net/kcm/kcmsock.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Kernel Connection Multiplexor
   4  *
   5  * Copyright (c) 2016 Tom Herbert <tom@herbertland.com>
   6  */
   7
   8 #include <linux/bpf.h>
   9 #include <linux/errno.h>
  10 #include <linux/errqueue.h>
  11 #include <linux/file.h>
  12 #include <linux/filter.h>
  13 #include <linux/in.h>
  14 #include <linux/kernel.h>
  15 #include <linux/module.h>
  16 #include <linux/net.h>
  17 #include <linux/netdevice.h>
  18 #include <linux/poll.h>
  19 #include <linux/rculist.h>
  20 #include <linux/skbuff.h>
  21 #include <linux/socket.h>
  22 #include <linux/uaccess.h>
  23 #include <linux/workqueue.h>
  24 #include <linux/syscalls.h>
  25 #include <linux/sched/signal.h>
  26
  27 #include <net/kcm.h>
  28 #include <net/netns/generic.h>
  29 #include <net/sock.h>
  30 #include <uapi/linux/kcm.h>
  31
  32 unsigned int kcm_net_id;
  33
  34 static struct kmem_cache *kcm_psockp __read_mostly;
  35 static struct kmem_cache *kcm_muxp __read_mostly;
  36 static struct workqueue_struct *kcm_wq;
  37
  38 static inline struct kcm_sock *kcm_sk(const struct sock *sk)
  39 {
  40         return (struct kcm_sock *)sk;
  41 }
  42
  43 static inline struct kcm_tx_msg *kcm_tx_msg(struct sk_buff *skb)
  44 {
  45         return (struct kcm_tx_msg *)skb->cb;
  46 }
  47
  48 static void report_csk_error(struct sock *csk, int err)
  49 {
  50         csk->sk_err = EPIPE;
  51         sk_error_report(csk);
  52 }
  53
  54 static void kcm_abort_tx_psock(struct kcm_psock *psock, int err,
  55                                bool wakeup_kcm)
  56 {
  57         struct sock *csk = psock->sk;
  58         struct kcm_mux *mux = psock->mux;
  59
  60         /* Unrecoverable error in transmit */
  61
  62         spin_lock_bh(&mux->lock);
  63
  64         if (psock->tx_stopped) {
  65                 spin_unlock_bh(&mux->lock);
  66                 return;
  67         }
  68
  69         psock->tx_stopped = 1;
  70         KCM_STATS_INCR(psock->stats.tx_aborts);
  71
  72         if (!psock->tx_kcm) {
  73                 /* Take off psocks_avail list */
  74                 list_del(&psock->psock_avail_list);
  75         } else if (wakeup_kcm) {
  76                 /* In this case psock is being aborted while outside of
  77                  * write_msgs and psock is reserved. Schedule tx_work
  78                  * to handle the failure there. Need to commit tx_stopped
  79                  * before queuing work.
  80                  */
  81                 smp_mb();
  82
  83                 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
  84         }
  85
  86         spin_unlock_bh(&mux->lock);
  87
  88         /* Report error on lower socket */
  89         report_csk_error(csk, err);
  90 }
  91
  92 /* RX mux lock held. */
  93 static void kcm_update_rx_mux_stats(struct kcm_mux *mux,
  94                                     struct kcm_psock *psock)
  95 {
  96         STRP_STATS_ADD(mux->stats.rx_bytes,
  97                        psock->strp.stats.bytes -
  98                        psock->saved_rx_bytes);
  99         mux->stats.rx_msgs +=
 100                 psock->strp.stats.msgs - psock->saved_rx_msgs;
 101         psock->saved_rx_msgs = psock->strp.stats.msgs;
 102         psock->saved_rx_bytes = psock->strp.stats.bytes;
 103 }
 104
 105 static void kcm_update_tx_mux_stats(struct kcm_mux *mux,
 106                                     struct kcm_psock *psock)
 107 {
 108         KCM_STATS_ADD(mux->stats.tx_bytes,
 109                       psock->stats.tx_bytes - psock->saved_tx_bytes);
 110         mux->stats.tx_msgs +=
 111                 psock->stats.tx_msgs - psock->saved_tx_msgs;
 112         psock->saved_tx_msgs = psock->stats.tx_msgs;
 113         psock->saved_tx_bytes = psock->stats.tx_bytes;
 114 }
 115
 116 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb);
 117
 118 /* KCM is ready to receive messages on its queue-- either the KCM is new or
 119  * has become unblocked after being blocked on full socket buffer. Queue any
 120  * pending ready messages on a psock. RX mux lock held.
 121  */
 122 static void kcm_rcv_ready(struct kcm_sock *kcm)
 123 {
 124         struct kcm_mux *mux = kcm->mux;
 125         struct kcm_psock *psock;
 126         struct sk_buff *skb;
 127
 128         if (unlikely(kcm->rx_wait || kcm->rx_psock || kcm->rx_disabled))
 129                 return;
 130
 131         while (unlikely((skb = __skb_dequeue(&mux->rx_hold_queue)))) {
 132                 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 133                         /* Assuming buffer limit has been reached */
 134                         skb_queue_head(&mux->rx_hold_queue, skb);
 135                         WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
 136                         return;
 137                 }
 138         }
 139
 140         while (!list_empty(&mux->psocks_ready)) {
 141                 psock = list_first_entry(&mux->psocks_ready, struct kcm_psock,
 142                                          psock_ready_list);
 143
 144                 if (kcm_queue_rcv_skb(&kcm->sk, psock->ready_rx_msg)) {
 145                         /* Assuming buffer limit has been reached */
 146                         WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
 147                         return;
 148                 }
 149
 150                 /* Consumed the ready message on the psock. Schedule rx_work to
 151                  * get more messages.
 152                  */
 153                 list_del(&psock->psock_ready_list);
 154                 psock->ready_rx_msg = NULL;
 155                 /* Commit clearing of ready_rx_msg for queuing work */
 156                 smp_mb();
 157
 158                 strp_unpause(&psock->strp);
 159                 strp_check_rcv(&psock->strp);
 160         }
 161
 162         /* Buffer limit is okay now, add to ready list */
 163         list_add_tail(&kcm->wait_rx_list,
 164                       &kcm->mux->kcm_rx_waiters);
 165         /* paired with lockless reads in kcm_rfree() */
 166         WRITE_ONCE(kcm->rx_wait, true);
 167 }
 168
 169 static void kcm_rfree(struct sk_buff *skb)
 170 {
 171         struct sock *sk = skb->sk;
 172         struct kcm_sock *kcm = kcm_sk(sk);
 173         struct kcm_mux *mux = kcm->mux;
 174         unsigned int len = skb->truesize;
 175
 176         sk_mem_uncharge(sk, len);
 177         atomic_sub(len, &sk->sk_rmem_alloc);
 178
 179         /* For reading rx_wait and rx_psock without holding lock */
 180         smp_mb__after_atomic();
 181
 182         if (!READ_ONCE(kcm->rx_wait) && !READ_ONCE(kcm->rx_psock) &&
 183             sk_rmem_alloc_get(sk) < sk->sk_rcvlowat) {
 184                 spin_lock_bh(&mux->rx_lock);
 185                 kcm_rcv_ready(kcm);
 186                 spin_unlock_bh(&mux->rx_lock);
 187         }
 188 }
 189
 190 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb)
 191 {
 192         struct sk_buff_head *list = &sk->sk_receive_queue;
 193
 194         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
 195                 return -ENOMEM;
 196
 197         if (!sk_rmem_schedule(sk, skb, skb->truesize))
 198                 return -ENOBUFS;
 199
 200         skb->dev = NULL;
 201
 202         skb_orphan(skb);
 203         skb->sk = sk;
 204         skb->destructor = kcm_rfree;
 205         atomic_add(skb->truesize, &sk->sk_rmem_alloc);
 206         sk_mem_charge(sk, skb->truesize);
 207
 208         skb_queue_tail(list, skb);
 209
 210         if (!sock_flag(sk, SOCK_DEAD))
 211                 sk->sk_data_ready(sk);
 212
 213         return 0;
 214 }
 215
 216 /* Requeue received messages for a kcm socket to other kcm sockets. This is
 217  * called with a kcm socket is receive disabled.
 218  * RX mux lock held.
