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 int sk_msg_wait_data(struct sock *sk, struct sk_psock *psock, int flags,
 403                      long timeo, int *err)
 404 {
 405         DEFINE_WAIT_FUNC(wait, woken_wake_function);
 406         int ret = 0;
 407
 408         if (sk->sk_shutdown & RCV_SHUTDOWN)
 409                 return 1;
 410
 411         if (!timeo)
 412                 return ret;
 413
 414         add_wait_queue(sk_sleep(sk), &wait);
 415         sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
 416         ret = sk_wait_event(sk, &timeo,
 417                             !list_empty(&psock->ingress_msg) ||
 418                             !skb_queue_empty(&sk->sk_receive_queue), &wait);
 419         sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
 420         remove_wait_queue(sk_sleep(sk), &wait);
 421         return ret;
 422 }
 423 EXPORT_SYMBOL_GPL(sk_msg_wait_data);
 424
 425 /* Receive sk_msg from psock->ingress_msg to @msg. */
 426 int sk_msg_recvmsg(struct sock *sk, struct sk_psock *psock, struct msghdr *msg,
 427                    int len, int flags)
 428 {
 429         struct iov_iter *iter = &msg->msg_iter;
 430         int peek = flags & MSG_PEEK;
 431         struct sk_msg *msg_rx;
 432         int i, copied = 0;
 433
 434         msg_rx = sk_psock_peek_msg(psock);
 435         while (copied != len) {
 436                 struct scatterlist *sge;
 437
 438                 if (unlikely(!msg_rx))
 439                         break;
 440
 441                 i = msg_rx->sg.start;
 442                 do {
 443                         struct page *page;
 444                         int copy;
 445
 446                         sge = sk_msg_elem(msg_rx, i);
 447                         copy = sge->length;
 448                         page = sg_page(sge);
 449                         if (copied + copy > len)
 450                                 copy = len - copied;
 451                         copy = copy_page_to_iter(page, sge->offset, copy, iter);
 452                         if (!copy)
 453                                 return copied ? copied : -EFAULT;
 454
 455                         copied += copy;
 456                         if (likely(!peek)) {
 457                                 sge->offset += copy;
 458                                 sge->length -= copy;
 459                                 if (!msg_rx->skb)
 460                                         sk_mem_uncharge(sk, copy);
 461                                 msg_rx->sg.size -= copy;
 462
 463                                 if (!sge->length) {
 464                                         sk_msg_iter_var_next(i);
 465                                         if (!msg_rx->skb)
 466                                                 put_page(page);
 467                                 }
 468                         } else {
 469                                 /* Lets not optimize peek case if copy_page_to_iter
 470                                  * didn't copy the entire length lets just break.
 471                                  */
 472                                 if (copy != sge->length)
 473                                         return copied;
 474                                 sk_msg_iter_var_next(i);
 475                         }
 476
 477                         if (copied == len)
 478                                 break;
 479                 } while (i != msg_rx->sg.end);
 480
 481                 if (unlikely(peek)) {
 482                         msg_rx = sk_psock_next_msg(psock, msg_rx);
 483                         if (!msg_rx)
 484                                 break;
 485                         continue;
 486                 }
 487
 488                 msg_rx->sg.start = i;
 489                 if (!sge->length && msg_rx->sg.start == msg_rx->sg.end) {
 490                         msg_rx = sk_psock_dequeue_msg(psock);
 491                         kfree_sk_msg(msg_rx);
 492                 }
 493                 msg_rx = sk_psock_peek_msg(psock);
 494         }
 495
 496         return copied;
 497 }
 498 EXPORT_SYMBOL_GPL(sk_msg_recvmsg);
 499
 500 static struct sk_msg *sk_psock_create_ingress_msg(struct sock *sk,
 501                                                   struct sk_buff *skb)
 502 {
 503         struct sk_msg *msg;
 504
 505         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
 506                 return NULL;
 507
 508         if (!sk_rmem_schedule(sk, skb, skb->truesize))
 509                 return NULL;
 510
 511         msg = kzalloc(sizeof(*msg), __GFP_NOWARN | GFP_KERNEL);
 512         if (unlikely(!msg))
 513                 return NULL;
 514
 515         sk_msg_init(msg);
 516         return msg;
 517 }
 518
 519 static int sk_psock_skb_ingress_enqueue(struct sk_buff *skb,
 520                                         struct sk_psock *psock,
 521                                         struct sock *sk,
 522                                         struct sk_msg *msg)
 523 {
 524         int num_sge, copied;
 525
 526         /* skb linearize may fail with ENOMEM, but lets simply try again
 527          * later if this happens. Under memory pressure we don't want to
 528          * drop the skb. We need to linearize the skb so that the mapping
 529          * in skb_to_sgvec can not error.
