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                 ret = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_pl,
 947                                                try_to_copy);
 948                 if (ret < 0)
 949                         goto trim_sgl;
 950
 951                 /* Open records defined only if successfully copied, otherwise
 952                  * we would trim the sg but not reset the open record frags.
 953                  */
 954                 tls_ctx->pending_open_record_frags = true;
 955                 copied += try_to_copy;
 956                 if (full_record || eor) {
 957                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
 958                                                   record_type, &copied,
 959                                                   msg->msg_flags);
 960                         if (ret) {
 961                                 if (ret == -EINPROGRESS)
 962                                         num_async++;
 963                                 else if (ret == -ENOMEM)
 964                                         goto wait_for_memory;
 965                                 else if (ret != -EAGAIN) {
 966                                         if (ret == -ENOSPC)
 967                                                 ret = 0;
 968                                         goto send_end;
 969                                 }
 970                         }
 971                 }
 972
 973                 continue;
 974
 975 wait_for_sndbuf:
 976                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 977 wait_for_memory:
 978                 ret = sk_stream_wait_memory(sk, &timeo);
 979                 if (ret) {
 980 trim_sgl:
 981                         tls_trim_both_msgs(sk, orig_size);
 982                         goto send_end;
 983                 }
 984
 985                 if (msg_en->sg.size < required_size)
 986                         goto alloc_encrypted;
 987         }
 988
 989         if (!num_async) {
 990                 goto send_end;
 991         } else if (num_zc) {
 992                 /* Wait for pending encryptions to get completed */
 993                 smp_store_mb(ctx->async_notify, true);
 994
 995                 if (atomic_read(&ctx->encrypt_pending))
 996                         crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
 997                 else
 998                         reinit_completion(&ctx->async_wait.completion);
 999
1000                 WRITE_ONCE(ctx->async_notify, false);
1001
1002                 if (ctx->async_wait.err) {
1003                         ret = ctx->async_wait.err;
1004                         copied = 0;
1005                 }
1006         }
1007
1008         /* Transmit if any encryptions have completed */
1009         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1010                 cancel_delayed_work(&ctx->tx_work.work);
1011                 tls_tx_records(sk, msg->msg_flags);
1012         }
1013
1014 send_end:
1015         ret = sk_stream_error(sk, msg->msg_flags, ret);
1016
1017         release_sock(sk);
1018         return copied ? copied : ret;
1019 }
1020
1021 int tls_sw_do_sendpage(struct sock *sk, struct page *page,
1022                        int offset, size_t size, int flags)
1023 {
1024         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1025         struct tls_context *tls_ctx = tls_get_ctx(sk);
1026         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1027         unsigned char record_type = TLS_RECORD_TYPE_DATA;
1028         struct sk_msg *msg_pl;
1029         struct tls_rec *rec;
1030         int num_async = 0;
1031         size_t copied = 0;
1032         bool full_record;
1033         int record_room;
1034         int ret = 0;
1035         bool eor;
1036
1037         eor = !(flags & (MSG_MORE | MSG_SENDPAGE_NOTLAST));
1038         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1039
1040         /* Wait till there is any pending write on socket */
1041         if (unlikely(sk->sk_write_pending)) {
1042                 ret = wait_on_pending_writer(sk, &timeo);
1043                 if (unlikely(ret))
1044                         goto sendpage_end;
1045         }
1046
1047         /* Call the sk_stream functions to manage the sndbuf mem. */
1048         while (size > 0) {
1049                 size_t copy, required_size;
1050
1051                 if (sk->sk_err) {
1052                         ret = -sk->sk_err;
1053                         goto sendpage_end;
1054                 }
1055
1056                 if (ctx->open_rec)
1057                         rec = ctx->open_rec;
1058                 else
1059                         rec = ctx->open_rec = tls_get_rec(sk);
1060                 if (!rec) {
1061                         ret = -ENOMEM;
1062                         goto sendpage_end;
1063                 }
1064
1065                 msg_pl = &rec->msg_plaintext;
1066
1067                 full_record = false;
1068                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
1069                 copied = 0;
1070                 copy = size;
1071                 if (copy >= record_room) {
1072                         copy = record_room;
1073                         full_record = true;
1074                 }
1075
1076                 required_size = msg_pl->sg.size + copy +
1077                                 tls_ctx->tx.overhead_size;
1078
1079                 if (!sk_stream_memory_free(sk))
1080                         goto wait_for_sndbuf;
1081 alloc_payload:
1082                 ret = tls_alloc_encrypted_msg(sk, required_size);
1083                 if (ret) {
1084                         if (ret != -ENOSPC)
1085                                 goto wait_for_memory;
1086
1087                         /* Adjust copy according to the amount that was
1088                          * actually allocated. The difference is due
1089                          * to max sg elements limit
1090                          */
1091                         copy -= required_size - msg_pl->sg.size;
1092                         full_record = true;
1093                 }
1094
1095                 sk_msg_page_add(msg_pl, page, copy, offset);
1096                 sk_mem_charge(sk, copy);
1097
1098                 offset += copy;
1099                 size -= copy;
1100                 copied += copy;
1101
1102                 tls_ctx->pending_open_record_frags = true;
1103                 if (full_record || eor || sk_msg_full(msg_pl)) {
1104                         rec->inplace_crypto = 0;
1105                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1106                                                   record_type, &copied, flags);
1107                         if (ret) {
