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         if (psock->work_state.skb && copied > 0)
 485                 schedule_work(&psock->work);
 486         return copied;
 487 }
 488 EXPORT_SYMBOL_GPL(sk_msg_recvmsg);
 489
 490 bool sk_msg_is_readable(struct sock *sk)
 491 {
 492         struct sk_psock *psock;
 493         bool empty = true;
 494
 495         rcu_read_lock();
 496         psock = sk_psock(sk);
 497         if (likely(psock))
 498                 empty = list_empty(&psock->ingress_msg);
 499         rcu_read_unlock();
 500         return !empty;
 501 }
 502 EXPORT_SYMBOL_GPL(sk_msg_is_readable);
 503
 504 static struct sk_msg *alloc_sk_msg(gfp_t gfp)
 505 {
 506         struct sk_msg *msg;
 507
 508         msg = kzalloc(sizeof(*msg), gfp | __GFP_NOWARN);
 509         if (unlikely(!msg))
 510                 return NULL;
 511         sg_init_marker(msg->sg.data, NR_MSG_FRAG_IDS);
 512         return msg;
 513 }
 514
 515 static struct sk_msg *sk_psock_create_ingress_msg(struct sock *sk,
 516                                                   struct sk_buff *skb)
 517 {
 518         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
 519                 return NULL;
 520
 521         if (!sk_rmem_schedule(sk, skb, skb->truesize))
 522                 return NULL;
 523
 524         return alloc_sk_msg(GFP_KERNEL);
 525 }
 526
 527 static int sk_psock_skb_ingress_enqueue(struct sk_buff *skb,
 528                                         u32 off, u32 len,
 529                                         struct sk_psock *psock,
 530                                         struct sock *sk,
 531                                         struct sk_msg *msg)
 532 {
 533         int num_sge, copied;
 534
 535         num_sge = skb_to_sgvec(skb, msg->sg.data, off, len);
 536         if (num_sge < 0) {
 537                 /* skb linearize may fail with ENOMEM, but lets simply try again
 538                  * later if this happens. Under memory pressure we don't want to
 539                  * drop the skb. We need to linearize the skb so that the mapping
 540                  * in skb_to_sgvec can not error.
 541                  */
 542                 if (skb_linearize(skb))
 543                         return -EAGAIN;
 544
 545                 num_sge = skb_to_sgvec(skb, msg->sg.data, off, len);
 546                 if (unlikely(num_sge < 0))
 547                         return num_sge;
 548         }
 549
 550         copied = len;
 551         msg->sg.start = 0;
 552         msg->sg.size = copied;
 553         msg->sg.end = num_sge;
 554         msg->skb = skb;
 555
 556         sk_psock_queue_msg(psock, msg);
 557         sk_psock_data_ready(sk, psock);
 558         return copied;
 559 }
 560
 561 static int sk_psock_skb_ingress_self(struct sk_psock *psock, struct sk_buff *skb,
 562                                      u32 off, u32 len);
 563
 564 static int sk_psock_skb_ingress(struct sk_psock *psock, struct sk_buff *skb,
 565                                 u32 off, u32 len)
 566 {
 567         struct sock *sk = psock->sk;
 568         struct sk_msg *msg;
 569         int err;
 570
 571         /* If we are receiving on the same sock skb->sk is already assigned,
 572          * skip memory accounting and owner transition seeing it already set
 573          * correctly.
 574          */
 575         if (unlikely(skb->sk == sk))
 576                 return sk_psock_skb_ingress_self(psock, skb, off, len);
 577         msg = sk_psock_create_ingress_msg(sk, skb);
 578         if (!msg)
 579                 return -EAGAIN;
 580
 581         /* This will transition ownership of the data from the socket where
 582          * the BPF program was run initiating the redirect to the socket
 583          * we will eventually receive this data on. The data will be released
 584          * from skb_consume found in __tcp_bpf_recvmsg() after its been copied
 585          * into user buffers.