 530          */
 531         if (skb_linearize(skb))
 532                 return -EAGAIN;
 533         num_sge = skb_to_sgvec(skb, msg->sg.data, 0, skb->len);
 534         if (unlikely(num_sge < 0))
 535                 return num_sge;
 536
 537         copied = skb->len;
 538         msg->sg.start = 0;
 539         msg->sg.size = copied;
 540         msg->sg.end = num_sge;
 541         msg->skb = skb;
 542
 543         sk_psock_queue_msg(psock, msg);
 544         sk_psock_data_ready(sk, psock);
 545         return copied;
 546 }
 547
 548 static int sk_psock_skb_ingress_self(struct sk_psock *psock, struct sk_buff *skb);
 549
 550 static int sk_psock_skb_ingress(struct sk_psock *psock, struct sk_buff *skb)
 551 {
 552         struct sock *sk = psock->sk;
 553         struct sk_msg *msg;
 554         int err;
 555
 556         /* If we are receiving on the same sock skb->sk is already assigned,
 557          * skip memory accounting and owner transition seeing it already set
 558          * correctly.
 559          */
 560         if (unlikely(skb->sk == sk))
 561                 return sk_psock_skb_ingress_self(psock, skb);
 562         msg = sk_psock_create_ingress_msg(sk, skb);
 563         if (!msg)
 564                 return -EAGAIN;
 565
 566         /* This will transition ownership of the data from the socket where
 567          * the BPF program was run initiating the redirect to the socket
 568          * we will eventually receive this data on. The data will be released
 569          * from skb_consume found in __tcp_bpf_recvmsg() after its been copied
 570          * into user buffers.
 571          */
 572         skb_set_owner_r(skb, sk);
 573         err = sk_psock_skb_ingress_enqueue(skb, psock, sk, msg);
 574         if (err < 0)
 575                 kfree(msg);
 576         return err;
 577 }
 578
 579 /* Puts an skb on the ingress queue of the socket already assigned to the
 580  * skb. In this case we do not need to check memory limits or skb_set_owner_r
 581  * because the skb is already accounted for here.
 582  */
 583 static int sk_psock_skb_ingress_self(struct sk_psock *psock, struct sk_buff *skb)
 584 {
 585         struct sk_msg *msg = kzalloc(sizeof(*msg), __GFP_NOWARN | GFP_ATOMIC);
 586         struct sock *sk = psock->sk;
 587         int err;
 588
 589         if (unlikely(!msg))
 590                 return -EAGAIN;
 591         sk_msg_init(msg);
 592         skb_set_owner_r(skb, sk);
 593         err = sk_psock_skb_ingress_enqueue(skb, psock, sk, msg);
 594         if (err < 0)
 595                 kfree(msg);
 596         return err;
 597 }
 598
 599 static int sk_psock_handle_skb(struct sk_psock *psock, struct sk_buff *skb,
 600                                u32 off, u32 len, bool ingress)
 601 {
 602         if (!ingress) {
 603                 if (!sock_writeable(psock->sk))
 604                         return -EAGAIN;
 605                 return skb_send_sock(psock->sk, skb, off, len);
 606         }
 607         return sk_psock_skb_ingress(psock, skb);
 608 }
 609
 610 static void sock_drop(struct sock *sk, struct sk_buff *skb)
 611 {
 612         sk_drops_add(sk, skb);
 613         kfree_skb(skb);
 614 }
 615
 616 static void sk_psock_skb_state(struct sk_psock *psock,
 617                                struct sk_psock_work_state *state,
 618                                struct sk_buff *skb,
 619                                int len, int off)
 620 {
 621         spin_lock_bh(&psock->ingress_lock);
 622         if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
 623                 state->skb = skb;
 624                 state->len = len;
 625                 state->off = off;
 626         } else {
 627                 sock_drop(psock->sk, skb);
 628         }
 629         spin_unlock_bh(&psock->ingress_lock);
 630 }
 631
 632 static void sk_psock_backlog(struct work_struct *work)
 633 {
 634         struct sk_psock *psock = container_of(work, struct sk_psock, work);
 635         struct sk_psock_work_state *state = &psock->work_state;
 636         struct sk_buff *skb = NULL;
 637         bool ingress;
 638         u32 len, off;
 639         int ret;
 640
 641         mutex_lock(&psock->work_mutex);
 642         if (unlikely(state->skb)) {
 643                 spin_lock_bh(&psock->ingress_lock);
 644                 skb = state->skb;
 645                 len = state->len;
 646                 off = state->off;
 647                 state->skb = NULL;
