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 #define MAX_IV_SIZE     TLS_CIPHER_AES_GCM_128_IV_SIZE
  46
  47 static int __skb_nsg(struct sk_buff *skb, int offset, int len,
  48                      unsigned int recursion_level)
  49 {
  50         int start = skb_headlen(skb);
  51         int i, chunk = start - offset;
  52         struct sk_buff *frag_iter;
  53         int elt = 0;
  54
  55         if (unlikely(recursion_level >= 24))
  56                 return -EMSGSIZE;
  57
  58         if (chunk > 0) {
  59                 if (chunk > len)
  60                         chunk = len;
  61                 elt++;
  62                 len -= chunk;
  63                 if (len == 0)
  64                         return elt;
  65                 offset += chunk;
  66         }
  67
  68         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
  69                 int end;
  70
  71                 WARN_ON(start > offset + len);
  72
  73                 end = start + skb_frag_size(&skb_shinfo(skb)->frags[i]);
  74                 chunk = end - offset;
  75                 if (chunk > 0) {
  76                         if (chunk > len)
  77                                 chunk = len;
  78                         elt++;
  79                         len -= chunk;
  80                         if (len == 0)
  81                                 return elt;
  82                         offset += chunk;
  83                 }
  84                 start = end;
  85         }
  86
  87         if (unlikely(skb_has_frag_list(skb))) {
  88                 skb_walk_frags(skb, frag_iter) {
  89                         int end, ret;
  90
  91                         WARN_ON(start > offset + len);
  92
  93                         end = start + frag_iter->len;
  94                         chunk = end - offset;
  95                         if (chunk > 0) {
  96                                 if (chunk > len)
  97                                         chunk = len;
  98                                 ret = __skb_nsg(frag_iter, offset - start, chunk,
  99                                                 recursion_level + 1);
 100                                 if (unlikely(ret < 0))
 101                                         return ret;
 102                                 elt += ret;
 103                                 len -= chunk;
 104                                 if (len == 0)
 105                                         return elt;
 106                                 offset += chunk;
 107                         }
 108                         start = end;
 109                 }
 110         }
 111         BUG_ON(len);
 112         return elt;
 113 }
 114
 115 /* Return the number of scatterlist elements required to completely map the
 116  * skb, or -EMSGSIZE if the recursion depth is exceeded.
 117  */
 118 static int skb_nsg(struct sk_buff *skb, int offset, int len)
 119 {
 120         return __skb_nsg(skb, offset, len, 0);
 121 }
 122
 123 static void tls_decrypt_done(struct crypto_async_request *req, int err)
 124 {
 125         struct aead_request *aead_req = (struct aead_request *)req;
 126         struct scatterlist *sgout = aead_req->dst;
 127         struct tls_sw_context_rx *ctx;
 128         struct tls_context *tls_ctx;
 129         struct scatterlist *sg;
 130         struct sk_buff *skb;
 131         unsigned int pages;
 132         int pending;
 133
 134         skb = (struct sk_buff *)req->data;
 135         tls_ctx = tls_get_ctx(skb->sk);
 136         ctx = tls_sw_ctx_rx(tls_ctx);
 137         pending = atomic_dec_return(&ctx->decrypt_pending);
 138
 139         /* Propagate if there was an err */
 140         if (err) {
 141                 ctx->async_wait.err = err;
 142                 tls_err_abort(skb->sk, err);
 143         }
 144
 145         /* After using skb->sk to propagate sk through crypto async callback
 146          * we need to NULL it again.
 147          */
 148         skb->sk = NULL;
 149
 150         /* Release the skb, pages and memory allocated for crypto req */
 151         kfree_skb(skb);
 152
 153         /* Skip the first S/G entry as it points to AAD */
 154         for_each_sg(sg_next(sgout), sg, UINT_MAX, pages) {
 155                 if (!sg)
 156                         break;
 157                 put_page(sg_page(sg));
 158         }
 159
 160         kfree(aead_req);
 161
 162         if (!pending && READ_ONCE(ctx->async_notify))
 163                 complete(&ctx->async_wait.completion);
 164 }
 165
 166 static int tls_do_decryption(struct sock *sk,
 167                              struct sk_buff *skb,
 168                              struct scatterlist *sgin,
 169                              struct scatterlist *sgout,
 170                              char *iv_recv,
 171                              size_t data_len,
 172                              struct aead_request *aead_req,
 173                              bool async)
 174 {
 175         struct tls_context *tls_ctx = tls_get_ctx(sk);
 176         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
 177         int ret;
 178
 179         aead_request_set_tfm(aead_req, ctx->aead_recv);
 180         aead_request_set_ad(aead_req, TLS_AAD_SPACE_SIZE);
 181         aead_request_set_crypt(aead_req, sgin, sgout,
 182                                data_len + tls_ctx->rx.tag_size,
 183                                (u8 *)iv_recv);
 184
 185         if (async) {
 186                 /* Using skb->sk to push sk through to crypto async callback
 187                  * handler. This allows propagating errors up to the socket
 188                  * if needed. It _must_ be cleared in the async handler
 189                  * before kfree_skb is called. We _know_ skb->sk is NULL
 190                  * because it is a clone from strparser.
 191                  */
 192                 skb->sk = sk;
 193                 aead_request_set_callback(aead_req,
 194                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 195                                           tls_decrypt_done, skb);
 196                 atomic_inc(&ctx->decrypt_pending);
 197         } else {
 198                 aead_request_set_callback(aead_req,
 199                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 200                                           crypto_req_done, &ctx->async_wait);
 201         }
 202
 203         ret = crypto_aead_decrypt(aead_req);
 204         if (ret == -EINPROGRESS) {
 205                 if (async)
 206                         return ret;
 207
 208                 ret = crypto_wait_req(ret, &ctx->async_wait);
 209         }
 210
 211         if (async)
 212                 atomic_dec(&ctx->decrypt_pending);
 213
 214         return ret;
 215 }
 216
 217 static void tls_trim_both_msgs(struct sock *sk, int target_size)
 218 {
 219         struct tls_context *tls_ctx = tls_get_ctx(sk);
 220         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 221         struct tls_rec *rec = ctx->open_rec;
 222
 223         sk_msg_trim(sk, &rec->msg_plaintext, target_size);
 224         if (target_size > 0)
 225                 target_size += tls_ctx->tx.overhead_size;
 226         sk_msg_trim(sk, &rec->msg_encrypted, target_size);
 227 }
 228
 229 static int tls_alloc_encrypted_msg(struct sock *sk, int len)
 230 {
 231         struct tls_context *tls_ctx = tls_get_ctx(sk);
 232         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 233         struct tls_rec *rec = ctx->open_rec;
 234         struct sk_msg *msg_en = &rec->msg_encrypted;
 235
 236         return sk_msg_alloc(sk, msg_en, len, 0);
 237 }
 238
 239 static int tls_clone_plaintext_msg(struct sock *sk, int required)
 240 {
 241         struct tls_context *tls_ctx = tls_get_ctx(sk);
 242         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 243         struct tls_rec *rec = ctx->open_rec;
 244         struct sk_msg *msg_pl = &rec->msg_plaintext;
 245         struct sk_msg *msg_en = &rec->msg_encrypted;
 246         int skip, len;
 247
 248         /* We add page references worth len bytes from encrypted sg
 249          * at the end of plaintext sg. It is guaranteed that msg_en
 250          * has enough required room (ensured by caller).
 251          */
 252         len = required - msg_pl->sg.size;
 253
 254         /* Skip initial bytes in msg_en's data to be able to use
 255          * same offset of both plain and encrypted data.
 256          */
 257         skip = tls_ctx->tx.prepend_size + msg_pl->sg.size;
 258
 259         return sk_msg_clone(sk, msg_pl, msg_en, skip, len);
 260 }
 261
 262 static struct tls_rec *tls_get_rec(struct sock *sk)
 263 {
 264         struct tls_context *tls_ctx = tls_get_ctx(sk);
 265         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 266         struct sk_msg *msg_pl, *msg_en;
 267         struct tls_rec *rec;
 268         int mem_size;
 269
 270         mem_size = sizeof(struct tls_rec) + crypto_aead_reqsize(ctx->aead_send);
 271
 272         rec = kzalloc(mem_size, sk->sk_allocation);
 273         if (!rec)
 274                 return NULL;
 275
 276         msg_pl = &rec->msg_plaintext;
 277         msg_en = &rec->msg_encrypted;
 278
 279         sk_msg_init(msg_pl);
 280         sk_msg_init(msg_en);
 281
 282         sg_init_table(rec->sg_aead_in, 2);
 283         sg_set_buf(&rec->sg_aead_in[0], rec->aad_space,
 284                    sizeof(rec->aad_space));
 285         sg_unmark_end(&rec->sg_aead_in[1]);
 286
 287         sg_init_table(rec->sg_aead_out, 2);
 288         sg_set_buf(&rec->sg_aead_out[0], rec->aad_space,
 289                    sizeof(rec->aad_space));
 290         sg_unmark_end(&rec->sg_aead_out[1]);
 291
 292         return rec;
 293 }
 294
 295 static void tls_free_rec(struct sock *sk, struct tls_rec *rec)
 296 {
 297         sk_msg_free(sk, &rec->msg_encrypted);
 298         sk_msg_free(sk, &rec->msg_plaintext);
 299         kfree(rec);
 300 }
 301
 302 static void tls_free_open_rec(struct sock *sk)
 303 {
 304         struct tls_context *tls_ctx = tls_get_ctx(sk);
 305         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 306         struct tls_rec *rec = ctx->open_rec;
 307
 308         if (rec) {
 309                 tls_free_rec(sk, rec);
 310                 ctx->open_rec = NULL;
 311         }
 312 }
 313
 314 int tls_tx_records(struct sock *sk, int flags)
 315 {
 316         struct tls_context *tls_ctx = tls_get_ctx(sk);
 317         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 318         struct tls_rec *rec, *tmp;
 319         struct sk_msg *msg_en;
 320         int tx_flags, rc = 0;
 321
 322         if (tls_is_partially_sent_record(tls_ctx)) {
 323                 rec = list_first_entry(&ctx->tx_list,
 324                                        struct tls_rec, list);
 325
 326                 if (flags == -1)
 327                         tx_flags = rec->tx_flags;
 328                 else
 329                         tx_flags = flags;
 330
 331                 rc = tls_push_partial_record(sk, tls_ctx, tx_flags);
 332                 if (rc)
 333                         goto tx_err;
 334
 335                 /* Full record has been transmitted.
