net/core/skmsg.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* Copyright (c) 2017 - 2018 Covalent IO, Inc. http://covalent.io */
   3
   4 #include <linux/skmsg.h>
   5 #include <linux/skbuff.h>
   6 #include <linux/scatterlist.h>
   7
   8 #include <net/sock.h>
   9 #include <net/tcp.h>
  10 #include <net/tls.h>
  11
  12 static bool sk_msg_try_coalesce_ok(struct sk_msg *msg, int elem_first_coalesce)
  13 {
  14         if (msg->sg.end > msg->sg.start &&
  15             elem_first_coalesce < msg->sg.end)
  16                 return true;
  17
  18         if (msg->sg.end < msg->sg.start &&
  19             (elem_first_coalesce > msg->sg.start ||
  20              elem_first_coalesce < msg->sg.end))
  21                 return true;
  22
  23         return false;
  24 }
  25
  26 int sk_msg_alloc(struct sock *sk, struct sk_msg *msg, int len,
  27                  int elem_first_coalesce)
  28 {
  29         struct page_frag *pfrag = sk_page_frag(sk);
  30         u32 osize = msg->sg.size;
  31         int ret = 0;
  32
  33         len -= msg->sg.size;
  34         while (len > 0) {
  35                 struct scatterlist *sge;
  36                 u32 orig_offset;
  37                 int use, i;
  38
  39                 if (!sk_page_frag_refill(sk, pfrag)) {
  40                         ret = -ENOMEM;
  41                         goto msg_trim;
  42                 }
  43
  44                 orig_offset = pfrag->offset;
  45                 use = min_t(int, len, pfrag->size - orig_offset);
  46                 if (!sk_wmem_schedule(sk, use)) {
  47                         ret = -ENOMEM;
  48                         goto msg_trim;
  49                 }
  50
  51                 i = msg->sg.end;
  52                 sk_msg_iter_var_prev(i);
  53                 sge = &msg->sg.data[i];
  54
  55                 if (sk_msg_try_coalesce_ok(msg, elem_first_coalesce) &&
  56                     sg_page(sge) == pfrag->page &&
  57                     sge->offset + sge->length == orig_offset) {
  58                         sge->length += use;
  59                 } else {
  60                         if (sk_msg_full(msg)) {
  61                                 ret = -ENOSPC;
  62                                 break;
  63                         }
  64
  65                         sge = &msg->sg.data[msg->sg.end];
  66                         sg_unmark_end(sge);
  67                         sg_set_page(sge, pfrag->page, use, orig_offset);
  68                         get_page(pfrag->page);
  69                         sk_msg_iter_next(msg, end);
  70                 }
  71
  72                 sk_mem_charge(sk, use);
  73                 msg->sg.size += use;
  74                 pfrag->offset += use;
  75                 len -= use;
  76         }
  77
  78         return ret;
  79
  80 msg_trim:
  81         sk_msg_trim(sk, msg, osize);
  82         return ret;
  83 }
  84 EXPORT_SYMBOL_GPL(sk_msg_alloc);
  85
  86 int sk_msg_clone(struct sock *sk, struct sk_msg *dst, struct sk_msg *src,
  87                  u32 off, u32 len)
  88 {
  89         int i = src->sg.start;
  90         struct scatterlist *sge = sk_msg_elem(src, i);
  91         struct scatterlist *sgd = NULL;
  92         u32 sge_len, sge_off;
  93
  94         while (off) {
  95                 if (sge->length > off)
  96                         break;
  97                 off -= sge->length;
  98                 sk_msg_iter_var_next(i);
  99                 if (i == src->sg.end && off)
 100                         return -ENOSPC;
 101                 sge = sk_msg_elem(src, i);
 102         }
 103
 104         while (len) {
 105                 sge_len = sge->length - off;
 106                 if (sge_len > len)
 107                         sge_len = len;
 108
 109                 if (dst->sg.end)
 110                         sgd = sk_msg_elem(dst, dst->sg.end - 1);
 111
 112                 if (sgd &&
 113                     (sg_page(sge) == sg_page(sgd)) &&
 114                     (sg_virt(sge) + off == sg_virt(sgd) + sgd->length)) {
 115                         sgd->length += sge_len;
 116                         dst->sg.size += sge_len;
 117                 } else if (!sk_msg_full(dst)) {
 118                         sge_off = sge->offset + off;
 119                         sk_msg_page_add(dst, sg_page(sge), sge_len, sge_off);
 120                 } else {
 121                         return -ENOSPC;
 122                 }
 123
 124                 off = 0;
 125                 len -= sge_len;
 126                 sk_mem_charge(sk, sge_len);
 127                 sk_msg_iter_var_next(i);
 128                 if (i == src->sg.end && len)
 129                         return -ENOSPC;
 130                 sge = sk_msg_elem(src, i);
 131         }
 132
 133         return 0;
 134 }
 135 EXPORT_SYMBOL_GPL(sk_msg_clone);
 136
 137 void sk_msg_return_zero(struct sock *sk, struct sk_msg *msg, int bytes)
 138 {
 139         int i = msg->sg.start;
 140
 141         do {
 142                 struct scatterlist *sge = sk_msg_elem(msg, i);
 143
 144                 if (bytes < sge->length) {
 145                         sge->length -= bytes;
 146                         sge->offset += bytes;
 147                         sk_mem_uncharge(sk, bytes);
 148                         break;
 149                 }