 219  */
 220 static void requeue_rx_msgs(struct kcm_mux *mux, struct sk_buff_head *head)
 221 {
 222         struct sk_buff *skb;
 223         struct kcm_sock *kcm;
 224
 225         while ((skb = skb_dequeue(head))) {
 226                 /* Reset destructor to avoid calling kcm_rcv_ready */
 227                 skb->destructor = sock_rfree;
 228                 skb_orphan(skb);
 229 try_again:
 230                 if (list_empty(&mux->kcm_rx_waiters)) {
 231                         skb_queue_tail(&mux->rx_hold_queue, skb);
 232                         continue;
 233                 }
 234
 235                 kcm = list_first_entry(&mux->kcm_rx_waiters,
 236                                        struct kcm_sock, wait_rx_list);
 237
 238                 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 239                         /* Should mean socket buffer full */
 240                         list_del(&kcm->wait_rx_list);
 241                         /* paired with lockless reads in kcm_rfree() */
 242                         WRITE_ONCE(kcm->rx_wait, false);
 243
 244                         /* Commit rx_wait to read in kcm_free */
 245                         smp_wmb();
 246
 247                         goto try_again;
 248                 }
 249         }
 250 }
 251
 252 /* Lower sock lock held */
 253 static struct kcm_sock *reserve_rx_kcm(struct kcm_psock *psock,
 254                                        struct sk_buff *head)
 255 {
 256         struct kcm_mux *mux = psock->mux;
 257         struct kcm_sock *kcm;
 258
 259         WARN_ON(psock->ready_rx_msg);
 260
 261         if (psock->rx_kcm)
 262                 return psock->rx_kcm;
 263
 264         spin_lock_bh(&mux->rx_lock);
 265
 266         if (psock->rx_kcm) {
 267                 spin_unlock_bh(&mux->rx_lock);
 268                 return psock->rx_kcm;
 269         }
 270
 271         kcm_update_rx_mux_stats(mux, psock);
 272
 273         if (list_empty(&mux->kcm_rx_waiters)) {
 274                 psock->ready_rx_msg = head;
 275                 strp_pause(&psock->strp);
 276                 list_add_tail(&psock->psock_ready_list,
 277                               &mux->psocks_ready);
 278                 spin_unlock_bh(&mux->rx_lock);
 279                 return NULL;
 280         }
 281
 282         kcm = list_first_entry(&mux->kcm_rx_waiters,
 283                                struct kcm_sock, wait_rx_list);
 284         list_del(&kcm->wait_rx_list);
 285         /* paired with lockless reads in kcm_rfree() */
 286         WRITE_ONCE(kcm->rx_wait, false);
 287
 288         psock->rx_kcm = kcm;
 289         /* paired with lockless reads in kcm_rfree() */
 290         WRITE_ONCE(kcm->rx_psock, psock);
 291
 292         spin_unlock_bh(&mux->rx_lock);
 293
 294         return kcm;
 295 }
 296
 297 static void kcm_done(struct kcm_sock *kcm);
 298
 299 static void kcm_done_work(struct work_struct *w)
 300 {
 301         kcm_done(container_of(w, struct kcm_sock, done_work));
 302 }
 303
 304 /* Lower sock held */
 305 static void unreserve_rx_kcm(struct kcm_psock *psock,
 306                              bool rcv_ready)
 307 {
 308         struct kcm_sock *kcm = psock->rx_kcm;
 309         struct kcm_mux *mux = psock->mux;
 310
 311         if (!kcm)
 312                 return;
 313
 314         spin_lock_bh(&mux->rx_lock);
 315
 316         psock->rx_kcm = NULL;
 317         /* paired with lockless reads in kcm_rfree() */
 318         WRITE_ONCE(kcm->rx_psock, NULL);
 319
 320         /* Commit kcm->rx_psock before sk_rmem_alloc_get to sync with
 321          * kcm_rfree
 322          */
 323         smp_mb();
 324
 325         if (unlikely(kcm->done)) {
 326                 spin_unlock_bh(&mux->rx_lock);
 327
 328                 /* Need to run kcm_done in a task since we need to qcquire
 329                  * callback locks which may already be held here.
 330                  */
 331                 INIT_WORK(&kcm->done_work, kcm_done_work);
 332                 schedule_work(&kcm->done_work);
 333                 return;
 334         }
 335
 336         if (unlikely(kcm->rx_disabled)) {
 337                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
 338         } else if (rcv_ready || unlikely(!sk_rmem_alloc_get(&kcm->sk))) {
 339                 /* Check for degenerative race with rx_wait that all
 340                  * data was dequeued (accounted for in kcm_rfree).
 341                  */
 342                 kcm_rcv_ready(kcm);
 343         }
 344         spin_unlock_bh(&mux->rx_lock);
 345 }
 346
 347 /* Lower sock lock held */
 348 static void psock_data_ready(struct sock *sk)
 349 {
 350         struct kcm_psock *psock;
 351
 352         read_lock_bh(&sk->sk_callback_lock);
 353
 354         psock = (struct kcm_psock *)sk->sk_user_data;
 355         if (likely(psock))
 356                 strp_data_ready(&psock->strp);
 357
 358         read_unlock_bh(&sk->sk_callback_lock);
 359 }
 360
 361 /* Called with lower sock held */
 362 static void kcm_rcv_strparser(struct strparser *strp, struct sk_buff *skb)
 363 {
 364         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 365         struct kcm_sock *kcm;
 366
 367 try_queue:
 368         kcm = reserve_rx_kcm(psock, skb);
 369         if (!kcm) {
 370                  /* Unable to reserve a KCM, message is held in psock and strp
 371                   * is paused.
 372                   */
 373                 return;
 374         }
 375
 376         if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 377                 /* Should mean socket buffer full */
 378                 unreserve_rx_kcm(psock, false);
 379                 goto try_queue;
 380         }
 381 }
 382
 383 static int kcm_parse_func_strparser(struct strparser *strp, struct sk_buff *skb)
 384 {
 385         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 386         struct bpf_prog *prog = psock->bpf_prog;
 387         int res;
 388
 389         res = bpf_prog_run_pin_on_cpu(prog, skb);
 390         return res;
 391 }
 392
 393 static int kcm_read_sock_done(struct strparser *strp, int err)
 394 {
 395         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 396
 397         unreserve_rx_kcm(psock, true);
 398
 399         return err;
 400 }
 401
 402 static void psock_state_change(struct sock *sk)
 403 {
 404         /* TCP only does a EPOLLIN for a half close. Do a EPOLLHUP here
 405          * since application will normally not poll with EPOLLIN
 406          * on the TCP sockets.
 407          */
 408
 409         report_csk_error(sk, EPIPE);
 410 }
 411
 412 static void psock_write_space(struct sock *sk)
 413 {
 414         struct kcm_psock *psock;
 415         struct kcm_mux *mux;
 416         struct kcm_sock *kcm;
 417
 418         read_lock_bh(&sk->sk_callback_lock);
 419
 420         psock = (struct kcm_psock *)sk->sk_user_data;
 421         if (unlikely(!psock))
 422                 goto out;
 423         mux = psock->mux;
 424
 425         spin_lock_bh(&mux->lock);
 426
 427         /* Check if the socket is reserved so someone is waiting for sending. */
 428         kcm = psock->tx_kcm;
 429         if (kcm && !unlikely(kcm->tx_stopped))
 430                 queue_work(kcm_wq, &kcm->tx_work);
 431
 432         spin_unlock_bh(&mux->lock);
 433 out:
 434         read_unlock_bh(&sk->sk_callback_lock);
 435 }
 436
 437 static void unreserve_psock(struct kcm_sock *kcm);
 438
 439 /* kcm sock is locked. */
 440 static struct kcm_psock *reserve_psock(struct kcm_sock *kcm)
 441 {
 442         struct kcm_mux *mux = kcm->mux;
 443         struct kcm_psock *psock;
 444
 445         psock = kcm->tx_psock;
 446
 447         smp_rmb(); /* Must read tx_psock before tx_wait */
 448
 449         if (psock) {
 450                 WARN_ON(kcm->tx_wait);
 451                 if (unlikely(psock->tx_stopped))
 452                         unreserve_psock(kcm);
 453                 else
 454                         return kcm->tx_psock;
 455         }
 456
 457         spin_lock_bh(&mux->lock);
 458
 459         /* Check again under lock to see if psock was reserved for this
 460          * psock via psock_unreserve.