 648                 spin_unlock_bh(&psock->ingress_lock);
 649         }
 650         if (skb)
 651                 goto start;
 652
 653         while ((skb = skb_dequeue(&psock->ingress_skb))) {
 654                 len = skb->len;
 655                 off = 0;
 656 start:
 657                 ingress = skb_bpf_ingress(skb);
 658                 skb_bpf_redirect_clear(skb);
 659                 do {
 660                         ret = -EIO;
 661                         if (!sock_flag(psock->sk, SOCK_DEAD))
 662                                 ret = sk_psock_handle_skb(psock, skb, off,
 663                                                           len, ingress);
 664                         if (ret <= 0) {
 665                                 if (ret == -EAGAIN) {
 666                                         sk_psock_skb_state(psock, state, skb,
 667                                                            len, off);
 668                                         goto end;
 669                                 }
 670                                 /* Hard errors break pipe and stop xmit. */
 671                                 sk_psock_report_error(psock, ret ? -ret : EPIPE);
 672                                 sk_psock_clear_state(psock, SK_PSOCK_TX_ENABLED);
 673                                 sock_drop(psock->sk, skb);
 674                                 goto end;
 675                         }
 676                         off += ret;
 677                         len -= ret;
 678                 } while (len);
 679
 680                 if (!ingress)
 681                         kfree_skb(skb);
 682         }
 683 end:
 684         mutex_unlock(&psock->work_mutex);
 685 }
 686
 687 struct sk_psock *sk_psock_init(struct sock *sk, int node)
 688 {
 689         struct sk_psock *psock;
 690         struct proto *prot;
 691
 692         write_lock_bh(&sk->sk_callback_lock);
 693
 694         if (sk->sk_user_data) {
 695                 psock = ERR_PTR(-EBUSY);
 696                 goto out;
 697         }
 698
 699         psock = kzalloc_node(sizeof(*psock), GFP_ATOMIC | __GFP_NOWARN, node);
 700         if (!psock) {
 701                 psock = ERR_PTR(-ENOMEM);
 702                 goto out;
 703         }
 704
 705         prot = READ_ONCE(sk->sk_prot);
 706         psock->sk = sk;
 707         psock->eval = __SK_NONE;
 708         psock->sk_proto = prot;
 709         psock->saved_unhash = prot->unhash;
 710         psock->saved_close = prot->close;
 711         psock->saved_write_space = sk->sk_write_space;
 712
 713         INIT_LIST_HEAD(&psock->link);
 714         spin_lock_init(&psock->link_lock);
 715
 716         INIT_WORK(&psock->work, sk_psock_backlog);
 717         mutex_init(&psock->work_mutex);
 718         INIT_LIST_HEAD(&psock->ingress_msg);
 719         spin_lock_init(&psock->ingress_lock);
 720         skb_queue_head_init(&psock->ingress_skb);
 721
 722         sk_psock_set_state(psock, SK_PSOCK_TX_ENABLED);
 723         refcount_set(&psock->refcnt, 1);
 724
 725         rcu_assign_sk_user_data_nocopy(sk, psock);
 726         sock_hold(sk);
 727
 728 out:
 729         write_unlock_bh(&sk->sk_callback_lock);
 730         return psock;
 731 }
 732 EXPORT_SYMBOL_GPL(sk_psock_init);
 733
 734 struct sk_psock_link *sk_psock_link_pop(struct sk_psock *psock)
 735 {
 736         struct sk_psock_link *link;
 737
 738         spin_lock_bh(&psock->link_lock);
 739         link = list_first_entry_or_null(&psock->link, struct sk_psock_link,
 740                                         list);
 741         if (link)
 742                 list_del(&link->list);
 743         spin_unlock_bh(&psock->link_lock);
 744         return link;
 745 }
 746
 747 static void __sk_psock_purge_ingress_msg(struct sk_psock *psock)
 748 {
 749         struct sk_msg *msg, *tmp;
 750
 751         list_for_each_entry_safe(msg, tmp, &psock->ingress_msg, list) {
 752                 list_del(&msg->list);
 753                 sk_msg_free(psock->sk, msg);
 754                 kfree(msg);
 755         }
 756 }
 757
 758 static void __sk_psock_zap_ingress(struct sk_psock *psock)
 759 {
 760         struct sk_buff *skb;
 761
 762         while ((skb = skb_dequeue(&psock->ingress_skb)) != NULL) {
 763                 skb_bpf_redirect_clear(skb);
 764                 sock_drop(psock->sk, skb);
 765         }
 766         kfree_skb(psock->work_state.skb);
 767         /* We null the skb here to ensure that calls to sk_psock_backlog
 768          * do not pick up the free'd skb.