 150
 151                 sk_mem_uncharge(sk, sge->length);
 152                 bytes -= sge->length;
 153                 sge->length = 0;
 154                 sge->offset = 0;
 155                 sk_msg_iter_var_next(i);
 156         } while (bytes && i != msg->sg.end);
 157         msg->sg.start = i;
 158 }
 159 EXPORT_SYMBOL_GPL(sk_msg_return_zero);
 160
 161 void sk_msg_return(struct sock *sk, struct sk_msg *msg, int bytes)
 162 {
 163         int i = msg->sg.start;
 164
 165         do {
 166                 struct scatterlist *sge = &msg->sg.data[i];
 167                 int uncharge = (bytes < sge->length) ? bytes : sge->length;
 168
 169                 sk_mem_uncharge(sk, uncharge);
 170                 bytes -= uncharge;
 171                 sk_msg_iter_var_next(i);
 172         } while (i != msg->sg.end);
 173 }
 174 EXPORT_SYMBOL_GPL(sk_msg_return);
 175
 176 static int sk_msg_free_elem(struct sock *sk, struct sk_msg *msg, u32 i,
 177                             bool charge)
 178 {
 179         struct scatterlist *sge = sk_msg_elem(msg, i);
 180         u32 len = sge->length;
 181
 182         /* When the skb owns the memory we free it from consume_skb path. */
 183         if (!msg->skb) {
 184                 if (charge)
 185                         sk_mem_uncharge(sk, len);
 186                 put_page(sg_page(sge));
 187         }
 188         memset(sge, 0, sizeof(*sge));
 189         return len;
 190 }
 191
 192 static int __sk_msg_free(struct sock *sk, struct sk_msg *msg, u32 i,
 193                          bool charge)
 194 {
 195         struct scatterlist *sge = sk_msg_elem(msg, i);
 196         int freed = 0;
 197
 198         while (msg->sg.size) {
 199                 msg->sg.size -= sge->length;
 200                 freed += sk_msg_free_elem(sk, msg, i, charge);
 201                 sk_msg_iter_var_next(i);
 202                 sk_msg_check_to_free(msg, i, msg->sg.size);
 203                 sge = sk_msg_elem(msg, i);
 204         }
 205         consume_skb(msg->skb);
 206         sk_msg_init(msg);
 207         return freed;
 208 }
 209
 210 int sk_msg_free_nocharge(struct sock *sk, struct sk_msg *msg)
 211 {
 212         return __sk_msg_free(sk, msg, msg->sg.start, false);
 213 }
 214 EXPORT_SYMBOL_GPL(sk_msg_free_nocharge);
 215
 216 int sk_msg_free(struct sock *sk, struct sk_msg *msg)
 217 {
 218         return __sk_msg_free(sk, msg, msg->sg.start, true);
 219 }
 220 EXPORT_SYMBOL_GPL(sk_msg_free);
 221
 222 static void __sk_msg_free_partial(struct sock *sk, struct sk_msg *msg,
 223                                   u32 bytes, bool charge)
 224 {
 225         struct scatterlist *sge;
 226         u32 i = msg->sg.start;
 227
 228         while (bytes) {
 229                 sge = sk_msg_elem(msg, i);
 230                 if (!sge->length)
 231                         break;
 232                 if (bytes < sge->length) {
 233                         if (charge)
 234                                 sk_mem_uncharge(sk, bytes);
 235                         sge->length -= bytes;
 236                         sge->offset += bytes;
 237                         msg->sg.size -= bytes;
 238                         break;
 239                 }
 240
 241                 msg->sg.size -= sge->length;
 242                 bytes -= sge->length;
 243                 sk_msg_free_elem(sk, msg, i, charge);
 244                 sk_msg_iter_var_next(i);
 245                 sk_msg_check_to_free(msg, i, bytes);
 246         }
 247         msg->sg.start = i;
 248 }
 249
 250 void sk_msg_free_partial(struct sock *sk, struct sk_msg *msg, u32 bytes)
 251 {
 252         __sk_msg_free_partial(sk, msg, bytes, true);
 253 }
 254 EXPORT_SYMBOL_GPL(sk_msg_free_partial);
 255
 256 void sk_msg_free_partial_nocharge(struct sock *sk, struct sk_msg *msg,
 257                                   u32 bytes)
 258 {
 259         __sk_msg_free_partial(sk, msg, bytes, false);
 260 }
 261
 262 void sk_msg_trim(struct sock *sk, struct sk_msg *msg, int len)
 263 {
 264         int trim = msg->sg.size - len;
 265         u32 i = msg->sg.end;
 266
 267         if (trim <= 0) {
 268                 WARN_ON(trim < 0);
 269                 return;
 270         }
 271
 272         sk_msg_iter_var_prev(i);
 273         msg->sg.size = len;
 274         while (msg->sg.data[i].length &&
 275                trim >= msg->sg.data[i].length) {
 276                 trim -= msg->sg.data[i].length;
 277                 sk_msg_free_elem(sk, msg, i, true);
 278                 sk_msg_iter_var_prev(i);
 279                 if (!trim)
 280                         goto out;
 281         }
 282
 283         msg->sg.data[i].length -= trim;
 284         sk_mem_uncharge(sk, trim);
 285         /* Adjust copybreak if it falls into the trimmed part of last buf */
 286         if (msg->sg.curr == i && msg->sg.copybreak > msg->sg.data[i].length)
 287                 msg->sg.copybreak = msg->sg.data[i].length;
 288 out:
 289         sk_msg_iter_var_next(i);
 290         msg->sg.end = i;
 291
 292         /* If we trim data a full sg elem before curr pointer update
 293          * copybreak and current so that any future copy operations
 294          * start at new copy location.
 295          * However trimed data that has not yet been used in a copy op
 296          * does not require an update.
