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 #include <trace/events/sock.h>
  12
  13 static bool sk_msg_try_coalesce_ok(struct sk_msg *msg, int elem_first_coalesce)
  14 {
  15         if (msg->sg.end > msg->sg.start &&
  16             elem_first_coalesce < msg->sg.end)
  17                 return true;
  18
  19         if (msg->sg.end < msg->sg.start &&
  20             (elem_first_coalesce > msg->sg.start ||
  21              elem_first_coalesce < msg->sg.end))
  22                 return true;
  23
  24         return false;
  25 }
  26
  27 int sk_msg_alloc(struct sock *sk, struct sk_msg *msg, int len,
  28                  int elem_first_coalesce)
  29 {
  30         struct page_frag *pfrag = sk_page_frag(sk);
  31         u32 osize = msg->sg.size;
  32         int ret = 0;
  33
  34         len -= msg->sg.size;
  35         while (len > 0) {
  36                 struct scatterlist *sge;
  37                 u32 orig_offset;
  38                 int use, i;
  39
  40                 if (!sk_page_frag_refill(sk, pfrag)) {
  41                         ret = -ENOMEM;
  42                         goto msg_trim;
  43                 }
  44
  45                 orig_offset = pfrag->offset;
  46                 use = min_t(int, len, pfrag->size - orig_offset);
  47                 if (!sk_wmem_schedule(sk, use)) {
  48                         ret = -ENOMEM;
  49                         goto msg_trim;
  50                 }
  51
  52                 i = msg->sg.end;
  53                 sk_msg_iter_var_prev(i);
  54                 sge = &msg->sg.data[i];
  55
  56                 if (sk_msg_try_coalesce_ok(msg, elem_first_coalesce) &&
  57                     sg_page(sge) == pfrag->page &&
  58                     sge->offset + sge->length == orig_offset) {
  59                         sge->length += use;
  60                 } else {
  61                         if (sk_msg_full(msg)) {
  62                                 ret = -ENOSPC;
  63                                 break;
  64                         }
  65
  66                         sge = &msg->sg.data[msg->sg.end];
  67                         sg_unmark_end(sge);
  68                         sg_set_page(sge, pfrag->page, use, orig_offset);
  69                         get_page(pfrag->page);
  70                         sk_msg_iter_next(msg, end);
  71                 }
  72
  73                 sk_mem_charge(sk, use);
  74                 msg->sg.size += use;
  75                 pfrag->offset += use;
  76                 len -= use;
  77         }
  78
  79         return ret;
  80
  81 msg_trim:
  82         sk_msg_trim(sk, msg, osize);
  83         return ret;
  84 }
  85 EXPORT_SYMBOL_GPL(sk_msg_alloc);
  86
  87 int sk_msg_clone(struct sock *sk, struct sk_msg *dst, struct sk_msg *src,
  88                  u32 off, u32 len)
  89 {
  90         int i = src->sg.start;
  91         struct scatterlist *sge = sk_msg_elem(src, i);
  92         struct scatterlist *sgd = NULL;
  93         u32 sge_len, sge_off;
  94
  95         while (off) {
  96                 if (sge->length > off)
  97                         break;
  98                 off -= sge->length;
  99                 sk_msg_iter_var_next(i);
 100                 if (i == src->sg.end && off)
 101                         return -ENOSPC;
 102                 sge = sk_msg_elem(src, i);
 103         }
 104
 105         while (len) {
 106                 sge_len = sge->length - off;
