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         u32 osize = msg->sg.size;
  31         int ret = 0;
  32
  33         len -= msg->sg.size;
  34         while (len > 0) {
  35                 struct scatterlist *sge;
  36                 u32 orig_offset;
  37                 int use, i;
  38
  39                 if (!sk_page_frag_refill(sk, pfrag)) {
  40                         ret = -ENOMEM;
  41                         goto msg_trim;
  42                 }
  43
  44                 orig_offset = pfrag->offset;
  45                 use = min_t(int, len, pfrag->size - orig_offset);
  46                 if (!sk_wmem_schedule(sk, use)) {
  47                         ret = -ENOMEM;
  48                         goto msg_trim;
  49                 }
  50
  51                 i = msg->sg.end;
  52                 sk_msg_iter_var_prev(i);
  53                 sge = &msg->sg.data[i];
  54
  55                 if (sk_msg_try_coalesce_ok(msg, elem_first_coalesce) &&
  56                     sg_page(sge) == pfrag->page &&
  57                     sge->offset + sge->length == orig_offset) {
  58                         sge->length += use;
  59                 } else {
  60                         if (sk_msg_full(msg)) {
  61                                 ret = -ENOSPC;
  62                                 break;
  63                         }
  64
  65                         sge = &msg->sg.data[msg->sg.end];
  66                         sg_unmark_end(sge);
  67                         sg_set_page(sge, pfrag->page, use, orig_offset);
  68                         get_page(pfrag->page);
  69                         sk_msg_iter_next(msg, end);
  70                 }
  71
  72                 sk_mem_charge(sk, use);
  73                 msg->sg.size += use;
  74                 pfrag->offset += use;
  75                 len -= use;
  76         }
  77
  78         return ret;
  79
  80 msg_trim:
  81         sk_msg_trim(sk, msg, osize);
  82         return ret;
  83 }
  84 EXPORT_SYMBOL_GPL(sk_msg_alloc);
  85
  86 int sk_msg_clone(struct sock *sk, struct sk_msg *dst, struct sk_msg *src,
  87                  u32 off, u32 len)
  88 {
  89         int i = src->sg.start;
  90         struct scatterlist *sge = sk_msg_elem(src, i);
  91         struct scatterlist *sgd = NULL;
  92         u32 sge_len, sge_off;
  93
  94         while (off) {
  95                 if (sge->length > off)
  96                         break;
  97                 off -= sge->length;
  98                 sk_msg_iter_var_next(i);
  99                 if (i == src->sg.end && off)
 100                         return -ENOSPC;
 101                 sge = sk_msg_elem(src, i);
 102         }
 103
 104         while (len) {
 105                 sge_len = sge->length - off;
 106                 if (sge_len > len)
 107                         sge_len = len;
 108
 109                 if (dst->sg.end)
 110                         sgd = sk_msg_elem(dst, dst->sg.end - 1);
 111
 112                 if (sgd &&
 113                     (sg_page(sge) == sg_page(sgd)) &&
 114                     (sg_virt(sge) + off == sg_virt(sgd) + sgd->length)) {
 115                         sgd->length += sge_len;
 116                         dst->sg.size += sge_len;
 117                 } else if (!sk_msg_full(dst)) {
 118                         sge_off = sge->offset + off;
 119                         sk_msg_page_add(dst, sg_page(sge), sge_len, sge_off);
 120                 } else {
 121                         return -ENOSPC;
 122                 }
 123
 124                 off = 0;
 125                 len -= sge_len;
 126                 sk_mem_charge(sk, sge_len);
 127                 sk_msg_iter_var_next(i);
 128                 if (i == src->sg.end && len)
 129                         return -ENOSPC;
 130                 sge = sk_msg_elem(src, i);
 131         }
 132
 133         return 0;
 134 }
 135 EXPORT_SYMBOL_GPL(sk_msg_clone);
 136
 137 void sk_msg_return_zero(struct sock *sk, struct sk_msg *msg, int bytes)
 138 {
 139         int i = msg->sg.start;
 140
 141         do {
 142                 struct scatterlist *sge = sk_msg_elem(msg, i);
 143
 144                 if (bytes < sge->length) {
 145                         sge->length -= bytes;
 146                         sge->offset += bytes;
 147                         sk_mem_uncharge(sk, bytes);
 148                         break;
 149                 }
 150
 151                 sk_mem_uncharge(sk, sge->length);
 152                 bytes -= sge->length;
 153                 sge->length = 0;
 154                 sge->offset = 0;
 155                 sk_msg_iter_var_next(i);
 156         } while (bytes && i != msg->sg.end);
 157         msg->sg.start = i;
 158 }
 159 EXPORT_SYMBOL_GPL(sk_msg_return_zero);
 160
 161 void sk_msg_return(struct sock *sk, struct sk_msg *msg, int bytes)
 162 {
 163         int i = msg->sg.start;
 164
 165         do {
 166                 struct scatterlist *sge = &msg->sg.data[i];
 167                 int uncharge = (bytes < sge->length) ? bytes : sge->length;
 168
 169                 sk_mem_uncharge(sk, uncharge);
 170                 bytes -= uncharge;
 171                 sk_msg_iter_var_next(i);
 172         } while (i != msg->sg.end);
 173 }
 174 EXPORT_SYMBOL_GPL(sk_msg_return);
 175
 176 static int sk_msg_free_elem(struct sock *sk, struct sk_msg *msg, u32 i,
 177                             bool charge)
 178 {
 179         struct scatterlist *sge = sk_msg_elem(msg, i);
 180         u32 len = sge->length;
 181
 182         /* When the skb owns the memory we free it from consume_skb path. */
 183         if (!msg->skb) {
 184                 if (charge)
 185                         sk_mem_uncharge(sk, len);
 186                 put_page(sg_page(sge));
 187         }
 188         memset(sge, 0, sizeof(*sge));
 189         return len;
 190 }
 191
 192 static int __sk_msg_free(struct sock *sk, struct sk_msg *msg, u32 i,
 193                          bool charge)
 194 {
 195         struct scatterlist *sge = sk_msg_elem(msg, i);
 196         int freed = 0;
 197
