov or out_sg must be non-NULL. In case both out_iov and out_sg are
1331  * NULL, then the decryption happens inside skb buffers itself, i.e.
1332  * zero-copy gets disabled and 'zc' is updated.
1333  */
1334
1335 static int decrypt_internal(struct sock *sk, struct sk_buff *skb,
1336                             struct iov_iter *out_iov,
1337                             struct scatterlist *out_sg,
1338                             int *chunk, bool *zc, bool async)
1339 {
1340         struct tls_context *tls_ctx = tls_get_ctx(sk);
1341         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1342         struct tls_prot_info *prot = &tls_ctx->prot_info;
1343         struct strp_msg *rxm = strp_msg(skb);
1344         int n_sgin, n_sgout, nsg, mem_size, aead_size, err, pages = 0;
1345         struct aead_request *aead_req;
1346         struct sk_buff *unused;
1347         u8 *aad, *iv, *mem = NULL;
1348         struct scatterlist *sgin = NULL;
1349         struct scatterlist *sgout = NULL;
1350         const int data_len = rxm->full_len - prot->overhead_size +
1351                              prot->tail_size;
1352         int iv_offset = 0;
1353
1354         if (*zc && (out_iov || out_sg)) {
1355                 if (out_iov)
1356                         n_sgout = iov_iter_npages(out_iov, INT_MAX) + 1;
1357                 else
1358                         n_sgout = sg_nents(out_sg);
1359                 n_sgin = skb_nsg(skb, rxm->offset + prot->prepend_size,
1360                                  rxm->full_len - prot->prepend_size);
1361         } else {
1362                 n_sgout = 0;
1363                 *zc = false;
1364                 n_sgin = skb_cow_data(skb, 0, &unused);
1365         }
1366
1367         if (n_sgin < 1)
1368                 return -EBADMSG;
1369
1370         /* Increment to accommodate AAD */
1371         n_sgin = n_sgin + 1;
1372
1373         nsg = n_sgin + n_sgout;
1374
1375         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(ctx->aead_recv);
1376         mem_size = aead_size + (nsg * sizeof(struct scatterlist));
1377         mem_size = mem_size + prot->aad_size;
1378         mem_size = mem_size + crypto_aead_ivsize(ctx->aead_recv);
1379
1380         /* Allocate a single block of memory which contains
1381          * aead_req || sgin[] || sgout[] || aad || iv.
1382          * This order achieves correct alignment for aead_req, sgin, sgout.
1383          */
1384         mem = kmalloc(mem_size, sk->sk_allocation);
1385         if (!mem)
1386                 return -ENOMEM;
1387
1388         /* Segment the allocated memory */
1389         aead_req = (struct aead_request *)mem;
1390         sgin = (struct scatterlist *)(mem + aead_size);
1391         sgout = sgin + n_sgin;
1392         aad = (u8 *)(sgout + n_sgout);
1393         iv = aad + prot->aad_size;
1394
1395         /* For CCM based ciphers, first byte of nonce+iv is always '2' */
1396         if (prot->cipher_type == TLS_CIPHER_AES_CCM_128) {
1397                 iv[0] = 2;
1398                 iv_offset = 1;
1399         }
1400
1401         /* Prepare IV */
1402         err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
1403                             iv + iv_offset + prot->salt_size,
1404                             prot->iv_size);
1405         if (err < 0) {
1406                 kfree(mem);
1407                 return err;
1408         }
1409         if (prot->version == TLS_1_3_VERSION)
1410                 memcpy(iv + iv_offset, tls_ctx->rx.iv,
1411                        crypto_aead_ivsize(ctx->aead_recv));
1412         else
1413                 memcpy(iv + iv_offset, tls_ctx->rx.iv, prot->salt_size);
1414
1415         xor_iv_with_seq(prot->version, iv, tls_ctx->rx.rec_seq);
1416
1417         /* Prepare AAD */
1418         tls_make_aad(aad, rxm->full_len - prot->overhead_size +
1419                      prot->tail_size,
1420                      tls_ctx->rx.rec_seq, prot->rec_seq_size,
1421                      ctx->control, prot->version);
1422
1423         /* Prepare sgin */
1424         sg_init_table(sgin, n_sgin);
1425         sg_set_buf(&sgin[0], aad, prot->aad_size);
1426         err = skb_to_sgvec(skb, &sgin[1],
1427                            rxm->offset + prot->prepend_size,
1428                            rxm->full_len - prot->prepend_size);
1429         if (err < 0) {
1430                 kfree(mem);
1431                 return err;
1432         }
1433
1434         if (n_sgout) {
1435                 if (out_iov) {
1436                         sg_init_table(sgout, n_sgout);
1437                         sg_set_buf(&sgout[0], aad, prot->aad_size);
1438
1439                         *chunk = 0;
1440                         err = tls_setup_from_iter(sk, out_iov, data_len,
1441                                                   &pages, chunk, &sgout[1],
1442                                                   (n_sgout - 1));
1443                         if (err < 0)
1444                                 goto fallback_to_reg_recv;
1445                 } else if (out_sg) {
1446                         memcpy(sgout, out_sg, n_sgout * sizeof(*sgout));
1447                 } else {
1448                         goto fallback_to_reg_recv;
1449                 }
1450         } else {
1451 fallback_to_reg_recv:
1452                 sgout = sgin;
1453                 pages = 0;
1454                 *chunk = data_len;
1455                 *zc = false;
1456         }
1457
1458         /* Prepare and submit AEAD request */
1459         err = tls_do_decryption(sk, skb, sgin, sgout, iv,