 336                  * Remove the head of tx_list
 337                  */
 338                 list_del(&rec->list);
 339                 sk_msg_free(sk, &rec->msg_plaintext);
 340                 kfree(rec);
 341         }
 342
 343         /* Tx all ready records */
 344         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
 345                 if (READ_ONCE(rec->tx_ready)) {
 346                         if (flags == -1)
 347                                 tx_flags = rec->tx_flags;
 348                         else
 349                                 tx_flags = flags;
 350
 351                         msg_en = &rec->msg_encrypted;
 352                         rc = tls_push_sg(sk, tls_ctx,
 353                                          &msg_en->sg.data[msg_en->sg.curr],
 354                                          0, tx_flags);
 355                         if (rc)
 356                                 goto tx_err;
 357
 358                         list_del(&rec->list);
 359                         sk_msg_free(sk, &rec->msg_plaintext);
 360                         kfree(rec);
 361                 } else {
 362                         break;
 363                 }
 364         }
 365
 366 tx_err:
 367         if (rc < 0 && rc != -EAGAIN)
 368                 tls_err_abort(sk, EBADMSG);
 369
 370         return rc;
 371 }
 372
 373 static void tls_encrypt_done(struct crypto_async_request *req, int err)
 374 {
 375         struct aead_request *aead_req = (struct aead_request *)req;
 376         struct sock *sk = req->data;
 377         struct tls_context *tls_ctx = tls_get_ctx(sk);
 378         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 379         struct scatterlist *sge;
 380         struct sk_msg *msg_en;
 381         struct tls_rec *rec;
 382         bool ready = false;
 383         int pending;
 384
 385         rec = container_of(aead_req, struct tls_rec, aead_req);
 386         msg_en = &rec->msg_encrypted;
 387
 388         sge = sk_msg_elem(msg_en, msg_en->sg.curr);
 389         sge->offset -= tls_ctx->tx.prepend_size;
 390         sge->length += tls_ctx->tx.prepend_size;
 391
 392         /* Check if error is previously set on socket */
 393         if (err || sk->sk_err) {
 394                 rec = NULL;
 395
 396                 /* If err is already set on socket, return the same code */
 397                 if (sk->sk_err) {
 398                         ctx->async_wait.err = sk->sk_err;
 399                 } else {
 400                         ctx->async_wait.err = err;
 401                         tls_err_abort(sk, err);
 402                 }
 403         }
 404
 405         if (rec) {
 406                 struct tls_rec *first_rec;
 407
 408                 /* Mark the record as ready for transmission */
 409                 smp_store_mb(rec->tx_ready, true);
 410
 411                 /* If received record is at head of tx_list, schedule tx */
 412                 first_rec = list_first_entry(&ctx->tx_list,
 413                                              struct tls_rec, list);
 414                 if (rec == first_rec)
 415                         ready = true;
 416         }
 417
 418         pending = atomic_dec_return(&ctx->encrypt_pending);
 419
 420         if (!pending && READ_ONCE(ctx->async_notify))
 421                 complete(&ctx->async_wait.completion);
 422
 423         if (!ready)
 424                 return;
 425
 426         /* Schedule the transmission */
 427         if (!test_and_set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
 428                 schedule_delayed_work(&ctx->tx_work.work, 1);
 429 }
 430
 431 static int tls_do_encryption(struct sock *sk,
 432                              struct tls_context *tls_ctx,
 433                              struct tls_sw_context_tx *ctx,
 434                              struct aead_request *aead_req,
 435                              size_t data_len, u32 start)
 436 {
 437         struct tls_rec *rec = ctx->open_rec;
 438         struct sk_msg *msg_en = &rec->msg_encrypted;
 439         struct scatterlist *sge = sk_msg_elem(msg_en, start);
 440         int rc;
 441
 442         sge->offset += tls_ctx->tx.prepend_size;
 443         sge->length -= tls_ctx->tx.prepend_size;
 444
 445         msg_en->sg.curr = start;
 446
 447         aead_request_set_tfm(aead_req, ctx->aead_send);
 448         aead_request_set_ad(aead_req, TLS_AAD_SPACE_SIZE);
 449         aead_request_set_crypt(aead_req, rec->sg_aead_in,
 450                                rec->sg_aead_out,
 451                                data_len, tls_ctx->tx.iv);
 452
 453         aead_request_set_callback(aead_req, CRYPTO_TFM_REQ_MAY_BACKLOG,
 454                                   tls_encrypt_done, sk);
 455
 456         /* Add the record in tx_list */
 457         list_add_tail((struct list_head *)&rec->list, &ctx->tx_list);
 458         atomic_inc(&ctx->encrypt_pending);
 459
 460         rc = crypto_aead_encrypt(aead_req);
 461         if (!rc || rc != -EINPROGRESS) {
 462                 atomic_dec(&ctx->encrypt_pending);
 463                 sge->offset -= tls_ctx->tx.prepend_size;
 464                 sge->length += tls_ctx->tx.prepend_size;
 465         }
 466
 467         if (!rc) {
 468                 WRITE_ONCE(rec->tx_ready, true);
 469         } else if (rc != -EINPROGRESS) {
 470                 list_del(&rec->list);
 471                 return rc;
 472         }
 473
 474         /* Unhook the record from context if encryption is not failure */
 475         ctx->open_rec = NULL;
 476         tls_advance_record_sn(sk, &tls_ctx->tx);
 477         return rc;
 478 }
 479
 480 static int tls_split_open_record(struct sock *sk, struct tls_rec *from,
 481                                  struct tls_rec **to, struct sk_msg *msg_opl,
 482                                  struct sk_msg *msg_oen, u32 split_point,
 483                                  u32 tx_overhead_size, u32 *orig_end)
 484 {
 485         u32 i, j, bytes = 0, apply = msg_opl->apply_bytes;
 486         struct scatterlist *sge, *osge, *nsge;
 487         u32 orig_size = msg_opl->sg.size;
 488         struct scatterlist tmp = { };
 489         struct sk_msg *msg_npl;
 490         struct tls_rec *new;
 491         int ret;
 492
 493         new = tls_get_rec(sk);
 494         if (!new)
 495                 return -ENOMEM;
 496         ret = sk_msg_alloc(sk, &new->msg_encrypted, msg_opl->sg.size +
 497                            tx_overhead_size, 0);
 498         if (ret < 0) {
 499                 tls_free_rec(sk, new);
 500                 return ret;
 501         }
 502
 503         *orig_end = msg_opl->sg.end;
 504         i = msg_opl->sg.start;
 505         sge = sk_msg_elem(msg_opl, i);
 506         while (apply && sge->length) {
 507                 if (sge->length > apply) {
 508                         u32 len = sge->length - apply;
 509
 510                         get_page(sg_page(sge));
 511                         sg_set_page(&tmp, sg_page(sge), len,
 512                                     sge->offset + apply);
 513                         sge->length = apply;
 514                         bytes += apply;
 515                         apply = 0;
 516                 } else {
 517                         apply -= sge->length;
 518                         bytes += sge->length;
 519                 }
 520
 521                 sk_msg_iter_var_next(i);
 522                 if (i == msg_opl->sg.end)
 523                         break;
 524                 sge = sk_msg_elem(msg_opl, i);
 525         }
 526
 527         msg_opl->sg.end = i;
 528         msg_opl->sg.curr = i;
 529         msg_opl->sg.copybreak = 0;
 530         msg_opl->apply_bytes = 0;
 531         msg_opl->sg.size = bytes;
 532
 533         msg_npl = &new->msg_plaintext;
 534         msg_npl->apply_bytes = apply;
 535         msg_npl->sg.size = orig_size - bytes;
 536
 537         j = msg_npl->sg.start;
 538         nsge = sk_msg_elem(msg_npl, j);
 539         if (tmp.length) {
 540                 memcpy(nsge, &tmp, sizeof(*nsge));
 541                 sk_msg_iter_var_next(j);
 542                 nsge = sk_msg_elem(msg_npl, j);
 543         }
 544
 545         osge = sk_msg_elem(msg_opl, i);
 546         while (osge->length) {
 547                 memcpy(nsge, osge, sizeof(*nsge));