 198         while (msg->sg.size) {
 199                 msg->sg.size -= sge->length;
 200                 freed += sk_msg_free_elem(sk, msg, i, charge);
 201                 sk_msg_iter_var_next(i);
 202                 sk_msg_check_to_free(msg, i, msg->sg.size);
 203                 sge = sk_msg_elem(msg, i);
 204         }
 205         consume_skb(msg->skb);
 206         sk_msg_init(msg);
 207         return freed;
 208 }
 209
 210 int sk_msg_free_nocharge(struct sock *sk, struct sk_msg *msg)
 211 {
 212         return __sk_msg_free(sk, msg, msg->sg.start, false);
 213 }
 214 EXPORT_SYMBOL_GPL(sk_msg_free_nocharge);
 215
 216 int sk_msg_free(struct sock *sk, struct sk_msg *msg)
 217 {
 218         return __sk_msg_free(sk, msg, msg->sg.start, true);
 219 }
 220 EXPORT_SYMBOL_GPL(sk_msg_free);
 221
 222 static void __sk_msg_free_partial(struct sock *sk, struct sk_msg *msg,
 223                                   u32 bytes, bool charge)
 224 {
 225         struct scatterlist *sge;
 226         u32 i = msg->sg.start;
 227
 228         while (bytes) {
 229                 sge = sk_msg_elem(msg, i);
 230                 if (!sge->length)
 231                         break;
 232                 if (bytes < sge->length) {
 233                         if (charge)
 234                                 sk_mem_uncharge(sk, bytes);
 235                         sge->length -= bytes;
 236                         sge->offset += bytes;
 237                         msg->sg.size -= bytes;
 238                         break;
 239                 }
 240
 241                 msg->sg.size -= sge->length;
 242                 bytes -= sge->length;
 243                 sk_msg_free_elem(sk, msg, i, charge);
 244                 sk_msg_iter_var_next(i);
 245                 sk_msg_check_to_free(msg, i, bytes);
 246         }
 247         msg->sg.start = i;
 248 }
 249
 250 void sk_msg_free_partial(struct sock *sk, struct sk_msg *msg, u32 bytes)
 251 {
 252         __sk_msg_free_partial(sk, msg, bytes, true);
 253 }
 254 EXPORT_SYMBOL_GPL(sk_msg_free_partial);
 255
 256 void sk_msg_free_partial_nocharge(struct sock *sk, struct sk_msg *msg,
 257                                   u32 bytes)
 258 {
 259         __sk_msg_free_partial(sk, msg, bytes, false);
 260 }
 261
 262 void sk_msg_trim(struct sock *sk, struct sk_msg *msg, int len)
 263 {
 264         int trim = msg->sg.size - len;
 265         u32 i = msg->sg.end;
 266
 267         if (trim <= 0) {
 268                 WARN_ON(trim < 0);
 269                 return;
 270         }
 271
 272         sk_msg_iter_var_prev(i);
 273         msg->sg.size = len;
 274         while (msg->sg.data[i].length &&
 275                trim >= msg->sg.data[i].length) {
 276                 trim -= msg->sg.data[i].length;
 277                 sk_msg_free_elem(sk, msg, i, true);
 278                 sk_msg_iter_var_prev(i);
 279                 if (!trim)
 280                         goto out;
 281         }
 282
 283         msg->sg.data[i].length -= trim;
 284         sk_mem_uncharge(sk, trim);
 285         /* Adjust copybreak if it falls into the trimmed part of last buf */
 286         if (msg->sg.curr == i && msg->sg.copybreak > msg->sg.data[i].length)
 287                 msg->sg.copybreak = msg->sg.data[i].length;
 288 out:
 289         sk_msg_iter_var_next(i);
 290         msg->sg.end = i;
 291
 292         /* If we trim data a full sg elem before curr pointer update
 293          * copybreak and current so that any future copy operations
 294          * start at new copy location.
 295          * However trimed data that has not yet been used in a copy op
 296          * does not require an update.
 297          */
 298         if (!msg->sg.size) {
 299                 msg->sg.curr = msg->sg.start;
 300                 msg->sg.copybreak = 0;
 301         } else if (sk_msg_iter_dist(msg->sg.start, msg->sg.curr) >=
 302                    sk_msg_iter_dist(msg->sg.start, msg->sg.end)) {
 303                 sk_msg_iter_var_prev(i);
 304                 msg->sg.curr = i;
 305                 msg->sg.copybreak = msg->sg.data[i].length;
 306         }
 307 }
 308 EXPORT_SYMBOL_GPL(sk_msg_trim);
 309
 310 int sk_msg_zerocopy_from_iter(struct sock *sk, struct iov_iter *from,
 311                               struct sk_msg *msg, u32 bytes)
 312 {
 313         int i, maxpages, ret = 0, num_elems = sk_msg_elem_used(msg);
 314         const int to_max_pages = MAX_MSG_FRAGS;
 315         struct page *pages[MAX_MSG_FRAGS];
 316         ssize_t orig, copied, use, offset;
 317
 318         orig = msg->sg.size;
 319         while (bytes > 0) {
 320                 i = 0;
 321                 maxpages = to_max_pages - num_elems;
 322                 if (maxpages == 0) {
 323                         ret = -EFAULT;
 324                         goto out;
 325                 }
 326
 327                 copied = iov_iter_get_pages(from, pages, bytes, maxpages,
 328                                             &offset);
 329                 if (copied <= 0) {
 330                         ret = -EFAULT;
 331                         goto out;
 332                 }
 333
 334                 iov_iter_advance(from, copied);
 335                 bytes -= copied;
 336                 msg->sg.size += copied;
 337
 338                 while (copied) {
 339                         use = min_t(int, copied, PAGE_SIZE - offset);
 340                         sg_set_page(&msg->sg.data[msg->sg.end],
 341                                     pages[i], use, offset);
 342                         sg_unmark_end(&msg->sg.data[msg->sg.end]);
 343                         sk_mem_charge(sk, use);
 344
 345                         offset = 0;
 346                         copied -= use;
 347                         sk_msg_iter_next(msg, end);
 348                         num_elems++;
 349                         i++;
 350                 }
 351                 /* When zerocopy is mixed with sk_msg_*copy* operations we
 352                  * may have a copybreak set in this case clear and prefer
 353                  * zerocopy remainder when possible.
