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_pages2(from, pages, bytes, maxpages,
 328                                             &offset);
 329                 if (copied <= 0) {
 330                         ret = -EFAULT;
 331                         goto out;
 332                 }
 333
 334                 bytes -= copied;
 335                 msg->sg.size += copied;
 336
 337                 while (copied) {
 338                         use = min_t(int, copied, PAGE_SIZE - offset);
 339                         sg_set_page(&msg->sg.data[msg->sg.end],
 340                                     pages[i], use, offset);
 341                         sg_unmark_end(&msg->sg.data[msg->sg.end]);
 342                         sk_mem_charge(sk, use);
 343
 344                         offset = 0;
 345                         copied -= use;
 346                         sk_msg_iter_next(msg, end);
 347                         num_elems++;
 348                         i++;
 349                 }
 350                 /* When zerocopy is mixed with sk_msg_*copy* operations we
 351                  * may have a copybreak set in this case clear and prefer
 352                  * zerocopy remainder when possible.
 353                  */
 354                 msg->sg.copybreak = 0;
 355                 msg->sg.curr = msg->sg.end;
 356         }
 357 out:
 358         /* Revert iov_iter updates, msg will need to use 'trim' later if it
 359          * also needs to be cleared.
 360          */
 361         if (ret)
 362                 iov_iter_revert(from, msg->sg.size - orig);
 363         return ret;
 364 }
 365 EXPORT_SYMBOL_GPL(sk_msg_zerocopy_from_iter);
 366
 367 int sk_msg_memcopy_from_iter(struct sock *sk, struct iov_iter *from,
 368                              struct sk_msg *msg, u32 bytes)
 369 {
 370         int ret = -ENOSPC, i = msg->sg.curr;
 371         struct scatterlist *sge;
 372         u32 copy, buf_size;
 373         void *to;
 374
 375         do {
 376                 sge = sk_msg_elem(msg, i);
 377                 /* This is possible if a trim operation shrunk the buffer */
 378                 if (msg->sg.copybreak >= sge->length) {
 379                         msg->sg.copybreak = 0;
 380                         sk_msg_iter_var_next(i);
 381                         if (i == msg->sg.end)
 382                                 break;
 383                         sge = sk_msg_elem(msg, i);
 384                 }
 385
 386                 buf_size = sge->length - msg->sg.copybreak;
 387                 copy = (buf_size > bytes) ? bytes : buf_size;
 388                 to = sg_virt(sge) + msg->sg.copybreak;
 389                 msg->sg.copybreak += copy;
 390                 if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY)
 391                         ret = copy_from_iter_nocache(to, copy, from);
 392                 else
 393                         ret = copy_from_iter(to, copy, from);
 394                 if (ret != copy) {
 395                         ret = -EFAULT;
 396                         goto out;
 397                 }
 398                 bytes -= copy;
 399                 if (!bytes)
 400                         break;
 401                 msg->sg.copybreak = 0;
 402                 sk_msg_iter_var_next(i);
 403         } while (i != msg->sg.end);
 404 out:
 405         msg->sg.curr = i;
 406         return ret;
 407 }
 408 EXPORT_SYMBOL_GPL(sk_msg_memcopy_from_iter);
 409
 410 /* Receive sk_msg from psock->ingress_msg to @msg. */
 411 int sk_msg_recvmsg(struct sock *sk, struct sk_psock *psock, struct msghdr *msg,
 412                    int len, int flags)
 413 {
 414         struct iov_iter *iter = &msg->msg_iter;
 415         int peek = flags & MSG_PEEK;
 416         struct sk_msg *msg_rx;
 417         int i, copied = 0;
 418
 419         msg_rx = sk_psock_peek_msg(psock);
 420         while (copied != len) {
 421                 struct scatterlist *sge;
 422
 423                 if (unlikely(!msg_rx))
 424                         break;
 425
 426                 i = msg_rx->sg.start;
 427                 do {
 428                         struct page *page;
 429                         int copy;
 430
 431                         sge = sk_msg_elem(msg_rx, i);
 432                         copy = sge->length;
 433                         page = sg_page(sge);
 434                         if (copied + copy > len)
 435                                 copy = len - copied;
 436                         copy = copy_page_to_iter(page, sge->offset, copy, iter);
 437                         if (!copy) {
 438                                 copied = copied ? copied : -EFAULT;
 439                                 goto out;
 440                         }
 441
 442                         copied += copy;
 443                         if (likely(!peek)) {
 444                                 sge->offset += copy;
 445                                 sge->length -= copy;
 446                                 if (!msg_rx->skb)
 447                                         sk_mem_uncharge(sk, copy);
 448                                 msg_rx->sg.size -= copy;
 449
 450                                 if (!sge->length) {
 451                                         sk_msg_iter_var_next(i);
 452                                         if (!msg_rx->skb)
 453                                                 put_page(page);
 454                                 }
 455                         } else {
 456                                 /* Lets not optimize peek case if copy_page_to_iter
 457                                  * didn't copy the entire length lets just break.