1460                                 data_len, aead_req, async);
1461         if (err == -EINPROGRESS)
1462                 return err;
1463
1464         /* Release the pages in case iov was mapped to pages */
1465         for (; pages > 0; pages--)
1466                 put_page(sg_page(&sgout[pages]));
1467
1468         kfree(mem);
1469         return err;
1470 }
1471
1472 static int decrypt_skb_update(struct sock *sk, struct sk_buff *skb,
1473                               struct iov_iter *dest, int *chunk, bool *zc,
1474                               bool async)
1475 {
1476         struct tls_context *tls_ctx = tls_get_ctx(sk);
1477         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1478         struct tls_prot_info *prot = &tls_ctx->prot_info;
1479         int version = prot->version;
1480         struct strp_msg *rxm = strp_msg(skb);
1481         int err = 0;
1482
1483         if (!ctx->decrypted) {
1484 #ifdef CONFIG_TLS_DEVICE
1485                 err = tls_device_decrypted(sk, skb);
1486                 if (err < 0)
1487                         return err;
1488 #endif
1489                 /* Still not decrypted after tls_device */
1490                 if (!ctx->decrypted) {
1491                         err = decrypt_internal(sk, skb, dest, NULL, chunk, zc,
1492                                                async);
1493                         if (err < 0) {
1494                                 if (err == -EINPROGRESS)
1495                                         tls_advance_record_sn(sk, &tls_ctx->rx,
1496                                                               version);
1497
1498                                 return err;
1499                         }
1500                 } else {
1501                         *zc = false;
1502                 }
1503
1504                 rxm->full_len -= padding_length(ctx, tls_ctx, skb);
1505                 rxm->offset += prot->prepend_size;
1506                 rxm->full_len -= prot->overhead_size;
1507                 tls_advance_record_sn(sk, &tls_ctx->rx, version);
1508                 ctx->decrypted = true;
1509                 ctx->saved_data_ready(sk);
1510         } else {
1511                 *zc = false;
1512         }
1513
1514         return err;
1515 }
1516
1517 int decrypt_skb(struct sock *sk, struct sk_buff *skb,
1518                 struct scatterlist *sgout)
1519 {
1520         bool zc = true;
1521         int chunk;
1522
1523         return decrypt_internal(sk, skb, NULL, sgout, &chunk, &zc, false);
1524 }
1525
1526 static bool tls_sw_advance_skb(struct sock *sk, struct sk_buff *skb,
1527                                unsigned int len)
1528 {
1529         struct tls_context *tls_ctx = tls_get_ctx(sk);
1530         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1531
1532         if (skb) {
1533                 struct strp_msg *rxm = strp_msg(skb);
1534
1535                 if (len < rxm->full_len) {
1536                         rxm->offset += len;
1537                         rxm->full_len -= len;
1538                         return false;
1539                 }
1540                 consume_skb(skb);
1541         }
1542
1543         /* Finished with message */
1544         ctx->recv_pkt = NULL;
1545         __strp_unpause(&ctx->strp);
1546
1547         return true;
1548 }
1549
1550 /* This function traverses the rx_list in tls receive context to copies the
1551  * decrypted records into the buffer provided by caller zero copy is not
1552  * true. Further, the records are removed from the rx_list if it is not a peek
1553  * case and the record has been consumed completely.
1554  */
1555 static int process_rx_list(struct tls_sw_context_rx *ctx,
1556                            struct msghdr *msg,
1557                            u8 *control,
1558                            bool *cmsg,
1559                            size_t skip,
1560                            size_t len,
1561                            bool zc,
1562                            bool is_peek)
1563 {
1564         struct sk_buff *skb = skb_peek(&ctx->rx_list);
1565         u8 ctrl = *control;
1566         u8 msgc = *cmsg;
1567         struct tls_msg *tlm;
1568         ssize_t copied = 0;
1569
1570         /* Set the record type in 'control' if caller didn't pass it */
1571         if (!ctrl && skb) {
1572                 tlm = tls_msg(skb);
1573                 ctrl = tlm->control;
1574         }
1575
1576         while (skip && skb) {
1577                 struct strp_msg *rxm = strp_msg(skb);
1578                 tlm = tls_msg(skb);
1579
1580                 /* Cannot process a record of different type */
1581                 if (ctrl != tlm->control)
1582                         return 0;
1583
1584                 if (skip < rxm->full_len)
1585                         break;
1586
1587                 skip = skip - rxm->full_len;
1588                 skb = skb_peek_next(skb, &ctx->rx_list);
1589         }
1590
1591         while (len && skb) {
1592                 struct sk_buff *next_skb;
1593                 struct strp_msg *rxm = strp_msg(skb);
1594                 int chunk = min_t(unsigned int, rxm->full_len - skip, len);
1595
1596                 tlm = tls_msg(skb);
1597
1598                 /* Cannot process a record of different type */
1599                 if (ctrl != tlm->control)
1600                         return 0;
1601
1602                 /* Set record type if not already done. For a non-data record,
1603                  * do not proceed if record type could not be copied.
