net/core/skmsg.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* Copyright (c) 2017 - 2018 Covalent IO, Inc. http://covalent.io */
   3
   4 #include <linux/skmsg.h>
   5 #include <linux/skbuff.h>
   6 #include <linux/scatterlist.h>
   7
   8 #include <net/sock.h>
   9 #include <net/tcp.h>
  10 #include <net/tls.h>
  11
  12 static bool sk_msg_try_coalesce_ok(struct sk_msg *msg, int elem_first_coalesce)
  13 {
  14         if (msg->sg.end > msg->sg.start &&
  15             elem_first_coalesce < msg->sg.end)
  16                 return true;
  17
  18         if (msg->sg.end < msg->sg.start &&
  19             (elem_first_coalesce > msg->sg.start ||
  20              elem_first_coalesce < msg->sg.end))
  21                 return true;
  22
  23         return false;
  24 }
  25
  26 int sk_msg_alloc(struct sock *sk, struct sk_msg *msg, int len,
  27                  int elem_first_coalesce)
  28 {
  29         struct page_frag *pfrag = sk_page_frag(sk);
  30         u32 osize = msg->sg.size;
  31         int ret = 0;
  32
  33         len -= msg->sg.size;
  34         while (len > 0) {
  35                 struct scatterlist *sge;
  36                 u32 orig_offset;
  37                 int use, i;
  38
  39                 if (!sk_page_frag_refill(sk, pfrag)) {
  40                         ret = -ENOMEM;
  41                         goto msg_trim;
  42                 }
  43
  44                 orig_offset = pfrag->offset;
  45                 use = min_t(int, len, pfrag->size - orig_offset);
  46                 if (!sk_wmem_schedule(sk, use)) {
  47                         ret = -ENOMEM;
  48                         goto msg_trim;
  49                 }
  50
  51                 i = msg->sg.end;
  52                 sk_msg_iter_var_prev(i);
  53                 sge = &msg->sg.data[i];
  54
  55                 if (sk_msg_try_coalesce_ok(msg, elem_first_coalesce) &&
  56                     sg_page(sge) == pfrag->page &&
  57                     sge->offset + sge->length == orig_offset) {
  58                         sge->length += use;
  59                 } else {
  60                         if (sk_msg_full(msg)) {
  61                                 ret = -ENOSPC;
  62                                 break;
  63                         }
  64
  65                         sge = &msg->sg.data[msg->sg.end];
  66                         sg_unmark_end(sge);
  67                         sg_set_page(sge, pfrag->page, use, orig_offset);
  68                         get_page(pfrag->page);
  69                         sk_msg_iter_next(msg, end);
  70                 }
  71
  72                 sk_mem_charge(sk, use);
  73                 msg->sg.size += use;
  74                 pfrag->offset += use;
  75                 len -= use;
  76         }
  77
  78         return ret;
  79
  80 msg_trim:
  81         sk_msg_trim(sk, msg, osize);
  82         return ret;
  83 }
  84 EXPORT_SYMBOL_GPL(sk_msg_alloc);
  85
  86 int sk_msg_clone(struct sock *sk, struct sk_msg *dst, struct sk_msg *src,
  87                  u32 off, u32 len)
  88 {
  89         int i = src->sg.start;
  90         struct scatterlist *sge = sk_msg_elem(src, i);
  91         struct scatterlist *sgd = NULL;
  92         u32 sge_len, sge_off;
  93
  94         while (off) {
  95                 if (sge->length > off)
  96                         break;
  97                 off -= sge->length;
  98                 sk_msg_iter_var_next(i);
  99                 if (i == src->sg.end && off)
 100                         return -ENOSPC;
 101                 sge = sk_msg_elem(src, i);
 102         }
 103
 104         while (len) {
 105                 sge_len = sge->length - off;
 106                 if (sge_len > len)
 107                         sge_len = len;
 108
 109                 if (dst->sg.end)
 110                         sgd = sk_msg_elem(dst, dst->sg.end - 1);
 111
 112                 if (sgd &&
 113                     (sg_page(sge) == sg_page(sgd)) &&
 114                     (sg_virt(sge) + off == sg_virt(sgd) + sgd->length)) {
 115                         sgd->length += sge_len;
 116                         dst->sg.size += sge_len;
 117                 } else if (!sk_msg_full(dst)) {
 118                         sge_off = sge->offset + off;
 119                         sk_msg_page_add(dst, sg_page(sge), sge_len, sge_off);
 120                 } else {
 121                         return -ENOSPC;
 122                 }
 123
 124                 off = 0;
 125                 len -= sge_len;
 126                 sk_mem_charge(sk, sge_len);
 127                 sk_msg_iter_var_next(i);
 128                 if (i == src->sg.end && len)
 129                         return -ENOSPC;
 130                 sge = sk_msg_elem(src, i);
 131         }
 132
 133         return 0;
 134 }
 135 EXPORT_SYMBOL_GPL(sk_msg_clone);
 136
 137 void sk_msg_return_zero(struct sock *sk, struct sk_msg *msg, int bytes)
 138 {
 139         int i = msg->sg.start;
 140
 141         do {
 142                 struct scatterlist *sge = sk_msg_elem(msg, i);
 143
 144                 if (bytes < sge->length) {
 145                         sge->length -= bytes;
 146                         sge->offset += bytes;
 147                         sk_mem_uncharge(sk, bytes);
 148                         break;
 149                 }