 297          */
 298         if (!msg->sg.size) {
 299                 msg->sg.curr = msg->sg.start;
 300                 msg->sg.copybreak = 0;
 301         } else if (sk_msg_iter_dist(msg->sg.start, msg->sg.curr) >=
 302                    sk_msg_iter_dist(msg->sg.start, msg->sg.end)) {
 303                 sk_msg_iter_var_prev(i);
 304                 msg->sg.curr = i;
 305                 msg->sg.copybreak = msg->sg.data[i].length;
 306         }
 307 }
 308 EXPORT_SYMBOL_GPL(sk_msg_trim);
 309
 310 int sk_msg_zerocopy_from_iter(struct sock *sk, struct iov_iter *from,
 311                               struct sk_msg *msg, u32 bytes)
 312 {
 313         int i, maxpages, ret = 0, num_elems = sk_msg_elem_used(msg);
 314         const int to_max_pages = MAX_MSG_FRAGS;
 315         struct page *pages[MAX_MSG_FRAGS];
 316         ssize_t orig, copied, use, offset;
 317
 318         orig = msg->sg.size;
 319         while (bytes > 0) {
 320                 i = 0;
 321                 maxpages = to_max_pages - num_elems;
 322                 if (maxpages == 0) {
 323                         ret = -EFAULT;
 324                         goto out;
 325                 }
 326
 327                 copied = iov_iter_get_pages2(from, pages, bytes, maxpages,
 328                                             &offset);
 329                 if (copied <= 0) {
 330                         ret = -EFAULT;
 331                         goto out;
 332                 }
 333
 334                 bytes -= copied;
 335                 msg->sg.size += copied;
 336
 337                 while (copied) {
 338                         use = min_t(int, copied, PAGE_SIZE - offset);
 339                         sg_set_page(&msg->sg.data[msg->sg.end],
 340                                     pages[i], use, offset);
 341                         sg_unmark_end(&msg->sg.data[msg->sg.end]);
 342                         sk_mem_charge(sk, use);
 343
 344                         offset = 0;
 345                         copied -= use;
 346                         sk_msg_iter_next(msg, end);
 347                         num_elems++;
 348                         i++;
 349                 }
 350                 /* When zerocopy is mixed with sk_msg_*copy* operations we
 351                  * may have a copybreak set in this case clear and prefer
 352                  * zerocopy remainder when possible.
 353                  */
 354                 msg->sg.copybreak = 0;
 355                 msg->sg.curr = msg->sg.end;
 356         }
 357 out:
 358         /* Revert iov_iter updates, msg will need to use 'trim' later if it
 359          * also needs to be cleared.
 360          */
 361         if (ret)
 362                 iov_iter_revert(from, msg->sg.size - orig);
 363         return ret;
 364 }
 365 EXPORT_SYMBOL_GPL(sk_msg_zerocopy_from_iter);
 366
 367 int sk_msg_memcopy_from_iter(struct sock *sk, struct iov_iter *from,
 368                              struct sk_msg *msg, u32 bytes)
 369 {
 370         int ret = -ENOSPC, i = msg->sg.curr;
 371         struct scatterlist *sge;
 372         u32 copy, buf_size;
 373         void *to;
 374
 375         do {
 376                 sge = sk_msg_elem(msg, i);
 377                 /* This is possible if a trim operation shrunk the buffer */
 378                 if (msg->sg.copybreak >= sge->length) {
 379                         msg->sg.copybreak = 0;
 380                         sk_msg_iter_var_next(i);
 381                         if (i == msg->sg.end)
 382                                 break;
 383                         sge = sk_msg_elem(msg, i);
 384                 }
 385
 386                 buf_size = sge->length - msg->sg.copybreak;
 387                 copy = (buf_size > bytes) ? bytes : buf_size;
 388                 to = sg_virt(sge) + msg->sg.copybreak;
 389                 msg->sg.copybreak += copy;
 390                 if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY)
 391                         ret = copy_from_iter_nocache(to, copy, from);
 392                 else
 393                         ret = copy_from_iter(to, copy, from);
 394                 if (ret != copy) {
 395                         ret = -EFAULT;
 396                         goto out;
 397                 }
 398                 bytes -= copy;
 399                 if (!bytes)
 400                         break;
 401                 msg->sg.copybreak = 0;
 402                 sk_msg_iter_var_next(i);
 403         } while (i != msg->sg.end);
 404 out:
 405         msg->sg.curr = i;
 406         return ret;
 407 }
 408 EXPORT_SYMBOL_GPL(sk_msg_memcopy_from_iter);
 409
 410 /* Receive sk_msg from psock->ingress_msg to @msg. */
 411 int sk_msg_recvmsg(struct sock *sk, struct sk_psock *psock, struct msghdr *msg,
 412                    int len, int flags)
 413 {
 414         struct iov_iter *iter = &msg->msg_iter;
 415         int peek = flags & MSG_PEEK;
 416         struct sk_msg *msg_rx;
 417         int i, copied = 0;
 418
 419         msg_rx = sk_psock_peek_msg(psock);
 420         while (copied != len) {
 421                 struct scatterlist *sge;
 422
 423                 if (unlikely(!msg_rx))
 424                         break;
 425
 426                 i = msg_rx->sg.start;
 427                 do {
 428                         struct page *page;
 429                         int copy;
 430
 431                         sge = sk_msg_elem(msg_rx, i);
 432                         copy = sge->length;
 433                         page = sg_page(sge);
 434                         if (copied + copy > len)
 435                                 copy = len - copied;
 436                         copy = copy_page_to_iter(page, sge->offset, copy, iter);
 437                         if (!copy)
 438                                 return copied ? copied : -EFAULT;
 439
 440                         copied += copy;
 441                         if (likely(!peek)) {
 442                                 sge->offset += copy;
 443                                 sge->length -= copy;
 444                                 if (!msg_rx->skb)
 445                                         sk_mem_uncharge(sk, copy);
 446                                 msg_rx->sg.size -= copy;
 447
 448                                 if (!sge->length) {
 449                                         sk_msg_iter_var_next(i);
 450                                         if (!msg_rx->skb)
 451                                                 put_page(page);
 452                                 }
 453                         } else {
 454                                 /* Lets not optimize peek case if copy_page_to_iter
 455                                  * didn't copy the entire length lets just break.