1604                  */
1605                 if (!msgc) {
1606                         int cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1607                                             sizeof(ctrl), &ctrl);
1608                         msgc = true;
1609                         if (ctrl != TLS_RECORD_TYPE_DATA) {
1610                                 if (cerr || msg->msg_flags & MSG_CTRUNC)
1611                                         return -EIO;
1612
1613                                 *cmsg = msgc;
1614                         }
1615                 }
1616
1617                 if (!zc || (rxm->full_len - skip) > len) {
1618                         int err = skb_copy_datagram_msg(skb, rxm->offset + skip,
1619                                                     msg, chunk);
1620                         if (err < 0)
1621                                 return err;
1622                 }
1623
1624                 len = len - chunk;
1625                 copied = copied + chunk;
1626
1627                 /* Consume the data from record if it is non-peek case*/
1628                 if (!is_peek) {
1629                         rxm->offset = rxm->offset + chunk;
1630                         rxm->full_len = rxm->full_len - chunk;
1631
1632                         /* Return if there is unconsumed data in the record */
1633                         if (rxm->full_len - skip)
1634                                 break;
1635                 }
1636
1637                 /* The remaining skip-bytes must lie in 1st record in rx_list.
1638                  * So from the 2nd record, 'skip' should be 0.
1639                  */
1640                 skip = 0;
1641
1642                 if (msg)
1643                         msg->msg_flags |= MSG_EOR;
1644
1645                 next_skb = skb_peek_next(skb, &ctx->rx_list);
1646
1647                 if (!is_peek) {
1648                         skb_unlink(skb, &ctx->rx_list);
1649                         consume_skb(skb);
1650                 }
1651
1652                 skb = next_skb;
1653         }
1654
1655         *control = ctrl;
1656         return copied;
1657 }
1658
1659 int tls_sw_recvmsg(struct sock *sk,
1660                    struct msghdr *msg,
1661                    size_t len,
1662                    int nonblock,
1663                    int flags,
1664                    int *addr_len)
1665 {
1666         struct tls_context *tls_ctx = tls_get_ctx(sk);
1667         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1668         struct tls_prot_info *prot = &tls_ctx->prot_info;
1669         struct sk_psock *psock;
1670         unsigned char control = 0;
1671         ssize_t decrypted = 0;
1672         struct strp_msg *rxm;
1673         struct tls_msg *tlm;
1674         struct sk_buff *skb;
1675         ssize_t copied = 0;
1676         bool cmsg = false;
1677         int target, err = 0;
1678         long timeo;
1679         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
1680         bool is_peek = flags & MSG_PEEK;
1681         int num_async = 0;
1682
1683         flags |= nonblock;
1684
1685         if (unlikely(flags & MSG_ERRQUEUE))
1686                 return sock_recv_errqueue(sk, msg, len, SOL_IP, IP_RECVERR);
1687
1688         psock = sk_psock_get(sk);
1689         lock_sock(sk);
1690
1691         /* Process pending decrypted records. It must be non-zero-copy */
1692         err = process_rx_list(ctx, msg, &control, &cmsg, 0, len, false,
1693                               is_peek);
1694         if (err < 0) {
1695                 tls_err_abort(sk, err);
1696                 goto end;
1697         } else {
1698                 copied = err;
1699         }
1700
1701         len = len - copied;
1702         if (len) {
1703                 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1704                 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1705         } else {
1706                 goto recv_end;
1707         }
1708
1709         do {
1710                 bool retain_skb = false;
1711                 bool zc = false;
1712                 int to_decrypt;
1713                 int chunk = 0;
1714                 bool async_capable;
1715                 bool async = false;
1716
1717                 skb = tls_wait_data(sk, psock, flags, timeo, &err);
1718                 if (!skb) {
1719                         if (psock) {
1720                                 int ret = __tcp_bpf_recvmsg(sk, psock,
1721                                                             msg, len, flags);
1722
1723                                 if (ret > 0) {
1724                                         decrypted += ret;
1725                                         len -= ret;
1726                                         continue;
1727                                 }
1728                         }
1729                         goto recv_end;
1730                 } else {
1731                         tlm = tls_msg(skb);
1732                         if (prot->version == TLS_1_3_VERSION)
1733                                 tlm->control = 0;
1734                         else
1735                                 tlm->control = ctx->control;
1736                 }
1737
1738                 rxm = strp_msg(skb);
1739
1740                 to_decrypt = rxm->full_len - prot->overhead_size;
1741
1742                 if (to_decrypt <= len && !is_kvec && !is_peek &&
1743                     ctx->control == TLS_RECORD_TYPE_DATA &&
1744                     prot->version != TLS_1_3_VERSION)
1745                         zc = true;
1746
1747                 /* Do not use async mode if record is non-data */
1748                 if (ctx->control == TLS_RECORD_TYPE_DATA)
1749                         async_capable = ctx->async_capable;
1750                 else
1751                         async_capable = false;
1752
1753                 err = decrypt_skb_update(sk, skb, &msg->msg_iter,
1754                                          &chunk, &zc, async_capable);
1755                 if (err < 0 && err != -EINPROGRESS) {
1756                         tls_err_abort(sk, EBADMSG);
1757                         goto recv_end;
1758                 }
1759
1760                 if (err == -EINPROGRESS) {
1761                         async = true;
1762                         num_async++;
1763                 } else if (prot->version == TLS_1_3_VERSION) {
1764                         tlm->control = ctx->control;
1765                 }
1766
1767                 /* If the type of records being processed is not known yet,
1768                  * set it to record type just dequeued. If it is already known,
1769                  * but does not match the record type just dequeued, go to end.