1108                                 if (ret == -EINPROGRESS)
1109                                         num_async++;
1110                                 else if (ret == -ENOMEM)
1111                                         goto wait_for_memory;
1112                                 else if (ret != -EAGAIN) {
1113                                         if (ret == -ENOSPC)
1114                                                 ret = 0;
1115                                         goto sendpage_end;
1116                                 }
1117                         }
1118                 }
1119                 continue;
1120 wait_for_sndbuf:
1121                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1122 wait_for_memory:
1123                 ret = sk_stream_wait_memory(sk, &timeo);
1124                 if (ret) {
1125                         tls_trim_both_msgs(sk, msg_pl->sg.size);
1126                         goto sendpage_end;
1127                 }
1128
1129                 goto alloc_payload;
1130         }
1131
1132         if (num_async) {
1133                 /* Transmit if any encryptions have completed */
1134                 if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1135                         cancel_delayed_work(&ctx->tx_work.work);
1136                         tls_tx_records(sk, flags);
1137                 }
1138         }
1139 sendpage_end:
1140         ret = sk_stream_error(sk, flags, ret);
1141         return copied ? copied : ret;
1142 }
1143
1144 int tls_sw_sendpage_locked(struct sock *sk, struct page *page,
1145                            int offset, size_t size, int flags)
1146 {
1147         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1148                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY))
1149                 return -ENOTSUPP;
1150
1151         return tls_sw_do_sendpage(sk, page, offset, size, flags);
1152 }
1153
1154 int tls_sw_sendpage(struct sock *sk, struct page *page,
1155                     int offset, size_t size, int flags)
1156 {
1157         int ret;
1158
1159         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1160                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY))
1161                 return -ENOTSUPP;
1162
1163         lock_sock(sk);
1164         ret = tls_sw_do_sendpage(sk, page, offset, size, flags);
1165         release_sock(sk);
1166         return ret;
1167 }
1168
1169 static struct sk_buff *tls_wait_data(struct sock *sk, struct sk_psock *psock,
1170                                      int flags, long timeo, int *err)
1171 {
1172         struct tls_context *tls_ctx = tls_get_ctx(sk);
1173         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1174         struct sk_buff *skb;
1175         DEFINE_WAIT_FUNC(wait, woken_wake_function);
1176
1177         while (!(skb = ctx->recv_pkt) && sk_psock_queue_empty(psock)) {
1178                 if (sk->sk_err) {
1179                         *err = sock_error(sk);
1180                         return NULL;
1181                 }
1182
1183                 if (sk->sk_shutdown & RCV_SHUTDOWN)
1184                         return NULL;
1185
1186                 if (sock_flag(sk, SOCK_DONE))
1187                         return NULL;
1188
1189                 if ((flags & MSG_DONTWAIT) || !timeo) {
1190                         *err = -EAGAIN;
1191                         return NULL;
1192                 }
1193
1194                 add_wait_queue(sk_sleep(sk), &wait);
1195                 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1196                 sk_wait_event(sk, &timeo,
1197                               ctx->recv_pkt != skb ||
1198                               !sk_psock_queue_empty(psock),
1199                               &wait);
1200                 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1201                 remove_wait_queue(sk_sleep(sk), &wait);
1202
1203                 /* Handle signals */
1204                 if (signal_pending(current)) {
1205                         *err = sock_intr_errno(timeo);
1206                         return NULL;
1207                 }
1208         }
1209
1210         return skb;
1211 }
1212
1213 static int tls_setup_from_iter(struct sock *sk, struct iov_iter *from,
1214                                int length, int *pages_used,
1215                                unsigned int *size_used,
1216                                struct scatterlist *to,
1217                                int to_max_pages)
1218 {
1219         int rc = 0, i = 0, num_elem = *pages_used, maxpages;
1220         struct page *pages[MAX_SKB_FRAGS];
1221         unsigned int size = *size_used;
1222         ssize_t copied, use;
1223         size_t offset;
1224
1225         while (length > 0) {
1226                 i = 0;
1227                 maxpages = to_max_pages - num_elem;
1228                 if (maxpages == 0) {
1229                         rc = -EFAULT;
1230                         goto out;
1231                 }
1232                 copied = iov_iter_get_pages(from, pages,
1233                                             length,
1234                                             maxpages, &offset);
1235                 if (copied <= 0) {
1236                         rc = -EFAULT;
1237                         goto out;
1238                 }
1239
1240                 iov_iter_advance(from, copied);
1241
1242                 length -= copied;
1243                 size += copied;
1244                 while (copied) {
1245                         use = min_t(int, copied, PAGE_SIZE - offset);
1246
1247                         sg_set_page(&to[num_elem],
1248                                     pages[i], use, offset);
1249                         sg_unmark_end(&to[num_elem]);
1250                         /* We do not uncharge memory from this API */
1251
1252                         offset = 0;
1253                         copied -= use;
1254
1255                         i++;
1256                         num_elem++;
1257                 }
1258         }
1259         /* Mark the end in the last sg entry if newly added */
1260         if (num_elem > *pages_used)
1261                 sg_mark_end(&to[num_elem - 1]);
1262 out:
1263         if (rc)
1264                 iov_iter_revert(from, size - *size_used);
1265         *size_used = size;
1266         *pages_used = num_elem;
1267
1268         return rc;
1269 }
1270
1271 /* This function decrypts the input skb into either out_iov or in out_sg
1272  * or in skb buffers itself. The input parameter 'zc' indicates if
1273  * zero-copy mode needs to be tried or not. With zero-copy mode, either
1274  * out_iov or out_sg must be non-NULL. In case both out_iov and out_sg are
1275  * NULL, then the decryption happens inside skb buffers itself, i.e.
