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/bug.h>
  39 #include <linux/sched/signal.h>
  40 #include <linux/module.h>
  41 #include <linux/kernel.h>
  42 #include <linux/splice.h>
  43 #include <crypto/aead.h>
  44
  45 #include <net/strparser.h>
  46 #include <net/tls.h>
  47 #include <trace/events/sock.h>
  48
  49 #include "tls.h"
  50
  51 struct tls_decrypt_arg {
  52         struct_group(inargs,
  53         bool zc;
  54         bool async;
  55         u8 tail;
  56         );
  57
  58         struct sk_buff *skb;
  59 };
  60
  61 struct tls_decrypt_ctx {
  62         struct sock *sk;
  63         u8 iv[MAX_IV_SIZE];
  64         u8 aad[TLS_MAX_AAD_SIZE];
  65         u8 tail;
  66         struct scatterlist sg[];
  67 };
  68
  69 noinline void tls_err_abort(struct sock *sk, int err)
  70 {
  71         WARN_ON_ONCE(err >= 0);
  72         /* sk->sk_err should contain a positive error code. */
  73         WRITE_ONCE(sk->sk_err, -err);
  74         /* Paired with smp_rmb() in tcp_poll() */
  75         smp_wmb();
  76         sk_error_report(sk);
  77 }
  78
  79 static int __skb_nsg(struct sk_buff *skb, int offset, int len,
  80                      unsigned int recursion_level)
  81 {
  82         int start = skb_headlen(skb);
  83         int i, chunk = start - offset;
  84         struct sk_buff *frag_iter;
  85         int elt = 0;
  86
  87         if (unlikely(recursion_level >= 24))
  88                 return -EMSGSIZE;
  89
  90         if (chunk > 0) {
  91                 if (chunk > len)
  92                         chunk = len;
  93                 elt++;
  94                 len -= chunk;
  95                 if (len == 0)
  96                         return elt;
  97                 offset += chunk;
  98         }
  99
 100         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
 101                 int end;
 102
 103                 WARN_ON(start > offset + len);
 104
 105                 end = start + skb_frag_size(&skb_shinfo(skb)->frags[i]);
 106                 chunk = end - offset;
 107                 if (chunk > 0) {
 108                         if (chunk > len)
 109                                 chunk = len;
 110                         elt++;
 111                         len -= chunk;
 112                         if (len == 0)
 113                                 return elt;
 114                         offset += chunk;
 115                 }
 116                 start = end;
 117         }
 118
 119         if (unlikely(skb_has_frag_list(skb))) {
 120                 skb_walk_frags(skb, frag_iter) {
 121                         int end, ret;
 122
 123                         WARN_ON(start > offset + len);
 124
 125                         end = start + frag_iter->len;
 126                         chunk = end - offset;
 127                         if (chunk > 0) {
 128                                 if (chunk > len)
 129                                         chunk = len;
 130                                 ret = __skb_nsg(frag_iter, offset - start, chunk,
 131                                                 recursion_level + 1);
 132                                 if (unlikely(ret < 0))
 133                                         return ret;
 134                                 elt += ret;
 135                                 len -= chunk;
 136                                 if (len == 0)
 137                                         return elt;
 138                                 offset += chunk;
 139                         }
 140                         start = end;
 141                 }
 142         }
 143         BUG_ON(len);
 144         return elt;
 145 }
 146
 147 /* Return the number of scatterlist elements required to completely map the
 148  * skb, or -EMSGSIZE if the recursion depth is exceeded.
 149  */
 150 static int skb_nsg(struct sk_buff *skb, int offset, int len)
 151 {
 152         return __skb_nsg(skb, offset, len, 0);
 153 }
 154
 155 static int tls_padding_length(struct tls_prot_info *prot, struct sk_buff *skb,
 156                               struct tls_decrypt_arg *darg)
 157 {
 158         struct strp_msg *rxm = strp_msg(skb);
 159         struct tls_msg *tlm = tls_msg(skb);
 160         int sub = 0;
 161
 162         /* Determine zero-padding length */
 163         if (prot->version == TLS_1_3_VERSION) {
 164                 int offset = rxm->full_len - TLS_TAG_SIZE - 1;
 165                 char content_type = darg->zc ? darg->tail : 0;
 166                 int err;
 167
 168                 while (content_type == 0) {
 169                         if (offset < prot->prepend_size)
 170                                 return -EBADMSG;
 171                         err = skb_copy_bits(skb, rxm->offset + offset,
 172                                             &content_type, 1);
 173                         if (err)
 174                                 return err;
 175                         if (content_type)
 176                                 break;
 177                         sub++;
 178                         offset--;
 179                 }
 180                 tlm->control = content_type;
 181         }
 182         return sub;
 183 }
 184
 185 static void tls_decrypt_done(void *data, int err)
 186 {
 187         struct aead_request *aead_req = data;
 188         struct crypto_aead *aead = crypto_aead_reqtfm(aead_req);
 189         struct scatterlist *sgout = aead_req->dst;
 190         struct scatterlist *sgin = aead_req->src;
 191         struct tls_sw_context_rx *ctx;
 192         struct tls_decrypt_ctx *dctx;
 193         struct tls_context *tls_ctx;
 194         struct scatterlist *sg;
 195         unsigned int pages;
 196         struct sock *sk;
 197         int aead_size;
 198
 199         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(aead);
 200         aead_size = ALIGN(aead_size, __alignof__(*dctx));
 201         dctx = (void *)((u8 *)aead_req + aead_size);
 202
 203         sk = dctx->sk;
 204         tls_ctx = tls_get_ctx(sk);
 205         ctx = tls_sw_ctx_rx(tls_ctx);
 206
 207         /* Propagate if there was an err */
 208         if (err) {
 209                 if (err == -EBADMSG)
 210                         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSDECRYPTERROR);
 211                 ctx->async_wait.err = err;
 212                 tls_err_abort(sk, err);
 213         }
 214
 215         /* Free the destination pages if skb was not decrypted inplace */
 216         if (sgout != sgin) {
 217                 /* Skip the first S/G entry as it points to AAD */
 218                 for_each_sg(sg_next(sgout), sg, UINT_MAX, pages) {
 219                         if (!sg)
 220                                 break;
 221                         put_page(sg_page(sg));
 222                 }
 223         }
 224
 225         kfree(aead_req);
 226
 227         spin_lock_bh(&ctx->decrypt_compl_lock);
 228         if (!atomic_dec_return(&ctx->decrypt_pending))
 229                 complete(&ctx->async_wait.completion);
 230         spin_unlock_bh(&ctx->decrypt_compl_lock);
 231 }
 232
 233 static int tls_do_decryption(struct sock *sk,
 234                              struct scatterlist *sgin,
 235                              struct scatterlist *sgout,
 236                              char *iv_recv,
 237                              size_t data_len,
 238                              struct aead_request *aead_req,
 239                              struct tls_decrypt_arg *darg)
 240 {
 241         struct tls_context *tls_ctx = tls_get_ctx(sk);
 242         struct tls_prot_info *prot = &tls_ctx->prot_info;
 243         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
 244         int ret;
 245
 246         aead_request_set_tfm(aead_req, ctx->aead_recv);
 247         aead_request_set_ad(aead_req, prot->aad_size);
 248         aead_request_set_crypt(aead_req, sgin, sgout,
 249                                data_len + prot->tag_size,
 250                                (u8 *)iv_recv);
 251
 252         if (darg->async) {
 253                 aead_request_set_callback(aead_req,
 254                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 255                                           tls_decrypt_done, aead_req);
 256                 atomic_inc(&ctx->decrypt_pending);
 257         } else {
 258                 aead_request_set_callback(aead_req,
 259                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 260                                           crypto_req_done, &ctx->async_wait);
 261         }
 262
 263         ret = crypto_aead_decrypt(aead_req);
 264         if (ret == -EINPROGRESS) {
 265                 if (darg->async)
 266                         return 0;
 267
 268                 ret = crypto_wait_req(ret, &ctx->async_wait);
 269         }
 270         darg->async = false;
 271
 272         return ret;
 273 }
 274
 275 static void tls_trim_both_msgs(struct sock *sk, int target_size)
 276 {
 277         struct tls_context *tls_ctx = tls_get_ctx(sk);
 278         struct tls_prot_info *prot = &tls_ctx->prot_info;
 279         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 280         struct tls_rec *rec = ctx->open_rec;
 281
 282         sk_msg_trim(sk, &rec->msg_plaintext, target_size);
 283         if (target_size > 0)
 284                 target_size += prot->overhead_size;
 285         sk_msg_trim(sk, &rec->msg_encrypted, target_size);
 286 }
 287
 288 static int tls_alloc_encrypted_msg(struct sock *sk, int len)
 289 {
 290         struct tls_context *tls_ctx = tls_get_ctx(sk);
 291         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 292         struct tls_rec *rec = ctx->open_rec;
 293         struct sk_msg *msg_en = &rec->msg_encrypted;
 294
 295         return sk_msg_alloc(sk, msg_en, len, 0);
 296 }
 297
 298 static int tls_clone_plaintext_msg(struct sock *sk, int required)
 299 {
 300         struct tls_context *tls_ctx = tls_get_ctx(sk);
 301         struct tls_prot_info *prot = &tls_ctx->prot_info;
 302         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 303         struct tls_rec *rec = ctx->open_rec;
 304         struct sk_msg *msg_pl = &rec->msg_plaintext;
 305         struct sk_msg *msg_en = &rec->msg_encrypted;
 306         int skip, len;
 307
 308         /* We add page references worth len bytes from encrypted sg
 309          * at the end of plaintext sg. It is guaranteed that msg_en
 310          * has enough required room (ensured by caller).
 311          */
 312         len = required - msg_pl->sg.size;
 313
 314         /* Skip initial bytes in msg_en's data to be able to use
 315          * same offset of both plain and encrypted data.
 316          */
 317         skip = prot->prepend_size + msg_pl->sg.size;
 318
 319         return sk_msg_clone(sk, msg_pl, msg_en, skip, len);
 320 }
 321
 322 static struct tls_rec *tls_get_rec(struct sock *sk)
 323 {
 324         struct tls_context *tls_ctx = tls_get_ctx(sk);
 325         struct tls_prot_info *prot = &tls_ctx->prot_info;
 326         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 327         struct sk_msg *msg_pl, *msg_en;
 328         struct tls_rec *rec;
 329         int mem_size;
 330
 331         mem_size = sizeof(struct tls_rec) + crypto_aead_reqsize(ctx->aead_send);
 332
 333         rec = kzalloc(mem_size, sk->sk_allocation);
 334         if (!rec)
 335                 return NULL;
 336
 337         msg_pl = &rec->msg_plaintext;
 338         msg_en = &rec->msg_encrypted;
 339
 340         sk_msg_init(msg_pl);
 341         sk_msg_init(msg_en);
 342
 343         sg_init_table(rec->sg_aead_in, 2);
 344         sg_set_buf(&rec->sg_aead_in[0], rec->aad_space, prot->aad_size);
 345         sg_unmark_end(&rec->sg_aead_in[1]);
 346
 347         sg_init_table(rec->sg_aead_out, 2);
 348         sg_set_buf(&rec->sg_aead_out[0], rec->aad_space, prot->aad_size);
 349         sg_unmark_end(&rec->sg_aead_out[1]);
 350
 351         rec->sk = sk;
 352
 353         return rec;
 354 }
 355
 356 static void tls_free_rec(struct sock *sk, struct tls_rec *rec)
 357 {
 358         sk_msg_free(sk, &rec->msg_encrypted);
 359         sk_msg_free(sk, &rec->msg_plaintext);
 360         kfree(rec);
 361 }
 362
 363 static void tls_free_open_rec(struct sock *sk)
 364 {
 365         struct tls_context *tls_ctx = tls_get_ctx(sk);
 366         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 367         struct tls_rec *rec = ctx->open_rec;
 368
 369         if (rec) {
 370                 tls_free_rec(sk, rec);
 371                 ctx->open_rec = NULL;
 372         }
 373 }
 374
 375 int tls_tx_records(struct sock *sk, int flags)
 376 {
 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 tls_rec *rec, *tmp;
 380         struct sk_msg *msg_en;
 381         int tx_flags, rc = 0;
 382
 383         if (tls_is_partially_sent_record(tls_ctx)) {
 384                 rec = list_first_entry(&ctx->tx_list,
 385                                        struct tls_rec, list);
 386
 387                 if (flags == -1)
 388                         tx_flags = rec->tx_flags;
 389                 else
 390                         tx_flags = flags;
 391
 392                 rc = tls_push_partial_record(sk, tls_ctx, tx_flags);
 393                 if (rc)
 394                         goto tx_err;
 395
 396                 /* Full record has been transmitted.