1770                  * We always get record type here since for tls1.2, record type
1771                  * is known just after record is dequeued from stream parser.
1772                  * For tls1.3, we disable async.
1773                  */
1774
1775                 if (!control)
1776                         control = tlm->control;
1777                 else if (control != tlm->control)
1778                         goto recv_end;
1779
1780                 if (!cmsg) {
1781                         int cerr;
1782
1783                         cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1784                                         sizeof(control), &control);
1785                         cmsg = true;
1786                         if (control != TLS_RECORD_TYPE_DATA) {
1787                                 if (cerr || msg->msg_flags & MSG_CTRUNC) {
1788                                         err = -EIO;
1789                                         goto recv_end;
1790                                 }
1791                         }
1792                 }
1793
1794                 if (async)
1795                         goto pick_next_record;
1796
1797                 if (!zc) {
1798                         if (rxm->full_len > len) {
1799                                 retain_skb = true;
1800                                 chunk = len;
1801                         } else {
1802                                 chunk = rxm->full_len;
1803                         }
1804
1805                         err = skb_copy_datagram_msg(skb, rxm->offset,
1806                                                     msg, chunk);
1807                         if (err < 0)
1808                                 goto recv_end;
1809
1810                         if (!is_peek) {
1811                                 rxm->offset = rxm->offset + chunk;
1812                                 rxm->full_len = rxm->full_len - chunk;
1813                         }
1814                 }
1815
1816 pick_next_record:
1817                 if (chunk > len)
1818                         chunk = len;
1819
1820                 decrypted += chunk;
1821                 len -= chunk;
1822
1823                 /* For async or peek case, queue the current skb */
1824                 if (async || is_peek || retain_skb) {
1825                         skb_queue_tail(&ctx->rx_list, skb);
1826                         skb = NULL;
1827                 }
1828
1829                 if (tls_sw_advance_skb(sk, skb, chunk)) {
1830                         /* Return full control message to
1831                          * userspace before trying to parse
1832                          * another message type
1833                          */
1834                         msg->msg_flags |= MSG_EOR;
1835                         if (ctx->control != TLS_RECORD_TYPE_DATA)
1836                                 goto recv_end;
1837                 } else {
1838                         break;
1839                 }
1840
1841                 /* If we have a new message from strparser, continue now. */
1842                 if (decrypted >= target && !ctx->recv_pkt)
1843                         break;
1844         } while (len);
1845
1846 recv_end:
1847         if (num_async) {
1848                 /* Wait for all previously submitted records to be decrypted */
1849                 smp_store_mb(ctx->async_notify, true);
1850                 if (atomic_read(&ctx->decrypt_pending)) {
1851                         err = crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1852                         if (err) {
1853                                 /* one of async decrypt failed */
1854                                 tls_err_abort(sk, err);
1855                                 copied = 0;
1856                                 decrypted = 0;
1857                                 goto end;
1858                         }
1859                 } else {
1860                         reinit_completion(&ctx->async_wait.completion);
1861                 }
1862                 WRITE_ONCE(ctx->async_notify, false);
1863
1864                 /* Drain records from the rx_list & copy if required */
1865                 if (is_peek || is_kvec)
1866                         err = process_rx_list(ctx, msg, &control, &cmsg, copied,
1867                                               decrypted, false, is_peek);
1868                 else
1869                         err = process_rx_list(ctx, msg, &control, &cmsg, 0,
1870                                               decrypted, true, is_peek);
1871                 if (err < 0) {
1872                         tls_err_abort(sk, err);
1873                         copied = 0;
1874                         goto end;
1875                 }
1876         }
1877
1878         copied += decrypted;
1879
1880 end:
1881         release_sock(sk);