 548                 sg_unmark_end(nsge);
 549                 sk_msg_iter_var_next(i);
 550                 sk_msg_iter_var_next(j);
 551                 if (i == *orig_end)
 552                         break;
 553                 osge = sk_msg_elem(msg_opl, i);
 554                 nsge = sk_msg_elem(msg_npl, j);
 555         }
 556
 557         msg_npl->sg.end = j;
 558         msg_npl->sg.curr = j;
 559         msg_npl->sg.copybreak = 0;
 560
 561         *to = new;
 562         return 0;
 563 }
 564
 565 static void tls_merge_open_record(struct sock *sk, struct tls_rec *to,
 566                                   struct tls_rec *from, u32 orig_end)
 567 {
 568         struct sk_msg *msg_npl = &from->msg_plaintext;
 569         struct sk_msg *msg_opl = &to->msg_plaintext;
 570         struct scatterlist *osge, *nsge;
 571         u32 i, j;
 572
 573         i = msg_opl->sg.end;
 574         sk_msg_iter_var_prev(i);
 575         j = msg_npl->sg.start;
 576
 577         osge = sk_msg_elem(msg_opl, i);
 578         nsge = sk_msg_elem(msg_npl, j);
 579
 580         if (sg_page(osge) == sg_page(nsge) &&
 581             osge->offset + osge->length == nsge->offset) {
 582                 osge->length += nsge->length;
 583                 put_page(sg_page(nsge));
 584         }
 585
 586         msg_opl->sg.end = orig_end;
 587         msg_opl->sg.curr = orig_end;
 588         msg_opl->sg.copybreak = 0;
 589         msg_opl->apply_bytes = msg_opl->sg.size + msg_npl->sg.size;
 590         msg_opl->sg.size += msg_npl->sg.size;
 591
 592         sk_msg_free(sk, &to->msg_encrypted);
 593         sk_msg_xfer_full(&to->msg_encrypted, &from->msg_encrypted);
 594
 595         kfree(from);
 596 }
 597
 598 static int tls_push_record(struct sock *sk, int flags,
 599                            unsigned char record_type)
 600 {
 601         struct tls_context *tls_ctx = tls_get_ctx(sk);
 602         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 603         struct tls_rec *rec = ctx->open_rec, *tmp = NULL;
 604         u32 i, split_point, uninitialized_var(orig_end);
 605         struct sk_msg *msg_pl, *msg_en;
 606         struct aead_request *req;
 607         bool split;
 608         int rc;
 609
 610         if (!rec)
 611                 return 0;
 612
 613         msg_pl = &rec->msg_plaintext;
 614         msg_en = &rec->msg_encrypted;
 615
 616         split_point = msg_pl->apply_bytes;
 617         split = split_point && split_point < msg_pl->sg.size;
 618         if (split) {
 619                 rc = tls_split_open_record(sk, rec, &tmp, msg_pl, msg_en,
 620                                            split_point, tls_ctx->tx.overhead_size,
 621                                            &orig_end);
 622                 if (rc < 0)
 623                         return rc;
 624                 sk_msg_trim(sk, msg_en, msg_pl->sg.size +
 625                             tls_ctx->tx.overhead_size);
 626         }
 627
 628         rec->tx_flags = flags;
 629         req = &rec->aead_req;
 630
 631         i = msg_pl->sg.end;
 632         sk_msg_iter_var_prev(i);
 633         sg_mark_end(sk_msg_elem(msg_pl, i));
 634
 635         i = msg_pl->sg.start;
 636         sg_chain(rec->sg_aead_in, 2, rec->inplace_crypto ?
 637                  &msg_en->sg.data[i] : &msg_pl->sg.data[i]);
 638
 639         i = msg_en->sg.end;
 640         sk_msg_iter_var_prev(i);
 641         sg_mark_end(sk_msg_elem(msg_en, i));
 642
 643         i = msg_en->sg.start;
 644         sg_chain(rec->sg_aead_out, 2, &msg_en->sg.data[i]);
 645
 646         tls_make_aad(rec->aad_space, msg_pl->sg.size,
 647                      tls_ctx->tx.rec_seq, tls_ctx->tx.rec_seq_size,
 648                      record_type);
 649
 650         tls_fill_prepend(tls_ctx,
 651                          page_address(sg_page(&msg_en->sg.data[i])) +
 652                          msg_en->sg.data[i].offset, msg_pl->sg.size,
 653                          record_type);
 654
 655         tls_ctx->pending_open_record_frags = false;
 656
 657         rc = tls_do_encryption(sk, tls_ctx, ctx, req, msg_pl->sg.size, i);
 658         if (rc < 0) {
 659                 if (rc != -EINPROGRESS) {
 660                         tls_err_abort(sk, EBADMSG);
 661                         if (split) {
 662                                 tls_ctx->pending_open_record_frags = true;
 663                                 tls_merge_open_record(sk, rec, tmp, orig_end);
 664                         }
 665                 }
 666                 return rc;
 667         } else if (split) {
 668                 msg_pl = &tmp->msg_plaintext;
 669                 msg_en = &tmp->msg_encrypted;
 670                 sk_msg_trim(sk, msg_en, msg_pl->sg.size +
 671                             tls_ctx->tx.overhead_size);
 672                 tls_ctx->pending_open_record_frags = true;
 673                 ctx->open_rec = tmp;
 674         }
 675
 676         return tls_tx_records(sk, flags);
 677 }
 678
 679 static int bpf_exec_tx_verdict(struct sk_msg *msg, struct sock *sk,
 680                                bool full_record, u8 record_type,
 681                                size_t *copied, int flags)
 682 {
 683         struct tls_context *tls_ctx = tls_get_ctx(sk);
 684         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 685         struct sk_msg msg_redir = { };
 686         struct sk_psock *psock;
 687         struct sock *sk_redir;
 688         struct tls_rec *rec;
 689         bool enospc, policy;
 690         int err = 0, send;
 691         u32 delta = 0;
 692
 693         policy = !(flags & MSG_SENDPAGE_NOPOLICY);
 694         psock = sk_psock_get(sk);
 695         if (!psock || !policy)
 696                 return tls_push_record(sk, flags, record_type);
 697 more_data:
 698         enospc = sk_msg_full(msg);
 699         if (psock->eval == __SK_NONE) {
 700                 delta = msg->sg.size;
 701                 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
 702                 if (delta < msg->sg.size)
 703                         delta -= msg->sg.size;
 704                 else
 705                         delta = 0;
 706         }
 707         if (msg->cork_bytes && msg->cork_bytes > msg->sg.size &&
 708             !enospc && !full_record) {
 709                 err = -ENOSPC;
 710                 goto out_err;
 711         }
 712         msg->cork_bytes = 0;
 713         send = msg->sg.size;
 714         if (msg->apply_bytes && msg->apply_bytes < send)
 715                 send = msg->apply_bytes;
 716
 717         switch (psock->eval) {
 718         case __SK_PASS:
 719                 err = tls_push_record(sk, flags, record_type);
 720                 if (err < 0) {
 721                         *copied -= sk_msg_free(sk, msg);
 722                         tls_free_open_rec(sk);
 723                         goto out_err;
 724                 }
 725                 break;
 726         case __SK_REDIRECT:
 727                 sk_redir = psock->sk_redir;
 728                 memcpy(&msg_redir, msg, sizeof(*msg));
 729                 if (msg->apply_bytes < send)
 730                         msg->apply_bytes = 0;
 731                 else
 732                         msg->apply_bytes -= send;
 733                 sk_msg_return_zero(sk, msg, send);
 734                 msg->sg.size -= send;
 735                 release_sock(sk);
 736                 err = tcp_bpf_sendmsg_redir(sk_redir, &msg_redir, send, flags);
 737                 lock_sock(sk);
 738                 if (err < 0) {
 739                         *copied -= sk_msg_free_nocharge(sk, &msg_redir);
 740                         msg->sg.size = 0;
 741                 }
 742                 if (msg->sg.size == 0)
 743                         tls_free_open_rec(sk);
 744                 break;
 745         case __SK_DROP:
 746         default:
 747                 sk_msg_free_partial(sk, msg, send);
 748                 if (msg->apply_bytes < send)
 749                         msg->apply_bytes = 0;
 750                 else
 751                         msg->apply_bytes -= send;
 752                 if (msg->sg.size == 0)
 753                         tls_free_open_rec(sk);
 754                 *copied -= (send + delta);
 755                 err = -EACCES;
 756         }
 757
 758         if (likely(!err)) {