 397                  * Remove the head of tx_list
 398                  */
 399                 list_del(&rec->list);
 400                 sk_msg_free(sk, &rec->msg_plaintext);
 401                 kfree(rec);
 402         }
 403
 404         /* Tx all ready records */
 405         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
 406                 if (READ_ONCE(rec->tx_ready)) {
 407                         if (flags == -1)
 408                                 tx_flags = rec->tx_flags;
 409                         else
 410                                 tx_flags = flags;
 411
 412                         msg_en = &rec->msg_encrypted;
 413                         rc = tls_push_sg(sk, tls_ctx,
 414                                          &msg_en->sg.data[msg_en->sg.curr],
 415                                          0, tx_flags);
 416                         if (rc)
 417                                 goto tx_err;
 418
 419                         list_del(&rec->list);
 420                         sk_msg_free(sk, &rec->msg_plaintext);
 421                         kfree(rec);
 422                 } else {
 423                         break;
 424                 }
 425         }
 426
 427 tx_err:
 428         if (rc < 0 && rc != -EAGAIN)
 429                 tls_err_abort(sk, -EBADMSG);
 430
 431         return rc;
 432 }
 433
 434 static void tls_encrypt_done(void *data, int err)
 435 {
 436         struct tls_sw_context_tx *ctx;
 437         struct tls_context *tls_ctx;
 438         struct tls_prot_info *prot;
 439         struct tls_rec *rec = data;
 440         struct scatterlist *sge;
 441         struct sk_msg *msg_en;
 442         bool ready = false;
 443         struct sock *sk;
 444         int pending;
 445
 446         msg_en = &rec->msg_encrypted;
 447
 448         sk = rec->sk;
 449         tls_ctx = tls_get_ctx(sk);
 450         prot = &tls_ctx->prot_info;
 451         ctx = tls_sw_ctx_tx(tls_ctx);
 452
 453         sge = sk_msg_elem(msg_en, msg_en->sg.curr);
 454         sge->offset -= prot->prepend_size;
 455         sge->length += prot->prepend_size;
 456
 457         /* Check if error is previously set on socket */
 458         if (err || sk->sk_err) {
 459                 rec = NULL;
 460
 461                 /* If err is already set on socket, return the same code */
 462                 if (sk->sk_err) {
 463                         ctx->async_wait.err = -sk->sk_err;
 464                 } else {
 465                         ctx->async_wait.err = err;
 466                         tls_err_abort(sk, err);
 467                 }
 468         }
 469
 470         if (rec) {
 471                 struct tls_rec *first_rec;
 472
 473                 /* Mark the record as ready for transmission */
 474                 smp_store_mb(rec->tx_ready, true);
 475
 476                 /* If received record is at head of tx_list, schedule tx */
 477                 first_rec = list_first_entry(&ctx->tx_list,
 478                                              struct tls_rec, list);
 479                 if (rec == first_rec)
 480                         ready = true;
 481         }
 482
 483         spin_lock_bh(&ctx->encrypt_compl_lock);
 484         pending = atomic_dec_return(&ctx->encrypt_pending);
 485
 486         if (!pending && ctx->async_notify)
 487                 complete(&ctx->async_wait.completion);
 488         spin_unlock_bh(&ctx->encrypt_compl_lock);
 489
 490         if (!ready)
 491                 return;
 492
 493         /* Schedule the transmission */
 494         if (!test_and_set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
 495                 schedule_delayed_work(&ctx->tx_work.work, 1);
 496 }
 497
 498 static int tls_do_encryption(struct sock *sk,
 499                              struct tls_context *tls_ctx,
 500                              struct tls_sw_context_tx *ctx,
 501                              struct aead_request *aead_req,
 502                              size_t data_len, u32 start)
 503 {
 504         struct tls_prot_info *prot = &tls_ctx->prot_info;
 505         struct tls_rec *rec = ctx->open_rec;
 506         struct sk_msg *msg_en = &rec->msg_encrypted;
 507         struct scatterlist *sge = sk_msg_elem(msg_en, start);
 508         int rc, iv_offset = 0;
 509
 510         /* For CCM based ciphers, first byte of IV is a constant */
 511         switch (prot->cipher_type) {
 512         case TLS_CIPHER_AES_CCM_128:
 513                 rec->iv_data[0] = TLS_AES_CCM_IV_B0_BYTE;
 514                 iv_offset = 1;
 515                 break;
 516         case TLS_CIPHER_SM4_CCM:
 517                 rec->iv_data[0] = TLS_SM4_CCM_IV_B0_BYTE;
 518                 iv_offset = 1;
 519                 break;
 520         }
 521
 522         memcpy(&rec->iv_data[iv_offset], tls_ctx->tx.iv,
 523                prot->iv_size + prot->salt_size);
 524
 525         tls_xor_iv_with_seq(prot, rec->iv_data + iv_offset,
 526                             tls_ctx->tx.rec_seq);
 527
 528         sge->offset += prot->prepend_size;
 529         sge->length -= prot->prepend_size;
 530
 531         msg_en->sg.curr = start;
 532
 533         aead_request_set_tfm(aead_req, ctx->aead_send);
 534         aead_request_set_ad(aead_req, prot->aad_size);
 535         aead_request_set_crypt(aead_req, rec->sg_aead_in,
 536                                rec->sg_aead_out,
 537                                data_len, rec->iv_data);
 538
 539         aead_request_set_callback(aead_req, CRYPTO_TFM_REQ_MAY_BACKLOG,
 540                                   tls_encrypt_done, rec);
 541
 542         /* Add the record in tx_list */
 543         list_add_tail((struct list_head *)&rec->list, &ctx->tx_list);
 544         atomic_inc(&ctx->encrypt_pending);
 545
 546         rc = crypto_aead_encrypt(aead_req);
 547         if (!rc || rc != -EINPROGRESS) {
 548                 atomic_dec(&ctx->encrypt_pending);
 549                 sge->offset -= prot->prepend_size;
 550                 sge->length += prot->prepend_size;
 551         }
 552
 553         if (!rc) {
 554                 WRITE_ONCE(rec->tx_ready, true);
 555         } else if (rc != -EINPROGRESS) {
 556                 list_del(&rec->list);
 557                 return rc;
 558         }
 559
 560         /* Unhook the record from context if encryption is not failure */
 561         ctx->open_rec = NULL;
 562         tls_advance_record_sn(sk, prot, &tls_ctx->tx);
 563         return rc;
 564 }
 565
 566 static int tls_split_open_record(struct sock *sk, struct tls_rec *from,
 567                                  struct tls_rec **to, struct sk_msg *msg_opl,
 568                                  struct sk_msg *msg_oen, u32 split_point,
 569                                  u32 tx_overhead_size, u32 *orig_end)
 570 {
 571         u32 i, j, bytes = 0, apply = msg_opl->apply_bytes;
 572         struct scatterlist *sge, *osge, *nsge;
 573         u32 orig_size = msg_opl->sg.size;
 574         struct scatterlist tmp = { };
 575         struct sk_msg *msg_npl;
 576         struct tls_rec *new;
 577         int ret;
 578
 579         new = tls_get_rec(sk);
 580         if (!new)
 581                 return -ENOMEM;
 582         ret = sk_msg_alloc(sk, &new->msg_encrypted, msg_opl->sg.size +
 583                            tx_overhead_size, 0);
 584         if (ret < 0) {
 585                 tls_free_rec(sk, new);
 586                 return ret;
 587         }
 588
 589         *orig_end = msg_opl->sg.end;
 590         i = msg_opl->sg.start;
 591         sge = sk_msg_elem(msg_opl, i);
 592         while (apply && sge->length) {
 593                 if (sge->length > apply) {
 594                         u32 len = sge->length - apply;
 595
 596                         get_page(sg_page(sge));
 597                         sg_set_page(&tmp, sg_page(sge), len,
 598                                     sge->offset + apply);
 599                         sge->length = apply;
 600                         bytes += apply;
 601                         apply = 0;
 602                 } else {
 603                         apply -= sge->length;
 604                         bytes += sge->length;
 605                 }
 606
 607                 sk_msg_iter_var_next(i);
 608                 if (i == msg_opl->sg.end)
 609                         break;
 610                 sge = sk_msg_elem(msg_opl, i);
 611         }
 612
 613         msg_opl->sg.end = i;
 614         msg_opl->sg.curr = i;
 615         msg_opl->sg.copybreak = 0;
 616         msg_opl->apply_bytes = 0;
 617         msg_opl->sg.size = bytes;
 618
 619         msg_npl = &new->msg_plaintext;
 620         msg_npl->apply_bytes = apply;
 621         msg_npl->sg.size = orig_size - bytes;
 622
 623         j = msg_npl->sg.start;
 624         nsge = sk_msg_elem(msg_npl, j);
 625         if (tmp.length) {
 626                 memcpy(nsge, &tmp, sizeof(*nsge));
 627                 sk_msg_iter_var_next(j);
 628                 nsge = sk_msg_elem(msg_npl, j);
 629         }
 630
 631         osge = sk_msg_elem(msg_opl, i);
 632         while (osge->length) {
 633                 memcpy(nsge, osge, sizeof(*nsge));
 634                 sg_unmark_end(nsge);
 635                 sk_msg_iter_var_next(i);
 636                 sk_msg_iter_var_next(j);
 637                 if (i == *orig_end)
 638                         break;
 639                 osge = sk_msg_elem(msg_opl, i);
 640                 nsge = sk_msg_elem(msg_npl, j);
 641         }
 642
 643         msg_npl->sg.end = j;
 644         msg_npl->sg.curr = j;
 645         msg_npl->sg.copybreak = 0;
 646
 647         *to = new;
 648         return 0;
 649 }
 650
 651 static void tls_merge_open_record(struct sock *sk, struct tls_rec *to,
 652                                   struct tls_rec *from, u32 orig_end)
 653 {
 654         struct sk_msg *msg_npl = &from->msg_plaintext;
 655         struct sk_msg *msg_opl = &to->msg_plaintext;
 656         struct scatterlist *osge, *nsge;
 657         u32 i, j;
 658
 659         i = msg_opl->sg.end;
 660         sk_msg_iter_var_prev(i);
 661         j = msg_npl->sg.start;
 662
 663         osge = sk_msg_elem(msg_opl, i);
 664         nsge = sk_msg_elem(msg_npl, j);
 665
 666         if (sg_page(osge) == sg_page(nsge) &&
 667             osge->offset + osge->length == nsge->offset) {
 668                 osge->length += nsge->length;
 669                 put_page(sg_page(nsge));
 670         }
 671
 672         msg_opl->sg.end = orig_end;
 673         msg_opl->sg.curr = orig_end;
 674         msg_opl->sg.copybreak = 0;
 675         msg_opl->apply_bytes = msg_opl->sg.size + msg_npl->sg.size;
 676         msg_opl->sg.size += msg_npl->sg.size;
 677
 678         sk_msg_free(sk, &to->msg_encrypted);
 679         sk_msg_xfer_full(&to->msg_encrypted, &from->msg_encrypted);
 680
 681         kfree(from);
 682 }
 683
 684 static int tls_push_record(struct sock *sk, int flags,
 685                            unsigned char record_type)
 686 {
 687         struct tls_context *tls_ctx = tls_get_ctx(sk);
 688         struct tls_prot_info *prot = &tls_ctx->prot_info;
 689         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 690         struct tls_rec *rec = ctx->open_rec, *tmp = NULL;
 691         u32 i, split_point, orig_end;
 692         struct sk_msg *msg_pl, *msg_en;
 693         struct aead_request *req;
 694         bool split;
 695         int rc;
 696
 697         if (!rec)
 698                 return 0;
 699
 700         msg_pl = &rec->msg_plaintext;
 701         msg_en = &rec->msg_encrypted;
 702
 703         split_point = msg_pl->apply_bytes;
 704         split = split_point && split_point < msg_pl->sg.size;
 705         if (unlikely((!split &&
 706                       msg_pl->sg.size +
 707                       prot->overhead_size > msg_en->sg.size) ||
 708                      (split &&
 709                       split_point +
 710                       prot->overhead_size > msg_en->sg.size))) {
 711                 split = true;
 712                 split_point = msg_en->sg.size;
 713         }
 714         if (split) {
 715                 rc = tls_split_open_record(sk, rec, &tmp, msg_pl, msg_en,
 716                                            split_point, prot->overhead_size,
 717                                            &orig_end);
 718                 if (rc < 0)
 719                         return rc;
 720                 /* This can happen if above tls_split_open_record allocates
 721                  * a single large encryption buffer instead of two smaller
 722                  * ones. In this case adjust pointers and continue without
 723                  * split.