 107                 if (sge_len > len)
 108                         sge_len = len;
 109
 110                 if (dst->sg.end)
 111                         sgd = sk_msg_elem(dst, dst->sg.end - 1);
 112
 113                 if (sgd &&
 114                     (sg_page(sge) == sg_page(sgd)) &&
 115                     (sg_virt(sge) + off == sg_virt(sgd) + sgd->length)) {
 116                         sgd->length += sge_len;
 117                         dst->sg.size += sge_len;
 118                 } else if (!sk_msg_full(dst)) {
 119                         sge_off = sge->offset + off;
 120                         sk_msg_page_add(dst, sg_page(sge), sge_len, sge_off);
 121                 } else {
 122                         return -ENOSPC;
 123                 }
 124
 125                 off = 0;
 126                 len -= sge_len;
 127                 sk_mem_charge(sk, sge_len);
 128                 sk_msg_iter_var_next(i);
 129                 if (i == src->sg.end && len)
 130                         return -ENOSPC;
 131                 sge = sk_msg_elem(src, i);
 132         }
 133
 134         return 0;
 135 }
 136 EXPORT_SYMBOL_GPL(sk_msg_clone);
 137
 138 void sk_msg_return_zero(struct sock *sk, struct sk_msg *msg, int bytes)
 139 {
 140         int i = msg->sg.start;
 141
 142         do {
 143                 struct scatterlist *sge = sk_msg_elem(msg, i);
 144
 145                 if (bytes < sge->length) {
 146                         sge->length -= bytes;
 147                         sge->offset += bytes;
 148                         sk_mem_uncharge(sk, bytes);
 149                         break;
 150                 }
 151
 152                 sk_mem_uncharge(sk, sge->length);
 153                 bytes -= sge->length;
 154                 sge->length = 0;
 155                 sge->offset = 0;
 156                 sk_msg_iter_var_next(i);
 157         } while (bytes && i != msg->sg.end);
 158         msg->sg.start = i;
 159 }
 160 EXPORT_SYMBOL_GPL(sk_msg_return_zero);
 161
 162 void sk_msg_return(struct sock *sk, struct sk_msg *msg, int bytes)
 163 {
 164         int i = msg->sg.start;
 165
 166         do {
 167                 struct scatterlist *sge = &msg->sg.data[i];
 168                 int uncharge = (bytes < sge->length) ? bytes : sge->length;
 169
 170                 sk_mem_uncharge(sk, uncharge);
 171                 bytes -= uncharge;
 172                 sk_msg_iter_var_next(i);
 173         } while (i != msg->sg.end);
 174 }
 175 EXPORT_SYMBOL_GPL(sk_msg_return);
 176
 177 static int sk_msg_free_elem(struct sock *sk, struct sk_msg *msg, u32 i,
 178                             bool charge)
 179 {
 180         struct scatterlist *sge = sk_msg_elem(msg, i);
 181         u32 len = sge->length;
 182
 183         /* When the skb owns the memory we free it from consume_skb path. */
 184         if (!msg->skb) {
 185                 if (charge)
 186                         sk_mem_uncharge(sk, len);
 187                 put_page(sg_page(sge));
 188         }
 189         memset(sge, 0, sizeof(*sge));
 190         return len;
 191 }
 192
 193 static int __sk_msg_free(struct sock *sk, struct sk_msg *msg, u32 i,
 194                          bool charge)
 195 {
 196         struct scatterlist *sge = sk_msg_elem(msg, i);
 197         int freed = 0;
 198
 199         while (msg->sg.size) {
 200                 msg->sg.size -= sge->length;
 201                 freed += sk_msg_free_elem(sk, msg, i, charge);
 202                 sk_msg_iter_var_next(i);