 456                                  */
 457                                 if (copy != sge->length)
 458                                         return copied;
 459                                 sk_msg_iter_var_next(i);
 460                         }
 461
 462                         if (copied == len)
 463                                 break;
 464                 } while ((i != msg_rx->sg.end) && !sg_is_last(sge));
 465
 466                 if (unlikely(peek)) {
 467                         msg_rx = sk_psock_next_msg(psock, msg_rx);
 468                         if (!msg_rx)
 469                                 break;
 470                         continue;
 471                 }
 472
 473                 msg_rx->sg.start = i;
 474                 if (!sge->length && (i == msg_rx->sg.end || sg_is_last(sge))) {
 475                         msg_rx = sk_psock_dequeue_msg(psock);
 476                         kfree_sk_msg(msg_rx);
 477                 }
 478                 msg_rx = sk_psock_peek_msg(psock);
 479         }
 480
 481         return copied;
 482 }
 483 EXPORT_SYMBOL_GPL(sk_msg_recvmsg);
 484
 485 bool sk_msg_is_readable(struct sock *sk)
 486 {
 487         struct sk_psock *psock;
 488         bool empty = true;
 489
 490         rcu_read_lock();
 491         psock = sk_psock(sk);
 492         if (likely(psock))
 493                 empty = list_empty(&psock->ingress_msg);
 494         rcu_read_unlock();
 495         return !empty;
 496 }
 497 EXPORT_SYMBOL_GPL(sk_msg_is_readable);
 498
 499 static struct sk_msg *alloc_sk_msg(void)
 500 {
 501         struct sk_msg *msg;
 502
 503         msg = kzalloc(sizeof(*msg), __GFP_NOWARN | GFP_KERNEL);
 504         if (unlikely(!msg))
 505                 return NULL;
 506         sg_init_marker(msg->sg.data, NR_MSG_FRAG_IDS);
 507         return msg;
 508 }
 509
 510 static struct sk_msg *sk_psock_create_ingress_msg(struct sock *sk,
 511                                                   struct sk_buff *skb)
 512 {
 513         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
 514                 return NULL;
 515
 516         if (!sk_rmem_schedule(sk, skb, skb->truesize))
 517                 return NULL;
 518
 519         return alloc_sk_msg();
 520 }
 521
 522 static int sk_psock_skb_ingress_enqueue(struct sk_buff *skb,
 523                                         u32 off, u32 len,
 524                                         struct sk_psock *psock,
 525                                         struct sock *sk,
 526                                         struct sk_msg *msg)
 527 {
 528         int num_sge, copied;
 529
 530         num_sge = skb_to_sgvec(skb, msg->sg.data, off, len);
 531         if (num_sge < 0) {
 532                 /* skb linearize may fail with ENOMEM, but lets simply try again
 533                  * later if this happens. Under memory pressure we don't want to
 534                  * drop the skb. We need to linearize the skb so that the mapping
 535                  * in skb_to_sgvec can not error.
 536                  */
 537                 if (skb_linearize(skb))
 538                         return -EAGAIN;
 539
 540                 num_sge = skb_to_sgvec(skb, msg->sg.data, off, len);
 541                 if (unlikely(num_sge < 0))
 542                         return num_sge;
 543         }
 544
 545         copied = len;
 546         msg->sg.start = 0;
 547         msg->sg.size = copied;
 548         msg->sg.end = num_sge;
 549         msg->skb = skb;
 550
 551         sk_psock_queue_msg(psock, msg);
 552         sk_psock_data_ready(sk, psock);
 553         return copied;
 554 }
 555
 556 static int sk_psock_skb_ingress_self(struct sk_psock *psock, struct sk_buff *skb,
 557                                      u32 off, u32 len);
 558
 559 static int sk_psock_skb_ingress(struct sk_psock *psock, struct sk_buff *skb,
 560                                 u32 off, u32 len)
 561 {
 562         struct sock *sk = psock->sk;
 563         struct sk_msg *msg;
 564         int err;
 565
 566         /* If we are receiving on the same sock skb->sk is already assigned,
 567          * skip memory accounting and owner transition seeing it already set
 568          * correctly.
 569          */
 570         if (unlikely(skb->sk == sk))
 571                 return sk_psock_skb_ingress_self(psock, skb, off, len);
 572         msg = sk_psock_create_ingress_msg(sk, skb);
 573         if (!msg)
 574                 return -EAGAIN;
 575
 576         /* This will transition ownership of the data from the socket where
 577          * the BPF program was run initiating the redirect to the socket
 578          * we will eventually receive this data on. The data will be released
 579          * from skb_consume found in __tcp_bpf_recvmsg() after its been copied
 580          * into user buffers.
 581          */
 582         skb_set_owner_r(skb, sk);
 583         err = sk_psock_skb_ingress_enqueue(skb, off, len, psock, sk, msg);
 584         if (err < 0)
 585                 kfree(msg);
 586         return err;
 587 }
 588
 589 /* Puts an skb on the ingress queue of the socket already assigned to the
 590  * skb. In this case we do not need to check memory limits or skb_set_owner_r
 591  * because the skb is already accounted for here.