 724                  */
 725                 if (!msg_pl->sg.size) {
 726                         tls_merge_open_record(sk, rec, tmp, orig_end);
 727                         msg_pl = &rec->msg_plaintext;
 728                         msg_en = &rec->msg_encrypted;
 729                         split = false;
 730                 }
 731                 sk_msg_trim(sk, msg_en, msg_pl->sg.size +
 732                             prot->overhead_size);
 733         }
 734
 735         rec->tx_flags = flags;
 736         req = &rec->aead_req;
 737
 738         i = msg_pl->sg.end;
 739         sk_msg_iter_var_prev(i);
 740
 741         rec->content_type = record_type;
 742         if (prot->version == TLS_1_3_VERSION) {
 743                 /* Add content type to end of message.  No padding added */
 744                 sg_set_buf(&rec->sg_content_type, &rec->content_type, 1);
 745                 sg_mark_end(&rec->sg_content_type);
 746                 sg_chain(msg_pl->sg.data, msg_pl->sg.end + 1,
 747                          &rec->sg_content_type);
 748         } else {
 749                 sg_mark_end(sk_msg_elem(msg_pl, i));
 750         }
 751
 752         if (msg_pl->sg.end < msg_pl->sg.start) {
 753                 sg_chain(&msg_pl->sg.data[msg_pl->sg.start],
 754                          MAX_SKB_FRAGS - msg_pl->sg.start + 1,
 755                          msg_pl->sg.data);
 756         }
 757
 758         i = msg_pl->sg.start;
 759         sg_chain(rec->sg_aead_in, 2, &msg_pl->sg.data[i]);
 760
 761         i = msg_en->sg.end;
 762         sk_msg_iter_var_prev(i);
 763         sg_mark_end(sk_msg_elem(msg_en, i));
 764
 765         i = msg_en->sg.start;
 766         sg_chain(rec->sg_aead_out, 2, &msg_en->sg.data[i]);
 767
 768         tls_make_aad(rec->aad_space, msg_pl->sg.size + prot->tail_size,
 769                      tls_ctx->tx.rec_seq, record_type, prot);
 770
 771         tls_fill_prepend(tls_ctx,
 772                          page_address(sg_page(&msg_en->sg.data[i])) +
 773                          msg_en->sg.data[i].offset,
 774                          msg_pl->sg.size + prot->tail_size,
 775                          record_type);
 776
 777         tls_ctx->pending_open_record_frags = false;
 778
 779         rc = tls_do_encryption(sk, tls_ctx, ctx, req,
 780                                msg_pl->sg.size + prot->tail_size, i);
 781         if (rc < 0) {
 782                 if (rc != -EINPROGRESS) {
 783                         tls_err_abort(sk, -EBADMSG);
 784                         if (split) {
 785                                 tls_ctx->pending_open_record_frags = true;
 786                                 tls_merge_open_record(sk, rec, tmp, orig_end);
 787                         }
 788                 }
 789                 ctx->async_capable = 1;
 790                 return rc;
 791         } else if (split) {
 792                 msg_pl = &tmp->msg_plaintext;
 793                 msg_en = &tmp->msg_encrypted;
 794                 sk_msg_trim(sk, msg_en, msg_pl->sg.size + prot->overhead_size);
 795                 tls_ctx->pending_open_record_frags = true;
 796                 ctx->open_rec = tmp;
 797         }
 798
 799         return tls_tx_records(sk, flags);
 800 }
 801
 802 static int bpf_exec_tx_verdict(struct sk_msg *msg, struct sock *sk,
 803                                bool full_record, u8 record_type,
 804                                ssize_t *copied, int flags)
 805 {
 806         struct tls_context *tls_ctx = tls_get_ctx(sk);
 807         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 808         struct sk_msg msg_redir = { };
 809         struct sk_psock *psock;
 810         struct sock *sk_redir;
 811         struct tls_rec *rec;
 812         bool enospc, policy, redir_ingress;
 813         int err = 0, send;
 814         u32 delta = 0;
 815
 816         policy = !(flags & MSG_SENDPAGE_NOPOLICY);
 817         psock = sk_psock_get(sk);
 818         if (!psock || !policy) {
 819                 err = tls_push_record(sk, flags, record_type);
 820                 if (err && err != -EINPROGRESS && sk->sk_err == EBADMSG) {
 821                         *copied -= sk_msg_free(sk, msg);
 822                         tls_free_open_rec(sk);
 823                         err = -sk->sk_err;
 824                 }
 825                 if (psock)
 826                         sk_psock_put(sk, psock);
 827                 return err;
 828         }
 829 more_data:
 830         enospc = sk_msg_full(msg);
 831         if (psock->eval == __SK_NONE) {
 832                 delta = msg->sg.size;
 833                 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
 834                 delta -= msg->sg.size;
 835         }
 836         if (msg->cork_bytes && msg->cork_bytes > msg->sg.size &&
 837             !enospc && !full_record) {
 838                 err = -ENOSPC;
 839                 goto out_err;
 840         }
 841         msg->cork_bytes = 0;
 842         send = msg->sg.size;
 843         if (msg->apply_bytes && msg->apply_bytes < send)
 844                 send = msg->apply_bytes;
 845
 846         switch (psock->eval) {
 847         case __SK_PASS:
 848                 err = tls_push_record(sk, flags, record_type);
 849                 if (err && err != -EINPROGRESS && sk->sk_err == EBADMSG) {
 850                         *copied -= sk_msg_free(sk, msg);
 851                         tls_free_open_rec(sk);
 852                         err = -sk->sk_err;
 853                         goto out_err;
 854                 }
 855                 break;
 856         case __SK_REDIRECT:
 857                 redir_ingress = psock->redir_ingress;
 858                 sk_redir = psock->sk_redir;
 859                 memcpy(&msg_redir, msg, sizeof(*msg));
 860                 if (msg->apply_bytes < send)
 861                         msg->apply_bytes = 0;
 862                 else
 863                         msg->apply_bytes -= send;
 864                 sk_msg_return_zero(sk, msg, send);
 865                 msg->sg.size -= send;
 866                 release_sock(sk);
 867                 err = tcp_bpf_sendmsg_redir(sk_redir, redir_ingress,
 868                                             &msg_redir, send, flags);
 869                 lock_sock(sk);
 870                 if (err < 0) {
 871                         *copied -= sk_msg_free_nocharge(sk, &msg_redir);
 872                         msg->sg.size = 0;
 873                 }
 874                 if (msg->sg.size == 0)
 875                         tls_free_open_rec(sk);
 876                 break;
 877         case __SK_DROP:
 878         default:
 879                 sk_msg_free_partial(sk, msg, send);
 880                 if (msg->apply_bytes < send)
 881                         msg->apply_bytes = 0;
 882                 else
 883                         msg->apply_bytes -= send;
 884                 if (msg->sg.size == 0)
 885                         tls_free_open_rec(sk);
 886                 *copied -= (send + delta);
 887                 err = -EACCES;
 888         }
 889
 890         if (likely(!err)) {
 891                 bool reset_eval = !ctx->open_rec;
 892
 893                 rec = ctx->open_rec;
 894                 if (rec) {
 895                         msg = &rec->msg_plaintext;
 896                         if (!msg->apply_bytes)
 897                                 reset_eval = true;
 898                 }
 899                 if (reset_eval) {
 900                         psock->eval = __SK_NONE;
 901                         if (psock->sk_redir) {
 902                                 sock_put(psock->sk_redir);
 903                                 psock->sk_redir = NULL;
 904                         }
 905                 }
 906                 if (rec)
 907                         goto more_data;
 908         }
 909  out_err:
 910         sk_psock_put(sk, psock);
 911         return err;
 912 }
 913
 914 static int tls_sw_push_pending_record(struct sock *sk, int flags)
 915 {
 916         struct tls_context *tls_ctx = tls_get_ctx(sk);
 917         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 918         struct tls_rec *rec = ctx->open_rec;
 919         struct sk_msg *msg_pl;
 920         size_t copied;
 921
 922         if (!rec)
 923                 return 0;
 924
 925         msg_pl = &rec->msg_plaintext;
 926         copied = msg_pl->sg.size;
 927         if (!copied)
 928                 return 0;
 929
 930         return bpf_exec_tx_verdict(msg_pl, sk, true, TLS_RECORD_TYPE_DATA,
 931                                    &copied, flags);
 932 }
 933
 934 static int tls_sw_sendmsg_splice(struct sock *sk, struct msghdr *msg,
 935                                  struct sk_msg *msg_pl, size_t try_to_copy,
 936                                  ssize_t *copied)
 937 {
 938         struct page *page = NULL, **pages = &page;
 939
 940         do {
 941                 ssize_t part;
 942                 size_t off;
 943
 944                 part = iov_iter_extract_pages(&msg->msg_iter, &pages,
 945                                               try_to_copy, 1, 0, &off);
 946                 if (part <= 0)
 947                         return part ?: -EIO;
 948
 949                 if (WARN_ON_ONCE(!sendpage_ok(page))) {
 950                         iov_iter_revert(&msg->msg_iter, part);
 951                         return -EIO;
 952                 }
 953
 954                 sk_msg_page_add(msg_pl, page, part, off);
 955                 sk_mem_charge(sk, part);
 956                 *copied += part;
 957                 try_to_copy -= part;
 958         } while (try_to_copy && !sk_msg_full(msg_pl));
 959
 960         return 0;
 961 }
 962
 963 static int tls_sw_sendmsg_locked(struct sock *sk, struct msghdr *msg,
 964                                  size_t size)
 965 {
 966         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 967         struct tls_context *tls_ctx = tls_get_ctx(sk);
 968         struct tls_prot_info *prot = &tls_ctx->prot_info;
 969         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 970         bool async_capable = ctx->async_capable;
 971         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 972         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
 973         bool eor = !(msg->msg_flags & MSG_MORE);
 974         size_t try_to_copy;
 975         ssize_t copied = 0;
 976         struct sk_msg *msg_pl, *msg_en;
 977         struct tls_rec *rec;
 978         int required_size;
 979         int num_async = 0;
 980         bool full_record;
 981         int record_room;
 982         int num_zc = 0;
 983         int orig_size;
 984         int ret = 0;
 985         int pending;
 986
 987         if (!eor && (msg->msg_flags & MSG_EOR))
 988                 return -EINVAL;
 989
 990         if (unlikely(msg->msg_controllen)) {
 991                 ret = tls_process_cmsg(sk, msg, &record_type);
 992                 if (ret) {
 993                         if (ret == -EINPROGRESS)
 994                                 num_async++;
 995                         else if (ret != -EAGAIN)
 996                                 goto send_end;
 997                 }
 998         }
 999
1000         while (msg_data_left(msg)) {
1001                 if (sk->sk_err) {
1002                         ret = -sk->sk_err;
1003                         goto send_end;
1004                 }
1005
1006                 if (ctx->open_rec)
1007                         rec = ctx->open_rec;
1008                 else
1009                         rec = ctx->open_rec = tls_get_rec(sk);
1010                 if (!rec) {
1011                         ret = -ENOMEM;
1012                         goto send_end;
1013                 }
1014
1015                 msg_pl = &rec->msg_plaintext;
1016                 msg_en = &rec->msg_encrypted;
1017
1018                 orig_size = msg_pl->sg.size;
1019                 full_record = false;
1020                 try_to_copy = msg_data_left(msg);
1021                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
1022                 if (try_to_copy >= record_room) {
1023                         try_to_copy = record_room;
1024                         full_record = true;
1025                 }
1026
1027                 required_size = msg_pl->sg.size + try_to_copy +
1028                                 prot->overhead_size;
1029
1030                 if (!sk_stream_memory_free(sk))
1031                         goto wait_for_sndbuf;
1032
1033 alloc_encrypted:
1034                 ret = tls_alloc_encrypted_msg(sk, required_size);
1035                 if (ret) {
1036                         if (ret != -ENOSPC)
1037                                 goto wait_for_memory;
1038
1039                         /* Adjust try_to_copy according to the amount that was
1040                          * actually allocated. The difference is due
1041                          * to max sg elements limit
1042                          */
1043                         try_to_copy -= required_size - msg_en->sg.size;
1044                         full_record = true;
1045                 }
1046
1047                 if (try_to_copy && (msg->msg_flags & MSG_SPLICE_PAGES)) {
1048                         ret = tls_sw_sendmsg_splice(sk, msg, msg_pl,
1049                                                     try_to_copy, &copied);
1050                         if (ret < 0)
1051                                 goto send_end;
1052                         tls_ctx->pending_open_record_frags = true;
1053                         if (full_record || eor || sk_msg_full(msg_pl))
1054                                 goto copied;
1055                         continue;
1056                 }
1057
1058                 if (!is_kvec && (full_record || eor) && !async_capable) {
1059                         u32 first = msg_pl->sg.end;
1060
1061                         ret = sk_msg_zerocopy_from_iter(sk, &msg->msg_iter,
1062                                                         msg_pl, try_to_copy);
1063                         if (ret)
1064                                 goto fallback_to_reg_send;
1065
1066                         num_zc++;
1067                         copied += try_to_copy;
1068
1069                         sk_msg_sg_copy_set(msg_pl, first);
1070                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1071                                                   record_type, &copied,
1072                                                   msg->msg_flags);
1073                         if (ret) {
1074                                 if (ret == -EINPROGRESS)
1075                                         num_async++;
1076                                 else if (ret == -ENOMEM)
1077                                         goto wait_for_memory;
1078                                 else if (ctx->open_rec && ret == -ENOSPC)
1079                                         goto rollback_iter;
1080                                 else if (ret != -EAGAIN)
1081                                         goto send_end;
1082                         }
1083                         continue;
1084 rollback_iter:
1085                         copied -= try_to_copy;
1086                         sk_msg_sg_copy_clear(msg_pl, first);
1087                         iov_iter_revert(&msg->msg_iter,
1088                                         msg_pl->sg.size - orig_size);
1089 fallback_to_reg_send:
1090                         sk_msg_trim(sk, msg_pl, orig_size);
1091                 }
1092
1093                 required_size = msg_pl->sg.size + try_to_copy;
1094
1095                 ret = tls_clone_plaintext_msg(sk, required_size);
1096                 if (ret) {
1097                         if (ret != -ENOSPC)
1098                                 goto send_end;
1099
1100                         /* Adjust try_to_copy according to the amount that was
1101                          * actually allocated. The difference is due
1102                          * to max sg elements limit
1103                          */
1104                         try_to_copy -= required_size - msg_pl->sg.size;
1105                         full_record = true;
1106                         sk_msg_trim(sk, msg_en,
1107                                     msg_pl->sg.size + prot->overhead_size);
1108                 }
1109
1110                 if (try_to_copy) {
1111                         ret = sk_msg_memcopy_from_iter(sk, &msg->msg_iter,
1112                                                        msg_pl, try_to_copy);
1113                         if (ret < 0)
1114                                 goto trim_sgl;
1115                 }
1116
1117                 /* Open records defined only if successfully copied, otherwise
1118                  * we would trim the sg but not reset the open record frags.