1882         if (psock)
1883                 sk_psock_put(sk, psock);
1884         return copied ? : err;
1885 }
1886
1887 ssize_t tls_sw_splice_read(struct socket *sock,  loff_t *ppos,
1888                            struct pipe_inode_info *pipe,
1889                            size_t len, unsigned int flags)
1890 {
1891         struct tls_context *tls_ctx = tls_get_ctx(sock->sk);
1892         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1893         struct strp_msg *rxm = NULL;
1894         struct sock *sk = sock->sk;
1895         struct sk_buff *skb;
1896         ssize_t copied = 0;
1897         int err = 0;
1898         long timeo;
1899         int chunk;
1900         bool zc = false;
1901
1902         lock_sock(sk);
1903
1904         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1905
1906         skb = tls_wait_data(sk, NULL, flags, timeo, &err);
1907         if (!skb)
1908                 goto splice_read_end;
1909
1910         if (!ctx->decrypted) {
1911                 err = decrypt_skb_update(sk, skb, NULL, &chunk, &zc, false);
1912
1913                 /* splice does not support reading control messages */
1914                 if (ctx->control != TLS_RECORD_TYPE_DATA) {
1915                         err = -ENOTSUPP;
1916                         goto splice_read_end;
1917                 }
1918
1919                 if (err < 0) {
1920                         tls_err_abort(sk, EBADMSG);
1921                         goto splice_read_end;
1922                 }
1923                 ctx->decrypted = true;
1924         }
1925         rxm = strp_msg(skb);
1926
1927         chunk = min_t(unsigned int, rxm->full_len, len);
1928         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, chunk, flags);
1929         if (copied < 0)
1930                 goto splice_read_end;
1931
1932         if (likely(!(flags & MSG_PEEK)))
1933                 tls_sw_advance_skb(sk, skb, copied);
1934
1935 splice_read_end:
1936         release_sock(sk);
1937         return copied ? : err;
1938 }
1939
1940 bool tls_sw_stream_read(const struct sock *sk)
1941 {
1942         struct tls_context *tls_ctx = tls_get_ctx(sk);
1943         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1944         bool ingress_empty = true;
1945         struct sk_psock *psock;
1946
1947         rcu_read_lock();
1948         psock = sk_psock(sk);
1949         if (psock)
1950                 ingress_empty = list_empty(&psock->ingress_msg);
1951         rcu_read_unlock();
1952
1953         return !ingress_empty || ctx->recv_pkt;
1954 }
1955
1956 static int tls_read_size(struct strparser *strp, struct sk_buff *skb)
1957 {
1958         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
1959         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1960         struct tls_prot_info *prot = &tls_ctx->prot_info;
1961         char header[TLS_HEADER_SIZE + MAX_IV_SIZE];
1962         struct strp_msg *rxm = strp_msg(skb);
1963         size_t cipher_overhead;
1964         size_t data_len = 0;
1965         int ret;
1966
1967         /* Verify that we have a full TLS header, or wait for more data */
1968         if (rxm->offset + prot->prepend_size > skb->len)
1969                 return 0;
1970
1971         /* Sanity-check size of on-stack buffer. */
1972         if (WARN_ON(prot->prepend_size > sizeof(header))) {
1973                 ret = -EINVAL;
1974                 goto read_failure;
1975         }
1976
1977         /* Linearize header to local buffer */
1978         ret = skb_copy_bits(skb, rxm->offset, header, prot->prepend_size);
1979
1980         if (ret < 0)
1981                 goto read_failure;
1982
1983         ctx->control = header[0];
1984
1985         data_len = ((header[4] & 0xFF) | (header[3] << 8));
1986
1987         cipher_overhead = prot->tag_size;
1988         if (prot->version != TLS_1_3_VERSION)
1989                 cipher_overhead += prot->iv_size;
1990
1991         if (data_len > TLS_MAX_PAYLOAD_SIZE + cipher_overhead +
1992             prot->tail_size) {
1993                 ret = -EMSGSIZE;
1994                 goto read_failure;
1995         }
1996         if (data_len < cipher_overhead) {
1997                 ret = -EBADMSG;
1998                 goto read_failure;
1999         }
2000
2001         /* Note that both TLS1.3 and TLS1.2 use TLS_1_2 version here */
2002         if (header[1] != TLS_1_2_VERSION_MINOR ||
2003             header[2] != TLS_1_2_VERSION_MAJOR) {
2004                 ret = -EINVAL;
2005                 goto read_failure;
2006         }
2007 #ifdef CONFIG_TLS_DEVICE
2008         handle_device_resync(strp->sk, TCP_SKB_CB(skb)->seq + rxm->offset,
2009                              *(u64*)tls_ctx->rx.rec_seq);
2010 #endif
2011         return data_len + TLS_HEADER_SIZE;
2012
2013 read_failure:
2014         tls_err_abort(strp->sk, ret);
2015
2016         return ret;
2017 }
2018