 759                 bool reset_eval = !ctx->open_rec;
 760
 761                 rec = ctx->open_rec;
 762                 if (rec) {
 763                         msg = &rec->msg_plaintext;
 764                         if (!msg->apply_bytes)
 765                                 reset_eval = true;
 766                 }
 767                 if (reset_eval) {
 768                         psock->eval = __SK_NONE;
 769                         if (psock->sk_redir) {
 770                                 sock_put(psock->sk_redir);
 771                                 psock->sk_redir = NULL;
 772                         }
 773                 }
 774                 if (rec)
 775                         goto more_data;
 776         }
 777  out_err:
 778         sk_psock_put(sk, psock);
 779         return err;
 780 }
 781
 782 static int tls_sw_push_pending_record(struct sock *sk, int flags)
 783 {
 784         struct tls_context *tls_ctx = tls_get_ctx(sk);
 785         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 786         struct tls_rec *rec = ctx->open_rec;
 787         struct sk_msg *msg_pl;
 788         size_t copied;
 789
 790         if (!rec)
 791                 return 0;
 792
 793         msg_pl = &rec->msg_plaintext;
 794         copied = msg_pl->sg.size;
 795         if (!copied)
 796                 return 0;
 797
 798         return bpf_exec_tx_verdict(msg_pl, sk, true, TLS_RECORD_TYPE_DATA,
 799                                    &copied, flags);
 800 }
 801
 802 int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 803 {
 804         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 805         struct tls_context *tls_ctx = tls_get_ctx(sk);
 806         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 807         struct crypto_tfm *tfm = crypto_aead_tfm(ctx->aead_send);
 808         bool async_capable = tfm->__crt_alg->cra_flags & CRYPTO_ALG_ASYNC;
 809         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 810         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
 811         bool eor = !(msg->msg_flags & MSG_MORE);
 812         size_t try_to_copy, copied = 0;
 813         struct sk_msg *msg_pl, *msg_en;
 814         struct tls_rec *rec;
 815         int required_size;
 816         int num_async = 0;
 817         bool full_record;
 818         int record_room;
 819         int num_zc = 0;
 820         int orig_size;
 821         int ret = 0;
 822
 823         if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
 824                 return -ENOTSUPP;
 825
 826         lock_sock(sk);
 827
 828         /* Wait till there is any pending write on socket */
 829         if (unlikely(sk->sk_write_pending)) {
 830                 ret = wait_on_pending_writer(sk, &timeo);
 831                 if (unlikely(ret))
 832                         goto send_end;
 833         }
 834
 835         if (unlikely(msg->msg_controllen)) {
 836                 ret = tls_proccess_cmsg(sk, msg, &record_type);
 837                 if (ret) {
 838                         if (ret == -EINPROGRESS)
 839                                 num_async++;
 840                         else if (ret != -EAGAIN)
 841                                 goto send_end;
 842                 }
 843         }
 844
 845         while (msg_data_left(msg)) {
 846                 if (sk->sk_err) {
 847                         ret = -sk->sk_err;
 848                         goto send_end;
 849                 }
 850
 851                 if (ctx->open_rec)
 852                         rec = ctx->open_rec;
 853                 else
 854                         rec = ctx->open_rec = tls_get_rec(sk);
 855                 if (!rec) {
 856                         ret = -ENOMEM;
 857                         goto send_end;
 858                 }
 859
 860                 msg_pl = &rec->msg_plaintext;
 861                 msg_en = &rec->msg_encrypted;
 862
 863                 orig_size = msg_pl->sg.size;
 864                 full_record = false;
 865                 try_to_copy = msg_data_left(msg);
 866                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
 867                 if (try_to_copy >= record_room) {
 868                         try_to_copy = record_room;
 869                         full_record = true;
 870                 }
 871
 872                 required_size = msg_pl->sg.size + try_to_copy +
 873                                 tls_ctx->tx.overhead_size;
 874
 875                 if (!sk_stream_memory_free(sk))
 876                         goto wait_for_sndbuf;
 877
 878 alloc_encrypted:
 879                 ret = tls_alloc_encrypted_msg(sk, required_size);
 880                 if (ret) {
 881                         if (ret != -ENOSPC)
 882                                 goto wait_for_memory;
 883
 884                         /* Adjust try_to_copy according to the amount that was
 885                          * actually allocated. The difference is due
 886                          * to max sg elements limit
 887                          */
 888                         try_to_copy -= required_size - msg_en->sg.size;
 889                         full_record = true;
 890                 }
 891
 892                 if (!is_kvec && (full_record || eor) && !async_capable) {
 893                         u32 first = msg_pl->sg.end;
 894
 895                         ret = sk_msg_zerocopy_from_iter(sk, &msg->msg_iter,
 896                                                         msg_pl, try_to_copy);
 897                         if (ret)
 898                                 goto fallback_to_reg_send;
 899
 900                         rec->inplace_crypto = 0;
 901
 902                         num_zc++;
 903                         copied += try_to_copy;
 904
 905                         sk_msg_sg_copy_set(msg_pl, first);
 906                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
 907                                                   record_type, &copied,
 908                                                   msg->msg_flags);
 909                         if (ret) {
 910                                 if (ret == -EINPROGRESS)
 911                                         num_async++;
 912                                 else if (ret == -ENOMEM)
 913                                         goto wait_for_memory;
 914                                 else if (ret == -ENOSPC)
 915                                         goto rollback_iter;
 916                                 else if (ret != -EAGAIN)
 917                                         goto send_end;
 918                         }
 919                         continue;
 920 rollback_iter:
 921                         copied -= try_to_copy;
 922                         sk_msg_sg_copy_clear(msg_pl, first);
 923                         iov_iter_revert(&msg->msg_iter,
 924                                         msg_pl->sg.size - orig_size);
 925 fallback_to_reg_send:
 926                         sk_msg_trim(sk, msg_pl, orig_size);
 927                 }
 928
 929                 required_size = msg_pl->sg.size + try_to_copy;
 930
 931                 ret = tls_clone_plaintext_msg(sk, required_size);
 932                 if (ret) {
 933                         if (ret != -ENOSPC)
 934                                 goto send_end;
 935
 936                         /* Adjust try_to_copy according to the amount that was
 937                          * actually allocated. The difference is due
 938                          * to max sg elements limit
 939                          */
 940                         try_to_copy -= required_size - msg_pl->sg.size;
 941                         full_record = true;
 942                         sk_msg_trim(sk, msg_en, msg_pl->sg.size +
 943                                     tls_ctx->tx.overhead_size);
 944                 }
 945
 946                 if (try_to_copy) {
 947                         ret = sk_msg_memcopy_from_iter(sk, &msg->msg_iter,
 948                                                        msg_pl, try_to_copy);
 949                         if (ret < 0)
 950                                 goto trim_sgl;
 951                 }
 952
 953                 /* Open records defined only if successfully copied, otherwise
 954                  * we would trim the sg but not reset the open record frags.
