net/core/skmsg.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* Copyright (c) 2017 - 2018 Covalent IO, Inc. http://covalent.io */
   3
   4 #include <linux/skmsg.h>
   5 #include <linux/skbuff.h>
   6 #include <linux/scatterlist.h>
   7
   8 #include <net/sock.h>
   9 #include <net/tcp.h>
  10 #include <net/tls.h>
  11
  12 static bool sk_msg_try_coalesce_ok(struct sk_msg *msg, int elem_first_coalesce)
  13 {
  14         if (msg->sg.end > msg->sg.start &&
  15             elem_first_coalesce < msg->sg.end)
  16                 return true;
  17
  18         if (msg->sg.end < msg->sg.start &&
  19             (elem_first_coalesce > msg->sg.start ||
  20              elem_first_coalesce < msg->sg.end))
  21                 return true;
  22
  23         return false;
  24 }
  25
  26 int sk_msg_alloc(struct sock *sk, struct sk_msg *msg, int len,
  27                  int elem_first_coalesce)
  28 {
  29         struct page_frag *pfrag = sk_page_frag(sk);
  30         int ret = 0;
  31
  32         len -= msg->sg.size;
  33         while (len > 0) {
  34                 struct scatterlist *sge;
  35                 u32 orig_offset;
  36                 int use, i;
  37
  38                 if (!sk_page_frag_refill(sk, pfrag))
  39                         return -ENOMEM;
  40
  41                 orig_offset = pfrag->offset;
  42                 use = min_t(int, len, pfrag->size - orig_offset);
  43                 if (!sk_wmem_schedule(sk, use))
  44                         return -ENOMEM;
  45
  46                 i = msg->sg.end;
  47                 sk_msg_iter_var_prev(i);
  48                 sge = &msg->sg.data[i];
  49
  50                 if (sk_msg_try_coalesce_ok(msg, elem_first_coalesce) &&
  51                     sg_page(sge) == pfrag->page &&
  52                     sge->offset + sge->length == orig_offset) {
  53                         sge->length += use;
  54                 } else {
  55                         if (sk_msg_full(msg)) {
  56                                 ret = -ENOSPC;
  57                                 break;
  58                         }
  59
  60                         sge = &msg->sg.data[msg->sg.end];
  61                         sg_unmark_end(sge);
  62                         sg_set_page(sge, pfrag->page, use, orig_offset);
  63                         get_page(pfrag->page);
  64                         sk_msg_iter_next(msg, end);
  65                 }
  66
  67                 sk_mem_charge(sk, use);
  68                 msg->sg.size += use;
  69                 pfrag->offset += use;
  70                 len -= use;
  71         }
  72
  73         return ret;
  74 }
  75 EXPORT_SYMBOL_GPL(sk_msg_alloc);
  76
  77 int sk_msg_clone(struct sock *sk, struct sk_msg *dst, struct sk_msg *src,
  78                  u32 off, u32 len)
  79 {
  80         int i = src->sg.start;
  81         struct scatterlist *sge = sk_msg_elem(src, i);
  82         struct scatterlist *sgd = NULL;
  83         u32 sge_len, sge_off;
  84
  85         while (off) {
  86                 if (sge->length > off)
  87                         break;
  88                 off -= sge->length;
  89                 sk_msg_iter_var_next(i);
  90                 if (i == src->sg.end && off)
  91                         return -ENOSPC;
  92                 sge = sk_msg_elem(src, i);
  93         }
  94
  95         while (len) {
  96                 sge_len = sge->length - off;
  97                 if (sge_len > len)
  98                         sge_len = len;
  99
 100                 if (dst->sg.end)
 101                         sgd = sk_msg_elem(dst, dst->sg.end - 1);
 102
 103                 if (sgd &&
 104                     (sg_page(sge) == sg_page(sgd)) &&
 105                     (sg_virt(sge) + off == sg_virt(sgd) + sgd->length)) {
 106                         sgd->length += sge_len;
 107                         dst->sg.size += sge_len;
 108                 } else if (!sk_msg_full(dst)) {
 109                         sge_off = sge->offset + off;
 110                         sk_msg_page_add(dst, sg_page(sge), sge_len, sge_off);
 111                 } else {
 112                         return -ENOSPC;
 113                 }
 114
 115                 off = 0;