 769          */
 770         psock->work_state.skb = NULL;
 771         __sk_psock_purge_ingress_msg(psock);
 772 }
 773
 774 static void sk_psock_link_destroy(struct sk_psock *psock)
 775 {
 776         struct sk_psock_link *link, *tmp;
 777
 778         list_for_each_entry_safe(link, tmp, &psock->link, list) {
 779                 list_del(&link->list);
 780                 sk_psock_free_link(link);
 781         }
 782 }
 783
 784 void sk_psock_stop(struct sk_psock *psock, bool wait)
 785 {
 786         spin_lock_bh(&psock->ingress_lock);
 787         sk_psock_clear_state(psock, SK_PSOCK_TX_ENABLED);
 788         sk_psock_cork_free(psock);
 789         __sk_psock_zap_ingress(psock);
 790         spin_unlock_bh(&psock->ingress_lock);
 791
 792         if (wait)
 793                 cancel_work_sync(&psock->work);
 794 }
 795
 796 static void sk_psock_done_strp(struct sk_psock *psock);
 797
 798 static void sk_psock_destroy(struct work_struct *work)
 799 {
 800         struct sk_psock *psock = container_of(to_rcu_work(work),
 801                                               struct sk_psock, rwork);
 802         /* No sk_callback_lock since already detached. */
 803
 804         sk_psock_done_strp(psock);
 805
 806         cancel_work_sync(&psock->work);
 807         mutex_destroy(&psock->work_mutex);
 808
 809         psock_progs_drop(&psock->progs);
 810
 811         sk_psock_link_destroy(psock);
 812         sk_psock_cork_free(psock);
 813
 814         if (psock->sk_redir)
 815                 sock_put(psock->sk_redir);
 816         sock_put(psock->sk);
 817         kfree(psock);
 818 }
 819
 820 void sk_psock_drop(struct sock *sk, struct sk_psock *psock)
 821 {
 822         write_lock_bh(&sk->sk_callback_lock);
 823         sk_psock_restore_proto(sk, psock);
 824         rcu_assign_sk_user_data(sk, NULL);
 825         if (psock->progs.stream_parser)
 826                 sk_psock_stop_strp(sk, psock);
 827         else if (psock->progs.stream_verdict || psock->progs.skb_verdict)
 828                 sk_psock_stop_verdict(sk, psock);
 829         write_unlock_bh(&sk->sk_callback_lock);
 830
 831         sk_psock_stop(psock, false);
 832
 833         INIT_RCU_WORK(&psock->rwork, sk_psock_destroy);
 834         queue_rcu_work(system_wq, &psock->rwork);
 835 }
 836 EXPORT_SYMBOL_GPL(sk_psock_drop);
 837
 838 static int sk_psock_map_verd(int verdict, bool redir)
 839 {
 840         switch (verdict) {
 841         case SK_PASS:
 842                 return redir ? __SK_REDIRECT : __SK_PASS;
 843         case SK_DROP:
 844         default:
 845                 break;
 846         }
 847
 848         return __SK_DROP;
 849 }
 850
 851 int sk_psock_msg_verdict(struct sock *sk, struct sk_psock *psock,
 852                          struct sk_msg *msg)
 853 {
 854         struct bpf_prog *prog;
 855         int ret;
 856
 857         rcu_read_lock();
 858         prog = READ_ONCE(psock->progs.msg_parser);
 859         if (unlikely(!prog)) {
 860                 ret = __SK_PASS;
 861                 goto out;
 862         }
 863
 864         sk_msg_compute_data_pointers(msg);
 865         msg->sk = sk;
 866         ret = bpf_prog_run_pin_on_cpu(prog, msg);
 867         ret = sk_psock_map_verd(ret, msg->sk_redir);
 868         psock->apply_bytes = msg->apply_bytes;
 869         if (ret == __SK_REDIRECT) {
 870                 if (psock->sk_redir)
 871                         sock_put(psock->sk_redir);
 872                 psock->sk_redir = msg->sk_redir;
 873                 if (!psock->sk_redir) {
 874                         ret = __SK_DROP;
 875                         goto out;
 876                 }
 877                 sock_hold(psock->sk_redir);
 878         }
 879 out:
 880         rcu_read_unlock();
 881         return ret;
 882 }
