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         memcpy(rec->iv_data, tls_ctx->tx.iv, sizeof(rec->iv_data));
 443
 444         sge->offset += tls_ctx->tx.prepend_size;
 445         sge->length -= tls_ctx->tx.prepend_size;
 446
 447         msg_en->sg.curr = start;
 448
 449         aead_request_set_tfm(aead_req, ctx->aead_send);
 450         aead_request_set_ad(aead_req, TLS_AAD_SPACE_SIZE);
 451         aead_request_set_crypt(aead_req, rec->sg_aead_in,
 452                                rec->sg_aead_out,
 453                                data_len, rec->iv_data);
 454
 455         aead_request_set_callback(aead_req, CRYPTO_TFM_REQ_MAY_BACKLOG,
 456                                   tls_encrypt_done, sk);
 457
 458         /* Add the record in tx_list */
 459         list_add_tail((struct list_head *)&rec->list, &ctx->tx_list);
 460         atomic_inc(&ctx->encrypt_pending);
 461
 462         rc = crypto_aead_encrypt(aead_req);
 463         if (!rc || rc != -EINPROGRESS) {
 464                 atomic_dec(&ctx->encrypt_pending);
 465                 sge->offset -= tls_ctx->tx.prepend_size;
 466                 sge->length += tls_ctx->tx.prepend_size;
 467         }
 468
 469         if (!rc) {
 470                 WRITE_ONCE(rec->tx_ready, true);
 471         } else if (rc != -EINPROGRESS) {
 472                 list_del(&rec->list);
 473                 return rc;
 474         }
 475
 476         /* Unhook the record from context if encryption is not failure */
 477         ctx->open_rec = NULL;
 478         tls_advance_record_sn(sk, &tls_ctx->tx);
 479         return rc;
 480 }
 481
 482 static int tls_split_open_record(struct sock *sk, struct tls_rec *from,
 483                                  struct tls_rec **to, struct sk_msg *msg_opl,
 484                                  struct sk_msg *msg_oen, u32 split_point,
 485                                  u32 tx_overhead_size, u32 *orig_end)
 486 {
 487         u32 i, j, bytes = 0, apply = msg_opl->apply_bytes;
 488         struct scatterlist *sge, *osge, *nsge;
 489         u32 orig_size = msg_opl->sg.size;
 490         struct scatterlist tmp = { };
 491         struct sk_msg *msg_npl;
 492         struct tls_rec *new;
 493         int ret;
 494
 495         new = tls_get_rec(sk);
 496         if (!new)
 497                 return -ENOMEM;
 498         ret = sk_msg_alloc(sk, &new->msg_encrypted, msg_opl->sg.size +
 499                            tx_overhead_size, 0);
 500         if (ret < 0) {
 501                 tls_free_rec(sk, new);
 502                 return ret;
 503         }
 504
 505         *orig_end = msg_opl->sg.end;
 506         i = msg_opl->sg.start;
 507         sge = sk_msg_elem(msg_opl, i);
 508         while (apply && sge->length) {
 509                 if (sge->length > apply) {
 510                         u32 len = sge->length - apply;
 511
 512                         get_page(sg_page(sge));
 513                         sg_set_page(&tmp, sg_page(sge), len,
 514                                     sge->offset + apply);
 515                         sge->length = apply;
 516                         bytes += apply;
 517                         apply = 0;
 518                 } else {
 519                         apply -= sge->length;
 520                         bytes += sge->length;
 521                 }
 522
 523                 sk_msg_iter_var_next(i);
 524                 if (i == msg_opl->sg.end)
 525                         break;
 526                 sge = sk_msg_elem(msg_opl, i);
 527         }
 528
 529         msg_opl->sg.end = i;
 530         msg_opl->sg.curr = i;
 531         msg_opl->sg.copybreak = 0;
 532         msg_opl->apply_bytes = 0;
 533         msg_opl->sg.size = bytes;
 534
 535         msg_npl = &new->msg_plaintext;
 536         msg_npl->apply_bytes = apply;
 537         msg_npl->sg.size = orig_size - bytes;
 538
 539         j = msg_npl->sg.start;
 540         nsge = sk_msg_elem(msg_npl, j);
 541         if (tmp.length) {
 542                 memcpy(nsge, &tmp, sizeof(*nsge));
 543                 sk_msg_iter_var_next(j);
 544                 nsge = sk_msg_elem(msg_npl, j);
 545         }
 546
 547         osge = sk_msg_elem(msg_opl, i);
 548         while (osge->length) {
 549                 memcpy(nsge, osge, sizeof(*nsge));
 550                 sg_unmark_end(nsge);
 551                 sk_msg_iter_var_next(i);
 552                 sk_msg_iter_var_next(j);
 553                 if (i == *orig_end)
 554                         break;
 555                 osge = sk_msg_elem(msg_opl, i);
 556                 nsge = sk_msg_elem(msg_npl, j);
 557         }
 558
 559         msg_npl->sg.end = j;
 560         msg_npl->sg.curr = j;
 561         msg_npl->sg.copybreak = 0;
 562
 563         *to = new;
 564         return 0;
 565 }
 566
 567 static void tls_merge_open_record(struct sock *sk, struct tls_rec *to,
 568                                   struct tls_rec *from, u32 orig_end)
 569 {
 570         struct sk_msg *msg_npl = &from->msg_plaintext;
 571         struct sk_msg *msg_opl = &to->msg_plaintext;
 572         struct scatterlist *osge, *nsge;
 573         u32 i, j;
 574
 575         i = msg_opl->sg.end;
 576         sk_msg_iter_var_prev(i);
 577         j = msg_npl->sg.start;
 578
 579         osge = sk_msg_elem(msg_opl, i);
 580         nsge = sk_msg_elem(msg_npl, j);
 581
 582         if (sg_page(osge) == sg_page(nsge) &&
 583             osge->offset + osge->length == nsge->offset) {
 584                 osge->length += nsge->length;
 585                 put_page(sg_page(nsge));
 586         }
 587
 588         msg_opl->sg.end = orig_end;
 589         msg_opl->sg.curr = orig_end;
 590         msg_opl->sg.copybreak = 0;
 591         msg_opl->apply_bytes = msg_opl->sg.size + msg_npl->sg.size;
 592         msg_opl->sg.size += msg_npl->sg.size;
 593
 594         sk_msg_free(sk, &to->msg_encrypted);
 595         sk_msg_xfer_full(&to->msg_encrypted, &from->msg_encrypted);
 596
 597         kfree(from);
 598 }
 599
 600 static int tls_push_record(struct sock *sk, int flags,
 601                            unsigned char record_type)
 602 {
 603         struct tls_context *tls_ctx = tls_get_ctx(sk);
 604         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 605         struct tls_rec *rec = ctx->open_rec, *tmp = NULL;
 606         u32 i, split_point, uninitialized_var(orig_end);
 607         struct sk_msg *msg_pl, *msg_en;
 608         struct aead_request *req;
 609         bool split;
 610         int rc;
 611
 612         if (!rec)
 613                 return 0;
 614
 615         msg_pl = &rec->msg_plaintext;
 616         msg_en = &rec->msg_encrypted;
 617
 618         split_point = msg_pl->apply_bytes;
 619         split = split_point && split_point < msg_pl->sg.size;
 620         if (split) {
 621                 rc = tls_split_open_record(sk, rec, &tmp, msg_pl, msg_en,
 622                                            split_point, tls_ctx->tx.overhead_size,
 623                                            &orig_end);
 624                 if (rc < 0)
 625                         return rc;
 626                 sk_msg_trim(sk, msg_en, msg_pl->sg.size +
 627                             tls_ctx->tx.overhead_size);
 628         }
 629
 630         rec->tx_flags = flags;
 631         req = &rec->aead_req;
 632
 633         i = msg_pl->sg.end;
 634         sk_msg_iter_var_prev(i);
 635         sg_mark_end(sk_msg_elem(msg_pl, i));
 636
 637         i = msg_pl->sg.start;
 638         sg_chain(rec->sg_aead_in, 2, rec->inplace_crypto ?