 458                                  */
 459                                 if (copy != sge->length)
 460                                         goto out;
 461                                 sk_msg_iter_var_next(i);
 462                         }
 463
 464                         if (copied == len)
 465                                 break;
 466                 } while ((i != msg_rx->sg.end) && !sg_is_last(sge));
 467
 468                 if (unlikely(peek)) {
 469                         msg_rx = sk_psock_next_msg(psock, msg_rx);
 470                         if (!msg_rx)
 471                                 break;
 472                         continue;
 473                 }
 474
 475                 msg_rx->sg.start = i;
 476                 if (!sge->length && (i == msg_rx->sg.end || sg_is_last(sge))) {
 477                         msg_rx = sk_psock_dequeue_msg(psock);
 478                         kfree_sk_msg(msg_rx);
 479                 }
 480                 msg_rx = sk_psock_peek_msg(psock);
 481         }
 482 out:
 483         if (psock->work_state.skb && copied > 0)
 484                 schedule_work(&psock->work);
 485         return copied;
 486 }
 487 EXPORT_SYMBOL_GPL(sk_msg_recvmsg);
 488
 489 bool sk_msg_is_readable(struct sock *sk)
 490 {
 491         struct sk_psock *psock;
 492         bool empty = true;
 493
 494         rcu_read_lock();
 495         psock = sk_psock(sk);
 496         if (likely(psock))
 497                 empty = list_empty(&psock->ingress_msg);
 498         rcu_read_unlock();
 499         return !empty;
 500 }
 501 EXPORT_SYMBOL_GPL(sk_msg_is_readable);
 502
 503 static struct sk_msg *alloc_sk_msg(gfp_t gfp)
 504 {
 505         struct sk_msg *msg;
 506
 507         msg = kzalloc(sizeof(*msg), gfp | __GFP_NOWARN);
 508         if (unlikely(!msg))
 509                 return NULL;
 510         sg_init_marker(msg->sg.data, NR_MSG_FRAG_IDS);
 511         return msg;
 512 }
 513
 514 static struct sk_msg *sk_psock_create_ingress_msg(struct sock *sk,
 515                                                   struct sk_buff *skb)
 516 {
 517         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
 518                 return NULL;
 519
 520         if (!sk_rmem_schedule(sk, skb, skb->truesize))
 521                 return NULL;
 522
 523         return alloc_sk_msg(GFP_KERNEL);
 524 }
 525
 526 static int sk_psock_skb_ingress_enqueue(struct sk_buff *skb,
 527                                         u32 off, u32 len,
 528                                         struct sk_psock *psock,
 529                                         struct sock *sk,
 530                                         struct sk_msg *msg)
 531 {
 532         int num_sge, copied;
 533
 534         num_sge = skb_to_sgvec(skb, msg->sg.data, off, len);
 535         if (num_sge < 0) {
 536                 /* skb linearize may fail with ENOMEM, but lets simply try again
 537                  * later if this happens. Under memory pressure we don't want to
 538                  * drop the skb. We need to linearize the skb so that the mapping
 539                  * in skb_to_sgvec can not error.