 116                 len -= sge_len;
 117                 sk_mem_charge(sk, sge_len);
 118                 sk_msg_iter_var_next(i);
 119                 if (i == src->sg.end && len)
 120                         return -ENOSPC;
 121                 sge = sk_msg_elem(src, i);
 122         }
 123
 124         return 0;
 125 }
 126 EXPORT_SYMBOL_GPL(sk_msg_clone);
 127
 128 void sk_msg_return_zero(struct sock *sk, struct sk_msg *msg, int bytes)
 129 {
 130         int i = msg->sg.start;
 131
 132         do {
 133                 struct scatterlist *sge = sk_msg_elem(msg, i);
 134
 135                 if (bytes < sge->length) {
 136                         sge->length -= bytes;
 137                         sge->offset += bytes;
 138                         sk_mem_uncharge(sk, bytes);
 139                         break;
 140                 }
 141
 142                 sk_mem_uncharge(sk, sge->length);
 143                 bytes -= sge->length;
 144                 sge->length = 0;
 145                 sge->offset = 0;
 146                 sk_msg_iter_var_next(i);
 147         } while (bytes && i != msg->sg.end);
 148         msg->sg.start = i;
 149 }
 150 EXPORT_SYMBOL_GPL(sk_msg_return_zero);
 151
 152 void sk_msg_return(struct sock *sk, struct sk_msg *msg, int bytes)
 153 {
 154         int i = msg->sg.start;
 155
 156         do {
 157                 struct scatterlist *sge = &msg->sg.data[i];
 158                 int uncharge = (bytes < sge->length) ? bytes : sge->length;
 159
 160                 sk_mem_uncharge(sk, uncharge);
 161                 bytes -= uncharge;
 162                 sk_msg_iter_var_next(i);
 163         } while (i != msg->sg.end);
 164 }
 165 EXPORT_SYMBOL_GPL(sk_msg_return);
 166
 167 static int sk_msg_free_elem(struct sock *sk, struct sk_msg *msg, u32 i,
 168                             bool charge)
 169 {
 170         struct scatterlist *sge = sk_msg_elem(msg, i);
 171         u32 len = sge->length;
 172
 173         if (charge)
 174                 sk_mem_uncharge(sk, len);
 175         if (!msg->skb)
 176                 put_page(sg_page(sge));
 177         memset(sge, 0, sizeof(*sge));
 178         return len;
 179 }
 180
 181 static int __sk_msg_free(struct sock *sk, struct sk_msg *msg, u32 i,
 182                          bool charge)
 183 {
 184         struct scatterlist *sge = sk_msg_elem(msg, i);
 185         int freed = 0;
 186
 187         while (msg->sg.size) {
 188                 msg->sg.size -= sge->length;
 189                 freed += sk_msg_free_elem(sk, msg, i, charge);
 190                 sk_msg_iter_var_next(i);
 191                 sk_msg_check_to_free(msg, i, msg->sg.size);
 192                 sge = sk_msg_elem(msg, i);
 193         }
 194         consume_skb(msg->skb);
 195         sk_msg_init(msg);
 196         return freed;
 197 }
 198
 199 int sk_msg_free_nocharge(struct sock *sk, struct sk_msg *msg)
 200 {
 201         return __sk_msg_free(sk, msg, msg->sg.start, false);
 202 }
 203 EXPORT_SYMBOL_GPL(sk_msg_free_nocharge);
 204
 205 int sk_msg_free(struct sock *sk, struct sk_msg *msg)
 206 {
 207         return __sk_msg_free(sk, msg, msg->sg.start, true);
 208 }
 209 EXPORT_SYMBOL_GPL(sk_msg_free);
 210
 211 static void __sk_msg_free_partial(struct sock *sk, struct sk_msg *msg,
 212                                   u32 bytes, bool charge)
 213 {
 214         struct scatterlist *sge;
 215         u32 i = msg->sg.start;
 216
 217         while (bytes) {
 218                 sge = sk_msg_elem(msg, i);
 219                 if (!sge->length)
 220                         break;
 221                 if (bytes < sge->length) {
 222                         if (charge)
 223                                 sk_mem_uncharge(sk, bytes);
 224                         sge->length -= bytes;
 225                         sge->offset += bytes;
 226                         msg->sg.size -= bytes;
 227                         break;
 228                 }
 229
 230                 msg->sg.size -= sge->length;
 231                 bytes -= sge->length;
 232                 sk_msg_free_elem(sk, msg, i, charge);
 233                 sk_msg_iter_var_next(i);
 234                 sk_msg_check_to_free(msg, i, bytes);
 235         }
 236         msg->sg.start = i;
 237 }