 639                  &msg_en->sg.data[i] : &msg_pl->sg.data[i]);
 640
 641         i = msg_en->sg.end;
 642         sk_msg_iter_var_prev(i);
 643         sg_mark_end(sk_msg_elem(msg_en, i));
 644
 645         i = msg_en->sg.start;
 646         sg_chain(rec->sg_aead_out, 2, &msg_en->sg.data[i]);
 647
 648         tls_make_aad(rec->aad_space, msg_pl->sg.size,
 649                      tls_ctx->tx.rec_seq, tls_ctx->tx.rec_seq_size,
 650                      record_type);
 651
 652         tls_fill_prepend(tls_ctx,
 653                          page_address(sg_page(&msg_en->sg.data[i])) +
 654                          msg_en->sg.data[i].offset, msg_pl->sg.size,
 655                          record_type);
 656
 657         tls_ctx->pending_open_record_frags = false;
 658
 659         rc = tls_do_encryption(sk, tls_ctx, ctx, req, msg_pl->sg.size, i);
 660         if (rc < 0) {
 661                 if (rc != -EINPROGRESS) {
 662                         tls_err_abort(sk, EBADMSG);
 663                         if (split) {
 664                                 tls_ctx->pending_open_record_frags = true;
 665                                 tls_merge_open_record(sk, rec, tmp, orig_end);
 666                         }
 667                 }
 668                 return rc;
 669         } else if (split) {
 670                 msg_pl = &tmp->msg_plaintext;
 671                 msg_en = &tmp->msg_encrypted;
 672                 sk_msg_trim(sk, msg_en, msg_pl->sg.size +
 673                             tls_ctx->tx.overhead_size);
 674                 tls_ctx->pending_open_record_frags = true;
 675                 ctx->open_rec = tmp;
 676         }
 677
 678         return tls_tx_records(sk, flags);
 679 }
 680
 681 static int bpf_exec_tx_verdict(struct sk_msg *msg, struct sock *sk,
 682                                bool full_record, u8 record_type,
 683                                size_t *copied, int flags)
 684 {
 685         struct tls_context *tls_ctx = tls_get_ctx(sk);
 686         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 687         struct sk_msg msg_redir = { };
 688         struct sk_psock *psock;
 689         struct sock *sk_redir;
 690         struct tls_rec *rec;
 691         bool enospc, policy;
 692         int err = 0, send;
 693         u32 delta = 0;
 694
 695         policy = !(flags & MSG_SENDPAGE_NOPOLICY);
 696         psock = sk_psock_get(sk);
 697         if (!psock || !policy)
 698                 return tls_push_record(sk, flags, record_type);
 699 more_data:
 700         enospc = sk_msg_full(msg);
 701         if (psock->eval == __SK_NONE) {
 702                 delta = msg->sg.size;
 703                 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
 704                 if (delta < msg->sg.size)
 705                         delta -= msg->sg.size;
 706                 else
 707                         delta = 0;
 708         }
 709         if (msg->cork_bytes && msg->cork_bytes > msg->sg.size &&
 710             !enospc && !full_record) {
 711                 err = -ENOSPC;
 712                 goto out_err;
 713         }
 714         msg->cork_bytes = 0;
 715         send = msg->sg.size;
 716         if (msg->apply_bytes && msg->apply_bytes < send)
 717                 send = msg->apply_bytes;
 718
 719         switch (psock->eval) {
 720         case __SK_PASS:
 721                 err = tls_push_record(sk, flags, record_type);
 722                 if (err < 0) {
 723                         *copied -= sk_msg_free(sk, msg);
 724                         tls_free_open_rec(sk);
 725                         goto out_err;
 726                 }
 727                 break;
 728         case __SK_REDIRECT:
 729                 sk_redir = psock->sk_redir;
 730                 memcpy(&msg_redir, msg, sizeof(*msg));
 731                 if (msg->apply_bytes < send)
 732                         msg->apply_bytes = 0;
 733                 else
 734                         msg->apply_bytes -= send;
 735                 sk_msg_return_zero(sk, msg, send);
 736                 msg->sg.size -= send;
 737                 release_sock(sk);
 738                 err = tcp_bpf_sendmsg_redir(sk_redir, &msg_redir, send, flags);
 739                 lock_sock(sk);
 740                 if (err < 0) {
 741                         *copied -= sk_msg_free_nocharge(sk, &msg_redir);
 742                         msg->sg.size = 0;
 743                 }
 744                 if (msg->sg.size == 0)
 745                         tls_free_open_rec(sk);
 746                 break;
 747         case __SK_DROP:
 748         default:
 749                 sk_msg_free_partial(sk, msg, send);
 750                 if (msg->apply_bytes < send)
 751                         msg->apply_bytes = 0;
 752                 else
 753                         msg->apply_bytes -= send;
 754                 if (msg->sg.size == 0)
 755                         tls_free_open_rec(sk);
 756                 *copied -= (send + delta);
 757                 err = -EACCES;
 758         }
 759
 760         if (likely(!err)) {
 761                 bool reset_eval = !ctx->open_rec;
 762
 763                 rec = ctx->open_rec;
 764                 if (rec) {
 765                         msg = &rec->msg_plaintext;
 766                         if (!msg->apply_bytes)
 767                                 reset_eval = true;
 768                 }
 769                 if (reset_eval) {
 770                         psock->eval = __SK_NONE;
 771                         if (psock->sk_redir) {
 772                                 sock_put(psock->sk_redir);
 773                                 psock->sk_redir = NULL;
 774                         }
 775                 }
 776                 if (rec)
 777                         goto more_data;
 778         }
 779  out_err:
 780         sk_psock_put(sk, psock);
 781         return err;
 782 }
 783
 784 static int tls_sw_push_pending_record(struct sock *sk, int flags)
 785 {
 786         struct tls_context *tls_ctx = tls_get_ctx(sk);
 787         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 788         struct tls_rec *rec = ctx->open_rec;
 789         struct sk_msg *msg_pl;
 790         size_t copied;
 791
 792         if (!rec)
 793                 return 0;
 794
 795         msg_pl = &rec->msg_plaintext;
 796         copied = msg_pl->sg.size;
 797         if (!copied)
 798                 return 0;
 799
 800         return bpf_exec_tx_verdict(msg_pl, sk, true, TLS_RECORD_TYPE_DATA,
 801                                    &copied, flags);
 802 }
 803
 804 int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 805 {
 806         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 807         struct tls_context *tls_ctx = tls_get_ctx(sk);