1119                  */
1120                 tls_ctx->pending_open_record_frags = true;
1121                 copied += try_to_copy;
1122 copied:
1123                 if (full_record || eor) {
1124                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1125                                                   record_type, &copied,
1126                                                   msg->msg_flags);
1127                         if (ret) {
1128                                 if (ret == -EINPROGRESS)
1129                                         num_async++;
1130                                 else if (ret == -ENOMEM)
1131                                         goto wait_for_memory;
1132                                 else if (ret != -EAGAIN) {
1133                                         if (ret == -ENOSPC)
1134                                                 ret = 0;
1135                                         goto send_end;
1136                                 }
1137                         }
1138                 }
1139
1140                 continue;
1141
1142 wait_for_sndbuf:
1143                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1144 wait_for_memory:
1145                 ret = sk_stream_wait_memory(sk, &timeo);
1146                 if (ret) {
1147 trim_sgl:
1148                         if (ctx->open_rec)
1149                                 tls_trim_both_msgs(sk, orig_size);
1150                         goto send_end;
1151                 }
1152
1153                 if (ctx->open_rec && msg_en->sg.size < required_size)
1154                         goto alloc_encrypted;
1155         }
1156
1157         if (!num_async) {
1158                 goto send_end;
1159         } else if (num_zc) {
1160                 /* Wait for pending encryptions to get completed */
1161                 spin_lock_bh(&ctx->encrypt_compl_lock);
1162                 ctx->async_notify = true;
1163
1164                 pending = atomic_read(&ctx->encrypt_pending);
1165                 spin_unlock_bh(&ctx->encrypt_compl_lock);
1166                 if (pending)
1167                         crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1168                 else
1169                         reinit_completion(&ctx->async_wait.completion);
1170
1171                 /* There can be no concurrent accesses, since we have no
1172                  * pending encrypt operations
1173                  */
1174                 WRITE_ONCE(ctx->async_notify, false);
1175
1176                 if (ctx->async_wait.err) {
1177                         ret = ctx->async_wait.err;
1178                         copied = 0;
1179                 }
1180         }
1181
1182         /* Transmit if any encryptions have completed */
1183         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1184                 cancel_delayed_work(&ctx->tx_work.work);
1185                 tls_tx_records(sk, msg->msg_flags);
1186         }
1187
1188 send_end:
1189         ret = sk_stream_error(sk, msg->msg_flags, ret);
1190         return copied > 0 ? copied : ret;
1191 }
1192
1193 int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1194 {
1195         struct tls_context *tls_ctx = tls_get_ctx(sk);
1196         int ret;
1197
1198         if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1199                                MSG_CMSG_COMPAT | MSG_SPLICE_PAGES | MSG_EOR |
1200                                MSG_SENDPAGE_NOPOLICY))
1201                 return -EOPNOTSUPP;
1202
1203         ret = mutex_lock_interruptible(&tls_ctx->tx_lock);
1204         if (ret)
1205                 return ret;
1206         lock_sock(sk);
1207         ret = tls_sw_sendmsg_locked(sk, msg, size);
1208         release_sock(sk);
1209         mutex_unlock(&tls_ctx->tx_lock);
1210         return ret;
1211 }
1212
1213 /*
1214  * Handle unexpected EOF during splice without SPLICE_F_MORE set.
1215  */
1216 void tls_sw_splice_eof(struct socket *sock)
1217 {
1218         struct sock *sk = sock->sk;
1219         struct tls_context *tls_ctx = tls_get_ctx(sk);
1220         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1221         struct tls_rec *rec;
1222         struct sk_msg *msg_pl;
1223         ssize_t copied = 0;
1224         bool retrying = false;
1225         int ret = 0;
1226         int pending;
1227
1228         if (!ctx->open_rec)
1229                 return;
1230
1231         mutex_lock(&tls_ctx->tx_lock);
1232         lock_sock(sk);
1233
1234 retry:
1235         rec = ctx->open_rec;
1236         if (!rec)
1237                 goto unlock;
1238
1239         msg_pl = &rec->msg_plaintext;
1240
1241         /* Check the BPF advisor and perform transmission. */
1242         ret = bpf_exec_tx_verdict(msg_pl, sk, false, TLS_RECORD_TYPE_DATA,
1243                                   &copied, 0);
1244         switch (ret) {
1245         case 0:
1246         case -EAGAIN:
1247                 if (retrying)
1248                         goto unlock;
1249                 retrying = true;
1250                 goto retry;
1251         case -EINPROGRESS:
1252                 break;
1253         default:
1254                 goto unlock;
1255         }
1256
1257         /* Wait for pending encryptions to get completed */
1258         spin_lock_bh(&ctx->encrypt_compl_lock);
1259         ctx->async_notify = true;
1260
1261         pending = atomic_read(&ctx->encrypt_pending);
1262         spin_unlock_bh(&ctx->encrypt_compl_lock);
1263         if (pending)
1264                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1265         else
1266                 reinit_completion(&ctx->async_wait.completion);
1267
1268         /* There can be no concurrent accesses, since we have no pending
1269          * encrypt operations
1270          */
1271         WRITE_ONCE(ctx->async_notify, false);
1272
1273         if (ctx->async_wait.err)
1274                 goto unlock;
1275
1276         /* Transmit if any encryptions have completed */
1277         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1278                 cancel_delayed_work(&ctx->tx_work.work);
1279                 tls_tx_records(sk, 0);
1280         }
1281
1282 unlock:
1283         release_sock(sk);
1284         mutex_unlock(&tls_ctx->tx_lock);
1285 }
1286
1287 static int
1288 tls_rx_rec_wait(struct sock *sk, struct sk_psock *psock, bool nonblock,
1289                 bool released)
1290 {
1291         struct tls_context *tls_ctx = tls_get_ctx(sk);
1292         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1293         DEFINE_WAIT_FUNC(wait, woken_wake_function);
1294         long timeo;
1295
1296         timeo = sock_rcvtimeo(sk, nonblock);
1297
1298         while (!tls_strp_msg_ready(ctx)) {
1299                 if (!sk_psock_queue_empty(psock))
1300                         return 0;
1301
1302                 if (sk->sk_err)
1303                         return sock_error(sk);
1304
1305                 if (!skb_queue_empty(&sk->sk_receive_queue)) {
1306                         tls_strp_check_rcv(&ctx->strp);
1307                         if (tls_strp_msg_ready(ctx))
1308                                 break;
1309                 }
1310
1311                 if (sk->sk_shutdown & RCV_SHUTDOWN)
1312                         return 0;
1313
1314                 if (sock_flag(sk, SOCK_DONE))
1315                         return 0;
1316
1317                 if (!timeo)
1318                         return -EAGAIN;
1319
1320                 released = true;
1321                 add_wait_queue(sk_sleep(sk), &wait);
1322                 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1323                 sk_wait_event(sk, &timeo,
1324                               tls_strp_msg_ready(ctx) ||
1325                               !sk_psock_queue_empty(psock),
1326                               &wait);
1327                 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1328                 remove_wait_queue(sk_sleep(sk), &wait);
1329
1330                 /* Handle signals */
1331                 if (signal_pending(current))
1332                         return sock_intr_errno(timeo);
1333         }
1334
1335         tls_strp_msg_load(&ctx->strp, released);
1336
1337         return 1;
1338 }
1339
1340 static int tls_setup_from_iter(struct iov_iter *from,
1341                                int length, int *pages_used,
1342                                struct scatterlist *to,
1343                                int to_max_pages)
1344 {
1345         int rc = 0, i = 0, num_elem = *pages_used, maxpages;
1346         struct page *pages[MAX_SKB_FRAGS];
1347         unsigned int size = 0;
1348         ssize_t copied, use;
1349         size_t offset;
1350
1351         while (length > 0) {
1352                 i = 0;
1353                 maxpages = to_max_pages - num_elem;
1354                 if (maxpages == 0) {
1355                         rc = -EFAULT;
1356                         goto out;
1357                 }
1358                 copied = iov_iter_get_pages2(from, pages,
1359                                             length,
1360                                             maxpages, &offset);
1361                 if (copied <= 0) {
1362                         rc = -EFAULT;
1363                         goto out;
1364                 }
1365
1366                 length -= copied;
1367                 size += copied;
1368                 while (copied) {
1369                         use = min_t(int, copied, PAGE_SIZE - offset);
1370
1371                         sg_set_page(&to[num_elem],
1372                                     pages[i], use, offset);
1373                         sg_unmark_end(&to[num_elem]);
1374                         /* We do not uncharge memory from this API */
1375
1376                         offset = 0;
1377                         copied -= use;
1378
1379                         i++;
1380                         num_elem++;
1381                 }
1382         }
1383         /* Mark the end in the last sg entry if newly added */
1384         if (num_elem > *pages_used)
1385                 sg_mark_end(&to[num_elem - 1]);
1386 out:
1387         if (rc)
1388                 iov_iter_revert(from, size);
1389         *pages_used = num_elem;
1390
1391         return rc;
1392 }
1393
1394 static struct sk_buff *
1395 tls_alloc_clrtxt_skb(struct sock *sk, struct sk_buff *skb,
1396                      unsigned int full_len)
1397 {
1398         struct strp_msg *clr_rxm;
1399         struct sk_buff *clr_skb;
1400         int err;
1401
1402         clr_skb = alloc_skb_with_frags(0, full_len, TLS_PAGE_ORDER,
1403                                        &err, sk->sk_allocation);
1404         if (!clr_skb)
1405                 return NULL;
1406
1407         skb_copy_header(clr_skb, skb);
1408         clr_skb->len = full_len;
1409         clr_skb->data_len = full_len;
1410
1411         clr_rxm = strp_msg(clr_skb);
1412         clr_rxm->offset = 0;
1413
1414         return clr_skb;
1415 }
1416
1417 /* Decrypt handlers
1418  *
1419  * tls_decrypt_sw() and tls_decrypt_device() are decrypt handlers.
1420  * They must transform the darg in/out argument are as follows:
1421  *       |          Input            |         Output
1422  * -------------------------------------------------------------------
1423  *    zc | Zero-copy decrypt allowed | Zero-copy performed
1424  * async | Async decrypt allowed     | Async crypto used / in progress
1425  *   skb |            *              | Output skb
1426  *
1427  * If ZC decryption was performed darg.skb will point to the input skb.