 203                 sk_msg_check_to_free(msg, i, msg->sg.size);
 204                 sge = sk_msg_elem(msg, i);
 205         }
 206         consume_skb(msg->skb);
 207         sk_msg_init(msg);
 208         return freed;
 209 }
 210
 211 int sk_msg_free_nocharge(struct sock *sk, struct sk_msg *msg)
 212 {
 213         return __sk_msg_free(sk, msg, msg->sg.start, false);
 214 }
 215 EXPORT_SYMBOL_GPL(sk_msg_free_nocharge);
 216
 217 int sk_msg_free(struct sock *sk, struct sk_msg *msg)
 218 {
 219         return __sk_msg_free(sk, msg, msg->sg.start, true);
 220 }
 221 EXPORT_SYMBOL_GPL(sk_msg_free);
 222
 223 static void __sk_msg_free_partial(struct sock *sk, struct sk_msg *msg,
 224                                   u32 bytes, bool charge)
 225 {
 226         struct scatterlist *sge;
 227         u32 i = msg->sg.start;
 228
 229         while (bytes) {
 230                 sge = sk_msg_elem(msg, i);
 231                 if (!sge->length)
 232                         break;
 233                 if (bytes < sge->length) {
 234                         if (charge)
 235                                 sk_mem_uncharge(sk, bytes);
 236                         sge->length -= bytes;
 237                         sge->offset += bytes;
 238                         msg->sg.size -= bytes;
 239                         break;
 240                 }
 241
 242                 msg->sg.size -= sge->length;
 243                 bytes -= sge->length;
 244                 sk_msg_free_elem(sk, msg, i, charge);
 245                 sk_msg_iter_var_next(i);
 246                 sk_msg_check_to_free(msg, i, bytes);
 247         }
 248         msg->sg.start = i;
 249 }
 250
 251 void sk_msg_free_partial(struct sock *sk, struct sk_msg *msg, u32 bytes)
 252 {
 253         __sk_msg_free_partial(sk, msg, bytes, true);
 254 }
 255 EXPORT_SYMBOL_GPL(sk_msg_free_partial);
 256
 257 void sk_msg_free_partial_nocharge(struct sock *sk, struct sk_msg *msg,
 258                                   u32 bytes)
 259 {
 260         __sk_msg_free_partial(sk, msg, bytes, false);
 261 }
 262
 263 void sk_msg_trim(struct sock *sk, struct sk_msg *msg, int len)
 264 {
 265         int trim = msg->sg.size - len;
 266         u32 i = msg->sg.end;
 267
 268         if (trim <= 0) {
 269                 WARN_ON(trim < 0);
 270                 return;
 271         }
 272
 273         sk_msg_iter_var_prev(i);
 274         msg->sg.size = len;
 275         while (msg->sg.data[i].length &&
 276                trim >= msg->sg.data[i].length) {
 277                 trim -= msg->sg.data[i].length;
 278                 sk_msg_free_elem(sk, msg, i, true);
 279                 sk_msg_iter_var_prev(i);
 280                 if (!trim)
 281                         goto out;
 282         }
 283
 284         msg->sg.data[i].length -= trim;
 285         sk_mem_uncharge(sk, trim);
 286         /* Adjust copybreak if it falls into the trimmed part of last buf */
 287         if (msg->sg.curr == i && msg->sg.copybreak > msg->sg.data[i].length)
 288                 msg->sg.copybreak = msg->sg.data[i].length;
 289 out:
 290         sk_msg_iter_var_next(i);
 291         msg->sg.end = i;
 292
 293         /* If we trim data a full sg elem before curr pointer update
 294          * copybreak and current so that any future copy operations
 295          * start at new copy location.
 296          * However trimed data that has not yet been used in a copy op
 297          * does not require an update.
