net/tls/tls_sw.c

   1 /*
   2  * Copyright (c) 2016-2017, Mellanox Technologies. All rights reserved.
   3  * Copyright (c) 2016-2017, Dave Watson <[email protected]>. All rights reserved.
   4  * Copyright (c) 2016-2017, Lance Chao <[email protected]>. All rights reserved.
   5  * Copyright (c) 2016, Fridolin Pokorny <[email protected]>. All rights reserved.
   6  * Copyright (c) 2016, Nikos Mavrogiannopoulos <[email protected]>. 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/splice.h>
  42 #include <crypto/aead.h>
  43
  44 #include <net/strparser.h>
  45 #include <net/tls.h>
  46
  47 #include "tls.h"
  48
  49 struct tls_decrypt_arg {
  50         struct_group(inargs,
  51         bool zc;
  52         bool async;
  53         u8 tail;
  54         );
  55
  56         struct sk_buff *skb;
  57 };
  58
  59 struct tls_decrypt_ctx {
  60         u8 iv[MAX_IV_SIZE];
  61         u8 aad[TLS_MAX_AAD_SIZE];
  62         u8 tail;
  63         struct scatterlist sg[];
  64 };
  65
  66 noinline void tls_err_abort(struct sock *sk, int err)
  67 {
  68         WARN_ON_ONCE(err >= 0);
  69         /* sk->sk_err should contain a positive error code. */
  70         sk->sk_err = -err;
  71         sk_error_report(sk);
  72 }
  73
  74 static int __skb_nsg(struct sk_buff *skb, int offset, int len,
  75                      unsigned int recursion_level)
  76 {
  77         int start = skb_headlen(skb);
  78         int i, chunk = start - offset;
  79         struct sk_buff *frag_iter;
  80         int elt = 0;
  81
  82         if (unlikely(recursion_level >= 24))
  83                 return -EMSGSIZE;
  84
  85         if (chunk > 0) {
  86                 if (chunk > len)
  87                         chunk = len;
  88                 elt++;
  89                 len -= chunk;
  90                 if (len == 0)
  91                         return elt;
  92                 offset += chunk;
  93         }
  94
  95         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
  96                 int end;
  97
  98                 WARN_ON(start > offset + len);
  99
 100                 end = start + skb_frag_size(&skb_shinfo(skb)->frags[i]);
 101                 chunk = end - offset;
 102                 if (chunk > 0) {
 103                         if (chunk > len)
 104                                 chunk = len;
 105                         elt++;
 106                         len -= chunk;
 107                         if (len == 0)
 108                                 return elt;
 109                         offset += chunk;
 110                 }
 111                 start = end;
 112         }
 113
 114         if (unlikely(skb_has_frag_list(skb))) {
 115                 skb_walk_frags(skb, frag_iter) {
 116                         int end, ret;
 117
 118                         WARN_ON(start > offset + len);
 119
 120                         end = start + frag_iter->len;
 121                         chunk = end - offset;
 122                         if (chunk > 0) {
 123                                 if (chunk > len)
 124                                         chunk = len;
 125                                 ret = __skb_nsg(frag_iter, offset - start, chunk,
 126                                                 recursion_level + 1);
 127                                 if (unlikely(ret < 0))
 128                                         return ret;
 129                                 elt += ret;
 130                                 len -= chunk;
 131                                 if (len == 0)
 132                                         return elt;
 133                                 offset += chunk;
 134                         }
 135                         start = end;
 136                 }
 137         }
 138         BUG_ON(len);
 139         return elt;
 140 }
 141
 142 /* Return the number of scatterlist elements required to completely map the
 143  * skb, or -EMSGSIZE if the recursion depth is exceeded.
 144  */
 145 static int skb_nsg(struct sk_buff *skb, int offset, int len)
 146 {
 147         return __skb_nsg(skb, offset, len, 0);
 148 }
 149
 150 static int tls_padding_length(struct tls_prot_info *prot, struct sk_buff *skb,
 151                               struct tls_decrypt_arg *darg)
 152 {
 153         struct strp_msg *rxm = strp_msg(skb);
 154         struct tls_msg *tlm = tls_msg(skb);
 155         int sub = 0;
 156
 157         /* Determine zero-padding length */
 158         if (prot->version == TLS_1_3_VERSION) {
 159                 int offset = rxm->full_len - TLS_TAG_SIZE - 1;
 160                 char content_type = darg->zc ? darg->tail : 0;
 161                 int err;
 162
 163                 while (content_type == 0) {
 164                         if (offset < prot->prepend_size)
 165                                 return -EBADMSG;
 166                         err = skb_copy_bits(skb, rxm->offset + offset,
 167                                             &content_type, 1);
 168                         if (err)
 169                                 return err;
 170                         if (content_type)
 171                                 break;
 172                         sub++;
 173                         offset--;
 174                 }
 175                 tlm->control = content_type;
 176         }
 177         return sub;
 178 }
 179
 180 static void tls_decrypt_done(struct crypto_async_request *req, int err)
 181 {
 182         struct aead_request *aead_req = (struct aead_request *)req;
 183         struct scatterlist *sgout = aead_req->dst;
 184         struct scatterlist *sgin = aead_req->src;
 185         struct tls_sw_context_rx *ctx;
 186         struct tls_context *tls_ctx;
 187         struct scatterlist *sg;
 188         unsigned int pages;
 189         struct sock *sk;
 190
 191         sk = (struct sock *)req->data;
 192         tls_ctx = tls_get_ctx(sk);
 193         ctx = tls_sw_ctx_rx(tls_ctx);
 194
 195         /* Propagate if there was an err */
 196         if (err) {
 197                 if (err == -EBADMSG)
 198                         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSDECRYPTERROR);
 199                 ctx->async_wait.err = err;
 200                 tls_err_abort(sk, err);
 201         }
 202
 203         /* Free the destination pages if skb was not decrypted inplace */
 204         if (sgout != sgin) {
 205                 /* Skip the first S/G entry as it points to AAD */
 206                 for_each_sg(sg_next(sgout), sg, UINT_MAX, pages) {
 207                         if (!sg)
 208                                 break;
 209                         put_page(sg_page(sg));
 210                 }
 211         }
 212
 213         kfree(aead_req);
 214
 215         spin_lock_bh(&ctx->decrypt_compl_lock);
 216         if (!atomic_dec_return(&ctx->decrypt_pending))
 217                 complete(&ctx->async_wait.completion);
 218         spin_unlock_bh(&ctx->decrypt_compl_lock);
 219 }
 220
 221 static int tls_do_decryption(struct sock *sk,
 222                              struct scatterlist *sgin,
 223                              struct scatterlist *sgout,
 224                              char *iv_recv,
 225                              size_t data_len,
 226                              struct aead_request *aead_req,
 227                              struct tls_decrypt_arg *darg)
 228 {
 229         struct tls_context *tls_ctx = tls_get_ctx(sk);
 230         struct tls_prot_info *prot = &tls_ctx->prot_info;
 231         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
 232         int ret;
 233
 234         aead_request_set_tfm(aead_req, ctx->aead_recv);
 235         aead_request_set_ad(aead_req, prot->aad_size);
 236         aead_request_set_crypt(aead_req, sgin, sgout,
 237                                data_len + prot->tag_size,
 238                                (u8 *)iv_recv);
 239
 240         if (darg->async) {
 241                 aead_request_set_callback(aead_req,
 242                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 243                                           tls_decrypt_done, sk);
 244                 atomic_inc(&ctx->decrypt_pending);
 245         } else {
 246                 aead_request_set_callback(aead_req,
 247                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 248                                           crypto_req_done, &ctx->async_wait);
 249         }
 250
 251         ret = crypto_aead_decrypt(aead_req);
 252         if (ret == -EINPROGRESS) {
 253                 if (darg->async)
 254                         return 0;
 255
 256                 ret = crypto_wait_req(ret, &ctx->async_wait);
 257         }
 258         darg->async = false;
 259
 260         return ret;
 261 }
 262
 263 static void tls_trim_both_msgs(struct sock *sk, int target_size)
 264 {
 265         struct tls_context *tls_ctx = tls_get_ctx(sk);
 266         struct tls_prot_info *prot = &tls_ctx->prot_info;
 267         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 268         struct tls_rec *rec = ctx->open_rec;
 269
 270         sk_msg_trim(sk, &rec->msg_plaintext, target_size);
 271         if (target_size > 0)
 272                 target_size += prot->overhead_size;
 273         sk_msg_trim(sk, &rec->msg_encrypted, target_size);
 274 }
 275
 276 static int tls_alloc_encrypted_msg(struct sock *sk, int len)
 277 {
 278         struct tls_context *tls_ctx = tls_get_ctx(sk);
 279         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 280         struct tls_rec *rec = ctx->open_rec;
 281         struct sk_msg *msg_en = &rec->msg_encrypted;
 282
 283         return sk_msg_alloc(sk, msg_en, len, 0);
 284 }
 285
 286 static int tls_clone_plaintext_msg(struct sock *sk, int required)
 287 {
 288         struct tls_context *tls_ctx = tls_get_ctx(sk);
 289         struct tls_prot_info *prot = &tls_ctx->prot_info;
 290         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 291         struct tls_rec *rec = ctx->open_rec;
 292         struct sk_msg *msg_pl = &rec->msg_plaintext;
 293         struct sk_msg *msg_en = &rec->msg_encrypted;
 294         int skip, len;
 295
 296         /* We add page references worth len bytes from encrypted sg
 297          * at the end of plaintext sg. It is guaranteed that msg_en
 298          * has enough required room (ensured by caller).
 299          */
 300         len = required - msg_pl->sg.size;
 301
 302         /* Skip initial bytes in msg_en's data to be able to use
 303          * same offset of both plain and encrypted data.
 304          */
 305         skip = prot->prepend_size + msg_pl->sg.size;
 306
 307         return sk_msg_clone(sk, msg_pl, msg_en, skip, len);
 308 }
 309
 310 static struct tls_rec *tls_get_rec(struct sock *sk)
 311 {
 312         struct tls_context *tls_ctx = tls_get_ctx(sk);
 313         struct tls_prot_info *prot = &tls_ctx->prot_info;
 314         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 315         struct sk_msg *msg_pl, *msg_en;
 316         struct tls_rec *rec;
 317         int mem_size;
 318
 319         mem_size = sizeof(struct tls_rec) + crypto_aead_reqsize(ctx->aead_send);
 320
 321         rec = kzalloc(mem_size, sk->sk_allocation);
 322         if (!rec)
 323                 return NULL;
 324
 325         msg_pl = &rec->msg_plaintext;
 326         msg_en = &rec->msg_encrypted;
 327
 328         sk_msg_init(msg_pl);
 329         sk_msg_init(msg_en);
 330
 331         sg_init_table(rec->sg_aead_in, 2);
 332         sg_set_buf(&rec->sg_aead_in[0], rec->aad_space, prot->aad_size);
 333         sg_unmark_end(&rec->sg_aead_in[1]);
 334
 335         sg_init_table(rec->sg_aead_out, 2);
 336         sg_set_buf(&rec->sg_aead_out[0], rec->aad_space, prot->aad_size);
 337         sg_unmark_end(&rec->sg_aead_out[1]);
 338
 339         return rec;
 340 }
 341
 342 static void tls_free_rec(struct sock *sk, struct tls_rec *rec)
 343 {
 344         sk_msg_free(sk, &rec->msg_encrypted);
 345         sk_msg_free(sk, &rec->msg_plaintext);
 346         kfree(rec);
 347 }
 348
 349 static void tls_free_open_rec(struct sock *sk)
 350 {
 351         struct tls_context *tls_ctx = tls_get_ctx(sk);
 352         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 353         struct tls_rec *rec = ctx->open_rec;
 354
 355         if (rec) {
 356                 tls_free_rec(sk, rec);
 357                 ctx->open_rec = NULL;
 358         }
 359 }
 360
 361 int tls_tx_records(struct sock *sk, int flags)
 362 {
 363         struct tls_context *tls_ctx = tls_get_ctx(sk);
 364         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 365         struct tls_rec *rec, *tmp;
 366         struct sk_msg *msg_en;
 367         int tx_flags, rc = 0;
 368
 369         if (tls_is_partially_sent_record(tls_ctx)) {
 370                 rec = list_first_entry(&ctx->tx_list,
 371                                        struct tls_rec, list);
 372
 373                 if (flags == -1)
 374                         tx_flags = rec->tx_flags;
 375                 else
 376                         tx_flags = flags;
 377
 378                 rc = tls_push_partial_record(sk, tls_ctx, tx_flags);
 379                 if (rc)
 380                         goto tx_err;
 381
 382                 /* Full record has been transmitted.
