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         u32 osize = msg->sg.size;
  31         int ret = 0;
  32
  33         len -= msg->sg.size;
  34         while (len > 0) {
  35                 struct scatterlist *sge;
  36                 u32 orig_offset;
  37                 int use, i;
  38
  39                 if (!sk_page_frag_refill(sk, pfrag)) {
  40                         ret = -ENOMEM;
  41                         goto msg_trim;
  42                 }
  43
  44                 orig_offset = pfrag->offset;
  45                 use = min_t(int, len, pfrag->size - orig_offset);
  46                 if (!sk_wmem_schedule(sk, use)) {
  47                         ret = -ENOMEM;
  48                         goto msg_trim;
  49                 }
  50
  51                 i = msg->sg.end;
  52                 sk_msg_iter_var_prev(i);
  53                 sge = &msg->sg.data[i];
  54
  55                 if (sk_msg_try_coalesce_ok(msg, elem_first_coalesce) &&
  56                     sg_page(sge) == pfrag->page &&
  57                     sge->offset + sge->length == orig_offset) {
  58                         sge->length += use;
  59                 } else {
  60                         if (sk_msg_full(msg)) {
  61                                 ret = -ENOSPC;
  62                                 break;
  63                         }
  64
  65                         sge = &msg->sg.data[msg->sg.end];
  66                         sg_unmark_end(sge);
  67                         sg_set_page(sge, pfrag->page, use, orig_offset);
  68                         get_page(pfrag->page);
  69                         sk_msg_iter_next(msg, end);
  70                 }
  71
  72                 sk_mem_charge(sk, use);
  73                 msg->sg.size += use;
  74                 pfrag->offset += use;
  75                 len -= use;
  76         }
  77
  78         return ret;
  79
  80 msg_trim:
  81         sk_msg_trim(sk, msg, osize);
  82         return ret;
  83 }
  84 EXPORT_SYMBOL_GPL(sk_msg_alloc);
  85
  86 int sk_msg_clone(struct sock *sk, struct sk_msg *dst, struct sk_msg *src,
  87                  u32 off, u32 len)
  88 {
  89         int i = src->sg.start;
  90         struct scatterlist *sge = sk_msg_elem(src, i);
  91         struct scatterlist *sgd = NULL;
  92         u32 sge_len, sge_off;
  93
  94         while (off) {
  95                 if (sge->length > off)
  96                         break;
  97                 off -= sge->length;
  98                 sk_msg_iter_var_next(i);
  99                 if (i == src->sg.end && off)
 100                         return -ENOSPC;
 101                 sge = sk_msg_elem(src, i);
 102         }
 103
 104         while (len) {
 105                 sge_len = sge->length - off;
 106                 if (sge_len > len)
 107                         sge_len = len;
 108
 109                 if (dst->sg.end)
 110                         sgd = sk_msg_elem(dst, dst->sg.end - 1);
 111
 112                 if (sgd &&
 113                     (sg_page(sge) == sg_page(sgd)) &&
 114                     (sg_virt(sge) + off == sg_virt(sgd) + sgd->length)) {
 115                         sgd->length += sge_len;
 116                         dst->sg.size += sge_len;
 117                 } else if (!sk_msg_full(dst)) {
 118                         sge_off = sge->offset + off;
 119                         sk_msg_page_add(dst, sg_page(sge), sge_len, sge_off);
 120                 } else {
 121                         return -ENOSPC;
 122                 }
 123
 124                 off = 0;
 125                 len -= sge_len;
 126                 sk_mem_charge(sk, sge_len);
 127                 sk_msg_iter_var_next(i);
 128                 if (i == src->sg.end && len)
 129                         return -ENOSPC;
 130                 sge = sk_msg_elem(src, i);
 131         }
 132
 133         return 0;
 134 }
 135 EXPORT_SYMBOL_GPL(sk_msg_clone);
 136
 137 void sk_msg_return_zero(struct sock *sk, struct sk_msg *msg, int bytes)
 138 {
 139         int i = msg->sg.start;
 140
 141         do {
 142                 struct scatterlist *sge = sk_msg_elem(msg, i);
 143
 144                 if (bytes < sge->length) {
 145                         sge->length -= bytes;
 146                         sge->offset += bytes;
 147                         sk_mem_uncharge(sk, bytes);
 148                         break;
 149                 }
 150
 151                 sk_mem_uncharge(sk, sge->length);
 152                 bytes -= sge->length;
 153                 sge->length = 0;
 154                 sge->offset = 0;
 155                 sk_msg_iter_var_next(i);
 156         } while (bytes && i != msg->sg.end);
 157         msg->sg.start = i;
 158 }
 159 EXPORT_SYMBOL_GPL(sk_msg_return_zero);
 160