 808         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 809         struct crypto_tfm *tfm = crypto_aead_tfm(ctx->aead_send);
 810         bool async_capable = tfm->__crt_alg->cra_flags & CRYPTO_ALG_ASYNC;
 811         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 812         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
 813         bool eor = !(msg->msg_flags & MSG_MORE);
 814         size_t try_to_copy, copied = 0;
 815         struct sk_msg *msg_pl, *msg_en;
 816         struct tls_rec *rec;
 817         int required_size;
 818         int num_async = 0;
 819         bool full_record;
 820         int record_room;
 821         int num_zc = 0;
 822         int orig_size;
 823         int ret = 0;
 824
 825         if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
 826                 return -ENOTSUPP;
 827
 828         lock_sock(sk);
 829
 830         /* Wait till there is any pending write on socket */
 831         if (unlikely(sk->sk_write_pending)) {
 832                 ret = wait_on_pending_writer(sk, &timeo);
 833                 if (unlikely(ret))
 834                         goto send_end;
 835         }
 836
 837         if (unlikely(msg->msg_controllen)) {
 838                 ret = tls_proccess_cmsg(sk, msg, &record_type);
 839                 if (ret) {
 840                         if (ret == -EINPROGRESS)
 841                                 num_async++;
 842                         else if (ret != -EAGAIN)
 843                                 goto send_end;
 844                 }
 845         }
 846
 847         while (msg_data_left(msg)) {
 848                 if (sk->sk_err) {
 849                         ret = -sk->sk_err;
 850                         goto send_end;
 851                 }
 852
 853                 if (ctx->open_rec)
 854                         rec = ctx->open_rec;
 855                 else
 856                         rec = ctx->open_rec = tls_get_rec(sk);
 857                 if (!rec) {
 858                         ret = -ENOMEM;
 859                         goto send_end;
 860                 }
 861
 862                 msg_pl = &rec->msg_plaintext;
 863                 msg_en = &rec->msg_encrypted;
 864
 865                 orig_size = msg_pl->sg.size;
 866                 full_record = false;
 867                 try_to_copy = msg_data_left(msg);
 868                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
 869                 if (try_to_copy >= record_room) {
 870                         try_to_copy = record_room;
 871                         full_record = true;
 872                 }
 873
 874                 required_size = msg_pl->sg.size + try_to_copy +
 875                                 tls_ctx->tx.overhead_size;
 876
 877                 if (!sk_stream_memory_free(sk))
 878                         goto wait_for_sndbuf;
 879
 880 alloc_encrypted:
 881                 ret = tls_alloc_encrypted_msg(sk, required_size);
 882                 if (ret) {
 883                         if (ret != -ENOSPC)
 884                                 goto wait_for_memory;
 885
 886                         /* Adjust try_to_copy according to the amount that was
 887                          * actually allocated. The difference is due
 888                          * to max sg elements limit
 889                          */
 890                         try_to_copy -= required_size - msg_en->sg.size;
 891                         full_record = true;
 892                 }
 893
 894                 if (!is_kvec && (full_record || eor) && !async_capable) {
 895                         u32 first = msg_pl->sg.end;
 896
 897                         ret = sk_msg_zerocopy_from_iter(sk, &msg->msg_iter,
 898                                                         msg_pl, try_to_copy);
 899                         if (ret)
 900                                 goto fallback_to_reg_send;
 901
 902                         rec->inplace_crypto = 0;
 903
 904                         num_zc++;
 905                         copied += try_to_copy;
 906
 907                         sk_msg_sg_copy_set(msg_pl, first);
 908                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
 909                                                   record_type, &copied,
 910                                                   msg->msg_flags);
 911                         if (ret) {
 912                                 if (ret == -EINPROGRESS)
 913                                         num_async++;
 914                                 else if (ret == -ENOMEM)
 915                                         goto wait_for_memory;
 916                                 else if (ret == -ENOSPC)
 917                                         goto rollback_iter;
 918                                 else if (ret != -EAGAIN)
 919                                         goto send_end;
 920                         }
 921                         continue;
 922 rollback_iter:
 923                         copied -= try_to_copy;
 924                         sk_msg_sg_copy_clear(msg_pl, first);
 925                         iov_iter_revert(&msg->msg_iter,
 926                                         msg_pl->sg.size - orig_size);
 927 fallback_to_reg_send:
 928                         sk_msg_trim(sk, msg_pl, orig_size);
 929                 }
 930
 931                 required_size = msg_pl->sg.size + try_to_copy;
 932
 933                 ret = tls_clone_plaintext_msg(sk, required_size);
 934                 if (ret) {
 935                         if (ret != -ENOSPC)
 936                                 goto send_end;
 937
 938                         /* Adjust try_to_copy according to the amount that was
 939                          * actually allocated. The difference is due
 940                          * to max sg elements limit
 941                          */
 942                         try_to_copy -= required_size - msg_pl->sg.size;
 943                         full_record = true;
 944                         sk_msg_trim(sk, msg_en, msg_pl->sg.size +
 945                                     tls_ctx->tx.overhead_size);
 946                 }
 947
 948                 if (try_to_copy) {
 949                         ret = sk_msg_memcopy_from_iter(sk, &msg->msg_iter,
 950                                                        msg_pl, try_to_copy);
 951                         if (ret < 0)
 952                                 goto trim_sgl;
 953                 }
 954
 955                 /* Open records defined only if successfully copied, otherwise
 956                  * we would trim the sg but not reset the open record frags.