 592  */
 593 static int sk_psock_skb_ingress_self(struct sk_psock *psock, struct sk_buff *skb,
 594                                      u32 off, u32 len)
 595 {
 596         struct sk_msg *msg = alloc_sk_msg();
 597         struct sock *sk = psock->sk;
 598         int err;
 599
 600         if (unlikely(!msg))
 601                 return -EAGAIN;
 602         skb_set_owner_r(skb, sk);
 603         err = sk_psock_skb_ingress_enqueue(skb, off, len, psock, sk, msg);
 604         if (err < 0)
 605                 kfree(msg);
 606         return err;
 607 }
 608
 609 static int sk_psock_handle_skb(struct sk_psock *psock, struct sk_buff *skb,
 610                                u32 off, u32 len, bool ingress)
 611 {
 612         if (!ingress) {
 613                 if (!sock_writeable(psock->sk))
 614                         return -EAGAIN;
 615                 return skb_send_sock(psock->sk, skb, off, len);
 616         }
 617         return sk_psock_skb_ingress(psock, skb, off, len);
 618 }
 619
 620 static void sk_psock_skb_state(struct sk_psock *psock,
 621                                struct sk_psock_work_state *state,
 622                                struct sk_buff *skb,
 623                                int len, int off)
 624 {
 625         spin_lock_bh(&psock->ingress_lock);
 626         if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
 627                 state->skb = skb;
 628                 state->len = len;
 629                 state->off = off;
 630         } else {
 631                 sock_drop(psock->sk, skb);
 632         }
 633         spin_unlock_bh(&psock->ingress_lock);
 634 }
 635
 636 static void sk_psock_backlog(struct work_struct *work)
 637 {
 638         struct sk_psock *psock = container_of(work, struct sk_psock, work);
 639         struct sk_psock_work_state *state = &psock->work_state;
 640         struct sk_buff *skb = NULL;
 641         bool ingress;
 642         u32 len, off;
 643         int ret;
 644
 645         mutex_lock(&psock->work_mutex);
 646         if (unlikely(state->skb)) {
 647                 spin_lock_bh(&psock->ingress_lock);
 648                 skb = state->skb;
 649                 len = state->len;
 650                 off = state->off;
 651                 state->skb = NULL;
 652                 spin_unlock_bh(&psock->ingress_lock);
 653         }
 654         if (skb)
 655                 goto start;
 656
 657         while ((skb = skb_dequeue(&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 start:
 667                 ingress = skb_bpf_ingress(skb);
 668                 skb_bpf_redirect_clear(skb);
 669                 do {
 670                         ret = -EIO;
 671                         if (!sock_flag(psock->sk, SOCK_DEAD))
 672                                 ret = sk_psock_handle_skb(psock, skb, off,
 673                                                           len, ingress);
 674                         if (ret <= 0) {
 675                                 if (ret == -EAGAIN) {
 676                                         sk_psock_skb_state(psock, state, skb,
 677                                                            len, off);
 678                                         goto end;
 679                                 }
 680                                 /* Hard errors break pipe and stop xmit. */
 681                                 sk_psock_report_error(psock, ret ? -ret : EPIPE);
 682                                 sk_psock_clear_state(psock, SK_PSOCK_TX_ENABLED);
 683                                 sock_drop(psock->sk, skb);
 684                                 goto end;
 685                         }
 686                         off += ret;
 687                         len -= ret;
 688                 } while (len);
 689
 690                 if (!ingress)
 691                         kfree_skb(skb);
 692         }
 693 end:
 694         mutex_unlock(&psock->work_mutex);
 695 }
 696
 697 struct sk_psock *sk_psock_init(struct sock *sk, int node)
 698 {
 699         struct sk_psock *psock;
 700         struct proto *prot;
 701
 702         write_lock_bh(&sk->sk_callback_lock);
 703
 704         if (sk_is_inet(sk) && inet_csk_has_ulp(sk)) {
 705                 psock = ERR_PTR(-EINVAL);
 706                 goto out;
 707         }
 708
 709         if (sk->sk_user_data) {
 710                 psock = ERR_PTR(-EBUSY);
 711                 goto out;
 712         }
 713
 714         psock = kzalloc_node(sizeof(*psock), GFP_ATOMIC | __GFP_NOWARN, node);
 715         if (!psock) {
 716                 psock = ERR_PTR(-ENOMEM);
 717                 goto out;
 718         }
 719
 720         prot = READ_ONCE(sk->sk_prot);
 721         psock->sk = sk;
 722         psock->eval = __SK_NONE;
 723         psock->sk_proto = prot;
 724         psock->saved_unhash = prot->unhash;
 725         psock->saved_destroy = prot->destroy;
 726         psock->saved_close = prot->close;
 727         psock->saved_write_space = sk->sk_write_space;
 728
 729         INIT_LIST_HEAD(&psock->link);
 730         spin_lock_init(&psock->link_lock);
 731
 732         INIT_WORK(&psock->work, sk_psock_backlog);
 733         mutex_init(&psock->work_mutex);
 734         INIT_LIST_HEAD(&psock->ingress_msg);
 735         spin_lock_init(&psock->ingress_lock);
 736         skb_queue_head_init(&psock->ingress_skb);
 737
 738         sk_psock_set_state(psock, SK_PSOCK_TX_ENABLED);
 739         refcount_set(&psock->refcnt, 1);
 740
 741         __rcu_assign_sk_user_data_with_flags(sk, psock,
 742                                              SK_USER_DATA_NOCOPY |
 743                                              SK_USER_DATA_PSOCK);
 744         sock_hold(sk);
 745
 746 out:
 747         write_unlock_bh(&sk->sk_callback_lock);
 748         return psock;
 749 }
 750 EXPORT_SYMBOL_GPL(sk_psock_init);
 751
 752 struct sk_psock_link *sk_psock_link_pop(struct sk_psock *psock)
 753 {
 754         struct sk_psock_link *link;
 755
 756         spin_lock_bh(&psock->link_lock);
 757         link = list_first_entry_or_null(&psock->link, struct sk_psock_link,
 758                                         list);
 759         if (link)
 760                 list_del(&link->list);
 761         spin_unlock_bh(&psock->link_lock);
 762         return link;
 763 }
 764
 765 static void __sk_psock_purge_ingress_msg(struct sk_psock *psock)
 766 {
 767         struct sk_msg *msg, *tmp;
 768
 769         list_for_each_entry_safe(msg, tmp, &psock->ingress_msg, list) {
 770                 list_del(&msg->list);
 771                 sk_msg_free(psock->sk, msg);
 772                 kfree(msg);
 773         }
 774 }
 775
 776 static void __sk_psock_zap_ingress(struct sk_psock *psock)
 777 {
 778         struct sk_buff *skb;
 779
 780         while ((skb = skb_dequeue(&psock->ingress_skb)) != NULL) {
 781                 skb_bpf_redirect_clear(skb);
 782                 sock_drop(psock->sk, skb);
 783         }
 784         kfree_skb(psock->work_state.skb);
 785         /* We null the skb here to ensure that calls to sk_psock_backlog
 786          * do not pick up the free'd skb.