 354                  */
 355                 msg->sg.copybreak = 0;
 356                 msg->sg.curr = msg->sg.end;
 357         }
 358 out:
 359         /* Revert iov_iter updates, msg will need to use 'trim' later if it
 360          * also needs to be cleared.
 361          */
 362         if (ret)
 363                 iov_iter_revert(from, msg->sg.size - orig);
 364         return ret;
 365 }
 366 EXPORT_SYMBOL_GPL(sk_msg_zerocopy_from_iter);
 367
 368 int sk_msg_memcopy_from_iter(struct sock *sk, struct iov_iter *from,
 369                              struct sk_msg *msg, u32 bytes)
 370 {
 371         int ret = -ENOSPC, i = msg->sg.curr;
 372         struct scatterlist *sge;
 373         u32 copy, buf_size;
 374         void *to;
 375
 376         do {
 377                 sge = sk_msg_elem(msg, i);
 378                 /* This is possible if a trim operation shrunk the buffer */
 379                 if (msg->sg.copybreak >= sge->length) {
 380                         msg->sg.copybreak = 0;
 381                         sk_msg_iter_var_next(i);
 382                         if (i == msg->sg.end)
 383                                 break;
 384                         sge = sk_msg_elem(msg, i);
 385                 }
 386
 387                 buf_size = sge->length - msg->sg.copybreak;
 388                 copy = (buf_size > bytes) ? bytes : buf_size;
 389                 to = sg_virt(sge) + msg->sg.copybreak;
 390                 msg->sg.copybreak += copy;
 391                 if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY)
 392                         ret = copy_from_iter_nocache(to, copy, from);
 393                 else
 394                         ret = copy_from_iter(to, copy, from);
 395                 if (ret != copy) {
 396                         ret = -EFAULT;
 397                         goto out;
 398                 }
 399                 bytes -= copy;
 400                 if (!bytes)
 401                         break;
 402                 msg->sg.copybreak = 0;
 403                 sk_msg_iter_var_next(i);
 404         } while (i != msg->sg.end);
 405 out:
 406         msg->sg.curr = i;
 407         return ret;
 408 }
 409 EXPORT_SYMBOL_GPL(sk_msg_memcopy_from_iter);
 410
 411 /* Receive sk_msg from psock->ingress_msg to @msg. */
 412 int sk_msg_recvmsg(struct sock *sk, struct sk_psock *psock, struct msghdr *msg,
 413                    int len, int flags)
 414 {
 415         struct iov_iter *iter = &msg->msg_iter;
 416         int peek = flags & MSG_PEEK;
 417         struct sk_msg *msg_rx;
 418         int i, copied = 0;
 419
 420         msg_rx = sk_psock_peek_msg(psock);
 421         while (copied != len) {
 422                 struct scatterlist *sge;
 423
 424                 if (unlikely(!msg_rx))
 425                         break;
 426
 427                 i = msg_rx->sg.start;
 428                 do {
 429                         struct page *page;
 430                         int copy;
 431
 432                         sge = sk_msg_elem(msg_rx, i);
 433                         copy = sge->length;
 434                         page = sg_page(sge);
 435                         if (copied + copy > len)
 436                                 copy = len - copied;
 437                         copy = copy_page_to_iter(page, sge->offset, copy, iter);
 438                         if (!copy) {
 439                                 copied = copied ? copied : -EFAULT;
 440                                 goto out;
 441                         }
 442
 443                         copied += copy;
 444                         if (likely(!peek)) {
 445                                 sge->offset += copy;
 446                                 sge->length -= copy;
 447                                 if (!msg_rx->skb)
 448                                         sk_mem_uncharge(sk, copy);
 449                                 msg_rx->sg.size -= copy;
 450
 451                                 if (!sge->length) {
 452                                         sk_msg_iter_var_next(i);
 453                                         if (!msg_rx->skb)
 454                                                 put_page(page);
 455                                 }
 456                         } else {
 457                                 /* Lets not optimize peek case if copy_page_to_iter
 458                                  * didn't copy the entire length lets just break.