1428  */
1429
1430 /* This function decrypts the input skb into either out_iov or in out_sg
1431  * or in skb buffers itself. The input parameter 'darg->zc' indicates if
1432  * zero-copy mode needs to be tried or not. With zero-copy mode, either
1433  * out_iov or out_sg must be non-NULL. In case both out_iov and out_sg are
1434  * NULL, then the decryption happens inside skb buffers itself, i.e.
1435  * zero-copy gets disabled and 'darg->zc' is updated.
1436  */
1437 static int tls_decrypt_sg(struct sock *sk, struct iov_iter *out_iov,
1438                           struct scatterlist *out_sg,
1439                           struct tls_decrypt_arg *darg)
1440 {
1441         struct tls_context *tls_ctx = tls_get_ctx(sk);
1442         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1443         struct tls_prot_info *prot = &tls_ctx->prot_info;
1444         int n_sgin, n_sgout, aead_size, err, pages = 0;
1445         struct sk_buff *skb = tls_strp_msg(ctx);
1446         const struct strp_msg *rxm = strp_msg(skb);
1447         const struct tls_msg *tlm = tls_msg(skb);
1448         struct aead_request *aead_req;
1449         struct scatterlist *sgin = NULL;
1450         struct scatterlist *sgout = NULL;
1451         const int data_len = rxm->full_len - prot->overhead_size;
1452         int tail_pages = !!prot->tail_size;
1453         struct tls_decrypt_ctx *dctx;
1454         struct sk_buff *clear_skb;
1455         int iv_offset = 0;
1456         u8 *mem;
1457
1458         n_sgin = skb_nsg(skb, rxm->offset + prot->prepend_size,
1459                          rxm->full_len - prot->prepend_size);
1460         if (n_sgin < 1)
1461                 return n_sgin ?: -EBADMSG;
1462
1463         if (darg->zc && (out_iov || out_sg)) {
1464                 clear_skb = NULL;
1465
1466                 if (out_iov)
1467                         n_sgout = 1 + tail_pages +
1468                                 iov_iter_npages_cap(out_iov, INT_MAX, data_len);
1469                 else
1470                         n_sgout = sg_nents(out_sg);
1471         } else {
1472                 darg->zc = false;
1473
1474                 clear_skb = tls_alloc_clrtxt_skb(sk, skb, rxm->full_len);
1475                 if (!clear_skb)
1476                         return -ENOMEM;
1477
1478                 n_sgout = 1 + skb_shinfo(clear_skb)->nr_frags;
1479         }
1480
1481         /* Increment to accommodate AAD */
1482         n_sgin = n_sgin + 1;
1483
1484         /* Allocate a single block of memory which contains
1485          *   aead_req || tls_decrypt_ctx.
1486          * Both structs are variable length.
1487          */
1488         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(ctx->aead_recv);
1489         aead_size = ALIGN(aead_size, __alignof__(*dctx));
1490         mem = kmalloc(aead_size + struct_size(dctx, sg, n_sgin + n_sgout),
1491                       sk->sk_allocation);
1492         if (!mem) {
1493                 err = -ENOMEM;
1494                 goto exit_free_skb;
1495         }
1496
1497         /* Segment the allocated memory */
1498         aead_req = (struct aead_request *)mem;
1499         dctx = (struct tls_decrypt_ctx *)(mem + aead_size);
1500         dctx->sk = sk;
1501         sgin = &dctx->sg[0];
1502         sgout = &dctx->sg[n_sgin];
1503
1504         /* For CCM based ciphers, first byte of nonce+iv is a constant */
1505         switch (prot->cipher_type) {
1506         case TLS_CIPHER_AES_CCM_128:
1507                 dctx->iv[0] = TLS_AES_CCM_IV_B0_BYTE;
1508                 iv_offset = 1;
1509                 break;
1510         case TLS_CIPHER_SM4_CCM:
1511                 dctx->iv[0] = TLS_SM4_CCM_IV_B0_BYTE;
1512                 iv_offset = 1;
1513                 break;
1514         }
1515
1516         /* Prepare IV */
1517         if (prot->version == TLS_1_3_VERSION ||
1518             prot->cipher_type == TLS_CIPHER_CHACHA20_POLY1305) {
1519                 memcpy(&dctx->iv[iv_offset], tls_ctx->rx.iv,
1520                        prot->iv_size + prot->salt_size);
1521         } else {
1522                 err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
1523                                     &dctx->iv[iv_offset] + prot->salt_size,
1524                                     prot->iv_size);
1525                 if (err < 0)
1526                         goto exit_free;
1527                 memcpy(&dctx->iv[iv_offset], tls_ctx->rx.iv, prot->salt_size);
1528         }
1529         tls_xor_iv_with_seq(prot, &dctx->iv[iv_offset], tls_ctx->rx.rec_seq);
1530
1531         /* Prepare AAD */
1532         tls_make_aad(dctx->aad, rxm->full_len - prot->overhead_size +
1533                      prot->tail_size,
1534                      tls_ctx->rx.rec_seq, tlm->control, prot);
1535
1536         /* Prepare sgin */
1537         sg_init_table(sgin, n_sgin);
1538         sg_set_buf(&sgin[0], dctx->aad, prot->aad_size);
1539         err = skb_to_sgvec(skb, &sgin[1],
1540                            rxm->offset + prot->prepend_size,
1541                            rxm->full_len - prot->prepend_size);
1542         if (err < 0)
1543                 goto exit_free;
1544
1545         if (clear_skb) {
1546                 sg_init_table(sgout, n_sgout);
1547                 sg_set_buf(&sgout[0], dctx->aad, prot->aad_size);
1548
1549                 err = skb_to_sgvec(clear_skb, &sgout[1], prot->prepend_size,
1550                                    data_len + prot->tail_size);
1551                 if (err < 0)
1552                         goto exit_free;
1553         } else if (out_iov) {
1554                 sg_init_table(sgout, n_sgout);
1555                 sg_set_buf(&sgout[0], dctx->aad, prot->aad_size);
1556
1557                 err = tls_setup_from_iter(out_iov, data_len, &pages, &sgout[1],
1558                                           (n_sgout - 1 - tail_pages));
1559                 if (err < 0)
1560                         goto exit_free_pages;
1561
1562                 if (prot->tail_size) {
1563                         sg_unmark_end(&sgout[pages]);
1564                         sg_set_buf(&sgout[pages + 1], &dctx->tail,
1565                                    prot->tail_size);
1566                         sg_mark_end(&sgout[pages + 1]);
1567                 }
1568         } else if (out_sg) {
1569                 memcpy(sgout, out_sg, n_sgout * sizeof(*sgout));
1570         }
1571
1572         /* Prepare and submit AEAD request */
1573         err = tls_do_decryption(sk, sgin, sgout, dctx->iv,
1574                                 data_len + prot->tail_size, aead_req, darg);
1575         if (err)
1576                 goto exit_free_pages;
1577
1578         darg->skb = clear_skb ?: tls_strp_msg(ctx);
1579         clear_skb = NULL;
1580
1581         if (unlikely(darg->async)) {
1582                 err = tls_strp_msg_hold(&ctx->strp, &ctx->async_hold);
1583                 if (err)
1584                         __skb_queue_tail(&ctx->async_hold, darg->skb);
1585                 return err;
1586         }
1587
1588         if (prot->tail_size)
1589                 darg->tail = dctx->tail;
1590
1591 exit_free_pages:
1592         /* Release the pages in case iov was mapped to pages */
1593         for (; pages > 0; pages--)
1594                 put_page(sg_page(&sgout[pages]));
1595 exit_free:
1596         kfree(mem);
1597 exit_free_skb:
1598         consume_skb(clear_skb);
1599         return err;
1600 }
1601
1602 static int
1603 tls_decrypt_sw(struct sock *sk, struct tls_context *tls_ctx,
1604                struct msghdr *msg, struct tls_decrypt_arg *darg)
1605 {
1606         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1607         struct tls_prot_info *prot = &tls_ctx->prot_info;
1608         struct strp_msg *rxm;
1609         int pad, err;
1610
1611         err = tls_decrypt_sg(sk, &msg->msg_iter, NULL, darg);
1612         if (err < 0) {
1613                 if (err == -EBADMSG)
1614                         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSDECRYPTERROR);
1615                 return err;
1616         }
1617         /* keep going even for ->async, the code below is TLS 1.3 */
1618
1619         /* If opportunistic TLS 1.3 ZC failed retry without ZC */
1620         if (unlikely(darg->zc && prot->version == TLS_1_3_VERSION &&
1621                      darg->tail != TLS_RECORD_TYPE_DATA)) {
1622                 darg->zc = false;
1623                 if (!darg->tail)
1624                         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSRXNOPADVIOL);
1625                 TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSDECRYPTRETRY);
1626                 return tls_decrypt_sw(sk, tls_ctx, msg, darg);
1627         }
1628
1629         pad = tls_padding_length(prot, darg->skb, darg);
1630         if (pad < 0) {
1631                 if (darg->skb != tls_strp_msg(ctx))
1632                         consume_skb(darg->skb);
1633                 return pad;
1634         }
1635
1636         rxm = strp_msg(darg->skb);
1637         rxm->full_len -= pad;
1638
1639         return 0;
1640 }
1641
1642 static int
1643 tls_decrypt_device(struct sock *sk, struct msghdr *msg,
1644                    struct tls_context *tls_ctx, struct tls_decrypt_arg *darg)
1645 {
1646         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1647         struct tls_prot_info *prot = &tls_ctx->prot_info;
1648         struct strp_msg *rxm;
1649         int pad, err;
1650
1651         if (tls_ctx->rx_conf != TLS_HW)
1652                 return 0;
1653
1654         err = tls_device_decrypted(sk, tls_ctx);
1655         if (err <= 0)
1656                 return err;
1657
1658         pad = tls_padding_length(prot, tls_strp_msg(ctx), darg);
1659         if (pad < 0)
1660                 return pad;
1661
1662         darg->async = false;
1663         darg->skb = tls_strp_msg(ctx);
1664         /* ->zc downgrade check, in case TLS 1.3 gets here */
1665         darg->zc &= !(prot->version == TLS_1_3_VERSION &&
1666                       tls_msg(darg->skb)->control != TLS_RECORD_TYPE_DATA);
1667
1668         rxm = strp_msg(darg->skb);
1669         rxm->full_len -= pad;
1670
1671         if (!darg->zc) {
1672                 /* Non-ZC case needs a real skb */
1673                 darg->skb = tls_strp_msg_detach(ctx);
1674                 if (!darg->skb)
1675                         return -ENOMEM;
1676         } else {
1677                 unsigned int off, len;
1678
1679                 /* In ZC case nobody cares about the output skb.
1680                  * Just copy the data here. Note the skb is not fully trimmed.
1681                  */
1682                 off = rxm->offset + prot->prepend_size;
1683                 len = rxm->full_len - prot->overhead_size;
1684
1685                 err = skb_copy_datagram_msg(darg->skb, off, msg, len);
1686                 if (err)
1687                         return err;
1688         }
1689         return 1;
1690 }
1691
1692 static int tls_rx_one_record(struct sock *sk, struct msghdr *msg,
1693                              struct tls_decrypt_arg *darg)
1694 {
1695         struct tls_context *tls_ctx = tls_get_ctx(sk);
1696         struct tls_prot_info *prot = &tls_ctx->prot_info;
1697         struct strp_msg *rxm;
1698         int err;
1699
1700         err = tls_decrypt_device(sk, msg, tls_ctx, darg);
1701         if (!err)
1702                 err = tls_decrypt_sw(sk, tls_ctx, msg, darg);
1703         if (err < 0)
1704                 return err;
1705
1706         rxm = strp_msg(darg->skb);
1707         rxm->offset += prot->prepend_size;
1708         rxm->full_len -= prot->overhead_size;
1709         tls_advance_record_sn(sk, prot, &tls_ctx->rx);
1710
1711         return 0;
1712 }
1713
1714 int decrypt_skb(struct sock *sk, struct scatterlist *sgout)
1715 {
1716         struct tls_decrypt_arg darg = { .zc = true, };
1717
1718         return tls_decrypt_sg(sk, NULL, sgout, &darg);
1719 }
1720
1721 static int tls_record_content_type(struct msghdr *msg, struct tls_msg *tlm,
1722                                    u8 *control)
1723 {
1724         int err;
1725
1726         if (!*control) {
1727                 *control = tlm->control;
1728                 if (!*control)
1729                         return -EBADMSG;
1730
1731                 err = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1732                                sizeof(*control), control);
1733                 if (*control != TLS_RECORD_TYPE_DATA) {
1734                         if (err || msg->msg_flags & MSG_CTRUNC)
1735                                 return -EIO;
1736                 }
1737         } else if (*control != tlm->control) {
1738                 return 0;
1739         }
1740
1741         return 1;
1742 }
1743
1744 static void tls_rx_rec_done(struct tls_sw_context_rx *ctx)
1745 {
1746         tls_strp_msg_done(&ctx->strp);
1747 }
1748
1749 /* This function traverses the rx_list in tls receive context to copies the
1750  * decrypted records into the buffer provided by caller zero copy is not
1751  * true. Further, the records are removed from the rx_list if it is not a peek
1752  * case and the record has been consumed completely.