1276  * zero-copy gets disabled and 'zc' is updated.
1277  */
1278
1279 static int decrypt_internal(struct sock *sk, struct sk_buff *skb,
1280                             struct iov_iter *out_iov,
1281                             struct scatterlist *out_sg,
1282                             int *chunk, bool *zc)
1283 {
1284         struct tls_context *tls_ctx = tls_get_ctx(sk);
1285         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1286         struct strp_msg *rxm = strp_msg(skb);
1287         int n_sgin, n_sgout, nsg, mem_size, aead_size, err, pages = 0;
1288         struct aead_request *aead_req;
1289         struct sk_buff *unused;
1290         u8 *aad, *iv, *mem = NULL;
1291         struct scatterlist *sgin = NULL;
1292         struct scatterlist *sgout = NULL;
1293         const int data_len = rxm->full_len - tls_ctx->rx.overhead_size;
1294
1295         if (*zc && (out_iov || out_sg)) {
1296                 if (out_iov)
1297                         n_sgout = iov_iter_npages(out_iov, INT_MAX) + 1;
1298                 else
1299                         n_sgout = sg_nents(out_sg);
1300                 n_sgin = skb_nsg(skb, rxm->offset + tls_ctx->rx.prepend_size,
1301                                  rxm->full_len - tls_ctx->rx.prepend_size);
1302         } else {
1303                 n_sgout = 0;
1304                 *zc = false;
1305                 n_sgin = skb_cow_data(skb, 0, &unused);
1306         }
1307
1308         if (n_sgin < 1)
1309                 return -EBADMSG;
1310
1311         /* Increment to accommodate AAD */
1312         n_sgin = n_sgin + 1;
1313
1314         nsg = n_sgin + n_sgout;
1315
1316         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(ctx->aead_recv);
1317         mem_size = aead_size + (nsg * sizeof(struct scatterlist));
1318         mem_size = mem_size + TLS_AAD_SPACE_SIZE;
1319         mem_size = mem_size + crypto_aead_ivsize(ctx->aead_recv);
1320
1321         /* Allocate a single block of memory which contains
1322          * aead_req || sgin[] || sgout[] || aad || iv.
1323          * This order achieves correct alignment for aead_req, sgin, sgout.
1324          */
1325         mem = kmalloc(mem_size, sk->sk_allocation);
1326         if (!mem)
1327                 return -ENOMEM;
1328
1329         /* Segment the allocated memory */
1330         aead_req = (struct aead_request *)mem;
1331         sgin = (struct scatterlist *)(mem + aead_size);
1332         sgout = sgin + n_sgin;
1333         aad = (u8 *)(sgout + n_sgout);
1334         iv = aad + TLS_AAD_SPACE_SIZE;
1335
1336         /* Prepare IV */
1337         err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
1338                             iv + TLS_CIPHER_AES_GCM_128_SALT_SIZE,
1339                             tls_ctx->rx.iv_size);
1340         if (err < 0) {
1341                 kfree(mem);
1342                 return err;
1343         }
1344         memcpy(iv, tls_ctx->rx.iv, TLS_CIPHER_AES_GCM_128_SALT_SIZE);
1345
1346         /* Prepare AAD */
1347         tls_make_aad(aad, rxm->full_len - tls_ctx->rx.overhead_size,
1348                      tls_ctx->rx.rec_seq, tls_ctx->rx.rec_seq_size,
1349                      ctx->control);
1350
1351         /* Prepare sgin */
1352         sg_init_table(sgin, n_sgin);
1353         sg_set_buf(&sgin[0], aad, TLS_AAD_SPACE_SIZE);
1354         err = skb_to_sgvec(skb, &sgin[1],
1355                            rxm->offset + tls_ctx->rx.prepend_size,
1356                            rxm->full_len - tls_ctx->rx.prepend_size);
1357         if (err < 0) {
1358                 kfree(mem);
1359                 return err;
1360         }
1361
1362         if (n_sgout) {
1363                 if (out_iov) {
1364                         sg_init_table(sgout, n_sgout);
1365                         sg_set_buf(&sgout[0], aad, TLS_AAD_SPACE_SIZE);
1366
1367                         *chunk = 0;
1368                         err = tls_setup_from_iter(sk, out_iov, data_len,
1369                                                   &pages, chunk, &sgout[1],
1370                                                   (n_sgout - 1));
1371                         if (err < 0)
1372                                 goto fallback_to_reg_recv;
1373                 } else if (out_sg) {
1374                         memcpy(sgout, out_sg, n_sgout * sizeof(*sgout));
1375                 } else {
1376                         goto fallback_to_reg_recv;
1377                 }
1378         } else {
1379 fallback_to_reg_recv:
1380                 sgout = sgin;
1381                 pages = 0;
1382                 *chunk = 0;
1383                 *zc = false;
1384         }
1385
1386         /* Prepare and submit AEAD request */
1387         err = tls_do_decryption(sk, skb, sgin, sgout, iv,
1388                                 data_len, aead_req, *zc);
1389         if (err == -EINPROGRESS)
1390                 return err;
1391
1392         /* Release the pages in case iov was mapped to pages */
1393         for (; pages > 0; pages--)
1394                 put_page(sg_page(&sgout[pages]));
1395
1396         kfree(mem);