 383                  * Remove the head of tx_list
 384                  */
 385                 list_del(&rec->list);
 386                 sk_msg_free(sk, &rec->msg_plaintext);
 387                 kfree(rec);
 388         }
 389
 390         /* Tx all ready records */
 391         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
 392                 if (READ_ONCE(rec->tx_ready)) {
 393                         if (flags == -1)
 394                                 tx_flags = rec->tx_flags;
 395                         else
 396                                 tx_flags = flags;
 397
 398                         msg_en = &rec->msg_encrypted;
 399                         rc = tls_push_sg(sk, tls_ctx,
 400                                          &msg_en->sg.data[msg_en->sg.curr],
 401                                          0, tx_flags);
 402                         if (rc)
 403                                 goto tx_err;
 404
 405                         list_del(&rec->list);
 406                         sk_msg_free(sk, &rec->msg_plaintext);
 407                         kfree(rec);
 408                 } else {
 409                         break;
 410                 }
 411         }
 412
 413 tx_err:
 414         if (rc < 0 && rc != -EAGAIN)
 415                 tls_err_abort(sk, -EBADMSG);
 416
 417         return rc;
 418 }
 419
 420 static void tls_encrypt_done(struct crypto_async_request *req, int err)
 421 {
 422         struct aead_request *aead_req = (struct aead_request *)req;
 423         struct sock *sk = req->data;
 424         struct tls_context *tls_ctx = tls_get_ctx(sk);
 425         struct tls_prot_info *prot = &tls_ctx->prot_info;
 426         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 427         struct scatterlist *sge;
 428         struct sk_msg *msg_en;
 429         struct tls_rec *rec;
 430         bool ready = false;
 431         int pending;
 432
 433         rec = container_of(aead_req, struct tls_rec, aead_req);
 434         msg_en = &rec->msg_encrypted;
 435
 436         sge = sk_msg_elem(msg_en, msg_en->sg.curr);
 437         sge->offset -= prot->prepend_size;
 438         sge->length += prot->prepend_size;
 439
 440         /* Check if error is previously set on socket */
 441         if (err || sk->sk_err) {
 442                 rec = NULL;
 443
 444                 /* If err is already set on socket, return the same code */
 445                 if (sk->sk_err) {
 446                         ctx->async_wait.err = -sk->sk_err;
 447                 } else {
 448                         ctx->async_wait.err = err;
 449                         tls_err_abort(sk, err);
 450                 }
 451         }
 452
 453         if (rec) {
 454                 struct tls_rec *first_rec;
 455
 456                 /* Mark the record as ready for transmission */
 457                 smp_store_mb(rec->tx_ready, true);
 458
 459                 /* If received record is at head of tx_list, schedule tx */
 460                 first_rec = list_first_entry(&ctx->tx_list,
 461                                              struct tls_rec, list);
 462                 if (rec == first_rec)
 463                         ready = true;
 464         }
 465
 466         spin_lock_bh(&ctx->encrypt_compl_lock);
 467         pending = atomic_dec_return(&ctx->encrypt_pending);
 468
 469         if (!pending && ctx->async_notify)
 470                 complete(&ctx->async_wait.completion);
 471         spin_unlock_bh(&ctx->encrypt_compl_lock);
 472
 473         if (!ready)
 474                 return;
 475
 476         /* Schedule the transmission */
 477         if (!test_and_set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
 478                 schedule_delayed_work(&ctx->tx_work.work, 1);
 479 }
 480
 481 static int tls_do_encryption(struct sock *sk,
 482                              struct tls_context *tls_ctx,
 483                              struct tls_sw_context_tx *ctx,
 484                              struct aead_request *aead_req,
 485                              size_t data_len, u32 start)
 486 {
 487         struct tls_prot_info *prot = &tls_ctx->prot_info;
 488         struct tls_rec *rec = ctx->open_rec;
 489         struct sk_msg *msg_en = &rec->msg_encrypted;
 490         struct scatterlist *sge = sk_msg_elem(msg_en, start);
 491         int rc, iv_offset = 0;
 492
 493         /* For CCM based ciphers, first byte of IV is a constant */
 494         switch (prot->cipher_type) {
 495         case TLS_CIPHER_AES_CCM_128:
 496                 rec->iv_data[0] = TLS_AES_CCM_IV_B0_BYTE;
 497                 iv_offset = 1;
 498                 break;
 499         case TLS_CIPHER_SM4_CCM:
 500                 rec->iv_data[0] = TLS_SM4_CCM_IV_B0_BYTE;
 501                 iv_offset = 1;
 502                 break;
 503         }
 504
 505         memcpy(&rec->iv_data[iv_offset], tls_ctx->tx.iv,
 506                prot->iv_size + prot->salt_size);
 507
 508         tls_xor_iv_with_seq(prot, rec->iv_data + iv_offset,
 509                             tls_ctx->tx.rec_seq);
 510
 511         sge->offset += prot->prepend_size;
 512         sge->length -= prot->prepend_size;
 513
 514         msg_en->sg.curr = start;
 515
 516         aead_request_set_tfm(aead_req, ctx->aead_send);
 517         aead_request_set_ad(aead_req, prot->aad_size);
 518         aead_request_set_crypt(aead_req, rec->sg_aead_in,
 519                                rec->sg_aead_out,
 520                                data_len, rec->iv_data);
 521
 522         aead_request_set_callback(aead_req, CRYPTO_TFM_REQ_MAY_BACKLOG,
 523                                   tls_encrypt_done, sk);
 524
 525         /* Add the record in tx_list */
 526         list_add_tail((struct list_head *)&rec->list, &ctx->tx_list);
 527         atomic_inc(&ctx->encrypt_pending);
 528
 529         rc = crypto_aead_encrypt(aead_req);
 530         if (!rc || rc != -EINPROGRESS) {
 531                 atomic_dec(&ctx->encrypt_pending);
 532                 sge->offset -= prot->prepend_size;
 533                 sge->length += prot->prepend_size;
 534         }
 535
 536         if (!rc) {
 537                 WRITE_ONCE(rec->tx_ready, true);
 538         } else if (rc != -EINPROGRESS) {
 539                 list_del(&rec->list);
 540                 return rc;
 541         }
 542
 543         /* Unhook the record from context if encryption is not failure */
 544         ctx->open_rec = NULL;
 545         tls_advance_record_sn(sk, prot, &tls_ctx->tx);
 546         return rc;
 547 }
 548
 549 static int tls_split_open_record(struct sock *sk, struct tls_rec *from,
 550                                  struct tls_rec **to, struct sk_msg *msg_opl,
 551                                  struct sk_msg *msg_oen, u32 split_point,
 552                                  u32 tx_overhead_size, u32 *orig_end)
 553 {
 554         u32 i, j, bytes = 0, apply = msg_opl->apply_bytes;
 555         struct scatterlist *sge, *osge, *nsge;
 556         u32 orig_size = msg_opl->sg.size;
 557         struct scatterlist tmp = { };
 558         struct sk_msg *msg_npl;
 559         struct tls_rec *new;
 560         int ret;
 561
 562         new = tls_get_rec(sk);
 563         if (!new)
 564                 return -ENOMEM;
 565         ret = sk_msg_alloc(sk, &new->msg_encrypted, msg_opl->sg.size +
 566                            tx_overhead_size, 0);
 567         if (ret < 0) {
 568                 tls_free_rec(sk, new);
 569                 return ret;
 570         }
 571
 572         *orig_end = msg_opl->sg.end;
 573         i = msg_opl->sg.start;
 574         sge = sk_msg_elem(msg_opl, i);
 575         while (apply && sge->length) {
 576                 if (sge->length > apply) {
 577                         u32 len = sge->length - apply;
 578
 579                         get_page(sg_page(sge));
 580                         sg_set_page(&tmp, sg_page(sge), len,
 581                                     sge->offset + apply);
 582                         sge->length = apply;
 583                         bytes += apply;
 584                         apply = 0;
 585                 } else {
 586                         apply -= sge->length;
 587                         bytes += sge->length;
 588                 }
 589
 590                 sk_msg_iter_var_next(i);
 591                 if (i == msg_opl->sg.end)
 592                         break;
 593                 sge = sk_msg_elem(msg_opl, i);
 594         }
 595
 596         msg_opl->sg.end = i;
 597         msg_opl->sg.curr = i;
 598         msg_opl->sg.copybreak = 0;
 599         msg_opl->apply_bytes = 0;
 600         msg_opl->sg.size = bytes;
 601
 602         msg_npl = &new->msg_plaintext;
 603         msg_npl->apply_bytes = apply;
 604         msg_npl->sg.size = orig_size - bytes;
 605
 606         j = msg_npl->sg.start;
 607         nsge = sk_msg_elem(msg_npl, j);
 608         if (tmp.length) {
 609                 memcpy(nsge, &tmp, sizeof(*nsge));
 610                 sk_msg_iter_var_next(j);
 611                 nsge = sk_msg_elem(msg_npl, j);
 612         }
 613
 614         osge = sk_msg_elem(msg_opl, i);
 615         while (osge->length) {
 616                 memcpy(nsge, osge, sizeof(*nsge));
 617                 sg_unmark_end(nsge);
 618                 sk_msg_iter_var_next(i);
 619                 sk_msg_iter_var_next(j);
 620                 if (i == *orig_end)
 621                         break;
 622                 osge = sk_msg_elem(msg_opl, i);
 623                 nsge = sk_msg_elem(msg_npl, j);
 624         }
 625
 626         msg_npl->sg.end = j;
 627         msg_npl->sg.curr = j;
 628         msg_npl->sg.copybreak = 0;
 629
 630         *to = new;
 631         return 0;
 632 }
 633
 634 static void tls_merge_open_record(struct sock *sk, struct tls_rec *to,
 635                                   struct tls_rec *from, u32 orig_end)
 636 {
 637         struct sk_msg *msg_npl = &from->msg_plaintext;
 638         struct sk_msg *msg_opl = &to->msg_plaintext;
 639         struct scatterlist *osge, *nsge;
 640         u32 i, j;
 641
 642         i = msg_opl->sg.end;
 643         sk_msg_iter_var_prev(i);
 644         j = msg_npl->sg.start;
 645
 646         osge = sk_msg_elem(msg_opl, i);
 647         nsge = sk_msg_elem(msg_npl, j);
 648
 649         if (sg_page(osge) == sg_page(nsge) &&
 650             osge->offset + osge->length == nsge->offset) {
 651                 osge->length += nsge->length;
 652                 put_page(sg_page(nsge));
 653         }
 654
 655         msg_opl->sg.end = orig_end;
 656         msg_opl->sg.curr = orig_end;
 657         msg_opl->sg.copybreak = 0;
 658         msg_opl->apply_bytes = msg_opl->sg.size + msg_npl->sg.size;
 659         msg_opl->sg.size += msg_npl->sg.size;
 660
 661         sk_msg_free(sk, &to->msg_encrypted);
 662         sk_msg_xfer_full(&to->msg_encrypted, &from->msg_encrypted);
 663
 664         kfree(from);
 665 }
 666
 667 static int tls_push_record(struct sock *sk, int flags,
 668                            unsigned char record_type)
 669 {
 670         struct tls_context *tls_ctx = tls_get_ctx(sk);
 671         struct tls_prot_info *prot = &tls_ctx->prot_info;
 672         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 673         struct tls_rec *rec = ctx->open_rec, *tmp = NULL;
 674         u32 i, split_point, orig_end;
 675         struct sk_msg *msg_pl, *msg_en;
 676         struct aead_request *req;
 677         bool split;
 678         int rc;
 679
 680         if (!rec)
 681                 return 0;
 682
 683         msg_pl = &rec->msg_plaintext;
 684         msg_en = &rec->msg_encrypted;
 685
 686         split_point = msg_pl->apply_bytes;
 687         split = split_point && split_point < msg_pl->sg.size;
 688         if (unlikely((!split &&
 689                       msg_pl->sg.size +
 690                       prot->overhead_size > msg_en->sg.size) ||
 691                      (split &&
 692                       split_point +
 693                       prot->overhead_size > msg_en->sg.size))) {
 694                 split = true;
 695                 split_point = msg_en->sg.size;
 696         }
 697         if (split) {
 698                 rc = tls_split_open_record(sk, rec, &tmp, msg_pl, msg_en,
 699                                            split_point, prot->overhead_size,
 700                                            &orig_end);
 701                 if (rc < 0)
 702                         return rc;
 703                 /* This can happen if above tls_split_open_record allocates
 704                  * a single large encryption buffer instead of two smaller
 705                  * ones. In this case adjust pointers and continue without
 706                  * split.