 150
 151                 sk_mem_uncharge(sk, sge->length);
 152                 bytes -= sge->length;
 153                 sge->length = 0;
 154                 sge->offset = 0;
 155                 sk_msg_iter_var_next(i);
 156         } while (bytes && i != msg->sg.end);
 157         msg->sg.start = i;
 158 }
 159 EXPORT_SYMBOL_GPL(sk_msg_return_zero);
 160
 161 void sk_msg_return(struct sock *sk, struct sk_msg *msg, int bytes)
 162 {
 163         int i = msg->sg.start;
 164
 165         do {
 166                 struct scatterlist *sge = &msg->sg.data[i];
 167                 int uncharge = (bytes < sge->length) ? bytes : sge->length;
 168
 169                 sk_mem_uncharge(sk, uncharge);
 170                 bytes -= uncharge;
 171                 sk_msg_iter_var_next(i);
 172         } while (i != msg->sg.end);
 173 }
 174 EXPORT_SYMBOL_GPL(sk_msg_return);
 175
 176 static int sk_msg_free_elem(struct sock *sk, struct sk_msg *msg, u32 i,
 177                             bool charge)
 178 {
 179         struct scatterlist *sge = sk_msg_elem(msg, i);
 180         u32 len = sge->length;
 181
 182         /* When the skb owns the memory we free it from consume_skb path. */
 183         if (!msg->skb) {
 184                 if (charge)
 185                         sk_mem_uncharge(sk, len);
 186                 put_page(sg_page(sge));
 187         }
 188         memset(sge, 0, sizeof(*sge));
 189         return len;
 190 }
 191
 192 static int __sk_msg_free(struct sock *sk, struct sk_msg *msg, u32 i,
 193                          bool charge)
 194 {
 195         struct scatterlist *sge = sk_msg_elem(msg, i);
 196         int freed = 0;
 197
 198         while (msg->sg.size) {
 199                 msg->sg.size -= sge->length;
 200                 freed += sk_msg_free_elem(sk, msg, i, charge);
 201                 sk_msg_iter_var_next(i);
 202                 sk_msg_check_to_free(msg, i, msg->sg.size);
 203                 sge = sk_msg_elem(msg, i);
 204         }
 205         consume_skb(msg->skb);
 206         sk_msg_init(msg);
 207         return freed;
 208 }
 209
 210 int sk_msg_free_nocharge(struct sock *sk, struct sk_msg *msg)
 211 {
 212         return __sk_msg_free(sk, msg, msg->sg.start, false);
 213 }
 214 EXPORT_SYMBOL_GPL(sk_msg_free_nocharge);
 215
 216 int sk_msg_free(struct sock *sk, struct sk_msg *msg)
 217 {
 218         return __sk_msg_free(sk, msg, msg->sg.start, true);
 219 }
 220 EXPORT_SYMBOL_GPL(sk_msg_free);
 221
 222 static void __sk_msg_free_partial(struct sock *sk, struct sk_msg *msg,
 223                                   u32 bytes, bool charge)
 224 {
 225         struct scatterlist *sge;
 226         u32 i = msg->sg.start;
 227
 228         while (bytes) {
 229                 sge = sk_msg_elem(msg, i);
 230                 if (!sge->length)
 231                         break;
 232                 if (bytes < sge->length) {
 233                         if (charge)
 234                                 sk_mem_uncharge(sk, bytes);
 235                         sge->length -= bytes;
 236                         sge->offset += bytes;
 237                         msg->sg.size -= bytes;
 238                         break;
 239                 }
 240
 241                 msg->sg.size -= sge->length;
 242                 bytes -= sge->length;
 243                 sk_msg_free_elem(sk, msg, i, charge);
 244                 sk_msg_iter_var_next(i);
 245                 sk_msg_check_to_free(msg, i, bytes);
 246         }
 247         msg->sg.start = i;
 248 }
 249
 250 void sk_msg_free_partial(struct sock *sk, struct sk_msg *msg, u32 bytes)
 251 {
 252         __sk_msg_free_partial(sk, msg, bytes, true);
 253 }
 254 EXPORT_SYMBOL_GPL(sk_msg_free_partial);
 255
 256 void sk_msg_free_partial_nocharge(struct sock *sk, struct sk_msg *msg,
 257                                   u32 bytes)
 258 {
 259         __sk_msg_free_partial(sk, msg, bytes, false);
 260 }
 261
 262 void sk_msg_trim(struct sock *sk, struct sk_msg *msg, int len)
 263 {
 264         int trim = msg->sg.size - len;
 265         u32 i = msg->sg.end;
 266
 267         if (trim <= 0) {
 268                 WARN_ON(trim < 0);
 269                 return;
 270         }
 271
 272         sk_msg_iter_var_prev(i);
 273         msg->sg.size = len;
 274         while (msg->sg.data[i].length &&
 275                trim >= msg->sg.data[i].length) {
 276                 trim -= msg->sg.data[i].length;
 277                 sk_msg_free_elem(sk, msg, i, true);
 278                 sk_msg_iter_var_prev(i);
 279                 if (!trim)
 280                         goto out;
 281         }
 282
 283         msg->sg.data[i].length -= trim;
 284         sk_mem_uncharge(sk, trim);
 285         /* Adjust copybreak if it falls into the trimmed part of last buf */
 286         if (msg->sg.curr == i && msg->sg.copybreak > msg->sg.data[i].length)
 287                 msg->sg.copybreak = msg->sg.data[i].length;
 288 out:
 289         sk_msg_iter_var_next(i);
 290         msg->sg.end = i;
 291
 292         /* If we trim data a full sg elem before curr pointer update
 293          * copybreak and current so that any future copy operations
 294          * start at new copy location.
 295          * However trimed data that has not yet been used in a copy op
 296          * does not require an update.