 955                  */
 956                 tls_ctx->pending_open_record_frags = true;
 957                 copied += try_to_copy;
 958                 if (full_record || eor) {
 959                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
 960                                                   record_type, &copied,
 961                                                   msg->msg_flags);
 962                         if (ret) {
 963                                 if (ret == -EINPROGRESS)
 964                                         num_async++;
 965                                 else if (ret == -ENOMEM)
 966                                         goto wait_for_memory;
 967                                 else if (ret != -EAGAIN) {
 968                                         if (ret == -ENOSPC)
 969                                                 ret = 0;
 970                                         goto send_end;
 971                                 }
 972                         }
 973                 }
 974
 975                 continue;
 976
 977 wait_for_sndbuf:
 978                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 979 wait_for_memory:
 980                 ret = sk_stream_wait_memory(sk, &timeo);
 981                 if (ret) {
 982 trim_sgl:
 983                         tls_trim_both_msgs(sk, orig_size);
 984                         goto send_end;
 985                 }
 986
 987                 if (msg_en->sg.size < required_size)
 988                         goto alloc_encrypted;
 989         }
 990
 991         if (!num_async) {
 992                 goto send_end;
 993         } else if (num_zc) {
 994                 /* Wait for pending encryptions to get completed */
 995                 smp_store_mb(ctx->async_notify, true);
 996
 997                 if (atomic_read(&ctx->encrypt_pending))
 998                         crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
 999                 else
1000                         reinit_completion(&ctx->async_wait.completion);
1001
1002                 WRITE_ONCE(ctx->async_notify, false);
1003
1004                 if (ctx->async_wait.err) {
1005                         ret = ctx->async_wait.err;
1006                         copied = 0;
1007                 }
1008         }
1009
1010         /* Transmit if any encryptions have completed */
1011         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1012                 cancel_delayed_work(&ctx->tx_work.work);
1013                 tls_tx_records(sk, msg->msg_flags);
1014         }
1015
1016 send_end:
1017         ret = sk_stream_error(sk, msg->msg_flags, ret);
1018
1019         release_sock(sk);
1020         return copied ? copied : ret;
1021 }
1022
1023 int tls_sw_do_sendpage(struct sock *sk, struct page *page,
1024                        int offset, size_t size, int flags)
1025 {
1026         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1027         struct tls_context *tls_ctx = tls_get_ctx(sk);
1028         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1029         unsigned char record_type = TLS_RECORD_TYPE_DATA;
1030         struct sk_msg *msg_pl;
1031         struct tls_rec *rec;
1032         int num_async = 0;
1033         size_t copied = 0;
1034         bool full_record;
1035         int record_room;
1036         int ret = 0;
1037         bool eor;
1038
1039         eor = !(flags & (MSG_MORE | MSG_SENDPAGE_NOTLAST));
1040         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1041
1042         /* Wait till there is any pending write on socket */
1043         if (unlikely(sk->sk_write_pending)) {
1044                 ret = wait_on_pending_writer(sk, &timeo);
1045                 if (unlikely(ret))
1046                         goto sendpage_end;
1047         }
1048
1049         /* Call the sk_stream functions to manage the sndbuf mem. */
1050         while (size > 0) {
1051                 size_t copy, required_size;
1052
1053                 if (sk->sk_err) {
1054                         ret = -sk->sk_err;
1055                         goto sendpage_end;
1056                 }
1057
1058                 if (ctx->open_rec)
1059                         rec = ctx->open_rec;
1060                 else
1061                         rec = ctx->open_rec = tls_get_rec(sk);
1062                 if (!rec) {
1063                         ret = -ENOMEM;
1064                         goto sendpage_end;
1065                 }
1066
1067                 msg_pl = &rec->msg_plaintext;
1068
1069                 full_record = false;
1070                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
1071                 copied = 0;
1072                 copy = size;
1073                 if (copy >= record_room) {
1074                         copy = record_room;
1075                         full_record = true;
1076                 }
1077
1078                 required_size = msg_pl->sg.size + copy +
1079                                 tls_ctx->tx.overhead_size;
1080
1081                 if (!sk_stream_memory_free(sk))
1082                         goto wait_for_sndbuf;
1083 alloc_payload:
1084                 ret = tls_alloc_encrypted_msg(sk, required_size);
1085                 if (ret) {
1086                         if (ret != -ENOSPC)
1087                                 goto wait_for_memory;
1088
1089                         /* Adjust copy according to the amount that was
1090                          * actually allocated. The difference is due
1091                          * to max sg elements limit
1092                          */
1093                         copy -= required_size - msg_pl->sg.size;
1094                         full_record = true;
1095                 }
1096
1097                 sk_msg_page_add(msg_pl, page, copy, offset);
1098                 sk_mem_charge(sk, copy);
1099
1100                 offset += copy;
1101                 size -= copy;
1102                 copied += copy;
1103
1104                 tls_ctx->pending_open_record_frags = true;
1105                 if (full_record || eor || sk_msg_full(msg_pl)) {
1106                         rec->inplace_crypto = 0;
1107                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1108                                                   record_type, &copied, flags);
1109                         if (ret) {
1110                                 if (ret == -EINPROGRESS)
1111                                         num_async++;
1112                                 else if (ret == -ENOMEM)
1113                                         goto wait_for_memory;
1114                                 else if (ret != -EAGAIN) {
1115                                         if (ret == -ENOSPC)
1116                                                 ret = 0;
1117                                         goto sendpage_end;
1118                                 }
1119                         }
1120                 }
1121                 continue;
1122 wait_for_sndbuf:
1123                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1124 wait_for_memory:
1125                 ret = sk_stream_wait_memory(sk, &timeo);
1126                 if (ret) {
1127                         tls_trim_both_msgs(sk, msg_pl->sg.size);
1128                         goto sendpage_end;
1129                 }
1130
1131                 goto alloc_payload;
1132         }
1133
1134         if (num_async) {
1135                 /* Transmit if any encryptions have completed */
1136                 if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1137                         cancel_delayed_work(&ctx->tx_work.work);
1138                         tls_tx_records(sk, flags);
1139                 }
1140         }
1141 sendpage_end:
1142         ret = sk_stream_error(sk, flags, ret);
1143         return copied ? copied : ret;
1144 }
1145
1146 int tls_sw_sendpage_locked(struct sock *sk, struct page *page,
1147                            int offset, size_t size, int flags)
1148 {
1149         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1150                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY))
1151                 return -ENOTSUPP;
1152
1153         return tls_sw_do_sendpage(sk, page, offset, size, flags);
1154 }
1155
1156 int tls_sw_sendpage(struct sock *sk, struct page *page,
1157                     int offset, size_t size, int flags)
1158 {
1159         int ret;
1160
1161         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1162                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY))
1163                 return -ENOTSUPP;
1164
1165         lock_sock(sk);
1166         ret = tls_sw_do_sendpage(sk, page, offset, size, flags);
1167         release_sock(sk);
1168         return ret;
1169 }
1170
1171 static struct sk_buff *tls_wait_data(struct sock *sk, struct sk_psock *psock,
1172                                      int flags, long timeo, int *err)
1173 {
1174         struct tls_context *tls_ctx = tls_get_ctx(sk);
1175         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1176         struct sk_buff *skb;
1177         DEFINE_WAIT_FUNC(wait, woken_wake_function);
1178
1179         while (!(skb = ctx->recv_pkt) && sk_psock_queue_empty(psock)) {
1180                 if (sk->sk_err) {
1181                         *err = sock_error(sk);
1182                         return NULL;
1183                 }
1184
1185                 if (sk->sk_shutdown & RCV_SHUTDOWN)
1186                         return NULL;
1187
1188                 if (sock_flag(sk, SOCK_DONE))
1189                         return NULL;
1190
1191                 if ((flags & MSG_DONTWAIT) || !timeo) {
1192                         *err = -EAGAIN;
1193                         return NULL;
1194                 }
1195
1196                 add_wait_queue(sk_sleep(sk), &wait);
1197                 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1198                 sk_wait_event(sk, &timeo,
1199                               ctx->recv_pkt != skb ||
1200                               !sk_psock_queue_empty(psock),
1201                               &wait);
1202                 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1203                 remove_wait_queue(sk_sleep(sk), &wait);
1204
1205                 /* Handle signals */
1206                 if (signal_pending(current)) {
1207                         *err = sock_intr_errno(timeo);
1208                         return NULL;
1209                 }
1210         }
1211
1212         return skb;
1213 }
1214
1215 static int tls_setup_from_iter(struct sock *sk, struct iov_iter *from,
1216                                int length, int *pages_used,
1217                                unsigned int *size_used,
1218                                struct scatterlist *to,
1219                                int to_max_pages)
1220 {
1221         int rc = 0, i = 0, num_elem = *pages_used, maxpages;
1222         struct page *pages[MAX_SKB_FRAGS];
1223         unsigned int size = *size_used;
1224         ssize_t copied, use;
1225         size_t offset;
1226
1227         while (length > 0) {
1228                 i = 0;
1229                 maxpages = to_max_pages - num_elem;
1230                 if (maxpages == 0) {
1231                         rc = -EFAULT;
1232                         goto out;
1233                 }
1234                 copied = iov_iter_get_pages(from, pages,
1235                                             length,
1236                                             maxpages, &offset);
1237                 if (copied <= 0) {
1238                         rc = -EFAULT;
1239                         goto out;
1240                 }
1241
1242                 iov_iter_advance(from, copied);
1243
1244                 length -= copied;
1245                 size += copied;
1246                 while (copied) {
1247                         use = min_t(int, copied, PAGE_SIZE - offset);
1248
1249                         sg_set_page(&to[num_elem],
1250                                     pages[i], use, offset);
1251                         sg_unmark_end(&to[num_elem]);
1252                         /* We do not uncharge memory from this API */
1253
1254                         offset = 0;
1255                         copied -= use;
1256
1257                         i++;
1258                         num_elem++;
1259                 }
1260         }
1261         /* Mark the end in the last sg entry if newly added */
1262         if (num_elem > *pages_used)
1263                 sg_mark_end(&to[num_elem - 1]);
1264 out:
1265         if (rc)
1266                 iov_iter_revert(from, size - *size_used);
1267         *size_used = size;
1268         *pages_used = num_elem;
1269
1270         return rc;
1271 }
1272
1273 /* This function decrypts the input skb into either out_iov or in out_sg
1274  * or in skb buffers itself. The input parameter 'zc' indicates if
1275  * zero-copy mode needs to be tried or not. With zero-copy mode, either
1276  * out_iov or out_sg must be non-NULL. In case both out_iov and out_sg are
1277  * NULL, then the decryption happens inside skb buffers itself, i.e.
1278  * zero-copy gets disabled and 'zc' is updated.