 461          */
 462         psock = kcm->tx_psock;
 463         if (unlikely(psock)) {
 464                 WARN_ON(kcm->tx_wait);
 465                 spin_unlock_bh(&mux->lock);
 466                 return kcm->tx_psock;
 467         }
 468
 469         if (!list_empty(&mux->psocks_avail)) {
 470                 psock = list_first_entry(&mux->psocks_avail,
 471                                          struct kcm_psock,
 472                                          psock_avail_list);
 473                 list_del(&psock->psock_avail_list);
 474                 if (kcm->tx_wait) {
 475                         list_del(&kcm->wait_psock_list);
 476                         kcm->tx_wait = false;
 477                 }
 478                 kcm->tx_psock = psock;
 479                 psock->tx_kcm = kcm;
 480                 KCM_STATS_INCR(psock->stats.reserved);
 481         } else if (!kcm->tx_wait) {
 482                 list_add_tail(&kcm->wait_psock_list,
 483                               &mux->kcm_tx_waiters);
 484                 kcm->tx_wait = true;
 485         }
 486
 487         spin_unlock_bh(&mux->lock);
 488
 489         return psock;
 490 }
 491
 492 /* mux lock held */
 493 static void psock_now_avail(struct kcm_psock *psock)
 494 {
 495         struct kcm_mux *mux = psock->mux;
 496         struct kcm_sock *kcm;
 497
 498         if (list_empty(&mux->kcm_tx_waiters)) {
 499                 list_add_tail(&psock->psock_avail_list,
 500                               &mux->psocks_avail);
 501         } else {
 502                 kcm = list_first_entry(&mux->kcm_tx_waiters,
 503                                        struct kcm_sock,
 504                                        wait_psock_list);
 505                 list_del(&kcm->wait_psock_list);
 506                 kcm->tx_wait = false;
 507                 psock->tx_kcm = kcm;
 508
 509                 /* Commit before changing tx_psock since that is read in
 510                  * reserve_psock before queuing work.
 511                  */
 512                 smp_mb();
 513
 514                 kcm->tx_psock = psock;
 515                 KCM_STATS_INCR(psock->stats.reserved);
 516                 queue_work(kcm_wq, &kcm->tx_work);
 517         }
 518 }
 519
 520 /* kcm sock is locked. */
 521 static void unreserve_psock(struct kcm_sock *kcm)
 522 {
 523         struct kcm_psock *psock;
 524         struct kcm_mux *mux = kcm->mux;
 525
 526         spin_lock_bh(&mux->lock);
 527
 528         psock = kcm->tx_psock;
 529
 530         if (WARN_ON(!psock)) {
 531                 spin_unlock_bh(&mux->lock);
 532                 return;
 533         }
 534
 535         smp_rmb(); /* Read tx_psock before tx_wait */
 536
 537         kcm_update_tx_mux_stats(mux, psock);
 538
 539         WARN_ON(kcm->tx_wait);
 540
 541         kcm->tx_psock = NULL;
 542         psock->tx_kcm = NULL;
 543         KCM_STATS_INCR(psock->stats.unreserved);
 544
 545         if (unlikely(psock->tx_stopped)) {
 546                 if (psock->done) {
 547                         /* Deferred free */
 548                         list_del(&psock->psock_list);
 549                         mux->psocks_cnt--;
 550                         sock_put(psock->sk);
 551                         fput(psock->sk->sk_socket->file);
 552                         kmem_cache_free(kcm_psockp, psock);
 553                 }
 554
 555                 /* Don't put back on available list */
 556
 557                 spin_unlock_bh(&mux->lock);
 558
 559                 return;
 560         }
 561
 562         psock_now_avail(psock);
 563
 564         spin_unlock_bh(&mux->lock);
 565 }
 566
 567 static void kcm_report_tx_retry(struct kcm_sock *kcm)
 568 {
 569         struct kcm_mux *mux = kcm->mux;
 570
 571         spin_lock_bh(&mux->lock);
 572         KCM_STATS_INCR(mux->stats.tx_retries);
 573         spin_unlock_bh(&mux->lock);
 574 }
 575
 576 /* Write any messages ready on the kcm socket.  Called with kcm sock lock
 577  * held.  Return bytes actually sent or error.
 578  */
 579 static int kcm_write_msgs(struct kcm_sock *kcm)
 580 {
 581         struct sock *sk = &kcm->sk;
 582         struct kcm_psock *psock;
 583         struct sk_buff *skb, *head;
 584         struct kcm_tx_msg *txm;
 585         unsigned short fragidx, frag_offset;
 586         unsigned int sent, total_sent = 0;
 587         int ret = 0;
 588
 589         kcm->tx_wait_more = false;
 590         psock = kcm->tx_psock;
 591         if (unlikely(psock && psock->tx_stopped)) {
 592                 /* A reserved psock was aborted asynchronously. Unreserve
 593                  * it and we'll retry the message.
 594                  */
 595                 unreserve_psock(kcm);
 596                 kcm_report_tx_retry(kcm);
 597                 if (skb_queue_empty(&sk->sk_write_queue))
 598                         return 0;
 599
 600                 kcm_tx_msg(skb_peek(&sk->sk_write_queue))->sent = 0;
 601
 602         } else if (skb_queue_empty(&sk->sk_write_queue)) {
 603                 return 0;
 604         }
 605
 606         head = skb_peek(&sk->sk_write_queue);
 607         txm = kcm_tx_msg(head);
 608
 609         if (txm->sent) {
 610                 /* Send of first skbuff in queue already in progress */
 611                 if (WARN_ON(!psock)) {
 612                         ret = -EINVAL;
 613                         goto out;
 614                 }
 615                 sent = txm->sent;
 616                 frag_offset = txm->frag_offset;
 617                 fragidx = txm->fragidx;
 618                 skb = txm->frag_skb;
 619
 620                 goto do_frag;
 621         }
 622
 623 try_again:
 624         psock = reserve_psock(kcm);
 625         if (!psock)
 626                 goto out;
 627
 628         do {
 629                 skb = head;
 630                 txm = kcm_tx_msg(head);
 631                 sent = 0;
 632
 633 do_frag_list:
 634                 if (WARN_ON(!skb_shinfo(skb)->nr_frags)) {
 635                         ret = -EINVAL;
 636                         goto out;
 637                 }
 638
 639                 for (fragidx = 0; fragidx < skb_shinfo(skb)->nr_frags;
 640                      fragidx++) {
 641                         skb_frag_t *frag;
 642
 643                         frag_offset = 0;
 644 do_frag:
 645                         frag = &skb_shinfo(skb)->frags[fragidx];
 646                         if (WARN_ON(!skb_frag_size(frag))) {
 647                                 ret = -EINVAL;
 648                                 goto out;
 649                         }
 650
 651                         ret = kernel_sendpage(psock->sk->sk_socket,
 652                                               skb_frag_page(frag),
 653                                               skb_frag_off(frag) + frag_offset,
 654                                               skb_frag_size(frag) - frag_offset,
 655                                               MSG_DONTWAIT);
 656                         if (ret <= 0) {
 657                                 if (ret == -EAGAIN) {
 658                                         /* Save state to try again when there's
 659                                          * write space on the socket
 660                                          */
 661                                         txm->sent = sent;
 662                                         txm->frag_offset = frag_offset;
 663                                         txm->fragidx = fragidx;
 664                                         txm->frag_skb = skb;
 665
 666                                         ret = 0;
 667                                         goto out;
 668                                 }
 669
 670                                 /* Hard failure in sending message, abort this
 671                                  * psock since it has lost framing
 672                                  * synchronization and retry sending the
 673                                  * message from the beginning.