 238
 239 void sk_msg_free_partial(struct sock *sk, struct sk_msg *msg, u32 bytes)
 240 {
 241         __sk_msg_free_partial(sk, msg, bytes, true);
 242 }
 243 EXPORT_SYMBOL_GPL(sk_msg_free_partial);
 244
 245 void sk_msg_free_partial_nocharge(struct sock *sk, struct sk_msg *msg,
 246                                   u32 bytes)
 247 {
 248         __sk_msg_free_partial(sk, msg, bytes, false);
 249 }
 250
 251 void sk_msg_trim(struct sock *sk, struct sk_msg *msg, int len)
 252 {
 253         int trim = msg->sg.size - len;
 254         u32 i = msg->sg.end;
 255
 256         if (trim <= 0) {
 257                 WARN_ON(trim < 0);
 258                 return;
 259         }
 260
 261         sk_msg_iter_var_prev(i);
 262         msg->sg.size = len;
 263         while (msg->sg.data[i].length &&
 264                trim >= msg->sg.data[i].length) {
 265                 trim -= msg->sg.data[i].length;
 266                 sk_msg_free_elem(sk, msg, i, true);
 267                 sk_msg_iter_var_prev(i);
 268                 if (!trim)
 269                         goto out;
 270         }
 271
 272         msg->sg.data[i].length -= trim;
 273         sk_mem_uncharge(sk, trim);
 274         /* Adjust copybreak if it falls into the trimmed part of last buf */
 275         if (msg->sg.curr == i && msg->sg.copybreak > msg->sg.data[i].length)
 276                 msg->sg.copybreak = msg->sg.data[i].length;
 277 out:
 278         sk_msg_iter_var_next(i);
 279         msg->sg.end = i;
 280
 281         /* If we trim data a full sg elem before curr pointer update
 282          * copybreak and current so that any future copy operations
 283          * start at new copy location.
 284          * However trimed data that has not yet been used in a copy op
 285          * does not require an update.
 286          */
 287         if (!msg->sg.size) {
 288                 msg->sg.curr = msg->sg.start;
 289                 msg->sg.copybreak = 0;
 290         } else if (sk_msg_iter_dist(msg->sg.start, msg->sg.curr) >=
 291                    sk_msg_iter_dist(msg->sg.start, msg->sg.end)) {
 292                 sk_msg_iter_var_prev(i);
 293                 msg->sg.curr = i;
 294                 msg->sg.copybreak = msg->sg.data[i].length;
 295         }
 296 }
 297 EXPORT_SYMBOL_GPL(sk_msg_trim);
 298
 299 int sk_msg_zerocopy_from_iter(struct sock *sk, struct iov_iter *from,
 300                               struct sk_msg *msg, u32 bytes)
 301 {
 302         int i, maxpages, ret = 0, num_elems = sk_msg_elem_used(msg);
 303         const int to_max_pages = MAX_MSG_FRAGS;
 304         struct page *pages[MAX_MSG_FRAGS];
 305         ssize_t orig, copied, use, offset;
 306
 307         orig = msg->sg.size;
 308         while (bytes > 0) {
 309                 i = 0;
 310                 maxpages = to_max_pages - num_elems;
 311                 if (maxpages == 0) {
 312                         ret = -EFAULT;
 313                         goto out;
 314                 }
 315
 316                 copied = iov_iter_get_pages(from, pages, bytes, maxpages,
 317                                             &offset);
 318                 if (copied <= 0) {
 319                         ret = -EFAULT;
 320                         goto out;
 321                 }
 322
 323                 iov_iter_advance(from, copied);
 324                 bytes -= copied;
 325                 msg->sg.size += copied;
 326
 327                 while (copied) {
 328                         use = min_t(int, copied, PAGE_SIZE - offset);
 329                         sg_set_page(&msg->sg.data[msg->sg.end],
 330                                     pages[i], use, offset);
 331                         sg_unmark_end(&msg->sg.data[msg->sg.end]);
 332                         sk_mem_charge(sk, use);
 333
 334                         offset = 0;
 335                         copied -= use;
 336                         sk_msg_iter_next(msg, end);
 337                         num_elems++;
 338                         i++;
 339                 }
 340                 /* When zerocopy is mixed with sk_msg_*copy* operations we
 341                  * may have a copybreak set in this case clear and prefer
 342                  * zerocopy remainder when possible.