 787          */
 788         psock->work_state.skb = NULL;
 789         __sk_psock_purge_ingress_msg(psock);
 790 }
 791
 792 static void sk_psock_link_destroy(struct sk_psock *psock)
 793 {
 794         struct sk_psock_link *link, *tmp;
 795
 796         list_for_each_entry_safe(link, tmp, &psock->link, list) {
 797                 list_del(&link->list);
 798                 sk_psock_free_link(link);
 799         }
 800 }
 801
 802 void sk_psock_stop(struct sk_psock *psock, bool wait)
 803 {
 804         spin_lock_bh(&psock->ingress_lock);
 805         sk_psock_clear_state(psock, SK_PSOCK_TX_ENABLED);
 806         sk_psock_cork_free(psock);
 807         __sk_psock_zap_ingress(psock);
 808         spin_unlock_bh(&psock->ingress_lock);
 809
 810         if (wait)
 811                 cancel_work_sync(&psock->work);
 812 }
 813
 814 static void sk_psock_done_strp(struct sk_psock *psock);
 815
 816 static void sk_psock_destroy(struct work_struct *work)
 817 {
 818         struct sk_psock *psock = container_of(to_rcu_work(work),
 819                                               struct sk_psock, rwork);
 820         /* No sk_callback_lock since already detached. */
 821
 822         sk_psock_done_strp(psock);
 823
 824         cancel_work_sync(&psock->work);
 825         mutex_destroy(&psock->work_mutex);
 826
 827         psock_progs_drop(&psock->progs);
 828
 829         sk_psock_link_destroy(psock);
 830         sk_psock_cork_free(psock);
 831
 832         if (psock->sk_redir)
 833                 sock_put(psock->sk_redir);
 834         sock_put(psock->sk);
 835         kfree(psock);
 836 }
 837
 838 void sk_psock_drop(struct sock *sk, struct sk_psock *psock)
 839 {
 840         write_lock_bh(&sk->sk_callback_lock);
 841         sk_psock_restore_proto(sk, psock);
 842         rcu_assign_sk_user_data(sk, NULL);
 843         if (psock->progs.stream_parser)
 844                 sk_psock_stop_strp(sk, psock);
 845         else if (psock->progs.stream_verdict || psock->progs.skb_verdict)
 846                 sk_psock_stop_verdict(sk, psock);
 847         write_unlock_bh(&sk->sk_callback_lock);
 848
 849         sk_psock_stop(psock, false);
 850
 851         INIT_RCU_WORK(&psock->rwork, sk_psock_destroy);
 852         queue_rcu_work(system_wq, &psock->rwork);
 853 }
 854 EXPORT_SYMBOL_GPL(sk_psock_drop);
 855
 856 static int sk_psock_map_verd(int verdict, bool redir)
 857 {
 858         switch (verdict) {
 859         case SK_PASS:
 860                 return redir ? __SK_REDIRECT : __SK_PASS;
 861         case SK_DROP:
 862         default:
 863                 break;
 864         }
 865
 866         return __SK_DROP;
 867 }
 868
 869 int sk_psock_msg_verdict(struct sock *sk, struct sk_psock *psock,
 870                          struct sk_msg *msg)
 871 {
 872         struct bpf_prog *prog;
 873         int ret;
 874
 875         rcu_read_lock();
 876         prog = READ_ONCE(psock->progs.msg_parser);
 877         if (unlikely(!prog)) {
 878                 ret = __SK_PASS;
 879                 goto out;
 880         }
 881
 882         sk_msg_compute_data_pointers(msg);
 883         msg->sk = sk;
 884         ret = bpf_prog_run_pin_on_cpu(prog, msg);
 885         ret = sk_psock_map_verd(ret, msg->sk_redir);
 886         psock->apply_bytes = msg->apply_bytes;
 887         if (ret == __SK_REDIRECT) {
 888                 if (psock->sk_redir)
 889                         sock_put(psock->sk_redir);
 890                 psock->sk_redir = msg->sk_redir;
 891                 if (!psock->sk_redir) {
 892                         ret = __SK_DROP;
 893                         goto out;
 894                 }
 895                 sock_hold(psock->sk_redir);
 896         }
 897 out:
 898         rcu_read_unlock();
 899         return ret;
 900 }
 901 EXPORT_SYMBOL_GPL(sk_psock_msg_verdict);
 902
 903 static int sk_psock_skb_redirect(struct sk_psock *from, struct sk_buff *skb)
 904 {
 905         struct sk_psock *psock_other;
 906         struct sock *sk_other;
 907
 908         sk_other = skb_bpf_redirect_fetch(skb);
 909         /* This error is a buggy BPF program, it returned a redirect
 910          * return code, but then didn't set a redirect interface.