1397         return err;
1398 }
1399
1400 static int decrypt_skb_update(struct sock *sk, struct sk_buff *skb,
1401                               struct iov_iter *dest, int *chunk, bool *zc)
1402 {
1403         struct tls_context *tls_ctx = tls_get_ctx(sk);
1404         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1405         struct strp_msg *rxm = strp_msg(skb);
1406         int err = 0;
1407
1408 #ifdef CONFIG_TLS_DEVICE
1409         err = tls_device_decrypted(sk, skb);
1410         if (err < 0)
1411                 return err;
1412 #endif
1413         if (!ctx->decrypted) {
1414                 err = decrypt_internal(sk, skb, dest, NULL, chunk, zc);
1415                 if (err < 0) {
1416                         if (err == -EINPROGRESS)
1417                                 tls_advance_record_sn(sk, &tls_ctx->rx);
1418
1419                         return err;
1420                 }
1421         } else {
1422                 *zc = false;
1423         }
1424
1425         rxm->offset += tls_ctx->rx.prepend_size;
1426         rxm->full_len -= tls_ctx->rx.overhead_size;
1427         tls_advance_record_sn(sk, &tls_ctx->rx);
1428         ctx->decrypted = true;
1429         ctx->saved_data_ready(sk);
1430
1431         return err;
1432 }
1433
1434 int decrypt_skb(struct sock *sk, struct sk_buff *skb,
1435                 struct scatterlist *sgout)
1436 {
1437         bool zc = true;
1438         int chunk;
1439
1440         return decrypt_internal(sk, skb, NULL, sgout, &chunk, &zc);
1441 }
1442
1443 static bool tls_sw_advance_skb(struct sock *sk, struct sk_buff *skb,
1444                                unsigned int len)
1445 {
1446         struct tls_context *tls_ctx = tls_get_ctx(sk);
1447         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1448
1449         if (skb) {
1450                 struct strp_msg *rxm = strp_msg(skb);
1451
1452                 if (len < rxm->full_len) {
1453                         rxm->offset += len;
1454                         rxm->full_len -= len;
1455                         return false;
1456                 }
1457                 kfree_skb(skb);
1458         }
1459
1460         /* Finished with message */
1461         ctx->recv_pkt = NULL;
1462         __strp_unpause(&ctx->strp);
1463
1464         return true;
1465 }
1466
1467 int tls_sw_recvmsg(struct sock *sk,
1468                    struct msghdr *msg,
1469                    size_t len,
1470                    int nonblock,
1471                    int flags,
1472                    int *addr_len)
1473 {
1474         struct tls_context *tls_ctx = tls_get_ctx(sk);
1475         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1476         struct sk_psock *psock;
1477         unsigned char control;
1478         struct strp_msg *rxm;
1479         struct sk_buff *skb;
1480         ssize_t copied = 0;
1481         bool cmsg = false;
1482         int target, err = 0;
1483         long timeo;
1484         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
1485         int num_async = 0;
1486
1487         flags |= nonblock;
1488
1489         if (unlikely(flags & MSG_ERRQUEUE))
1490                 return sock_recv_errqueue(sk, msg, len, SOL_IP, IP_RECVERR);
1491
1492         psock = sk_psock_get(sk);
1493         lock_sock(sk);
1494
1495         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1496         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1497         do {
1498                 bool zc = false;
1499                 bool async = false;
1500                 int chunk = 0;
1501
1502                 skb = tls_wait_data(sk, psock, flags, timeo, &err);
1503                 if (!skb) {
1504                         if (psock) {
1505                                 int ret = __tcp_bpf_recvmsg(sk, psock,
1506                                                             msg, len, flags);
1507
1508                                 if (ret > 0) {
1509                                         copied += ret;
1510                                         len -= ret;
1511                                         continue;
1512                                 }
1513                         }
1514                         goto recv_end;
1515                 }
1516
1517                 rxm = strp_msg(skb);
1518
1519                 if (!cmsg) {
1520                         int cerr;
1521
1522                         cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1523                                         sizeof(ctx->control), &ctx->control);
1524                         cmsg = true;
1525                         control = ctx->control;
1526                         if (ctx->control != TLS_RECORD_TYPE_DATA) {
1527                                 if (cerr || msg->msg_flags & MSG_CTRUNC) {
1528                                         err = -EIO;
1529                                         goto recv_end;
1530                                 }
1531                         }
1532                 } else if (control != ctx->control) {
1533                         goto recv_end;
1534                 }
1535