 459                                  */
 460                                 if (copy != sge->length)
 461                                         goto out;
 462                                 sk_msg_iter_var_next(i);
 463                         }
 464
 465                         if (copied == len)
 466                                 break;
 467                 } while ((i != msg_rx->sg.end) && !sg_is_last(sge));
 468
 469                 if (unlikely(peek)) {
 470                         msg_rx = sk_psock_next_msg(psock, msg_rx);
 471                         if (!msg_rx)
 472                                 break;
 473                         continue;
 474                 }
 475
 476                 msg_rx->sg.start = i;
 477                 if (!sge->length && (i == msg_rx->sg.end || sg_is_last(sge))) {
 478                         msg_rx = sk_psock_dequeue_msg(psock);
 479                         kfree_sk_msg(msg_rx);
 480                 }
 481                 msg_rx = sk_psock_peek_msg(psock);
 482         }
 483 out:
 484         if (psock->work_state.skb && copied > 0)
 485                 schedule_work(&psock->work);
 486         return copied;
 487 }
 488 EXPORT_SYMBOL_GPL(sk_msg_recvmsg);
 489
 490 bool sk_msg_is_readable(struct sock *sk)
 491 {
 492         struct sk_psock *psock;
 493         bool empty = true;
 494
 495         rcu_read_lock();
 496         psock = sk_psock(sk);
 497         if (likely(psock))
 498                 empty = list_empty(&psock->ingress_msg);
 499         rcu_read_unlock();
 500         return !empty;
 501 }
 502 EXPORT_SYMBOL_GPL(sk_msg_is_readable);
 503
 504 static struct sk_msg *sk_psock_create_ingress_msg(struct sock *sk,
 505                                                   struct sk_buff *skb)
 506 {
 507         struct sk_msg *msg;
 508
 509         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
 510                 return NULL;
 511
 512         if (!sk_rmem_schedule(sk, skb, skb->truesize))
 513                 return NULL;
 514
 515         msg = kzalloc(sizeof(*msg), __GFP_NOWARN | GFP_KERNEL);
 516         if (unlikely(!msg))
 517                 return NULL;
 518
 519         sk_msg_init(msg);
 520         return msg;
 521 }
 522
 523 static int sk_psock_skb_ingress_enqueue(struct sk_buff *skb,
 524                                         u32 off, u32 len,
 525                                         struct sk_psock *psock,
 526                                         struct sock *sk,
 527                                         struct sk_msg *msg)
 528 {
 529         int num_sge, copied;
 530
 531         /* skb linearize may fail with ENOMEM, but lets simply try again
 532          * later if this happens. Under memory pressure we don't want to
 533          * drop the skb. We need to linearize the skb so that the mapping
 534          * in skb_to_sgvec can not error.
 535          */
 536         if (skb_linearize(skb))
 537                 return -EAGAIN;
 538         num_sge = skb_to_sgvec(skb, msg->sg.data, off, len);
 539         if (unlikely(num_sge < 0))
 540                 return num_sge;
 541
 542         copied = len;
 543         msg->sg.start = 0;
 544         msg->sg.size = copied;
 545         msg->sg.end = num_sge;
 546         msg->skb = skb;
 547
 548         sk_psock_queue_msg(psock, msg);
 549         sk_psock_data_ready(sk, psock);
 550         return copied;
 551 }
 552
 553 static int sk_psock_skb_ingress_self(struct sk_psock *psock, struct sk_buff *skb,
 554                                      u32 off, u32 len);
 555
 556 static int sk_psock_skb_ingress(struct sk_psock *psock, struct sk_buff *skb,
 557                                 u32 off, u32 len)
 558 {
 559         struct sock *sk = psock->sk;
 560         struct sk_msg *msg;
 561         int err;
 562
 563         /* If we are receiving on the same sock skb->sk is already assigned,
 564          * skip memory accounting and owner transition seeing it already set
 565          * correctly.
 566          */
 567         if (unlikely(skb->sk == sk))
 568                 return sk_psock_skb_ingress_self(psock, skb, off, len);
 569         msg = sk_psock_create_ingress_msg(sk, skb);
 570         if (!msg)
 571                 return -EAGAIN;
 572
 573         /* This will transition ownership of the data from the socket where
 574          * the BPF program was run initiating the redirect to the socket
 575          * we will eventually receive this data on. The data will be released
 576          * from skb_consume found in __tcp_bpf_recvmsg() after its been copied
 577          * into user buffers.
 578          */
 579         skb_set_owner_r(skb, sk);
 580         err = sk_psock_skb_ingress_enqueue(skb, off, len, psock, sk, msg);
 581         if (err < 0)
 582                 kfree(msg);
 583         return err;
 584 }
 585
 586 /* Puts an skb on the ingress queue of the socket already assigned to the
 587  * skb. In this case we do not need to check memory limits or skb_set_owner_r
 588  * because the skb is already accounted for here.