 883 EXPORT_SYMBOL_GPL(sk_psock_msg_verdict);
 884
 885 static int sk_psock_skb_redirect(struct sk_psock *from, struct sk_buff *skb)
 886 {
 887         struct sk_psock *psock_other;
 888         struct sock *sk_other;
 889
 890         sk_other = skb_bpf_redirect_fetch(skb);
 891         /* This error is a buggy BPF program, it returned a redirect
 892          * return code, but then didn't set a redirect interface.
 893          */
 894         if (unlikely(!sk_other)) {
 895                 sock_drop(from->sk, skb);
 896                 return -EIO;
 897         }
 898         psock_other = sk_psock(sk_other);
 899         /* This error indicates the socket is being torn down or had another
 900          * error that caused the pipe to break. We can't send a packet on
 901          * a socket that is in this state so we drop the skb.
 902          */
 903         if (!psock_other || sock_flag(sk_other, SOCK_DEAD)) {
 904                 skb_bpf_redirect_clear(skb);
 905                 sock_drop(from->sk, skb);
 906                 return -EIO;
 907         }
 908         spin_lock_bh(&psock_other->ingress_lock);
 909         if (!sk_psock_test_state(psock_other, SK_PSOCK_TX_ENABLED)) {
 910                 spin_unlock_bh(&psock_other->ingress_lock);
 911                 skb_bpf_redirect_clear(skb);
 912                 sock_drop(from->sk, skb);
 913                 return -EIO;
 914         }
 915
 916         skb_queue_tail(&psock_other->ingress_skb, skb);
 917         schedule_work(&psock_other->work);
 918         spin_unlock_bh(&psock_other->ingress_lock);
 919         return 0;
 920 }
 921
 922 static void sk_psock_tls_verdict_apply(struct sk_buff *skb,
 923                                        struct sk_psock *from, int verdict)
 924 {
 925         switch (verdict) {
 926         case __SK_REDIRECT:
 927                 sk_psock_skb_redirect(from, skb);
 928                 break;
 929         case __SK_PASS:
 930         case __SK_DROP:
 931         default:
 932                 break;
 933         }
 934 }
 935
 936 int sk_psock_tls_strp_read(struct sk_psock *psock, struct sk_buff *skb)
 937 {
 938         struct bpf_prog *prog;
 939         int ret = __SK_PASS;
 940
 941         rcu_read_lock();
 942         prog = READ_ONCE(psock->progs.stream_verdict);
 943         if (likely(prog)) {
 944                 skb->sk = psock->sk;
 945                 skb_dst_drop(skb);
 946                 skb_bpf_redirect_clear(skb);
 947                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
 948                 ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb));
 949                 skb->sk = NULL;
 950         }
 951         sk_psock_tls_verdict_apply(skb, psock, ret);
 952         rcu_read_unlock();
 953         return ret;
 954 }
 955 EXPORT_SYMBOL_GPL(sk_psock_tls_strp_read);
 956
 957 static int sk_psock_verdict_apply(struct sk_psock *psock, struct sk_buff *skb,
 958                                   int verdict)
 959 {
 960         struct sock *sk_other;
 961         int err = 0;
 962
 963         switch (verdict) {
 964         case __SK_PASS:
 965                 err = -EIO;
 966                 sk_other = psock->sk;
 967                 if (sock_flag(sk_other, SOCK_DEAD) ||
 968                     !sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
 969                         goto out_free;
 970                 }
 971
 972                 skb_bpf_set_ingress(skb);
 973
 974                 /* If the queue is empty then we can submit directly
 975                  * into the msg queue. If its not empty we have to
 976                  * queue work otherwise we may get OOO data. Otherwise,
 977                  * if sk_psock_skb_ingress errors will be handled by
 978                  * retrying later from workqueue.