1536                 if (!ctx->decrypted) {
1537                         int to_copy = rxm->full_len - tls_ctx->rx.overhead_size;
1538
1539                         if (!is_kvec && to_copy <= len &&
1540                             likely(!(flags & MSG_PEEK)))
1541                                 zc = true;
1542
1543                         err = decrypt_skb_update(sk, skb, &msg->msg_iter,
1544                                                  &chunk, &zc);
1545                         if (err < 0 && err != -EINPROGRESS) {
1546                                 tls_err_abort(sk, EBADMSG);
1547                                 goto recv_end;
1548                         }
1549
1550                         if (err == -EINPROGRESS) {
1551                                 async = true;
1552                                 num_async++;
1553                                 goto pick_next_record;
1554                         }
1555
1556                         ctx->decrypted = true;
1557                 }
1558
1559                 if (!zc) {
1560                         chunk = min_t(unsigned int, rxm->full_len, len);
1561
1562                         err = skb_copy_datagram_msg(skb, rxm->offset, msg,
1563                                                     chunk);
1564                         if (err < 0)
1565                                 goto recv_end;
1566                 }
1567
1568 pick_next_record:
1569                 copied += chunk;
1570                 len -= chunk;
1571                 if (likely(!(flags & MSG_PEEK))) {
1572                         u8 control = ctx->control;
1573
1574                         /* For async, drop current skb reference */
1575                         if (async)
1576                                 skb = NULL;
1577
1578                         if (tls_sw_advance_skb(sk, skb, chunk)) {
1579                                 /* Return full control message to
1580                                  * userspace before trying to parse
1581                                  * another message type
1582                                  */
1583                                 msg->msg_flags |= MSG_EOR;
1584                                 if (control != TLS_RECORD_TYPE_DATA)
1585                                         goto recv_end;
1586                         } else {
1587                                 break;
1588                         }
1589                 } else {
1590                         /* MSG_PEEK right now cannot look beyond current skb
1591                          * from strparser, meaning we cannot advance skb here
1592                          * and thus unpause strparser since we'd loose original
1593                          * one.
1594                          */
1595                         break;
1596                 }
1597
1598                 /* If we have a new message from strparser, continue now. */
1599                 if (copied >= target && !ctx->recv_pkt)
1600                         break;
1601         } while (len);
1602
1603 recv_end:
1604         if (num_async) {
1605                 /* Wait for all previously submitted records to be decrypted */
1606                 smp_store_mb(ctx->async_notify, true);
1607                 if (atomic_read(&ctx->decrypt_pending)) {
1608                         err = crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1609                         if (err) {
1610                                 /* one of async decrypt failed */
1611                                 tls_err_abort(sk, err);
1612                                 copied = 0;
1613                         }
1614                 } else {
1615                         reinit_completion(&ctx->async_wait.completion);
1616                 }
1617                 WRITE_ONCE(ctx->async_notify, false);
1618         }
1619
1620         release_sock(sk);
1621         if (psock)
1622                 sk_psock_put(sk, psock);
1623         return copied ? : err;
1624 }
1625
1626 ssize_t tls_sw_splice_read(struct socket *sock,  loff_t *ppos,
1627                            struct pipe_inode_info *pipe,
1628                            size_t len, unsigned int flags)
1629 {
1630         struct tls_context *tls_ctx = tls_get_ctx(sock->sk);
1631         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1632         struct strp_msg *rxm = NULL;
1633         struct sock *sk = sock->sk;
1634         struct sk_buff *skb;
1635         ssize_t copied = 0;
1636         int err = 0;
1637         long timeo;
1638         int chunk;
1639         bool zc = false;
1640
1641         lock_sock(sk);
1642
1643         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1644
1645         skb = tls_wait_data(sk, NULL, flags, timeo, &err);
1646         if (!skb)
1647                 goto splice_read_end;
1648
1649         /* splice does not support reading control messages */
1650         if (ctx->control != TLS_RECORD_TYPE_DATA) {
1651                 err = -ENOTSUPP;
1652                 goto splice_read_end;
1653         }
1654
1655         if (!ctx->decrypted) {