 297          */
 298         if (!msg->sg.size) {
 299                 msg->sg.curr = msg->sg.start;
 300                 msg->sg.copybreak = 0;
 301         } else if (sk_msg_iter_dist(msg->sg.start, msg->sg.curr) >=
 302                    sk_msg_iter_dist(msg->sg.start, msg->sg.end)) {
 303                 sk_msg_iter_var_prev(i);
 304                 msg->sg.curr = i;
 305                 msg->sg.copybreak = msg->sg.data[i].length;
 306         }
 307 }
 308 EXPORT_SYMBOL_GPL(sk_msg_trim);
 309
 310 int sk_msg_zerocopy_from_iter(struct sock *sk, struct iov_iter *from,
 311                               struct sk_msg *msg, u32 bytes)
 312 {
 313         int i, maxpages, ret = 0, num_elems = sk_msg_elem_used(msg);
 314         const int to_max_pages = MAX_MSG_FRAGS;
 315         struct page *pages[MAX_MSG_FRAGS];
 316         ssize_t orig, copied, use, offset;
 317
 318         orig = msg->sg.size;
 319         while (bytes > 0) {
 320                 i = 0;
 321                 maxpages = to_max_pages - num_elems;
 322                 if (maxpages == 0) {
 323                         ret = -EFAULT;
 324                         goto out;
 325                 }
 326
 327                 copied = iov_iter_get_pages(from, pages, bytes, maxpages,
 328                                             &offset);
 329                 if (copied <= 0) {
 330                         ret = -EFAULT;
 331                         goto out;
 332                 }
 333
 334                 iov_iter_advance(from, copied);
 335                 bytes -= copied;
 336                 msg->sg.size += copied;
 337
 338                 while (copied) {
 339                         use = min_t(int, copied, PAGE_SIZE - offset);
 340                         sg_set_page(&msg->sg.data[msg->sg.end],
 341                                     pages[i], use, offset);
 342                         sg_unmark_end(&msg->sg.data[msg->sg.end]);
 343                         sk_mem_charge(sk, use);
 344
 345                         offset = 0;
 346                         copied -= use;
 347                         sk_msg_iter_next(msg, end);
 348                         num_elems++;
 349                         i++;
 350                 }
 351                 /* When zerocopy is mixed with sk_msg_*copy* operations we
 352                  * may have a copybreak set in this case clear and prefer
 353                  * zerocopy remainder when possible.
 354                  */
 355                 msg->sg.copybreak = 0;
 356                 msg->sg.curr = msg->sg.end;
 357         }
 358 out:
 359         /* Revert iov_iter updates, msg will need to use 'trim' later if it
 360          * also needs to be cleared.
 361          */
 362         if (ret)
 363                 iov_iter_revert(from, msg->sg.size - orig);
 364         return ret;
 365 }
 366 EXPORT_SYMBOL_GPL(sk_msg_zerocopy_from_iter);
 367
 368 int sk_msg_memcopy_from_iter(struct sock *sk, struct iov_iter *from,
 369                              struct sk_msg *msg, u32 bytes)
 370 {
 371         int ret = -ENOSPC, i = msg->sg.curr;
 372         struct scatterlist *sge;
 373         u32 copy, buf_size;
 374         void *to;
 375
 376         do {
 377                 sge = sk_msg_elem(msg, i);
 378                 /* This is possible if a trim operation shrunk the buffer */
 379                 if (msg->sg.copybreak >= sge->length) {
 380                         msg->sg.copybreak = 0;
 381                         sk_msg_iter_var_next(i);
 382                         if (i == msg->sg.end)
 383                                 break;
 384                         sge = sk_msg_elem(msg, i);
 385                 }
 386
 387                 buf_size = sge->length - msg->sg.copybreak;
 388                 copy = (buf_size > bytes) ? bytes : buf_size;
 389                 to = sg_virt(sge) + msg->sg.copybreak;
 390                 msg->sg.copybreak += copy;
 391                 if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY)
 392                         ret = copy_from_iter_nocache(to, copy, from);
 393                 else
 394                         ret = copy_from_iter(to, copy, from);
 395                 if (ret != copy) {
 396                         ret = -EFAULT;
 397                         goto out;
 398                 }
 399                 bytes -= copy;
 400                 if (!bytes)
 401                         break;
 402                 msg->sg.copybreak = 0;
 403                 sk_msg_iter_var_next(i);
 404         } while (i != msg->sg.end);
 405 out:
 406         msg->sg.curr = i;
 407         return ret;
 408 }
 409 EXPORT_SYMBOL_GPL(sk_msg_memcopy_from_iter);
 410
 411 /* Receive sk_msg from psock->ingress_msg to @msg. */
 412 int sk_msg_recvmsg(struct sock *sk, struct sk_psock *psock, struct msghdr *msg,
 413                    int len, int flags)
 414 {
 415         struct iov_iter *iter = &msg->msg_iter;
 416         int peek = flags & MSG_PEEK;
 417         struct sk_msg *msg_rx;
 418         int i, copied = 0;
 419
 420         msg_rx = sk_psock_peek_msg(psock);
 421         while (copied != len) {
 422                 struct scatterlist *sge;
 423
 424                 if (unlikely(!msg_rx))
 425                         break;
 426
 427                 i = msg_rx->sg.start;
 428                 do {
 429                         struct page *page;
 430                         int copy;
 431
 432                         sge = sk_msg_elem(msg_rx, i);
 433                         copy = sge->length;
 434                         page = sg_page(sge);
 435                         if (copied + copy > len)
 436                                 copy = len - copied;
 437                         copy = copy_page_to_iter(page, sge->offset, copy, iter);
 438                         if (!copy)
 439                                 return copied ? copied : -EFAULT;
 440
 441                         copied += copy;
 442                         if (likely(!peek)) {
 443                                 sge->offset += copy;
 444                                 sge->length -= copy;
 445                                 if (!msg_rx->skb)
 446                                         sk_mem_uncharge(sk, copy);
 447                                 msg_rx->sg.size -= copy;
 448
 449                                 if (!sge->length) {
 450                                         sk_msg_iter_var_next(i);
 451                                         if (!msg_rx->skb)
 452                                                 put_page(page);
 453                                 }
 454                         } else {
 455                                 /* Lets not optimize peek case if copy_page_to_iter
 456                                  * didn't copy the entire length lets just break.