 586          */
 587         skb_set_owner_r(skb, sk);
 588         err = sk_psock_skb_ingress_enqueue(skb, off, len, psock, sk, msg);
 589         if (err < 0)
 590                 kfree(msg);
 591         return err;
 592 }
 593
 594 /* Puts an skb on the ingress queue of the socket already assigned to the
 595  * skb. In this case we do not need to check memory limits or skb_set_owner_r
 596  * because the skb is already accounted for here.
 597  */
 598 static int sk_psock_skb_ingress_self(struct sk_psock *psock, struct sk_buff *skb,
 599                                      u32 off, u32 len)
 600 {
 601         struct sk_msg *msg = alloc_sk_msg(GFP_ATOMIC);
 602         struct sock *sk = psock->sk;
 603         int err;
 604
 605         if (unlikely(!msg))
 606                 return -EAGAIN;
 607         skb_set_owner_r(skb, sk);
 608         err = sk_psock_skb_ingress_enqueue(skb, off, len, psock, sk, msg);
 609         if (err < 0)
 610                 kfree(msg);
 611         return err;
 612 }
 613
 614 static int sk_psock_handle_skb(struct sk_psock *psock, struct sk_buff *skb,
 615                                u32 off, u32 len, bool ingress)
 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         return sk_psock_skb_ingress(psock, skb, off, len);
 623 }
 624
 625 static void sk_psock_skb_state(struct sk_psock *psock,
 626                                struct sk_psock_work_state *state,
 627                                struct sk_buff *skb,
 628                                int len, int off)
 629 {
 630         spin_lock_bh(&psock->ingress_lock);
 631         if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
 632                 state->skb = skb;
 633                 state->len = len;
 634                 state->off = off;
 635         } else {
 636                 sock_drop(psock->sk, skb);
 637         }
 638         spin_unlock_bh(&psock->ingress_lock);
 639 }
 640
 641 static void sk_psock_backlog(struct work_struct *work)
 642 {
 643         struct sk_psock *psock = container_of(work, struct sk_psock, work);
 644         struct sk_psock_work_state *state = &psock->work_state;
 645         struct sk_buff *skb = NULL;
 646         bool ingress;
 647         u32 len, off;
 648         int ret;
 649
 650         mutex_lock(&psock->work_mutex);
 651         if (unlikely(state->skb)) {
 652                 spin_lock_bh(&psock->ingress_lock);
 653                 skb = state->skb;
 654                 len = state->len;
 655                 off = state->off;
 656                 state->skb = NULL;
 657                 spin_unlock_bh(&psock->ingress_lock);
 658         }
 659         if (skb)
 660                 goto start;
 661
 662         while ((skb = skb_dequeue(&psock->ingress_skb))) {
 663                 len = skb->len;
 664                 off = 0;
 665                 if (skb_bpf_strparser(skb)) {
 666                         struct strp_msg *stm = strp_msg(skb);
 667
 668                         off = stm->offset;
 669                         len = stm->full_len;
 670                 }
 671 start:
 672                 ingress = skb_bpf_ingress(skb);
 673                 skb_bpf_redirect_clear(skb);
 674                 do {
 675                         ret = -EIO;
 676                         if (!sock_flag(psock->sk, SOCK_DEAD))
 677                                 ret = sk_psock_handle_skb(psock, skb, off,
 678                                                           len, ingress);
 679                         if (ret <= 0) {
 680                                 if (ret == -EAGAIN) {
 681                                         sk_psock_skb_state(psock, state, skb,
 682                                                            len, off);
 683                                         goto end;
 684                                 }