1279  */
1280
1281 static int decrypt_internal(struct sock *sk, struct sk_buff *skb,
1282                             struct iov_iter *out_iov,
1283                             struct scatterlist *out_sg,
1284                             int *chunk, bool *zc)
1285 {
1286         struct tls_context *tls_ctx = tls_get_ctx(sk);
1287         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1288         struct strp_msg *rxm = strp_msg(skb);
1289         int n_sgin, n_sgout, nsg, mem_size, aead_size, err, pages = 0;
1290         struct aead_request *aead_req;
1291         struct sk_buff *unused;
1292         u8 *aad, *iv, *mem = NULL;
1293         struct scatterlist *sgin = NULL;
1294         struct scatterlist *sgout = NULL;
1295         const int data_len = rxm->full_len - tls_ctx->rx.overhead_size;
1296
1297         if (*zc && (out_iov || out_sg)) {
1298                 if (out_iov)
1299                         n_sgout = iov_iter_npages(out_iov, INT_MAX) + 1;
1300                 else
1301                         n_sgout = sg_nents(out_sg);
1302                 n_sgin = skb_nsg(skb, rxm->offset + tls_ctx->rx.prepend_size,
1303                                  rxm->full_len - tls_ctx->rx.prepend_size);
1304         } else {
1305                 n_sgout = 0;
1306                 *zc = false;
1307                 n_sgin = skb_cow_data(skb, 0, &unused);
1308         }
1309
1310         if (n_sgin < 1)
1311                 return -EBADMSG;
1312
1313         /* Increment to accommodate AAD */
1314         n_sgin = n_sgin + 1;
1315
1316         nsg = n_sgin + n_sgout;
1317
1318         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(ctx->aead_recv);
1319         mem_size = aead_size + (nsg * sizeof(struct scatterlist));
1320         mem_size = mem_size + TLS_AAD_SPACE_SIZE;
1321         mem_size = mem_size + crypto_aead_ivsize(ctx->aead_recv);
1322
1323         /* Allocate a single block of memory which contains
1324          * aead_req || sgin[] || sgout[] || aad || iv.
1325          * This order achieves correct alignment for aead_req, sgin, sgout.
1326          */
1327         mem = kmalloc(mem_size, sk->sk_allocation);
1328         if (!mem)
1329                 return -ENOMEM;
1330
1331         /* Segment the allocated memory */
1332         aead_req = (struct aead_request *)mem;
1333         sgin = (struct scatterlist *)(mem + aead_size);
1334         sgout = sgin + n_sgin;
1335         aad = (u8 *)(sgout + n_sgout);
1336         iv = aad + TLS_AAD_SPACE_SIZE;
1337
1338         /* Prepare IV */
1339         err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
1340                             iv + TLS_CIPHER_AES_GCM_128_SALT_SIZE,
1341                             tls_ctx->rx.iv_size);
1342         if (err < 0) {
1343                 kfree(mem);
1344                 return err;
1345         }
1346         memcpy(iv, tls_ctx->rx.iv, TLS_CIPHER_AES_GCM_128_SALT_SIZE);
1347
1348         /* Prepare AAD */
1349         tls_make_aad(aad, rxm->full_len - tls_ctx->rx.overhead_size,
1350                      tls_ctx->rx.rec_seq, tls_ctx->rx.rec_seq_size,
1351                      ctx->control);
1352
1353         /* Prepare sgin */
1354         sg_init_table(sgin, n_sgin);
1355         sg_set_buf(&sgin[0], aad, TLS_AAD_SPACE_SIZE);
1356         err = skb_to_sgvec(skb, &sgin[1],
1357                            rxm->offset + tls_ctx->rx.prepend_size,
1358                            rxm->full_len - tls_ctx->rx.prepend_size);
1359         if (err < 0) {
1360                 kfree(mem);
1361                 return err;
1362         }
1363
1364         if (n_sgout) {
1365                 if (out_iov) {
1366                         sg_init_table(sgout, n_sgout);
1367                         sg_set_buf(&sgout[0], aad, TLS_AAD_SPACE_SIZE);
1368
1369                         *chunk = 0;
1370                         err = tls_setup_from_iter(sk, out_iov, data_len,
1371                                                   &pages, chunk, &sgout[1],
1372                                                   (n_sgout - 1));
1373                         if (err < 0)
1374                                 goto fallback_to_reg_recv;
1375                 } else if (out_sg) {
1376                         memcpy(sgout, out_sg, n_sgout * sizeof(*sgout));
1377                 } else {
1378                         goto fallback_to_reg_recv;
1379                 }
1380         } else {
1381 fallback_to_reg_recv:
1382                 sgout = sgin;
1383                 pages = 0;
1384                 *chunk = 0;
1385                 *zc = false;
1386         }
1387
1388         /* Prepare and submit AEAD request */
1389         err = tls_do_decryption(sk, skb, sgin, sgout, iv,
1390                                 data_len, aead_req, *zc);
1391         if (err == -EINPROGRESS)
1392                 return err;
1393
1394         /* Release the pages in case iov was mapped to pages */
1395         for (; pages > 0; pages--)
1396                 put_page(sg_page(&sgout[pages]));
1397
1398         kfree(mem);
1399         return err;
1400 }
1401
1402 static int decrypt_skb_update(struct sock *sk, struct sk_buff *skb,
1403                               struct iov_iter *dest, int *chunk, bool *zc)
1404 {
1405         struct tls_context *tls_ctx = tls_get_ctx(sk);
1406         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1407         struct strp_msg *rxm = strp_msg(skb);
1408         int err = 0;
1409
1410 #ifdef CONFIG_TLS_DEVICE
1411         err = tls_device_decrypted(sk, skb);
1412         if (err < 0)
1413                 return err;
1414 #endif
1415         if (!ctx->decrypted) {
1416                 err = decrypt_internal(sk, skb, dest, NULL, chunk, zc);
1417                 if (err < 0) {
1418                         if (err == -EINPROGRESS)
1419                                 tls_advance_record_sn(sk, &tls_ctx->rx);
1420
1421                         return err;
1422                 }
1423         } else {
1424                 *zc = false;
1425         }
1426
1427         rxm->offset += tls_ctx->rx.prepend_size;
1428         rxm->full_len -= tls_ctx->rx.overhead_size;
1429         tls_advance_record_sn(sk, &tls_ctx->rx);
1430         ctx->decrypted = true;
1431         ctx->saved_data_ready(sk);
1432
1433         return err;
1434 }
1435
1436 int decrypt_skb(struct sock *sk, struct sk_buff *skb,
1437                 struct scatterlist *sgout)
1438 {
1439         bool zc = true;
1440         int chunk;
1441
1442         return decrypt_internal(sk, skb, NULL, sgout, &chunk, &zc);
1443 }
1444
1445 static bool tls_sw_advance_skb(struct sock *sk, struct sk_buff *skb,
1446                                unsigned int len)
1447 {
1448         struct tls_context *tls_ctx = tls_get_ctx(sk);
1449         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1450
1451         if (skb) {
1452                 struct strp_msg *rxm = strp_msg(skb);
1453
1454                 if (len < rxm->full_len) {
1455                         rxm->offset += len;
1456                         rxm->full_len -= len;
1457                         return false;
1458                 }
1459                 kfree_skb(skb);
1460         }
1461
1462         /* Finished with message */
1463         ctx->recv_pkt = NULL;
1464         __strp_unpause(&ctx->strp);
1465
1466         return true;
1467 }
1468
1469 int tls_sw_recvmsg(struct sock *sk,
1470                    struct msghdr *msg,
1471                    size_t len,
1472                    int nonblock,
1473                    int flags,
1474                    int *addr_len)
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 sk_psock *psock;
1479         unsigned char control;
1480         struct strp_msg *rxm;
1481         struct sk_buff *skb;
1482         ssize_t copied = 0;
1483         bool cmsg = false;
1484         int target, err = 0;
1485         long timeo;
1486         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
1487         int num_async = 0;
1488
1489         flags |= nonblock;
1490
1491         if (unlikely(flags & MSG_ERRQUEUE))
1492                 return sock_recv_errqueue(sk, msg, len, SOL_IP, IP_RECVERR);
1493
1494         psock = sk_psock_get(sk);
1495         lock_sock(sk);
1496
1497         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1498         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1499         do {
1500                 bool zc = false;
1501                 bool async = false;
1502                 int chunk = 0;
1503
1504                 skb = tls_wait_data(sk, psock, flags, timeo, &err);
1505                 if (!skb) {
1506                         if (psock) {
1507                                 int ret = __tcp_bpf_recvmsg(sk, psock,
1508                                                             msg, len, flags);
1509
1510                                 if (ret > 0) {
1511                                         copied += ret;
1512                                         len -= ret;
1513                                         continue;
1514                                 }
1515                         }
1516                         goto recv_end;
1517                 }
1518
1519                 rxm = strp_msg(skb);
1520
1521                 if (!cmsg) {
1522                         int cerr;
1523
1524                         cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1525                                         sizeof(ctx->control), &ctx->control);
1526                         cmsg = true;
1527                         control = ctx->control;
1528                         if (ctx->control != TLS_RECORD_TYPE_DATA) {
1529                                 if (cerr || msg->msg_flags & MSG_CTRUNC) {
1530                                         err = -EIO;
1531                                         goto recv_end;
1532                                 }
1533                         }
1534                 } else if (control != ctx->control) {
1535                         goto recv_end;
1536                 }
1537
1538                 if (!ctx->decrypted) {
1539                         int to_copy = rxm->full_len - tls_ctx->rx.overhead_size;
1540
1541                         if (!is_kvec && to_copy <= len &&
1542                             likely(!(flags & MSG_PEEK)))
1543                                 zc = true;
1544
1545                         err = decrypt_skb_update(sk, skb, &msg->msg_iter,
1546                                                  &chunk, &zc);
1547                         if (err < 0 && err != -EINPROGRESS) {
1548                                 tls_err_abort(sk, EBADMSG);
1549                                 goto recv_end;
1550                         }
1551
1552                         if (err == -EINPROGRESS) {
1553                                 async = true;
1554                                 num_async++;
1555                                 goto pick_next_record;
1556                         }
1557
1558                         ctx->decrypted = true;
1559                 }
1560
1561                 if (!zc) {
1562                         chunk = min_t(unsigned int, rxm->full_len, len);
1563
1564                         err = skb_copy_datagram_msg(skb, rxm->offset, msg,
1565                                                     chunk);
1566                         if (err < 0)
1567                                 goto recv_end;
1568                 }
1569
1570 pick_next_record:
1571                 copied += chunk;
1572                 len -= chunk;
1573                 if (likely(!(flags & MSG_PEEK))) {
1574                         u8 control = ctx->control;
1575
1576                         /* For async, drop current skb reference */
1577                         if (async)
1578                                 skb = NULL;
1579
1580                         if (tls_sw_advance_skb(sk, skb, chunk)) {
1581                                 /* Return full control message to
1582                                  * userspace before trying to parse
1583                                  * another message type
1584                                  */
1585                                 msg->msg_flags |= MSG_EOR;
1586                                 if (control != TLS_RECORD_TYPE_DATA)
1587                                         goto recv_end;
1588                         } else {
1589                                 break;
1590                         }
1591                 } else {
1592                         /* MSG_PEEK right now cannot look beyond current skb
1593                          * from strparser, meaning we cannot advance skb here
1594                          * and thus unpause strparser since we'd loose original
1595                          * one.