 343                  */
 344                 msg->sg.copybreak = 0;
 345                 msg->sg.curr = msg->sg.end;
 346         }
 347 out:
 348         /* Revert iov_iter updates, msg will need to use 'trim' later if it
 349          * also needs to be cleared.
 350          */
 351         if (ret)
 352                 iov_iter_revert(from, msg->sg.size - orig);
 353         return ret;
 354 }
 355 EXPORT_SYMBOL_GPL(sk_msg_zerocopy_from_iter);
 356
 357 int sk_msg_memcopy_from_iter(struct sock *sk, struct iov_iter *from,
 358                              struct sk_msg *msg, u32 bytes)
 359 {
 360         int ret = -ENOSPC, i = msg->sg.curr;
 361         struct scatterlist *sge;
 362         u32 copy, buf_size;
 363         void *to;
 364
 365         do {
 366                 sge = sk_msg_elem(msg, i);
 367                 /* This is possible if a trim operation shrunk the buffer */
 368                 if (msg->sg.copybreak >= sge->length) {
 369                         msg->sg.copybreak = 0;
 370                         sk_msg_iter_var_next(i);
 371                         if (i == msg->sg.end)
 372                                 break;
 373                         sge = sk_msg_elem(msg, i);
 374                 }
 375
 376                 buf_size = sge->length - msg->sg.copybreak;
 377                 copy = (buf_size > bytes) ? bytes : buf_size;
 378                 to = sg_virt(sge) + msg->sg.copybreak;
 379                 msg->sg.copybreak += copy;
 380                 if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY)
 381                         ret = copy_from_iter_nocache(to, copy, from);
 382                 else
 383                         ret = copy_from_iter(to, copy, from);
 384                 if (ret != copy) {
 385                         ret = -EFAULT;
 386                         goto out;
 387                 }
 388                 bytes -= copy;
 389                 if (!bytes)
 390                         break;
 391                 msg->sg.copybreak = 0;
 392                 sk_msg_iter_var_next(i);
 393         } while (i != msg->sg.end);
 394 out:
 395         msg->sg.curr = i;
 396         return ret;
 397 }
 398 EXPORT_SYMBOL_GPL(sk_msg_memcopy_from_iter);
 399
 400 static int sk_psock_skb_ingress(struct sk_psock *psock, struct sk_buff *skb)
 401 {
 402         struct sock *sk = psock->sk;
 403         int copied = 0, num_sge;
 404         struct sk_msg *msg;
 405
 406         msg = kzalloc(sizeof(*msg), __GFP_NOWARN | GFP_ATOMIC);
 407         if (unlikely(!msg))
 408                 return -EAGAIN;
 409         if (!sk_rmem_schedule(sk, skb, skb->len)) {
 410                 kfree(msg);
 411                 return -EAGAIN;
 412         }
 413
 414         sk_msg_init(msg);
 415         num_sge = skb_to_sgvec(skb, msg->sg.data, 0, skb->len);
 416         if (unlikely(num_sge < 0)) {
 417                 kfree(msg);
 418                 return num_sge;
 419         }
 420
 421         sk_mem_charge(sk, skb->len);
 422         copied = skb->len;
 423         msg->sg.start = 0;
 424         msg->sg.size = copied;
 425         msg->sg.end = num_sge;
 426         msg->skb = skb;
 427
 428         sk_psock_queue_msg(psock, msg);
 429         sk_psock_data_ready(sk, psock);
 430         return copied;
 431 }
 432
 433 static int sk_psock_handle_skb(struct sk_psock *psock, struct sk_buff *skb,
 434                                u32 off, u32 len, bool ingress)
 435 {
 436         if (!ingress) {
 437                 if (!sock_writeable(psock->sk))
 438                         return -EAGAIN;
 439                 return skb_send_sock_locked(psock->sk, skb, off, len);
 440         }
 441         return sk_psock_skb_ingress(psock, skb);
 442 }
 443
 444 static void sk_psock_backlog(struct work_struct *work)
 445 {
 446         struct sk_psock *psock = container_of(work, struct sk_psock, work);
 447         struct sk_psock_work_state *state = &psock->work_state;
 448         struct sk_buff *skb;
 449         bool ingress;
 450         u32 len, off;
 451         int ret;
 452
 453         /* Lock sock to avoid losing sk_socket during loop. */
 454         lock_sock(psock->sk);
 455         if (state->skb) {
 456                 skb = state->skb;
 457                 len = state->len;