 674                                  */
 675                                 kcm_abort_tx_psock(psock, ret ? -ret : EPIPE,
 676                                                    true);
 677                                 unreserve_psock(kcm);
 678
 679                                 txm->sent = 0;
 680                                 kcm_report_tx_retry(kcm);
 681                                 ret = 0;
 682
 683                                 goto try_again;
 684                         }
 685
 686                         sent += ret;
 687                         frag_offset += ret;
 688                         KCM_STATS_ADD(psock->stats.tx_bytes, ret);
 689                         if (frag_offset < skb_frag_size(frag)) {
 690                                 /* Not finished with this frag */
 691                                 goto do_frag;
 692                         }
 693                 }
 694
 695                 if (skb == head) {
 696                         if (skb_has_frag_list(skb)) {
 697                                 skb = skb_shinfo(skb)->frag_list;
 698                                 goto do_frag_list;
 699                         }
 700                 } else if (skb->next) {
 701                         skb = skb->next;
 702                         goto do_frag_list;
 703                 }
 704
 705                 /* Successfully sent the whole packet, account for it. */
 706                 skb_dequeue(&sk->sk_write_queue);
 707                 kfree_skb(head);
 708                 sk->sk_wmem_queued -= sent;
 709                 total_sent += sent;
 710                 KCM_STATS_INCR(psock->stats.tx_msgs);
 711         } while ((head = skb_peek(&sk->sk_write_queue)));
 712 out:
 713         if (!head) {
 714                 /* Done with all queued messages. */
 715                 WARN_ON(!skb_queue_empty(&sk->sk_write_queue));
 716                 unreserve_psock(kcm);
 717         }
 718
 719         /* Check if write space is available */
 720         sk->sk_write_space(sk);
 721
 722         return total_sent ? : ret;
 723 }
 724
 725 static void kcm_tx_work(struct work_struct *w)
 726 {
 727         struct kcm_sock *kcm = container_of(w, struct kcm_sock, tx_work);
 728         struct sock *sk = &kcm->sk;
 729         int err;
 730
 731         lock_sock(sk);
 732
 733         /* Primarily for SOCK_DGRAM sockets, also handle asynchronous tx
 734          * aborts
 735          */
 736         err = kcm_write_msgs(kcm);
 737         if (err < 0) {
 738                 /* Hard failure in write, report error on KCM socket */
 739                 pr_warn("KCM: Hard failure on kcm_write_msgs %d\n", err);
 740                 report_csk_error(&kcm->sk, -err);
 741                 goto out;
 742         }
 743
 744         /* Primarily for SOCK_SEQPACKET sockets */
 745         if (likely(sk->sk_socket) &&
 746             test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
 747                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 748                 sk->sk_write_space(sk);
 749         }
 750
 751 out:
 752         release_sock(sk);
 753 }
 754
 755 static void kcm_push(struct kcm_sock *kcm)
 756 {
 757         if (kcm->tx_wait_more)
 758                 kcm_write_msgs(kcm);
 759 }
 760
 761 static ssize_t kcm_sendpage(struct socket *sock, struct page *page,
 762                             int offset, size_t size, int flags)
 763
 764 {
 765         struct sock *sk = sock->sk;
 766         struct kcm_sock *kcm = kcm_sk(sk);
 767         struct sk_buff *skb = NULL, *head = NULL;
 768         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 769         bool eor;
 770         int err = 0;
 771         int i;
 772
 773         if (flags & MSG_SENDPAGE_NOTLAST)
 774                 flags |= MSG_MORE;
 775
 776         /* No MSG_EOR from splice, only look at MSG_MORE */
 777         eor = !(flags & MSG_MORE);
 778
 779         lock_sock(sk);
 780
 781         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 782
 783         err = -EPIPE;
 784         if (sk->sk_err)
 785                 goto out_error;
 786
 787         if (kcm->seq_skb) {
 788                 /* Previously opened message */
 789                 head = kcm->seq_skb;
 790                 skb = kcm_tx_msg(head)->last_skb;
 791                 i = skb_shinfo(skb)->nr_frags;
 792
 793                 if (skb_can_coalesce(skb, i, page, offset)) {
 794                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], size);
 795                         skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
 796                         goto coalesced;
 797                 }
 798
 799                 if (i >= MAX_SKB_FRAGS) {
 800                         struct sk_buff *tskb;
 801
 802                         tskb = alloc_skb(0, sk->sk_allocation);
 803                         while (!tskb) {
 804                                 kcm_push(kcm);
 805                                 err = sk_stream_wait_memory(sk, &timeo);
 806                                 if (err)
 807                                         goto out_error;
 808                         }
 809
 810                         if (head == skb)
 811                                 skb_shinfo(head)->frag_list = tskb;
 812                         else
 813                                 skb->next = tskb;
 814
 815                         skb = tskb;
 816                         skb->ip_summed = CHECKSUM_UNNECESSARY;
 817                         i = 0;
 818                 }
 819         } else {
 820                 /* Call the sk_stream functions to manage the sndbuf mem. */
 821                 if (!sk_stream_memory_free(sk)) {
 822                         kcm_push(kcm);
 823                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 824                         err = sk_stream_wait_memory(sk, &timeo);
 825                         if (err)
 826                                 goto out_error;
 827                 }
 828
 829                 head = alloc_skb(0, sk->sk_allocation);
 830                 while (!head) {
 831                         kcm_push(kcm);
 832                         err = sk_stream_wait_memory(sk, &timeo);
 833                         if (err)
 834                                 goto out_error;
 835                 }
 836
 837                 skb = head;
 838                 i = 0;
 839         }
 840
 841         get_page(page);
 842         skb_fill_page_desc_noacc(skb, i, page, offset, size);
 843         skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
 844
 845 coalesced:
 846         skb->len += size;
 847         skb->data_len += size;
 848         skb->truesize += size;
 849         sk->sk_wmem_queued += size;
 850         sk_mem_charge(sk, size);
 851
 852         if (head != skb) {
 853                 head->len += size;
 854                 head->data_len += size;
 855                 head->truesize += size;
 856         }
 857
 858         if (eor) {
 859                 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
 860
 861                 /* Message complete, queue it on send buffer */
 862                 __skb_queue_tail(&sk->sk_write_queue, head);
 863                 kcm->seq_skb = NULL;
 864                 KCM_STATS_INCR(kcm->stats.tx_msgs);
 865
 866                 if (flags & MSG_BATCH) {
 867                         kcm->tx_wait_more = true;
 868                 } else if (kcm->tx_wait_more || not_busy) {
 869                         err = kcm_write_msgs(kcm);
 870                         if (err < 0) {
 871                                 /* We got a hard error in write_msgs but have
 872                                  * already queued this message. Report an error
 873                                  * in the socket, but don't affect return value
 874                                  * from sendmsg
 875                                  */
 876                                 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
 877                                 report_csk_error(&kcm->sk, -err);
 878                         }
 879                 }
 880         } else {
 881                 /* Message not complete, save state */
 882                 kcm->seq_skb = head;
 883                 kcm_tx_msg(head)->last_skb = skb;
 884         }
 885
 886         KCM_STATS_ADD(kcm->stats.tx_bytes, size);
 887
 888         release_sock(sk);
 889         return size;
 890
 891 out_error:
 892         kcm_push(kcm);
 893
 894         err = sk_stream_error(sk, flags, err);
 895
 896         /* make sure we wake any epoll edge trigger waiter */
 897         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
 898                 sk->sk_write_space(sk);
 899
 900         release_sock(sk);
 901         return err;
 902 }
 903
 904 static int kcm_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
 905 {
 906         struct sock *sk = sock->sk;
 907         struct kcm_sock *kcm = kcm_sk(sk);
 908         struct sk_buff *skb = NULL, *head = NULL;
 909         size_t copy, copied = 0;
 910         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 911         int eor = (sock->type == SOCK_DGRAM) ?
 912                   !(msg->msg_flags & MSG_MORE) : !!(msg->msg_flags & MSG_EOR);
 913         int err = -EPIPE;
 914
 915         lock_sock(sk);
 916
 917         /* Per tcp_sendmsg this should be in poll */
 918         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 919
 920         if (sk->sk_err)
 921                 goto out_error;
 922
 923         if (kcm->seq_skb) {
 924                 /* Previously opened message */
 925                 head = kcm->seq_skb;
 926                 skb = kcm_tx_msg(head)->last_skb;
 927                 goto start;
 928         }
 929
 930         /* Call the sk_stream functions to manage the sndbuf mem. */
 931         if (!sk_stream_memory_free(sk)) {
 932                 kcm_push(kcm);
 933                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 934                 err = sk_stream_wait_memory(sk, &timeo);
 935                 if (err)
 936                         goto out_error;
 937         }
 938
 939         if (msg_data_left(msg)) {
 940                 /* New message, alloc head skb */
 941                 head = alloc_skb(0, sk->sk_allocation);
 942                 while (!head) {
 943                         kcm_push(kcm);
 944                         err = sk_stream_wait_memory(sk, &timeo);
 945                         if (err)
 946                                 goto out_error;
 947
 948                         head = alloc_skb(0, sk->sk_allocation);
 949                 }
 950
 951                 skb = head;
 952
 953                 /* Set ip_summed to CHECKSUM_UNNECESSARY to avoid calling
 954                  * csum_and_copy_from_iter from skb_do_copy_data_nocache.