 589  */
 590 static int sk_psock_skb_ingress_self(struct sk_psock *psock, struct sk_buff *skb,
 591                                      u32 off, u32 len)
 592 {
 593         struct sk_msg *msg = kzalloc(sizeof(*msg), __GFP_NOWARN | GFP_ATOMIC);
 594         struct sock *sk = psock->sk;
 595         int err;
 596
 597         if (unlikely(!msg))
 598                 return -EAGAIN;
 599         sk_msg_init(msg);
 600         skb_set_owner_r(skb, sk);
 601         err = sk_psock_skb_ingress_enqueue(skb, off, len, psock, sk, msg);
 602         if (err < 0)
 603                 kfree(msg);
 604         return err;
 605 }
 606
 607 static int sk_psock_handle_skb(struct sk_psock *psock, struct sk_buff *skb,
 608                                u32 off, u32 len, bool ingress)
 609 {
 610         if (!ingress) {
 611                 if (!sock_writeable(psock->sk))
 612                         return -EAGAIN;
 613                 return skb_send_sock(psock->sk, skb, off, len);
 614         }
 615         return sk_psock_skb_ingress(psock, skb, off, len);
 616 }
 617
 618 static void sk_psock_skb_state(struct sk_psock *psock,
 619                                struct sk_psock_work_state *state,
 620                                struct sk_buff *skb,
 621                                int len, int off)
 622 {
 623         spin_lock_bh(&psock->ingress_lock);
 624         if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
 625                 state->skb = skb;
 626                 state->len = len;
 627                 state->off = off;
 628         } else {
 629                 sock_drop(psock->sk, skb);
 630         }
 631         spin_unlock_bh(&psock->ingress_lock);
 632 }
 633
 634 static void sk_psock_backlog(struct work_struct *work)
 635 {
 636         struct sk_psock *psock = container_of(work, struct sk_psock, work);
 637         struct sk_psock_work_state *state = &psock->work_state;
 638         struct sk_buff *skb = NULL;
 639         bool ingress;
 640         u32 len, off;
 641         int ret;
 642
 643         mutex_lock(&psock->work_mutex);
 644         if (unlikely(state->skb)) {
 645                 spin_lock_bh(&psock->ingress_lock);
 646                 skb = state->skb;
 647                 len = state->len;
 648                 off = state->off;
 649                 state->skb = NULL;
 650                 spin_unlock_bh(&psock->ingress_lock);
 651         }
 652         if (skb)
 653                 goto start;
 654
 655         while ((skb = skb_dequeue(&psock->ingress_skb))) {
 656                 len = skb->len;
 657                 off = 0;
 658                 if (skb_bpf_strparser(skb)) {
 659                         struct strp_msg *stm = strp_msg(skb);
 660
 661                         off = stm->offset;
 662                         len = stm->full_len;
 663                 }
 664 start:
 665                 ingress = skb_bpf_ingress(skb);
 666                 skb_bpf_redirect_clear(skb);
 667                 do {
 668                         ret = -EIO;
 669                         if (!sock_flag(psock->sk, SOCK_DEAD))
 670                                 ret = sk_psock_handle_skb(psock, skb, off,
 671                                                           len, ingress);
 672                         if (ret <= 0) {
 673                                 if (ret == -EAGAIN) {
 674                                         sk_psock_skb_state(psock, state, skb,
 675                                                            len, off);
 676                                         goto end;
 677                                 }
 678                                 /* Hard errors break pipe and stop xmit. */
 679                                 sk_psock_report_error(psock, ret ? -ret : EPIPE);
 680                                 sk_psock_clear_state(psock, SK_PSOCK_TX_ENABLED);
 681                                 sock_drop(psock->sk, skb);
 682                                 goto end;
 683                         }
 684                         off += ret;
 685                         len -= ret;
 686                 } while (len);
 687
 688                 if (!ingress)
 689                         kfree_skb(skb);
 690         }
 691 end:
 692         mutex_unlock(&psock->work_mutex);
 693 }
 694
 695 struct sk_psock *sk_psock_init(struct sock *sk, int node)
 696 {
 697         struct sk_psock *psock;
 698         struct proto *prot;
 699
 700         write_lock_bh(&sk->sk_callback_lock);
 701
 702         if (sk_is_inet(sk) && inet_csk_has_ulp(sk)) {
 703                 psock = ERR_PTR(-EINVAL);
 704                 goto out;
 705         }
 706
 707         if (sk->sk_user_data) {
 708                 psock = ERR_PTR(-EBUSY);
 709                 goto out;
 710         }
 711
 712         psock = kzalloc_node(sizeof(*psock), GFP_ATOMIC | __GFP_NOWARN, node);
 713         if (!psock) {
 714                 psock = ERR_PTR(-ENOMEM);
 715                 goto out;
 716         }
 717
 718         prot = READ_ONCE(sk->sk_prot);
 719         psock->sk = sk;
 720         psock->eval = __SK_NONE;
 721         psock->sk_proto = prot;
 722         psock->saved_unhash = prot->unhash;
 723         psock->saved_destroy = prot->destroy;
 724         psock->saved_close = prot->close;
 725         psock->saved_write_space = sk->sk_write_space;
 726
 727         INIT_LIST_HEAD(&psock->link);
 728         spin_lock_init(&psock->link_lock);
 729
 730         INIT_WORK(&psock->work, sk_psock_backlog);
 731         mutex_init(&psock->work_mutex);
 732         INIT_LIST_HEAD(&psock->ingress_msg);
 733         spin_lock_init(&psock->ingress_lock);
 734         skb_queue_head_init(&psock->ingress_skb);
 735
 736         sk_psock_set_state(psock, SK_PSOCK_TX_ENABLED);
 737         refcount_set(&psock->refcnt, 1);
 738
 739         __rcu_assign_sk_user_data_with_flags(sk, psock,
 740                                              SK_USER_DATA_NOCOPY |
 741                                              SK_USER_DATA_PSOCK);
 742         sock_hold(sk);
 743
 744 out:
 745         write_unlock_bh(&sk->sk_callback_lock);
 746         return psock;
 747 }
 748 EXPORT_SYMBOL_GPL(sk_psock_init);
 749
 750 struct sk_psock_link *sk_psock_link_pop(struct sk_psock *psock)
 751 {
 752         struct sk_psock_link *link;
 753
 754         spin_lock_bh(&psock->link_lock);
 755         link = list_first_entry_or_null(&psock->link, struct sk_psock_link,
 756                                         list);
 757         if (link)
 758                 list_del(&link->list);
 759         spin_unlock_bh(&psock->link_lock);
 760         return link;
 761 }
 762
 763 static void __sk_psock_purge_ingress_msg(struct sk_psock *psock)
 764 {
 765         struct sk_msg *msg, *tmp;
 766
 767         list_for_each_entry_safe(msg, tmp, &psock->ingress_msg, list) {
 768                 list_del(&msg->list);
 769                 sk_msg_free(psock->sk, msg);
 770                 kfree(msg);
 771         }
 772 }
 773
 774 static void __sk_psock_zap_ingress(struct sk_psock *psock)
 775 {
 776         struct sk_buff *skb;
 777
 778         while ((skb = skb_dequeue(&psock->ingress_skb)) != NULL) {
 779                 skb_bpf_redirect_clear(skb);
 780                 sock_drop(psock->sk, skb);
 781         }
 782         kfree_skb(psock->work_state.skb);
 783         /* We null the skb here to ensure that calls to sk_psock_backlog
 784          * do not pick up the free'd skb.