 707                  */
 708                 if (!msg_pl->sg.size) {
 709                         tls_merge_open_record(sk, rec, tmp, orig_end);
 710                         msg_pl = &rec->msg_plaintext;
 711                         msg_en = &rec->msg_encrypted;
 712                         split = false;
 713                 }
 714                 sk_msg_trim(sk, msg_en, msg_pl->sg.size +
 715                             prot->overhead_size);
 716         }
 717
 718         rec->tx_flags = flags;
 719         req = &rec->aead_req;
 720
 721         i = msg_pl->sg.end;
 722         sk_msg_iter_var_prev(i);
 723
 724         rec->content_type = record_type;
 725         if (prot->version == TLS_1_3_VERSION) {
 726                 /* Add content type to end of message.  No padding added */
 727                 sg_set_buf(&rec->sg_content_type, &rec->content_type, 1);
 728                 sg_mark_end(&rec->sg_content_type);
 729                 sg_chain(msg_pl->sg.data, msg_pl->sg.end + 1,
 730                          &rec->sg_content_type);
 731         } else {
 732                 sg_mark_end(sk_msg_elem(msg_pl, i));
 733         }
 734
 735         if (msg_pl->sg.end < msg_pl->sg.start) {
 736                 sg_chain(&msg_pl->sg.data[msg_pl->sg.start],
 737                          MAX_SKB_FRAGS - msg_pl->sg.start + 1,
 738                          msg_pl->sg.data);
 739         }
 740
 741         i = msg_pl->sg.start;
 742         sg_chain(rec->sg_aead_in, 2, &msg_pl->sg.data[i]);
 743
 744         i = msg_en->sg.end;
 745         sk_msg_iter_var_prev(i);
 746         sg_mark_end(sk_msg_elem(msg_en, i));
 747
 748         i = msg_en->sg.start;
 749         sg_chain(rec->sg_aead_out, 2, &msg_en->sg.data[i]);
 750
 751         tls_make_aad(rec->aad_space, msg_pl->sg.size + prot->tail_size,
 752                      tls_ctx->tx.rec_seq, record_type, prot);
 753
 754         tls_fill_prepend(tls_ctx,
 755                          page_address(sg_page(&msg_en->sg.data[i])) +
 756                          msg_en->sg.data[i].offset,
 757                          msg_pl->sg.size + prot->tail_size,
 758                          record_type);
 759
 760         tls_ctx->pending_open_record_frags = false;
 761
 762         rc = tls_do_encryption(sk, tls_ctx, ctx, req,
 763                                msg_pl->sg.size + prot->tail_size, i);
 764         if (rc < 0) {
 765                 if (rc != -EINPROGRESS) {
 766                         tls_err_abort(sk, -EBADMSG);
 767                         if (split) {
 768                                 tls_ctx->pending_open_record_frags = true;
 769                                 tls_merge_open_record(sk, rec, tmp, orig_end);
 770                         }
 771                 }
 772                 ctx->async_capable = 1;
 773                 return rc;
 774         } else if (split) {
 775                 msg_pl = &tmp->msg_plaintext;
 776                 msg_en = &tmp->msg_encrypted;
 777                 sk_msg_trim(sk, msg_en, msg_pl->sg.size + prot->overhead_size);
 778                 tls_ctx->pending_open_record_frags = true;
 779                 ctx->open_rec = tmp;
 780         }
 781
 782         return tls_tx_records(sk, flags);
 783 }
 784
 785 static int bpf_exec_tx_verdict(struct sk_msg *msg, struct sock *sk,
 786                                bool full_record, u8 record_type,
 787                                ssize_t *copied, int flags)
 788 {
 789         struct tls_context *tls_ctx = tls_get_ctx(sk);
 790         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 791         struct sk_msg msg_redir = { };
 792         struct sk_psock *psock;
 793         struct sock *sk_redir;
 794         struct tls_rec *rec;
 795         bool enospc, policy, redir_ingress;
 796         int err = 0, send;
 797         u32 delta = 0;
 798
 799         policy = !(flags & MSG_SENDPAGE_NOPOLICY);
 800         psock = sk_psock_get(sk);
 801         if (!psock || !policy) {
 802                 err = tls_push_record(sk, flags, record_type);
 803                 if (err && sk->sk_err == EBADMSG) {
 804                         *copied -= sk_msg_free(sk, msg);
 805                         tls_free_open_rec(sk);
 806                         err = -sk->sk_err;
 807                 }
 808                 if (psock)
 809                         sk_psock_put(sk, psock);
 810                 return err;
 811         }
 812 more_data:
 813         enospc = sk_msg_full(msg);
 814         if (psock->eval == __SK_NONE) {
 815                 delta = msg->sg.size;
 816                 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
 817                 delta -= msg->sg.size;
 818         }
 819         if (msg->cork_bytes && msg->cork_bytes > msg->sg.size &&
 820             !enospc && !full_record) {
 821                 err = -ENOSPC;
 822                 goto out_err;
 823         }
 824         msg->cork_bytes = 0;
 825         send = msg->sg.size;
 826         if (msg->apply_bytes && msg->apply_bytes < send)
 827                 send = msg->apply_bytes;
 828
 829         switch (psock->eval) {
 830         case __SK_PASS:
 831                 err = tls_push_record(sk, flags, record_type);
 832                 if (err && sk->sk_err == EBADMSG) {
 833                         *copied -= sk_msg_free(sk, msg);
 834                         tls_free_open_rec(sk);
 835                         err = -sk->sk_err;
 836                         goto out_err;
 837                 }
 838                 break;
 839         case __SK_REDIRECT:
 840                 redir_ingress = psock->redir_ingress;
 841                 sk_redir = psock->sk_redir;
 842                 memcpy(&msg_redir, msg, sizeof(*msg));
 843                 if (msg->apply_bytes < send)
 844                         msg->apply_bytes = 0;
 845                 else
 846                         msg->apply_bytes -= send;
 847                 sk_msg_return_zero(sk, msg, send);
 848                 msg->sg.size -= send;
 849                 release_sock(sk);
 850                 err = tcp_bpf_sendmsg_redir(sk_redir, redir_ingress,
 851                                             &msg_redir, send, flags);
 852                 lock_sock(sk);
 853                 if (err < 0) {
 854                         *copied -= sk_msg_free_nocharge(sk, &msg_redir);
 855                         msg->sg.size = 0;
 856                 }
 857                 if (msg->sg.size == 0)
 858                         tls_free_open_rec(sk);
 859                 break;
 860         case __SK_DROP:
 861         default:
 862                 sk_msg_free_partial(sk, msg, send);
 863                 if (msg->apply_bytes < send)
 864                         msg->apply_bytes = 0;
 865                 else
 866                         msg->apply_bytes -= send;
 867                 if (msg->sg.size == 0)
 868                         tls_free_open_rec(sk);
 869                 *copied -= (send + delta);
 870                 err = -EACCES;
 871         }
 872
 873         if (likely(!err)) {
 874                 bool reset_eval = !ctx->open_rec;
 875
 876                 rec = ctx->open_rec;
 877                 if (rec) {
 878                         msg = &rec->msg_plaintext;
 879                         if (!msg->apply_bytes)
 880                                 reset_eval = true;
 881                 }
 882                 if (reset_eval) {
 883                         psock->eval = __SK_NONE;
 884                         if (psock->sk_redir) {
 885                                 sock_put(psock->sk_redir);
 886                                 psock->sk_redir = NULL;
 887                         }
 888                 }
 889                 if (rec)
 890                         goto more_data;
 891         }
 892  out_err:
 893         sk_psock_put(sk, psock);
 894         return err;
 895 }
 896
 897 static int tls_sw_push_pending_record(struct sock *sk, int flags)
 898 {
 899         struct tls_context *tls_ctx = tls_get_ctx(sk);
 900         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 901         struct tls_rec *rec = ctx->open_rec;
 902         struct sk_msg *msg_pl;
 903         size_t copied;
 904
 905         if (!rec)
 906                 return 0;
 907
 908         msg_pl = &rec->msg_plaintext;
 909         copied = msg_pl->sg.size;
 910         if (!copied)
 911                 return 0;
 912
 913         return bpf_exec_tx_verdict(msg_pl, sk, true, TLS_RECORD_TYPE_DATA,
 914                                    &copied, flags);
 915 }
 916
 917 int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 918 {
 919         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 920         struct tls_context *tls_ctx = tls_get_ctx(sk);
 921         struct tls_prot_info *prot = &tls_ctx->prot_info;
 922         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 923         bool async_capable = ctx->async_capable;
 924         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 925         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
 926         bool eor = !(msg->msg_flags & MSG_MORE);
 927         size_t try_to_copy;
 928         ssize_t copied = 0;
 929         struct sk_msg *msg_pl, *msg_en;
 930         struct tls_rec *rec;
 931         int required_size;
 932         int num_async = 0;
 933         bool full_record;
 934         int record_room;
 935         int num_zc = 0;
 936         int orig_size;
 937         int ret = 0;
 938         int pending;
 939
 940         if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
 941                                MSG_CMSG_COMPAT))
 942                 return -EOPNOTSUPP;
 943
 944         mutex_lock(&tls_ctx->tx_lock);
 945         lock_sock(sk);
 946
 947         if (unlikely(msg->msg_controllen)) {
 948                 ret = tls_process_cmsg(sk, msg, &record_type);
 949                 if (ret) {
 950                         if (ret == -EINPROGRESS)
 951                                 num_async++;
 952                         else if (ret != -EAGAIN)
 953                                 goto send_end;
 954                 }
 955         }
 956
 957         while (msg_data_left(msg)) {
 958                 if (sk->sk_err) {
 959                         ret = -sk->sk_err;
 960                         goto send_end;
 961                 }
 962
 963                 if (ctx->open_rec)
 964                         rec = ctx->open_rec;
 965                 else
 966                         rec = ctx->open_rec = tls_get_rec(sk);
 967                 if (!rec) {
 968                         ret = -ENOMEM;
 969                         goto send_end;
 970                 }
 971
 972                 msg_pl = &rec->msg_plaintext;
 973                 msg_en = &rec->msg_encrypted;
 974
 975                 orig_size = msg_pl->sg.size;
 976                 full_record = false;
 977                 try_to_copy = msg_data_left(msg);
 978                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
 979                 if (try_to_copy >= record_room) {
 980                         try_to_copy = record_room;
 981                         full_record = true;
 982                 }
 983
 984                 required_size = msg_pl->sg.size + try_to_copy +
 985                                 prot->overhead_size;
 986
 987                 if (!sk_stream_memory_free(sk))
 988                         goto wait_for_sndbuf;
 989
 990 alloc_encrypted:
 991                 ret = tls_alloc_encrypted_msg(sk, required_size);
 992                 if (ret) {
 993                         if (ret != -ENOSPC)
 994                                 goto wait_for_memory;
 995
 996                         /* Adjust try_to_copy according to the amount that was
 997                          * actually allocated. The difference is due
 998                          * to max sg elements limit
 999                          */
1000                         try_to_copy -= required_size - msg_en->sg.size;
1001                         full_record = true;
1002                 }
1003
1004                 if (!is_kvec && (full_record || eor) && !async_capable) {
1005                         u32 first = msg_pl->sg.end;
1006
1007                         ret = sk_msg_zerocopy_from_iter(sk, &msg->msg_iter,
1008                                                         msg_pl, try_to_copy);
1009                         if (ret)
1010                                 goto fallback_to_reg_send;
1011
1012                         num_zc++;
1013                         copied += try_to_copy;
1014
1015                         sk_msg_sg_copy_set(msg_pl, first);
1016                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1017                                                   record_type, &copied,
1018                                                   msg->msg_flags);
1019                         if (ret) {
1020                                 if (ret == -EINPROGRESS)
1021                                         num_async++;
1022                                 else if (ret == -ENOMEM)
1023                                         goto wait_for_memory;
1024                                 else if (ctx->open_rec && ret == -ENOSPC)
1025                                         goto rollback_iter;
1026                                 else if (ret != -EAGAIN)
1027                                         goto send_end;
1028                         }
1029                         continue;
1030 rollback_iter:
1031                         copied -= try_to_copy;
1032                         sk_msg_sg_copy_clear(msg_pl, first);
1033                         iov_iter_revert(&msg->msg_iter,
1034                                         msg_pl->sg.size - orig_size);
1035 fallback_to_reg_send:
1036                         sk_msg_trim(sk, msg_pl, orig_size);
1037                 }
1038
1039                 required_size = msg_pl->sg.size + try_to_copy;
1040
1041                 ret = tls_clone_plaintext_msg(sk, required_size);
1042                 if (ret) {
1043                         if (ret != -ENOSPC)
1044                                 goto send_end;
1045
1046                         /* Adjust try_to_copy according to the amount that was
1047                          * actually allocated. The difference is due
1048                          * to max sg elements limit
1049                          */
1050                         try_to_copy -= required_size - msg_pl->sg.size;
1051                         full_record = true;
1052                         sk_msg_trim(sk, msg_en,
1053                                     msg_pl->sg.size + prot->overhead_size);
1054                 }
1055
1056                 if (try_to_copy) {
1057                         ret = sk_msg_memcopy_from_iter(sk, &msg->msg_iter,
1058                                                        msg_pl, try_to_copy);
1059                         if (ret < 0)
1060                                 goto trim_sgl;
1061                 }
1062
1063                 /* Open records defined only if successfully copied, otherwise
1064                  * we would trim the sg but not reset the open record frags.