 955                  */
 956                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 957         }
 958
 959 start:
 960         while (msg_data_left(msg)) {
 961                 bool merge = true;
 962                 int i = skb_shinfo(skb)->nr_frags;
 963                 struct page_frag *pfrag = sk_page_frag(sk);
 964
 965                 if (!sk_page_frag_refill(sk, pfrag))
 966                         goto wait_for_memory;
 967
 968                 if (!skb_can_coalesce(skb, i, pfrag->page,
 969                                       pfrag->offset)) {
 970                         if (i == MAX_SKB_FRAGS) {
 971                                 struct sk_buff *tskb;
 972
 973                                 tskb = alloc_skb(0, sk->sk_allocation);
 974                                 if (!tskb)
 975                                         goto wait_for_memory;
 976
 977                                 if (head == skb)
 978                                         skb_shinfo(head)->frag_list = tskb;
 979                                 else
 980                                         skb->next = tskb;
 981
 982                                 skb = tskb;
 983                                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 984                                 continue;
 985                         }
 986                         merge = false;
 987                 }
 988
 989                 copy = min_t(int, msg_data_left(msg),
 990                              pfrag->size - pfrag->offset);
 991
 992                 if (!sk_wmem_schedule(sk, copy))
 993                         goto wait_for_memory;
 994
 995                 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
 996                                                pfrag->page,
 997                                                pfrag->offset,
 998                                                copy);
 999                 if (err)
1000                         goto out_error;
1001
1002                 /* Update the skb. */
1003                 if (merge) {
1004                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1005                 } else {
1006                         skb_fill_page_desc(skb, i, pfrag->page,
1007                                            pfrag->offset, copy);
1008                         get_page(pfrag->page);
1009                 }
1010
1011                 pfrag->offset += copy;
1012                 copied += copy;
1013                 if (head != skb) {
1014                         head->len += copy;
1015                         head->data_len += copy;
1016                 }
1017
1018                 continue;
1019
1020 wait_for_memory:
1021                 kcm_push(kcm);
1022                 err = sk_stream_wait_memory(sk, &timeo);
1023                 if (err)
1024                         goto out_error;
1025         }
1026
1027         if (eor) {
1028                 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
1029
1030                 if (head) {
1031                         /* Message complete, queue it on send buffer */
1032                         __skb_queue_tail(&sk->sk_write_queue, head);
1033                         kcm->seq_skb = NULL;
1034                         KCM_STATS_INCR(kcm->stats.tx_msgs);
1035                 }
1036
1037                 if (msg->msg_flags & MSG_BATCH) {
1038                         kcm->tx_wait_more = true;
1039                 } else if (kcm->tx_wait_more || not_busy) {
1040                         err = kcm_write_msgs(kcm);
1041                         if (err < 0) {
1042                                 /* We got a hard error in write_msgs but have
1043                                  * already queued this message. Report an error
1044                                  * in the socket, but don't affect return value
1045                                  * from sendmsg
1046                                  */
1047                                 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
1048                                 report_csk_error(&kcm->sk, -err);
1049                         }
1050                 }
1051         } else {
1052                 /* Message not complete, save state */
1053 partial_message:
1054                 if (head) {
1055                         kcm->seq_skb = head;
1056                         kcm_tx_msg(head)->last_skb = skb;
1057                 }
1058         }
1059
1060         KCM_STATS_ADD(kcm->stats.tx_bytes, copied);
1061
1062         release_sock(sk);
1063         return copied;
1064
1065 out_error:
1066         kcm_push(kcm);
1067
1068         if (sock->type == SOCK_SEQPACKET) {
1069                 /* Wrote some bytes before encountering an
1070                  * error, return partial success.
1071                  */
1072                 if (copied)
1073                         goto partial_message;
1074                 if (head != kcm->seq_skb)
1075                         kfree_skb(head);
1076         } else {
1077                 kfree_skb(head);
1078                 kcm->seq_skb = NULL;
1079         }
1080
1081         err = sk_stream_error(sk, msg->msg_flags, err);
1082
1083         /* make sure we wake any epoll edge trigger waiter */
1084         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1085                 sk->sk_write_space(sk);
1086
1087         release_sock(sk);
1088         return err;
1089 }
1090
1091 static int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
1092                        size_t len, int flags)
1093 {
1094         struct sock *sk = sock->sk;
1095         struct kcm_sock *kcm = kcm_sk(sk);
1096         int err = 0;
1097         struct strp_msg *stm;
1098         int copied = 0;
1099         struct sk_buff *skb;
1100
1101         skb = skb_recv_datagram(sk, flags, &err);
1102         if (!skb)
1103                 goto out;
1104
1105         /* Okay, have a message on the receive queue */
1106
1107         stm = strp_msg(skb);
1108
1109         if (len > stm->full_len)
1110                 len = stm->full_len;
1111
1112         err = skb_copy_datagram_msg(skb, stm->offset, msg, len);
1113         if (err < 0)
1114                 goto out;
1115
1116         copied = len;
1117         if (likely(!(flags & MSG_PEEK))) {
1118                 KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1119                 if (copied < stm->full_len) {
1120                         if (sock->type == SOCK_DGRAM) {
1121                                 /* Truncated message */
1122                                 msg->msg_flags |= MSG_TRUNC;
1123                                 goto msg_finished;
1124                         }
1125                         stm->offset += copied;
1126                         stm->full_len -= copied;
1127                 } else {
1128 msg_finished:
1129                         /* Finished with message */
1130                         msg->msg_flags |= MSG_EOR;
1131                         KCM_STATS_INCR(kcm->stats.rx_msgs);
1132                 }
1133         }
1134
1135 out:
1136         skb_free_datagram(sk, skb);
1137         return copied ? : err;
1138 }
1139
1140 static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1141                                struct pipe_inode_info *pipe, size_t len,
1142                                unsigned int flags)
1143 {
1144         struct sock *sk = sock->sk;
1145         struct kcm_sock *kcm = kcm_sk(sk);
1146         struct strp_msg *stm;
1147         int err = 0;
1148         ssize_t copied;
1149         struct sk_buff *skb;
1150
1151         /* Only support splice for SOCKSEQPACKET */
1152
1153         skb = skb_recv_datagram(sk, flags, &err);
1154         if (!skb)
1155                 goto err_out;
1156
1157         /* Okay, have a message on the receive queue */
1158
1159         stm = strp_msg(skb);
1160
1161         if (len > stm->full_len)
1162                 len = stm->full_len;
1163
1164         copied = skb_splice_bits(skb, sk, stm->offset, pipe, len, flags);
1165         if (copied < 0) {
1166                 err = copied;
1167                 goto err_out;
1168         }
1169
1170         KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1171
1172         stm->offset += copied;
1173         stm->full_len -= copied;
1174
1175         /* We have no way to return MSG_EOR. If all the bytes have been
1176          * read we still leave the message in the receive socket buffer.
1177          * A subsequent recvmsg needs to be done to return MSG_EOR and
1178          * finish reading the message.
1179          */
1180
1181         skb_free_datagram(sk, skb);
1182         return copied;
1183
1184 err_out:
1185         skb_free_datagram(sk, skb);
1186         return err;
1187 }
1188
1189 /* kcm sock lock held */
1190 static void kcm_recv_disable(struct kcm_sock *kcm)
1191 {
1192         struct kcm_mux *mux = kcm->mux;
1193
1194         if (kcm->rx_disabled)
1195                 return;
1196
1197         spin_lock_bh(&mux->rx_lock);
1198
1199         kcm->rx_disabled = 1;
1200
1201         /* If a psock is reserved we'll do cleanup in unreserve */
1202         if (!kcm->rx_psock) {
1203                 if (kcm->rx_wait) {
1204                         list_del(&kcm->wait_rx_list);
1205                         /* paired with lockless reads in kcm_rfree() */
1206                         WRITE_ONCE(kcm->rx_wait, false);
1207                 }
1208
1209                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1210         }
1211
1212         spin_unlock_bh(&mux->rx_lock);
1213 }
1214
1215 /* kcm sock lock held */
1216 static void kcm_recv_enable(struct kcm_sock *kcm)
1217 {
1218         struct kcm_mux *mux = kcm->mux;
1219
1220         if (!kcm->rx_disabled)
1221                 return;
1222
1223         spin_lock_bh(&mux->rx_lock);
1224
1225         kcm->rx_disabled = 0;
1226         kcm_rcv_ready(kcm);
1227
1228         spin_unlock_bh(&mux->rx_lock);
1229 }
1230
1231 static int kcm_setsockopt(struct socket *sock, int level, int optname,