1656                 err = decrypt_skb_update(sk, skb, NULL, &chunk, &zc);
1657
1658                 if (err < 0) {
1659                         tls_err_abort(sk, EBADMSG);
1660                         goto splice_read_end;
1661                 }
1662                 ctx->decrypted = true;
1663         }
1664         rxm = strp_msg(skb);
1665
1666         chunk = min_t(unsigned int, rxm->full_len, len);
1667         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, chunk, flags);
1668         if (copied < 0)
1669                 goto splice_read_end;
1670
1671         if (likely(!(flags & MSG_PEEK)))
1672                 tls_sw_advance_skb(sk, skb, copied);
1673
1674 splice_read_end:
1675         release_sock(sk);
1676         return copied ? : err;
1677 }
1678
1679 bool tls_sw_stream_read(const struct sock *sk)
1680 {
1681         struct tls_context *tls_ctx = tls_get_ctx(sk);
1682         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1683         bool ingress_empty = true;
1684         struct sk_psock *psock;
1685
1686         rcu_read_lock();
1687         psock = sk_psock(sk);
1688         if (psock)
1689                 ingress_empty = list_empty(&psock->ingress_msg);
1690         rcu_read_unlock();
1691
1692         return !ingress_empty || ctx->recv_pkt;
1693 }
1694
1695 static int tls_read_size(struct strparser *strp, struct sk_buff *skb)
1696 {
1697         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
1698         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1699         char header[TLS_HEADER_SIZE + MAX_IV_SIZE];
1700         struct strp_msg *rxm = strp_msg(skb);
1701         size_t cipher_overhead;
1702         size_t data_len = 0;
1703         int ret;
1704
1705         /* Verify that we have a full TLS header, or wait for more data */
1706         if (rxm->offset + tls_ctx->rx.prepend_size > skb->len)
1707                 return 0;
1708
1709         /* Sanity-check size of on-stack buffer. */
1710         if (WARN_ON(tls_ctx->rx.prepend_size > sizeof(header))) {
1711                 ret = -EINVAL;
1712                 goto read_failure;
1713         }
1714
1715         /* Linearize header to local buffer */
1716         ret = skb_copy_bits(skb, rxm->offset, header, tls_ctx->rx.prepend_size);
1717
1718         if (ret < 0)
1719                 goto read_failure;
1720
1721         ctx->control = header[0];
1722
1723         data_len = ((header[4] & 0xFF) | (header[3] << 8));
1724
1725         cipher_overhead = tls_ctx->rx.tag_size + tls_ctx->rx.iv_size;
1726
1727         if (data_len > TLS_MAX_PAYLOAD_SIZE + cipher_overhead) {
1728                 ret = -EMSGSIZE;
1729                 goto read_failure;
1730         }
1731         if (data_len < cipher_overhead) {
1732                 ret = -EBADMSG;
1733                 goto read_failure;
1734         }
1735
1736         if (header[1] != TLS_VERSION_MINOR(tls_ctx->crypto_recv.info.version) ||
1737             header[2] != TLS_VERSION_MAJOR(tls_ctx->crypto_recv.info.version)) {
1738                 ret = -EINVAL;
1739                 goto read_failure;
1740         }
1741
1742 #ifdef CONFIG_TLS_DEVICE
1743         handle_device_resync(strp->sk, TCP_SKB_CB(skb)->seq + rxm->offset,
1744                              *(u64*)tls_ctx->rx.rec_seq);
1745 #endif
1746         return data_len + TLS_HEADER_SIZE;
1747
1748 read_failure:
1749         tls_err_abort(strp->sk, ret);
1750
1751         return ret;
1752 }
1753
1754 static void tls_queue(struct strparser *strp, struct sk_buff *skb)
1755 {
1756         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
1757         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1758
1759         ctx->decrypted = false;
1760
1761         ctx->recv_pkt = skb;
1762         strp_pause(strp);
1763
1764         ctx->saved_data_ready(strp->sk);
1765 }
1766
1767 static void tls_data_ready(struct sock *sk)
1768 {
1769         struct tls_context *tls_ctx = tls_get_ctx(sk);
1770         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1771         struct sk_psock *psock;
1772
1773         strp_data_ready(&ctx->strp);
1774
1775         psock = sk_psock_get(sk);
1776         if (psock && !list_empty(&psock->ingress_msg)) {
1777                 ctx->saved_data_ready(sk);
1778                 sk_psock_put(sk, psock);
1779         }
1780 }
1781
1782 void tls_sw_free_resources_tx(struct sock *sk)
1783 {
1784         struct tls_context *tls_ctx = tls_get_ctx(sk);
1785         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1786         struct tls_rec *rec, *tmp;
1787
1788         /* Wait for any pending async encryptions to complete */
1789         smp_store_mb(ctx->async_notify, true);
1790         if (atomic_read(&ctx->encrypt_pending))
1791                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1792
1793         cancel_delayed_work_sync(&ctx->tx_work.work);
1794
1795         /* Tx whatever records we can transmit and abandon the rest */
1796         tls_tx_records(sk, -1);
1797
1798         /* Free up un-sent records in tx_list. First, free
1799          * the partially sent record if any at head of tx_list.