 540                  */
 541                 if (skb_linearize(skb))
 542                         return -EAGAIN;
 543
 544                 num_sge = skb_to_sgvec(skb, msg->sg.data, off, len);
 545                 if (unlikely(num_sge < 0))
 546                         return num_sge;
 547         }
 548
 549         copied = len;
 550         msg->sg.start = 0;
 551         msg->sg.size = copied;
 552         msg->sg.end = num_sge;
 553         msg->skb = skb;
 554
 555         sk_psock_queue_msg(psock, msg);
 556         sk_psock_data_ready(sk, psock);
 557         return copied;
 558 }
 559
 560 static int sk_psock_skb_ingress_self(struct sk_psock *psock, struct sk_buff *skb,
 561                                      u32 off, u32 len);
 562
 563 static int sk_psock_skb_ingress(struct sk_psock *psock, struct sk_buff *skb,
 564                                 u32 off, u32 len)
 565 {
 566         struct sock *sk = psock->sk;
 567         struct sk_msg *msg;
 568         int err;
 569
 570         /* If we are receiving on the same sock skb->sk is already assigned,
 571          * skip memory accounting and owner transition seeing it already set
 572          * correctly.
 573          */
 574         if (unlikely(skb->sk == sk))
 575                 return sk_psock_skb_ingress_self(psock, skb, off, len);
 576         msg = sk_psock_create_ingress_msg(sk, skb);
 577         if (!msg)
 578                 return -EAGAIN;
 579
 580         /* This will transition ownership of the data from the socket where
 581          * the BPF program was run initiating the redirect to the socket
 582          * we will eventually receive this data on. The data will be released
 583          * from skb_consume found in __tcp_bpf_recvmsg() after its been copied
 584          * into user buffers.
 585          */
 586         skb_set_owner_r(skb, sk);
 587         err = sk_psock_skb_ingress_enqueue(skb, off, len, psock, sk, msg);
 588         if (err < 0)
 589                 kfree(msg);
 590         return err;
 591 }
 592
 593 /* Puts an skb on the ingress queue of the socket already assigned to the
 594  * skb. In this case we do not need to check memory limits or skb_set_owner_r
 595  * because the skb is already accounted for here.
 596  */
 597 static int sk_psock_skb_ingress_self(struct sk_psock *psock, struct sk_buff *skb,
 598                                      u32 off, u32 len)
 599 {
 600         struct sk_msg *msg = alloc_sk_msg(GFP_ATOMIC);
 601         struct sock *sk = psock->sk;
 602         int err;
 603
 604         if (unlikely(!msg))
 605                 return -EAGAIN;
 606         skb_set_owner_r(skb, sk);
 607         err = sk_psock_skb_ingress_enqueue(skb, off, len, psock, sk, msg);
 608         if (err < 0)
 609                 kfree(msg);
 610         return err;
 611 }
 612
 613 static int sk_psock_handle_skb(struct sk_psock *psock, struct sk_buff *skb,
 614                                u32 off, u32 len, bool ingress)
 615 {
 616         if (!ingress) {
 617                 if (!sock_writeable(psock->sk))
 618                         return -EAGAIN;
 619                 return skb_send_sock(psock->sk, skb, off, len);
 620         }
 621         return sk_psock_skb_ingress(psock, skb, off, len);
 622 }
 623
 624 static void sk_psock_skb_state(struct sk_psock *psock,
 625                                struct sk_psock_work_state *state,
 626                                struct sk_buff *skb,
 627                                int len, int off)
 628 {
 629         spin_lock_bh(&psock->ingress_lock);
 630         if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
 631                 state->skb = skb;
 632                 state->len = len;
 633                 state->off = off;
 634         } else {
 635                 sock_drop(psock->sk, skb);
 636         }
 637         spin_unlock_bh(&psock->ingress_lock);
 638 }
 639
 640 static void sk_psock_backlog(struct work_struct *work)
 641 {
 642         struct sk_psock *psock = container_of(work, struct sk_psock, work);
 643         struct sk_psock_work_state *state = &psock->work_state;
 644         struct sk_buff *skb = NULL;
 645         bool ingress;
 646         u32 len, off;
 647         int ret;
 648
 649         mutex_lock(&psock->work_mutex);
 650         if (unlikely(state->skb)) {
 651                 spin_lock_bh(&psock->ingress_lock);
 652                 skb = state->skb;
 653                 len = state->len;
 654                 off = state->off;
 655                 state->skb = NULL;
 656                 spin_unlock_bh(&psock->ingress_lock);
 657         }
 658         if (skb)
 659                 goto start;
 660
 661         while ((skb = skb_dequeue(&psock->ingress_skb))) {
 662                 len = skb->len;
 663                 off = 0;
 664                 if (skb_bpf_strparser(skb)) {