 979                  */
 980                 if (skb_queue_empty(&psock->ingress_skb)) {
 981                         err = sk_psock_skb_ingress_self(psock, skb);
 982                 }
 983                 if (err < 0) {
 984                         spin_lock_bh(&psock->ingress_lock);
 985                         if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
 986                                 skb_queue_tail(&psock->ingress_skb, skb);
 987                                 schedule_work(&psock->work);
 988                                 err = 0;
 989                         }
 990                         spin_unlock_bh(&psock->ingress_lock);
 991                         if (err < 0) {
 992                                 skb_bpf_redirect_clear(skb);
 993                                 goto out_free;
 994                         }
 995                 }
 996                 break;
 997         case __SK_REDIRECT:
 998                 err = sk_psock_skb_redirect(psock, skb);
 999                 break;
1000         case __SK_DROP:
1001         default:
1002 out_free:
1003                 sock_drop(psock->sk, skb);
1004         }
1005
1006         return err;
1007 }
1008
1009 static void sk_psock_write_space(struct sock *sk)
1010 {
1011         struct sk_psock *psock;
1012         void (*write_space)(struct sock *sk) = NULL;
1013
1014         rcu_read_lock();
1015         psock = sk_psock(sk);
1016         if (likely(psock)) {
1017                 if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED))
1018                         schedule_work(&psock->work);
1019                 write_space = psock->saved_write_space;
1020         }
1021         rcu_read_unlock();
1022         if (write_space)
1023                 write_space(sk);
1024 }
1025
1026 #if IS_ENABLED(CONFIG_BPF_STREAM_PARSER)
1027 static void sk_psock_strp_read(struct strparser *strp, struct sk_buff *skb)
1028 {
1029         struct sk_psock *psock;
1030         struct bpf_prog *prog;
1031         int ret = __SK_DROP;
1032         struct sock *sk;
1033
1034         rcu_read_lock();
1035         sk = strp->sk;
1036         psock = sk_psock(sk);
1037         if (unlikely(!psock)) {
1038                 sock_drop(sk, skb);
1039                 goto out;
1040         }
1041         prog = READ_ONCE(psock->progs.stream_verdict);
1042         if (likely(prog)) {
1043                 skb->sk = sk;
1044                 skb_dst_drop(skb);
1045                 skb_bpf_redirect_clear(skb);
1046                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1047                 ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb));
1048                 skb->sk = NULL;
1049         }
1050         sk_psock_verdict_apply(psock, skb, ret);
1051 out:
1052         rcu_read_unlock();
1053 }
1054
1055 static int sk_psock_strp_read_done(struct strparser *strp, int err)
1056 {
1057         return err;
1058 }
1059
1060 static int sk_psock_strp_parse(struct strparser *strp, struct sk_buff *skb)
1061 {
1062         struct sk_psock *psock = container_of(strp, struct sk_psock, strp);
1063         struct bpf_prog *prog;
1064         int ret = skb->len;
1065
1066         rcu_read_lock();
1067         prog = READ_ONCE(psock->progs.stream_parser);
1068         if (likely(prog)) {
1069                 skb->sk = psock->sk;
1070                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1071                 skb->sk = NULL;
1072         }
1073         rcu_read_unlock();
1074         return ret;
1075 }
1076
1077 /* Called with socket lock held. */
1078 static void sk_psock_strp_data_ready(struct sock *sk)
1079 {
1080         struct sk_psock *psock;
1081
1082         rcu_read_lock();
1083         psock = sk_psock(sk);
1084         if (likely(psock)) {
1085                 if (tls_sw_has_ctx_rx(sk)) {
1086                         psock->saved_data_ready(sk);
1087                 } else {
1088                         write_lock_bh(&sk->sk_callback_lock);
1089                         strp_data_ready(&psock->strp);
1090                         write_unlock_bh(&sk->sk_callback_lock);