1232                           sockptr_t optval, unsigned int optlen)
1233 {
1234         struct kcm_sock *kcm = kcm_sk(sock->sk);
1235         int val, valbool;
1236         int err = 0;
1237
1238         if (level != SOL_KCM)
1239                 return -ENOPROTOOPT;
1240
1241         if (optlen < sizeof(int))
1242                 return -EINVAL;
1243
1244         if (copy_from_sockptr(&val, optval, sizeof(int)))
1245                 return -EFAULT;
1246
1247         valbool = val ? 1 : 0;
1248
1249         switch (optname) {
1250         case KCM_RECV_DISABLE:
1251                 lock_sock(&kcm->sk);
1252                 if (valbool)
1253                         kcm_recv_disable(kcm);
1254                 else
1255                         kcm_recv_enable(kcm);
1256                 release_sock(&kcm->sk);
1257                 break;
1258         default:
1259                 err = -ENOPROTOOPT;
1260         }
1261
1262         return err;
1263 }
1264
1265 static int kcm_getsockopt(struct socket *sock, int level, int optname,
1266                           char __user *optval, int __user *optlen)
1267 {
1268         struct kcm_sock *kcm = kcm_sk(sock->sk);
1269         int val, len;
1270
1271         if (level != SOL_KCM)
1272                 return -ENOPROTOOPT;
1273
1274         if (get_user(len, optlen))
1275                 return -EFAULT;
1276
1277         if (len < 0)
1278                 return -EINVAL;
1279
1280         len = min_t(unsigned int, len, sizeof(int));
1281
1282         switch (optname) {
1283         case KCM_RECV_DISABLE:
1284                 val = kcm->rx_disabled;
1285                 break;
1286         default:
1287                 return -ENOPROTOOPT;
1288         }
1289
1290         if (put_user(len, optlen))
1291                 return -EFAULT;
1292         if (copy_to_user(optval, &val, len))
1293                 return -EFAULT;
1294         return 0;
1295 }
1296
1297 static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1298 {
1299         struct kcm_sock *tkcm;
1300         struct list_head *head;
1301         int index = 0;
1302
1303         /* For SOCK_SEQPACKET sock type, datagram_poll checks the sk_state, so
1304          * we set sk_state, otherwise epoll_wait always returns right away with
1305          * EPOLLHUP
1306          */
1307         kcm->sk.sk_state = TCP_ESTABLISHED;
1308
1309         /* Add to mux's kcm sockets list */
1310         kcm->mux = mux;
1311         spin_lock_bh(&mux->lock);
1312
1313         head = &mux->kcm_socks;
1314         list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1315                 if (tkcm->index != index)
1316                         break;
1317                 head = &tkcm->kcm_sock_list;
1318                 index++;
1319         }
1320
1321         list_add(&kcm->kcm_sock_list, head);
1322         kcm->index = index;
1323
1324         mux->kcm_socks_cnt++;
1325         spin_unlock_bh(&mux->lock);
1326
1327         INIT_WORK(&kcm->tx_work, kcm_tx_work);
1328
1329         spin_lock_bh(&mux->rx_lock);
1330         kcm_rcv_ready(kcm);
1331         spin_unlock_bh(&mux->rx_lock);
1332 }
1333
1334 static int kcm_attach(struct socket *sock, struct socket *csock,
1335                       struct bpf_prog *prog)
1336 {
1337         struct kcm_sock *kcm = kcm_sk(sock->sk);
1338         struct kcm_mux *mux = kcm->mux;
1339         struct sock *csk;
1340         struct kcm_psock *psock = NULL, *tpsock;
1341         struct list_head *head;
1342         int index = 0;
1343         static const struct strp_callbacks cb = {
1344                 .rcv_msg = kcm_rcv_strparser,
1345                 .parse_msg = kcm_parse_func_strparser,
1346                 .read_sock_done = kcm_read_sock_done,
1347         };
1348         int err = 0;
1349
1350         csk = csock->sk;
1351         if (!csk)
1352                 return -EINVAL;
1353
1354         lock_sock(csk);
1355
1356         /* Only allow TCP sockets to be attached for now */
1357         if ((csk->sk_family != AF_INET && csk->sk_family != AF_INET6) ||
1358             csk->sk_protocol != IPPROTO_TCP) {
1359                 err = -EOPNOTSUPP;
1360                 goto out;
1361         }
1362
1363         /* Don't allow listeners or closed sockets */
1364         if (csk->sk_state == TCP_LISTEN || csk->sk_state == TCP_CLOSE) {
1365                 err = -EOPNOTSUPP;
1366                 goto out;
1367         }
1368
1369         psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1370         if (!psock) {
1371                 err = -ENOMEM;
1372                 goto out;
1373         }
1374
1375         psock->mux = mux;
1376         psock->sk = csk;
1377         psock->bpf_prog = prog;
1378
1379         write_lock_bh(&csk->sk_callback_lock);
1380
1381         /* Check if sk_user_data is already by KCM or someone else.
1382          * Must be done under lock to prevent race conditions.
1383          */
1384         if (csk->sk_user_data) {
1385                 write_unlock_bh(&csk->sk_callback_lock);
1386                 kmem_cache_free(kcm_psockp, psock);
1387                 err = -EALREADY;
1388                 goto out;
1389         }
1390
1391         err = strp_init(&psock->strp, csk, &cb);
1392         if (err) {
1393                 write_unlock_bh(&csk->sk_callback_lock);
1394                 kmem_cache_free(kcm_psockp, psock);
1395                 goto out;
1396         }
1397
1398         psock->save_data_ready = csk->sk_data_ready;
1399         psock->save_write_space = csk->sk_write_space;
1400         psock->save_state_change = csk->sk_state_change;
1401         csk->sk_user_data = psock;
1402         csk->sk_data_ready = psock_data_ready;
1403         csk->sk_write_space = psock_write_space;
1404         csk->sk_state_change = psock_state_change;
1405
1406         write_unlock_bh(&csk->sk_callback_lock);
1407
1408         sock_hold(csk);
1409
1410         /* Finished initialization, now add the psock to the MUX. */
1411         spin_lock_bh(&mux->lock);
1412         head = &mux->psocks;
1413         list_for_each_entry(tpsock, &mux->psocks, psock_list) {
1414                 if (tpsock->index != index)
1415                         break;
1416                 head = &tpsock->psock_list;
1417                 index++;
1418         }
1419
1420         list_add(&psock->psock_list, head);
1421         psock->index = index;
1422
1423         KCM_STATS_INCR(mux->stats.psock_attach);
1424         mux->psocks_cnt++;
1425         psock_now_avail(psock);
1426         spin_unlock_bh(&mux->lock);
1427
1428         /* Schedule RX work in case there are already bytes queued */
1429         strp_check_rcv(&psock->strp);
1430
1431 out:
1432         release_sock(csk);
1433
1434         return err;
1435 }
1436
1437 static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1438 {
1439         struct socket *csock;
1440         struct bpf_prog *prog;
1441         int err;
1442
1443         csock = sockfd_lookup(info->fd, &err);
1444         if (!csock)
1445                 return -ENOENT;
1446
1447         prog = bpf_prog_get_type(info->bpf_fd, BPF_PROG_TYPE_SOCKET_FILTER);
1448         if (IS_ERR(prog)) {
1449                 err = PTR_ERR(prog);
1450                 goto out;
1451         }
1452
1453         err = kcm_attach(sock, csock, prog);
1454         if (err) {
1455                 bpf_prog_put(prog);
1456                 goto out;
1457         }
1458
1459         /* Keep reference on file also */
1460
1461         return 0;
1462 out:
1463         sockfd_put(csock);
1464         return err;
1465 }
1466
1467 static void kcm_unattach(struct kcm_psock *psock)
1468 {
1469         struct sock *csk = psock->sk;
1470         struct kcm_mux *mux = psock->mux;
1471
1472         lock_sock(csk);
1473
1474         /* Stop getting callbacks from TCP socket. After this there should
1475          * be no way to reserve a kcm for this psock.
1476          */
1477         write_lock_bh(&csk->sk_callback_lock);
1478         csk->sk_user_data = NULL;
1479         csk->sk_data_ready = psock->save_data_ready;
1480         csk->sk_write_space = psock->save_write_space;
1481         csk->sk_state_change = psock->save_state_change;
1482         strp_stop(&psock->strp);
1483
1484         if (WARN_ON(psock->rx_kcm)) {
1485                 write_unlock_bh(&csk->sk_callback_lock);
1486                 release_sock(csk);
1487                 return;
1488         }
1489
1490         spin_lock_bh(&mux->rx_lock);
1491
1492         /* Stop receiver activities. After this point psock should not be
1493          * able to get onto ready list either through callbacks or work.
1494          */
1495         if (psock->ready_rx_msg) {
1496                 list_del(&psock->psock_ready_list);
1497                 kfree_skb(psock->ready_rx_msg);
1498                 psock->ready_rx_msg = NULL;
1499                 KCM_STATS_INCR(mux->stats.rx_ready_drops);
1500         }
1501
1502         spin_unlock_bh(&mux->rx_lock);
1503
1504         write_unlock_bh(&csk->sk_callback_lock);
1505
1506         /* Call strp_done without sock lock */
1507         release_sock(csk);
1508         strp_done(&psock->strp);
1509         lock_sock(csk);
1510
1511         bpf_prog_put(psock->bpf_prog);
1512
1513         spin_lock_bh(&mux->lock);
1514
1515         aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1516         save_strp_stats(&psock->strp, &mux->aggregate_strp_stats);
1517
1518         KCM_STATS_INCR(mux->stats.psock_unattach);
1519
1520         if (psock->tx_kcm) {
1521                 /* psock was reserved.  Just mark it finished and we will clean
1522                  * up in the kcm paths, we need kcm lock which can not be
1523                  * acquired here.
1524                  */
1525                 KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1526                 spin_unlock_bh(&mux->lock);
1527
1528                 /* We are unattaching a socket that is reserved. Abort the
1529                  * socket since we may be out of sync in sending on it. We need
1530                  * to do this without the mux lock.