 298          */
 299         if (!msg->sg.size) {
 300                 msg->sg.curr = msg->sg.start;
 301                 msg->sg.copybreak = 0;
 302         } else if (sk_msg_iter_dist(msg->sg.start, msg->sg.curr) >=
 303                    sk_msg_iter_dist(msg->sg.start, msg->sg.end)) {
 304                 sk_msg_iter_var_prev(i);
 305                 msg->sg.curr = i;
 306                 msg->sg.copybreak = msg->sg.data[i].length;
 307         }
 308 }
 309 EXPORT_SYMBOL_GPL(sk_msg_trim);
 310
 311 int sk_msg_zerocopy_from_iter(struct sock *sk, struct iov_iter *from,
 312                               struct sk_msg *msg, u32 bytes)
 313 {
 314         int i, maxpages, ret = 0, num_elems = sk_msg_elem_used(msg);
 315         const int to_max_pages = MAX_MSG_FRAGS;
 316         struct page *pages[MAX_MSG_FRAGS];
 317         ssize_t orig, copied, use, offset;
 318
 319         orig = msg->sg.size;
 320         while (bytes > 0) {
 321                 i = 0;
 322                 maxpages = to_max_pages - num_elems;
 323                 if (maxpages == 0) {
 324                         ret = -EFAULT;
 325                         goto out;
 326                 }
 327
 328                 copied = iov_iter_get_pages2(from, pages, bytes, maxpages,
 329                                             &offset);
 330                 if (copied <= 0) {
 331                         ret = -EFAULT;
 332                         goto out;
 333                 }
 334
 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         return copied;
 485 }
 486 EXPORT_SYMBOL_GPL(sk_msg_recvmsg);
 487
 488 bool sk_msg_is_readable(struct sock *sk)
 489 {
 490         struct sk_psock *psock;
 491         bool empty = true;
 492
 493         rcu_read_lock();
 494         psock = sk_psock(sk);
 495         if (likely(psock))
 496                 empty = list_empty(&psock->ingress_msg);
 497         rcu_read_unlock();
 498         return !empty;
 499 }
 500 EXPORT_SYMBOL_GPL(sk_msg_is_readable);
 501
 502 static struct sk_msg *alloc_sk_msg(gfp_t gfp)
 503 {
 504         struct sk_msg *msg;
 505
 506         msg = kzalloc(sizeof(*msg), gfp | __GFP_NOWARN);
 507         if (unlikely(!msg))
 508                 return NULL;
 509         sg_init_marker(msg->sg.data, NR_MSG_FRAG_IDS);
 510         return msg;
 511 }
 512
 513 static struct sk_msg *sk_psock_create_ingress_msg(struct sock *sk,
 514                                                   struct sk_buff *skb)
 515 {
 516         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
 517                 return NULL;
 518
 519         if (!sk_rmem_schedule(sk, skb, skb->truesize))
 520                 return NULL;
 521
 522         return alloc_sk_msg(GFP_KERNEL);
 523 }
 524
 525 static int sk_psock_skb_ingress_enqueue(struct sk_buff *skb,
 526                                         u32 off, u32 len,
 527                                         struct sk_psock *psock,
 528                                         struct sock *sk,
 529                                         struct sk_msg *msg)
 530 {
 531         int num_sge, copied;
 532
 533         num_sge = skb_to_sgvec(skb, msg->sg.data, off, len);
 534         if (num_sge < 0) {
 535                 /* skb linearize may fail with ENOMEM, but lets simply try again
 536                  * later if this happens. Under memory pressure we don't want to
 537                  * drop the skb. We need to linearize the skb so that the mapping
 538                  * in skb_to_sgvec can not error.
 539                  */
 540                 if (skb_linearize(skb))
 541                         return -EAGAIN;
 542
 543                 num_sge = skb_to_sgvec(skb, msg->sg.data, off, len);
 544                 if (unlikely(num_sge < 0))
 545                         return num_sge;
 546         }
 547
 548         copied = len;
 549         msg->sg.start = 0;
 550         msg->sg.size = copied;
 551         msg->sg.end = num_sge;
 552         msg->skb = skb;
 553
 554         sk_psock_queue_msg(psock, msg);
 555         sk_psock_data_ready(sk, psock);
 556         return copied;
 557 }
 558
 559 static int sk_psock_skb_ingress_self(struct sk_psock *psock, struct sk_buff *skb,
 560                                      u32 off, u32 len);
 561
 562 static int sk_psock_skb_ingress(struct sk_psock *psock, struct sk_buff *skb,
 563                                 u32 off, u32 len)
 564 {
 565         struct sock *sk = psock->sk;
 566         struct sk_msg *msg;
 567         int err;
 568
 569         /* If we are receiving on the same sock skb->sk is already assigned,
 570          * skip memory accounting and owner transition seeing it already set
 571          * correctly.