 457                                  */
 458                                 if (copy != sge->length)
 459                                         return copied;
 460                                 sk_msg_iter_var_next(i);
 461                         }
 462
 463                         if (copied == len)
 464                                 break;
 465                 } while (!sg_is_last(sge));
 466
 467                 if (unlikely(peek)) {
 468                         msg_rx = sk_psock_next_msg(psock, msg_rx);
 469                         if (!msg_rx)
 470                                 break;
 471                         continue;
 472                 }
 473
 474                 msg_rx->sg.start = i;
 475                 if (!sge->length && sg_is_last(sge)) {
 476                         msg_rx = sk_psock_dequeue_msg(psock);
 477                         kfree_sk_msg(msg_rx);
 478                 }
 479                 msg_rx = sk_psock_peek_msg(psock);
 480         }
 481
 482         return copied;
 483 }
 484 EXPORT_SYMBOL_GPL(sk_msg_recvmsg);
 485
 486 bool sk_msg_is_readable(struct sock *sk)
 487 {
 488         struct sk_psock *psock;
 489         bool empty = true;
 490
 491         rcu_read_lock();
 492         psock = sk_psock(sk);
 493         if (likely(psock))
 494                 empty = list_empty(&psock->ingress_msg);
 495         rcu_read_unlock();
 496         return !empty;
 497 }
 498 EXPORT_SYMBOL_GPL(sk_msg_is_readable);
 499
 500 static struct sk_msg *sk_psock_create_ingress_msg(struct sock *sk,
 501                                                   struct sk_buff *skb)
 502 {
 503         struct sk_msg *msg;
 504
 505         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
 506                 return NULL;
 507
 508         if (!sk_rmem_schedule(sk, skb, skb->truesize))
 509                 return NULL;
 510
 511         msg = kzalloc(sizeof(*msg), __GFP_NOWARN | GFP_KERNEL);
 512         if (unlikely(!msg))
 513                 return NULL;
 514
 515         sk_msg_init(msg);
 516         return msg;
 517 }
 518
 519 static int sk_psock_skb_ingress_enqueue(struct sk_buff *skb,
 520                                         u32 off, u32 len,
 521                                         struct sk_psock *psock,
 522                                         struct sock *sk,
 523                                         struct sk_msg *msg)
 524 {
 525         int num_sge, copied;
 526
 527         /* skb linearize may fail with ENOMEM, but lets simply try again
 528          * later if this happens. Under memory pressure we don't want to
 529          * drop the skb. We need to linearize the skb so that the mapping
 530          * in skb_to_sgvec can not error.
 531          */
 532         if (skb_linearize(skb))
 533                 return -EAGAIN;
 534         num_sge = skb_to_sgvec(skb, msg->sg.data, off, len);
 535         if (unlikely(num_sge < 0))
 536                 return num_sge;
 537
 538         copied = len;
 539         msg->sg.start = 0;
 540         msg->sg.size = copied;
 541         msg->sg.end = num_sge;
 542         msg->skb = skb;
 543
 544         sk_psock_queue_msg(psock, msg);
 545         sk_psock_data_ready(sk, psock);
 546         return copied;
 547 }
 548
 549 static int sk_psock_skb_ingress_self(struct sk_psock *psock, struct sk_buff *skb,
 550                                      u32 off, u32 len);
 551
 552 static int sk_psock_skb_ingress(struct sk_psock *psock, struct sk_buff *skb,
 553                                 u32 off, u32 len)
 554 {
 555         struct sock *sk = psock->sk;
 556         struct sk_msg *msg;
 557         int err;
 558
 559         /* If we are receiving on the same sock skb->sk is already assigned,
 560          * skip memory accounting and owner transition seeing it already set
 561          * correctly.
 562          */
 563         if (unlikely(skb->sk == sk))
 564                 return sk_psock_skb_ingress_self(psock, skb, off, len);
 565         msg = sk_psock_create_ingress_msg(sk, skb);
 566         if (!msg)
 567                 return -EAGAIN;
 568
 569         /* This will transition ownership of the data from the socket where
 570          * the BPF program was run initiating the redirect to the socket
 571          * we will eventually receive this data on. The data will be released
 572          * from skb_consume found in __tcp_bpf_recvmsg() after its been copied
 573          * into user buffers.
 574          */
 575         skb_set_owner_r(skb, sk);
 576         err = sk_psock_skb_ingress_enqueue(skb, off, len, psock, sk, msg);
 577         if (err < 0)
 578                 kfree(msg);
 579         return err;
 580 }
 581
 582 /* Puts an skb on the ingress queue of the socket already assigned to the
 583  * skb. In this case we do not need to check memory limits or skb_set_owner_r
 584  * because the skb is already accounted for here.