 161 void sk_msg_return(struct sock *sk, struct sk_msg *msg, int bytes)
 162 {
 163         int i = msg->sg.start;
 164
 165         do {
 166                 struct scatterlist *sge = &msg->sg.data[i];
 167                 int uncharge = (bytes < sge->length) ? bytes : sge->length;
 168
 169                 sk_mem_uncharge(sk, uncharge);
 170                 bytes -= uncharge;
 171                 sk_msg_iter_var_next(i);
 172         } while (i != msg->sg.end);
 173 }
 174 EXPORT_SYMBOL_GPL(sk_msg_return);
 175
 176 static int sk_msg_free_elem(struct sock *sk, struct sk_msg *msg, u32 i,
 177                             bool charge)
 178 {
 179         struct scatterlist *sge = sk_msg_elem(msg, i);
 180         u32 len = sge->length;
 181
 182         /* When the skb owns the memory we free it from consume_skb path. */
 183         if (!msg->skb) {
 184                 if (charge)
 185                         sk_mem_uncharge(sk, len);
 186                 put_page(sg_page(sge));
 187         }
 188         memset(sge, 0, sizeof(*sge));
 189         return len;
 190 }
 191
 192 static int __sk_msg_free(struct sock *sk, struct sk_msg *msg, u32 i,
 193                          bool charge)
 194 {
 195         struct scatterlist *sge = sk_msg_elem(msg, i);
 196         int freed = 0;
 197
 198         while (msg->sg.size) {
 199                 msg->sg.size -= sge->length;
 200                 freed += sk_msg_free_elem(sk, msg, i, charge);
 201                 sk_msg_iter_var_next(i);
 202                 sk_msg_check_to_free(msg, i, msg->sg.size);
 203                 sge = sk_msg_elem(msg, i);
 204         }
 205         consume_skb(msg->skb);
 206         sk_msg_init(msg);
 207         return freed;
 208 }
 209
 210 int sk_msg_free_nocharge(struct sock *sk, struct sk_msg *msg)
 211 {
 212         return __sk_msg_free(sk, msg, msg->sg.start, false);
 213 }
 214 EXPORT_SYMBOL_GPL(sk_msg_free_nocharge);
 215
 216 int sk_msg_free(struct sock *sk, struct sk_msg *msg)
 217 {
 218         return __sk_msg_free(sk, msg, msg->sg.start, true);
 219 }
 220 EXPORT_SYMBOL_GPL(sk_msg_free);
 221
 222 static void __sk_msg_free_partial(struct sock *sk, struct sk_msg *msg,
 223                                   u32 bytes, bool charge)
 224 {
 225         struct scatterlist *sge;
 226         u32 i = msg->sg.start;
 227
 228         while (bytes) {
 229                 sge = sk_msg_elem(msg, i);
 230                 if (!sge->length)
 231                         break;
 232                 if (bytes < sge->length) {
 233                         if (charge)
 234                                 sk_mem_uncharge(sk, bytes);
 235                         sge->length -= bytes;
 236                         sge->offset += bytes;
 237                         msg->sg.size -= bytes;
 238                         break;
 239                 }
 240
 241                 msg->sg.size -= sge->length;
 242                 bytes -= sge->length;
 243                 sk_msg_free_elem(sk, msg, i, charge);
 244                 sk_msg_iter_var_next(i);
 245                 sk_msg_check_to_free(msg, i, bytes);
 246         }
 247         msg->sg.start = i;
 248 }
 249
 250 void sk_msg_free_partial(struct sock *sk, struct sk_msg *msg, u32 bytes)
 251 {
 252         __sk_msg_free_partial(sk, msg, bytes, true);
 253 }
 254 EXPORT_SYMBOL_GPL(sk_msg_free_partial);
 255
 256 void sk_msg_free_partial_nocharge(struct sock *sk, struct sk_msg *msg,
 257                                   u32 bytes)
 258 {
 259         __sk_msg_free_partial(sk, msg, bytes, false);
 260 }
 261
 262 void sk_msg_trim(struct sock *sk, struct sk_msg *msg, int len)
 263 {
 264         int trim = msg->sg.size - len;
 265         u32 i = msg->sg.end;
 266
 267         if (trim <= 0) {
 268                 WARN_ON(trim < 0);
 269                 return;
 270         }
 271
 272         sk_msg_iter_var_prev(i);
 273         msg->sg.size = len;
 274         while (msg->sg.data[i].length &&
 275                trim >= msg->sg.data[i].length) {
 276                 trim -= msg->sg.data[i].length;
 277                 sk_msg_free_elem(sk, msg, i, true);
 278                 sk_msg_iter_var_prev(i);
 279                 if (!trim)
 280                         goto out;
 281         }
 282
 283         msg->sg.data[i].length -= trim;
 284         sk_mem_uncharge(sk, trim);
 285         /* Adjust copybreak if it falls into the trimmed part of last buf */
 286         if (msg->sg.curr == i && msg->sg.copybreak > msg->sg.data[i].length)
 287                 msg->sg.copybreak = msg->sg.data[i].length;
 288 out:
 289         sk_msg_iter_var_next(i);
 290         msg->sg.end = i;
 291
 292         /* If we trim data a full sg elem before curr pointer update
 293          * copybreak and current so that any future copy operations
 294          * start at new copy location.