 665                         struct strp_msg *stm = strp_msg(skb);
 666
 667                         off = stm->offset;
 668                         len = stm->full_len;
 669                 }
 670 start:
 671                 ingress = skb_bpf_ingress(skb);
 672                 skb_bpf_redirect_clear(skb);
 673                 do {
 674                         ret = -EIO;
 675                         if (!sock_flag(psock->sk, SOCK_DEAD))
 676                                 ret = sk_psock_handle_skb(psock, skb, off,
 677                                                           len, ingress);
 678                         if (ret <= 0) {
 679                                 if (ret == -EAGAIN) {
 680                                         sk_psock_skb_state(psock, state, skb,
 681                                                            len, off);
 682                                         goto end;
 683                                 }
 684                                 /* Hard errors break pipe and stop xmit. */
 685                                 sk_psock_report_error(psock, ret ? -ret : EPIPE);
 686                                 sk_psock_clear_state(psock, SK_PSOCK_TX_ENABLED);
 687                                 sock_drop(psock->sk, skb);
 688                                 goto end;
 689                         }
 690                         off += ret;
 691                         len -= ret;
 692                 } while (len);
 693
 694                 if (!ingress)
 695                         kfree_skb(skb);
 696         }
 697 end:
 698         mutex_unlock(&psock->work_mutex);
 699 }
 700
 701 struct sk_psock *sk_psock_init(struct sock *sk, int node)
 702 {
 703         struct sk_psock *psock;
 704         struct proto *prot;
 705
 706         write_lock_bh(&sk->sk_callback_lock);
 707
 708         if (sk_is_inet(sk) && inet_csk_has_ulp(sk)) {
 709                 psock = ERR_PTR(-EINVAL);
 710                 goto out;
 711         }
 712
 713         if (sk->sk_user_data) {
 714                 psock = ERR_PTR(-EBUSY);
 715                 goto out;
 716         }
 717
 718         psock = kzalloc_node(sizeof(*psock), GFP_ATOMIC | __GFP_NOWARN, node);
 719         if (!psock) {
 720                 psock = ERR_PTR(-ENOMEM);
 721                 goto out;
 722         }
 723
 724         prot = READ_ONCE(sk->sk_prot);
 725         psock->sk = sk;
 726         psock->eval = __SK_NONE;
 727         psock->sk_proto = prot;
 728         psock->saved_unhash = prot->unhash;
 729         psock->saved_destroy = prot->destroy;
 730         psock->saved_close = prot->close;
 731         psock->saved_write_space = sk->sk_write_space;
 732
 733         INIT_LIST_HEAD(&psock->link);
 734         spin_lock_init(&psock->link_lock);
 735
 736         INIT_WORK(&psock->work, sk_psock_backlog);
 737         mutex_init(&psock->work_mutex);
 738         INIT_LIST_HEAD(&psock->ingress_msg);
 739         spin_lock_init(&psock->ingress_lock);
 740         skb_queue_head_init(&psock->ingress_skb);
 741
 742         sk_psock_set_state(psock, SK_PSOCK_TX_ENABLED);
 743         refcount_set(&psock->refcnt, 1);
 744
 745         __rcu_assign_sk_user_data_with_flags(sk, psock,
 746                                              SK_USER_DATA_NOCOPY |
 747                                              SK_USER_DATA_PSOCK);
 748         sock_hold(sk);
 749
 750 out:
 751         write_unlock_bh(&sk->sk_callback_lock);
 752         return psock;
 753 }
 754 EXPORT_SYMBOL_GPL(sk_psock_init);
 755
 756 struct sk_psock_link *sk_psock_link_pop(struct sk_psock *psock)
 757 {
 758         struct sk_psock_link *link;
 759
 760         spin_lock_bh(&psock->link_lock);
 761         link = list_first_entry_or_null(&psock->link, struct sk_psock_link,
 762                                         list);
 763         if (link)
 764                 list_del(&link->list);
 765         spin_unlock_bh(&psock->link_lock);
 766         return link;
 767 }
 768
 769 static void __sk_psock_purge_ingress_msg(struct sk_psock *psock)
 770 {
 771         struct sk_msg *msg, *tmp;
 772
 773         list_for_each_entry_safe(msg, tmp, &psock->ingress_msg, list) {
 774                 list_del(&msg->list);
 775                 sk_msg_free(psock->sk, msg);
 776                 kfree(msg);
 777         }
 778 }
 779
 780 static void __sk_psock_zap_ingress(struct sk_psock *psock)
 781 {
 782         struct sk_buff *skb;
 783
 784         while ((skb = skb_dequeue(&psock->ingress_skb)) != NULL) {
 785                 skb_bpf_redirect_clear(skb);
 786                 sock_drop(psock->sk, skb);
 787         }
 788         kfree_skb(psock->work_state.skb);
 789         /* We null the skb here to ensure that calls to sk_psock_backlog
 790          * do not pick up the free'd skb.