1065                  */
1066                 tls_ctx->pending_open_record_frags = true;
1067                 copied += try_to_copy;
1068                 if (full_record || eor) {
1069                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1070                                                   record_type, &copied,
1071                                                   msg->msg_flags);
1072                         if (ret) {
1073                                 if (ret == -EINPROGRESS)
1074                                         num_async++;
1075                                 else if (ret == -ENOMEM)
1076                                         goto wait_for_memory;
1077                                 else if (ret != -EAGAIN) {
1078                                         if (ret == -ENOSPC)
1079                                                 ret = 0;
1080                                         goto send_end;
1081                                 }
1082                         }
1083                 }
1084
1085                 continue;
1086
1087 wait_for_sndbuf:
1088                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1089 wait_for_memory:
1090                 ret = sk_stream_wait_memory(sk, &timeo);
1091                 if (ret) {
1092 trim_sgl:
1093                         if (ctx->open_rec)
1094                                 tls_trim_both_msgs(sk, orig_size);
1095                         goto send_end;
1096                 }
1097
1098                 if (ctx->open_rec && msg_en->sg.size < required_size)
1099                         goto alloc_encrypted;
1100         }
1101
1102         if (!num_async) {
1103                 goto send_end;
1104         } else if (num_zc) {
1105                 /* Wait for pending encryptions to get completed */
1106                 spin_lock_bh(&ctx->encrypt_compl_lock);
1107                 ctx->async_notify = true;
1108
1109                 pending = atomic_read(&ctx->encrypt_pending);
1110                 spin_unlock_bh(&ctx->encrypt_compl_lock);
1111                 if (pending)
1112                         crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1113                 else
1114                         reinit_completion(&ctx->async_wait.completion);
1115
1116                 /* There can be no concurrent accesses, since we have no
1117                  * pending encrypt operations
1118                  */
1119                 WRITE_ONCE(ctx->async_notify, false);
1120
1121                 if (ctx->async_wait.err) {
1122                         ret = ctx->async_wait.err;
1123                         copied = 0;
1124                 }
1125         }
1126
1127         /* Transmit if any encryptions have completed */
1128         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1129                 cancel_delayed_work(&ctx->tx_work.work);
1130                 tls_tx_records(sk, msg->msg_flags);
1131         }
1132
1133 send_end:
1134         ret = sk_stream_error(sk, msg->msg_flags, ret);
1135
1136         release_sock(sk);
1137         mutex_unlock(&tls_ctx->tx_lock);
1138         return copied > 0 ? copied : ret;
1139 }
1140
1141 static int tls_sw_do_sendpage(struct sock *sk, struct page *page,
1142                               int offset, size_t size, int flags)
1143 {
1144         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1145         struct tls_context *tls_ctx = tls_get_ctx(sk);
1146         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1147         struct tls_prot_info *prot = &tls_ctx->prot_info;
1148         unsigned char record_type = TLS_RECORD_TYPE_DATA;
1149         struct sk_msg *msg_pl;
1150         struct tls_rec *rec;
1151         int num_async = 0;
1152         ssize_t copied = 0;
1153         bool full_record;
1154         int record_room;
1155         int ret = 0;
1156         bool eor;
1157
1158         eor = !(flags & MSG_SENDPAGE_NOTLAST);
1159         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1160
1161         /* Call the sk_stream functions to manage the sndbuf mem. */
1162         while (size > 0) {
1163                 size_t copy, required_size;
1164
1165                 if (sk->sk_err) {
1166                         ret = -sk->sk_err;
1167                         goto sendpage_end;
1168                 }
1169
1170                 if (ctx->open_rec)
1171                         rec = ctx->open_rec;
1172                 else
1173                         rec = ctx->open_rec = tls_get_rec(sk);
1174                 if (!rec) {
1175                         ret = -ENOMEM;
1176                         goto sendpage_end;
1177                 }
1178
1179                 msg_pl = &rec->msg_plaintext;
1180
1181                 full_record = false;
1182                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
1183                 copy = size;
1184                 if (copy >= record_room) {
1185                         copy = record_room;
1186                         full_record = true;
1187                 }
1188
1189                 required_size = msg_pl->sg.size + copy + prot->overhead_size;
1190
1191                 if (!sk_stream_memory_free(sk))
1192                         goto wait_for_sndbuf;
1193 alloc_payload:
1194                 ret = tls_alloc_encrypted_msg(sk, required_size);
1195                 if (ret) {
1196                         if (ret != -ENOSPC)
1197                                 goto wait_for_memory;
1198
1199                         /* Adjust copy according to the amount that was
1200                          * actually allocated. The difference is due
1201                          * to max sg elements limit
1202                          */
1203                         copy -= required_size - msg_pl->sg.size;
1204                         full_record = true;
1205                 }
1206
1207                 sk_msg_page_add(msg_pl, page, copy, offset);
1208                 sk_mem_charge(sk, copy);
1209
1210                 offset += copy;
1211                 size -= copy;
1212                 copied += copy;
1213
1214                 tls_ctx->pending_open_record_frags = true;
1215                 if (full_record || eor || sk_msg_full(msg_pl)) {
1216                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1217                                                   record_type, &copied, flags);
1218                         if (ret) {
1219                                 if (ret == -EINPROGRESS)
1220                                         num_async++;
1221                                 else if (ret == -ENOMEM)
1222                                         goto wait_for_memory;
1223                                 else if (ret != -EAGAIN) {
1224                                         if (ret == -ENOSPC)
1225                                                 ret = 0;
1226                                         goto sendpage_end;
1227                                 }
1228                         }
1229                 }
1230                 continue;
1231 wait_for_sndbuf:
1232                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1233 wait_for_memory:
1234                 ret = sk_stream_wait_memory(sk, &timeo);
1235                 if (ret) {
1236                         if (ctx->open_rec)
1237                                 tls_trim_both_msgs(sk, msg_pl->sg.size);
1238                         goto sendpage_end;
1239                 }
1240
1241                 if (ctx->open_rec)
1242                         goto alloc_payload;
1243         }
1244
1245         if (num_async) {
1246                 /* Transmit if any encryptions have completed */
1247                 if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1248                         cancel_delayed_work(&ctx->tx_work.work);
1249                         tls_tx_records(sk, flags);
1250                 }
1251         }
1252 sendpage_end:
1253         ret = sk_stream_error(sk, flags, ret);
1254         return copied > 0 ? copied : ret;
1255 }
1256
1257 int tls_sw_sendpage_locked(struct sock *sk, struct page *page,
1258                            int offset, size_t size, int flags)
1259 {
1260         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1261                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY |
1262                       MSG_NO_SHARED_FRAGS))
1263                 return -EOPNOTSUPP;
1264
1265         return tls_sw_do_sendpage(sk, page, offset, size, flags);
1266 }
1267
1268 int tls_sw_sendpage(struct sock *sk, struct page *page,
1269                     int offset, size_t size, int flags)
1270 {
1271         struct tls_context *tls_ctx = tls_get_ctx(sk);
1272         int ret;
1273
1274         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1275                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY))
1276                 return -EOPNOTSUPP;
1277
1278         mutex_lock(&tls_ctx->tx_lock);
1279         lock_sock(sk);
1280         ret = tls_sw_do_sendpage(sk, page, offset, size, flags);
1281         release_sock(sk);
1282         mutex_unlock(&tls_ctx->tx_lock);
1283         return ret;
1284 }
1285
1286 static int
1287 tls_rx_rec_wait(struct sock *sk, struct sk_psock *psock, bool nonblock,
1288                 bool released)
1289 {
1290         struct tls_context *tls_ctx = tls_get_ctx(sk);
1291         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1292         DEFINE_WAIT_FUNC(wait, woken_wake_function);
1293         long timeo;
1294
1295         timeo = sock_rcvtimeo(sk, nonblock);
1296
1297         while (!tls_strp_msg_ready(ctx)) {
1298                 if (!sk_psock_queue_empty(psock))
1299                         return 0;
1300
1301                 if (sk->sk_err)
1302                         return sock_error(sk);
1303
1304                 if (!skb_queue_empty(&sk->sk_receive_queue)) {
1305                         tls_strp_check_rcv(&ctx->strp);
1306                         if (tls_strp_msg_ready(ctx))
1307                                 break;
1308                 }
1309
1310                 if (sk->sk_shutdown & RCV_SHUTDOWN)
1311                         return 0;
1312
1313                 if (sock_flag(sk, SOCK_DONE))
1314                         return 0;
1315
1316                 if (!timeo)
1317                         return -EAGAIN;
1318
1319                 released = true;
1320                 add_wait_queue(sk_sleep(sk), &wait);
1321                 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1322                 sk_wait_event(sk, &timeo,
1323                               tls_strp_msg_ready(ctx) ||
1324                               !sk_psock_queue_empty(psock),
1325                               &wait);
1326                 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1327                 remove_wait_queue(sk_sleep(sk), &wait);
1328
1329                 /* Handle signals */
1330                 if (signal_pending(current))
1331                         return sock_intr_errno(timeo);
1332         }
1333
1334         tls_strp_msg_load(&ctx->strp, released);
1335
1336         return 1;
1337 }
1338
1339 static int tls_setup_from_iter(struct iov_iter *from,
1340                                int length, int *pages_used,
1341                                struct scatterlist *to,
1342                                int to_max_pages)
1343 {
1344         int rc = 0, i = 0, num_elem = *pages_used, maxpages;
1345         struct page *pages[MAX_SKB_FRAGS];
1346         unsigned int size = 0;
1347         ssize_t copied, use;
1348         size_t offset;
1349
1350         while (length > 0) {
1351                 i = 0;
1352                 maxpages = to_max_pages - num_elem;
1353                 if (maxpages == 0) {
1354                         rc = -EFAULT;
1355                         goto out;
1356                 }
1357                 copied = iov_iter_get_pages2(from, pages,
1358                                             length,
1359                                             maxpages, &offset);
1360                 if (copied <= 0) {
1361                         rc = -EFAULT;
1362                         goto out;
1363                 }
1364
1365                 length -= copied;
1366                 size += copied;
1367                 while (copied) {
1368                         use = min_t(int, copied, PAGE_SIZE - offset);
1369
1370                         sg_set_page(&to[num_elem],
1371                                     pages[i], use, offset);
1372                         sg_unmark_end(&to[num_elem]);
1373                         /* We do not uncharge memory from this API */
1374
1375                         offset = 0;
1376                         copied -= use;
1377
1378                         i++;
1379                         num_elem++;
1380                 }
1381         }
1382         /* Mark the end in the last sg entry if newly added */
1383         if (num_elem > *pages_used)
1384                 sg_mark_end(&to[num_elem - 1]);
1385 out:
1386         if (rc)
1387                 iov_iter_revert(from, size);
1388         *pages_used = num_elem;
1389
1390         return rc;
1391 }
1392
1393 static struct sk_buff *
1394 tls_alloc_clrtxt_skb(struct sock *sk, struct sk_buff *skb,
1395                      unsigned int full_len)
1396 {
1397         struct strp_msg *clr_rxm;
1398         struct sk_buff *clr_skb;
1399         int err;
1400
1401         clr_skb = alloc_skb_with_frags(0, full_len, TLS_PAGE_ORDER,
1402                                        &err, sk->sk_allocation);
1403         if (!clr_skb)
1404                 return NULL;
1405
1406         skb_copy_header(clr_skb, skb);
1407         clr_skb->len = full_len;
1408         clr_skb->data_len = full_len;
1409
1410         clr_rxm = strp_msg(clr_skb);
1411         clr_rxm->offset = 0;
1412
1413         return clr_skb;
1414 }
1415
1416 /* Decrypt handlers
1417  *
1418  * tls_decrypt_sw() and tls_decrypt_device() are decrypt handlers.
1419  * They must transform the darg in/out argument are as follows:
1420  *       |          Input            |         Output
1421  * -------------------------------------------------------------------
1422  *    zc | Zero-copy decrypt allowed | Zero-copy performed
1423  * async | Async decrypt allowed     | Async crypto used / in progress
1424  *   skb |            *              | Output skb
1425  *
1426  * If ZC decryption was performed darg.skb will point to the input skb.
1427  */
1428
1429 /* This function decrypts the input skb into either out_iov or in out_sg
1430  * or in skb buffers itself. The input parameter 'darg->zc' indicates if
1431  * zero-copy mode needs to be tried or not. With zero-copy mode, either
1432  * out_iov or out_sg must be non-NULL. In case both out_iov and out_sg are
1433  * NULL, then the decryption happens inside skb buffers itself, i.e.
1434  * zero-copy gets disabled and 'darg->zc' is updated.