 911          */
 912         if (unlikely(!sk_other)) {
 913                 skb_bpf_redirect_clear(skb);
 914                 sock_drop(from->sk, skb);
 915                 return -EIO;
 916         }
 917         psock_other = sk_psock(sk_other);
 918         /* This error indicates the socket is being torn down or had another
 919          * error that caused the pipe to break. We can't send a packet on
 920          * a socket that is in this state so we drop the skb.
 921          */
 922         if (!psock_other || sock_flag(sk_other, SOCK_DEAD)) {
 923                 skb_bpf_redirect_clear(skb);
 924                 sock_drop(from->sk, skb);
 925                 return -EIO;
 926         }
 927         spin_lock_bh(&psock_other->ingress_lock);
 928         if (!sk_psock_test_state(psock_other, SK_PSOCK_TX_ENABLED)) {
 929                 spin_unlock_bh(&psock_other->ingress_lock);
 930                 skb_bpf_redirect_clear(skb);
 931                 sock_drop(from->sk, skb);
 932                 return -EIO;
 933         }
 934
 935         skb_queue_tail(&psock_other->ingress_skb, skb);
 936         schedule_work(&psock_other->work);
 937         spin_unlock_bh(&psock_other->ingress_lock);
 938         return 0;
 939 }
 940
 941 static void sk_psock_tls_verdict_apply(struct sk_buff *skb,
 942                                        struct sk_psock *from, int verdict)
 943 {
 944         switch (verdict) {
 945         case __SK_REDIRECT:
 946                 sk_psock_skb_redirect(from, skb);
 947                 break;
 948         case __SK_PASS:
 949         case __SK_DROP:
 950         default:
 951                 break;
 952         }
 953 }
 954
 955 int sk_psock_tls_strp_read(struct sk_psock *psock, struct sk_buff *skb)
 956 {
 957         struct bpf_prog *prog;
 958         int ret = __SK_PASS;
 959
 960         rcu_read_lock();
 961         prog = READ_ONCE(psock->progs.stream_verdict);
 962         if (likely(prog)) {
 963                 skb->sk = psock->sk;
 964                 skb_dst_drop(skb);
 965                 skb_bpf_redirect_clear(skb);
 966                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
 967                 ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb));
 968                 skb->sk = NULL;
 969         }
 970         sk_psock_tls_verdict_apply(skb, psock, ret);
 971         rcu_read_unlock();
 972         return ret;
 973 }
 974 EXPORT_SYMBOL_GPL(sk_psock_tls_strp_read);
 975
 976 static int sk_psock_verdict_apply(struct sk_psock *psock, struct sk_buff *skb,
 977                                   int verdict)
 978 {
 979         struct sock *sk_other;
 980         int err = 0;
 981         u32 len, off;
 982
 983         switch (verdict) {
 984         case __SK_PASS:
 985                 err = -EIO;
 986                 sk_other = psock->sk;
 987                 if (sock_flag(sk_other, SOCK_DEAD) ||
 988                     !sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
 989                         skb_bpf_redirect_clear(skb);
 990                         goto out_free;
 991                 }
 992
 993                 skb_bpf_set_ingress(skb);
 994
 995                 /* If the queue is empty then we can submit directly
 996                  * into the msg queue. If its not empty we have to
 997                  * queue work otherwise we may get OOO data. Otherwise,
 998                  * if sk_psock_skb_ingress errors will be handled by
 999                  * retrying later from workqueue.
1000                  */
1001                 if (skb_queue_empty(&psock->ingress_skb)) {
1002                         len = skb->len;
1003                         off = 0;
1004                         if (skb_bpf_strparser(skb)) {
1005                                 struct strp_msg *stm = strp_msg(skb);
1006
1007                                 off = stm->offset;
1008                                 len = stm->full_len;
1009                         }
1010                         err = sk_psock_skb_ingress_self(psock, skb, off, len);
1011                 }
1012                 if (err < 0) {
1013                         spin_lock_bh(&psock->ingress_lock);
1014                         if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
1015                                 skb_queue_tail(&psock->ingress_skb, skb);
1016                                 schedule_work(&psock->work);
1017                                 err = 0;
1018                         }
1019                         spin_unlock_bh(&psock->ingress_lock);
1020                         if (err < 0) {
1021                                 skb_bpf_redirect_clear(skb);
1022                                 goto out_free;
1023                         }
1024                 }
1025                 break;
1026         case __SK_REDIRECT:
1027                 err = sk_psock_skb_redirect(psock, skb);
1028                 break;
1029         case __SK_DROP:
1030         default:
1031 out_free:
1032                 sock_drop(psock->sk, skb);
1033         }
1034
1035         return err;
1036 }
1037
1038 static void sk_psock_write_space(struct sock *sk)
1039 {
1040         struct sk_psock *psock;
1041         void (*write_space)(struct sock *sk) = NULL;
1042
1043         rcu_read_lock();
1044         psock = sk_psock(sk);
1045         if (likely(psock)) {
1046                 if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED))
1047                         schedule_work(&psock->work);
1048                 write_space = psock->saved_write_space;
1049         }
1050         rcu_read_unlock();
1051         if (write_space)
1052                 write_space(sk);
1053 }
1054
1055 #if IS_ENABLED(CONFIG_BPF_STREAM_PARSER)
1056 static void sk_psock_strp_read(struct strparser *strp, struct sk_buff *skb)