 791          */
 792         psock->work_state.skb = NULL;
 793         __sk_psock_purge_ingress_msg(psock);
 794 }
 795
 796 static void sk_psock_link_destroy(struct sk_psock *psock)
 797 {
 798         struct sk_psock_link *link, *tmp;
 799
 800         list_for_each_entry_safe(link, tmp, &psock->link, list) {
 801                 list_del(&link->list);
 802                 sk_psock_free_link(link);
 803         }
 804 }
 805
 806 void sk_psock_stop(struct sk_psock *psock, bool wait)
 807 {
 808         spin_lock_bh(&psock->ingress_lock);
 809         sk_psock_clear_state(psock, SK_PSOCK_TX_ENABLED);
 810         sk_psock_cork_free(psock);
 811         __sk_psock_zap_ingress(psock);
 812         spin_unlock_bh(&psock->ingress_lock);
 813
 814         if (wait)
 815                 cancel_work_sync(&psock->work);
 816 }
 817
 818 static void sk_psock_done_strp(struct sk_psock *psock);
 819
 820 static void sk_psock_destroy(struct work_struct *work)
 821 {
 822         struct sk_psock *psock = container_of(to_rcu_work(work),
 823                                               struct sk_psock, rwork);
 824         /* No sk_callback_lock since already detached. */
 825
 826         sk_psock_done_strp(psock);
 827
 828         cancel_work_sync(&psock->work);
 829         mutex_destroy(&psock->work_mutex);
 830
 831         psock_progs_drop(&psock->progs);
 832
 833         sk_psock_link_destroy(psock);
 834         sk_psock_cork_free(psock);
 835
 836         if (psock->sk_redir)
 837                 sock_put(psock->sk_redir);
 838         sock_put(psock->sk);
 839         kfree(psock);
 840 }
 841
 842 void sk_psock_drop(struct sock *sk, struct sk_psock *psock)
 843 {
 844         write_lock_bh(&sk->sk_callback_lock);
 845         sk_psock_restore_proto(sk, psock);
 846         rcu_assign_sk_user_data(sk, NULL);
 847         if (psock->progs.stream_parser)
 848                 sk_psock_stop_strp(sk, psock);
 849         else if (psock->progs.stream_verdict || psock->progs.skb_verdict)
 850                 sk_psock_stop_verdict(sk, psock);
 851         write_unlock_bh(&sk->sk_callback_lock);
 852
 853         sk_psock_stop(psock, false);
 854
 855         INIT_RCU_WORK(&psock->rwork, sk_psock_destroy);
 856         queue_rcu_work(system_wq, &psock->rwork);
 857 }
 858 EXPORT_SYMBOL_GPL(sk_psock_drop);
 859
 860 static int sk_psock_map_verd(int verdict, bool redir)
 861 {
 862         switch (verdict) {
 863         case SK_PASS:
 864                 return redir ? __SK_REDIRECT : __SK_PASS;
 865         case SK_DROP:
 866         default:
 867                 break;
 868         }
 869
 870         return __SK_DROP;
 871 }
 872
 873 int sk_psock_msg_verdict(struct sock *sk, struct sk_psock *psock,
 874                          struct sk_msg *msg)
 875 {
 876         struct bpf_prog *prog;
 877         int ret;
 878
 879         rcu_read_lock();
 880         prog = READ_ONCE(psock->progs.msg_parser);
 881         if (unlikely(!prog)) {
 882                 ret = __SK_PASS;
 883                 goto out;
 884         }
 885
 886         sk_msg_compute_data_pointers(msg);
 887         msg->sk = sk;
 888         ret = bpf_prog_run_pin_on_cpu(prog, msg);
 889         ret = sk_psock_map_verd(ret, msg->sk_redir);
 890         psock->apply_bytes = msg->apply_bytes;
 891         if (ret == __SK_REDIRECT) {
 892                 if (psock->sk_redir)
 893                         sock_put(psock->sk_redir);
 894                 psock->sk_redir = msg->sk_redir;
 895                 if (!psock->sk_redir) {
 896                         ret = __SK_DROP;
 897                         goto out;
 898                 }
 899                 sock_hold(psock->sk_redir);
 900         }
 901 out:
 902         rcu_read_unlock();
 903         return ret;
 904 }
 905 EXPORT_SYMBOL_GPL(sk_psock_msg_verdict);
 906
 907 static int sk_psock_skb_redirect(struct sk_psock *from, struct sk_buff *skb)
 908 {
 909         struct sk_psock *psock_other;
 910         struct sock *sk_other;
 911
 912         sk_other = skb_bpf_redirect_fetch(skb);
 913         /* This error is a buggy BPF program, it returned a redirect
 914          * return code, but then didn't set a redirect interface.