 685                                 /* Hard errors break pipe and stop xmit. */
 686                                 sk_psock_report_error(psock, ret ? -ret : EPIPE);
 687                                 sk_psock_clear_state(psock, SK_PSOCK_TX_ENABLED);
 688                                 sock_drop(psock->sk, skb);
 689                                 goto end;
 690                         }
 691                         off += ret;
 692                         len -= ret;
 693                 } while (len);
 694
 695                 if (!ingress)
 696                         kfree_skb(skb);
 697         }
 698 end:
 699         mutex_unlock(&psock->work_mutex);
 700 }
 701
 702 struct sk_psock *sk_psock_init(struct sock *sk, int node)
 703 {
 704         struct sk_psock *psock;
 705         struct proto *prot;
 706
 707         write_lock_bh(&sk->sk_callback_lock);
 708
 709         if (sk_is_inet(sk) && inet_csk_has_ulp(sk)) {
 710                 psock = ERR_PTR(-EINVAL);
 711                 goto out;
 712         }
 713
 714         if (sk->sk_user_data) {
 715                 psock = ERR_PTR(-EBUSY);
 716                 goto out;
 717         }
 718
 719         psock = kzalloc_node(sizeof(*psock), GFP_ATOMIC | __GFP_NOWARN, node);
 720         if (!psock) {
 721                 psock = ERR_PTR(-ENOMEM);
 722                 goto out;
 723         }
 724
 725         prot = READ_ONCE(sk->sk_prot);
 726         psock->sk = sk;
 727         psock->eval = __SK_NONE;
 728         psock->sk_proto = prot;
 729         psock->saved_unhash = prot->unhash;
 730         psock->saved_destroy = prot->destroy;
 731         psock->saved_close = prot->close;
 732         psock->saved_write_space = sk->sk_write_space;
 733
 734         INIT_LIST_HEAD(&psock->link);
 735         spin_lock_init(&psock->link_lock);
 736
 737         INIT_WORK(&psock->work, sk_psock_backlog);
 738         mutex_init(&psock->work_mutex);
 739         INIT_LIST_HEAD(&psock->ingress_msg);
 740         spin_lock_init(&psock->ingress_lock);
 741         skb_queue_head_init(&psock->ingress_skb);
 742
 743         sk_psock_set_state(psock, SK_PSOCK_TX_ENABLED);
 744         refcount_set(&psock->refcnt, 1);
 745
 746         __rcu_assign_sk_user_data_with_flags(sk, psock,
 747                                              SK_USER_DATA_NOCOPY |
 748                                              SK_USER_DATA_PSOCK);
 749         sock_hold(sk);
 750
 751 out:
 752         write_unlock_bh(&sk->sk_callback_lock);
 753         return psock;
 754 }
 755 EXPORT_SYMBOL_GPL(sk_psock_init);
 756
 757 struct sk_psock_link *sk_psock_link_pop(struct sk_psock *psock)
 758 {
 759         struct sk_psock_link *link;
 760
 761         spin_lock_bh(&psock->link_lock);
 762         link = list_first_entry_or_null(&psock->link, struct sk_psock_link,
 763                                         list);
 764         if (link)
 765                 list_del(&link->list);
 766         spin_unlock_bh(&psock->link_lock);
 767         return link;
 768 }
 769
 770 static void __sk_psock_purge_ingress_msg(struct sk_psock *psock)
 771 {
 772         struct sk_msg *msg, *tmp;
 773
 774         list_for_each_entry_safe(msg, tmp, &psock->ingress_msg, list) {
 775                 list_del(&msg->list);
 776                 sk_msg_free(psock->sk, msg);
 777                 kfree(msg);
 778         }
 779 }
 780
 781 static void __sk_psock_zap_ingress(struct sk_psock *psock)
 782 {
 783         struct sk_buff *skb;
 784
 785         while ((skb = skb_dequeue(&psock->ingress_skb)) != NULL) {
 786                 skb_bpf_redirect_clear(skb);
 787                 sock_drop(psock->sk, skb);
 788         }
 789         kfree_skb(psock->work_state.skb);
 790         /* We null the skb here to ensure that calls to sk_psock_backlog
 791          * do not pick up the free'd skb.
