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