 585  */
 586 static int sk_psock_skb_ingress_self(struct sk_psock *psock, struct sk_buff *skb,
 587                                      u32 off, u32 len)
 588 {
 589         struct sk_msg *msg = kzalloc(sizeof(*msg), __GFP_NOWARN | GFP_ATOMIC);
 590         struct sock *sk = psock->sk;
 591         int err;
 592
 593         if (unlikely(!msg))
 594                 return -EAGAIN;
 595         sk_msg_init(msg);
 596         skb_set_owner_r(skb, sk);
 597         err = sk_psock_skb_ingress_enqueue(skb, off, len, psock, sk, msg);
 598         if (err < 0)
 599                 kfree(msg);
 600         return err;
 601 }
 602
 603 static int sk_psock_handle_skb(struct sk_psock *psock, struct sk_buff *skb,
 604                                u32 off, u32 len, bool ingress)
 605 {
 606         if (!ingress) {
 607                 if (!sock_writeable(psock->sk))
 608                         return -EAGAIN;
 609                 return skb_send_sock(psock->sk, skb, off, len);
 610         }
 611         return sk_psock_skb_ingress(psock, skb, off, len);
 612 }
 613
 614 static void sk_psock_skb_state(struct sk_psock *psock,
 615                                struct sk_psock_work_state *state,
 616                                struct sk_buff *skb,
 617                                int len, int off)
 618 {
 619         spin_lock_bh(&psock->ingress_lock);
 620         if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
 621                 state->skb = skb;
 622                 state->len = len;
 623                 state->off = off;
 624         } else {
 625                 sock_drop(psock->sk, skb);
 626         }
 627         spin_unlock_bh(&psock->ingress_lock);
 628 }
 629
 630 static void sk_psock_backlog(struct work_struct *work)
 631 {
 632         struct sk_psock *psock = container_of(work, struct sk_psock, work);
 633         struct sk_psock_work_state *state = &psock->work_state;
 634         struct sk_buff *skb = NULL;
 635         bool ingress;
 636         u32 len, off;
 637         int ret;
 638
 639         mutex_lock(&psock->work_mutex);
 640         if (unlikely(state->skb)) {
 641                 spin_lock_bh(&psock->ingress_lock);
 642                 skb = state->skb;
 643                 len = state->len;
 644                 off = state->off;
 645                 state->skb = NULL;
 646                 spin_unlock_bh(&psock->ingress_lock);
 647         }
 648         if (skb)
 649                 goto start;
 650
 651         while ((skb = skb_dequeue(&psock->ingress_skb))) {
 652                 len = skb->len;
 653                 off = 0;
 654                 if (skb_bpf_strparser(skb)) {
 655                         struct strp_msg *stm = strp_msg(skb);
 656
 657                         off = stm->offset;
 658                         len = stm->full_len;
 659                 }
 660 start:
 661                 ingress = skb_bpf_ingress(skb);
 662                 skb_bpf_redirect_clear(skb);
 663                 do {
 664                         ret = -EIO;
 665                         if (!sock_flag(psock->sk, SOCK_DEAD))
 666                                 ret = sk_psock_handle_skb(psock, skb, off,
 667                                                           len, ingress);
 668                         if (ret <= 0) {
 669                                 if (ret == -EAGAIN) {
 670                                         sk_psock_skb_state(psock, state, skb,
 671                                                            len, off);
 672                                         goto end;
 673                                 }
 674                                 /* Hard errors break pipe and stop xmit. */
 675                                 sk_psock_report_error(psock, ret ? -ret : EPIPE);
 676                                 sk_psock_clear_state(psock, SK_PSOCK_TX_ENABLED);
 677                                 sock_drop(psock->sk, skb);
 678                                 goto end;
 679                         }
 680                         off += ret;
 681                         len -= ret;
 682                 } while (len);
 683
 684                 if (!ingress)
 685                         kfree_skb(skb);
 686         }
 687 end:
 688         mutex_unlock(&psock->work_mutex);
 689 }
 690
 691 struct sk_psock *sk_psock_init(struct sock *sk, int node)
 692 {
 693         struct sk_psock *psock;
 694         struct proto *prot;
 695
 696         write_lock_bh(&sk->sk_callback_lock);
 697
 698         if (sk_is_inet(sk) && inet_csk_has_ulp(sk)) {
 699                 psock = ERR_PTR(-EINVAL);
 700                 goto out;
 701         }
 702
 703         if (sk->sk_user_data) {
 704                 psock = ERR_PTR(-EBUSY);
 705                 goto out;
 706         }
 707
 708         psock = kzalloc_node(sizeof(*psock), GFP_ATOMIC | __GFP_NOWARN, node);
 709         if (!psock) {
 710                 psock = ERR_PTR(-ENOMEM);
 711                 goto out;
 712         }
 713
 714         prot = READ_ONCE(sk->sk_prot);
 715         psock->sk = sk;
 716         psock->eval = __SK_NONE;
 717         psock->sk_proto = prot;
 718         psock->saved_unhash = prot->unhash;
 719         psock->saved_close = prot->close;
 720         psock->saved_write_space = sk->sk_write_space;
 721
 722         INIT_LIST_HEAD(&psock->link);
 723         spin_lock_init(&psock->link_lock);
 724
 725         INIT_WORK(&psock->work, sk_psock_backlog);
 726         mutex_init(&psock->work_mutex);
 727         INIT_LIST_HEAD(&psock->ingress_msg);
 728         spin_lock_init(&psock->ingress_lock);
 729         skb_queue_head_init(&psock->ingress_skb);
 730
 731         sk_psock_set_state(psock, SK_PSOCK_TX_ENABLED);
 732         refcount_set(&psock->refcnt, 1);
 733
 734         __rcu_assign_sk_user_data_with_flags(sk, psock,
 735                                              SK_USER_DATA_NOCOPY |
 736                                              SK_USER_DATA_PSOCK);
 737         sock_hold(sk);
 738
 739 out:
 740         write_unlock_bh(&sk->sk_callback_lock);
 741         return psock;
 742 }
 743 EXPORT_SYMBOL_GPL(sk_psock_init);
 744
 745 struct sk_psock_link *sk_psock_link_pop(struct sk_psock *psock)
 746 {
 747         struct sk_psock_link *link;
 748
 749         spin_lock_bh(&psock->link_lock);
 750         link = list_first_entry_or_null(&psock->link, struct sk_psock_link,
 751                                         list);
 752         if (link)
 753                 list_del(&link->list);
 754         spin_unlock_bh(&psock->link_lock);
 755         return link;
 756 }
 757
 758 static void __sk_psock_purge_ingress_msg(struct sk_psock *psock)
 759 {
 760         struct sk_msg *msg, *tmp;
 761
 762         list_for_each_entry_safe(msg, tmp, &psock->ingress_msg, list) {
 763                 list_del(&msg->list);
 764                 sk_msg_free(psock->sk, msg);
 765                 kfree(msg);
 766         }
 767 }
 768
 769 static void __sk_psock_zap_ingress(struct sk_psock *psock)
 770 {
 771         struct sk_buff *skb;
 772
 773         while ((skb = skb_dequeue(&psock->ingress_skb)) != NULL) {
 774                 skb_bpf_redirect_clear(skb);
 775                 sock_drop(psock->sk, skb);
 776         }
 777         kfree_skb(psock->work_state.skb);
 778         /* We null the skb here to ensure that calls to sk_psock_backlog
 779          * do not pick up the free'd skb.