1435  */
1436 static int tls_decrypt_sg(struct sock *sk, struct iov_iter *out_iov,
1437                           struct scatterlist *out_sg,
1438                           struct tls_decrypt_arg *darg)
1439 {
1440         struct tls_context *tls_ctx = tls_get_ctx(sk);
1441         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1442         struct tls_prot_info *prot = &tls_ctx->prot_info;
1443         int n_sgin, n_sgout, aead_size, err, pages = 0;
1444         struct sk_buff *skb = tls_strp_msg(ctx);
1445         const struct strp_msg *rxm = strp_msg(skb);
1446         const struct tls_msg *tlm = tls_msg(skb);
1447         struct aead_request *aead_req;
1448         struct scatterlist *sgin = NULL;
1449         struct scatterlist *sgout = NULL;
1450         const int data_len = rxm->full_len - prot->overhead_size;
1451         int tail_pages = !!prot->tail_size;
1452         struct tls_decrypt_ctx *dctx;
1453         struct sk_buff *clear_skb;
1454         int iv_offset = 0;
1455         u8 *mem;
1456
1457         n_sgin = skb_nsg(skb, rxm->offset + prot->prepend_size,
1458                          rxm->full_len - prot->prepend_size);
1459         if (n_sgin < 1)
1460                 return n_sgin ?: -EBADMSG;
1461
1462         if (darg->zc && (out_iov || out_sg)) {
1463                 clear_skb = NULL;
1464
1465                 if (out_iov)
1466                         n_sgout = 1 + tail_pages +
1467                                 iov_iter_npages_cap(out_iov, INT_MAX, data_len);
1468                 else
1469                         n_sgout = sg_nents(out_sg);
1470         } else {
1471                 darg->zc = false;
1472
1473                 clear_skb = tls_alloc_clrtxt_skb(sk, skb, rxm->full_len);
1474                 if (!clear_skb)
1475                         return -ENOMEM;
1476
1477                 n_sgout = 1 + skb_shinfo(clear_skb)->nr_frags;
1478         }
1479
1480         /* Increment to accommodate AAD */
1481         n_sgin = n_sgin + 1;
1482
1483         /* Allocate a single block of memory which contains
1484          *   aead_req || tls_decrypt_ctx.
1485          * Both structs are variable length.
1486          */
1487         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(ctx->aead_recv);
1488         mem = kmalloc(aead_size + struct_size(dctx, sg, n_sgin + n_sgout),
1489                       sk->sk_allocation);
1490         if (!mem) {
1491                 err = -ENOMEM;
1492                 goto exit_free_skb;
1493         }
1494
1495         /* Segment the allocated memory */
1496         aead_req = (struct aead_request *)mem;
1497         dctx = (struct tls_decrypt_ctx *)(mem + aead_size);
1498         sgin = &dctx->sg[0];
1499         sgout = &dctx->sg[n_sgin];
1500
1501         /* For CCM based ciphers, first byte of nonce+iv is a constant */
1502         switch (prot->cipher_type) {
1503         case TLS_CIPHER_AES_CCM_128:
1504                 dctx->iv[0] = TLS_AES_CCM_IV_B0_BYTE;
1505                 iv_offset = 1;
1506                 break;
1507         case TLS_CIPHER_SM4_CCM:
1508                 dctx->iv[0] = TLS_SM4_CCM_IV_B0_BYTE;
1509                 iv_offset = 1;
1510                 break;
1511         }
1512
1513         /* Prepare IV */
1514         if (prot->version == TLS_1_3_VERSION ||
1515             prot->cipher_type == TLS_CIPHER_CHACHA20_POLY1305) {
1516                 memcpy(&dctx->iv[iv_offset], tls_ctx->rx.iv,
1517                        prot->iv_size + prot->salt_size);
1518         } else {
1519                 err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
1520                                     &dctx->iv[iv_offset] + prot->salt_size,
1521                                     prot->iv_size);
1522                 if (err < 0)
1523                         goto exit_free;
1524                 memcpy(&dctx->iv[iv_offset], tls_ctx->rx.iv, prot->salt_size);
1525         }
1526         tls_xor_iv_with_seq(prot, &dctx->iv[iv_offset], tls_ctx->rx.rec_seq);
1527
1528         /* Prepare AAD */
1529         tls_make_aad(dctx->aad, rxm->full_len - prot->overhead_size +
1530                      prot->tail_size,
1531                      tls_ctx->rx.rec_seq, tlm->control, prot);
1532
1533         /* Prepare sgin */
1534         sg_init_table(sgin, n_sgin);
1535         sg_set_buf(&sgin[0], dctx->aad, prot->aad_size);
1536         err = skb_to_sgvec(skb, &sgin[1],
1537                            rxm->offset + prot->prepend_size,
1538                            rxm->full_len - prot->prepend_size);
1539         if (err < 0)
1540                 goto exit_free;
1541
1542         if (clear_skb) {
1543                 sg_init_table(sgout, n_sgout);
1544                 sg_set_buf(&sgout[0], dctx->aad, prot->aad_size);
1545
1546                 err = skb_to_sgvec(clear_skb, &sgout[1], prot->prepend_size,
1547                                    data_len + prot->tail_size);
1548                 if (err < 0)
1549                         goto exit_free;
1550         } else if (out_iov) {
1551                 sg_init_table(sgout, n_sgout);
1552                 sg_set_buf(&sgout[0], dctx->aad, prot->aad_size);
1553
1554                 err = tls_setup_from_iter(out_iov, data_len, &pages, &sgout[1],
1555                                           (n_sgout - 1 - tail_pages));
1556                 if (err < 0)
1557                         goto exit_free_pages;
1558
1559                 if (prot->tail_size) {
1560                         sg_unmark_end(&sgout[pages]);
1561                         sg_set_buf(&sgout[pages + 1], &dctx->tail,
1562                                    prot->tail_size);
1563                         sg_mark_end(&sgout[pages + 1]);
1564                 }
1565         } else if (out_sg) {
1566                 memcpy(sgout, out_sg, n_sgout * sizeof(*sgout));
1567         }
1568
1569         /* Prepare and submit AEAD request */
1570         err = tls_do_decryption(sk, sgin, sgout, dctx->iv,
1571                                 data_len + prot->tail_size, aead_req, darg);
1572         if (err)
1573                 goto exit_free_pages;
1574
1575         darg->skb = clear_skb ?: tls_strp_msg(ctx);
1576         clear_skb = NULL;
1577
1578         if (unlikely(darg->async)) {
1579                 err = tls_strp_msg_hold(&ctx->strp, &ctx->async_hold);
1580                 if (err)
1581                         __skb_queue_tail(&ctx->async_hold, darg->skb);
1582                 return err;
1583         }
1584
1585         if (prot->tail_size)
1586                 darg->tail = dctx->tail;
1587
1588 exit_free_pages:
1589         /* Release the pages in case iov was mapped to pages */
1590         for (; pages > 0; pages--)
1591                 put_page(sg_page(&sgout[pages]));
1592 exit_free:
1593         kfree(mem);
1594 exit_free_skb:
1595         consume_skb(clear_skb);
1596         return err;
1597 }
1598
1599 static int
1600 tls_decrypt_sw(struct sock *sk, struct tls_context *tls_ctx,
1601                struct msghdr *msg, struct tls_decrypt_arg *darg)
1602 {
1603         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1604         struct tls_prot_info *prot = &tls_ctx->prot_info;
1605         struct strp_msg *rxm;
1606         int pad, err;
1607
1608         err = tls_decrypt_sg(sk, &msg->msg_iter, NULL, darg);
1609         if (err < 0) {
1610                 if (err == -EBADMSG)
1611                         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSDECRYPTERROR);
1612                 return err;
1613         }
1614         /* keep going even for ->async, the code below is TLS 1.3 */
1615
1616         /* If opportunistic TLS 1.3 ZC failed retry without ZC */
1617         if (unlikely(darg->zc && prot->version == TLS_1_3_VERSION &&
1618                      darg->tail != TLS_RECORD_TYPE_DATA)) {
1619                 darg->zc = false;
1620                 if (!darg->tail)
1621                         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSRXNOPADVIOL);
1622                 TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSDECRYPTRETRY);
1623                 return tls_decrypt_sw(sk, tls_ctx, msg, darg);
1624         }
1625
1626         pad = tls_padding_length(prot, darg->skb, darg);
1627         if (pad < 0) {
1628                 if (darg->skb != tls_strp_msg(ctx))
1629                         consume_skb(darg->skb);
1630                 return pad;
1631         }
1632
1633         rxm = strp_msg(darg->skb);
1634         rxm->full_len -= pad;
1635
1636         return 0;
1637 }
1638
1639 static int
1640 tls_decrypt_device(struct sock *sk, struct msghdr *msg,
1641                    struct tls_context *tls_ctx, struct tls_decrypt_arg *darg)
1642 {
1643         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1644         struct tls_prot_info *prot = &tls_ctx->prot_info;
1645         struct strp_msg *rxm;
1646         int pad, err;
1647
1648         if (tls_ctx->rx_conf != TLS_HW)
1649                 return 0;
1650
1651         err = tls_device_decrypted(sk, tls_ctx);
1652         if (err <= 0)
1653                 return err;
1654
1655         pad = tls_padding_length(prot, tls_strp_msg(ctx), darg);
1656         if (pad < 0)
1657                 return pad;
1658
1659         darg->async = false;
1660         darg->skb = tls_strp_msg(ctx);
1661         /* ->zc downgrade check, in case TLS 1.3 gets here */
1662         darg->zc &= !(prot->version == TLS_1_3_VERSION &&
1663                       tls_msg(darg->skb)->control != TLS_RECORD_TYPE_DATA);
1664
1665         rxm = strp_msg(darg->skb);
1666         rxm->full_len -= pad;
1667
1668         if (!darg->zc) {
1669                 /* Non-ZC case needs a real skb */
1670                 darg->skb = tls_strp_msg_detach(ctx);
1671                 if (!darg->skb)
1672                         return -ENOMEM;
1673         } else {
1674                 unsigned int off, len;
1675
1676                 /* In ZC case nobody cares about the output skb.
1677                  * Just copy the data here. Note the skb is not fully trimmed.
1678                  */
1679                 off = rxm->offset + prot->prepend_size;
1680                 len = rxm->full_len - prot->overhead_size;
1681
1682                 err = skb_copy_datagram_msg(darg->skb, off, msg, len);
1683                 if (err)
1684                         return err;
1685         }
1686         return 1;
1687 }
1688
1689 static int tls_rx_one_record(struct sock *sk, struct msghdr *msg,
1690                              struct tls_decrypt_arg *darg)
1691 {
1692         struct tls_context *tls_ctx = tls_get_ctx(sk);
1693         struct tls_prot_info *prot = &tls_ctx->prot_info;
1694         struct strp_msg *rxm;
1695         int err;
1696
1697         err = tls_decrypt_device(sk, msg, tls_ctx, darg);
1698         if (!err)
1699                 err = tls_decrypt_sw(sk, tls_ctx, msg, darg);
1700         if (err < 0)
1701                 return err;
1702
1703         rxm = strp_msg(darg->skb);
1704         rxm->offset += prot->prepend_size;
1705         rxm->full_len -= prot->overhead_size;
1706         tls_advance_record_sn(sk, prot, &tls_ctx->rx);
1707
1708         return 0;
1709 }
1710
1711 int decrypt_skb(struct sock *sk, struct scatterlist *sgout)
1712 {
1713         struct tls_decrypt_arg darg = { .zc = true, };
1714
1715         return tls_decrypt_sg(sk, NULL, sgout, &darg);
1716 }
1717
1718 static int tls_record_content_type(struct msghdr *msg, struct tls_msg *tlm,
1719                                    u8 *control)
1720 {
1721         int err;
1722
1723         if (!*control) {
1724                 *control = tlm->control;
1725                 if (!*control)
1726                         return -EBADMSG;
1727
1728                 err = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1729                                sizeof(*control), control);
1730                 if (*control != TLS_RECORD_TYPE_DATA) {
1731                         if (err || msg->msg_flags & MSG_CTRUNC)
1732                                 return -EIO;
1733                 }
1734         } else if (*control != tlm->control) {
1735                 return 0;
1736         }
1737
1738         return 1;
1739 }
1740
1741 static void tls_rx_rec_done(struct tls_sw_context_rx *ctx)
1742 {
1743         tls_strp_msg_done(&ctx->strp);
1744 }
1745
1746 /* This function traverses the rx_list in tls receive context to copies the
1747  * decrypted records into the buffer provided by caller zero copy is not
1748  * true. Further, the records are removed from the rx_list if it is not a peek
1749  * case and the record has been consumed completely.
1750  */
1751 static int process_rx_list(struct tls_sw_context_rx *ctx,
1752                            struct msghdr *msg,
1753                            u8 *control,
1754                            size_t skip,
1755                            size_t len,
1756                            bool is_peek)
1757 {
1758         struct sk_buff *skb = skb_peek(&ctx->rx_list);
1759         struct tls_msg *tlm;
1760         ssize_t copied = 0;
1761         int err;
1762
1763         while (skip && skb) {
1764                 struct strp_msg *rxm = strp_msg(skb);
1765                 tlm = tls_msg(skb);
1766
1767                 err = tls_record_content_type(msg, tlm, control);
1768                 if (err <= 0)
1769                         goto out;
1770
1771                 if (skip < rxm->full_len)
1772                         break;
1773
1774                 skip = skip - rxm->full_len;
1775                 skb = skb_peek_next(skb, &ctx->rx_list);
1776         }
1777
1778         while (len && skb) {
1779                 struct sk_buff *next_skb;
1780                 struct strp_msg *rxm = strp_msg(skb);
1781                 int chunk = min_t(unsigned int, rxm->full_len - skip, len);
1782
1783                 tlm = tls_msg(skb);
1784
1785                 err = tls_record_content_type(msg, tlm, control);
1786                 if (err <= 0)
1787                         goto out;
1788
1789                 err = skb_copy_datagram_msg(skb, rxm->offset + skip,
1790                                             msg, chunk);
1791                 if (err < 0)
1792                         goto out;
1793
1794                 len = len - chunk;
1795                 copied = copied + chunk;
1796
1797                 /* Consume the data from record if it is non-peek case*/
1798                 if (!is_peek) {
1799                         rxm->offset = rxm->offset + chunk;
1800                         rxm->full_len = rxm->full_len - chunk;
1801
1802                         /* Return if there is unconsumed data in the record */
1803                         if (rxm->full_len - skip)
1804                                 break;
1805                 }
1806
1807                 /* The remaining skip-bytes must lie in 1st record in rx_list.