 957                  */
 958                 tls_ctx->pending_open_record_frags = true;
 959                 copied += try_to_copy;
 960                 if (full_record || eor) {
 961                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
 962                                                   record_type, &copied,
 963                                                   msg->msg_flags);
 964                         if (ret) {
 965                                 if (ret == -EINPROGRESS)
 966                                         num_async++;
 967                                 else if (ret == -ENOMEM)
 968                                         goto wait_for_memory;
 969                                 else if (ret != -EAGAIN) {
 970                                         if (ret == -ENOSPC)
 971                                                 ret = 0;
 972                                         goto send_end;
 973                                 }
 974                         }
 975                 }
 976
 977                 continue;
 978
 979 wait_for_sndbuf:
 980                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 981 wait_for_memory:
 982                 ret = sk_stream_wait_memory(sk, &timeo);
 983                 if (ret) {
 984 trim_sgl:
 985                         tls_trim_both_msgs(sk, orig_size);
 986                         goto send_end;
 987                 }
 988
 989                 if (msg_en->sg.size < required_size)
 990                         goto alloc_encrypted;
 991         }
 992
 993         if (!num_async) {
 994                 goto send_end;
 995         } else if (num_zc) {
 996                 /* Wait for pending encryptions to get completed */
 997                 smp_store_mb(ctx->async_notify, true);
 998
 999                 if (atomic_read(&ctx->encrypt_pending))
1000                         crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1001                 else
1002                         reinit_completion(&ctx->async_wait.completion);
1003
1004                 WRITE_ONCE(ctx->async_notify, false);
1005
1006                 if (ctx->async_wait.err) {
1007                         ret = ctx->async_wait.err;
1008                         copied = 0;
1009                 }
1010         }
1011
1012         /* Transmit if any encryptions have completed */
1013         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1014                 cancel_delayed_work(&ctx->tx_work.work);
1015                 tls_tx_records(sk, msg->msg_flags);
1016         }
1017
1018 send_end:
1019         ret = sk_stream_error(sk, msg->msg_flags, ret);
1020
1021         release_sock(sk);
1022         return copied ? copied : ret;
1023 }
1024
1025 int tls_sw_do_sendpage(struct sock *sk, struct page *page,
1026                        int offset, size_t size, int flags)
1027 {
1028         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1029         struct tls_context *tls_ctx = tls_get_ctx(sk);
1030         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1031         unsigned char record_type = TLS_RECORD_TYPE_DATA;
1032         struct sk_msg *msg_pl;
1033         struct tls_rec *rec;
1034         int num_async = 0;
1035         size_t copied = 0;
1036         bool full_record;
1037         int record_room;
1038         int ret = 0;
1039         bool eor;
1040
1041         eor = !(flags & (MSG_MORE | MSG_SENDPAGE_NOTLAST));
1042         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1043
1044         /* Wait till there is any pending write on socket */
1045         if (unlikely(sk->sk_write_pending)) {
1046                 ret = wait_on_pending_writer(sk, &timeo);
1047                 if (unlikely(ret))
1048                         goto sendpage_end;
1049         }
1050
1051         /* Call the sk_stream functions to manage the sndbuf mem. */
1052         while (size > 0) {
1053                 size_t copy, required_size;
1054
1055                 if (sk->sk_err) {
1056                         ret = -sk->sk_err;
1057                         goto sendpage_end;
1058                 }
1059
1060                 if (ctx->open_rec)
1061                         rec = ctx->open_rec;
1062                 else
1063                         rec = ctx->open_rec = tls_get_rec(sk);
1064                 if (!rec) {
1065                         ret = -ENOMEM;
1066                         goto sendpage_end;
1067                 }
1068
1069                 msg_pl = &rec->msg_plaintext;
1070
1071                 full_record = false;
1072                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
1073                 copied = 0;
1074                 copy = size;
1075                 if (copy >= record_room) {
1076                         copy = record_room;
1077                         full_record = true;
1078                 }
1079
1080                 required_size = msg_pl->sg.size + copy +
1081                                 tls_ctx->tx.overhead_size;
1082
1083                 if (!sk_stream_memory_free(sk))
1084                         goto wait_for_sndbuf;
1085 alloc_payload:
1086                 ret = tls_alloc_encrypted_msg(sk, required_size);
1087                 if (ret) {
1088                         if (ret != -ENOSPC)
1089                                 goto wait_for_memory;
1090
1091                         /* Adjust copy according to the amount that was
1092                          * actually allocated. The difference is due
1093                          * to max sg elements limit
1094                          */
1095                         copy -= required_size - msg_pl->sg.size;
1096                         full_record = true;
1097                 }
1098
1099                 sk_msg_page_add(msg_pl, page, copy, offset);
1100                 sk_mem_charge(sk, copy);
1101
1102                 offset += copy;
1103                 size -= copy;
1104                 copied += copy;
1105
1106                 tls_ctx->pending_open_record_frags = true;
1107                 if (full_record || eor || sk_msg_full(msg_pl)) {
1108                         rec->inplace_crypto = 0;
1109                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1110                                                   record_type, &copied, flags);
1111                         if (ret) {
1112                                 if (ret == -EINPROGRESS)
1113                                         num_async++;
1114                                 else if (ret == -ENOMEM)
1115                                         goto wait_for_memory;
1116                                 else if (ret != -EAGAIN) {
1117                                         if (ret == -ENOSPC)
1118                                                 ret = 0;
1119                                         goto sendpage_end;
1120                                 }
1121                         }
1122                 }
1123                 continue;
1124 wait_for_sndbuf:
1125                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1126 wait_for_memory:
1127                 ret = sk_stream_wait_memory(sk, &timeo);
1128                 if (ret) {
1129                         tls_trim_both_msgs(sk, msg_pl->sg.size);
1130                         goto sendpage_end;
1131                 }
1132
1133                 goto alloc_payload;
1134         }
1135
1136         if (num_async) {
1137                 /* Transmit if any encryptions have completed */
1138                 if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1139                         cancel_delayed_work(&ctx->tx_work.work);
1140                         tls_tx_records(sk, flags);
1141                 }
1142         }
1143 sendpage_end:
1144         ret = sk_stream_error(sk, flags, ret);
1145         return copied ? copied : ret;
1146 }
1147
1148 int tls_sw_sendpage_locked(struct sock *sk, struct page *page,
1149                            int offset, size_t size, int flags)
1150 {
1151         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1152                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY))
1153                 return -ENOTSUPP;
1154
1155         return tls_sw_do_sendpage(sk, page, offset, size, flags);
1156 }
1157
1158 int tls_sw_sendpage(struct sock *sk, struct page *page,
1159                     int offset, size_t size, int flags)
1160 {
1161         int ret;
1162
1163         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1164                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY))
1165                 return -ENOTSUPP;
1166
1167         lock_sock(sk);
1168         ret = tls_sw_do_sendpage(sk, page, offset, size, flags);
1169         release_sock(sk);
1170         return ret;
1171 }
1172
1173 static struct sk_buff *tls_wait_data(struct sock *sk, struct sk_psock *psock,
1174                                      int flags, long timeo, int *err)
1175 {
1176         struct tls_context *tls_ctx = tls_get_ctx(sk);
1177         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1178         struct sk_buff *skb;
1179         DEFINE_WAIT_FUNC(wait, woken_wake_function);
1180
1181         while (!(skb = ctx->recv_pkt) && sk_psock_queue_empty(psock)) {
1182                 if (sk->sk_err) {
1183                         *err = sock_error(sk);
1184                         return NULL;
1185                 }
1186
1187                 if (sk->sk_shutdown & RCV_SHUTDOWN)
1188                         return NULL;
1189
1190                 if (sock_flag(sk, SOCK_DONE))
1191                         return NULL;
1192
1193                 if ((flags & MSG_DONTWAIT) || !timeo) {
1194                         *err = -EAGAIN;
1195                         return NULL;
1196                 }
1197
1198                 add_wait_queue(sk_sleep(sk), &wait);
1199                 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1200                 sk_wait_event(sk, &timeo,
1201                               ctx->recv_pkt != skb ||
1202                               !sk_psock_queue_empty(psock),
1203                               &wait);
1204                 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1205                 remove_wait_queue(sk_sleep(sk), &wait);
1206
1207                 /* Handle signals */
1208                 if (signal_pending(current)) {
1209                         *err = sock_intr_errno(timeo);
1210                         return NULL;
1211                 }
1212         }
1213
1214         return skb;
1215 }
1216
1217 static int tls_setup_from_iter(struct sock *sk, struct iov_iter *from,
1218                                int length, int *pages_used,
1219                                unsigned int *size_used,
1220                                struct scatterlist *to,
1221                                int to_max_pages)
1222 {
1223         int rc = 0, i = 0, num_elem = *pages_used, maxpages;
1224         struct page *pages[MAX_SKB_FRAGS];
1225         unsigned int size = *size_used;
1226         ssize_t copied, use;
1227         size_t offset;
1228
1229         while (length > 0) {
1230                 i = 0;
1231                 maxpages = to_max_pages - num_elem;
1232                 if (maxpages == 0) {
1233                         rc = -EFAULT;
1234                         goto out;
1235                 }
1236                 copied = iov_iter_get_pages(from, pages,
1237                                             length,
1238                                             maxpages, &offset);
1239                 if (copied <= 0) {
1240                         rc = -EFAULT;
1241                         goto out;
1242                 }
1243
1244                 iov_iter_advance(from, copied);
1245
1246                 length -= copied;
1247                 size += copied;
1248                 while (copied) {
1249                         use = min_t(int, copied, PAGE_SIZE - offset);
1250
1251                         sg_set_page(&to[num_elem],
1252                                     pages[i], use, offset);
1253                         sg_unmark_end(&to[num_elem]);
1254                         /* We do not uncharge memory from this API */
1255
1256                         offset = 0;
1257                         copied -= use;
1258
1259                         i++;
1260                         num_elem++;
1261                 }
1262         }
1263         /* Mark the end in the last sg entry if newly added */
1264         if (num_elem > *pages_used)
1265                 sg_mark_end(&to[num_elem - 1]);
1266 out:
1267         if (rc)
1268                 iov_iter_revert(from, size - *size_used);
1269         *size_used = size;
1270         *pages_used = num_elem;
1271
1272         return rc;
1273 }
1274
1275 /* This function decrypts the input skb into either out_iov or in out_sg
1276  * or in skb buffers itself. The input parameter 'zc' indicates if
1277  * zero-copy mode needs to be tried or not. With zero-copy mode, either
1278  * out_iov or out_sg must be non-NULL. In case both out_iov and out_sg are
1279  * NULL, then the decryption happens inside skb buffers itself, i.e.