 295          * However trimed data that has not yet been used in a copy op
 296          * does not require an update.
 297          */
 298         if (!msg->sg.size) {
 299                 msg->sg.curr = msg->sg.start;
 300                 msg->sg.copybreak = 0;
 301         } else if (sk_msg_iter_dist(msg->sg.start, msg->sg.curr) >=
 302                    sk_msg_iter_dist(msg->sg.start, msg->sg.end)) {
 303                 sk_msg_iter_var_prev(i);
 304                 msg->sg.curr = i;
 305                 msg->sg.copybreak = msg->sg.data[i].length;
 306         }
 307 }
 308 EXPORT_SYMBOL_GPL(sk_msg_trim);
 309
 310 int sk_msg_zerocopy_from_iter(struct sock *sk, struct iov_iter *from,
 311                               struct sk_msg *msg, u32 bytes)
 312 {
 313         int i, maxpages, ret = 0, num_elems = sk_msg_elem_used(msg);
 314         const int to_max_pages = MAX_MSG_FRAGS;
 315         struct page *pages[MAX_MSG_FRAGS];
 316         ssize_t orig, copied, use, offset;
 317
 318         orig = msg->sg.size;
 319         while (bytes > 0) {
 320                 i = 0;
 321                 maxpages = to_max_pages - num_elems;
 322                 if (maxpages == 0) {
 323                         ret = -EFAULT;
 324                         goto out;
 325                 }
 326
 327                 copied = iov_iter_get_pages(from, pages, bytes, maxpages,
 328                                             &offset);
 329                 if (copied <= 0) {
 330                         ret = -EFAULT;
 331                         goto out;
 332                 }
 333
 334                 iov_iter_advance(from, copied);
 335                 bytes -= copied;
 336                 msg->sg.size += copied;
 337
 338                 while (copied) {
 339                         use = min_t(int, copied, PAGE_SIZE - offset);
 340                         sg_set_page(&msg->sg.data[msg->sg.end],
 341                                     pages[i], use, offset);
 342                         sg_unmark_end(&msg->sg.data[msg->sg.end]);
 343                         sk_mem_charge(sk, use);
 344
 345                         offset = 0;
 346                         copied -= use;
 347                         sk_msg_iter_next(msg, end);
 348                         num_elems++;
 349                         i++;
 350                 }
 351                 /* When zerocopy is mixed with sk_msg_*copy* operations we
 352                  * may have a copybreak set in this case clear and prefer
 353                  * zerocopy remainder when possible.
 354                  */
 355                 msg->sg.copybreak = 0;
 356                 msg->sg.curr = msg->sg.end;
 357         }
 358 out:
 359         /* Revert iov_iter updates, msg will need to use 'trim' later if it
 360          * also needs to be cleared.
 361          */
 362         if (ret)
 363                 iov_iter_revert(from, msg->sg.size - orig);
 364         return ret;
 365 }
 366 EXPORT_SYMBOL_GPL(sk_msg_zerocopy_from_iter);
 367
 368 int sk_msg_memcopy_from_iter(struct sock *sk, struct iov_iter *from,
 369                              struct sk_msg *msg, u32 bytes)
 370 {
 371         int ret = -ENOSPC, i = msg->sg.curr;
 372         struct scatterlist *sge;
 373         u32 copy, buf_size;
 374         void *to;
 375
 376         do {
 377                 sge = sk_msg_elem(msg, i);
 378                 /* This is possible if a trim operation shrunk the buffer */
 379                 if (msg->sg.copybreak >= sge->length) {
 380                         msg->sg.copybreak = 0;
 381                         sk_msg_iter_var_next(i);
 382                         if (i == msg->sg.end)
 383                                 break;
 384                         sge = sk_msg_elem(msg, i);
 385                 }
 386
 387                 buf_size = sge->length - msg->sg.copybreak;
 388                 copy = (buf_size > bytes) ? bytes : buf_size;
 389                 to = sg_virt(sge) + msg->sg.copybreak;
 390                 msg->sg.copybreak += copy;
 391                 if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY)
 392                         ret = copy_from_iter_nocache(to, copy, from);
 393                 else
 394                         ret = copy_from_iter(to, copy, from);
 395                 if (ret != copy) {
 396                         ret = -EFAULT;
 397                         goto out;
 398                 }
 399                 bytes -= copy;
 400                 if (!bytes)
 401                         break;
 402                 msg->sg.copybreak = 0;
 403                 sk_msg_iter_var_next(i);
 404         } while (i != msg->sg.end);
 405 out:
 406         msg->sg.curr = i;
 407         return ret;
 408 }
 409 EXPORT_SYMBOL_GPL(sk_msg_memcopy_from_iter);
 410
 411 static struct sk_msg *sk_psock_create_ingress_msg(struct sock *sk,
 412                                                   struct sk_buff *skb)
 413 {
 414         struct sk_msg *msg;
 415
 416         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
 417                 return NULL;
 418
 419         if (!sk_rmem_schedule(sk, skb, skb->truesize))
 420                 return NULL;
 421
 422         msg = kzalloc(sizeof(*msg), __GFP_NOWARN | GFP_ATOMIC);
 423         if (unlikely(!msg))
 424                 return NULL;
 425
 426         sk_msg_init(msg);
 427         return msg;
 428 }
 429
 430 static int sk_psock_skb_ingress_enqueue(struct sk_buff *skb,
 431                                         struct sk_psock *psock,
 432                                         struct sock *sk,
 433                                         struct sk_msg *msg)
 434 {
 435         int num_sge, copied;
 436
 437         /* skb linearize may fail with ENOMEM, but lets simply try again
 438          * later if this happens. Under memory pressure we don't want to
 439          * drop the skb. We need to linearize the skb so that the mapping
 440          * in skb_to_sgvec can not error.