1808                  * So from the 2nd record, 'skip' should be 0.
1809                  */
1810                 skip = 0;
1811
1812                 if (msg)
1813                         msg->msg_flags |= MSG_EOR;
1814
1815                 next_skb = skb_peek_next(skb, &ctx->rx_list);
1816
1817                 if (!is_peek) {
1818                         __skb_unlink(skb, &ctx->rx_list);
1819                         consume_skb(skb);
1820                 }
1821
1822                 skb = next_skb;
1823         }
1824         err = 0;
1825
1826 out:
1827         return copied ? : err;
1828 }
1829
1830 static bool
1831 tls_read_flush_backlog(struct sock *sk, struct tls_prot_info *prot,
1832                        size_t len_left, size_t decrypted, ssize_t done,
1833                        size_t *flushed_at)
1834 {
1835         size_t max_rec;
1836
1837         if (len_left <= decrypted)
1838                 return false;
1839
1840         max_rec = prot->overhead_size - prot->tail_size + TLS_MAX_PAYLOAD_SIZE;
1841         if (done - *flushed_at < SZ_128K && tcp_inq(sk) > max_rec)
1842                 return false;
1843
1844         *flushed_at = done;
1845         return sk_flush_backlog(sk);
1846 }
1847
1848 static int tls_rx_reader_lock(struct sock *sk, struct tls_sw_context_rx *ctx,
1849                               bool nonblock)
1850 {
1851         long timeo;
1852         int err;
1853
1854         lock_sock(sk);
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                         err = -EAGAIN;
1870                         goto err_unlock;
1871                 }
1872                 if (signal_pending(current)) {
1873                         err = sock_intr_errno(timeo);
1874                         goto err_unlock;
1875                 }
1876         }
1877
1878         WRITE_ONCE(ctx->reader_present, 1);
1879
1880         return 0;
1881
1882 err_unlock:
1883         release_sock(sk);
1884         return err;
1885 }
1886
1887 static void tls_rx_reader_unlock(struct sock *sk, struct tls_sw_context_rx *ctx)
1888 {
1889         if (unlikely(ctx->reader_contended)) {
1890                 if (wq_has_sleeper(&ctx->wq))
1891                         wake_up(&ctx->wq);
1892                 else
1893                         ctx->reader_contended = 0;
1894
1895                 WARN_ON_ONCE(!ctx->reader_present);
1896         }
1897
1898         WRITE_ONCE(ctx->reader_present, 0);
1899         release_sock(sk);
1900 }
1901
1902 int tls_sw_recvmsg(struct sock *sk,
1903                    struct msghdr *msg,
1904                    size_t len,
1905                    int flags,
1906                    int *addr_len)
1907 {
1908         struct tls_context *tls_ctx = tls_get_ctx(sk);
1909         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1910         struct tls_prot_info *prot = &tls_ctx->prot_info;
1911         ssize_t decrypted = 0, async_copy_bytes = 0;
1912         struct sk_psock *psock;
1913         unsigned char control = 0;
1914         size_t flushed_at = 0;
1915         struct strp_msg *rxm;
1916         struct tls_msg *tlm;
1917         ssize_t copied = 0;
1918         bool async = false;
1919         int target, err;
1920         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
1921         bool is_peek = flags & MSG_PEEK;
1922         bool released = true;
1923         bool bpf_strp_enabled;
1924         bool zc_capable;
1925
1926         if (unlikely(flags & MSG_ERRQUEUE))
1927                 return sock_recv_errqueue(sk, msg, len, SOL_IP, IP_RECVERR);
1928
1929         psock = sk_psock_get(sk);
1930         err = tls_rx_reader_lock(sk, ctx, flags & MSG_DONTWAIT);
1931         if (err < 0)
1932                 return err;
1933         bpf_strp_enabled = sk_psock_strp_enabled(psock);
1934
1935         /* If crypto failed the connection is broken */
1936         err = ctx->async_wait.err;
1937         if (err)
1938                 goto end;
1939
1940         /* Process pending decrypted records. It must be non-zero-copy */
1941         err = process_rx_list(ctx, msg, &control, 0, len, is_peek);
1942         if (err < 0)
1943                 goto end;
1944
1945         copied = err;
1946         if (len <= copied)
1947                 goto end;
1948
1949         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1950         len = len - copied;
1951
1952         zc_capable = !bpf_strp_enabled && !is_kvec && !is_peek &&
1953                 ctx->zc_capable;
1954         decrypted = 0;
1955         while (len && (decrypted + copied < target || tls_strp_msg_ready(ctx))) {
1956                 struct tls_decrypt_arg darg;
1957                 int to_decrypt, chunk;
1958
1959                 err = tls_rx_rec_wait(sk, psock, flags & MSG_DONTWAIT,
1960                                       released);
1961                 if (err <= 0) {
1962                         if (psock) {
1963                                 chunk = sk_msg_recvmsg(sk, psock, msg, len,
1964                                                        flags);
1965                                 if (chunk > 0) {
1966                                         decrypted += chunk;
1967                                         len -= chunk;
1968                                         continue;
1969                                 }
1970                         }
1971                         goto recv_end;
1972                 }
1973
1974                 memset(&darg.inargs, 0, sizeof(darg.inargs));
1975
1976                 rxm = strp_msg(tls_strp_msg(ctx));
1977                 tlm = tls_msg(tls_strp_msg(ctx));
1978
1979                 to_decrypt = rxm->full_len - prot->overhead_size;
1980
1981                 if (zc_capable && to_decrypt <= len &&
1982                     tlm->control == TLS_RECORD_TYPE_DATA)
1983                         darg.zc = true;
1984
1985                 /* Do not use async mode if record is non-data */
1986                 if (tlm->control == TLS_RECORD_TYPE_DATA && !bpf_strp_enabled)
1987                         darg.async = ctx->async_capable;
1988                 else
1989                         darg.async = false;
1990
1991                 err = tls_rx_one_record(sk, msg, &darg);
1992                 if (err < 0) {
1993                         tls_err_abort(sk, -EBADMSG);
1994                         goto recv_end;
1995                 }
1996
1997                 async |= darg.async;
1998
1999                 /* If the type of records being processed is not known yet,
2000                  * set it to record type just dequeued. If it is already known,
2001                  * but does not match the record type just dequeued, go to end.
2002                  * We always get record type here since for tls1.2, record type
2003                  * is known just after record is dequeued from stream parser.
2004                  * For tls1.3, we disable async.
2005                  */
2006                 err = tls_record_content_type(msg, tls_msg(darg.skb), &control);
2007                 if (err <= 0) {
2008                         DEBUG_NET_WARN_ON_ONCE(darg.zc);
2009                         tls_rx_rec_done(ctx);
2010 put_on_rx_list_err:
2011                         __skb_queue_tail(&ctx->rx_list, darg.skb);
2012                         goto recv_end;
2013                 }
2014
2015                 /* periodically flush backlog, and feed strparser */
2016                 released = tls_read_flush_backlog(sk, prot, len, to_decrypt,
2017                                                   decrypted + copied,
2018                                                   &flushed_at);
2019
2020                 /* TLS 1.3 may have updated the length by more than overhead */
2021                 rxm = strp_msg(darg.skb);
2022                 chunk = rxm->full_len;
2023                 tls_rx_rec_done(ctx);
2024
2025                 if (!darg.zc) {
2026                         bool partially_consumed = chunk > len;
2027                         struct sk_buff *skb = darg.skb;
2028
2029                         DEBUG_NET_WARN_ON_ONCE(darg.skb == ctx->strp.anchor);
2030
2031                         if (async) {
2032                                 /* TLS 1.2-only, to_decrypt must be text len */
2033                                 chunk = min_t(int, to_decrypt, len);
2034                                 async_copy_bytes += chunk;
2035 put_on_rx_list:
2036                                 decrypted += chunk;
2037                                 len -= chunk;
2038                                 __skb_queue_tail(&ctx->rx_list, skb);
2039                                 continue;
2040                         }
2041
2042                         if (bpf_strp_enabled) {
2043                                 released = true;
2044                                 err = sk_psock_tls_strp_read(psock, skb);
2045                                 if (err != __SK_PASS) {
2046                                         rxm->offset = rxm->offset + rxm->full_len;
2047                                         rxm->full_len = 0;
2048                                         if (err == __SK_DROP)
2049                                                 consume_skb(skb);
2050                                         continue;
2051                                 }
2052                         }
2053
2054                         if (partially_consumed)
2055                                 chunk = len;
2056
2057                         err = skb_copy_datagram_msg(skb, rxm->offset,
2058                                                     msg, chunk);
2059                         if (err < 0)
2060                                 goto put_on_rx_list_err;
2061
2062                         if (is_peek)
2063                                 goto put_on_rx_list;
2064
2065                         if (partially_consumed) {
2066                                 rxm->offset += chunk;
2067                                 rxm->full_len -= chunk;
2068                                 goto put_on_rx_list;
2069                         }
2070
2071                         consume_skb(skb);
2072                 }
2073
2074                 decrypted += chunk;
2075                 len -= chunk;
2076
2077                 /* Return full control message to userspace before trying
2078                  * to parse another message type
2079                  */
2080                 msg->msg_flags |= MSG_EOR;
2081                 if (control != TLS_RECORD_TYPE_DATA)
2082                         break;
2083         }
2084
2085 recv_end:
2086         if (async) {
2087                 int ret, pending;
2088
2089                 /* Wait for all previously submitted records to be decrypted */
2090                 spin_lock_bh(&ctx->decrypt_compl_lock);
2091                 reinit_completion(&ctx->async_wait.completion);
2092                 pending = atomic_read(&ctx->decrypt_pending);
2093                 spin_unlock_bh(&ctx->decrypt_compl_lock);
2094                 ret = 0;
2095                 if (pending)
2096                         ret = crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
2097                 __skb_queue_purge(&ctx->async_hold);
2098
2099                 if (ret) {
2100                         if (err >= 0 || err == -EINPROGRESS)
2101                                 err = ret;
2102                         decrypted = 0;
2103                         goto end;
2104                 }
2105
2106                 /* Drain records from the rx_list & copy if required */
2107                 if (is_peek || is_kvec)
2108                         err = process_rx_list(ctx, msg, &control, copied,
2109                                               decrypted, is_peek);
2110                 else
2111                         err = process_rx_list(ctx, msg, &control, 0,
2112                                               async_copy_bytes, is_peek);
2113                 decrypted = max(err, 0);
2114         }
2115
2116         copied += decrypted;
2117
2118 end:
2119         tls_rx_reader_unlock(sk, ctx);
2120         if (psock)
2121                 sk_psock_put(sk, psock);
2122         return copied ? : err;
2123 }
2124
2125 ssize_t tls_sw_splice_read(struct socket *sock,  loff_t *ppos,
2126                            struct pipe_inode_info *pipe,
2127                            size_t len, unsigned int flags)
2128 {
2129         struct tls_context *tls_ctx = tls_get_ctx(sock->sk);
2130         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2131         struct strp_msg *rxm = NULL;
2132         struct sock *sk = sock->sk;
2133         struct tls_msg *tlm;
2134         struct sk_buff *skb;