1091                 }
1092         }
1093         rcu_read_unlock();
1094 }
1095
1096 int sk_psock_init_strp(struct sock *sk, struct sk_psock *psock)
1097 {
1098         static const struct strp_callbacks cb = {
1099                 .rcv_msg        = sk_psock_strp_read,
1100                 .read_sock_done = sk_psock_strp_read_done,
1101                 .parse_msg      = sk_psock_strp_parse,
1102         };
1103
1104         return strp_init(&psock->strp, sk, &cb);
1105 }
1106
1107 void sk_psock_start_strp(struct sock *sk, struct sk_psock *psock)
1108 {
1109         if (psock->saved_data_ready)
1110                 return;
1111
1112         psock->saved_data_ready = sk->sk_data_ready;
1113         sk->sk_data_ready = sk_psock_strp_data_ready;
1114         sk->sk_write_space = sk_psock_write_space;
1115 }
1116
1117 void sk_psock_stop_strp(struct sock *sk, struct sk_psock *psock)
1118 {
1119         if (!psock->saved_data_ready)
1120                 return;
1121
1122         sk->sk_data_ready = psock->saved_data_ready;
1123         psock->saved_data_ready = NULL;
1124         strp_stop(&psock->strp);
1125 }
1126
1127 static void sk_psock_done_strp(struct sk_psock *psock)
1128 {
1129         /* Parser has been stopped */
1130         if (psock->progs.stream_parser)
1131                 strp_done(&psock->strp);
1132 }
1133 #else
1134 static void sk_psock_done_strp(struct sk_psock *psock)
1135 {
1136 }
1137 #endif /* CONFIG_BPF_STREAM_PARSER */
1138
1139 static int sk_psock_verdict_recv(read_descriptor_t *desc, struct sk_buff *skb,
1140                                  unsigned int offset, size_t orig_len)
1141 {
1142         struct sock *sk = (struct sock *)desc->arg.data;
1143         struct sk_psock *psock;
1144         struct bpf_prog *prog;
1145         int ret = __SK_DROP;
1146         int len = skb->len;
1147
1148         /* clone here so sk_eat_skb() in tcp_read_sock does not drop our data */
1149         skb = skb_clone(skb, GFP_ATOMIC);
1150         if (!skb) {
1151                 desc->error = -ENOMEM;
1152                 return 0;
1153         }
1154
1155         rcu_read_lock();
1156         psock = sk_psock(sk);
1157         if (unlikely(!psock)) {
1158                 len = 0;
1159                 sock_drop(sk, skb);
1160                 goto out;
1161         }
1162         prog = READ_ONCE(psock->progs.stream_verdict);
1163         if (!prog)
1164                 prog = READ_ONCE(psock->progs.skb_verdict);
1165         if (likely(prog)) {
1166                 skb->sk = sk;
1167                 skb_dst_drop(skb);
1168                 skb_bpf_redirect_clear(skb);
1169                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1170                 ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb));
1171                 skb->sk = NULL;
1172         }
1173         if (sk_psock_verdict_apply(psock, skb, ret) < 0)
1174                 len = 0;
1175 out:
1176         rcu_read_unlock();
1177         return len;
1178 }
1179
1180 static void sk_psock_verdict_data_ready(struct sock *sk)
1181 {
1182         struct socket *sock = sk->sk_socket;
1183         read_descriptor_t desc;
1184
1185         if (unlikely(!sock || !sock->ops || !sock->ops->read_sock))
1186                 return;
1187
1188         desc.arg.data = sk;
1189         desc.error = 0;
1190         desc.count = 1;
1191
1192         sock->ops->read_sock(sk, &desc, sk_psock_verdict_recv);
1193 }
1194
1195 void sk_psock_start_verdict(struct sock *sk, struct sk_psock *psock)
1196 {
1197         if (psock->saved_data_ready)
1198                 return;
1199
1200         psock->saved_data_ready = sk->sk_data_ready;
1201         sk->sk_data_ready = sk_psock_verdict_data_ready;
1202         sk->sk_write_space = sk_psock_write_space;
1203 }
1204
1205 void sk_psock_stop_verdict(struct sock *sk, struct sk_psock *psock)
1206 {
1207         if (!psock->saved_data_ready)
1208                 return;
1209
1210         sk->sk_data_ready = psock->saved_data_ready;
1211         psock->saved_data_ready = NULL;
1212 }