1057 {
1058         struct sk_psock *psock;
1059         struct bpf_prog *prog;
1060         int ret = __SK_DROP;
1061         struct sock *sk;
1062
1063         rcu_read_lock();
1064         sk = strp->sk;
1065         psock = sk_psock(sk);
1066         if (unlikely(!psock)) {
1067                 sock_drop(sk, skb);
1068                 goto out;
1069         }
1070         prog = READ_ONCE(psock->progs.stream_verdict);
1071         if (likely(prog)) {
1072                 skb->sk = sk;
1073                 skb_dst_drop(skb);
1074                 skb_bpf_redirect_clear(skb);
1075                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1076                 if (ret == SK_PASS)
1077                         skb_bpf_set_strparser(skb);
1078                 ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb));
1079                 skb->sk = NULL;
1080         }
1081         sk_psock_verdict_apply(psock, skb, ret);
1082 out:
1083         rcu_read_unlock();
1084 }
1085
1086 static int sk_psock_strp_read_done(struct strparser *strp, int err)
1087 {
1088         return err;
1089 }
1090
1091 static int sk_psock_strp_parse(struct strparser *strp, struct sk_buff *skb)
1092 {
1093         struct sk_psock *psock = container_of(strp, struct sk_psock, strp);
1094         struct bpf_prog *prog;
1095         int ret = skb->len;
1096
1097         rcu_read_lock();
1098         prog = READ_ONCE(psock->progs.stream_parser);
1099         if (likely(prog)) {
1100                 skb->sk = psock->sk;
1101                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1102                 skb->sk = NULL;
1103         }
1104         rcu_read_unlock();
1105         return ret;
1106 }
1107
1108 /* Called with socket lock held. */
1109 static void sk_psock_strp_data_ready(struct sock *sk)
1110 {
1111         struct sk_psock *psock;
1112
1113         rcu_read_lock();
1114         psock = sk_psock(sk);
1115         if (likely(psock)) {
1116                 if (tls_sw_has_ctx_rx(sk)) {
1117                         psock->saved_data_ready(sk);
1118                 } else {
1119                         write_lock_bh(&sk->sk_callback_lock);
1120                         strp_data_ready(&psock->strp);
1121                         write_unlock_bh(&sk->sk_callback_lock);
1122                 }
1123         }
1124         rcu_read_unlock();
1125 }
1126
1127 int sk_psock_init_strp(struct sock *sk, struct sk_psock *psock)
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         return strp_init(&psock->strp, sk, &cb);
1136 }
1137
1138 void sk_psock_start_strp(struct sock *sk, struct sk_psock *psock)
1139 {
1140         if (psock->saved_data_ready)
1141                 return;
1142
1143         psock->saved_data_ready = sk->sk_data_ready;
1144         sk->sk_data_ready = sk_psock_strp_data_ready;
1145         sk->sk_write_space = sk_psock_write_space;
1146 }
1147
1148 void sk_psock_stop_strp(struct sock *sk, struct sk_psock *psock)
1149 {
1150         psock_set_prog(&psock->progs.stream_parser, NULL);
1151
1152         if (!psock->saved_data_ready)
1153                 return;
1154
1155         sk->sk_data_ready = psock->saved_data_ready;
1156         psock->saved_data_ready = NULL;
1157         strp_stop(&psock->strp);
1158 }
1159
1160 static void sk_psock_done_strp(struct sk_psock *psock)
1161 {
1162         /* Parser has been stopped */
1163         if (psock->progs.stream_parser)
1164                 strp_done(&psock->strp);
1165 }
1166 #else
1167 static void sk_psock_done_strp(struct sk_psock *psock)
1168 {
1169 }
1170 #endif /* CONFIG_BPF_STREAM_PARSER */
1171
1172 static int sk_psock_verdict_recv(struct sock *sk, struct sk_buff *skb)
1173 {
1174         struct sk_psock *psock;
1175         struct bpf_prog *prog;
1176         int ret = __SK_DROP;
1177         int len = skb->len;
1178
1179         skb_get(skb);
1180
1181         rcu_read_lock();
1182         psock = sk_psock(sk);
1183         if (unlikely(!psock)) {
1184                 len = 0;
1185                 sock_drop(sk, skb);
1186                 goto out;
1187         }
1188         prog = READ_ONCE(psock->progs.stream_verdict);
1189         if (!prog)
1190                 prog = READ_ONCE(psock->progs.skb_verdict);
1191         if (likely(prog)) {
1192                 skb_dst_drop(skb);
1193                 skb_bpf_redirect_clear(skb);
1194                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1195                 ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb));
1196         }
1197         ret = sk_psock_verdict_apply(psock, skb, ret);
1198         if (ret < 0)
1199                 len = ret;
1200 out:
1201         rcu_read_unlock();
1202         return len;
1203 }
1204
1205 static void sk_psock_verdict_data_ready(struct sock *sk)
1206 {
1207         struct socket *sock = sk->sk_socket;
1208
1209         if (unlikely(!sock || !sock->ops || !sock->ops->read_skb))
1210                 return;
1211         sock->ops->read_skb(sk, sk_psock_verdict_recv);
1212 }
1213
1214 void sk_psock_start_verdict(struct sock *sk, struct sk_psock *psock)
1215 {
1216         if (psock->saved_data_ready)
1217                 return;
1218
1219         psock->saved_data_ready = sk->sk_data_ready;
1220         sk->sk_data_ready = sk_psock_verdict_data_ready;
1221         sk->sk_write_space = sk_psock_write_space;
1222 }
1223
1224 void sk_psock_stop_verdict(struct sock *sk, struct sk_psock *psock)
1225 {
1226         psock_set_prog(&psock->progs.stream_verdict, NULL);
1227         psock_set_prog(&psock->progs.skb_verdict, NULL);
1228
1229         if (!psock->saved_data_ready)
1230                 return;
1231
1232         sk->sk_data_ready = psock->saved_data_ready;
1233         psock->saved_data_ready = NULL;
1234 }