1531                  */
1532                 kcm_abort_tx_psock(psock, EPIPE, false);
1533
1534                 spin_lock_bh(&mux->lock);
1535                 if (!psock->tx_kcm) {
1536                         /* psock now unreserved in window mux was unlocked */
1537                         goto no_reserved;
1538                 }
1539                 psock->done = 1;
1540
1541                 /* Commit done before queuing work to process it */
1542                 smp_mb();
1543
1544                 /* Queue tx work to make sure psock->done is handled */
1545                 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
1546                 spin_unlock_bh(&mux->lock);
1547         } else {
1548 no_reserved:
1549                 if (!psock->tx_stopped)
1550                         list_del(&psock->psock_avail_list);
1551                 list_del(&psock->psock_list);
1552                 mux->psocks_cnt--;
1553                 spin_unlock_bh(&mux->lock);
1554
1555                 sock_put(csk);
1556                 fput(csk->sk_socket->file);
1557                 kmem_cache_free(kcm_psockp, psock);
1558         }
1559
1560         release_sock(csk);
1561 }
1562
1563 static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1564 {
1565         struct kcm_sock *kcm = kcm_sk(sock->sk);
1566         struct kcm_mux *mux = kcm->mux;
1567         struct kcm_psock *psock;
1568         struct socket *csock;
1569         struct sock *csk;
1570         int err;
1571
1572         csock = sockfd_lookup(info->fd, &err);
1573         if (!csock)
1574                 return -ENOENT;
1575
1576         csk = csock->sk;
1577         if (!csk) {
1578                 err = -EINVAL;
1579                 goto out;
1580         }
1581
1582         err = -ENOENT;
1583
1584         spin_lock_bh(&mux->lock);
1585
1586         list_for_each_entry(psock, &mux->psocks, psock_list) {
1587                 if (psock->sk != csk)
1588                         continue;
1589
1590                 /* Found the matching psock */
1591
1592                 if (psock->unattaching || WARN_ON(psock->done)) {
1593                         err = -EALREADY;
1594                         break;
1595                 }
1596
1597                 psock->unattaching = 1;
1598
1599                 spin_unlock_bh(&mux->lock);
1600
1601                 /* Lower socket lock should already be held */
1602                 kcm_unattach(psock);
1603
1604                 err = 0;
1605                 goto out;
1606         }
1607
1608         spin_unlock_bh(&mux->lock);
1609
1610 out:
1611         sockfd_put(csock);
1612         return err;
1613 }
1614
1615 static struct proto kcm_proto = {
1616         .name   = "KCM",
1617         .owner  = THIS_MODULE,
1618         .obj_size = sizeof(struct kcm_sock),
1619 };
1620
1621 /* Clone a kcm socket. */
1622 static struct file *kcm_clone(struct socket *osock)
1623 {
1624         struct socket *newsock;
1625         struct sock *newsk;
1626
1627         newsock = sock_alloc();
1628         if (!newsock)
1629                 return ERR_PTR(-ENFILE);
1630
1631         newsock->type = osock->type;
1632         newsock->ops = osock->ops;
1633
1634         __module_get(newsock->ops->owner);
1635
1636         newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1637                          &kcm_proto, false);
1638         if (!newsk) {
1639                 sock_release(newsock);
1640                 return ERR_PTR(-ENOMEM);
1641         }
1642         sock_init_data(newsock, newsk);
1643         init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1644
1645         return sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1646 }
1647
1648 static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
1649 {
1650         int err;
1651
1652         switch (cmd) {
1653         case SIOCKCMATTACH: {
1654                 struct kcm_attach info;
1655
1656                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1657                         return -EFAULT;
1658
1659                 err = kcm_attach_ioctl(sock, &info);
1660
1661                 break;
1662         }
1663         case SIOCKCMUNATTACH: {
1664                 struct kcm_unattach info;
1665
1666                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1667                         return -EFAULT;
1668
1669                 err = kcm_unattach_ioctl(sock, &info);
1670
1671                 break;
1672         }
1673         case SIOCKCMCLONE: {
1674                 struct kcm_clone info;
1675                 struct file *file;
1676
1677                 info.fd = get_unused_fd_flags(0);
1678                 if (unlikely(info.fd < 0))
1679                         return info.fd;
1680
1681                 file = kcm_clone(sock);
1682                 if (IS_ERR(file)) {
1683                         put_unused_fd(info.fd);
1684                         return PTR_ERR(file);
1685                 }
1686                 if (copy_to_user((void __user *)arg, &info,
1687                                  sizeof(info))) {
1688                         put_unused_fd(info.fd);
1689                         fput(file);
1690                         return -EFAULT;
1691                 }
1692                 fd_install(info.fd, file);
1693                 err = 0;
1694                 break;
1695         }
1696         default:
1697                 err = -ENOIOCTLCMD;
1698                 break;
1699         }
1700
1701         return err;
1702 }
1703
1704 static void free_mux(struct rcu_head *rcu)
1705 {
1706         struct kcm_mux *mux = container_of(rcu,
1707             struct kcm_mux, rcu);
1708
1709         kmem_cache_free(kcm_muxp, mux);
1710 }
1711
1712 static void release_mux(struct kcm_mux *mux)
1713 {
1714         struct kcm_net *knet = mux->knet;
1715         struct kcm_psock *psock, *tmp_psock;
1716
1717         /* Release psocks */
1718         list_for_each_entry_safe(psock, tmp_psock,
1719                                  &mux->psocks, psock_list) {
1720                 if (!WARN_ON(psock->unattaching))
1721                         kcm_unattach(psock);
1722         }
1723
1724         if (WARN_ON(mux->psocks_cnt))
1725                 return;
1726
1727         __skb_queue_purge(&mux->rx_hold_queue);
1728
1729         mutex_lock(&knet->mutex);
1730         aggregate_mux_stats(&mux->stats, &knet->aggregate_mux_stats);
1731         aggregate_psock_stats(&mux->aggregate_psock_stats,
1732                               &knet->aggregate_psock_stats);
1733         aggregate_strp_stats(&mux->aggregate_strp_stats,
1734                              &knet->aggregate_strp_stats);
1735         list_del_rcu(&mux->kcm_mux_list);
1736         knet->count--;
1737         mutex_unlock(&knet->mutex);
1738
1739         call_rcu(&mux->rcu, free_mux);
1740 }
1741
1742 static void kcm_done(struct kcm_sock *kcm)
1743 {
1744         struct kcm_mux *mux = kcm->mux;
1745         struct sock *sk = &kcm->sk;
1746         int socks_cnt;
1747
1748         spin_lock_bh(&mux->rx_lock);
1749         if (kcm->rx_psock) {
1750                 /* Cleanup in unreserve_rx_kcm */
1751                 WARN_ON(kcm->done);
1752                 kcm->rx_disabled = 1;
1753                 kcm->done = 1;
1754                 spin_unlock_bh(&mux->rx_lock);
1755                 return;
1756         }
1757
1758         if (kcm->rx_wait) {
1759                 list_del(&kcm->wait_rx_list);
1760                 /* paired with lockless reads in kcm_rfree() */
1761                 WRITE_ONCE(kcm->rx_wait, false);
1762         }
1763         /* Move any pending receive messages to other kcm sockets */
1764         requeue_rx_msgs(mux, &sk->sk_receive_queue);
1765
1766         spin_unlock_bh(&mux->rx_lock);
1767
1768         if (WARN_ON(sk_rmem_alloc_get(sk)))
1769                 return;
1770
1771         /* Detach from MUX */
1772         spin_lock_bh(&mux->lock);
1773
1774         list_del(&kcm->kcm_sock_list);
1775         mux->kcm_socks_cnt--;
1776         socks_cnt = mux->kcm_socks_cnt;
1777
1778         spin_unlock_bh(&mux->lock);
1779
1780         if (!socks_cnt) {
1781                 /* We are done with the mux now. */
1782                 release_mux(mux);
1783         }
1784
1785         WARN_ON(kcm->rx_wait);
1786
1787         sock_put(&kcm->sk);
1788 }
1789
1790 /* Called by kcm_release to close a KCM socket.
1791  * If this is the last KCM socket on the MUX, destroy the MUX.
1792  */
1793 static int kcm_release(struct socket *sock)
1794 {
1795         struct sock *sk = sock->sk;
1796         struct kcm_sock *kcm;
1797         struct kcm_mux *mux;
1798         struct kcm_psock *psock;
1799
1800         if (!sk)
1801                 return 0;
1802
1803         kcm = kcm_sk(sk);
1804         mux = kcm->mux;
1805
1806         lock_sock(sk);
1807         sock_orphan(sk);
1808         kfree_skb(kcm->seq_skb);
1809
1810         /* Purge queue under lock to avoid race condition with tx_work trying
1811          * to act when queue is nonempty. If tx_work runs after this point
1812          * it will just return.
1813          */
1814         __skb_queue_purge(&sk->sk_write_queue);
1815
1816         /* Set tx_stopped. This is checked when psock is bound to a kcm and we
1817          * get a writespace callback. This prevents further work being queued
1818          * from the callback (unbinding the psock occurs after canceling work.