1800          */
1801         if (tls_ctx->partially_sent_record) {
1802                 struct scatterlist *sg = tls_ctx->partially_sent_record;
1803
1804                 while (1) {
1805                         put_page(sg_page(sg));
1806                         sk_mem_uncharge(sk, sg->length);
1807
1808                         if (sg_is_last(sg))
1809                                 break;
1810                         sg++;
1811                 }
1812
1813                 tls_ctx->partially_sent_record = NULL;
1814
1815                 rec = list_first_entry(&ctx->tx_list,
1816                                        struct tls_rec, list);
1817                 list_del(&rec->list);
1818                 sk_msg_free(sk, &rec->msg_plaintext);
1819                 kfree(rec);
1820         }
1821
1822         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
1823                 list_del(&rec->list);
1824                 sk_msg_free(sk, &rec->msg_encrypted);
1825                 sk_msg_free(sk, &rec->msg_plaintext);
1826                 kfree(rec);
1827         }
1828
1829         crypto_free_aead(ctx->aead_send);
1830         tls_free_open_rec(sk);
1831
1832         kfree(ctx);
1833 }
1834
1835 void tls_sw_release_resources_rx(struct sock *sk)
1836 {
1837         struct tls_context *tls_ctx = tls_get_ctx(sk);
1838         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1839
1840         if (ctx->aead_recv) {
1841                 kfree_skb(ctx->recv_pkt);
1842                 ctx->recv_pkt = NULL;
1843                 crypto_free_aead(ctx->aead_recv);
1844                 strp_stop(&ctx->strp);
1845                 write_lock_bh(&sk->sk_callback_lock);
1846                 sk->sk_data_ready = ctx->saved_data_ready;
1847                 write_unlock_bh(&sk->sk_callback_lock);
1848                 release_sock(sk);
1849                 strp_done(&ctx->strp);
1850                 lock_sock(sk);
1851         }
1852 }
1853
1854 void tls_sw_free_resources_rx(struct sock *sk)
1855 {
1856         struct tls_context *tls_ctx = tls_get_ctx(sk);
1857         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1858
1859         tls_sw_release_resources_rx(sk);
1860
1861         kfree(ctx);
1862 }
1863
1864 /* The work handler to transmitt the encrypted records in tx_list */
1865 static void tx_work_handler(struct work_struct *work)
1866 {
1867         struct delayed_work *delayed_work = to_delayed_work(work);
1868         struct tx_work *tx_work = container_of(delayed_work,
1869                                                struct tx_work, work);
1870         struct sock *sk = tx_work->sk;
1871         struct tls_context *tls_ctx = tls_get_ctx(sk);
1872         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1873
1874         if (!test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
1875                 return;
1876
1877         lock_sock(sk);
1878         tls_tx_records(sk, -1);
1879         release_sock(sk);
1880 }
1881
1882 int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
1883 {
1884         struct tls_crypto_info *crypto_info;
1885         struct tls12_crypto_info_aes_gcm_128 *gcm_128_info;
1886         struct tls_sw_context_tx *sw_ctx_tx = NULL;
1887         struct tls_sw_context_rx *sw_ctx_rx = NULL;
1888         struct cipher_context *cctx;
1889         struct crypto_aead **aead;
1890         struct strp_callbacks cb;
1891         u16 nonce_size, tag_size, iv_size, rec_seq_size;
1892         char *iv, *rec_seq;
1893         int rc = 0;
1894
1895         if (!ctx) {
1896                 rc = -EINVAL;
1897                 goto out;
1898         }
1899
1900         if (tx) {
1901                 if (!ctx->priv_ctx_tx) {
1902                         sw_ctx_tx = kzalloc(sizeof(*sw_ctx_tx), GFP_KERNEL);
1903                         if (!sw_ctx_tx) {
1904                                 rc = -ENOMEM;
1905                                 goto out;
1906                         }
1907                         ctx->priv_ctx_tx = sw_ctx_tx;
1908                 } else {
1909                         sw_ctx_tx =
1910                                 (struct tls_sw_context_tx *)ctx->priv_ctx_tx;
1911                 }
1912         } else {
1913                 if (!ctx->priv_ctx_rx) {
1914                         sw_ctx_rx = kzalloc(sizeof(*sw_ctx_rx), GFP_KERNEL);
1915                         if (!sw_ctx_rx) {
1916                                 rc = -ENOMEM;
1917                                 goto out;
1918                         }
1919                         ctx->priv_ctx_rx = sw_ctx_rx;
1920                 } else {
1921                         sw_ctx_rx =