 458                 off = state->off;
 459                 state->skb = NULL;
 460                 goto start;
 461         }
 462
 463         while ((skb = skb_dequeue(&psock->ingress_skb))) {
 464                 len = skb->len;
 465                 off = 0;
 466 start:
 467                 ingress = tcp_skb_bpf_ingress(skb);
 468                 do {
 469                         ret = -EIO;
 470                         if (likely(psock->sk->sk_socket))
 471                                 ret = sk_psock_handle_skb(psock, skb, off,
 472                                                           len, ingress);
 473                         if (ret <= 0) {
 474                                 if (ret == -EAGAIN) {
 475                                         state->skb = skb;
 476                                         state->len = len;
 477                                         state->off = off;
 478                                         goto end;
 479                                 }
 480                                 /* Hard errors break pipe and stop xmit. */
 481                                 sk_psock_report_error(psock, ret ? -ret : EPIPE);
 482                                 sk_psock_clear_state(psock, SK_PSOCK_TX_ENABLED);
 483                                 kfree_skb(skb);
 484                                 goto end;
 485                         }
 486                         off += ret;
 487                         len -= ret;
 488                 } while (len);
 489
 490                 if (!ingress)
 491                         kfree_skb(skb);
 492         }
 493 end:
 494         release_sock(psock->sk);
 495 }
 496
 497 struct sk_psock *sk_psock_init(struct sock *sk, int node)
 498 {
 499         struct sk_psock *psock;
 500         struct proto *prot;
 501
 502         write_lock_bh(&sk->sk_callback_lock);
 503
 504         if (inet_csk_has_ulp(sk)) {
 505                 psock = ERR_PTR(-EINVAL);
 506                 goto out;
 507         }
 508
 509         if (sk->sk_user_data) {
 510                 psock = ERR_PTR(-EBUSY);
 511                 goto out;
 512         }
 513
 514         psock = kzalloc_node(sizeof(*psock), GFP_ATOMIC | __GFP_NOWARN, node);
 515         if (!psock) {
 516                 psock = ERR_PTR(-ENOMEM);
 517                 goto out;
 518         }
 519
 520         prot = READ_ONCE(sk->sk_prot);
 521         psock->sk = sk;
 522         psock->eval = __SK_NONE;
 523         psock->sk_proto = prot;
 524         psock->saved_unhash = prot->unhash;
 525         psock->saved_close = prot->close;
 526         psock->saved_write_space = sk->sk_write_space;
 527
 528         INIT_LIST_HEAD(&psock->link);
 529         spin_lock_init(&psock->link_lock);
 530
 531         INIT_WORK(&psock->work, sk_psock_backlog);
 532         INIT_LIST_HEAD(&psock->ingress_msg);
 533         skb_queue_head_init(&psock->ingress_skb);
 534
 535         sk_psock_set_state(psock, SK_PSOCK_TX_ENABLED);
 536         refcount_set(&psock->refcnt, 1);
 537
 538         rcu_assign_sk_user_data_nocopy(sk, psock);
 539         sock_hold(sk);
 540
 541 out:
 542         write_unlock_bh(&sk->sk_callback_lock);
 543         return psock;
 544 }
 545 EXPORT_SYMBOL_GPL(sk_psock_init);
 546
 547 struct sk_psock_link *sk_psock_link_pop(struct sk_psock *psock)
 548 {
 549         struct sk_psock_link *link;
 550
 551         spin_lock_bh(&psock->link_lock);
 552         link = list_first_entry_or_null(&psock->link, struct sk_psock_link,
 553                                         list);
 554         if (link)
 555                 list_del(&link->list);
 556         spin_unlock_bh(&psock->link_lock);
 557         return link;
 558 }
 559
 560 void __sk_psock_purge_ingress_msg(struct sk_psock *psock)
 561 {
 562         struct sk_msg *msg, *tmp;
 563
 564         list_for_each_entry_safe(msg, tmp, &psock->ingress_msg, list) {
 565                 list_del(&msg->list);
 566                 sk_msg_free(psock->sk, msg);
 567                 kfree(msg);
 568         }
 569 }
 570
 571 static void sk_psock_zap_ingress(struct sk_psock *psock)
 572 {
 573         __skb_queue_purge(&psock->ingress_skb);
 574         __sk_psock_purge_ingress_msg(psock);
 575 }
 576
 577 static void sk_psock_link_destroy(struct sk_psock *psock)
 578 {