2019 static void tls_queue(struct strparser *strp, struct sk_buff *skb)
2020 {
2021         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2022         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2023
2024         ctx->decrypted = false;
2025
2026         ctx->recv_pkt = skb;
2027         strp_pause(strp);
2028
2029         ctx->saved_data_ready(strp->sk);
2030 }
2031
2032 static void tls_data_ready(struct sock *sk)
2033 {
2034         struct tls_context *tls_ctx = tls_get_ctx(sk);
2035         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2036         struct sk_psock *psock;
2037
2038         strp_data_ready(&ctx->strp);
2039
2040         psock = sk_psock_get(sk);
2041         if (psock && !list_empty(&psock->ingress_msg)) {
2042                 ctx->saved_data_ready(sk);
2043                 sk_psock_put(sk, psock);
2044         }
2045 }
2046
2047 void tls_sw_free_resources_tx(struct sock *sk)
2048 {
2049         struct tls_context *tls_ctx = tls_get_ctx(sk);
2050         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2051         struct tls_rec *rec, *tmp;
2052
2053         /* Wait for any pending async encryptions to complete */
2054         smp_store_mb(ctx->async_notify, true);
2055         if (atomic_read(&ctx->encrypt_pending))
2056                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
2057
2058         release_sock(sk);
2059         cancel_delayed_work_sync(&ctx->tx_work.work);
2060         lock_sock(sk);
2061
2062         /* Tx whatever records we can transmit and abandon the rest */
2063         tls_tx_records(sk, -1);
2064
2065         /* Free up un-sent records in tx_list. First, free
2066          * the partially sent record if any at head of tx_list.
2067          */
2068         if (tls_free_partial_record(sk, tls_ctx)) {
2069                 rec = list_first_entry(&ctx->tx_list,
2070                                        struct tls_rec, list);
2071                 list_del(&rec->list);
2072                 sk_msg_free(sk, &rec->msg_plaintext);
2073                 kfree(rec);
2074         }
2075
2076         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
2077                 list_del(&rec->list);
2078                 sk_msg_free(sk, &rec->msg_encrypted);
2079                 sk_msg_free(sk, &rec->msg_plaintext);
2080                 kfree(rec);
2081         }
2082
2083         crypto_free_aead(ctx->aead_send);
2084         tls_free_open_rec(sk);
2085
2086         kfree(ctx);
2087 }
2088
2089 void tls_sw_release_resources_rx(struct sock *sk)
2090 {
2091         struct tls_context *tls_ctx = tls_get_ctx(sk);
2092         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2093
2094         if (ctx->aead_recv) {
2095                 kfree_skb(ctx->recv_pkt);
2096                 ctx->recv_pkt = NULL;
2097                 skb_queue_purge(&ctx->rx_list);
2098                 crypto_free_aead(ctx->aead_recv);
2099                 strp_stop(&ctx->strp);
2100                 write_lock_bh(&sk->sk_callback_lock);
2101                 sk->sk_data_ready = ctx->saved_data_ready;
2102                 write_unlock_bh(&sk->sk_callback_lock);
2103                 release_sock(sk);
2104                 strp_done(&ctx->strp);
2105                 lock_sock(sk);
2106         }
2107 }
2108
2109 void tls_sw_free_resources_rx(struct sock *sk)
2110 {
2111         struct tls_context *tls_ctx = tls_get_ctx(sk);
2112         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2113
2114         tls_sw_release_resources_rx(sk);
2115
2116         kfree(ctx);
2117 }
2118
2119 /* The work handler to transmitt the encrypted records in tx_list */
2120 static void tx_work_handler(struct work_struct *work)
2121 {
2122         struct delayed_work *delayed_work = to_delayed_work(work);
2123         struct tx_work *tx_work = container_of(delayed_work,
2124                                                struct tx_work, work);
2125         struct sock *sk = tx_work->sk;
2126         struct tls_context *tls_ctx = tls_get_ctx(sk);
2127         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2128
2129         if (!test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
2130                 return;
2131
2132         lock_sock(sk);
2133         tls_tx_records(sk, -1);
2134         release_sock(sk);
2135 }
2136
2137 void tls_sw_write_space(struct sock *sk, struct tls_context *ctx)
2138 {
2139         struct tls_sw_context_tx *tx_ctx = tls_sw_ctx_tx(ctx);
2140
2141         /* Schedule the transmission if tx list is ready */
2142         if (is_tx_ready(tx_ctx) && !sk->sk_write_pending) {
2143                 /* Schedule the transmission */
2144                 if (!test_and_set_bit(BIT_TX_SCHEDULED,
2145                                       &tx_ctx->tx_bitmask))
2146                         schedule_delayed_work(&tx_ctx->tx_work.work, 0);
2147         }
2148 }