 792          */
 793         psock->work_state.skb = NULL;
 794         __sk_psock_purge_ingress_msg(psock);
 795 }
 796
 797 static void sk_psock_link_destroy(struct sk_psock *psock)
 798 {
 799         struct sk_psock_link *link, *tmp;
 800
 801         list_for_each_entry_safe(link, tmp, &psock->link, list) {
 802                 list_del(&link->list);
 803                 sk_psock_free_link(link);
 804         }
 805 }
 806
 807 void sk_psock_stop(struct sk_psock *psock)
 808 {
 809         spin_lock_bh(&psock->ingress_lock);
 810         sk_psock_clear_state(psock, SK_PSOCK_TX_ENABLED);
 811         sk_psock_cork_free(psock);
 812         __sk_psock_zap_ingress(psock);
 813         spin_unlock_bh(&psock->ingress_lock);
 814 }
 815
 816 static void sk_psock_done_strp(struct sk_psock *psock);
 817
 818 static void sk_psock_destroy(struct work_struct *work)
 819 {
 820         struct sk_psock *psock = container_of(to_rcu_work(work),
 821                                               struct sk_psock, rwork);
 822         /* No sk_callback_lock since already detached. */
 823
 824         sk_psock_done_strp(psock);
 825
 826         cancel_work_sync(&psock->work);
 827         mutex_destroy(&psock->work_mutex);
 828
 829         psock_progs_drop(&psock->progs);
 830
 831         sk_psock_link_destroy(psock);
 832         sk_psock_cork_free(psock);
 833
 834         if (psock->sk_redir)
 835                 sock_put(psock->sk_redir);
 836         sock_put(psock->sk);
 837         kfree(psock);
 838 }
 839
 840 void sk_psock_drop(struct sock *sk, struct sk_psock *psock)
 841 {
 842         write_lock_bh(&sk->sk_callback_lock);
 843         sk_psock_restore_proto(sk, psock);
 844         rcu_assign_sk_user_data(sk, NULL);
 845         if (psock->progs.stream_parser)
 846                 sk_psock_stop_strp(sk, psock);
 847         else if (psock->progs.stream_verdict || psock->progs.skb_verdict)
 848                 sk_psock_stop_verdict(sk, psock);
 849         write_unlock_bh(&sk->sk_callback_lock);
 850
 851         sk_psock_stop(psock);
 852
 853         INIT_RCU_WORK(&psock->rwork, sk_psock_destroy);
 854         queue_rcu_work(system_wq, &psock->rwork);
 855 }
 856 EXPORT_SYMBOL_GPL(sk_psock_drop);
 857
 858 static int sk_psock_map_verd(int verdict, bool redir)
 859 {
 860         switch (verdict) {
 861         case SK_PASS:
 862                 return redir ? __SK_REDIRECT : __SK_PASS;
 863         case SK_DROP:
 864         default:
 865                 break;
 866         }
 867
 868         return __SK_DROP;
 869 }
 870
 871 int sk_psock_msg_verdict(struct sock *sk, struct sk_psock *psock,
 872                          struct sk_msg *msg)
 873 {
 874         struct bpf_prog *prog;
 875         int ret;
 876
 877         rcu_read_lock();
 878         prog = READ_ONCE(psock->progs.msg_parser);
 879         if (unlikely(!prog)) {
 880                 ret = __SK_PASS;
 881                 goto out;
 882         }
 883
 884         sk_msg_compute_data_pointers(msg);
 885         msg->sk = sk;
 886         ret = bpf_prog_run_pin_on_cpu(prog, msg);
 887         ret = sk_psock_map_verd(ret, msg->sk_redir);
 888         psock->apply_bytes = msg->apply_bytes;
 889         if (ret == __SK_REDIRECT) {
 890                 if (psock->sk_redir) {
 891                         sock_put(psock->sk_redir);
 892                         psock->sk_redir = NULL;
 893                 }
 894                 if (!msg->sk_redir) {
 895                         ret = __SK_DROP;
 896                         goto out;
 897                 }
 898                 psock->redir_ingress = sk_msg_to_ingress(msg);
 899                 psock->sk_redir = msg->sk_redir;
 900                 sock_hold(psock->sk_redir);
 901         }
 902 out:
 903         rcu_read_unlock();
 904         return ret;
 905 }
 906 EXPORT_SYMBOL_GPL(sk_psock_msg_verdict);
 907
 908 static int sk_psock_skb_redirect(struct sk_psock *from, struct sk_buff *skb)
 909 {
 910         struct sk_psock *psock_other;
 911         struct sock *sk_other;
 912
 913         sk_other = skb_bpf_redirect_fetch(skb);
 914         /* This error is a buggy BPF program, it returned a redirect
 915          * return code, but then didn't set a redirect interface.