 441          */
 442         if (skb_linearize(skb))
 443                 return -EAGAIN;
 444         num_sge = skb_to_sgvec(skb, msg->sg.data, 0, skb->len);
 445         if (unlikely(num_sge < 0))
 446                 return num_sge;
 447
 448         copied = skb->len;
 449         msg->sg.start = 0;
 450         msg->sg.size = copied;
 451         msg->sg.end = num_sge;
 452         msg->skb = skb;
 453
 454         sk_psock_queue_msg(psock, msg);
 455         sk_psock_data_ready(sk, psock);
 456         return copied;
 457 }
 458
 459 static int sk_psock_skb_ingress_self(struct sk_psock *psock, struct sk_buff *skb);
 460
 461 static int sk_psock_skb_ingress(struct sk_psock *psock, struct sk_buff *skb)
 462 {
 463         struct sock *sk = psock->sk;
 464         struct sk_msg *msg;
 465         int err;
 466
 467         /* If we are receiving on the same sock skb->sk is already assigned,
 468          * skip memory accounting and owner transition seeing it already set
 469          * correctly.
 470          */
 471         if (unlikely(skb->sk == sk))
 472                 return sk_psock_skb_ingress_self(psock, skb);
 473         msg = sk_psock_create_ingress_msg(sk, skb);
 474         if (!msg)
 475                 return -EAGAIN;
 476
 477         /* This will transition ownership of the data from the socket where
 478          * the BPF program was run initiating the redirect to the socket
 479          * we will eventually receive this data on. The data will be released
 480          * from skb_consume found in __tcp_bpf_recvmsg() after its been copied
 481          * into user buffers.
 482          */
 483         skb_set_owner_r(skb, sk);
 484         err = sk_psock_skb_ingress_enqueue(skb, psock, sk, msg);
 485         if (err < 0)
 486                 kfree(msg);
 487         return err;
 488 }
 489
 490 /* Puts an skb on the ingress queue of the socket already assigned to the
 491  * skb. In this case we do not need to check memory limits or skb_set_owner_r
 492  * because the skb is already accounted for here.
 493  */
 494 static int sk_psock_skb_ingress_self(struct sk_psock *psock, struct sk_buff *skb)
 495 {
 496         struct sk_msg *msg = kzalloc(sizeof(*msg), __GFP_NOWARN | GFP_ATOMIC);
 497         struct sock *sk = psock->sk;
 498         int err;
 499
 500         if (unlikely(!msg))
 501                 return -EAGAIN;
 502         sk_msg_init(msg);
 503         skb_set_owner_r(skb, sk);
 504         err = sk_psock_skb_ingress_enqueue(skb, psock, sk, msg);
 505         if (err < 0)
 506                 kfree(msg);
 507         return err;
 508 }
 509
 510 static int sk_psock_handle_skb(struct sk_psock *psock, struct sk_buff *skb,
 511                                u32 off, u32 len, bool ingress)
 512 {
 513         if (!ingress) {
 514                 if (!sock_writeable(psock->sk))
 515                         return -EAGAIN;
 516                 return skb_send_sock_locked(psock->sk, skb, off, len);
 517         }
 518         return sk_psock_skb_ingress(psock, skb);
 519 }
 520
 521 static void sk_psock_backlog(struct work_struct *work)
 522 {
 523         struct sk_psock *psock = container_of(work, struct sk_psock, work);
 524         struct sk_psock_work_state *state = &psock->work_state;
 525         struct sk_buff *skb;
 526         bool ingress;
 527         u32 len, off;
 528         int ret;
 529
 530         /* Lock sock to avoid losing sk_socket during loop. */
 531         lock_sock(psock->sk);
 532         if (state->skb) {
 533                 skb = state->skb;
 534                 len = state->len;
 535                 off = state->off;
 536                 state->skb = NULL;
 537                 goto start;
 538         }
 539
 540         while ((skb = skb_dequeue(&psock->ingress_skb))) {
 541                 len = skb->len;
 542                 off = 0;
 543 start:
 544                 ingress = tcp_skb_bpf_ingress(skb);
 545                 do {
 546                         ret = -EIO;
 547                         if (likely(psock->sk->sk_socket))
 548                                 ret = sk_psock_handle_skb(psock, skb, off,
 549                                                           len, ingress);
 550                         if (ret <= 0) {