2135         ssize_t copied = 0;
2136         int chunk;
2137         int err;
2138
2139         err = tls_rx_reader_lock(sk, ctx, flags & SPLICE_F_NONBLOCK);
2140         if (err < 0)
2141                 return err;
2142
2143         if (!skb_queue_empty(&ctx->rx_list)) {
2144                 skb = __skb_dequeue(&ctx->rx_list);
2145         } else {
2146                 struct tls_decrypt_arg darg;
2147
2148                 err = tls_rx_rec_wait(sk, NULL, flags & SPLICE_F_NONBLOCK,
2149                                       true);
2150                 if (err <= 0)
2151                         goto splice_read_end;
2152
2153                 memset(&darg.inargs, 0, sizeof(darg.inargs));
2154
2155                 err = tls_rx_one_record(sk, NULL, &darg);
2156                 if (err < 0) {
2157                         tls_err_abort(sk, -EBADMSG);
2158                         goto splice_read_end;
2159                 }
2160
2161                 tls_rx_rec_done(ctx);
2162                 skb = darg.skb;
2163         }
2164
2165         rxm = strp_msg(skb);
2166         tlm = tls_msg(skb);
2167
2168         /* splice does not support reading control messages */
2169         if (tlm->control != TLS_RECORD_TYPE_DATA) {
2170                 err = -EINVAL;
2171                 goto splice_requeue;
2172         }
2173
2174         chunk = min_t(unsigned int, rxm->full_len, len);
2175         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, chunk, flags);
2176         if (copied < 0)
2177                 goto splice_requeue;
2178
2179         if (chunk < rxm->full_len) {
2180                 rxm->offset += len;
2181                 rxm->full_len -= len;
2182                 goto splice_requeue;
2183         }
2184
2185         consume_skb(skb);
2186
2187 splice_read_end:
2188         tls_rx_reader_unlock(sk, ctx);
2189         return copied ? : err;
2190
2191 splice_requeue:
2192         __skb_queue_head(&ctx->rx_list, skb);
2193         goto splice_read_end;
2194 }
2195
2196 bool tls_sw_sock_is_readable(struct sock *sk)
2197 {
2198         struct tls_context *tls_ctx = tls_get_ctx(sk);
2199         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2200         bool ingress_empty = true;
2201         struct sk_psock *psock;
2202
2203         rcu_read_lock();
2204         psock = sk_psock(sk);
2205         if (psock)
2206                 ingress_empty = list_empty(&psock->ingress_msg);
2207         rcu_read_unlock();
2208
2209         return !ingress_empty || tls_strp_msg_ready(ctx) ||
2210                 !skb_queue_empty(&ctx->rx_list);
2211 }
2212
2213 int tls_rx_msg_size(struct tls_strparser *strp, struct sk_buff *skb)
2214 {
2215         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2216         struct tls_prot_info *prot = &tls_ctx->prot_info;
2217         char header[TLS_HEADER_SIZE + MAX_IV_SIZE];
2218         size_t cipher_overhead;
2219         size_t data_len = 0;
2220         int ret;
2221
2222         /* Verify that we have a full TLS header, or wait for more data */
2223         if (strp->stm.offset + prot->prepend_size > skb->len)
2224                 return 0;
2225
2226         /* Sanity-check size of on-stack buffer. */
2227         if (WARN_ON(prot->prepend_size > sizeof(header))) {
2228                 ret = -EINVAL;
2229                 goto read_failure;
2230         }
2231
2232         /* Linearize header to local buffer */
2233         ret = skb_copy_bits(skb, strp->stm.offset, header, prot->prepend_size);
2234         if (ret < 0)
2235                 goto read_failure;
2236
2237         strp->mark = header[0];
2238
2239         data_len = ((header[4] & 0xFF) | (header[3] << 8));
2240
2241         cipher_overhead = prot->tag_size;
2242         if (prot->version != TLS_1_3_VERSION &&
2243             prot->cipher_type != TLS_CIPHER_CHACHA20_POLY1305)
2244                 cipher_overhead += prot->iv_size;
2245
2246         if (data_len > TLS_MAX_PAYLOAD_SIZE + cipher_overhead +
2247             prot->tail_size) {
2248                 ret = -EMSGSIZE;
2249                 goto read_failure;
2250         }
2251         if (data_len < cipher_overhead) {
2252                 ret = -EBADMSG;
2253                 goto read_failure;
2254         }
2255
2256         /* Note that both TLS1.3 and TLS1.2 use TLS_1_2 version here */
2257         if (header[1] != TLS_1_2_VERSION_MINOR ||
2258             header[2] != TLS_1_2_VERSION_MAJOR) {
2259                 ret = -EINVAL;
2260                 goto read_failure;
2261         }
2262
2263         tls_device_rx_resync_new_rec(strp->sk, data_len + TLS_HEADER_SIZE,
2264                                      TCP_SKB_CB(skb)->seq + strp->stm.offset);
2265         return data_len + TLS_HEADER_SIZE;
2266
2267 read_failure:
2268         tls_err_abort(strp->sk, ret);
2269
2270         return ret;
2271 }
2272
2273 void tls_rx_msg_ready(struct tls_strparser *strp)
2274 {
2275         struct tls_sw_context_rx *ctx;
2276
2277         ctx = container_of(strp, struct tls_sw_context_rx, strp);
2278         ctx->saved_data_ready(strp->sk);
2279 }
2280
2281 static void tls_data_ready(struct sock *sk)
2282 {
2283         struct tls_context *tls_ctx = tls_get_ctx(sk);
2284         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2285         struct sk_psock *psock;
2286
2287         tls_strp_data_ready(&ctx->strp);
2288
2289         psock = sk_psock_get(sk);
2290         if (psock) {
2291                 if (!list_empty(&psock->ingress_msg))
2292                         ctx->saved_data_ready(sk);
2293                 sk_psock_put(sk, psock);
2294         }
2295 }
2296
2297 void tls_sw_cancel_work_tx(struct tls_context *tls_ctx)
2298 {
2299         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2300
2301         set_bit(BIT_TX_CLOSING, &ctx->tx_bitmask);
2302         set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask);
2303         cancel_delayed_work_sync(&ctx->tx_work.work);
2304 }
2305
2306 void tls_sw_release_resources_tx(struct sock *sk)
2307 {
2308         struct tls_context *tls_ctx = tls_get_ctx(sk);
2309         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2310         struct tls_rec *rec, *tmp;
2311         int pending;
2312
2313         /* Wait for any pending async encryptions to complete */
2314         spin_lock_bh(&ctx->encrypt_compl_lock);
2315         ctx->async_notify = true;
2316         pending = atomic_read(&ctx->encrypt_pending);
2317         spin_unlock_bh(&ctx->encrypt_compl_lock);
2318
2319         if (pending)
2320                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
2321
2322         tls_tx_records(sk, -1);
2323
2324         /* Free up un-sent records in tx_list. First, free
2325          * the partially sent record if any at head of tx_list.
2326          */
2327         if (tls_ctx->partially_sent_record) {
2328                 tls_free_partial_record(sk, tls_ctx);
2329                 rec = list_first_entry(&ctx->tx_list,
2330                                        struct tls_rec, list);
2331                 list_del(&rec->list);
2332                 sk_msg_free(sk, &rec->msg_plaintext);
2333                 kfree(rec);
2334         }
2335
2336         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
2337                 list_del(&rec->list);
2338                 sk_msg_free(sk, &rec->msg_encrypted);
2339                 sk_msg_free(sk, &rec->msg_plaintext);
2340                 kfree(rec);
2341         }
2342
2343         crypto_free_aead(ctx->aead_send);
2344         tls_free_open_rec(sk);
2345 }
2346
2347 void tls_sw_free_ctx_tx(struct tls_context *tls_ctx)
2348 {
2349         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2350
2351         kfree(ctx);
2352 }
2353
2354 void tls_sw_release_resources_rx(struct sock *sk)
2355 {
2356         struct tls_context *tls_ctx = tls_get_ctx(sk);
2357         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2358
2359         kfree(tls_ctx->rx.rec_seq);
2360         kfree(tls_ctx->rx.iv);
2361
2362         if (ctx->aead_recv) {
2363                 __skb_queue_purge(&ctx->rx_list);
2364                 crypto_free_aead(ctx->aead_recv);
2365                 tls_strp_stop(&ctx->strp);
2366                 /* If tls_sw_strparser_arm() was not called (cleanup paths)
2367                  * we still want to tls_strp_stop(), but sk->sk_data_ready was
2368                  * never swapped.
2369                  */
2370                 if (ctx->saved_data_ready) {
2371                         write_lock_bh(&sk->sk_callback_lock);
2372                         sk->sk_data_ready = ctx->saved_data_ready;
2373                         write_unlock_bh(&sk->sk_callback_lock);
2374                 }
2375         }
2376 }
2377
2378 void tls_sw_strparser_done(struct tls_context *tls_ctx)
2379 {
2380         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2381
2382         tls_strp_done(&ctx->strp);
2383 }
2384
2385 void tls_sw_free_ctx_rx(struct tls_context *tls_ctx)
2386 {
2387         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2388
2389         kfree(ctx);
2390 }
2391
2392 void tls_sw_free_resources_rx(struct sock *sk)
2393 {
2394         struct tls_context *tls_ctx = tls_get_ctx(sk);
2395
2396         tls_sw_release_resources_rx(sk);
2397         tls_sw_free_ctx_rx(tls_ctx);
2398 }
2399
2400 /* The work handler to transmitt the encrypted records in tx_list */
2401 static void tx_work_handler(struct work_struct *work)
2402 {
2403         struct delayed_work *delayed_work = to_delayed_work(work);
2404         struct tx_work *tx_work = container_of(delayed_work,
2405                                                struct tx_work, work);
2406         struct sock *sk = tx_work->sk;
2407         struct tls_context *tls_ctx = tls_get_ctx(sk);
2408         struct tls_sw_context_tx *ctx;
2409
2410         if (unlikely(!tls_ctx))
2411                 return;
2412
2413         ctx = tls_sw_ctx_tx(tls_ctx);
2414         if (test_bit(BIT_TX_CLOSING, &ctx->tx_bitmask))
2415                 return;
2416
2417         if (!test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
2418                 return;
2419         mutex_lock(&tls_ctx->tx_lock);
2420         lock_sock(sk);
2421         tls_tx_records(sk, -1);
2422         release_sock(sk);
2423         mutex_unlock(&tls_ctx->tx_lock);
2424 }
2425
2426 static bool tls_is_tx_ready(struct tls_sw_context_tx *ctx)
2427 {
2428         struct tls_rec *rec;
2429
2430         rec = list_first_entry(&ctx->tx_list, struct tls_rec, list);
2431         if (!rec)
2432                 return false;
2433
2434         return READ_ONCE(rec->tx_ready);
2435 }
2436
2437 void tls_sw_write_space(struct sock *sk, struct tls_context *ctx)
2438 {
2439         struct tls_sw_context_tx *tx_ctx = tls_sw_ctx_tx(ctx);
2440
2441         /* Schedule the transmission if tx list is ready */
2442         if (tls_is_tx_ready(tx_ctx) &&
2443             !test_and_set_bit(BIT_TX_SCHEDULED, &tx_ctx->tx_bitmask))
2444                 schedule_delayed_work(&tx_ctx->tx_work.work, 0);
2445 }
2446
2447 void tls_sw_strparser_arm(struct sock *sk, struct tls_context *tls_ctx)
2448 {
2449         struct tls_sw_context_rx *rx_ctx = tls_sw_ctx_rx(tls_ctx);
2450
2451         write_lock_bh(&sk->sk_callback_lock);
2452         rx_ctx->saved_data_ready = sk->sk_data_ready;
2453         sk->sk_data_ready = tls_data_ready;
2454         write_unlock_bh(&sk->sk_callback_lock);
2455 }
2456
2457 void tls_update_rx_zc_capable(struct tls_context *tls_ctx)
2458 {
2459         struct tls_sw_context_rx *rx_ctx = tls_sw_ctx_rx(tls_ctx);
2460
2461         rx_ctx->zc_capable = tls_ctx->rx_no_pad ||
2462                 tls_ctx->prot_info.version != TLS_1_3_VERSION;
2463 }
2464
2465 int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
2466 {
2467         struct tls_context *tls_ctx = tls_get_ctx(sk);
2468         struct tls_prot_info *prot = &tls_ctx->prot_info;
2469         struct tls_crypto_info *crypto_info;
2470         struct tls_sw_context_tx *sw_ctx_tx = NULL;
2471         struct tls_sw_context_rx *sw_ctx_rx = NULL;
2472         struct cipher_context *cctx;
2473         struct crypto_aead **aead;
2474         u16 nonce_size, tag_size, iv_size, rec_seq_size, salt_size;
2475         struct crypto_tfm *tfm;
2476         char *iv, *rec_seq, *key, *salt, *cipher_name;
2477         size_t keysize;
2478         int rc = 0;
2479
2480         if (!ctx) {
2481                 rc = -EINVAL;
2482                 goto out;
2483         }
2484
2485         if (tx) {
2486                 if (!ctx->priv_ctx_tx) {
2487                         sw_ctx_tx = kzalloc(sizeof(*sw_ctx_tx), GFP_KERNEL);