 916          */
 917         if (unlikely(!sk_other)) {
 918                 skb_bpf_redirect_clear(skb);
 919                 sock_drop(from->sk, skb);
 920                 return -EIO;
 921         }
 922         psock_other = sk_psock(sk_other);
 923         /* This error indicates the socket is being torn down or had another
 924          * error that caused the pipe to break. We can't send a packet on
 925          * a socket that is in this state so we drop the skb.
 926          */
 927         if (!psock_other || sock_flag(sk_other, SOCK_DEAD)) {
 928                 skb_bpf_redirect_clear(skb);
 929                 sock_drop(from->sk, skb);
 930                 return -EIO;
 931         }
 932         spin_lock_bh(&psock_other->ingress_lock);
 933         if (!sk_psock_test_state(psock_other, SK_PSOCK_TX_ENABLED)) {
 934                 spin_unlock_bh(&psock_other->ingress_lock);
 935                 skb_bpf_redirect_clear(skb);
 936                 sock_drop(from->sk, skb);
 937                 return -EIO;
 938         }
 939
 940         skb_queue_tail(&psock_other->ingress_skb, skb);
 941         schedule_work(&psock_other->work);
 942         spin_unlock_bh(&psock_other->ingress_lock);
 943         return 0;
 944 }
 945
 946 static void sk_psock_tls_verdict_apply(struct sk_buff *skb,
 947                                        struct sk_psock *from, int verdict)
 948 {
 949         switch (verdict) {
 950         case __SK_REDIRECT:
 951                 sk_psock_skb_redirect(from, skb);
 952                 break;
 953         case __SK_PASS:
 954         case __SK_DROP:
 955         default:
 956                 break;
 957         }
 958 }
 959
 960 int sk_psock_tls_strp_read(struct sk_psock *psock, struct sk_buff *skb)
 961 {
 962         struct bpf_prog *prog;
 963         int ret = __SK_PASS;
 964
 965         rcu_read_lock();
 966         prog = READ_ONCE(psock->progs.stream_verdict);
 967         if (likely(prog)) {
 968                 skb->sk = psock->sk;
 969                 skb_dst_drop(skb);
 970                 skb_bpf_redirect_clear(skb);
 971                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
 972                 ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb));
 973                 skb->sk = NULL;
 974         }
 975         sk_psock_tls_verdict_apply(skb, psock, ret);
 976         rcu_read_unlock();
 977         return ret;
 978 }
 979 EXPORT_SYMBOL_GPL(sk_psock_tls_strp_read);
 980
 981 static int sk_psock_verdict_apply(struct sk_psock *psock, struct sk_buff *skb,
 982                                   int verdict)
 983 {
 984         struct sock *sk_other;
 985         int err = 0;
 986         u32 len, off;
 987
 988         switch (verdict) {
 989         case __SK_PASS:
 990                 err = -EIO;
 991                 sk_other = psock->sk;
 992                 if (sock_flag(sk_other, SOCK_DEAD) ||
 993                     !sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
 994                         skb_bpf_redirect_clear(skb);
 995                         goto out_free;
 996                 }
 997
 998                 skb_bpf_set_ingress(skb);
 999
1000                 /* If the queue is empty then we can submit directly
1001                  * into the msg queue. If its not empty we have to
1002                  * queue work otherwise we may get OOO data. Otherwise,
1003                  * if sk_psock_skb_ingress errors will be handled by
1004                  * retrying later from workqueue.
