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