2488                         if (!sw_ctx_tx) {
2489                                 rc = -ENOMEM;
2490                                 goto out;
2491                         }
2492                         ctx->priv_ctx_tx = sw_ctx_tx;
2493                 } else {
2494                         sw_ctx_tx =
2495                                 (struct tls_sw_context_tx *)ctx->priv_ctx_tx;
2496                 }
2497         } else {
2498                 if (!ctx->priv_ctx_rx) {
2499                         sw_ctx_rx = kzalloc(sizeof(*sw_ctx_rx), GFP_KERNEL);
2500                         if (!sw_ctx_rx) {
2501                                 rc = -ENOMEM;
2502                                 goto out;
2503                         }
2504                         ctx->priv_ctx_rx = sw_ctx_rx;
2505                 } else {
2506                         sw_ctx_rx =
2507                                 (struct tls_sw_context_rx *)ctx->priv_ctx_rx;
2508                 }
2509         }
2510
2511         if (tx) {
2512                 crypto_init_wait(&sw_ctx_tx->async_wait);
2513                 spin_lock_init(&sw_ctx_tx->encrypt_compl_lock);
2514                 crypto_info = &ctx->crypto_send.info;
2515                 cctx = &ctx->tx;
2516                 aead = &sw_ctx_tx->aead_send;
2517                 INIT_LIST_HEAD(&sw_ctx_tx->tx_list);
2518                 INIT_DELAYED_WORK(&sw_ctx_tx->tx_work.work, tx_work_handler);
2519                 sw_ctx_tx->tx_work.sk = sk;
2520         } else {
2521                 crypto_init_wait(&sw_ctx_rx->async_wait);
2522                 spin_lock_init(&sw_ctx_rx->decrypt_compl_lock);
2523                 init_waitqueue_head(&sw_ctx_rx->wq);
2524                 crypto_info = &ctx->crypto_recv.info;
2525                 cctx = &ctx->rx;
2526                 skb_queue_head_init(&sw_ctx_rx->rx_list);
2527                 skb_queue_head_init(&sw_ctx_rx->async_hold);
2528                 aead = &sw_ctx_rx->aead_recv;
2529         }
2530
2531         switch (crypto_info->cipher_type) {
2532         case TLS_CIPHER_AES_GCM_128: {
2533                 struct tls12_crypto_info_aes_gcm_128 *gcm_128_info;
2534
2535                 gcm_128_info = (void *)crypto_info;
2536                 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2537                 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
2538                 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2539                 iv = gcm_128_info->iv;
2540                 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
2541                 rec_seq = gcm_128_info->rec_seq;
2542                 keysize = TLS_CIPHER_AES_GCM_128_KEY_SIZE;
2543                 key = gcm_128_info->key;
2544                 salt = gcm_128_info->salt;
2545                 salt_size = TLS_CIPHER_AES_GCM_128_SALT_SIZE;
2546                 cipher_name = "gcm(aes)";
2547                 break;
2548         }
2549         case TLS_CIPHER_AES_GCM_256: {
2550                 struct tls12_crypto_info_aes_gcm_256 *gcm_256_info;
2551
2552                 gcm_256_info = (void *)crypto_info;
2553                 nonce_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2554                 tag_size = TLS_CIPHER_AES_GCM_256_TAG_SIZE;
2555                 iv_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2556                 iv = gcm_256_info->iv;
2557                 rec_seq_size = TLS_CIPHER_AES_GCM_256_REC_SEQ_SIZE;
2558                 rec_seq = gcm_256_info->rec_seq;
2559                 keysize = TLS_CIPHER_AES_GCM_256_KEY_SIZE;
2560                 key = gcm_256_info->key;
2561                 salt = gcm_256_info->salt;
2562                 salt_size = TLS_CIPHER_AES_GCM_256_SALT_SIZE;
2563                 cipher_name = "gcm(aes)";
2564                 break;
2565         }
2566         case TLS_CIPHER_AES_CCM_128: {
2567                 struct tls12_crypto_info_aes_ccm_128 *ccm_128_info;
2568
2569                 ccm_128_info = (void *)crypto_info;
2570                 nonce_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2571                 tag_size = TLS_CIPHER_AES_CCM_128_TAG_SIZE;
2572                 iv_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2573                 iv = ccm_128_info->iv;
2574                 rec_seq_size = TLS_CIPHER_AES_CCM_128_REC_SEQ_SIZE;
2575                 rec_seq = ccm_128_info->rec_seq;
2576                 keysize = TLS_CIPHER_AES_CCM_128_KEY_SIZE;
2577                 key = ccm_128_info->key;
2578                 salt = ccm_128_info->salt;
2579                 salt_size = TLS_CIPHER_AES_CCM_128_SALT_SIZE;
2580                 cipher_name = "ccm(aes)";
2581                 break;
2582         }
2583         case TLS_CIPHER_CHACHA20_POLY1305: {
2584                 struct tls12_crypto_info_chacha20_poly1305 *chacha20_poly1305_info;
2585
2586                 chacha20_poly1305_info = (void *)crypto_info;
2587                 nonce_size = 0;
2588                 tag_size = TLS_CIPHER_CHACHA20_POLY1305_TAG_SIZE;
2589                 iv_size = TLS_CIPHER_CHACHA20_POLY1305_IV_SIZE;
2590                 iv = chacha20_poly1305_info->iv;
2591                 rec_seq_size = TLS_CIPHER_CHACHA20_POLY1305_REC_SEQ_SIZE;
2592                 rec_seq = chacha20_poly1305_info->rec_seq;
2593                 keysize = TLS_CIPHER_CHACHA20_POLY1305_KEY_SIZE;
2594                 key = chacha20_poly1305_info->key;
2595                 salt = chacha20_poly1305_info->salt;
2596                 salt_size = TLS_CIPHER_CHACHA20_POLY1305_SALT_SIZE;
2597                 cipher_name = "rfc7539(chacha20,poly1305)";
2598                 break;
2599         }
2600         case TLS_CIPHER_SM4_GCM: {
2601                 struct tls12_crypto_info_sm4_gcm *sm4_gcm_info;
2602
2603                 sm4_gcm_info = (void *)crypto_info;
2604                 nonce_size = TLS_CIPHER_SM4_GCM_IV_SIZE;
2605                 tag_size = TLS_CIPHER_SM4_GCM_TAG_SIZE;
2606                 iv_size = TLS_CIPHER_SM4_GCM_IV_SIZE;
2607                 iv = sm4_gcm_info->iv;
2608                 rec_seq_size = TLS_CIPHER_SM4_GCM_REC_SEQ_SIZE;
2609                 rec_seq = sm4_gcm_info->rec_seq;
2610                 keysize = TLS_CIPHER_SM4_GCM_KEY_SIZE;
2611                 key = sm4_gcm_info->key;
2612                 salt = sm4_gcm_info->salt;
2613                 salt_size = TLS_CIPHER_SM4_GCM_SALT_SIZE;
2614                 cipher_name = "gcm(sm4)";
2615                 break;
2616         }
2617         case TLS_CIPHER_SM4_CCM: {
2618                 struct tls12_crypto_info_sm4_ccm *sm4_ccm_info;
2619
2620                 sm4_ccm_info = (void *)crypto_info;
2621                 nonce_size = TLS_CIPHER_SM4_CCM_IV_SIZE;
2622                 tag_size = TLS_CIPHER_SM4_CCM_TAG_SIZE;
2623                 iv_size = TLS_CIPHER_SM4_CCM_IV_SIZE;
2624                 iv = sm4_ccm_info->iv;
2625                 rec_seq_size = TLS_CIPHER_SM4_CCM_REC_SEQ_SIZE;
2626                 rec_seq = sm4_ccm_info->rec_seq;
2627                 keysize = TLS_CIPHER_SM4_CCM_KEY_SIZE;
2628                 key = sm4_ccm_info->key;
2629                 salt = sm4_ccm_info->salt;
2630                 salt_size = TLS_CIPHER_SM4_CCM_SALT_SIZE;
2631                 cipher_name = "ccm(sm4)";
2632                 break;
2633         }
2634         case TLS_CIPHER_ARIA_GCM_128: {
2635                 struct tls12_crypto_info_aria_gcm_128 *aria_gcm_128_info;
2636
2637                 aria_gcm_128_info = (void *)crypto_info;
2638                 nonce_size = TLS_CIPHER_ARIA_GCM_128_IV_SIZE;
2639                 tag_size = TLS_CIPHER_ARIA_GCM_128_TAG_SIZE;
2640                 iv_size = TLS_CIPHER_ARIA_GCM_128_IV_SIZE;
2641                 iv = aria_gcm_128_info->iv;
2642                 rec_seq_size = TLS_CIPHER_ARIA_GCM_128_REC_SEQ_SIZE;
2643                 rec_seq = aria_gcm_128_info->rec_seq;
2644                 keysize = TLS_CIPHER_ARIA_GCM_128_KEY_SIZE;
2645                 key = aria_gcm_128_info->key;
2646                 salt = aria_gcm_128_info->salt;
2647                 salt_size = TLS_CIPHER_ARIA_GCM_128_SALT_SIZE;
2648                 cipher_name = "gcm(aria)";
2649                 break;
2650         }
2651         case TLS_CIPHER_ARIA_GCM_256: {
2652                 struct tls12_crypto_info_aria_gcm_256 *gcm_256_info;
2653
2654                 gcm_256_info = (void *)crypto_info;
2655                 nonce_size = TLS_CIPHER_ARIA_GCM_256_IV_SIZE;
2656                 tag_size = TLS_CIPHER_ARIA_GCM_256_TAG_SIZE;
2657                 iv_size = TLS_CIPHER_ARIA_GCM_256_IV_SIZE;
2658                 iv = gcm_256_info->iv;
2659                 rec_seq_size = TLS_CIPHER_ARIA_GCM_256_REC_SEQ_SIZE;
2660                 rec_seq = gcm_256_info->rec_seq;
2661                 keysize = TLS_CIPHER_ARIA_GCM_256_KEY_SIZE;
2662                 key = gcm_256_info->key;
2663                 salt = gcm_256_info->salt;
2664                 salt_size = TLS_CIPHER_ARIA_GCM_256_SALT_SIZE;
2665                 cipher_name = "gcm(aria)";
2666                 break;
2667         }
2668         default:
2669                 rc = -EINVAL;
2670                 goto free_priv;
2671         }
2672
2673         if (crypto_info->version == TLS_1_3_VERSION) {
2674                 nonce_size = 0;
2675                 prot->aad_size = TLS_HEADER_SIZE;
2676                 prot->tail_size = 1;
2677         } else {
2678                 prot->aad_size = TLS_AAD_SPACE_SIZE;
2679                 prot->tail_size = 0;
2680         }
2681
2682         /* Sanity-check the sizes for stack allocations. */
2683         if (iv_size > MAX_IV_SIZE || nonce_size > MAX_IV_SIZE ||
2684             rec_seq_size > TLS_MAX_REC_SEQ_SIZE || tag_size != TLS_TAG_SIZE ||
2685             prot->aad_size > TLS_MAX_AAD_SIZE) {
2686                 rc = -EINVAL;
2687                 goto free_priv;
2688         }
2689
2690         prot->version = crypto_info->version;
2691         prot->cipher_type = crypto_info->cipher_type;
2692         prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
2693         prot->tag_size = tag_size;
2694         prot->overhead_size = prot->prepend_size +
2695                               prot->tag_size + prot->tail_size;
2696         prot->iv_size = iv_size;
2697         prot->salt_size = salt_size;
2698         cctx->iv = kmalloc(iv_size + salt_size, GFP_KERNEL);
2699         if (!cctx->iv) {
2700                 rc = -ENOMEM;
2701                 goto free_priv;
2702         }
2703         /* Note: 128 & 256 bit salt are the same size */
2704         prot->rec_seq_size = rec_seq_size;
2705         memcpy(cctx->iv, salt, salt_size);
2706         memcpy(cctx->iv + salt_size, iv, iv_size);
2707         cctx->rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
2708         if (!cctx->rec_seq) {
2709                 rc = -ENOMEM;
2710                 goto free_iv;
2711         }
2712
2713         if (!*aead) {
2714                 *aead = crypto_alloc_aead(cipher_name, 0, 0);
2715                 if (IS_ERR(*aead)) {
2716                         rc = PTR_ERR(*aead);
2717                         *aead = NULL;
2718                         goto free_rec_seq;
2719                 }
2720         }
2721
2722         ctx->push_pending_record = tls_sw_push_pending_record;
2723
2724         rc = crypto_aead_setkey(*aead, key, keysize);
2725
2726         if (rc)
2727                 goto free_aead;
2728
2729         rc = crypto_aead_setauthsize(*aead, prot->tag_size);
2730         if (rc)
2731                 goto free_aead;
2732
2733         if (sw_ctx_rx) {
2734                 tfm = crypto_aead_tfm(sw_ctx_rx->aead_recv);
2735
2736                 tls_update_rx_zc_capable(ctx);
2737                 sw_ctx_rx->async_capable =
2738                         crypto_info->version != TLS_1_3_VERSION &&
2739                         !!(tfm->__crt_alg->cra_flags & CRYPTO_ALG_ASYNC);
2740
2741                 rc = tls_strp_init(&sw_ctx_rx->strp, sk);
2742                 if (rc)
2743                         goto free_aead;
2744         }
2745
2746         goto out;
2747
2748 free_aead:
2749         crypto_free_aead(*aead);
2750         *aead = NULL;
2751 free_rec_seq:
2752         kfree(cctx->rec_seq);
2753         cctx->rec_seq = NULL;
2754 free_iv:
2755         kfree(cctx->iv);
2756         cctx->iv = NULL;
2757 free_priv:
2758         if (tx) {
2759                 kfree(ctx->priv_ctx_tx);
2760                 ctx->priv_ctx_tx = NULL;
2761         } else {
2762                 kfree(ctx->priv_ctx_rx);
2763                 ctx->priv_ctx_rx = NULL;
2764         }
2765 out:
2766         return rc;
2767 }