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