2149
2150 int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
2151 {
2152         struct tls_context *tls_ctx = tls_get_ctx(sk);
2153         struct tls_prot_info *prot = &tls_ctx->prot_info;
2154         struct tls_crypto_info *crypto_info;
2155         struct tls12_crypto_info_aes_gcm_128 *gcm_128_info;
2156         struct tls12_crypto_info_aes_gcm_256 *gcm_256_info;
2157         struct tls12_crypto_info_aes_ccm_128 *ccm_128_info;
2158         struct tls_sw_context_tx *sw_ctx_tx = NULL;
2159         struct tls_sw_context_rx *sw_ctx_rx = NULL;
2160         struct cipher_context *cctx;
2161         struct crypto_aead **aead;
2162         struct strp_callbacks cb;
2163         u16 nonce_size, tag_size, iv_size, rec_seq_size, salt_size;
2164         struct crypto_tfm *tfm;
2165         char *iv, *rec_seq, *key, *salt, *cipher_name;
2166         size_t keysize;
2167         int rc = 0;
2168
2169         if (!ctx) {
2170                 rc = -EINVAL;
2171                 goto out;
2172         }
2173
2174         if (tx) {
2175                 if (!ctx->priv_ctx_tx) {
2176                         sw_ctx_tx = kzalloc(sizeof(*sw_ctx_tx), GFP_KERNEL);
2177                         if (!sw_ctx_tx) {
2178                                 rc = -ENOMEM;
2179                                 goto out;
2180                         }
2181                         ctx->priv_ctx_tx = sw_ctx_tx;
2182                 } else {
2183                         sw_ctx_tx =
2184                                 (struct tls_sw_context_tx *)ctx->priv_ctx_tx;
2185                 }
2186         } else {
2187                 if (!ctx->priv_ctx_rx) {
2188                         sw_ctx_rx = kzalloc(sizeof(*sw_ctx_rx), GFP_KERNEL);
2189                         if (!sw_ctx_rx) {
2190                                 rc = -ENOMEM;
2191                                 goto out;
2192                         }
2193                         ctx->priv_ctx_rx = sw_ctx_rx;
2194                 } else {
2195                         sw_ctx_rx =
2196                                 (struct tls_sw_context_rx *)ctx->priv_ctx_rx;
2197                 }
2198         }
2199
2200         if (tx) {
2201                 crypto_init_wait(&sw_ctx_tx->async_wait);
2202                 crypto_info = &ctx->crypto_send.info;
2203                 cctx = &ctx->tx;
2204                 aead = &sw_ctx_tx->aead_send;
2205                 INIT_LIST_HEAD(&sw_ctx_tx->tx_list);
2206                 INIT_DELAYED_WORK(&sw_ctx_tx->tx_work.work, tx_work_handler);
2207                 sw_ctx_tx->tx_work.sk = sk;
2208         } else {
2209                 crypto_init_wait(&sw_ctx_rx->async_wait);
2210                 crypto_info = &ctx->crypto_recv.info;
2211                 cctx = &ctx->rx;
2212                 skb_queue_head_init(&sw_ctx_rx->rx_list);
2213                 aead = &sw_ctx_rx->aead_recv;
2214         }
2215
2216         switch (crypto_info->cipher_type) {
2217         case TLS_CIPHER_AES_GCM_128: {
2218                 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2219                 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
2220                 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2221                 iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
2222                 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
2223                 rec_seq =
2224                  ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
2225                 gcm_128_info =
2226                         (struct tls12_crypto_info_aes_gcm_128 *)crypto_info;
2227                 keysize = TLS_CIPHER_AES_GCM_128_KEY_SIZE;
2228                 key = gcm_128_info->key;
2229                 salt = gcm_128_info->salt;
2230                 salt_size = TLS_CIPHER_AES_GCM_128_SALT_SIZE;
2231                 cipher_name = "gcm(aes)";
2232                 break;
2233         }
2234         case TLS_CIPHER_AES_GCM_256: {
2235                 nonce_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2236                 tag_size = TLS_CIPHER_AES_GCM_256_TAG_SIZE;
2237                 iv_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2238                 iv = ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->iv;
2239                 rec_seq_size = TLS_CIPHER_AES_GCM_256_REC_SEQ_SIZE;
2240                 rec_seq =
2241                  ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->rec_seq;
2242                 gcm_256_info =
2243                         (struct tls12_crypto_info_aes_gcm_256 *)crypto_info;
2244                 keysize = TLS_CIPHER_AES_GCM_256_KEY_SIZE;
2245                 key = gcm_256_info->key;
2246                 salt = gcm_256_info->salt;
2247                 salt_size = TLS_CIPHER_AES_GCM_256_SALT_SIZE;
2248                 cipher_name = "gcm(aes)";
2249                 break;