 785          */
 786         psock->work_state.skb = NULL;
 787         __sk_psock_purge_ingress_msg(psock);
 788 }
 789
 790 static void sk_psock_link_destroy(struct sk_psock *psock)
 791 {
 792         struct sk_psock_link *link, *tmp;
 793
 794         list_for_each_entry_safe(link, tmp, &psock->link, list) {
 795                 list_del(&link->list);
 796                 sk_psock_free_link(link);
 797         }
 798 }
 799
 800 void sk_psock_stop(struct sk_psock *psock)
 801 {
 802         spin_lock_bh(&psock->ingress_lock);
 803         sk_psock_clear_state(psock, SK_PSOCK_TX_ENABLED);
 804         sk_psock_cork_free(psock);
 805         __sk_psock_zap_ingress(psock);
 806         spin_unlock_bh(&psock->ingress_lock);
 807 }
 808
 809 static void sk_psock_done_strp(struct sk_psock *psock);
 810
 811 static void sk_psock_destroy(struct work_struct *work)
 812 {
 813         struct sk_psock *psock = container_of(to_rcu_work(work),
 814                                               struct sk_psock, rwork);
 815         /* No sk_callback_lock since already detached. */
 816
 817         sk_psock_done_strp(psock);
 818
 819         cancel_work_sync(&psock->work);
 820         mutex_destroy(&psock->work_mutex);
 821
 822         psock_progs_drop(&psock->progs);
 823
 824         sk_psock_link_destroy(psock);
 825         sk_psock_cork_free(psock);
 826
 827         if (psock->sk_redir)
 828                 sock_put(psock->sk_redir);
 829         sock_put(psock->sk);
 830         kfree(psock);
 831 }
 832
 833 void sk_psock_drop(struct sock *sk, struct sk_psock *psock)
 834 {
 835         write_lock_bh(&sk->sk_callback_lock);
 836         sk_psock_restore_proto(sk, psock);
 837         rcu_assign_sk_user_data(sk, NULL);
 838         if (psock->progs.stream_parser)
 839                 sk_psock_stop_strp(sk, psock);
 840         else if (psock->progs.stream_verdict || psock->progs.skb_verdict)
 841                 sk_psock_stop_verdict(sk, psock);
 842         write_unlock_bh(&sk->sk_callback_lock);
 843
 844         sk_psock_stop(psock);
 845
 846         INIT_RCU_WORK(&psock->rwork, sk_psock_destroy);
 847         queue_rcu_work(system_wq, &psock->rwork);
 848 }
 849 EXPORT_SYMBOL_GPL(sk_psock_drop);
 850
 851 static int sk_psock_map_verd(int verdict, bool redir)
 852 {
 853         switch (verdict) {
 854         case SK_PASS:
 855                 return redir ? __SK_REDIRECT : __SK_PASS;
 856         case SK_DROP:
 857         default:
 858                 break;
 859         }
 860
 861         return __SK_DROP;
 862 }
 863
 864 int sk_psock_msg_verdict(struct sock *sk, struct sk_psock *psock,
 865                          struct sk_msg *msg)
 866 {
 867         struct bpf_prog *prog;
 868         int ret;
 869
 870         rcu_read_lock();
 871         prog = READ_ONCE(psock->progs.msg_parser);
 872         if (unlikely(!prog)) {
 873                 ret = __SK_PASS;
 874                 goto out;
 875         }
 876
 877         sk_msg_compute_data_pointers(msg);
 878         msg->sk = sk;
 879         ret = bpf_prog_run_pin_on_cpu(prog, msg);
 880         ret = sk_psock_map_verd(ret, msg->sk_redir);
 881         psock->apply_bytes = msg->apply_bytes;
 882         if (ret == __SK_REDIRECT) {
 883                 if (psock->sk_redir) {
 884                         sock_put(psock->sk_redir);
 885                         psock->sk_redir = NULL;
 886                 }
 887                 if (!msg->sk_redir) {
 888                         ret = __SK_DROP;
 889                         goto out;
 890                 }
 891                 psock->redir_ingress = sk_msg_to_ingress(msg);
 892                 psock->sk_redir = msg->sk_redir;
 893                 sock_hold(psock->sk_redir);
 894         }
 895 out:
 896         rcu_read_unlock();
 897         return ret;
 898 }
 899 EXPORT_SYMBOL_GPL(sk_psock_msg_verdict);
 900
 901 static int sk_psock_skb_redirect(struct sk_psock *from, struct sk_buff *skb)
 902 {
 903         struct sk_psock *psock_other;
 904         struct sock *sk_other;
 905
 906         sk_other = skb_bpf_redirect_fetch(skb);
 907         /* This error is a buggy BPF program, it returned a redirect
 908          * return code, but then didn't set a redirect interface.
 909          */
 910         if (unlikely(!sk_other)) {
 911                 skb_bpf_redirect_clear(skb);
 912                 sock_drop(from->sk, skb);
 913                 return -EIO;
 914         }
 915         psock_other = sk_psock(sk_other);
 916         /* This error indicates the socket is being torn down or had another
 917          * error that caused the pipe to break. We can't send a packet on
 918          * a socket that is in this state so we drop the skb.