 579         struct sk_psock_link *link, *tmp;
 580
 581         list_for_each_entry_safe(link, tmp, &psock->link, list) {
 582                 list_del(&link->list);
 583                 sk_psock_free_link(link);
 584         }
 585 }
 586
 587 static void sk_psock_destroy_deferred(struct work_struct *gc)
 588 {
 589         struct sk_psock *psock = container_of(gc, struct sk_psock, gc);
 590
 591         /* No sk_callback_lock since already detached. */
 592
 593         /* Parser has been stopped */
 594         if (psock->progs.skb_parser)
 595                 strp_done(&psock->parser.strp);
 596
 597         cancel_work_sync(&psock->work);
 598
 599         psock_progs_drop(&psock->progs);
 600
 601         sk_psock_link_destroy(psock);
 602         sk_psock_cork_free(psock);
 603         sk_psock_zap_ingress(psock);
 604
 605         if (psock->sk_redir)
 606                 sock_put(psock->sk_redir);
 607         sock_put(psock->sk);
 608         kfree(psock);
 609 }
 610
 611 void sk_psock_destroy(struct rcu_head *rcu)
 612 {
 613         struct sk_psock *psock = container_of(rcu, struct sk_psock, rcu);
 614
 615         INIT_WORK(&psock->gc, sk_psock_destroy_deferred);
 616         schedule_work(&psock->gc);
 617 }
 618 EXPORT_SYMBOL_GPL(sk_psock_destroy);
 619
 620 void sk_psock_drop(struct sock *sk, struct sk_psock *psock)
 621 {
 622         sk_psock_cork_free(psock);
 623         sk_psock_zap_ingress(psock);
 624
 625         write_lock_bh(&sk->sk_callback_lock);
 626         sk_psock_restore_proto(sk, psock);
 627         rcu_assign_sk_user_data(sk, NULL);
 628         if (psock->progs.skb_parser)
 629                 sk_psock_stop_strp(sk, psock);
 630         write_unlock_bh(&sk->sk_callback_lock);
 631         sk_psock_clear_state(psock, SK_PSOCK_TX_ENABLED);
 632
 633         call_rcu(&psock->rcu, sk_psock_destroy);
 634 }
 635 EXPORT_SYMBOL_GPL(sk_psock_drop);
 636
 637 static int sk_psock_map_verd(int verdict, bool redir)
 638 {
 639         switch (verdict) {
 640         case SK_PASS:
 641                 return redir ? __SK_REDIRECT : __SK_PASS;
 642         case SK_DROP:
 643         default:
 644                 break;
 645         }
 646
 647         return __SK_DROP;
 648 }
 649
 650 int sk_psock_msg_verdict(struct sock *sk, struct sk_psock *psock,
 651                          struct sk_msg *msg)
 652 {
 653         struct bpf_prog *prog;
 654         int ret;
 655
 656         rcu_read_lock();
 657         prog = READ_ONCE(psock->progs.msg_parser);
 658         if (unlikely(!prog)) {
 659                 ret = __SK_PASS;
 660                 goto out;
 661         }
 662
 663         sk_msg_compute_data_pointers(msg);
 664         msg->sk = sk;
 665         ret = bpf_prog_run_pin_on_cpu(prog, msg);
 666         ret = sk_psock_map_verd(ret, msg->sk_redir);
 667         psock->apply_bytes = msg->apply_bytes;
 668         if (ret == __SK_REDIRECT) {
 669                 if (psock->sk_redir)
 670                         sock_put(psock->sk_redir);
 671                 psock->sk_redir = msg->sk_redir;
 672                 if (!psock->sk_redir) {
 673                         ret = __SK_DROP;
 674                         goto out;
 675                 }
 676                 sock_hold(psock->sk_redir);
 677         }
 678 out:
 679         rcu_read_unlock();
 680         return ret;
 681 }
 682 EXPORT_SYMBOL_GPL(sk_psock_msg_verdict);
 683
 684 static int sk_psock_bpf_run(struct sk_psock *psock, struct bpf_prog *prog,
 685                             struct sk_buff *skb)
 686 {
 687         bpf_compute_data_end_sk_skb(skb);
 688         return bpf_prog_run_pin_on_cpu(prog, skb);
 689 }
 690
 691 static struct sk_psock *sk_psock_from_strp(struct strparser *strp)
 692 {
 693         struct sk_psock_parser *parser;
 694
 695         parser = container_of(strp, struct sk_psock_parser, strp);
 696         return container_of(parser, struct sk_psock, parser);
 697 }
 698
 699 static void sk_psock_skb_redirect(struct sk_buff *skb)
 700 {
 701         struct sk_psock *psock_other;
 702         struct sock *sk_other;
 703
 704         sk_other = tcp_skb_bpf_redirect_fetch(skb);
 705         /* This error is a buggy BPF program, it returned a redirect
 706          * return code, but then didn't set a redirect interface.