1753  */
1754 static int process_rx_list(struct tls_sw_context_rx *ctx,
1755                            struct msghdr *msg,
1756                            u8 *control,
1757                            size_t skip,
1758                            size_t len,
1759                            bool is_peek)
1760 {
1761         struct sk_buff *skb = skb_peek(&ctx->rx_list);
1762         struct tls_msg *tlm;
1763         ssize_t copied = 0;
1764         int err;
1765
1766         while (skip && skb) {
1767                 struct strp_msg *rxm = strp_msg(skb);
1768                 tlm = tls_msg(skb);
1769
1770                 err = tls_record_content_type(msg, tlm, control);
1771                 if (err <= 0)
1772                         goto out;
1773
1774                 if (skip < rxm->full_len)
1775                         break;
1776
1777                 skip = skip - rxm->full_len;
1778                 skb = skb_peek_next(skb, &ctx->rx_list);
1779         }
1780
1781         while (len && skb) {
1782                 struct sk_buff *next_skb;
1783                 struct strp_msg *rxm = strp_msg(skb);
1784                 int chunk = min_t(unsigned int, rxm->full_len - skip, len);
1785
1786                 tlm = tls_msg(skb);
1787
1788                 err = tls_record_content_type(msg, tlm, control);
1789                 if (err <= 0)
1790                         goto out;
1791
1792                 err = skb_copy_datagram_msg(skb, rxm->offset + skip,
1793                                             msg, chunk);
1794                 if (err < 0)
1795                         goto out;
1796
1797                 len = len - chunk;
1798                 copied = copied + chunk;
1799
1800                 /* Consume the data from record if it is non-peek case*/
1801                 if (!is_peek) {
1802                         rxm->offset = rxm->offset + chunk;
1803                         rxm->full_len = rxm->full_len - chunk;
1804
1805                         /* Return if there is unconsumed data in the record */
1806                         if (rxm->full_len - skip)
1807                                 break;
1808                 }
1809
1810                 /* The remaining skip-bytes must lie in 1st record in rx_list.
1811                  * So from the 2nd record, 'skip' should be 0.
1812                  */
1813                 skip = 0;
1814
1815                 if (msg)
1816                         msg->msg_flags |= MSG_EOR;
1817
1818                 next_skb = skb_peek_next(skb, &ctx->rx_list);
1819
1820                 if (!is_peek) {
1821                         __skb_unlink(skb, &ctx->rx_list);
1822                         consume_skb(skb);
1823                 }
1824
1825                 skb = next_skb;
1826         }
1827         err = 0;
1828
1829 out:
1830         return copied ? : err;
1831 }
1832
1833 static bool
1834 tls_read_flush_backlog(struct sock *sk, struct tls_prot_info *prot,
1835                        size_t len_left, size_t decrypted, ssize_t done,
1836                        size_t *flushed_at)
1837 {
1838         size_t max_rec;
1839
1840         if (len_left <= decrypted)
1841                 return false;
1842
1843         max_rec = prot->overhead_size - prot->tail_size + TLS_MAX_PAYLOAD_SIZE;
1844         if (done - *flushed_at < SZ_128K && tcp_inq(sk) > max_rec)
1845                 return false;
1846
1847         *flushed_at = done;
1848         return sk_flush_backlog(sk);
1849 }
1850
1851 static int tls_rx_reader_acquire(struct sock *sk, struct tls_sw_context_rx *ctx,
1852                                  bool nonblock)
1853 {
1854         long timeo;
1855
1856         timeo = sock_rcvtimeo(sk, nonblock);
1857
1858         while (unlikely(ctx->reader_present)) {
1859                 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1860
1861                 ctx->reader_contended = 1;
1862
1863                 add_wait_queue(&ctx->wq, &wait);
1864                 sk_wait_event(sk, &timeo,
1865                               !READ_ONCE(ctx->reader_present), &wait);
1866                 remove_wait_queue(&ctx->wq, &wait);
1867
1868                 if (timeo <= 0)
1869                         return -EAGAIN;
1870                 if (signal_pending(current))
1871                         return sock_intr_errno(timeo);
1872         }
1873
1874         WRITE_ONCE(ctx->reader_present, 1);
1875
1876         return 0;
1877 }
1878
1879 static int tls_rx_reader_lock(struct sock *sk, struct tls_sw_context_rx *ctx,
1880                               bool nonblock)
1881 {
1882         int err;
1883
1884         lock_sock(sk);
1885         err = tls_rx_reader_acquire(sk, ctx, nonblock);
1886         if (err)
1887                 release_sock(sk);
1888         return err;
1889 }
1890
1891 static void tls_rx_reader_release(struct sock *sk, struct tls_sw_context_rx *ctx)
1892 {
1893         if (unlikely(ctx->reader_contended)) {
1894                 if (wq_has_sleeper(&ctx->wq))
1895                         wake_up(&ctx->wq);
1896                 else
1897                         ctx->reader_contended = 0;
1898
1899                 WARN_ON_ONCE(!ctx->reader_present);
1900         }
1901
1902         WRITE_ONCE(ctx->reader_present, 0);
1903 }
1904
1905 static void tls_rx_reader_unlock(struct sock *sk, struct tls_sw_context_rx *ctx)
1906 {
1907         tls_rx_reader_release(sk, ctx);
1908         release_sock(sk);
1909 }
1910
1911 int tls_sw_recvmsg(struct sock *sk,
1912                    struct msghdr *msg,
1913                    size_t len,
1914                    int flags,
1915                    int *addr_len)
1916 {
1917         struct tls_context *tls_ctx = tls_get_ctx(sk);
1918         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1919         struct tls_prot_info *prot = &tls_ctx->prot_info;
1920         ssize_t decrypted = 0, async_copy_bytes = 0;
1921         struct sk_psock *psock;
1922         unsigned char control = 0;
1923         size_t flushed_at = 0;
1924         struct strp_msg *rxm;
1925         struct tls_msg *tlm;
1926         ssize_t copied = 0;
1927         bool async = false;
1928         int target, err;
1929         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
1930         bool is_peek = flags & MSG_PEEK;
1931         bool released = true;
1932         bool bpf_strp_enabled;
1933         bool zc_capable;
1934
1935         if (unlikely(flags & MSG_ERRQUEUE))
1936                 return sock_recv_errqueue(sk, msg, len, SOL_IP, IP_RECVERR);
1937
1938         psock = sk_psock_get(sk);
1939         err = tls_rx_reader_lock(sk, ctx, flags & MSG_DONTWAIT);
1940         if (err < 0)
1941                 return err;
1942         bpf_strp_enabled = sk_psock_strp_enabled(psock);
1943
1944         /* If crypto failed the connection is broken */
1945         err = ctx->async_wait.err;
1946         if (err)
1947                 goto end;
1948
1949         /* Process pending decrypted records. It must be non-zero-copy */
1950         err = process_rx_list(ctx, msg, &control, 0, len, is_peek);
1951         if (err < 0)
1952                 goto end;
1953
1954         copied = err;
1955         if (len <= copied)
1956                 goto end;
1957
1958         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1959         len = len - copied;
1960
1961         zc_capable = !bpf_strp_enabled && !is_kvec && !is_peek &&
1962                 ctx->zc_capable;
1963         decrypted = 0;
1964         while (len && (decrypted + copied < target || tls_strp_msg_ready(ctx))) {
1965                 struct tls_decrypt_arg darg;
1966                 int to_decrypt, chunk;
1967
1968                 err = tls_rx_rec_wait(sk, psock, flags & MSG_DONTWAIT,
1969                                       released);
1970                 if (err <= 0) {
1971                         if (psock) {
1972                                 chunk = sk_msg_recvmsg(sk, psock, msg, len,
1973                                                        flags);
1974                                 if (chunk > 0) {
1975                                         decrypted += chunk;
1976                                         len -= chunk;
1977                                         continue;
1978                                 }
1979                         }
1980                         goto recv_end;
1981                 }
1982
1983                 memset(&darg.inargs, 0, sizeof(darg.inargs));
1984
1985                 rxm = strp_msg(tls_strp_msg(ctx));
1986                 tlm = tls_msg(tls_strp_msg(ctx));
1987
1988                 to_decrypt = rxm->full_len - prot->overhead_size;
1989
1990                 if (zc_capable && to_decrypt <= len &&
1991                     tlm->control == TLS_RECORD_TYPE_DATA)
1992                         darg.zc = true;
1993
1994                 /* Do not use async mode if record is non-data */
1995                 if (tlm->control == TLS_RECORD_TYPE_DATA && !bpf_strp_enabled)
1996                         darg.async = ctx->async_capable;
1997                 else
1998                         darg.async = false;
1999
2000                 err = tls_rx_one_record(sk, msg, &darg);
2001                 if (err < 0) {
2002                         tls_err_abort(sk, -EBADMSG);
2003                         goto recv_end;
2004                 }
2005
2006                 async |= darg.async;
2007
2008                 /* If the type of records being processed is not known yet,
2009                  * set it to record type just dequeued. If it is already known,
2010                  * but does not match the record type just dequeued, go to end.
2011                  * We always get record type here since for tls1.2, record type
2012                  * is known just after record is dequeued from stream parser.
2013                  * For tls1.3, we disable async.