1596                          */
1597                         break;
1598                 }
1599
1600                 /* If we have a new message from strparser, continue now. */
1601                 if (copied >= target && !ctx->recv_pkt)
1602                         break;
1603         } while (len);
1604
1605 recv_end:
1606         if (num_async) {
1607                 /* Wait for all previously submitted records to be decrypted */
1608                 smp_store_mb(ctx->async_notify, true);
1609                 if (atomic_read(&ctx->decrypt_pending)) {
1610                         err = crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1611                         if (err) {
1612                                 /* one of async decrypt failed */
1613                                 tls_err_abort(sk, err);
1614                                 copied = 0;
1615                         }
1616                 } else {
1617                         reinit_completion(&ctx->async_wait.completion);
1618                 }
1619                 WRITE_ONCE(ctx->async_notify, false);
1620         }
1621
1622         release_sock(sk);
1623         if (psock)
1624                 sk_psock_put(sk, psock);
1625         return copied ? : err;
1626 }
1627
1628 ssize_t tls_sw_splice_read(struct socket *sock,  loff_t *ppos,
1629                            struct pipe_inode_info *pipe,
1630                            size_t len, unsigned int flags)
1631 {
1632         struct tls_context *tls_ctx = tls_get_ctx(sock->sk);
1633         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1634         struct strp_msg *rxm = NULL;
1635         struct sock *sk = sock->sk;
1636         struct sk_buff *skb;
1637         ssize_t copied = 0;
1638         int err = 0;
1639         long timeo;
1640         int chunk;
1641         bool zc = false;
1642
1643         lock_sock(sk);
1644
1645         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1646
1647         skb = tls_wait_data(sk, NULL, flags, timeo, &err);
1648         if (!skb)
1649                 goto splice_read_end;
1650
1651         /* splice does not support reading control messages */
1652         if (ctx->control != TLS_RECORD_TYPE_DATA) {
1653                 err = -ENOTSUPP;
1654                 goto splice_read_end;
1655         }
1656
1657         if (!ctx->decrypted) {
1658                 err = decrypt_skb_update(sk, skb, NULL, &chunk, &zc);
1659
1660                 if (err < 0) {
1661                         tls_err_abort(sk, EBADMSG);
1662                         goto splice_read_end;
1663                 }
1664                 ctx->decrypted = true;
1665         }
1666         rxm = strp_msg(skb);
1667
1668         chunk = min_t(unsigned int, rxm->full_len, len);
1669         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, chunk, flags);
1670         if (copied < 0)
1671                 goto splice_read_end;
1672
1673         if (likely(!(flags & MSG_PEEK)))
1674                 tls_sw_advance_skb(sk, skb, copied);
1675
1676 splice_read_end:
1677         release_sock(sk);
1678         return copied ? : err;
1679 }
1680
1681 bool tls_sw_stream_read(const struct sock *sk)
1682 {
1683         struct tls_context *tls_ctx = tls_get_ctx(sk);
1684         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1685         bool ingress_empty = true;
1686         struct sk_psock *psock;
1687
1688         rcu_read_lock();
1689         psock = sk_psock(sk);
1690         if (psock)
1691                 ingress_empty = list_empty(&psock->ingress_msg);
1692         rcu_read_unlock();
1693
1694         return !ingress_empty || ctx->recv_pkt;
1695 }
1696
1697 static int tls_read_size(struct strparser *strp, struct sk_buff *skb)
1698 {
1699         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
1700         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1701         char header[TLS_HEADER_SIZE + MAX_IV_SIZE];
1702         struct strp_msg *rxm = strp_msg(skb);
1703         size_t cipher_overhead;
1704         size_t data_len = 0;
1705         int ret;
1706
1707         /* Verify that we have a full TLS header, or wait for more data */
1708         if (rxm->offset + tls_ctx->rx.prepend_size > skb->len)
1709                 return 0;
1710
1711         /* Sanity-check size of on-stack buffer. */
1712         if (WARN_ON(tls_ctx->rx.prepend_size > sizeof(header))) {
1713                 ret = -EINVAL;
1714                 goto read_failure;
1715         }
1716
1717         /* Linearize header to local buffer */
1718         ret = skb_copy_bits(skb, rxm->offset, header, tls_ctx->rx.prepend_size);
1719
1720         if (ret < 0)
1721                 goto read_failure;
1722
1723         ctx->control = header[0];
1724
1725         data_len = ((header[4] & 0xFF) | (header[3] << 8));
1726
1727         cipher_overhead = tls_ctx->rx.tag_size + tls_ctx->rx.iv_size;
1728
1729         if (data_len > TLS_MAX_PAYLOAD_SIZE + cipher_overhead) {
1730                 ret = -EMSGSIZE;
1731                 goto read_failure;
1732         }
1733         if (data_len < cipher_overhead) {
1734                 ret = -EBADMSG;
1735                 goto read_failure;
1736         }
1737
1738         if (header[1] != TLS_VERSION_MINOR(tls_ctx->crypto_recv.info.version) ||
1739             header[2] != TLS_VERSION_MAJOR(tls_ctx->crypto_recv.info.version)) {
1740                 ret = -EINVAL;
1741                 goto read_failure;
1742         }
1743
1744 #ifdef CONFIG_TLS_DEVICE
1745         handle_device_resync(strp->sk, TCP_SKB_CB(skb)->seq + rxm->offset,
1746                              *(u64*)tls_ctx->rx.rec_seq);
1747 #endif
1748         return data_len + TLS_HEADER_SIZE;
1749
1750 read_failure:
1751         tls_err_abort(strp->sk, ret);
1752
1753         return ret;
1754 }
1755
1756 static void tls_queue(struct strparser *strp, struct sk_buff *skb)
1757 {
1758         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
1759         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1760
1761         ctx->decrypted = false;
1762
1763         ctx->recv_pkt = skb;
1764         strp_pause(strp);
1765
1766         ctx->saved_data_ready(strp->sk);
1767 }
1768
1769 static void tls_data_ready(struct sock *sk)
1770 {
1771         struct tls_context *tls_ctx = tls_get_ctx(sk);
1772         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1773         struct sk_psock *psock;
1774
1775         strp_data_ready(&ctx->strp);
1776
1777         psock = sk_psock_get(sk);
1778         if (psock && !list_empty(&psock->ingress_msg)) {
1779                 ctx->saved_data_ready(sk);
1780                 sk_psock_put(sk, psock);
1781         }
1782 }
1783
1784 void tls_sw_free_resources_tx(struct sock *sk)
1785 {
1786         struct tls_context *tls_ctx = tls_get_ctx(sk);
1787         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1788         struct tls_rec *rec, *tmp;
1789
1790         /* Wait for any pending async encryptions to complete */
1791         smp_store_mb(ctx->async_notify, true);
1792         if (atomic_read(&ctx->encrypt_pending))
1793                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1794
1795         cancel_delayed_work_sync(&ctx->tx_work.work);
1796
1797         /* Tx whatever records we can transmit and abandon the rest */
1798         tls_tx_records(sk, -1);
1799
1800         /* Free up un-sent records in tx_list. First, free
1801          * the partially sent record if any at head of tx_list.