 572          */
 573         if (unlikely(skb->sk == sk))
 574                 return sk_psock_skb_ingress_self(psock, skb, off, len);
 575         msg = sk_psock_create_ingress_msg(sk, skb);
 576         if (!msg)
 577                 return -EAGAIN;
 578
 579         /* This will transition ownership of the data from the socket where
 580          * the BPF program was run initiating the redirect to the socket
 581          * we will eventually receive this data on. The data will be released
 582          * from skb_consume found in __tcp_bpf_recvmsg() after its been copied
 583          * into user buffers.
 584          */
 585         skb_set_owner_r(skb, sk);
 586         err = sk_psock_skb_ingress_enqueue(skb, off, len, psock, sk, msg);
 587         if (err < 0)
 588                 kfree(msg);
 589         return err;
 590 }
 591
 592 /* Puts an skb on the ingress queue of the socket already assigned to the
 593  * skb. In this case we do not need to check memory limits or skb_set_owner_r
 594  * because the skb is already accounted for here.
 595  */
 596 static int sk_psock_skb_ingress_self(struct sk_psock *psock, struct sk_buff *skb,
 597                                      u32 off, u32 len)
 598 {
 599         struct sk_msg *msg = alloc_sk_msg(GFP_ATOMIC);
 600         struct sock *sk = psock->sk;
 601         int err;
 602
 603         if (unlikely(!msg))
 604                 return -EAGAIN;
 605         skb_set_owner_r(skb, sk);
 606         err = sk_psock_skb_ingress_enqueue(skb, off, len, psock, sk, msg);
 607         if (err < 0)
 608                 kfree(msg);
 609         return err;
 610 }
 611
 612 static int sk_psock_handle_skb(struct sk_psock *psock, struct sk_buff *skb,
 613                                u32 off, u32 len, bool ingress)
 614 {
 615         int err = 0;
 616
 617         if (!ingress) {
 618                 if (!sock_writeable(psock->sk))
 619                         return -EAGAIN;
 620                 return skb_send_sock(psock->sk, skb, off, len);
 621         }
 622         skb_get(skb);
 623         err = sk_psock_skb_ingress(psock, skb, off, len);
 624         if (err < 0)
 625                 kfree_skb(skb);
 626         return err;
 627 }
 628
 629 static void sk_psock_skb_state(struct sk_psock *psock,
 630                                struct sk_psock_work_state *state,
 631                                int len, int off)
 632 {
 633         spin_lock_bh(&psock->ingress_lock);
 634         if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
 635                 state->len = len;
 636                 state->off = off;
 637         }
 638         spin_unlock_bh(&psock->ingress_lock);
 639 }
 640
 641 static void sk_psock_backlog(struct work_struct *work)
 642 {
 643         struct delayed_work *dwork = to_delayed_work(work);
 644         struct sk_psock *psock = container_of(dwork, struct sk_psock, work);
 645         struct sk_psock_work_state *state = &psock->work_state;
 646         struct sk_buff *skb = NULL;
 647         u32 len = 0, off = 0;
 648         bool ingress;
 649         int ret;
 650
 651         mutex_lock(&psock->work_mutex);
 652         if (unlikely(state->len)) {
 653                 len = state->len;
 654                 off = state->off;
 655         }
 656
 657         while ((skb = skb_peek(&psock->ingress_skb))) {
 658                 len = skb->len;
 659                 off = 0;
 660                 if (skb_bpf_strparser(skb)) {
 661                         struct strp_msg *stm = strp_msg(skb);
 662
 663                         off = stm->offset;
 664                         len = stm->full_len;
 665                 }
 666                 ingress = skb_bpf_ingress(skb);
 667                 skb_bpf_redirect_clear(skb);
 668                 do {
 669                         ret = -EIO;
 670                         if (!sock_flag(psock->sk, SOCK_DEAD))
 671                                 ret = sk_psock_handle_skb(psock, skb, off,
 672                                                           len, ingress);
 673                         if (ret <= 0) {
 674                                 if (ret == -EAGAIN) {
 675                                         sk_psock_skb_state(psock, state, len, off);
 676
 677                                         /* Delay slightly to prioritize any
 678                                          * other work that might be here.