 707          */
 708         if (unlikely(!sk_other)) {
 709                 kfree_skb(skb);
 710                 return;
 711         }
 712         psock_other = sk_psock(sk_other);
 713         /* This error indicates the socket is being torn down or had another
 714          * error that caused the pipe to break. We can't send a packet on
 715          * a socket that is in this state so we drop the skb.
 716          */
 717         if (!psock_other || sock_flag(sk_other, SOCK_DEAD) ||
 718             !sk_psock_test_state(psock_other, SK_PSOCK_TX_ENABLED)) {
 719                 kfree_skb(skb);
 720                 return;
 721         }
 722
 723         skb_queue_tail(&psock_other->ingress_skb, skb);
 724         schedule_work(&psock_other->work);
 725 }
 726
 727 static void sk_psock_tls_verdict_apply(struct sk_buff *skb, struct sock *sk, int verdict)
 728 {
 729         switch (verdict) {
 730         case __SK_REDIRECT:
 731                 skb_set_owner_r(skb, sk);
 732                 sk_psock_skb_redirect(skb);
 733                 break;
 734         case __SK_PASS:
 735         case __SK_DROP:
 736         default:
 737                 break;
 738         }
 739 }
 740
 741 int sk_psock_tls_strp_read(struct sk_psock *psock, struct sk_buff *skb)
 742 {
 743         struct bpf_prog *prog;
 744         int ret = __SK_PASS;
 745
 746         rcu_read_lock();
 747         prog = READ_ONCE(psock->progs.skb_verdict);
 748         if (likely(prog)) {
 749                 /* We skip full set_owner_r here because if we do a SK_PASS
 750                  * or SK_DROP we can skip skb memory accounting and use the
 751                  * TLS context.
 752                  */
 753                 skb->sk = psock->sk;
 754                 tcp_skb_bpf_redirect_clear(skb);
 755                 ret = sk_psock_bpf_run(psock, prog, skb);
 756                 ret = sk_psock_map_verd(ret, tcp_skb_bpf_redirect_fetch(skb));
 757                 skb->sk = NULL;
 758         }
 759         sk_psock_tls_verdict_apply(skb, psock->sk, ret);
 760         rcu_read_unlock();
 761         return ret;
 762 }
 763 EXPORT_SYMBOL_GPL(sk_psock_tls_strp_read);
 764
 765 static void sk_psock_verdict_apply(struct sk_psock *psock,
 766                                    struct sk_buff *skb, int verdict)
 767 {
 768         struct tcp_skb_cb *tcp;
 769         struct sock *sk_other;
 770         int err = -EIO;
 771
 772         switch (verdict) {
 773         case __SK_PASS:
 774                 sk_other = psock->sk;
 775                 if (sock_flag(sk_other, SOCK_DEAD) ||
 776                     !sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
 777                         goto out_free;
 778                 }
 779
 780                 tcp = TCP_SKB_CB(skb);
 781                 tcp->bpf.flags |= BPF_F_INGRESS;
 782
 783                 /* If the queue is empty then we can submit directly
 784                  * into the msg queue. If its not empty we have to
 785                  * queue work otherwise we may get OOO data. Otherwise,
 786                  * if sk_psock_skb_ingress errors will be handled by
 787                  * retrying later from workqueue.