 551                                 if (ret == -EAGAIN) {
 552                                         state->skb = skb;
 553                                         state->len = len;
 554                                         state->off = off;
 555                                         goto end;
 556                                 }
 557                                 /* Hard errors break pipe and stop xmit. */
 558                                 sk_psock_report_error(psock, ret ? -ret : EPIPE);
 559                                 sk_psock_clear_state(psock, SK_PSOCK_TX_ENABLED);
 560                                 kfree_skb(skb);
 561                                 goto end;
 562                         }
 563                         off += ret;
 564                         len -= ret;
 565                 } while (len);
 566
 567                 if (!ingress)
 568                         kfree_skb(skb);
 569         }
 570 end:
 571         release_sock(psock->sk);
 572 }
 573
 574 struct sk_psock *sk_psock_init(struct sock *sk, int node)
 575 {
 576         struct sk_psock *psock;
 577         struct proto *prot;
 578
 579         write_lock_bh(&sk->sk_callback_lock);
 580
 581         if (inet_csk_has_ulp(sk)) {
 582                 psock = ERR_PTR(-EINVAL);
 583                 goto out;
 584         }
 585
 586         if (sk->sk_user_data) {
 587                 psock = ERR_PTR(-EBUSY);
 588                 goto out;
 589         }
 590
 591         psock = kzalloc_node(sizeof(*psock), GFP_ATOMIC | __GFP_NOWARN, node);
 592         if (!psock) {
 593                 psock = ERR_PTR(-ENOMEM);
 594                 goto out;
 595         }
 596
 597         prot = READ_ONCE(sk->sk_prot);
 598         psock->sk = sk;
 599         psock->eval = __SK_NONE;
 600         psock->sk_proto = prot;
 601         psock->saved_unhash = prot->unhash;
 602         psock->saved_close = prot->close;
 603         psock->saved_write_space = sk->sk_write_space;
 604
 605         INIT_LIST_HEAD(&psock->link);
 606         spin_lock_init(&psock->link_lock);
 607
 608         INIT_WORK(&psock->work, sk_psock_backlog);
 609         INIT_LIST_HEAD(&psock->ingress_msg);
 610         skb_queue_head_init(&psock->ingress_skb);
 611
 612         sk_psock_set_state(psock, SK_PSOCK_TX_ENABLED);
 613         refcount_set(&psock->refcnt, 1);
 614
 615         __rcu_assign_sk_user_data_with_flags(sk, psock,
 616                                              SK_USER_DATA_NOCOPY |
 617                                              SK_USER_DATA_PSOCK);
 618         sock_hold(sk);
 619
 620 out:
 621         write_unlock_bh(&sk->sk_callback_lock);
 622         return psock;
 623 }
 624 EXPORT_SYMBOL_GPL(sk_psock_init);
 625
 626 struct sk_psock_link *sk_psock_link_pop(struct sk_psock *psock)
 627 {
 628         struct sk_psock_link *link;
 629
 630         spin_lock_bh(&psock->link_lock);
 631         link = list_first_entry_or_null(&psock->link, struct sk_psock_link,
 632                                         list);
 633         if (link)
 634                 list_del(&link->list);
 635         spin_unlock_bh(&psock->link_lock);
 636         return link;
 637 }
 638
 639 void __sk_psock_purge_ingress_msg(struct sk_psock *psock)
 640 {
 641         struct sk_msg *msg, *tmp;
 642
 643         list_for_each_entry_safe(msg, tmp, &psock->ingress_msg, list) {
 644                 list_del(&msg->list);
 645                 sk_msg_free(psock->sk, msg);
 646                 kfree(msg);
 647         }
 648 }
 649
 650 static void sk_psock_zap_ingress(struct sk_psock *psock)
 651 {
 652         __skb_queue_purge(&psock->ingress_skb);
 653         __sk_psock_purge_ingress_msg(psock);
 654 }
 655
 656 static void sk_psock_link_destroy(struct sk_psock *psock)
 657 {
 658         struct sk_psock_link *link, *tmp;
 659
 660         list_for_each_entry_safe(link, tmp, &psock->link, list) {
 661                 list_del(&link->list);
 662                 sk_psock_free_link(link);
 663         }
 664 }
 665
 666 static void sk_psock_destroy_deferred(struct work_struct *gc)
 667 {
 668         struct sk_psock *psock = container_of(gc, struct sk_psock, gc);
 669
 670         /* No sk_callback_lock since already detached. */
 671
 672         /* Parser has been stopped */