 780          */
 781         psock->work_state.skb = NULL;
 782         __sk_psock_purge_ingress_msg(psock);
 783 }
 784
 785 static void sk_psock_link_destroy(struct sk_psock *psock)
 786 {
 787         struct sk_psock_link *link, *tmp;
 788
 789         list_for_each_entry_safe(link, tmp, &psock->link, list) {
 790                 list_del(&link->list);
 791                 sk_psock_free_link(link);
 792         }
 793 }
 794
 795 void sk_psock_stop(struct sk_psock *psock, bool wait)
 796 {
 797         spin_lock_bh(&psock->ingress_lock);
 798         sk_psock_clear_state(psock, SK_PSOCK_TX_ENABLED);
 799         sk_psock_cork_free(psock);
 800         __sk_psock_zap_ingress(psock);
 801         spin_unlock_bh(&psock->ingress_lock);
 802
 803         if (wait)
 804                 cancel_work_sync(&psock->work);
 805 }
 806
 807 static void sk_psock_done_strp(struct sk_psock *psock);
 808
 809 static void sk_psock_destroy(struct work_struct *work)
 810 {
 811         struct sk_psock *psock = container_of(to_rcu_work(work),
 812                                               struct sk_psock, rwork);
 813         /* No sk_callback_lock since already detached. */
 814
 815         sk_psock_done_strp(psock);
 816
 817         cancel_work_sync(&psock->work);
 818         mutex_destroy(&psock->work_mutex);
 819
 820         psock_progs_drop(&psock->progs);
 821
 822         sk_psock_link_destroy(psock);
 823         sk_psock_cork_free(psock);
 824
 825         if (psock->sk_redir)
 826                 sock_put(psock->sk_redir);
 827         sock_put(psock->sk);
 828         kfree(psock);
 829 }
 830
 831 void sk_psock_drop(struct sock *sk, struct sk_psock *psock)
 832 {
 833         write_lock_bh(&sk->sk_callback_lock);
 834         sk_psock_restore_proto(sk, psock);
 835         rcu_assign_sk_user_data(sk, NULL);
 836         if (psock->progs.stream_parser)
 837                 sk_psock_stop_strp(sk, psock);
 838         else if (psock->progs.stream_verdict || psock->progs.skb_verdict)
 839                 sk_psock_stop_verdict(sk, psock);
 840         write_unlock_bh(&sk->sk_callback_lock);
 841
 842         sk_psock_stop(psock, false);
 843
 844         INIT_RCU_WORK(&psock->rwork, sk_psock_destroy);
 845         queue_rcu_work(system_wq, &psock->rwork);
 846 }
 847 EXPORT_SYMBOL_GPL(sk_psock_drop);
 848
 849 static int sk_psock_map_verd(int verdict, bool redir)
 850 {
 851         switch (verdict) {
 852         case SK_PASS:
 853                 return redir ? __SK_REDIRECT : __SK_PASS;
 854         case SK_DROP:
 855         default:
 856                 break;
 857         }
 858
 859         return __SK_DROP;
 860 }
 861
 862 int sk_psock_msg_verdict(struct sock *sk, struct sk_psock *psock,
 863                          struct sk_msg *msg)
 864 {
 865         struct bpf_prog *prog;
 866         int ret;
 867
 868         rcu_read_lock();
 869         prog = READ_ONCE(psock->progs.msg_parser);
 870         if (unlikely(!prog)) {
 871                 ret = __SK_PASS;
 872                 goto out;
 873         }
 874
 875         sk_msg_compute_data_pointers(msg);
 876         msg->sk = sk;
 877         ret = bpf_prog_run_pin_on_cpu(prog, msg);
 878         ret = sk_psock_map_verd(ret, msg->sk_redir);
 879         psock->apply_bytes = msg->apply_bytes;
 880         if (ret == __SK_REDIRECT) {
 881                 if (psock->sk_redir)
 882                         sock_put(psock->sk_redir);
 883                 psock->sk_redir = msg->sk_redir;
 884                 if (!psock->sk_redir) {
 885                         ret = __SK_DROP;
 886                         goto out;
 887                 }
 888                 sock_hold(psock->sk_redir);
 889         }
 890 out:
 891         rcu_read_unlock();
 892         return ret;
 893 }
 894 EXPORT_SYMBOL_GPL(sk_psock_msg_verdict);
 895
 896 static int sk_psock_skb_redirect(struct sk_psock *from, struct sk_buff *skb)
 897 {
 898         struct sk_psock *psock_other;
 899         struct sock *sk_other;
 900
 901         sk_other = skb_bpf_redirect_fetch(skb);
 902         /* This error is a buggy BPF program, it returned a redirect
 903          * return code, but then didn't set a redirect interface.
 904          */
 905         if (unlikely(!sk_other)) {
 906                 skb_bpf_redirect_clear(skb);
 907                 sock_drop(from->sk, skb);
 908                 return -EIO;
 909         }
 910         psock_other = sk_psock(sk_other);
 911         /* This error indicates the socket is being torn down or had another
 912          * error that caused the pipe to break. We can't send a packet on
 913          * a socket that is in this state so we drop the skb.