1280  * zero-copy gets disabled and 'zc' is updated.
1281  */
1282
1283 static int decrypt_internal(struct sock *sk, struct sk_buff *skb,
1284                             struct iov_iter *out_iov,
1285                             struct scatterlist *out_sg,
1286                             int *chunk, bool *zc)
1287 {
1288         struct tls_context *tls_ctx = tls_get_ctx(sk);
1289         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1290         struct strp_msg *rxm = strp_msg(skb);
1291         int n_sgin, n_sgout, nsg, mem_size, aead_size, err, pages = 0;
1292         struct aead_request *aead_req;
1293         struct sk_buff *unused;
1294         u8 *aad, *iv, *mem = NULL;
1295         struct scatterlist *sgin = NULL;
1296         struct scatterlist *sgout = NULL;
1297         const int data_len = rxm->full_len - tls_ctx->rx.overhead_size;
1298
1299         if (*zc && (out_iov || out_sg)) {
1300                 if (out_iov)
1301                         n_sgout = iov_iter_npages(out_iov, INT_MAX) + 1;
1302                 else
1303                         n_sgout = sg_nents(out_sg);
1304                 n_sgin = skb_nsg(skb, rxm->offset + tls_ctx->rx.prepend_size,
1305                                  rxm->full_len - tls_ctx->rx.prepend_size);
1306         } else {
1307                 n_sgout = 0;
1308                 *zc = false;
1309                 n_sgin = skb_cow_data(skb, 0, &unused);
1310         }
1311
1312         if (n_sgin < 1)
1313                 return -EBADMSG;
1314
1315         /* Increment to accommodate AAD */
1316         n_sgin = n_sgin + 1;
1317
1318         nsg = n_sgin + n_sgout;
1319
1320         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(ctx->aead_recv);
1321         mem_size = aead_size + (nsg * sizeof(struct scatterlist));
1322         mem_size = mem_size + TLS_AAD_SPACE_SIZE;
1323         mem_size = mem_size + crypto_aead_ivsize(ctx->aead_recv);
1324
1325         /* Allocate a single block of memory which contains
1326          * aead_req || sgin[] || sgout[] || aad || iv.
1327          * This order achieves correct alignment for aead_req, sgin, sgout.
1328          */
1329         mem = kmalloc(mem_size, sk->sk_allocation);
1330         if (!mem)
1331                 return -ENOMEM;
1332
1333         /* Segment the allocated memory */
1334         aead_req = (struct aead_request *)mem;
1335         sgin = (struct scatterlist *)(mem + aead_size);
1336         sgout = sgin + n_sgin;
1337         aad = (u8 *)(sgout + n_sgout);
1338         iv = aad + TLS_AAD_SPACE_SIZE;
1339
1340         /* Prepare IV */
1341         err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
1342                             iv + TLS_CIPHER_AES_GCM_128_SALT_SIZE,
1343                             tls_ctx->rx.iv_size);
1344         if (err < 0) {
1345                 kfree(mem);
1346                 return err;
1347         }
1348         memcpy(iv, tls_ctx->rx.iv, TLS_CIPHER_AES_GCM_128_SALT_SIZE);
1349
1350         /* Prepare AAD */
1351         tls_make_aad(aad, rxm->full_len - tls_ctx->rx.overhead_size,
1352                      tls_ctx->rx.rec_seq, tls_ctx->rx.rec_seq_size,
1353                      ctx->control);
1354
1355         /* Prepare sgin */
1356         sg_init_table(sgin, n_sgin);
1357         sg_set_buf(&sgin[0], aad, TLS_AAD_SPACE_SIZE);
1358         err = skb_to_sgvec(skb, &sgin[1],
1359                            rxm->offset + tls_ctx->rx.prepend_size,
1360                            rxm->full_len - tls_ctx->rx.prepend_size);
1361         if (err < 0) {
1362                 kfree(mem);
1363                 return err;
1364         }
1365
1366         if (n_sgout) {
1367                 if (out_iov) {
1368                         sg_init_table(sgout, n_sgout);
1369                         sg_set_buf(&sgout[0], aad, TLS_AAD_SPACE_SIZE);
1370
1371                         *chunk = 0;
1372                         err = tls_setup_from_iter(sk, out_iov, data_len,
1373                                                   &pages, chunk, &sgout[1],
1374                                                   (n_sgout - 1));
1375                         if (err < 0)
1376                                 goto fallback_to_reg_recv;
1377                 } else if (out_sg) {
1378                         memcpy(sgout, out_sg, n_sgout * sizeof(*sgout));
1379                 } else {
1380                         goto fallback_to_reg_recv;
1381                 }
1382         } else {
1383 fallback_to_reg_recv:
1384                 sgout = sgin;
1385                 pages = 0;
1386                 *chunk = 0;
1387                 *zc = false;
1388         }
1389
1390         /* Prepare and submit AEAD request */
1391         err = tls_do_decryption(sk, skb, sgin, sgout, iv,
1392                                 data_len, aead_req, *zc);
1393         if (err == -EINPROGRESS)
1394                 return err;
1395
1396         /* Release the pages in case iov was mapped to pages */
1397         for (; pages > 0; pages--)
1398                 put_page(sg_page(&sgout[pages]));
1399
1400         kfree(mem);
1401         return err;
1402 }
1403
1404 static int decrypt_skb_update(struct sock *sk, struct sk_buff *skb,
1405                               struct iov_iter *dest, int *chunk, bool *zc)
1406 {
1407         struct tls_context *tls_ctx = tls_get_ctx(sk);
1408         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1409         struct strp_msg *rxm = strp_msg(skb);
1410         int err = 0;
1411
1412 #ifdef CONFIG_TLS_DEVICE
1413         err = tls_device_decrypted(sk, skb);
1414         if (err < 0)
1415                 return err;
1416 #endif
1417         if (!ctx->decrypted) {
1418                 err = decrypt_internal(sk, skb, dest, NULL, chunk, zc);
1419                 if (err < 0) {
1420                         if (err == -EINPROGRESS)
1421                                 tls_advance_record_sn(sk, &tls_ctx->rx);
1422
1423                         return err;
1424                 }
1425         } else {
1426                 *zc = false;
1427         }
1428
1429         rxm->offset += tls_ctx->rx.prepend_size;
1430         rxm->full_len -= tls_ctx->rx.overhead_size;
1431         tls_advance_record_sn(sk, &tls_ctx->rx);
1432         ctx->decrypted = true;
1433         ctx->saved_data_ready(sk);
1434
1435         return err;
1436 }
1437
1438 int decrypt_skb(struct sock *sk, struct sk_buff *skb,
1439                 struct scatterlist *sgout)