 673         if (psock->progs.skb_parser)
 674                 strp_done(&psock->parser.strp);
 675
 676         cancel_work_sync(&psock->work);
 677
 678         psock_progs_drop(&psock->progs);
 679
 680         sk_psock_link_destroy(psock);
 681         sk_psock_cork_free(psock);
 682         sk_psock_zap_ingress(psock);
 683
 684         if (psock->sk_redir)
 685                 sock_put(psock->sk_redir);
 686         sock_put(psock->sk);
 687         kfree(psock);
 688 }
 689
 690 static void sk_psock_destroy(struct rcu_head *rcu)
 691 {
 692         struct sk_psock *psock = container_of(rcu, struct sk_psock, rcu);
 693
 694         INIT_WORK(&psock->gc, sk_psock_destroy_deferred);
 695         schedule_work(&psock->gc);
 696 }
 697
 698 void sk_psock_drop(struct sock *sk, struct sk_psock *psock)
 699 {
 700         sk_psock_cork_free(psock);
 701         sk_psock_zap_ingress(psock);
 702
 703         write_lock_bh(&sk->sk_callback_lock);
 704         sk_psock_restore_proto(sk, psock);
 705         rcu_assign_sk_user_data(sk, NULL);
 706         if (psock->progs.skb_parser)
 707                 sk_psock_stop_strp(sk, psock);
 708         else if (psock->progs.skb_verdict)
 709                 sk_psock_stop_verdict(sk, psock);
 710         write_unlock_bh(&sk->sk_callback_lock);
 711         sk_psock_clear_state(psock, SK_PSOCK_TX_ENABLED);
 712
 713         call_rcu(&psock->rcu, sk_psock_destroy);
 714 }
 715 EXPORT_SYMBOL_GPL(sk_psock_drop);
 716
 717 static int sk_psock_map_verd(int verdict, bool redir)
 718 {
 719         switch (verdict) {
 720         case SK_PASS:
 721                 return redir ? __SK_REDIRECT : __SK_PASS;
 722         case SK_DROP:
 723         default:
 724                 break;
 725         }
 726
 727         return __SK_DROP;
 728 }
 729
 730 int sk_psock_msg_verdict(struct sock *sk, struct sk_psock *psock,
 731                          struct sk_msg *msg)
 732 {
 733         struct bpf_prog *prog;
 734         int ret;
 735
 736         rcu_read_lock();
 737         prog = READ_ONCE(psock->progs.msg_parser);
 738         if (unlikely(!prog)) {
 739                 ret = __SK_PASS;
 740                 goto out;
 741         }
 742
 743         sk_msg_compute_data_pointers(msg);
 744         msg->sk = sk;
 745         ret = bpf_prog_run_pin_on_cpu(prog, msg);
 746         ret = sk_psock_map_verd(ret, msg->sk_redir);
 747         psock->apply_bytes = msg->apply_bytes;
 748         if (ret == __SK_REDIRECT) {
 749                 if (psock->sk_redir)
 750                         sock_put(psock->sk_redir);
 751                 psock->sk_redir = msg->sk_redir;
 752                 if (!psock->sk_redir) {
 753                         ret = __SK_DROP;
 754                         goto out;
 755                 }
 756                 sock_hold(psock->sk_redir);
 757         }
 758 out:
 759         rcu_read_unlock();
 760         return ret;
 761 }
 762 EXPORT_SYMBOL_GPL(sk_psock_msg_verdict);
 763
 764 static int sk_psock_bpf_run(struct sk_psock *psock, struct bpf_prog *prog,
 765                             struct sk_buff *skb)
 766 {
 767         bpf_compute_data_end_sk_skb(skb);
 768         return bpf_prog_run_pin_on_cpu(prog, skb);
 769 }
 770
 771 static struct sk_psock *sk_psock_from_strp(struct strparser *strp)
 772 {
 773         struct sk_psock_parser *parser;
 774
 775         parser = container_of(strp, struct sk_psock_parser, strp);
 776         return container_of(parser, struct sk_psock, parser);
 777 }
 778
 779 static void sk_psock_skb_redirect(struct sk_buff *skb)
 780 {
 781         struct sk_psock *psock_other;
 782         struct sock *sk_other;
 783
 784         sk_other = tcp_skb_bpf_redirect_fetch(skb);
 785         /* This error is a buggy BPF program, it returned a redirect
 786          * return code, but then didn't set a redirect interface.
 787          */
 788         if (unlikely(!sk_other)) {
 789                 kfree_skb(skb);
 790                 return;
 791         }
 792         psock_other = sk_psock(sk_other);
 793         /* This error indicates the socket is being torn down or had another
 794          * error that caused the pipe to break. We can't send a packet on
 795          * a socket that is in this state so we drop the skb.