2014                  */
2015                 err = tls_record_content_type(msg, tls_msg(darg.skb), &control);
2016                 if (err <= 0) {
2017                         DEBUG_NET_WARN_ON_ONCE(darg.zc);
2018                         tls_rx_rec_done(ctx);
2019 put_on_rx_list_err:
2020                         __skb_queue_tail(&ctx->rx_list, darg.skb);
2021                         goto recv_end;
2022                 }
2023
2024                 /* periodically flush backlog, and feed strparser */
2025                 released = tls_read_flush_backlog(sk, prot, len, to_decrypt,
2026                                                   decrypted + copied,
2027                                                   &flushed_at);
2028
2029                 /* TLS 1.3 may have updated the length by more than overhead */
2030                 rxm = strp_msg(darg.skb);
2031                 chunk = rxm->full_len;
2032                 tls_rx_rec_done(ctx);
2033
2034                 if (!darg.zc) {
2035                         bool partially_consumed = chunk > len;
2036                         struct sk_buff *skb = darg.skb;
2037
2038                         DEBUG_NET_WARN_ON_ONCE(darg.skb == ctx->strp.anchor);
2039
2040                         if (async) {
2041                                 /* TLS 1.2-only, to_decrypt must be text len */
2042                                 chunk = min_t(int, to_decrypt, len);
2043                                 async_copy_bytes += chunk;
2044 put_on_rx_list:
2045                                 decrypted += chunk;
2046                                 len -= chunk;
2047                                 __skb_queue_tail(&ctx->rx_list, skb);
2048                                 continue;
2049                         }
2050
2051                         if (bpf_strp_enabled) {
2052                                 released = true;
2053                                 err = sk_psock_tls_strp_read(psock, skb);
2054                                 if (err != __SK_PASS) {
2055                                         rxm->offset = rxm->offset + rxm->full_len;
2056                                         rxm->full_len = 0;
2057                                         if (err == __SK_DROP)
2058                                                 consume_skb(skb);
2059                                         continue;
2060                                 }
2061                         }
2062
2063                         if (partially_consumed)
2064                                 chunk = len;
2065
2066                         err = skb_copy_datagram_msg(skb, rxm->offset,
2067                                                     msg, chunk);
2068                         if (err < 0)
2069                                 goto put_on_rx_list_err;
2070
2071                         if (is_peek)
2072                                 goto put_on_rx_list;
2073
2074                         if (partially_consumed) {
2075                                 rxm->offset += chunk;
2076                                 rxm->full_len -= chunk;
2077                                 goto put_on_rx_list;
2078                         }
2079
2080                         consume_skb(skb);
2081                 }
2082
2083                 decrypted += chunk;
2084                 len -= chunk;
2085
2086                 /* Return full control message to userspace before trying
2087                  * to parse another message type
2088                  */
2089                 msg->msg_flags |= MSG_EOR;
2090                 if (control != TLS_RECORD_TYPE_DATA)
2091                         break;
2092         }
2093
2094 recv_end:
2095         if (async) {
2096                 int ret, pending;
2097
2098                 /* Wait for all previously submitted records to be decrypted */
2099                 spin_lock_bh(&ctx->decrypt_compl_lock);
2100                 reinit_completion(&ctx->async_wait.completion);
2101                 pending = atomic_read(&ctx->decrypt_pending);
2102                 spin_unlock_bh(&ctx->decrypt_compl_lock);
2103                 ret = 0;
2104                 if (pending)
2105                         ret = crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
2106                 __skb_queue_purge(&ctx->async_hold);
2107
2108                 if (ret) {
2109                         if (err >= 0 || err == -EINPROGRESS)
2110                                 err = ret;
2111                         decrypted = 0;
2112                         goto end;
2113                 }
2114
2115                 /* Drain records from the rx_list & copy if required */
2116                 if (is_peek || is_kvec)
2117                         err = process_rx_list(ctx, msg, &control, copied,
2118                                               decrypted, is_peek);
2119                 else
2120                         err = process_rx_list(ctx, msg, &control, 0,
2121                                               async_copy_bytes, is_peek);
2122                 decrypted += max(err, 0);
2123         }
2124
2125         copied += decrypted;
2126
2127 end:
2128         tls_rx_reader_unlock(sk, ctx);
2129         if (psock)
2130                 sk_psock_put(sk, psock);
2131         return copied ? : err;
2132 }
2133
2134 ssize_t tls_sw_splice_read(struct socket *sock,  loff_t *ppos,
2135                            struct pipe_inode_info *pipe,
2136                            size_t len, unsigned int flags)
2137 {
2138         struct tls_context *tls_ctx = tls_get_ctx(sock->sk);
2139         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2140         struct strp_msg *rxm = NULL;
2141         struct sock *sk = sock->sk;
2142         struct tls_msg *tlm;
2143         struct sk_buff *skb;
2144         ssize_t copied = 0;
2145         int chunk;
2146         int err;
2147
2148         err = tls_rx_reader_lock(sk, ctx, flags & SPLICE_F_NONBLOCK);
2149         if (err < 0)
2150                 return err;
2151
2152         if (!skb_queue_empty(&ctx->rx_list)) {
2153                 skb = __skb_dequeue(&ctx->rx_list);
2154         } else {
2155                 struct tls_decrypt_arg darg;
2156
2157                 err = tls_rx_rec_wait(sk, NULL, flags & SPLICE_F_NONBLOCK,
2158                                       true);
2159                 if (err <= 0)
2160                         goto splice_read_end;
2161
2162                 memset(&darg.inargs, 0, sizeof(darg.inargs));
2163
2164                 err = tls_rx_one_record(sk, NULL, &darg);
2165                 if (err < 0) {
2166                         tls_err_abort(sk, -EBADMSG);
2167                         goto splice_read_end;
2168                 }
2169
2170                 tls_rx_rec_done(ctx);
2171                 skb = darg.skb;
2172         }
2173
2174         rxm = strp_msg(skb);
2175         tlm = tls_msg(skb);
2176
2177         /* splice does not support reading control messages */
2178         if (tlm->control != TLS_RECORD_TYPE_DATA) {
2179                 err = -EINVAL;
2180                 goto splice_requeue;
2181         }
2182
2183         chunk = min_t(unsigned int, rxm->full_len, len);
2184         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, chunk, flags);
2185         if (copied < 0)
2186                 goto splice_requeue;
2187
2188         if (chunk < rxm->full_len) {
2189                 rxm->offset += len;
2190                 rxm->full_len -= len;
2191                 goto splice_requeue;
2192         }
2193
2194         consume_skb(skb);
2195
2196 splice_read_end:
2197         tls_rx_reader_unlock(sk, ctx);
2198         return copied ? : err;
2199
2200 splice_requeue:
2201         __skb_queue_head(&ctx->rx_list, skb);
2202         goto splice_read_end;
2203 }
2204
2205 int tls_sw_read_sock(struct sock *sk, read_descriptor_t *desc,
2206                      sk_read_actor_t read_actor)
2207 {
2208         struct tls_context *tls_ctx = tls_get_ctx(sk);
2209         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2210         struct tls_prot_info *prot = &tls_ctx->prot_info;
2211         struct strp_msg *rxm = NULL;
2212         struct sk_buff *skb = NULL;
2213         struct sk_psock *psock;
2214         size_t flushed_at = 0;
2215         bool released = true;
2216         struct tls_msg *tlm;
2217         ssize_t copied = 0;
2218         ssize_t decrypted;
2219         int err, used;
2220
2221         psock = sk_psock_get(sk);
2222         if (psock) {
2223                 sk_psock_put(sk, psock);
2224                 return -EINVAL;
2225         }
2226         err = tls_rx_reader_acquire(sk, ctx, true);
2227         if (err < 0)
2228                 return err;
2229
2230         /* If crypto failed the connection is broken */
2231         err = ctx->async_wait.err;
2232         if (err)
2233                 goto read_sock_end;
2234
2235         decrypted = 0;
2236         do {
2237                 if (!skb_queue_empty(&ctx->rx_list)) {
2238                         skb = __skb_dequeue(&ctx->rx_list);
2239                         rxm = strp_msg(skb);
2240                         tlm = tls_msg(skb);
2241                 } else {
2242                         struct tls_decrypt_arg darg;
2243
2244                         err = tls_rx_rec_wait(sk, NULL, true, released);
2245                         if (err <= 0)
2246                                 goto read_sock_end;
2247
2248                         memset(&darg.inargs, 0, sizeof(darg.inargs));
2249
2250                         err = tls_rx_one_record(sk, NULL, &darg);
2251                         if (err < 0) {
2252                                 tls_err_abort(sk, -EBADMSG);
2253                                 goto read_sock_end;
2254                         }
2255
2256                         released = tls_read_flush_backlog(sk, prot, INT_MAX,
2257                                                           0, decrypted,
2258                                                           &flushed_at);
2259                         skb = darg.skb;
2260                         rxm = strp_msg(skb);
2261                         tlm = tls_msg(skb);
2262                         decrypted += rxm->full_len;
2263
2264                         tls_rx_rec_done(ctx);
2265                 }
2266
2267                 /* read_sock does not support reading control messages */
2268                 if (tlm->control != TLS_RECORD_TYPE_DATA) {
2269                         err = -EINVAL;
2270                         goto read_sock_requeue;
2271                 }
2272
2273                 used = read_actor(desc, skb, rxm->offset, rxm->full_len);
2274                 if (used <= 0) {
2275                         if (!copied)
2276                                 err = used;
2277                         goto read_sock_requeue;
2278                 }
2279                 copied += used;
2280                 if (used < rxm->full_len) {
2281                         rxm->offset += used;
2282                         rxm->full_len -= used;
2283                         if (!desc->count)
2284                                 goto read_sock_requeue;
2285                 } else {
2286                         consume_skb(skb);
2287                         if (!desc->count)
2288                                 skb = NULL;
2289                 }
2290         } while (skb);
2291
2292 read_sock_end:
2293         tls_rx_reader_release(sk, ctx);
2294         return copied ? : err;
2295
2296 read_sock_requeue:
2297         __skb_queue_head(&ctx->rx_list, skb);
2298         goto read_sock_end;
2299 }
2300
2301 bool tls_sw_sock_is_readable(struct sock *sk)
2302 {
2303         struct tls_context *tls_ctx = tls_get_ctx(sk);
2304         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2305         bool ingress_empty = true;
2306         struct sk_psock *psock;
2307
2308         rcu_read_lock();
2309         psock = sk_psock(sk);
2310         if (psock)
2311                 ingress_empty = list_empty(&psock->ingress_msg);
2312         rcu_read_unlock();
2313
2314         return !ingress_empty || tls_strp_msg_ready(ctx) ||
2315                 !skb_queue_empty(&ctx->rx_list);
2316 }
2317
2318 int tls_rx_msg_size(struct tls_strparser *strp, struct sk_buff *skb)
2319 {
2320         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2321         struct tls_prot_info *prot = &tls_ctx->prot_info;
2322         char header[TLS_HEADER_SIZE + MAX_IV_SIZE];
2323         size_t cipher_overhead;
2324         size_t data_len = 0;
2325         int ret;
2326
2327         /* Verify that we have a full TLS header, or wait for more data */
2328         if (strp->stm.offset + prot->prepend_size > skb->len)
2329                 return 0;
2330
2331         /* Sanity-check size of on-stack buffer. */
2332         if (WARN_ON(prot->prepend_size > sizeof(header))) {
2333                 ret = -EINVAL;
2334                 goto read_failure;
2335         }
2336
2337         /* Linearize header to local buffer */
2338         ret = skb_copy_bits(skb, strp->stm.offset, header, prot->prepend_size);
2339         if (ret < 0)
2340                 goto read_failure;
2341
2342         strp->mark = header[0];
2343
2344         data_len = ((header[4] & 0xFF) | (header[3] << 8));
2345
2346         cipher_overhead = prot->tag_size;
2347         if (prot->version != TLS_1_3_VERSION &&
2348             prot->cipher_type != TLS_CIPHER_CHACHA20_POLY1305)
2349                 cipher_overhead += prot->iv_size;
2350
2351         if (data_len > TLS_MAX_PAYLOAD_SIZE + cipher_overhead +
2352             prot->tail_size) {
2353                 ret = -EMSGSIZE;
2354                 goto read_failure;
2355         }
2356         if (data_len < cipher_overhead) {
2357                 ret = -EBADMSG;
2358                 goto read_failure;
2359         }
2360
2361         /* Note that both TLS1.3 and TLS1.2 use TLS_1_2 version here */
2362         if (header[1] != TLS_1_2_VERSION_MINOR ||
2363             header[2] != TLS_1_2_VERSION_MAJOR) {
2364                 ret = -EINVAL;
2365                 goto read_failure;
2366         }
2367
2368         tls_device_rx_resync_new_rec(strp->sk, data_len + TLS_HEADER_SIZE,
2369                                      TCP_SKB_CB(skb)->seq + strp->stm.offset);
2370         return data_len + TLS_HEADER_SIZE;
2371
2372 read_failure:
2373         tls_err_abort(strp->sk, ret);
2374
2375         return ret;
2376 }
2377
2378 void tls_rx_msg_ready(struct tls_strparser *strp)
2379 {
2380         struct tls_sw_context_rx *ctx;
2381
2382         ctx = container_of(strp, struct tls_sw_context_rx, strp);
2383         ctx->saved_data_ready(strp->sk);
2384 }
2385
2386 static void tls_data_ready(struct sock *sk)
2387 {
2388         struct tls_context *tls_ctx = tls_get_ctx(sk);
2389         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2390         struct sk_psock *psock;
2391         gfp_t alloc_save;
2392
2393         trace_sk_data_ready(sk);
2394
2395         alloc_save = sk->sk_allocation;
2396         sk->sk_allocation = GFP_ATOMIC;
2397         tls_strp_data_ready(&ctx->strp);
2398         sk->sk_allocation = alloc_save;
2399
2400         psock = sk_psock_get(sk);
2401         if (psock) {
2402                 if (!list_empty(&psock->ingress_msg))
2403                         ctx->saved_data_ready(sk);
2404                 sk_psock_put(sk, psock);
2405         }
2406 }
2407
2408 void tls_sw_cancel_work_tx(struct tls_context *tls_ctx)
2409 {
2410         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2411
2412         set_bit(BIT_TX_CLOSING, &ctx->tx_bitmask);
2413         set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask);
2414         cancel_delayed_work_sync(&ctx->tx_work.work);
2415 }
2416
2417 void tls_sw_release_resources_tx(struct sock *sk)
2418 {
2419         struct tls_context *tls_ctx = tls_get_ctx(sk);
2420         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2421         struct tls_rec *rec, *tmp;
2422         int pending;
2423
2424         /* Wait for any pending async encryptions to complete */
2425         spin_lock_bh(&ctx->encrypt_compl_lock);
2426         ctx->async_notify = true;
2427         pending = atomic_read(&ctx->encrypt_pending);
2428         spin_unlock_bh(&ctx->encrypt_compl_lock);
2429
2430         if (pending)
2431                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
2432
2433         tls_tx_records(sk, -1);
2434
2435         /* Free up un-sent records in tx_list. First, free
2436          * the partially sent record if any at head of tx_list.