 915          */
 916         if (unlikely(!sk_other)) {
 917                 skb_bpf_redirect_clear(skb);
 918                 sock_drop(from->sk, skb);
 919                 return -EIO;
 920         }
 921         psock_other = sk_psock(sk_other);
 922         /* This error indicates the socket is being torn down or had another
 923          * error that caused the pipe to break. We can't send a packet on
 924          * a socket that is in this state so we drop the skb.
 925          */
 926         if (!psock_other || sock_flag(sk_other, SOCK_DEAD)) {
 927                 skb_bpf_redirect_clear(skb);
 928                 sock_drop(from->sk, skb);
 929                 return -EIO;
 930         }
 931         spin_lock_bh(&psock_other->ingress_lock);
 932         if (!sk_psock_test_state(psock_other, SK_PSOCK_TX_ENABLED)) {
 933                 spin_unlock_bh(&psock_other->ingress_lock);
 934                 skb_bpf_redirect_clear(skb);
 935                 sock_drop(from->sk, skb);
 936                 return -EIO;
 937         }
 938
 939         skb_queue_tail(&psock_other->ingress_skb, skb);
 940         schedule_work(&psock_other->work);
 941         spin_unlock_bh(&psock_other->ingress_lock);
 942         return 0;
 943 }
 944
 945 static void sk_psock_tls_verdict_apply(struct sk_buff *skb,
 946                                        struct sk_psock *from, int verdict)
 947 {
 948         switch (verdict) {
 949         case __SK_REDIRECT:
 950                 sk_psock_skb_redirect(from, skb);
 951                 break;
 952         case __SK_PASS:
 953         case __SK_DROP:
 954         default:
 955                 break;
 956         }
 957 }
 958
 959 int sk_psock_tls_strp_read(struct sk_psock *psock, struct sk_buff *skb)
 960 {
 961         struct bpf_prog *prog;
 962         int ret = __SK_PASS;
 963
 964         rcu_read_lock();
 965         prog = READ_ONCE(psock->progs.stream_verdict);
 966         if (likely(prog)) {
 967                 skb->sk = psock->sk;
 968                 skb_dst_drop(skb);
 969                 skb_bpf_redirect_clear(skb);
 970                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
 971                 ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb));
 972                 skb->sk = NULL;
 973         }
 974         sk_psock_tls_verdict_apply(skb, psock, ret);
 975         rcu_read_unlock();
 976         return ret;
 977 }
 978 EXPORT_SYMBOL_GPL(sk_psock_tls_strp_read);
 979
 980 static int sk_psock_verdict_apply(struct sk_psock *psock, struct sk_buff *skb,
 981                                   int verdict)
 982 {
 983         struct sock *sk_other;
 984         int err = 0;
 985         u32 len, off;
 986
 987         switch (verdict) {
 988         case __SK_PASS:
 989                 err = -EIO;
 990                 sk_other = psock->sk;
 991                 if (sock_flag(sk_other, SOCK_DEAD) ||
 992                     !sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
 993                         skb_bpf_redirect_clear(skb);
 994                         goto out_free;
 995                 }
 996
 997                 skb_bpf_set_ingress(skb);
 998
 999                 /* If the queue is empty then we can submit directly
1000                  * into the msg queue. If its not empty we have to
1001                  * queue work otherwise we may get OOO data. Otherwise,
1002                  * if sk_psock_skb_ingress errors will be handled by
1003                  * retrying later from workqueue.