1440 {
1441         bool zc = true;
1442         int chunk;
1443
1444         return decrypt_internal(sk, skb, NULL, sgout, &chunk, &zc);
1445 }
1446
1447 static bool tls_sw_advance_skb(struct sock *sk, struct sk_buff *skb,
1448                                unsigned int len)
1449 {
1450         struct tls_context *tls_ctx = tls_get_ctx(sk);
1451         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1452
1453         if (skb) {
1454                 struct strp_msg *rxm = strp_msg(skb);
1455
1456                 if (len < rxm->full_len) {
1457                         rxm->offset += len;
1458                         rxm->full_len -= len;
1459                         return false;
1460                 }
1461                 kfree_skb(skb);
1462         }
1463
1464         /* Finished with message */
1465         ctx->recv_pkt = NULL;
1466         __strp_unpause(&ctx->strp);
1467
1468         return true;
1469 }
1470
1471 int tls_sw_recvmsg(struct sock *sk,
1472                    struct msghdr *msg,
1473                    size_t len,
1474                    int nonblock,
1475                    int flags,
1476                    int *addr_len)
1477 {
1478         struct tls_context *tls_ctx = tls_get_ctx(sk);
1479         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1480         struct sk_psock *psock;
1481         unsigned char control;
1482         struct strp_msg *rxm;
1483         struct sk_buff *skb;
1484         ssize_t copied = 0;
1485         bool cmsg = false;
1486         int target, err = 0;
1487         long timeo;
1488         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
1489         int num_async = 0;
1490
1491         flags |= nonblock;
1492
1493         if (unlikely(flags & MSG_ERRQUEUE))
1494                 return sock_recv_errqueue(sk, msg, len, SOL_IP, IP_RECVERR);
1495
1496         psock = sk_psock_get(sk);
1497         lock_sock(sk);
1498
1499         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1500         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1501         do {
1502                 bool zc = false;
1503                 bool async = false;
1504                 int chunk = 0;
1505
1506                 skb = tls_wait_data(sk, psock, flags, timeo, &err);
1507                 if (!skb) {
1508                         if (psock) {
1509                                 int ret = __tcp_bpf_recvmsg(sk, psock,
1510                                                             msg, len, flags);
1511
1512                                 if (ret > 0) {
1513                                         copied += ret;
1514                                         len -= ret;
1515                                         continue;
1516                                 }
1517                         }
1518                         goto recv_end;
1519                 }
1520
1521                 rxm = strp_msg(skb);
1522
1523                 if (!cmsg) {
1524                         int cerr;
1525
1526                         cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1527                                         sizeof(ctx->control), &ctx->control);
1528                         cmsg = true;
1529                         control = ctx->control;
1530                         if (ctx->control != TLS_RECORD_TYPE_DATA) {
1531                                 if (cerr || msg->msg_flags & MSG_CTRUNC) {
1532                                         err = -EIO;
1533                                         goto recv_end;
1534                                 }
1535                         }
1536                 } else if (control != ctx->control) {
1537                         goto recv_end;
1538                 }
1539
1540                 if (!ctx->decrypted) {
1541                         int to_copy = rxm->full_len - tls_ctx->rx.overhead_size;
1542
1543                         if (!is_kvec && to_copy <= len &&
1544                             likely(!(flags & MSG_PEEK)))
1545                                 zc = true;
1546
1547                         err = decrypt_skb_update(sk, skb, &msg->msg_iter,
1548                                                  &chunk, &zc);
1549                         if (err < 0 && err != -EINPROGRESS) {
1550                                 tls_err_abort(sk, EBADMSG);
1551                                 goto recv_end;
1552                         }
1553
1554                         if (err == -EINPROGRESS) {
1555                                 async = true;
1556                                 num_async++;
1557                                 goto pick_next_record;
1558                         }
1559
1560                         ctx->decrypted = true;
1561                 }
1562
1563                 if (!zc) {
1564                         chunk = min_t(unsigned int, rxm->full_len, len);
1565
1566                         err = skb_copy_datagram_msg(skb, rxm->offset, msg,
1567                                                     chunk);
1568                         if (err < 0)
1569                                 goto recv_end;
1570                 }
1571
1572 pick_next_record:
1573                 copied += chunk;
1574                 len -= chunk;
1575                 if (likely(!(flags & MSG_PEEK))) {
1576                         u8 control = ctx->control;
1577
1578                         /* For async, drop current skb reference */
1579                         if (async)
1580                                 skb = NULL;
1581
1582                         if (tls_sw_advance_skb(sk, skb, chunk)) {
1583                                 /* Return full control message to
1584                                  * userspace before trying to parse
1585                                  * another message type
1586                                  */
1587                                 msg->msg_flags |= MSG_EOR;
1588                                 if (control != TLS_RECORD_TYPE_DATA)
1589                                         goto recv_end;
1590                         } else {
1591                                 break;
1592                         }
1593                 } else {
1594                         /* MSG_PEEK right now cannot look beyond current skb
1595                          * from strparser, meaning we cannot advance skb here
1596                          * and thus unpause strparser since we'd loose original
1597                          * one.