1922                                 (struct tls_sw_context_rx *)ctx->priv_ctx_rx;
1923                 }
1924         }
1925
1926         if (tx) {
1927                 crypto_init_wait(&sw_ctx_tx->async_wait);
1928                 crypto_info = &ctx->crypto_send.info;
1929                 cctx = &ctx->tx;
1930                 aead = &sw_ctx_tx->aead_send;
1931                 INIT_LIST_HEAD(&sw_ctx_tx->tx_list);
1932                 INIT_DELAYED_WORK(&sw_ctx_tx->tx_work.work, tx_work_handler);
1933                 sw_ctx_tx->tx_work.sk = sk;
1934         } else {
1935                 crypto_init_wait(&sw_ctx_rx->async_wait);
1936                 crypto_info = &ctx->crypto_recv.info;
1937                 cctx = &ctx->rx;
1938                 aead = &sw_ctx_rx->aead_recv;
1939         }
1940
1941         switch (crypto_info->cipher_type) {
1942         case TLS_CIPHER_AES_GCM_128: {
1943                 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
1944                 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
1945                 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
1946                 iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
1947                 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
1948                 rec_seq =
1949                  ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
1950                 gcm_128_info =
1951                         (struct tls12_crypto_info_aes_gcm_128 *)crypto_info;
1952                 break;
1953         }
1954         default:
1955                 rc = -EINVAL;
1956                 goto free_priv;
1957         }
1958
1959         /* Sanity-check the IV size for stack allocations. */
1960         if (iv_size > MAX_IV_SIZE || nonce_size > MAX_IV_SIZE) {
1961                 rc = -EINVAL;
1962                 goto free_priv;
1963         }
1964
1965         cctx->prepend_size = TLS_HEADER_SIZE + nonce_size;
1966         cctx->tag_size = tag_size;
1967         cctx->overhead_size = cctx->prepend_size + cctx->tag_size;
1968         cctx->iv_size = iv_size;
1969         cctx->iv = kmalloc(iv_size + TLS_CIPHER_AES_GCM_128_SALT_SIZE,
1970                            GFP_KERNEL);
1971         if (!cctx->iv) {
1972                 rc = -ENOMEM;
1973                 goto free_priv;
1974         }
1975         memcpy(cctx->iv, gcm_128_info->salt, TLS_CIPHER_AES_GCM_128_SALT_SIZE);
1976         memcpy(cctx->iv + TLS_CIPHER_AES_GCM_128_SALT_SIZE, iv, iv_size);
1977         cctx->rec_seq_size = rec_seq_size;
1978         cctx->rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
1979         if (!cctx->rec_seq) {
1980                 rc = -ENOMEM;
1981                 goto free_iv;
1982         }
1983
1984         if (!*aead) {
1985                 *aead = crypto_alloc_aead("gcm(aes)", 0, 0);
1986                 if (IS_ERR(*aead)) {
1987                         rc = PTR_ERR(*aead);
1988                         *aead = NULL;
1989                         goto free_rec_seq;
1990                 }
1991         }
1992
1993         ctx->push_pending_record = tls_sw_push_pending_record;
1994
1995         rc = crypto_aead_setkey(*aead, gcm_128_info->key,
1996                                 TLS_CIPHER_AES_GCM_128_KEY_SIZE);
1997         if (rc)
1998                 goto free_aead;
1999
2000         rc = crypto_aead_setauthsize(*aead, cctx->tag_size);
2001         if (rc)
2002                 goto free_aead;
2003
2004         if (sw_ctx_rx) {
2005                 /* Set up strparser */
2006                 memset(&cb, 0, sizeof(cb));
2007                 cb.rcv_msg = tls_queue;
2008                 cb.parse_msg = tls_read_size;
2009
2010                 strp_init(&sw_ctx_rx->strp, sk, &cb);
2011
2012                 write_lock_bh(&sk->sk_callback_lock);
2013                 sw_ctx_rx->saved_data_ready = sk->sk_data_ready;
2014                 sk->sk_data_ready = tls_data_ready;
2015                 write_unlock_bh(&sk->sk_callback_lock);
2016
2017                 strp_check_rcv(&sw_ctx_rx->strp);
2018         }
2019
2020         goto out;
2021
2022 free_aead:
2023         crypto_free_aead(*aead);
2024         *aead = NULL;
2025 free_rec_seq:
2026         kfree(cctx->rec_seq);
2027         cctx->rec_seq = NULL;
2028 free_iv:
2029         kfree(cctx->iv);
2030         cctx->iv = NULL;
2031 free_priv:
2032         if (tx) {
2033                 kfree(ctx->priv_ctx_tx);
2034                 ctx->priv_ctx_tx = NULL;
2035         } else {
2036                 kfree(ctx->priv_ctx_rx);
2037                 ctx->priv_ctx_rx = NULL;
2038         }
2039 out:
2040         return rc;
2041 }