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