 679                                          */
 680                                         if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED))
 681                                                 schedule_delayed_work(&psock->work, 1);
 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                                 goto end;
 688                         }
 689                         off += ret;
 690                         len -= ret;
 691                 } while (len);
 692
 693                 skb = skb_dequeue(&psock->ingress_skb);
 694                 kfree_skb(skb);
 695         }
 696 end:
 697         mutex_unlock(&psock->work_mutex);
 698 }
 699
 700 struct sk_psock *sk_psock_init(struct sock *sk, int node)
 701 {
 702         struct sk_psock *psock;
 703         struct proto *prot;
 704
 705         write_lock_bh(&sk->sk_callback_lock);
 706
 707         if (sk_is_inet(sk) && inet_csk_has_ulp(sk)) {
 708                 psock = ERR_PTR(-EINVAL);
 709                 goto out;
 710         }
 711
 712         if (sk->sk_user_data) {
 713                 psock = ERR_PTR(-EBUSY);
 714                 goto out;
 715         }
 716
 717         psock = kzalloc_node(sizeof(*psock), GFP_ATOMIC | __GFP_NOWARN, node);
 718         if (!psock) {
 719                 psock = ERR_PTR(-ENOMEM);
 720                 goto out;
 721         }
 722
 723         prot = READ_ONCE(sk->sk_prot);
 724         psock->sk = sk;
 725         psock->eval = __SK_NONE;
 726         psock->sk_proto = prot;
 727         psock->saved_unhash = prot->unhash;
 728         psock->saved_destroy = prot->destroy;
 729         psock->saved_close = prot->close;
 730         psock->saved_write_space = sk->sk_write_space;
 731
 732         INIT_LIST_HEAD(&psock->link);
 733         spin_lock_init(&psock->link_lock);
 734
 735         INIT_DELAYED_WORK(&psock->work, sk_psock_backlog);
 736         mutex_init(&psock->work_mutex);
 737         INIT_LIST_HEAD(&psock->ingress_msg);
 738         spin_lock_init(&psock->ingress_lock);
 739         skb_queue_head_init(&psock->ingress_skb);
 740
 741         sk_psock_set_state(psock, SK_PSOCK_TX_ENABLED);
 742         refcount_set(&psock->refcnt, 1);
 743
 744         __rcu_assign_sk_user_data_with_flags(sk, psock,
 745                                              SK_USER_DATA_NOCOPY |
 746                                              SK_USER_DATA_PSOCK);
 747         sock_hold(sk);
 748
 749 out:
 750         write_unlock_bh(&sk->sk_callback_lock);
 751         return psock;
 752 }
 753 EXPORT_SYMBOL_GPL(sk_psock_init);
 754
 755 struct sk_psock_link *sk_psock_link_pop(struct sk_psock *psock)
 756 {
 757         struct sk_psock_link *link;
 758
 759         spin_lock_bh(&psock->link_lock);
 760         link = list_first_entry_or_null(&psock->link, struct sk_psock_link,
 761                                         list);
 762         if (link)
 763                 list_del(&link->list);
 764         spin_unlock_bh(&psock->link_lock);
 765         return link;
 766 }
 767
 768 static void __sk_psock_purge_ingress_msg(struct sk_psock *psock)
 769 {
 770         struct sk_msg *msg, *tmp;
 771
 772         list_for_each_entry_safe(msg, tmp, &psock->ingress_msg, list) {
 773                 list_del(&msg->list);
 774                 sk_msg_free(psock->sk, msg);
 775                 kfree(msg);
 776         }
 777 }
 778
 779 static void __sk_psock_zap_ingress(struct sk_psock *psock)
 780 {
 781         struct sk_buff *skb;
 782
 783         while ((skb = skb_dequeue(&psock->ingress_skb)) != NULL) {
 784                 skb_bpf_redirect_clear(skb);
 785                 sock_drop(psock->sk, skb);
 786         }
 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         spin_unlock_bh(&psock->ingress_lock);
 806 }
 807
 808 static void sk_psock_done_strp(struct sk_psock *psock);
 809
 810 static void sk_psock_destroy(struct work_struct *work)
 811 {
 812         struct sk_psock *psock = container_of(to_rcu_work(work),
 813                                               struct sk_psock, rwork);
 814         /* No sk_callback_lock since already detached. */
 815
 816         sk_psock_done_strp(psock);
 817
 818         cancel_delayed_work_sync(&psock->work);
 819         __sk_psock_zap_ingress(psock);
 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_delayed_work(&psock_other->work, 0);
 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                         goto out_free;
 988
 989                 skb_bpf_set_ingress(skb);
 990
 991                 /* If the queue is empty then we can submit directly
 992                  * into the msg queue. If its not empty we have to
 993                  * queue work otherwise we may get OOO data. Otherwise,
 994                  * if sk_psock_skb_ingress errors will be handled by
 995                  * retrying later from workqueue.