 914          */
 915         if (!psock_other || sock_flag(sk_other, SOCK_DEAD)) {
 916                 skb_bpf_redirect_clear(skb);
 917                 sock_drop(from->sk, skb);
 918                 return -EIO;
 919         }
 920         spin_lock_bh(&psock_other->ingress_lock);
 921         if (!sk_psock_test_state(psock_other, SK_PSOCK_TX_ENABLED)) {
 922                 spin_unlock_bh(&psock_other->ingress_lock);
 923                 skb_bpf_redirect_clear(skb);
 924                 sock_drop(from->sk, skb);
 925                 return -EIO;
 926         }
 927
 928         skb_queue_tail(&psock_other->ingress_skb, skb);
 929         schedule_work(&psock_other->work);
 930         spin_unlock_bh(&psock_other->ingress_lock);
 931         return 0;
 932 }
 933
 934 static void sk_psock_tls_verdict_apply(struct sk_buff *skb,
 935                                        struct sk_psock *from, int verdict)
 936 {
 937         switch (verdict) {
 938         case __SK_REDIRECT:
 939                 sk_psock_skb_redirect(from, skb);
 940                 break;
 941         case __SK_PASS:
 942         case __SK_DROP:
 943         default:
 944                 break;
 945         }
 946 }
 947
 948 int sk_psock_tls_strp_read(struct sk_psock *psock, struct sk_buff *skb)
 949 {
 950         struct bpf_prog *prog;
 951         int ret = __SK_PASS;
 952
 953         rcu_read_lock();
 954         prog = READ_ONCE(psock->progs.stream_verdict);
 955         if (likely(prog)) {
 956                 skb->sk = psock->sk;
 957                 skb_dst_drop(skb);
 958                 skb_bpf_redirect_clear(skb);
 959                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
 960                 ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb));
 961                 skb->sk = NULL;
 962         }
 963         sk_psock_tls_verdict_apply(skb, psock, ret);
 964         rcu_read_unlock();
 965         return ret;
 966 }
 967 EXPORT_SYMBOL_GPL(sk_psock_tls_strp_read);
 968
 969 static int sk_psock_verdict_apply(struct sk_psock *psock, struct sk_buff *skb,
 970                                   int verdict)
 971 {
 972         struct sock *sk_other;
 973         int err = 0;
 974         u32 len, off;
 975
 976         switch (verdict) {
 977         case __SK_PASS:
 978                 err = -EIO;
 979                 sk_other = psock->sk;
 980                 if (sock_flag(sk_other, SOCK_DEAD) ||
 981                     !sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
 982                         skb_bpf_redirect_clear(skb);
 983                         goto out_free;
 984                 }
 985
 986                 skb_bpf_set_ingress(skb);
 987
 988                 /* If the queue is empty then we can submit directly
 989                  * into the msg queue. If its not empty we have to
 990                  * queue work otherwise we may get OOO data. Otherwise,
 991                  * if sk_psock_skb_ingress errors will be handled by
 992                  * retrying later from workqueue.
 993                  */
 994                 if (skb_queue_empty(&psock->ingress_skb)) {
 995                         len = skb->len;
 996                         off = 0;
 997                         if (skb_bpf_strparser(skb)) {
 998                                 struct strp_msg *stm = strp_msg(skb);
 999
1000                                 off = stm->offset;
1001                                 len = stm->full_len;
1002                         }
1003                         err = sk_psock_skb_ingress_self(psock, skb, off, len);
1004                 }
1005                 if (err < 0) {
1006                         spin_lock_bh(&psock->ingress_lock);
1007                         if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
1008                                 skb_queue_tail(&psock->ingress_skb, skb);
1009                                 schedule_work(&psock->work);
1010                                 err = 0;
1011                         }
1012                         spin_unlock_bh(&psock->ingress_lock);
1013                         if (err < 0) {
1014                                 skb_bpf_redirect_clear(skb);
1015                                 goto out_free;
1016                         }
1017                 }
1018                 break;
1019         case __SK_REDIRECT:
1020                 err = sk_psock_skb_redirect(psock, skb);
1021                 break;
1022         case __SK_DROP:
1023         default:
1024 out_free:
1025                 sock_drop(psock->sk, skb);
1026         }
1027
1028         return err;
1029 }
1030
1031 static void sk_psock_write_space(struct sock *sk)
1032 {
1033         struct sk_psock *psock;
1034         void (*write_space)(struct sock *sk) = NULL;
1035
1036         rcu_read_lock();
1037         psock = sk_psock(sk);
1038         if (likely(psock)) {
1039                 if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED))
1040                         schedule_work(&psock->work);
1041                 write_space = psock->saved_write_space;
1042         }
1043         rcu_read_unlock();
1044         if (write_space)
1045                 write_space(sk);
1046 }
1047
1048 #if IS_ENABLED(CONFIG_BPF_STREAM_PARSER)
1049 static void sk_psock_strp_read(struct strparser *strp, struct sk_buff *skb)
1050 {
1051         struct sk_psock *psock;
1052         struct bpf_prog *prog;
1053         int ret = __SK_DROP;
1054         struct sock *sk;
1055
1056         rcu_read_lock();
1057         sk = strp->sk;
1058         psock = sk_psock(sk);
1059         if (unlikely(!psock)) {
1060                 sock_drop(sk, skb);
1061                 goto out;
1062         }
1063         prog = READ_ONCE(psock->progs.stream_verdict);