 796          */
 797         if (!psock_other || sock_flag(sk_other, SOCK_DEAD) ||
 798             !sk_psock_test_state(psock_other, SK_PSOCK_TX_ENABLED)) {
 799                 kfree_skb(skb);
 800                 return;
 801         }
 802
 803         skb_queue_tail(&psock_other->ingress_skb, skb);
 804         schedule_work(&psock_other->work);
 805 }
 806
 807 static void sk_psock_tls_verdict_apply(struct sk_buff *skb, struct sock *sk, int verdict)
 808 {
 809         switch (verdict) {
 810         case __SK_REDIRECT:
 811                 sk_psock_skb_redirect(skb);
 812                 break;
 813         case __SK_PASS:
 814         case __SK_DROP:
 815         default:
 816                 break;
 817         }
 818 }
 819
 820 int sk_psock_tls_strp_read(struct sk_psock *psock, struct sk_buff *skb)
 821 {
 822         struct bpf_prog *prog;
 823         int ret = __SK_PASS;
 824
 825         rcu_read_lock();
 826         prog = READ_ONCE(psock->progs.skb_verdict);
 827         if (likely(prog)) {
 828                 skb->sk = psock->sk;
 829                 tcp_skb_bpf_redirect_clear(skb);
 830                 ret = sk_psock_bpf_run(psock, prog, skb);
 831                 ret = sk_psock_map_verd(ret, tcp_skb_bpf_redirect_fetch(skb));
 832                 skb->sk = NULL;
 833         }
 834         sk_psock_tls_verdict_apply(skb, psock->sk, ret);
 835         rcu_read_unlock();
 836         return ret;
 837 }
 838 EXPORT_SYMBOL_GPL(sk_psock_tls_strp_read);
 839
 840 static void sk_psock_verdict_apply(struct sk_psock *psock,
 841                                    struct sk_buff *skb, int verdict)
 842 {
 843         struct tcp_skb_cb *tcp;
 844         struct sock *sk_other;
 845         int err = -EIO;
 846
 847         switch (verdict) {
 848         case __SK_PASS:
 849                 sk_other = psock->sk;
 850                 if (sock_flag(sk_other, SOCK_DEAD) ||
 851                     !sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
 852                         goto out_free;
 853                 }
 854
 855                 tcp = TCP_SKB_CB(skb);
 856                 tcp->bpf.flags |= BPF_F_INGRESS;
 857
 858                 /* If the queue is empty then we can submit directly
 859                  * into the msg queue. If its not empty we have to
 860                  * queue work otherwise we may get OOO data. Otherwise,
 861                  * if sk_psock_skb_ingress errors will be handled by
 862                  * retrying later from workqueue.
 863                  */
 864                 if (skb_queue_empty(&psock->ingress_skb)) {
 865                         err = sk_psock_skb_ingress_self(psock, skb);
 866                 }
 867                 if (err < 0) {
 868                         skb_queue_tail(&psock->ingress_skb, skb);
 869                         schedule_work(&psock->work);
 870                 }
 871                 break;
 872         case __SK_REDIRECT:
 873                 sk_psock_skb_redirect(skb);
 874                 break;
 875         case __SK_DROP:
 876         default:
 877 out_free:
 878                 kfree_skb(skb);
 879         }
 880 }
 881
 882 static void sk_psock_strp_read(struct strparser *strp, struct sk_buff *skb)
 883 {
 884         struct sk_psock *psock;
 885         struct bpf_prog *prog;
 886         int ret = __SK_DROP;
 887         struct sock *sk;
 888
 889         rcu_read_lock();
 890         sk = strp->sk;
 891         psock = sk_psock(sk);
 892         if (unlikely(!psock)) {
 893                 kfree_skb(skb);
 894                 goto out;
 895         }
 896         prog = READ_ONCE(psock->progs.skb_verdict);
 897         if (likely(prog)) {
 898                 skb->sk = sk;
 899                 tcp_skb_bpf_redirect_clear(skb);
 900                 ret = sk_psock_bpf_run(psock, prog, skb);
 901                 ret = sk_psock_map_verd(ret, tcp_skb_bpf_redirect_fetch(skb));
 902                 skb->sk = NULL;
 903         }
 904         sk_psock_verdict_apply(psock, skb, ret);
 905 out:
 906         rcu_read_unlock();
 907 }
 908
 909 static int sk_psock_strp_read_done(struct strparser *strp, int err)
 910 {
 911         return err;
 912 }
 913
 914 static int sk_psock_strp_parse(struct strparser *strp, struct sk_buff *skb)
 915 {
 916         struct sk_psock *psock = sk_psock_from_strp(strp);
 917         struct bpf_prog *prog;
 918         int ret = skb->len;
 919
 920         rcu_read_lock();
 921         prog = READ_ONCE(psock->progs.skb_parser);
 922         if (likely(prog)) {
 923                 skb->sk = psock->sk;
 924                 ret = sk_psock_bpf_run(psock, prog, skb);
 925                 skb->sk = NULL;
 926         }
 927         rcu_read_unlock();
 928         return ret;
 929 }
 930
 931 /* Called with socket lock held. */
 932 static void sk_psock_strp_data_ready(struct sock *sk)
 933 {
 934         struct sk_psock *psock;
 935
 936         rcu_read_lock();