1598                          */
1599                         break;
1600                 }
1601
1602                 /* If we have a new message from strparser, continue now. */
1603                 if (copied >= target && !ctx->recv_pkt)
1604                         break;
1605         } while (len);
1606
1607 recv_end:
1608         if (num_async) {
1609                 /* Wait for all previously submitted records to be decrypted */
1610                 smp_store_mb(ctx->async_notify, true);
1611                 if (atomic_read(&ctx->decrypt_pending)) {
1612                         err = crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1613                         if (err) {
1614                                 /* one of async decrypt failed */
1615                                 tls_err_abort(sk, err);
1616                                 copied = 0;
1617                         }
1618                 } else {
1619                         reinit_completion(&ctx->async_wait.completion);
1620                 }
1621                 WRITE_ONCE(ctx->async_notify, false);
1622         }
1623
1624         release_sock(sk);
1625         if (psock)
1626                 sk_psock_put(sk, psock);
1627         return copied ? : err;
1628 }
1629
1630 ssize_t tls_sw_splice_read(struct socket *sock,  loff_t *ppos,
1631                            struct pipe_inode_info *pipe,
1632                            size_t len, unsigned int flags)
1633 {
1634         struct tls_context *tls_ctx = tls_get_ctx(sock->sk);
1635         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1636         struct strp_msg *rxm = NULL;
1637         struct sock *sk = sock->sk;
1638         struct sk_buff *skb;
1639         ssize_t copied = 0;
1640         int err = 0;
1641         long timeo;
1642         int chunk;
1643         bool zc = false;
1644
1645         lock_sock(sk);
1646
1647         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1648
1649         skb = tls_wait_data(sk, NULL, flags, timeo, &err);
1650         if (!skb)
1651                 goto splice_read_end;
1652
1653         /* splice does not support reading control messages */
1654         if (ctx->control != TLS_RECORD_TYPE_DATA) {
1655                 err = -ENOTSUPP;
1656                 goto splice_read_end;
1657         }
1658
1659         if (!ctx->decrypted) {
1660                 err = decrypt_skb_update(sk, skb, NULL, &chunk, &zc);
1661
1662                 if (err < 0) {
1663                         tls_err_abort(sk, EBADMSG);
1664                         goto splice_read_end;
1665                 }
1666                 ctx->decrypted = true;
1667         }
1668         rxm = strp_msg(skb);
1669
1670         chunk = min_t(unsigned int, rxm->full_len, len);
1671         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, chunk, flags);
1672         if (copied < 0)
1673                 goto splice_read_end;
1674
1675         if (likely(!(flags & MSG_PEEK)))
1676                 tls_sw_advance_skb(sk, skb, copied);
1677
1678 splice_read_end:
1679         release_sock(sk);
1680         return copied ? : err;
1681 }
1682
1683 bool tls_sw_stream_read(const struct sock *sk)
1684 {
1685         struct tls_context *tls_ctx = tls_get_ctx(sk);
1686         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1687         bool ingress_empty = true;
1688         struct sk_psock *psock;
1689
1690         rcu_read_lock();
1691         psock = sk_psock(sk);
1692         if (psock)
1693                 ingress_empty = list_empty(&psock->ingress_msg);
1694         rcu_read_unlock();
1695
1696         return !ingress_empty || ctx->recv_pkt;
1697 }
1698
1699 static int tls_read_size(struct strparser *strp, struct sk_buff *skb)
1700 {
1701         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
1702         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1703         char header[TLS_HEADER_SIZE + MAX_IV_SIZE];
1704         struct strp_msg *rxm = strp_msg(skb);
1705         size_t cipher_overhead;
1706         size_t data_len = 0;
1707         int ret;
1708
1709         /* Verify that we have a full TLS header, or wait for more data */
1710         if (rxm->offset + tls_ctx->rx.prepend_size > skb->len)
1711                 return 0;
1712
1713         /* Sanity-check size of on-stack buffer. */
1714         if (WARN_ON(tls_ctx->rx.prepend_size > sizeof(header))) {
1715                 ret = -EINVAL;
1716                 goto read_failure;
1717         }
1718
1719         /* Linearize header to local buffer */
1720         ret = skb_copy_bits(skb, rxm->offset, header, tls_ctx->rx.prepend_size);
1721
1722         if (ret < 0)
1723                 goto read_failure;
1724
1725         ctx->control = header[0];
1726
1727         data_len = ((header[4] & 0xFF) | (header[3] << 8));
1728
1729         cipher_overhead = tls_ctx->rx.tag_size + tls_ctx->rx.iv_size;
1730
1731         if (data_len > TLS_MAX_PAYLOAD_SIZE + cipher_overhead) {
1732                 ret = -EMSGSIZE;
1733                 goto read_failure;
1734         }
1735         if (data_len < cipher_overhead) {
1736                 ret = -EBADMSG;
1737                 goto read_failure;
1738         }
1739
1740         if (header[1] != TLS_VERSION_MINOR(tls_ctx->crypto_recv.info.version) ||
1741             header[2] != TLS_VERSION_MAJOR(tls_ctx->crypto_recv.info.version)) {
1742                 ret = -EINVAL;
1743                 goto read_failure;
1744         }
1745
1746 #ifdef CONFIG_TLS_DEVICE
1747         handle_device_resync(strp->sk, TCP_SKB_CB(skb)->seq + rxm->offset,
1748                              *(u64*)tls_ctx->rx.rec_seq);
1749 #endif
1750         return data_len + TLS_HEADER_SIZE;
1751
1752 read_failure:
1753         tls_err_abort(strp->sk, ret);
1754
1755         return ret;
1756 }
1757
1758 static void tls_queue(struct strparser *strp, struct sk_buff *skb)
1759 {
1760         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
1761         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1762
1763         ctx->decrypted = false;
1764
1765         ctx->recv_pkt = skb;
1766         strp_pause(strp);
1767
1768         ctx->saved_data_ready(strp->sk);
1769 }
1770
1771 static void tls_data_ready(struct sock *sk)
1772 {
1773         struct tls_context *tls_ctx = tls_get_ctx(sk);
1774         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1775         struct sk_psock *psock;
1776
1777         strp_data_ready(&ctx->strp);
1778
1779         psock = sk_psock_get(sk);
1780         if (psock && !list_empty(&psock->ingress_msg)) {
1781                 ctx->saved_data_ready(sk);
1782                 sk_psock_put(sk, psock);
1783         }
1784 }
1785
1786 void tls_sw_free_resources_tx(struct sock *sk)
1787 {
1788         struct tls_context *tls_ctx = tls_get_ctx(sk);
1789         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1790         struct tls_rec *rec, *tmp;
1791
1792         /* Wait for any pending async encryptions to complete */
1793         smp_store_mb(ctx->async_notify, true);
1794         if (atomic_read(&ctx->encrypt_pending))
1795                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1796
1797         release_sock(sk);
1798         cancel_delayed_work_sync(&ctx->tx_work.work);
1799         lock_sock(sk);
1800
1801         /* Tx whatever records we can transmit and abandon the rest */
1802         tls_tx_records(sk, -1);
1803
1804         /* Free up un-sent records in tx_list. First, free
1805          * the partially sent record if any at head of tx_list.