 996                  */
 997                 if (skb_queue_empty(&psock->ingress_skb)) {
 998                         len = skb->len;
 999                         off = 0;
1000                         if (skb_bpf_strparser(skb)) {
1001                                 struct strp_msg *stm = strp_msg(skb);
1002
1003                                 off = stm->offset;
1004                                 len = stm->full_len;
1005                         }
1006                         err = sk_psock_skb_ingress_self(psock, skb, off, len);
1007                 }
1008                 if (err < 0) {
1009                         spin_lock_bh(&psock->ingress_lock);
1010                         if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
1011                                 skb_queue_tail(&psock->ingress_skb, skb);
1012                                 schedule_delayed_work(&psock->work, 0);
1013                                 err = 0;
1014                         }
1015                         spin_unlock_bh(&psock->ingress_lock);
1016                         if (err < 0)
1017                                 goto out_free;
1018                 }
1019                 break;
1020         case __SK_REDIRECT:
1021                 tcp_eat_skb(psock->sk, skb);
1022                 err = sk_psock_skb_redirect(psock, skb);
1023                 break;
1024         case __SK_DROP:
1025         default:
1026 out_free:
1027                 skb_bpf_redirect_clear(skb);
1028                 tcp_eat_skb(psock->sk, skb);
1029                 sock_drop(psock->sk, skb);
1030         }
1031
1032         return err;
1033 }
1034
1035 static void sk_psock_write_space(struct sock *sk)
1036 {
1037         struct sk_psock *psock;
1038         void (*write_space)(struct sock *sk) = NULL;
1039
1040         rcu_read_lock();
1041         psock = sk_psock(sk);
1042         if (likely(psock)) {
1043                 if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED))
1044                         schedule_delayed_work(&psock->work, 0);
1045                 write_space = psock->saved_write_space;
1046         }
1047         rcu_read_unlock();
1048         if (write_space)
1049                 write_space(sk);
1050 }
1051
1052 #if IS_ENABLED(CONFIG_BPF_STREAM_PARSER)
1053 static void sk_psock_strp_read(struct strparser *strp, struct sk_buff *skb)
1054 {
1055         struct sk_psock *psock;
1056         struct bpf_prog *prog;
1057         int ret = __SK_DROP;
1058         struct sock *sk;
1059
1060         rcu_read_lock();
1061         sk = strp->sk;
1062         psock = sk_psock(sk);
1063         if (unlikely(!psock)) {
1064                 sock_drop(sk, skb);
1065                 goto out;
1066         }
1067         prog = READ_ONCE(psock->progs.stream_verdict);
1068         if (likely(prog)) {
1069                 skb->sk = sk;
1070                 skb_dst_drop(skb);
1071                 skb_bpf_redirect_clear(skb);
1072                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1073                 skb_bpf_set_strparser(skb);
1074                 ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb));
1075                 skb->sk = NULL;
1076         }
1077         sk_psock_verdict_apply(psock, skb, ret);
1078 out:
1079         rcu_read_unlock();
1080 }
1081
1082 static int sk_psock_strp_read_done(struct strparser *strp, int err)
1083 {
1084         return err;
1085 }
1086
1087 static int sk_psock_strp_parse(struct strparser *strp, struct sk_buff *skb)
1088 {
1089         struct sk_psock *psock = container_of(strp, struct sk_psock, strp);
1090         struct bpf_prog *prog;
1091         int ret = skb->len;