 937         psock = sk_psock(sk);
 938         if (likely(psock)) {
 939                 if (tls_sw_has_ctx_rx(sk)) {
 940                         psock->parser.saved_data_ready(sk);
 941                 } else {
 942                         write_lock_bh(&sk->sk_callback_lock);
 943                         strp_data_ready(&psock->parser.strp);
 944                         write_unlock_bh(&sk->sk_callback_lock);
 945                 }
 946         }
 947         rcu_read_unlock();
 948 }
 949
 950 static int sk_psock_verdict_recv(read_descriptor_t *desc, struct sk_buff *skb,
 951                                  unsigned int offset, size_t orig_len)
 952 {
 953         struct sock *sk = (struct sock *)desc->arg.data;
 954         struct sk_psock *psock;
 955         struct bpf_prog *prog;
 956         int ret = __SK_DROP;
 957         int len = orig_len;
 958
 959         /* clone here so sk_eat_skb() in tcp_read_sock does not drop our data */
 960         skb = skb_clone(skb, GFP_ATOMIC);
 961         if (!skb) {
 962                 desc->error = -ENOMEM;
 963                 return 0;
 964         }
 965
 966         rcu_read_lock();
 967         psock = sk_psock(sk);
 968         if (unlikely(!psock)) {
 969                 len = 0;
 970                 kfree_skb(skb);
 971                 goto out;
 972         }
 973         prog = READ_ONCE(psock->progs.skb_verdict);
 974         if (likely(prog)) {
 975                 skb->sk = sk;
 976                 tcp_skb_bpf_redirect_clear(skb);
 977                 ret = sk_psock_bpf_run(psock, prog, skb);
 978                 ret = sk_psock_map_verd(ret, tcp_skb_bpf_redirect_fetch(skb));
 979                 skb->sk = NULL;
 980         }
 981         sk_psock_verdict_apply(psock, skb, ret);
 982 out:
 983         rcu_read_unlock();
 984         return len;
 985 }
 986
 987 static void sk_psock_verdict_data_ready(struct sock *sk)
 988 {
 989         struct socket *sock = sk->sk_socket;
 990         read_descriptor_t desc;
 991
 992         if (unlikely(!sock || !sock->ops || !sock->ops->read_sock))
 993                 return;
 994
 995         desc.arg.data = sk;
 996         desc.error = 0;
 997         desc.count = 1;
 998
 999         sock->ops->read_sock(sk, &desc, sk_psock_verdict_recv);
1000 }
1001
1002 static void sk_psock_write_space(struct sock *sk)
1003 {
1004         struct sk_psock *psock;
1005         void (*write_space)(struct sock *sk) = NULL;
1006
1007         rcu_read_lock();
1008         psock = sk_psock(sk);
1009         if (likely(psock)) {
1010                 if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED))
1011                         schedule_work(&psock->work);
1012                 write_space = psock->saved_write_space;
1013         }
1014         rcu_read_unlock();
1015         if (write_space)
1016                 write_space(sk);
1017 }
1018
1019 int sk_psock_init_strp(struct sock *sk, struct sk_psock *psock)
1020 {
1021         static const struct strp_callbacks cb = {
1022                 .rcv_msg        = sk_psock_strp_read,
1023                 .read_sock_done = sk_psock_strp_read_done,
1024                 .parse_msg      = sk_psock_strp_parse,
1025         };
1026
1027         psock->parser.enabled = false;
1028         return strp_init(&psock->parser.strp, sk, &cb);
1029 }
1030
1031 void sk_psock_start_verdict(struct sock *sk, struct sk_psock *psock)
1032 {
1033         struct sk_psock_parser *parser = &psock->parser;
1034
1035         if (parser->enabled)
1036                 return;
1037
1038         parser->saved_data_ready = sk->sk_data_ready;
1039         sk->sk_data_ready = sk_psock_verdict_data_ready;
1040         sk->sk_write_space = sk_psock_write_space;
1041         parser->enabled = true;
1042 }
1043
1044 void sk_psock_start_strp(struct sock *sk, struct sk_psock *psock)
1045 {
1046         struct sk_psock_parser *parser = &psock->parser;
1047
1048         if (parser->enabled)
1049                 return;
1050
1051         parser->saved_data_ready = sk->sk_data_ready;
1052         sk->sk_data_ready = sk_psock_strp_data_ready;
1053         sk->sk_write_space = sk_psock_write_space;
1054         parser->enabled = true;
1055 }
1056
1057 void sk_psock_stop_strp(struct sock *sk, struct sk_psock *psock)
1058 {
1059         struct sk_psock_parser *parser = &psock->parser;
1060
1061         if (!parser->enabled)
1062                 return;
1063
1064         sk->sk_data_ready = parser->saved_data_ready;
1065         parser->saved_data_ready = NULL;
1066         strp_stop(&parser->strp);
1067         parser->enabled = false;
1068 }
1069
1070 void sk_psock_stop_verdict(struct sock *sk, struct sk_psock *psock)
1071 {
1072         struct sk_psock_parser *parser = &psock->parser;
1073
1074         if (!parser->enabled)
1075                 return;
1076
1077         sk->sk_data_ready = parser->saved_data_ready;
1078         parser->saved_data_ready = NULL;
1079         parser->enabled = false;
1080 }