1064         if (likely(prog)) {
1065                 skb->sk = sk;
1066                 skb_dst_drop(skb);
1067                 skb_bpf_redirect_clear(skb);
1068                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1069                 if (ret == SK_PASS)
1070                         skb_bpf_set_strparser(skb);
1071                 ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb));
1072                 skb->sk = NULL;
1073         }
1074         sk_psock_verdict_apply(psock, skb, ret);
1075 out:
1076         rcu_read_unlock();
1077 }
1078
1079 static int sk_psock_strp_read_done(struct strparser *strp, int err)
1080 {
1081         return err;
1082 }
1083
1084 static int sk_psock_strp_parse(struct strparser *strp, struct sk_buff *skb)
1085 {
1086         struct sk_psock *psock = container_of(strp, struct sk_psock, strp);
1087         struct bpf_prog *prog;
1088         int ret = skb->len;
1089
1090         rcu_read_lock();
1091         prog = READ_ONCE(psock->progs.stream_parser);
1092         if (likely(prog)) {
1093                 skb->sk = psock->sk;
1094                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1095                 skb->sk = NULL;
1096         }
1097         rcu_read_unlock();
1098         return ret;
1099 }
1100
1101 /* Called with socket lock held. */
1102 static void sk_psock_strp_data_ready(struct sock *sk)
1103 {
1104         struct sk_psock *psock;
1105
1106         rcu_read_lock();
1107         psock = sk_psock(sk);
1108         if (likely(psock)) {
1109                 if (tls_sw_has_ctx_rx(sk)) {
1110                         psock->saved_data_ready(sk);
1111                 } else {
1112                         write_lock_bh(&sk->sk_callback_lock);
1113                         strp_data_ready(&psock->strp);
1114                         write_unlock_bh(&sk->sk_callback_lock);
1115                 }
1116         }
1117         rcu_read_unlock();
1118 }
1119
1120 int sk_psock_init_strp(struct sock *sk, struct sk_psock *psock)
1121 {
1122         static const struct strp_callbacks cb = {
1123                 .rcv_msg        = sk_psock_strp_read,
1124                 .read_sock_done = sk_psock_strp_read_done,
1125                 .parse_msg      = sk_psock_strp_parse,
1126         };
1127
1128         return strp_init(&psock->strp, sk, &cb);
1129 }
1130
1131 void sk_psock_start_strp(struct sock *sk, struct sk_psock *psock)
1132 {
1133         if (psock->saved_data_ready)
1134                 return;
1135
1136         psock->saved_data_ready = sk->sk_data_ready;
1137         sk->sk_data_ready = sk_psock_strp_data_ready;
1138         sk->sk_write_space = sk_psock_write_space;
1139 }
1140
1141 void sk_psock_stop_strp(struct sock *sk, struct sk_psock *psock)
1142 {
1143         psock_set_prog(&psock->progs.stream_parser, NULL);
1144
1145         if (!psock->saved_data_ready)
1146                 return;
1147
1148         sk->sk_data_ready = psock->saved_data_ready;
1149         psock->saved_data_ready = NULL;
1150         strp_stop(&psock->strp);
1151 }
1152
1153 static void sk_psock_done_strp(struct sk_psock *psock)
1154 {
1155         /* Parser has been stopped */
1156         if (psock->progs.stream_parser)
1157                 strp_done(&psock->strp);
1158 }
1159 #else
1160 static void sk_psock_done_strp(struct sk_psock *psock)
1161 {
1162 }
1163 #endif /* CONFIG_BPF_STREAM_PARSER */
1164
1165 static int sk_psock_verdict_recv(read_descriptor_t *desc, struct sk_buff *skb,
1166                                  unsigned int offset, size_t orig_len)
1167 {
1168         struct sock *sk = (struct sock *)desc->arg.data;
1169         struct sk_psock *psock;
1170         struct bpf_prog *prog;
1171         int ret = __SK_DROP;
1172         int len = orig_len;
1173
1174         /* clone here so sk_eat_skb() in tcp_read_sock does not drop our data */
1175         skb = skb_clone(skb, GFP_ATOMIC);
1176         if (!skb) {
1177                 desc->error = -ENOMEM;
1178                 return 0;
1179         }
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->sk = sk;
1193                 skb_dst_drop(skb);
1194                 skb_bpf_redirect_clear(skb);
1195                 ret = bpf_prog_run_pin_on_cpu(prog, skb);
1196                 ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb));
1197                 skb->sk = NULL;
1198         }
1199         if (sk_psock_verdict_apply(psock, skb, ret) < 0)
1200                 len = 0;
1201 out:
1202         rcu_read_unlock();
1203         return len;
1204 }
1205
1206 static void sk_psock_verdict_data_ready(struct sock *sk)
1207 {
1208         struct socket *sock = sk->sk_socket;
1209         read_descriptor_t desc;
1210
1211         if (unlikely(!sock || !sock->ops || !sock->ops->read_sock))
1212                 return;
1213
1214         desc.arg.data = sk;
1215         desc.error = 0;
1216         desc.count = 1;
1217
1218         sock->ops->read_sock(sk, &desc, sk_psock_verdict_recv);
1219 }
1220
1221 void sk_psock_start_verdict(struct sock *sk, struct sk_psock *psock)
1222 {
1223         if (psock->saved_data_ready)
1224                 return;
1225
1226         psock->saved_data_ready = sk->sk_data_ready;
1227         sk->sk_data_ready = sk_psock_verdict_data_ready;
1228         sk->sk_write_space = sk_psock_write_space;
1229 }
1230
1231 void sk_psock_stop_verdict(struct sock *sk, struct sk_psock *psock)
1232 {
1233         psock_set_prog(&psock->progs.stream_verdict, NULL);
1234         psock_set_prog(&psock->progs.skb_verdict, NULL);
1235
1236         if (!psock->saved_data_ready)
1237                 return;
1238
1239         sk->sk_data_ready = psock->saved_data_ready;
1240         psock->saved_data_ready = NULL;
1241 }