1806          */
1807         if (tls_ctx->partially_sent_record) {
1808                 struct scatterlist *sg = tls_ctx->partially_sent_record;
1809
1810                 while (1) {
1811                         put_page(sg_page(sg));
1812                         sk_mem_uncharge(sk, sg->length);
1813
1814                         if (sg_is_last(sg))
1815                                 break;
1816                         sg++;
1817                 }
1818
1819                 tls_ctx->partially_sent_record = NULL;
1820
1821                 rec = list_first_entry(&ctx->tx_list,
1822                                        struct tls_rec, list);
1823                 list_del(&rec->list);
1824                 sk_msg_free(sk, &rec->msg_plaintext);
1825                 kfree(rec);
1826         }
1827
1828         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
1829                 list_del(&rec->list);
1830                 sk_msg_free(sk, &rec->msg_encrypted);
1831                 sk_msg_free(sk, &rec->msg_plaintext);
1832                 kfree(rec);
1833         }
1834
1835         crypto_free_aead(ctx->aead_send);
1836         tls_free_open_rec(sk);
1837
1838         kfree(ctx);
1839 }
1840
1841 void tls_sw_release_resources_rx(struct sock *sk)
1842 {
1843         struct tls_context *tls_ctx = tls_get_ctx(sk);
1844         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1845
1846         if (ctx->aead_recv) {
1847                 kfree_skb(ctx->recv_pkt);
1848                 ctx->recv_pkt = NULL;
1849                 crypto_free_aead(ctx->aead_recv);
1850                 strp_stop(&ctx->strp);
1851                 write_lock_bh(&sk->sk_callback_lock);
1852                 sk->sk_data_ready = ctx->saved_data_ready;
1853                 write_unlock_bh(&sk->sk_callback_lock);
1854                 release_sock(sk);
1855                 strp_done(&ctx->strp);
1856                 lock_sock(sk);
1857         }
1858 }
1859
1860 void tls_sw_free_resources_rx(struct sock *sk)
1861 {
1862         struct tls_context *tls_ctx = tls_get_ctx(sk);
1863         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1864
1865         tls_sw_release_resources_rx(sk);
1866
1867         kfree(ctx);
1868 }
1869
1870 /* The work handler to transmitt the encrypted records in tx_list */
1871 static void tx_work_handler(struct work_struct *work)
1872 {
1873         struct delayed_work *delayed_work = to_delayed_work(work);
1874         struct tx_work *tx_work = container_of(delayed_work,
1875                                                struct tx_work, work);
1876         struct sock *sk = tx_work->sk;
1877         struct tls_context *tls_ctx = tls_get_ctx(sk);
1878         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1879
1880         if (!test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
1881                 return;
1882
1883         lock_sock(sk);
1884         tls_tx_records(sk, -1);
1885         release_sock(sk);
1886 }
1887
1888 int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
1889 {
1890         struct tls_crypto_info *crypto_info;
1891         struct tls12_crypto_info_aes_gcm_128 *gcm_128_info;
1892         struct tls_sw_context_tx *sw_ctx_tx = NULL;
1893         struct tls_sw_context_rx *sw_ctx_rx = NULL;
1894         struct cipher_context *cctx;
1895         struct crypto_aead **aead;
1896         struct strp_callbacks cb;
1897         u16 nonce_size, tag_size, iv_size, rec_seq_size;
1898         char *iv, *rec_seq;
1899         int rc = 0;
1900
1901         if (!ctx) {
1902                 rc = -EINVAL;
1903                 goto out;
1904         }
1905
1906         if (tx) {
1907                 if (!ctx->priv_ctx_tx) {
1908                         sw_ctx_tx = kzalloc(sizeof(*sw_ctx_tx), GFP_KERNEL);
1909                         if (!sw_ctx_tx) {
1910                                 rc = -ENOMEM;
1911                                 goto out;
1912                         }
1913                         ctx->priv_ctx_tx = sw_ctx_tx;
1914                 } else {
1915                         sw_ctx_tx =
1916                                 (struct tls_sw_context_tx *)ctx->priv_ctx_tx;
1917                 }
1918         } else {
1919                 if (!ctx->priv_ctx_rx) {
1920                         sw_ctx_rx = kzalloc(sizeof(*sw_ctx_rx), GFP_KERNEL);
1921                         if (!sw_ctx_rx) {
1922                                 rc = -ENOMEM;
1923                                 goto out;
1924                         }
1925                         ctx->priv_ctx_rx = sw_ctx_rx;
1926                 } else {
1927                         sw_ctx_rx =
1928                                 (struct tls_sw_context_rx *)ctx->priv_ctx_rx;
1929                 }
1930         }
1931
1932         if (tx) {
1933                 crypto_init_wait(&sw_ctx_tx->async_wait);
1934                 crypto_info = &ctx->crypto_send.info;
1935                 cctx = &ctx->tx;
1936                 aead = &sw_ctx_tx->aead_send;
1937                 INIT_LIST_HEAD(&sw_ctx_tx->tx_list);
1938                 INIT_DELAYED_WORK(&sw_ctx_tx->tx_work.work, tx_work_handler);
1939                 sw_ctx_tx->tx_work.sk = sk;
1940         } else {
1941                 crypto_init_wait(&sw_ctx_rx->async_wait);
1942                 crypto_info = &ctx->crypto_recv.info;
1943                 cctx = &ctx->rx;
1944                 aead = &sw_ctx_rx->aead_recv;
1945         }
1946
1947         switch (crypto_info->cipher_type) {
1948         case TLS_CIPHER_AES_GCM_128: {
1949                 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
1950                 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
1951                 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
1952                 iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
1953                 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
1954                 rec_seq =
1955                  ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
1956                 gcm_128_info =
1957                         (struct tls12_crypto_info_aes_gcm_128 *)crypto_info;
1958                 break;
1959         }
1960         default:
1961                 rc = -EINVAL;
1962                 goto free_priv;
1963         }
1964
1965         /* Sanity-check the IV size for stack allocations. */
1966         if (iv_size > MAX_IV_SIZE || nonce_size > MAX_IV_SIZE) {
1967                 rc = -EINVAL;
1968                 goto free_priv;
1969         }
1970
1971         cctx->prepend_size = TLS_HEADER_SIZE + nonce_size;
1972         cctx->tag_size = tag_size;
1973         cctx->overhead_size = cctx->prepend_size + cctx->tag_size;
1974         cctx->iv_size = iv_size;
1975         cctx->iv = kmalloc(iv_size + TLS_CIPHER_AES_GCM_128_SALT_SIZE,
1976                            GFP_KERNEL);
1977         if (!cctx->iv) {
1978                 rc = -ENOMEM;
1979                 goto free_priv;
1980         }
1981         memcpy(cctx->iv, gcm_128_info->salt, TLS_CIPHER_AES_GCM_128_SALT_SIZE);
1982         memcpy(cctx->iv + TLS_CIPHER_AES_GCM_128_SALT_SIZE, iv, iv_size);
1983         cctx->rec_seq_size = rec_seq_size;
1984         cctx->rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
1985         if (!cctx->rec_seq) {
1986                 rc = -ENOMEM;
1987                 goto free_iv;
1988         }
1989
1990         if (!*aead) {
1991                 *aead = crypto_alloc_aead("gcm(aes)", 0, 0);
1992                 if (IS_ERR(*aead)) {
1993                         rc = PTR_ERR(*aead);
1994                         *aead = NULL;
1995                         goto free_rec_seq;
1996                 }
1997         }
1998
1999         ctx->push_pending_record = tls_sw_push_pending_record;
2000
2001         rc = crypto_aead_setkey(*aead, gcm_128_info->key,
2002                                 TLS_CIPHER_AES_GCM_128_KEY_SIZE);
2003         if (rc)
2004                 goto free_aead;
2005
2006         rc = crypto_aead_setauthsize(*aead, cctx->tag_size);
2007         if (rc)
2008                 goto free_aead;
2009
2010         if (sw_ctx_rx) {
2011                 /* Set up strparser */
2012                 memset(&cb, 0, sizeof(cb));
2013                 cb.rcv_msg = tls_queue;
2014                 cb.parse_msg = tls_read_size;
2015
2016                 strp_init(&sw_ctx_rx->strp, sk, &cb);
2017
2018                 write_lock_bh(&sk->sk_callback_lock);
2019                 sw_ctx_rx->saved_data_ready = sk->sk_data_ready;
2020                 sk->sk_data_ready = tls_data_ready;
2021                 write_unlock_bh(&sk->sk_callback_lock);
2022
2023                 strp_check_rcv(&sw_ctx_rx->strp);
2024         }
2025
2026         goto out;
2027
2028 free_aead:
2029         crypto_free_aead(*aead);
2030         *aead = NULL;
2031 free_rec_seq:
2032         kfree(cctx->rec_seq);
2033         cctx->rec_seq = NULL;
2034 free_iv:
2035         kfree(cctx->iv);
2036         cctx->iv = NULL;
2037 free_priv:
2038         if (tx) {
2039                 kfree(ctx->priv_ctx_tx);
2040                 ctx->priv_ctx_tx = NULL;
2041         } else {
2042                 kfree(ctx->priv_ctx_rx);
2043                 ctx->priv_ctx_rx = NULL;
2044         }
2045 out:
2046         return rc;
2047 }