1004                  */
1005                 if (skb_queue_empty(&psock->ingress_skb)) {
1006                         len = skb->len;
1007                         off = 0;
1008                         if (skb_bpf_strparser(skb)) {
1009                                 struct strp_msg *stm = strp_msg(skb);
1010
1011                                 off = stm->offset;
1012                                 len = stm->full_len;
1013                         }
1014                         err = sk_psock_skb_ingress_self(psock, skb, off, len);
1015                 }
1016                 if (err < 0) {
1017                         spin_lock_bh(&psock->ingress_lock);
1018                         if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
1019                                 skb_queue_tail(&psock->ingress_skb, skb);
1020                                 schedule_work(&psock->work);
1021                                 err = 0;
1022                         }
1023                         spin_unlock_bh(&psock->ingress_lock);
1024                         if (err < 0) {
1025                                 skb_bpf_redirect_clear(skb);
1026                                 goto out_free;
1027                         }
1028                 }
1029                 break;
1030         case __SK_REDIRECT:
1031                 err = sk_psock_skb_redirect(psock, skb);
1032                 break;
1033         case __SK_DROP:
1034         default:
1035 out_free:
1036                 sock_drop(psock->sk, skb);
1037         }
1038
1039         return err;
1040 }
1041
1042 static void sk_psock_write_space(struct sock *sk)
1043 {
1044         struct sk_psock *psock;
1045         void (*write_space)(struct sock *sk) = NULL;
1046
1047         rcu_read_lock();
1048         psock = sk_psock(sk);
1049         if (likely(psock)) {
1050                 if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED))
1051                         schedule_work(&psock->work);
1052                 write_space = psock->saved_write_space;
1053         }
1054         rcu_read_unlock();
1055         if (write_space)
1056                 write_space(sk);
1057 }
1058
1059 #if IS_ENABLED(CONFIG_BPF_STREAM_PARSER)
1060 static void sk_psock_strp_read(struct strparser *strp, struct sk_buff *skb)
1061 {
1062         struct sk_psock *psock;
1063         struct bpf_prog *prog;
1064         int ret = __SK_DROP;
1065         struct sock *sk;
1066
1067         rcu_read_lock();
1068         sk = strp->sk;
1069         psock = sk_psock(sk);
1070         if (unlikely(!psock)) {
1071                 sock_drop(sk, skb);
1072                 goto out;
1073         }
1074         prog = READ_ONCE(psock->progs.stream_verdict);
1075         if (likely(prog)) {
1076                 skb->sk = sk;
1077                 skb_dst_drop(skb);
1078                 skb_bpf_redirect_clear(skb);
1079                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1080                 if (ret == SK_PASS)
1081                         skb_bpf_set_strparser(skb);
1082                 ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb));
1083                 skb->sk = NULL;
1084         }
1085         sk_psock_verdict_apply(psock, skb, ret);
1086 out:
1087         rcu_read_unlock();
1088 }
1089
1090 static int sk_psock_strp_read_done(struct strparser *strp, int err)
1091 {
1092         return err;
1093 }
1094
1095 static int sk_psock_strp_parse(struct strparser *strp, struct sk_buff *skb)
1096 {
1097         struct sk_psock *psock = container_of(strp, struct sk_psock, strp);
1098         struct bpf_prog *prog;
1099         int ret = skb->len;
1100
1101         rcu_read_lock();
1102         prog = READ_ONCE(psock->progs.stream_parser);
1103         if (likely(prog)) {
1104                 skb->sk = psock->sk;
1105                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1106                 skb->sk = NULL;
1107         }
1108         rcu_read_unlock();
1109         return ret;
1110 }
1111
1112 /* Called with socket lock held. */
1113 static void sk_psock_strp_data_ready(struct sock *sk)
1114 {
1115         struct sk_psock *psock;