2250         }
2251         case TLS_CIPHER_AES_CCM_128: {
2252                 nonce_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2253                 tag_size = TLS_CIPHER_AES_CCM_128_TAG_SIZE;
2254                 iv_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2255                 iv = ((struct tls12_crypto_info_aes_ccm_128 *)crypto_info)->iv;
2256                 rec_seq_size = TLS_CIPHER_AES_CCM_128_REC_SEQ_SIZE;
2257                 rec_seq =
2258                 ((struct tls12_crypto_info_aes_ccm_128 *)crypto_info)->rec_seq;
2259                 ccm_128_info =
2260                 (struct tls12_crypto_info_aes_ccm_128 *)crypto_info;
2261                 keysize = TLS_CIPHER_AES_CCM_128_KEY_SIZE;
2262                 key = ccm_128_info->key;
2263                 salt = ccm_128_info->salt;
2264                 salt_size = TLS_CIPHER_AES_CCM_128_SALT_SIZE;
2265                 cipher_name = "ccm(aes)";
2266                 break;
2267         }
2268         default:
2269                 rc = -EINVAL;
2270                 goto free_priv;
2271         }
2272
2273         /* Sanity-check the IV size for stack allocations. */
2274         if (iv_size > MAX_IV_SIZE || nonce_size > MAX_IV_SIZE) {
2275                 rc = -EINVAL;
2276                 goto free_priv;
2277         }
2278
2279         if (crypto_info->version == TLS_1_3_VERSION) {
2280                 nonce_size = 0;
2281                 prot->aad_size = TLS_HEADER_SIZE;
2282                 prot->tail_size = 1;
2283         } else {
2284                 prot->aad_size = TLS_AAD_SPACE_SIZE;
2285                 prot->tail_size = 0;
2286         }
2287
2288         prot->version = crypto_info->version;
2289         prot->cipher_type = crypto_info->cipher_type;
2290         prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
2291         prot->tag_size = tag_size;
2292         prot->overhead_size = prot->prepend_size +
2293                               prot->tag_size + prot->tail_size;
2294         prot->iv_size = iv_size;
2295         prot->salt_size = salt_size;
2296         cctx->iv = kmalloc(iv_size + salt_size, GFP_KERNEL);
2297         if (!cctx->iv) {
2298                 rc = -ENOMEM;
2299                 goto free_priv;
2300         }
2301         /* Note: 128 & 256 bit salt are the same size */
2302         prot->rec_seq_size = rec_seq_size;
2303         memcpy(cctx->iv, salt, salt_size);
2304         memcpy(cctx->iv + salt_size, iv, iv_size);
2305         cctx->rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
2306         if (!cctx->rec_seq) {
2307                 rc = -ENOMEM;
2308                 goto free_iv;
2309         }
2310
2311         if (!*aead) {
2312                 *aead = crypto_alloc_aead(cipher_name, 0, 0);
2313                 if (IS_ERR(*aead)) {
2314                         rc = PTR_ERR(*aead);
2315                         *aead = NULL;
2316                         goto free_rec_seq;
2317                 }
2318         }
2319
2320         ctx->push_pending_record = tls_sw_push_pending_record;
2321
2322         rc = crypto_aead_setkey(*aead, key, keysize);
2323
2324         if (rc)
2325                 goto free_aead;
2326
2327         rc = crypto_aead_setauthsize(*aead, prot->tag_size);
2328         if (rc)
2329                 goto free_aead;
2330
2331         if (sw_ctx_rx) {
2332                 tfm = crypto_aead_tfm(sw_ctx_rx->aead_recv);
2333
2334                 if (crypto_info->version == TLS_1_3_VERSION)
2335                         sw_ctx_rx->async_capable = false;
2336                 else
2337                         sw_ctx_rx->async_capable =
2338                                 tfm->__crt_alg->cra_flags & CRYPTO_ALG_ASYNC;
2339
2340                 /* Set up strparser */
2341                 memset(&cb, 0, sizeof(cb));
2342                 cb.rcv_msg = tls_queue;
2343                 cb.parse_msg = tls_read_size;
2344
2345                 strp_init(&sw_ctx_rx->strp, sk, &cb);
2346
2347                 write_lock_bh(&sk->sk_callback_lock);
2348                 sw_ctx_rx->saved_data_ready = sk->sk_data_ready;
2349                 sk->sk_data_ready = tls_data_ready;
2350                 write_unlock_bh(&sk->sk_callback_lock);
2351
2352                 strp_check_rcv(&sw_ctx_rx->strp);
2353         }
2354
2355         goto out;
2356
2357 free_aead:
2358         crypto_free_aead(*aead);
2359         *aead = NULL;
2360 free_rec_seq:
2361         kfree(cctx->rec_seq);
2362         cctx->rec_seq = NULL;
2363 free_iv:
2364         kfree(cctx->iv);
2365         cctx->iv = NULL;
2366 free_priv:
2367         if (tx) {
2368                 kfree(ctx->priv_ctx_tx);
2369                 ctx->priv_ctx_tx = NULL;
2370         } else {
2371                 kfree(ctx->priv_ctx_rx);
2372                 ctx->priv_ctx_rx = NULL;
2373         }
2374 out:
2375         return rc;
2376 }