1802          */
1803         if (tls_ctx->partially_sent_record) {
1804                 struct scatterlist *sg = tls_ctx->partially_sent_record;
1805
1806                 while (1) {
1807                         put_page(sg_page(sg));
1808                         sk_mem_uncharge(sk, sg->length);
1809
1810                         if (sg_is_last(sg))
1811                                 break;
1812                         sg++;
1813                 }
1814
1815                 tls_ctx->partially_sent_record = NULL;
1816
1817                 rec = list_first_entry(&ctx->tx_list,
1818                                        struct tls_rec, list);
1819                 list_del(&rec->list);
1820                 sk_msg_free(sk, &rec->msg_plaintext);
1821                 kfree(rec);
1822         }
1823
1824         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
1825                 list_del(&rec->list);
1826                 sk_msg_free(sk, &rec->msg_encrypted);
1827                 sk_msg_free(sk, &rec->msg_plaintext);
1828                 kfree(rec);
1829         }
1830
1831         crypto_free_aead(ctx->aead_send);
1832         tls_free_open_rec(sk);
1833
1834         kfree(ctx);
1835 }
1836
1837 void tls_sw_release_resources_rx(struct sock *sk)
1838 {
1839         struct tls_context *tls_ctx = tls_get_ctx(sk);
1840         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1841
1842         if (ctx->aead_recv) {
1843                 kfree_skb(ctx->recv_pkt);
1844                 ctx->recv_pkt = NULL;
1845                 crypto_free_aead(ctx->aead_recv);
1846                 strp_stop(&ctx->strp);
1847                 write_lock_bh(&sk->sk_callback_lock);
1848                 sk->sk_data_ready = ctx->saved_data_ready;
1849                 write_unlock_bh(&sk->sk_callback_lock);
1850                 release_sock(sk);
1851                 strp_done(&ctx->strp);
1852                 lock_sock(sk);
1853         }
1854 }
1855
1856 void tls_sw_free_resources_rx(struct sock *sk)
1857 {
1858         struct tls_context *tls_ctx = tls_get_ctx(sk);
1859         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1860
1861         tls_sw_release_resources_rx(sk);
1862
1863         kfree(ctx);
1864 }
1865
1866 /* The work handler to transmitt the encrypted records in tx_list */
1867 static void tx_work_handler(struct work_struct *work)
1868 {
1869         struct delayed_work *delayed_work = to_delayed_work(work);
1870         struct tx_work *tx_work = container_of(delayed_work,
1871                                                struct tx_work, work);
1872         struct sock *sk = tx_work->sk;
1873         struct tls_context *tls_ctx = tls_get_ctx(sk);
1874         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1875
1876         if (!test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
1877                 return;
1878
1879         lock_sock(sk);
1880         tls_tx_records(sk, -1);
1881         release_sock(sk);
1882 }
1883
1884 int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
1885 {
1886         struct tls_crypto_info *crypto_info;
1887         struct tls12_crypto_info_aes_gcm_128 *gcm_128_info;
1888         struct tls_sw_context_tx *sw_ctx_tx = NULL;
1889         struct tls_sw_context_rx *sw_ctx_rx = NULL;
1890         struct cipher_context *cctx;
1891         struct crypto_aead **aead;
1892         struct strp_callbacks cb;
1893         u16 nonce_size, tag_size, iv_size, rec_seq_size;
1894         char *iv, *rec_seq;
1895         int rc = 0;
1896
1897         if (!ctx) {
1898                 rc = -EINVAL;
1899                 goto out;
1900         }
1901
1902         if (tx) {
1903                 if (!ctx->priv_ctx_tx) {
1904                         sw_ctx_tx = kzalloc(sizeof(*sw_ctx_tx), GFP_KERNEL);
1905                         if (!sw_ctx_tx) {
1906                                 rc = -ENOMEM;
1907                                 goto out;
1908                         }
1909                         ctx->priv_ctx_tx = sw_ctx_tx;
1910                 } else {
1911                         sw_ctx_tx =
1912                                 (struct tls_sw_context_tx *)ctx->priv_ctx_tx;
1913                 }
1914         } else {
1915                 if (!ctx->priv_ctx_rx) {
1916                         sw_ctx_rx = kzalloc(sizeof(*sw_ctx_rx), GFP_KERNEL);
1917                         if (!sw_ctx_rx) {
1918                                 rc = -ENOMEM;
1919                                 goto out;
1920                         }
1921                         ctx->priv_ctx_rx = sw_ctx_rx;
1922                 } else {
1923                         sw_ctx_rx =
1924                                 (struct tls_sw_context_rx *)ctx->priv_ctx_rx;
1925                 }
1926         }
1927
1928         if (tx) {
1929                 crypto_init_wait(&sw_ctx_tx->async_wait);
1930                 crypto_info = &ctx->crypto_send.info;
1931                 cctx = &ctx->tx;
1932                 aead = &sw_ctx_tx->aead_send;
1933                 INIT_LIST_HEAD(&sw_ctx_tx->tx_list);
1934                 INIT_DELAYED_WORK(&sw_ctx_tx->tx_work.work, tx_work_handler);
1935                 sw_ctx_tx->tx_work.sk = sk;
1936         } else {
1937                 crypto_init_wait(&sw_ctx_rx->async_wait);
1938                 crypto_info = &ctx->crypto_recv.info;
1939                 cctx = &ctx->rx;
1940                 aead = &sw_ctx_rx->aead_recv;
1941         }
1942
1943         switch (crypto_info->cipher_type) {
1944         case TLS_CIPHER_AES_GCM_128: {
1945                 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
1946                 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
1947                 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
1948                 iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
1949                 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
1950                 rec_seq =
1951                  ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
1952                 gcm_128_info =
1953                         (struct tls12_crypto_info_aes_gcm_128 *)crypto_info;
1954                 break;
1955         }
1956         default:
1957                 rc = -EINVAL;
1958                 goto free_priv;
1959         }
1960
1961         /* Sanity-check the IV size for stack allocations. */
1962         if (iv_size > MAX_IV_SIZE || nonce_size > MAX_IV_SIZE) {
1963                 rc = -EINVAL;
1964                 goto free_priv;
1965         }
1966
1967         cctx->prepend_size = TLS_HEADER_SIZE + nonce_size;
1968         cctx->tag_size = tag_size;
1969         cctx->overhead_size = cctx->prepend_size + cctx->tag_size;
1970         cctx->iv_size = iv_size;
1971         cctx->iv = kmalloc(iv_size + TLS_CIPHER_AES_GCM_128_SALT_SIZE,
1972                            GFP_KERNEL);
1973         if (!cctx->iv) {
1974                 rc = -ENOMEM;
1975                 goto free_priv;
1976         }
1977         memcpy(cctx->iv, gcm_128_info->salt, TLS_CIPHER_AES_GCM_128_SALT_SIZE);
1978         memcpy(cctx->iv + TLS_CIPHER_AES_GCM_128_SALT_SIZE, iv, iv_size);
1979         cctx->rec_seq_size = rec_seq_size;
1980         cctx->rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
1981         if (!cctx->rec_seq) {
1982                 rc = -ENOMEM;
1983                 goto free_iv;
1984         }
1985
1986         if (!*aead) {
1987                 *aead = crypto_alloc_aead("gcm(aes)", 0, 0);
1988                 if (IS_ERR(*aead)) {
1989                         rc = PTR_ERR(*aead);
1990                         *aead = NULL;
1991                         goto free_rec_seq;
1992                 }
1993         }
1994
1995         ctx->push_pending_record = tls_sw_push_pending_record;
1996
1997         rc = crypto_aead_setkey(*aead, gcm_128_info->key,
1998                                 TLS_CIPHER_AES_GCM_128_KEY_SIZE);
1999         if (rc)
2000                 goto free_aead;
2001
2002         rc = crypto_aead_setauthsize(*aead, cctx->tag_size);
2003         if (rc)
2004                 goto free_aead;
2005
2006         if (sw_ctx_rx) {
2007                 /* Set up strparser */
2008                 memset(&cb, 0, sizeof(cb));
2009                 cb.rcv_msg = tls_queue;
2010                 cb.parse_msg = tls_read_size;
2011
2012                 strp_init(&sw_ctx_rx->strp, sk, &cb);
2013
2014                 write_lock_bh(&sk->sk_callback_lock);
2015                 sw_ctx_rx->saved_data_ready = sk->sk_data_ready;
2016                 sk->sk_data_ready = tls_data_ready;
2017                 write_unlock_bh(&sk->sk_callback_lock);
2018
2019                 strp_check_rcv(&sw_ctx_rx->strp);
2020         }
2021
2022         goto out;
2023
2024 free_aead:
2025         crypto_free_aead(*aead);
2026         *aead = NULL;
2027 free_rec_seq:
2028         kfree(cctx->rec_seq);
2029         cctx->rec_seq = NULL;
2030 free_iv:
2031         kfree(cctx->iv);
2032         cctx->iv = NULL;
2033 free_priv:
2034         if (tx) {
2035                 kfree(ctx->priv_ctx_tx);
2036                 ctx->priv_ctx_tx = NULL;
2037         } else {
2038                 kfree(ctx->priv_ctx_rx);
2039                 ctx->priv_ctx_rx = NULL;
2040         }
2041 out:
2042         return rc;
2043 }