2437          */
2438         if (tls_ctx->partially_sent_record) {
2439                 tls_free_partial_record(sk, tls_ctx);
2440                 rec = list_first_entry(&ctx->tx_list,
2441                                        struct tls_rec, list);
2442                 list_del(&rec->list);
2443                 sk_msg_free(sk, &rec->msg_plaintext);
2444                 kfree(rec);
2445         }
2446
2447         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
2448                 list_del(&rec->list);
2449                 sk_msg_free(sk, &rec->msg_encrypted);
2450                 sk_msg_free(sk, &rec->msg_plaintext);
2451                 kfree(rec);
2452         }
2453
2454         crypto_free_aead(ctx->aead_send);
2455         tls_free_open_rec(sk);
2456 }
2457
2458 void tls_sw_free_ctx_tx(struct tls_context *tls_ctx)
2459 {
2460         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2461
2462         kfree(ctx);
2463 }
2464
2465 void tls_sw_release_resources_rx(struct sock *sk)
2466 {
2467         struct tls_context *tls_ctx = tls_get_ctx(sk);
2468         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2469
2470         kfree(tls_ctx->rx.rec_seq);
2471         kfree(tls_ctx->rx.iv);
2472
2473         if (ctx->aead_recv) {
2474                 __skb_queue_purge(&ctx->rx_list);
2475                 crypto_free_aead(ctx->aead_recv);
2476                 tls_strp_stop(&ctx->strp);
2477                 /* If tls_sw_strparser_arm() was not called (cleanup paths)
2478                  * we still want to tls_strp_stop(), but sk->sk_data_ready was
2479                  * never swapped.
2480                  */
2481                 if (ctx->saved_data_ready) {
2482                         write_lock_bh(&sk->sk_callback_lock);
2483                         sk->sk_data_ready = ctx->saved_data_ready;
2484                         write_unlock_bh(&sk->sk_callback_lock);
2485                 }
2486         }
2487 }
2488
2489 void tls_sw_strparser_done(struct tls_context *tls_ctx)
2490 {
2491         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2492
2493         tls_strp_done(&ctx->strp);
2494 }
2495
2496 void tls_sw_free_ctx_rx(struct tls_context *tls_ctx)
2497 {
2498         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2499
2500         kfree(ctx);
2501 }
2502
2503 void tls_sw_free_resources_rx(struct sock *sk)
2504 {
2505         struct tls_context *tls_ctx = tls_get_ctx(sk);
2506
2507         tls_sw_release_resources_rx(sk);
2508         tls_sw_free_ctx_rx(tls_ctx);
2509 }
2510
2511 /* The work handler to transmitt the encrypted records in tx_list */
2512 static void tx_work_handler(struct work_struct *work)
2513 {
2514         struct delayed_work *delayed_work = to_delayed_work(work);
2515         struct tx_work *tx_work = container_of(delayed_work,
2516                                                struct tx_work, work);
2517         struct sock *sk = tx_work->sk;
2518         struct tls_context *tls_ctx = tls_get_ctx(sk);
2519         struct tls_sw_context_tx *ctx;
2520
2521         if (unlikely(!tls_ctx))
2522                 return;
2523
2524         ctx = tls_sw_ctx_tx(tls_ctx);
2525         if (test_bit(BIT_TX_CLOSING, &ctx->tx_bitmask))
2526                 return;
2527
2528         if (!test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
2529                 return;
2530
2531         if (mutex_trylock(&tls_ctx->tx_lock)) {
2532                 lock_sock(sk);
2533                 tls_tx_records(sk, -1);
2534                 release_sock(sk);
2535                 mutex_unlock(&tls_ctx->tx_lock);
2536         } else if (!test_and_set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
2537                 /* Someone is holding the tx_lock, they will likely run Tx
2538                  * and cancel the work on their way out of the lock section.
2539                  * Schedule a long delay just in case.
2540                  */
2541                 schedule_delayed_work(&ctx->tx_work.work, msecs_to_jiffies(10));
2542         }
2543 }
2544
2545 static bool tls_is_tx_ready(struct tls_sw_context_tx *ctx)
2546 {
2547         struct tls_rec *rec;
2548
2549         rec = list_first_entry_or_null(&ctx->tx_list, struct tls_rec, list);
2550         if (!rec)
2551                 return false;
2552
2553         return READ_ONCE(rec->tx_ready);
2554 }
2555
2556 void tls_sw_write_space(struct sock *sk, struct tls_context *ctx)
2557 {
2558         struct tls_sw_context_tx *tx_ctx = tls_sw_ctx_tx(ctx);
2559
2560         /* Schedule the transmission if tx list is ready */
2561         if (tls_is_tx_ready(tx_ctx) &&
2562             !test_and_set_bit(BIT_TX_SCHEDULED, &tx_ctx->tx_bitmask))
2563                 schedule_delayed_work(&tx_ctx->tx_work.work, 0);
2564 }
2565
2566 void tls_sw_strparser_arm(struct sock *sk, struct tls_context *tls_ctx)
2567 {
2568         struct tls_sw_context_rx *rx_ctx = tls_sw_ctx_rx(tls_ctx);
2569
2570         write_lock_bh(&sk->sk_callback_lock);
2571         rx_ctx->saved_data_ready = sk->sk_data_ready;
2572         sk->sk_data_ready = tls_data_ready;
2573         write_unlock_bh(&sk->sk_callback_lock);
2574 }
2575
2576 void tls_update_rx_zc_capable(struct tls_context *tls_ctx)
2577 {
2578         struct tls_sw_context_rx *rx_ctx = tls_sw_ctx_rx(tls_ctx);
2579
2580         rx_ctx->zc_capable = tls_ctx->rx_no_pad ||
2581                 tls_ctx->prot_info.version != TLS_1_3_VERSION;
2582 }
2583
2584 int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
2585 {
2586         struct tls_context *tls_ctx = tls_get_ctx(sk);
2587         struct tls_prot_info *prot = &tls_ctx->prot_info;
2588         struct tls_crypto_info *crypto_info;
2589         struct tls_sw_context_tx *sw_ctx_tx = NULL;
2590         struct tls_sw_context_rx *sw_ctx_rx = NULL;
2591         struct cipher_context *cctx;
2592         struct crypto_aead **aead;
2593         struct crypto_tfm *tfm;
2594         char *iv, *rec_seq, *key, *salt;
2595         const struct tls_cipher_desc *cipher_desc;
2596         u16 nonce_size;
2597         int rc = 0;
2598
2599         if (!ctx) {
2600                 rc = -EINVAL;
2601                 goto out;
2602         }
2603
2604         if (tx) {
2605                 if (!ctx->priv_ctx_tx) {
2606                         sw_ctx_tx = kzalloc(sizeof(*sw_ctx_tx), GFP_KERNEL);
2607                         if (!sw_ctx_tx) {
2608                                 rc = -ENOMEM;
2609                                 goto out;
2610                         }
2611                         ctx->priv_ctx_tx = sw_ctx_tx;
2612                 } else {
2613                         sw_ctx_tx =
2614                                 (struct tls_sw_context_tx *)ctx->priv_ctx_tx;
2615                 }
2616         } else {
2617                 if (!ctx->priv_ctx_rx) {
2618                         sw_ctx_rx = kzalloc(sizeof(*sw_ctx_rx), GFP_KERNEL);
2619                         if (!sw_ctx_rx) {
2620                                 rc = -ENOMEM;
2621                                 goto out;
2622                         }
2623                         ctx->priv_ctx_rx = sw_ctx_rx;
2624                 } else {
2625                         sw_ctx_rx =
2626                                 (struct tls_sw_context_rx *)ctx->priv_ctx_rx;
2627                 }
2628         }
2629
2630         if (tx) {
2631                 crypto_init_wait(&sw_ctx_tx->async_wait);
2632                 spin_lock_init(&sw_ctx_tx->encrypt_compl_lock);
2633                 crypto_info = &ctx->crypto_send.info;
2634                 cctx = &ctx->tx;
2635                 aead = &sw_ctx_tx->aead_send;
2636                 INIT_LIST_HEAD(&sw_ctx_tx->tx_list);
2637                 INIT_DELAYED_WORK(&sw_ctx_tx->tx_work.work, tx_work_handler);
2638                 sw_ctx_tx->tx_work.sk = sk;
2639         } else {
2640                 crypto_init_wait(&sw_ctx_rx->async_wait);
2641                 spin_lock_init(&sw_ctx_rx->decrypt_compl_lock);
2642                 init_waitqueue_head(&sw_ctx_rx->wq);
2643                 crypto_info = &ctx->crypto_recv.info;
2644                 cctx = &ctx->rx;
2645                 skb_queue_head_init(&sw_ctx_rx->rx_list);
2646                 skb_queue_head_init(&sw_ctx_rx->async_hold);
2647                 aead = &sw_ctx_rx->aead_recv;
2648         }
2649
2650         cipher_desc = get_cipher_desc(crypto_info->cipher_type);
2651         if (!cipher_desc) {
2652                 rc = -EINVAL;
2653                 goto free_priv;
2654         }
2655
2656         nonce_size = cipher_desc->nonce;
2657
2658         iv = crypto_info_iv(crypto_info, cipher_desc);
2659         key = crypto_info_key(crypto_info, cipher_desc);
2660         salt = crypto_info_salt(crypto_info, cipher_desc);
2661         rec_seq = crypto_info_rec_seq(crypto_info, cipher_desc);
2662
2663         if (crypto_info->version == TLS_1_3_VERSION) {
2664                 nonce_size = 0;
2665                 prot->aad_size = TLS_HEADER_SIZE;
2666                 prot->tail_size = 1;
2667         } else {
2668                 prot->aad_size = TLS_AAD_SPACE_SIZE;
2669                 prot->tail_size = 0;
2670         }
2671
2672         /* Sanity-check the sizes for stack allocations. */
2673         if (nonce_size > MAX_IV_SIZE || prot->aad_size > TLS_MAX_AAD_SIZE) {
2674                 rc = -EINVAL;
2675                 goto free_priv;
2676         }
2677
2678         prot->version = crypto_info->version;
2679         prot->cipher_type = crypto_info->cipher_type;
2680         prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
2681         prot->tag_size = cipher_desc->tag;
2682         prot->overhead_size = prot->prepend_size +
2683                               prot->tag_size + prot->tail_size;
2684         prot->iv_size = cipher_desc->iv;
2685         prot->salt_size = cipher_desc->salt;
2686         cctx->iv = kmalloc(cipher_desc->iv + cipher_desc->salt, GFP_KERNEL);
2687         if (!cctx->iv) {
2688                 rc = -ENOMEM;
2689                 goto free_priv;
2690         }
2691         /* Note: 128 & 256 bit salt are the same size */
2692         prot->rec_seq_size = cipher_desc->rec_seq;
2693         memcpy(cctx->iv, salt, cipher_desc->salt);
2694         memcpy(cctx->iv + cipher_desc->salt, iv, cipher_desc->iv);
2695
2696         cctx->rec_seq = kmemdup(rec_seq, cipher_desc->rec_seq, GFP_KERNEL);
2697         if (!cctx->rec_seq) {
2698                 rc = -ENOMEM;
2699                 goto free_iv;
2700         }
2701
2702         if (!*aead) {
2703                 *aead = crypto_alloc_aead(cipher_desc->cipher_name, 0, 0);
2704                 if (IS_ERR(*aead)) {
2705                         rc = PTR_ERR(*aead);
2706                         *aead = NULL;
2707                         goto free_rec_seq;
2708                 }
2709         }
2710
2711         ctx->push_pending_record = tls_sw_push_pending_record;
2712
2713         rc = crypto_aead_setkey(*aead, key, cipher_desc->key);
2714         if (rc)
2715                 goto free_aead;
2716
2717         rc = crypto_aead_setauthsize(*aead, prot->tag_size);
2718         if (rc)
2719                 goto free_aead;
2720
2721         if (sw_ctx_rx) {
2722                 tfm = crypto_aead_tfm(sw_ctx_rx->aead_recv);
2723
2724                 tls_update_rx_zc_capable(ctx);
2725                 sw_ctx_rx->async_capable =
2726                         crypto_info->version != TLS_1_3_VERSION &&
2727                         !!(tfm->__crt_alg->cra_flags & CRYPTO_ALG_ASYNC);
2728
2729                 rc = tls_strp_init(&sw_ctx_rx->strp, sk);
2730                 if (rc)
2731                         goto free_aead;
2732         }
2733
2734         goto out;
2735
2736 free_aead:
2737         crypto_free_aead(*aead);
2738         *aead = NULL;
2739 free_rec_seq:
2740         kfree(cctx->rec_seq);
2741         cctx->rec_seq = NULL;
2742 free_iv:
2743         kfree(cctx->iv);
2744         cctx->iv = NULL;
2745 free_priv:
2746         if (tx) {
2747                 kfree(ctx->priv_ctx_tx);
2748                 ctx->priv_ctx_tx = NULL;
2749         } else {
2750                 kfree(ctx->priv_ctx_rx);
2751                 ctx->priv_ctx_rx = NULL;
2752         }
2753 out:
2754         return rc;
2755 }