1092
1093         rcu_read_lock();
1094         prog = READ_ONCE(psock->progs.stream_parser);
1095         if (likely(prog)) {
1096                 skb->sk = psock->sk;
1097                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1098                 skb->sk = NULL;
1099         }
1100         rcu_read_unlock();
1101         return ret;
1102 }
1103
1104 /* Called with socket lock held. */
1105 static void sk_psock_strp_data_ready(struct sock *sk)
1106 {
1107         struct sk_psock *psock;
1108
1109         trace_sk_data_ready(sk);
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         int ret;
1128
1129         static const struct strp_callbacks cb = {
1130                 .rcv_msg        = sk_psock_strp_read,
1131                 .read_sock_done = sk_psock_strp_read_done,
1132                 .parse_msg      = sk_psock_strp_parse,
1133         };
1134
1135         ret = strp_init(&psock->strp, sk, &cb);
1136         if (!ret)
1137                 sk_psock_set_state(psock, SK_PSOCK_RX_STRP_ENABLED);
1138
1139         return ret;
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 (sk_psock_test_state(psock, SK_PSOCK_RX_STRP_ENABLED))
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         rcu_read_lock();
1184         psock = sk_psock(sk);
1185         if (unlikely(!psock)) {
1186                 len = 0;
1187                 tcp_eat_skb(sk, skb);
1188                 sock_drop(sk, skb);
1189                 goto out;
1190         }
1191         prog = READ_ONCE(psock->progs.stream_verdict);
1192         if (!prog)
1193                 prog = READ_ONCE(psock->progs.skb_verdict);
1194         if (likely(prog)) {
1195                 skb_dst_drop(skb);
1196                 skb_bpf_redirect_clear(skb);
1197                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1198                 ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb));
1199         }
1200         ret = sk_psock_verdict_apply(psock, skb, ret);
1201         if (ret < 0)
1202                 len = ret;
1203 out:
1204         rcu_read_unlock();
1205         return len;
1206 }
1207
1208 static void sk_psock_verdict_data_ready(struct sock *sk)
1209 {
1210         struct socket *sock = sk->sk_socket;
1211         const struct proto_ops *ops;
1212         int copied;
1213
1214         trace_sk_data_ready(sk);
1215
1216         if (unlikely(!sock))
1217                 return;
1218         ops = READ_ONCE(sock->ops);
1219         if (!ops || !ops->read_skb)
1220                 return;
1221         copied = ops->read_skb(sk, sk_psock_verdict_recv);
1222         if (copied >= 0) {
1223                 struct sk_psock *psock;
1224
1225                 rcu_read_lock();
1226                 psock = sk_psock(sk);
1227                 if (psock)
1228                         psock->saved_data_ready(sk);
1229                 rcu_read_unlock();
1230         }
1231 }
1232
1233 void sk_psock_start_verdict(struct sock *sk, struct sk_psock *psock)
1234 {
1235         if (psock->saved_data_ready)
1236                 return;
1237
1238         psock->saved_data_ready = sk->sk_data_ready;
1239         sk->sk_data_ready = sk_psock_verdict_data_ready;
1240         sk->sk_write_space = sk_psock_write_space;
1241 }
1242
1243 void sk_psock_stop_verdict(struct sock *sk, struct sk_psock *psock)
1244 {
1245         psock_set_prog(&psock->progs.stream_verdict, NULL);
1246         psock_set_prog(&psock->progs.skb_verdict, NULL);
1247
1248         if (!psock->saved_data_ready)
1249                 return;
1250
1251         sk->sk_data_ready = psock->saved_data_ready;
1252         psock->saved_data_ready = NULL;
1253 }