1005                  */
1006                 if (skb_queue_empty(&psock->ingress_skb)) {
1007                         len = skb->len;
1008                         off = 0;
1009                         if (skb_bpf_strparser(skb)) {
1010                                 struct strp_msg *stm = strp_msg(skb);
1011
1012                                 off = stm->offset;
1013                                 len = stm->full_len;
1014                         }
1015                         err = sk_psock_skb_ingress_self(psock, skb, off, len);
1016                 }
1017                 if (err < 0) {
1018                         spin_lock_bh(&psock->ingress_lock);
1019                         if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
1020                                 skb_queue_tail(&psock->ingress_skb, skb);
1021                                 schedule_work(&psock->work);
1022                                 err = 0;
1023                         }
1024                         spin_unlock_bh(&psock->ingress_lock);
1025                         if (err < 0) {
1026                                 skb_bpf_redirect_clear(skb);
1027                                 goto out_free;
1028                         }
1029                 }
1030                 break;
1031         case __SK_REDIRECT:
1032                 err = sk_psock_skb_redirect(psock, skb);
1033                 break;
1034         case __SK_DROP:
1035         default:
1036 out_free:
1037                 sock_drop(psock->sk, skb);
1038         }
1039
1040         return err;
1041 }
1042
1043 static void sk_psock_write_space(struct sock *sk)
1044 {
1045         struct sk_psock *psock;
1046         void (*write_space)(struct sock *sk) = NULL;
1047
1048         rcu_read_lock();
1049         psock = sk_psock(sk);
1050         if (likely(psock)) {
1051                 if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED))
1052                         schedule_work(&psock->work);
1053                 write_space = psock->saved_write_space;
1054         }
1055         rcu_read_unlock();
1056         if (write_space)
1057                 write_space(sk);
1058 }
1059
1060 #if IS_ENABLED(CONFIG_BPF_STREAM_PARSER)
1061 static void sk_psock_strp_read(struct strparser *strp, struct sk_buff *skb)
1062 {
1063         struct sk_psock *psock;
1064         struct bpf_prog *prog;
1065         int ret = __SK_DROP;
1066         struct sock *sk;
1067
1068         rcu_read_lock();
1069         sk = strp->sk;
1070         psock = sk_psock(sk);
1071         if (unlikely(!psock)) {
1072                 sock_drop(sk, skb);
1073                 goto out;
1074         }
1075         prog = READ_ONCE(psock->progs.stream_verdict);
1076         if (likely(prog)) {
1077                 skb->sk = sk;
1078                 skb_dst_drop(skb);
1079                 skb_bpf_redirect_clear(skb);
1080                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1081                 if (ret == SK_PASS)
1082                         skb_bpf_set_strparser(skb);
1083                 ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb));
1084                 skb->sk = NULL;
1085         }
1086         sk_psock_verdict_apply(psock, skb, ret);
1087 out:
1088         rcu_read_unlock();
1089 }
1090
1091 static int sk_psock_strp_read_done(struct strparser *strp, int err)
1092 {
1093         return err;
1094 }
1095
1096 static int sk_psock_strp_parse(struct strparser *strp, struct sk_buff *skb)
1097 {
1098         struct sk_psock *psock = container_of(strp, struct sk_psock, strp);
1099         struct bpf_prog *prog;
1100         int ret = skb->len;
1101
1102         rcu_read_lock();
1103         prog = READ_ONCE(psock->progs.stream_parser);
1104         if (likely(prog)) {
1105                 skb->sk = psock->sk;
1106                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1107                 skb->sk = NULL;
1108         }
1109         rcu_read_unlock();
1110         return ret;
1111 }
1112
1113 /* Called with socket lock held. */
1114 static void sk_psock_strp_data_ready(struct sock *sk)
1115 {
1116         struct sk_psock *psock;
1117
1118         trace_sk_data_ready(sk);