 788                  */
 789                 if (skb_queue_empty(&psock->ingress_skb)) {
 790                         err = sk_psock_skb_ingress(psock, skb);
 791                 }
 792                 if (err < 0) {
 793                         skb_queue_tail(&psock->ingress_skb, skb);
 794                         schedule_work(&psock->work);
 795                 }
 796                 break;
 797         case __SK_REDIRECT:
 798                 sk_psock_skb_redirect(skb);
 799                 break;
 800         case __SK_DROP:
 801         default:
 802 out_free:
 803                 kfree_skb(skb);
 804         }
 805 }
 806
 807 static void sk_psock_strp_read(struct strparser *strp, struct sk_buff *skb)
 808 {
 809         struct sk_psock *psock;
 810         struct bpf_prog *prog;
 811         int ret = __SK_DROP;
 812         struct sock *sk;
 813
 814         rcu_read_lock();
 815         sk = strp->sk;
 816         psock = sk_psock(sk);
 817         if (unlikely(!psock)) {
 818                 kfree_skb(skb);
 819                 goto out;
 820         }
 821         skb_set_owner_r(skb, sk);
 822         prog = READ_ONCE(psock->progs.skb_verdict);
 823         if (likely(prog)) {
 824                 tcp_skb_bpf_redirect_clear(skb);
 825                 ret = sk_psock_bpf_run(psock, prog, skb);
 826                 ret = sk_psock_map_verd(ret, tcp_skb_bpf_redirect_fetch(skb));
 827         }
 828         sk_psock_verdict_apply(psock, skb, ret);
 829 out:
 830         rcu_read_unlock();
 831 }
 832
 833 static int sk_psock_strp_read_done(struct strparser *strp, int err)
 834 {
 835         return err;
 836 }
 837
 838 static int sk_psock_strp_parse(struct strparser *strp, struct sk_buff *skb)
 839 {
 840         struct sk_psock *psock = sk_psock_from_strp(strp);
 841         struct bpf_prog *prog;
 842         int ret = skb->len;
 843
 844         rcu_read_lock();
 845         prog = READ_ONCE(psock->progs.skb_parser);
 846         if (likely(prog)) {
 847                 skb->sk = psock->sk;
 848                 ret = sk_psock_bpf_run(psock, prog, skb);
 849                 skb->sk = NULL;
 850         }
 851         rcu_read_unlock();
 852         return ret;
 853 }
 854
 855 /* Called with socket lock held. */
 856 static void sk_psock_strp_data_ready(struct sock *sk)
 857 {
 858         struct sk_psock *psock;
 859
 860         rcu_read_lock();
 861         psock = sk_psock(sk);
 862         if (likely(psock)) {
 863                 if (tls_sw_has_ctx_rx(sk)) {
 864                         psock->parser.saved_data_ready(sk);
 865                 } else {
 866                         write_lock_bh(&sk->sk_callback_lock);
 867                         strp_data_ready(&psock->parser.strp);
 868                         write_unlock_bh(&sk->sk_callback_lock);
 869                 }
 870         }
 871         rcu_read_unlock();
 872 }
 873
 874 static void sk_psock_write_space(struct sock *sk)
 875 {
 876         struct sk_psock *psock;
 877         void (*write_space)(struct sock *sk) = NULL;
 878
 879         rcu_read_lock();
 880         psock = sk_psock(sk);
 881         if (likely(psock)) {
 882                 if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED))
 883                         schedule_work(&psock->work);
 884                 write_space = psock->saved_write_space;
 885         }
 886         rcu_read_unlock();
 887         if (write_space)
 888                 write_space(sk);
 889 }
 890
 891 int sk_psock_init_strp(struct sock *sk, struct sk_psock *psock)
 892 {
 893         static const struct strp_callbacks cb = {
 894                 .rcv_msg        = sk_psock_strp_read,
 895                 .read_sock_done = sk_psock_strp_read_done,
 896                 .parse_msg      = sk_psock_strp_parse,
 897         };
 898
 899         psock->parser.enabled = false;
 900         return strp_init(&psock->parser.strp, sk, &cb);
 901 }
 902
 903 void sk_psock_start_strp(struct sock *sk, struct sk_psock *psock)
 904 {
 905         struct sk_psock_parser *parser = &psock->parser;
 906
 907         if (parser->enabled)
 908                 return;
 909
 910         parser->saved_data_ready = sk->sk_data_ready;
 911         sk->sk_data_ready = sk_psock_strp_data_ready;
 912         sk->sk_write_space = sk_psock_write_space;
 913         parser->enabled = true;
 914 }
 915
 916 void sk_psock_stop_strp(struct sock *sk, struct sk_psock *psock)
 917 {
 918         struct sk_psock_parser *parser = &psock->parser;
 919
 920         if (!parser->enabled)
 921                 return;
 922
 923         sk->sk_data_ready = parser->saved_data_ready;
 924         parser->saved_data_ready = NULL;
 925         strp_stop(&parser->strp);
 926         parser->enabled = false;
 927 }