 919          */
 920         if (!psock_other || sock_flag(sk_other, SOCK_DEAD)) {
 921                 skb_bpf_redirect_clear(skb);
 922                 sock_drop(from->sk, skb);
 923                 return -EIO;
 924         }
 925         spin_lock_bh(&psock_other->ingress_lock);
 926         if (!sk_psock_test_state(psock_other, SK_PSOCK_TX_ENABLED)) {
 927                 spin_unlock_bh(&psock_other->ingress_lock);
 928                 skb_bpf_redirect_clear(skb);
 929                 sock_drop(from->sk, skb);
 930                 return -EIO;
 931         }
 932
 933         skb_queue_tail(&psock_other->ingress_skb, skb);
 934         schedule_work(&psock_other->work);
 935         spin_unlock_bh(&psock_other->ingress_lock);
 936         return 0;
 937 }
 938
 939 static void sk_psock_tls_verdict_apply(struct sk_buff *skb,
 940                                        struct sk_psock *from, int verdict)
 941 {
 942         switch (verdict) {
 943         case __SK_REDIRECT:
 944                 sk_psock_skb_redirect(from, skb);
 945                 break;
 946         case __SK_PASS:
 947         case __SK_DROP:
 948         default:
 949                 break;
 950         }
 951 }
 952
 953 int sk_psock_tls_strp_read(struct sk_psock *psock, struct sk_buff *skb)
 954 {
 955         struct bpf_prog *prog;
 956         int ret = __SK_PASS;
 957
 958         rcu_read_lock();
 959         prog = READ_ONCE(psock->progs.stream_verdict);
 960         if (likely(prog)) {
 961                 skb->sk = psock->sk;
 962                 skb_dst_drop(skb);
 963                 skb_bpf_redirect_clear(skb);
 964                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
 965                 ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb));
 966                 skb->sk = NULL;
 967         }
 968         sk_psock_tls_verdict_apply(skb, psock, ret);
 969         rcu_read_unlock();
 970         return ret;
 971 }
 972 EXPORT_SYMBOL_GPL(sk_psock_tls_strp_read);
 973
 974 static int sk_psock_verdict_apply(struct sk_psock *psock, struct sk_buff *skb,
 975                                   int verdict)
 976 {
 977         struct sock *sk_other;
 978         int err = 0;
 979         u32 len, off;
 980
 981         switch (verdict) {
 982         case __SK_PASS:
 983                 err = -EIO;
 984                 sk_other = psock->sk;
 985                 if (sock_flag(sk_other, SOCK_DEAD) ||
 986                     !sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
 987                         skb_bpf_redirect_clear(skb);
 988                         goto out_free;
 989                 }
 990
 991                 skb_bpf_set_ingress(skb);
 992
 993                 /* If the queue is empty then we can submit directly
 994                  * into the msg queue. If its not empty we have to
 995                  * queue work otherwise we may get OOO data. Otherwise,
 996                  * if sk_psock_skb_ingress errors will be handled by
 997                  * retrying later from workqueue.
 998                  */
 999                 if (skb_queue_empty(&psock->ingress_skb)) {
1000                         len = skb->len;
1001                         off = 0;
1002                         if (skb_bpf_strparser(skb)) {
1003                                 struct strp_msg *stm = strp_msg(skb);
1004
1005                                 off = stm->offset;
1006                                 len = stm->full_len;
1007                         }
1008                         err = sk_psock_skb_ingress_self(psock, skb, off, len);
1009                 }
1010                 if (err < 0) {
1011                         spin_lock_bh(&psock->ingress_lock);
1012                         if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
1013                                 skb_queue_tail(&psock->ingress_skb, skb);
1014                                 schedule_work(&psock->work);
1015                                 err = 0;
1016                         }
1017                         spin_unlock_bh(&psock->ingress_lock);
1018                         if (err < 0) {
1019                                 skb_bpf_redirect_clear(skb);
1020                                 goto out_free;
1021                         }
1022                 }
1023                 break;
1024         case __SK_REDIRECT:
1025                 err = sk_psock_skb_redirect(psock, skb);
1026                 break;
1027         case __SK_DROP:
1028         default:
1029 out_free:
1030                 sock_drop(psock->sk, skb);
1031         }
1032
1033         return err;
1034 }
1035
1036 static void sk_psock_write_space(struct sock *sk)
1037 {
1038         struct sk_psock *psock;
1039         void (*write_space)(struct sock *sk) = NULL;
1040
1041         rcu_read_lock();
1042         psock = sk_psock(sk);
1043         if (likely(psock)) {
1044                 if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED))
1045                         schedule_work(&psock->work);
1046                 write_space = psock->saved_write_space;
1047         }
1048         rcu_read_unlock();
1049         if (write_space)
1050                 write_space(sk);
1051 }
1052
1053 #if IS_ENABLED(CONFIG_BPF_STREAM_PARSER)
1054 static void sk_psock_strp_read(struct strparser *strp, struct sk_buff *skb)
1055 {
1056         struct sk_psock *psock;
1057         struct bpf_prog *prog;
1058         int ret = __SK_DROP;
1059         struct sock *sk;
1060
1061         rcu_read_lock();
1062         sk = strp->sk;
1063         psock = sk_psock(sk);
1064         if (unlikely(!psock)) {
1065                 sock_drop(sk, skb);
1066                 goto out;
1067         }
1068         prog = READ_ONCE(psock->progs.stream_verdict);
1069         if (likely(prog)) {
1070                 skb->sk = sk;
1071                 skb_dst_drop(skb);
1072                 skb_bpf_redirect_clear(skb);