1119
1120         rcu_read_lock();
1121         psock = sk_psock(sk);
1122         if (likely(psock)) {
1123                 if (tls_sw_has_ctx_rx(sk)) {
1124                         psock->saved_data_ready(sk);
1125                 } else {
1126                         write_lock_bh(&sk->sk_callback_lock);
1127                         strp_data_ready(&psock->strp);
1128                         write_unlock_bh(&sk->sk_callback_lock);
1129                 }
1130         }
1131         rcu_read_unlock();
1132 }
1133
1134 int sk_psock_init_strp(struct sock *sk, struct sk_psock *psock)
1135 {
1136         static const struct strp_callbacks cb = {
1137                 .rcv_msg        = sk_psock_strp_read,
1138                 .read_sock_done = sk_psock_strp_read_done,
1139                 .parse_msg      = sk_psock_strp_parse,
1140         };
1141
1142         return strp_init(&psock->strp, sk, &cb);
1143 }
1144
1145 void sk_psock_start_strp(struct sock *sk, struct sk_psock *psock)
1146 {
1147         if (psock->saved_data_ready)
1148                 return;
1149
1150         psock->saved_data_ready = sk->sk_data_ready;
1151         sk->sk_data_ready = sk_psock_strp_data_ready;
1152         sk->sk_write_space = sk_psock_write_space;
1153 }
1154
1155 void sk_psock_stop_strp(struct sock *sk, struct sk_psock *psock)
1156 {
1157         psock_set_prog(&psock->progs.stream_parser, NULL);
1158
1159         if (!psock->saved_data_ready)
1160                 return;
1161
1162         sk->sk_data_ready = psock->saved_data_ready;
1163         psock->saved_data_ready = NULL;
1164         strp_stop(&psock->strp);
1165 }
1166
1167 static void sk_psock_done_strp(struct sk_psock *psock)
1168 {
1169         /* Parser has been stopped */
1170         if (psock->progs.stream_parser)
1171                 strp_done(&psock->strp);
1172 }
1173 #else
1174 static void sk_psock_done_strp(struct sk_psock *psock)
1175 {
1176 }
1177 #endif /* CONFIG_BPF_STREAM_PARSER */
1178
1179 static int sk_psock_verdict_recv(struct sock *sk, struct sk_buff *skb)
1180 {
1181         struct sk_psock *psock;
1182         struct bpf_prog *prog;
1183         int ret = __SK_DROP;
1184         int len = skb->len;
1185
1186         skb_get(skb);
1187
1188         rcu_read_lock();
1189         psock = sk_psock(sk);
1190         if (unlikely(!psock)) {
1191                 len = 0;
1192                 sock_drop(sk, skb);
1193                 goto out;
1194         }
1195         prog = READ_ONCE(psock->progs.stream_verdict);
1196         if (!prog)
1197                 prog = READ_ONCE(psock->progs.skb_verdict);
1198         if (likely(prog)) {
1199                 skb_dst_drop(skb);
1200                 skb_bpf_redirect_clear(skb);
1201                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1202                 ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb));
1203         }
1204         ret = sk_psock_verdict_apply(psock, skb, ret);
1205         if (ret < 0)
1206                 len = ret;
1207 out:
1208         rcu_read_unlock();
1209         return len;
1210 }
1211
1212 static void sk_psock_verdict_data_ready(struct sock *sk)
1213 {
1214         struct socket *sock = sk->sk_socket;
1215
1216         trace_sk_data_ready(sk);
1217
1218         if (unlikely(!sock || !sock->ops || !sock->ops->read_skb))
1219                 return;
1220         sock->ops->read_skb(sk, sk_psock_verdict_recv);
1221 }
1222
1223 void sk_psock_start_verdict(struct sock *sk, struct sk_psock *psock)
1224 {
1225         if (psock->saved_data_ready)
1226                 return;
1227
1228         psock->saved_data_ready = sk->sk_data_ready;
1229         sk->sk_data_ready = sk_psock_verdict_data_ready;
1230         sk->sk_write_space = sk_psock_write_space;
1231 }
1232
1233 void sk_psock_stop_verdict(struct sock *sk, struct sk_psock *psock)
1234 {
1235         psock_set_prog(&psock->progs.stream_verdict, NULL);
1236         psock_set_prog(&psock->progs.skb_verdict, NULL);
1237
1238         if (!psock->saved_data_ready)
1239                 return;
1240
1241         sk->sk_data_ready = psock->saved_data_ready;
1242         psock->saved_data_ready = NULL;
1243 }