1116
1117         rcu_read_lock();
1118         psock = sk_psock(sk);
1119         if (likely(psock)) {
1120                 if (tls_sw_has_ctx_rx(sk)) {
1121                         psock->saved_data_ready(sk);
1122                 } else {
1123                         write_lock_bh(&sk->sk_callback_lock);
1124                         strp_data_ready(&psock->strp);
1125                         write_unlock_bh(&sk->sk_callback_lock);
1126                 }
1127         }
1128         rcu_read_unlock();
1129 }
1130
1131 int sk_psock_init_strp(struct sock *sk, struct sk_psock *psock)
1132 {
1133         static const struct strp_callbacks cb = {
1134                 .rcv_msg        = sk_psock_strp_read,
1135                 .read_sock_done = sk_psock_strp_read_done,
1136                 .parse_msg      = sk_psock_strp_parse,
1137         };
1138
1139         return strp_init(&psock->strp, sk, &cb);
1140 }
1141
1142 void sk_psock_start_strp(struct sock *sk, struct sk_psock *psock)
1143 {
1144         if (psock->saved_data_ready)
1145                 return;
1146
1147         psock->saved_data_ready = sk->sk_data_ready;
1148         sk->sk_data_ready = sk_psock_strp_data_ready;
1149         sk->sk_write_space = sk_psock_write_space;
1150 }
1151
1152 void sk_psock_stop_strp(struct sock *sk, struct sk_psock *psock)
1153 {
1154         psock_set_prog(&psock->progs.stream_parser, NULL);
1155
1156         if (!psock->saved_data_ready)
1157                 return;
1158
1159         sk->sk_data_ready = psock->saved_data_ready;
1160         psock->saved_data_ready = NULL;
1161         strp_stop(&psock->strp);
1162 }
1163
1164 static void sk_psock_done_strp(struct sk_psock *psock)
1165 {
1166         /* Parser has been stopped */
1167         if (psock->progs.stream_parser)
1168                 strp_done(&psock->strp);
1169 }
1170 #else
1171 static void sk_psock_done_strp(struct sk_psock *psock)
1172 {
1173 }
1174 #endif /* CONFIG_BPF_STREAM_PARSER */
1175
1176 static int sk_psock_verdict_recv(struct sock *sk, struct sk_buff *skb)
1177 {
1178         struct sk_psock *psock;
1179         struct bpf_prog *prog;
1180         int ret = __SK_DROP;
1181         int len = skb->len;
1182
1183         skb_get(skb);
1184
1185         rcu_read_lock();
1186         psock = sk_psock(sk);
1187         if (unlikely(!psock)) {
1188                 len = 0;
1189                 sock_drop(sk, skb);
1190                 goto out;
1191         }
1192         prog = READ_ONCE(psock->progs.stream_verdict);
1193         if (!prog)
1194                 prog = READ_ONCE(psock->progs.skb_verdict);
1195         if (likely(prog)) {
1196                 skb_dst_drop(skb);
1197                 skb_bpf_redirect_clear(skb);
1198                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1199                 ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb));
1200         }
1201         ret = sk_psock_verdict_apply(psock, skb, ret);
1202         if (ret < 0)
1203                 len = ret;
1204 out:
1205         rcu_read_unlock();
1206         return len;
1207 }
1208
1209 static void sk_psock_verdict_data_ready(struct sock *sk)
1210 {
1211         struct socket *sock = sk->sk_socket;
1212
1213         if (unlikely(!sock || !sock->ops || !sock->ops->read_skb))
1214                 return;
1215         sock->ops->read_skb(sk, sk_psock_verdict_recv);
1216 }
1217
1218 void sk_psock_start_verdict(struct sock *sk, struct sk_psock *psock)
1219 {
1220         if (psock->saved_data_ready)
1221                 return;
1222
1223         psock->saved_data_ready = sk->sk_data_ready;
1224         sk->sk_data_ready = sk_psock_verdict_data_ready;
1225         sk->sk_write_space = sk_psock_write_space;
1226 }
1227
1228 void sk_psock_stop_verdict(struct sock *sk, struct sk_psock *psock)
1229 {
1230         psock_set_prog(&psock->progs.stream_verdict, NULL);
1231         psock_set_prog(&psock->progs.skb_verdict, NULL);
1232
1233         if (!psock->saved_data_ready)
1234                 return;
1235
1236         sk->sk_data_ready = psock->saved_data_ready;
1237         psock->saved_data_ready = NULL;
1238 }