1073                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1074                 if (ret == SK_PASS)
1075                         skb_bpf_set_strparser(skb);
1076                 ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb));
1077                 skb->sk = NULL;
1078         }
1079         sk_psock_verdict_apply(psock, skb, ret);
1080 out:
1081         rcu_read_unlock();
1082 }
1083
1084 static int sk_psock_strp_read_done(struct strparser *strp, int err)
1085 {
1086         return err;
1087 }
1088
1089 static int sk_psock_strp_parse(struct strparser *strp, struct sk_buff *skb)
1090 {
1091         struct sk_psock *psock = container_of(strp, struct sk_psock, strp);
1092         struct bpf_prog *prog;
1093         int ret = skb->len;
1094
1095         rcu_read_lock();
1096         prog = READ_ONCE(psock->progs.stream_parser);
1097         if (likely(prog)) {
1098                 skb->sk = psock->sk;
1099                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1100                 skb->sk = NULL;
1101         }
1102         rcu_read_unlock();
1103         return ret;
1104 }
1105
1106 /* Called with socket lock held. */
1107 static void sk_psock_strp_data_ready(struct sock *sk)
1108 {
1109         struct sk_psock *psock;
1110
1111         rcu_read_lock();
1112         psock = sk_psock(sk);
1113         if (likely(psock)) {
1114                 if (tls_sw_has_ctx_rx(sk)) {
1115                         psock->saved_data_ready(sk);
1116                 } else {
1117                         write_lock_bh(&sk->sk_callback_lock);
1118                         strp_data_ready(&psock->strp);
1119                         write_unlock_bh(&sk->sk_callback_lock);
1120                 }
1121         }
1122         rcu_read_unlock();
1123 }
1124
1125 int sk_psock_init_strp(struct sock *sk, struct sk_psock *psock)
1126 {
1127         static const struct strp_callbacks cb = {
1128                 .rcv_msg        = sk_psock_strp_read,
1129                 .read_sock_done = sk_psock_strp_read_done,
1130                 .parse_msg      = sk_psock_strp_parse,
1131         };
1132
1133         return strp_init(&psock->strp, sk, &cb);
1134 }
1135
1136 void sk_psock_start_strp(struct sock *sk, struct sk_psock *psock)
1137 {
1138         if (psock->saved_data_ready)
1139                 return;
1140
1141         psock->saved_data_ready = sk->sk_data_ready;
1142         sk->sk_data_ready = sk_psock_strp_data_ready;
1143         sk->sk_write_space = sk_psock_write_space;
1144 }
1145
1146 void sk_psock_stop_strp(struct sock *sk, struct sk_psock *psock)
1147 {
1148         psock_set_prog(&psock->progs.stream_parser, NULL);
1149
1150         if (!psock->saved_data_ready)
1151                 return;
1152
1153         sk->sk_data_ready = psock->saved_data_ready;
1154         psock->saved_data_ready = NULL;
1155         strp_stop(&psock->strp);
1156 }
1157
1158 static void sk_psock_done_strp(struct sk_psock *psock)
1159 {
1160         /* Parser has been stopped */
1161         if (psock->progs.stream_parser)
1162                 strp_done(&psock->strp);
1163 }
1164 #else
1165 static void sk_psock_done_strp(struct sk_psock *psock)
1166 {
1167 }
1168 #endif /* CONFIG_BPF_STREAM_PARSER */
1169
1170 static int sk_psock_verdict_recv(read_descriptor_t *desc, struct sk_buff *skb,
1171                                  unsigned int offset, size_t orig_len)
1172 {
1173         struct sock *sk = (struct sock *)desc->arg.data;
1174         struct sk_psock *psock;
1175         struct bpf_prog *prog;
1176         int ret = __SK_DROP;
1177         int len = orig_len;
1178
1179         /* clone here so sk_eat_skb() in tcp_read_sock does not drop our data */
1180         skb = skb_clone(skb, GFP_ATOMIC);
1181         if (!skb) {
1182                 desc->error = -ENOMEM;
1183                 return 0;
1184         }
1185
1186         rcu_read_lock();
1187         psock = sk_psock(sk);
1188         if (unlikely(!psock)) {
1189                 len = 0;
1190                 sock_drop(sk, skb);
1191                 goto out;
1192         }
1193         prog = READ_ONCE(psock->progs.stream_verdict);
1194         if (!prog)
1195                 prog = READ_ONCE(psock->progs.skb_verdict);
1196         if (likely(prog)) {
1197                 skb->sk = sk;
1198                 skb_dst_drop(skb);
1199                 skb_bpf_redirect_clear(skb);
1200                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1201                 ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb));
1202                 skb->sk = NULL;
1203         }
1204         if (sk_psock_verdict_apply(psock, skb, ret) < 0)
1205                 len = 0;
1206 out:
1207         rcu_read_unlock();
1208         return len;
1209 }
1210
1211 static void sk_psock_verdict_data_ready(struct sock *sk)
1212 {
1213         struct socket *sock = sk->sk_socket;
1214         read_descriptor_t desc;
1215
1216         if (unlikely(!sock || !sock->ops || !sock->ops->read_sock))
1217                 return;
1218
1219         desc.arg.data = sk;
1220         desc.error = 0;
1221         desc.count = 1;
1222
1223         sock->ops->read_sock(sk, &desc, sk_psock_verdict_recv);
1224 }
1225
1226 void sk_psock_start_verdict(struct sock *sk, struct sk_psock *psock)
1227 {
1228         if (psock->saved_data_ready)
1229                 return;
1230
1231         psock->saved_data_ready = sk->sk_data_ready;
1232         sk->sk_data_ready = sk_psock_verdict_data_ready;
1233         sk->sk_write_space = sk_psock_write_space;
1234 }
1235
1236 void sk_psock_stop_verdict(struct sock *sk, struct sk_psock *psock)
1237 {
1238         psock_set_prog(&psock->progs.stream_verdict, NULL);
1239         psock_set_prog(&psock->progs.skb_verdict, NULL);
1240
1241         if (!psock->saved_data_ready)
1242                 return;
1243
1244         sk->sk_data_ready = psock->saved_data_ready;
1245         psock->saved_data_ready = NULL;
1246 }