1819          */
1820         kcm->tx_stopped = 1;
1821
1822         release_sock(sk);
1823
1824         spin_lock_bh(&mux->lock);
1825         if (kcm->tx_wait) {
1826                 /* Take of tx_wait list, after this point there should be no way
1827                  * that a psock will be assigned to this kcm.
1828                  */
1829                 list_del(&kcm->wait_psock_list);
1830                 kcm->tx_wait = false;
1831         }
1832         spin_unlock_bh(&mux->lock);
1833
1834         /* Cancel work. After this point there should be no outside references
1835          * to the kcm socket.
1836          */
1837         cancel_work_sync(&kcm->tx_work);
1838
1839         lock_sock(sk);
1840         psock = kcm->tx_psock;
1841         if (psock) {
1842                 /* A psock was reserved, so we need to kill it since it
1843                  * may already have some bytes queued from a message. We
1844                  * need to do this after removing kcm from tx_wait list.
1845                  */
1846                 kcm_abort_tx_psock(psock, EPIPE, false);
1847                 unreserve_psock(kcm);
1848         }
1849         release_sock(sk);
1850
1851         WARN_ON(kcm->tx_wait);
1852         WARN_ON(kcm->tx_psock);
1853
1854         sock->sk = NULL;
1855
1856         kcm_done(kcm);
1857
1858         return 0;
1859 }
1860
1861 static const struct proto_ops kcm_dgram_ops = {
1862         .family =       PF_KCM,
1863         .owner =        THIS_MODULE,
1864         .release =      kcm_release,
1865         .bind =         sock_no_bind,
1866         .connect =      sock_no_connect,
1867         .socketpair =   sock_no_socketpair,
1868         .accept =       sock_no_accept,
1869         .getname =      sock_no_getname,
1870         .poll =         datagram_poll,
1871         .ioctl =        kcm_ioctl,
1872         .listen =       sock_no_listen,
1873         .shutdown =     sock_no_shutdown,
1874         .setsockopt =   kcm_setsockopt,
1875         .getsockopt =   kcm_getsockopt,
1876         .sendmsg =      kcm_sendmsg,
1877         .recvmsg =      kcm_recvmsg,
1878         .mmap =         sock_no_mmap,
1879         .sendpage =     kcm_sendpage,
1880 };
1881
1882 static const struct proto_ops kcm_seqpacket_ops = {
1883         .family =       PF_KCM,
1884         .owner =        THIS_MODULE,
1885         .release =      kcm_release,
1886         .bind =         sock_no_bind,
1887         .connect =      sock_no_connect,
1888         .socketpair =   sock_no_socketpair,
1889         .accept =       sock_no_accept,
1890         .getname =      sock_no_getname,
1891         .poll =         datagram_poll,
1892         .ioctl =        kcm_ioctl,
1893         .listen =       sock_no_listen,
1894         .shutdown =     sock_no_shutdown,
1895         .setsockopt =   kcm_setsockopt,
1896         .getsockopt =   kcm_getsockopt,
1897         .sendmsg =      kcm_sendmsg,
1898         .recvmsg =      kcm_recvmsg,
1899         .mmap =         sock_no_mmap,
1900         .sendpage =     kcm_sendpage,
1901         .splice_read =  kcm_splice_read,
1902 };
1903
1904 /* Create proto operation for kcm sockets */
1905 static int kcm_create(struct net *net, struct socket *sock,
1906                       int protocol, int kern)
1907 {
1908         struct kcm_net *knet = net_generic(net, kcm_net_id);
1909         struct sock *sk;
1910         struct kcm_mux *mux;
1911
1912         switch (sock->type) {
1913         case SOCK_DGRAM:
1914                 sock->ops = &kcm_dgram_ops;
1915                 break;
1916         case SOCK_SEQPACKET:
1917                 sock->ops = &kcm_seqpacket_ops;
1918                 break;
1919         default:
1920                 return -ESOCKTNOSUPPORT;
1921         }
1922
1923         if (protocol != KCMPROTO_CONNECTED)
1924                 return -EPROTONOSUPPORT;
1925
1926         sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
1927         if (!sk)
1928                 return -ENOMEM;
1929
1930         /* Allocate a kcm mux, shared between KCM sockets */
1931         mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
1932         if (!mux) {
1933                 sk_free(sk);
1934                 return -ENOMEM;
1935         }
1936
1937         spin_lock_init(&mux->lock);
1938         spin_lock_init(&mux->rx_lock);
1939         INIT_LIST_HEAD(&mux->kcm_socks);
1940         INIT_LIST_HEAD(&mux->kcm_rx_waiters);
1941         INIT_LIST_HEAD(&mux->kcm_tx_waiters);
1942
1943         INIT_LIST_HEAD(&mux->psocks);
1944         INIT_LIST_HEAD(&mux->psocks_ready);
1945         INIT_LIST_HEAD(&mux->psocks_avail);
1946
1947         mux->knet = knet;
1948
1949         /* Add new MUX to list */
1950         mutex_lock(&knet->mutex);
1951         list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
1952         knet->count++;
1953         mutex_unlock(&knet->mutex);
1954
1955         skb_queue_head_init(&mux->rx_hold_queue);
1956
1957         /* Init KCM socket */
1958         sock_init_data(sock, sk);
1959         init_kcm_sock(kcm_sk(sk), mux);
1960
1961         return 0;
1962 }
1963
1964 static const struct net_proto_family kcm_family_ops = {
1965         .family = PF_KCM,
1966         .create = kcm_create,
1967         .owner  = THIS_MODULE,
1968 };
1969
1970 static __net_init int kcm_init_net(struct net *net)
1971 {
1972         struct kcm_net *knet = net_generic(net, kcm_net_id);
1973
1974         INIT_LIST_HEAD_RCU(&knet->mux_list);
1975         mutex_init(&knet->mutex);
1976
1977         return 0;
1978 }
1979
1980 static __net_exit void kcm_exit_net(struct net *net)
1981 {
1982         struct kcm_net *knet = net_generic(net, kcm_net_id);
1983
1984         /* All KCM sockets should be closed at this point, which should mean
1985          * that all multiplexors and psocks have been destroyed.
1986          */
1987         WARN_ON(!list_empty(&knet->mux_list));
1988
1989         mutex_destroy(&knet->mutex);
1990 }
1991
1992 static struct pernet_operations kcm_net_ops = {
1993         .init = kcm_init_net,
1994         .exit = kcm_exit_net,
1995         .id   = &kcm_net_id,
1996         .size = sizeof(struct kcm_net),
1997 };
1998
1999 static int __init kcm_init(void)
2000 {
2001         int err = -ENOMEM;
2002
2003         kcm_muxp = kmem_cache_create("kcm_mux_cache",
2004                                      sizeof(struct kcm_mux), 0,
2005                                      SLAB_HWCACHE_ALIGN, NULL);
2006         if (!kcm_muxp)
2007                 goto fail;
2008
2009         kcm_psockp = kmem_cache_create("kcm_psock_cache",
2010                                        sizeof(struct kcm_psock), 0,
2011                                         SLAB_HWCACHE_ALIGN, NULL);
2012         if (!kcm_psockp)
2013                 goto fail;
2014
2015         kcm_wq = create_singlethread_workqueue("kkcmd");
2016         if (!kcm_wq)
2017                 goto fail;
2018
2019         err = proto_register(&kcm_proto, 1);
2020         if (err)
2021                 goto fail;
2022
2023         err = register_pernet_device(&kcm_net_ops);
2024         if (err)
2025                 goto net_ops_fail;
2026
2027         err = sock_register(&kcm_family_ops);
2028         if (err)
2029                 goto sock_register_fail;
2030
2031         err = kcm_proc_init();
2032         if (err)
2033                 goto proc_init_fail;
2034
2035         return 0;
2036
2037 proc_init_fail:
2038         sock_unregister(PF_KCM);
2039
2040 sock_register_fail:
2041         unregister_pernet_device(&kcm_net_ops);
2042
2043 net_ops_fail:
2044         proto_unregister(&kcm_proto);
2045
2046 fail:
2047         kmem_cache_destroy(kcm_muxp);
2048         kmem_cache_destroy(kcm_psockp);
2049
2050         if (kcm_wq)
2051                 destroy_workqueue(kcm_wq);
2052
2053         return err;
2054 }
2055
2056 static void __exit kcm_exit(void)
2057 {
2058         kcm_proc_exit();
2059         sock_unregister(PF_KCM);
2060         unregister_pernet_device(&kcm_net_ops);
2061         proto_unregister(&kcm_proto);
2062         destroy_workqueue(kcm_wq);
2063
2064         kmem_cache_destroy(kcm_muxp);
2065         kmem_cache_destroy(kcm_psockp);
2066 }
2067
2068 module_init(kcm_init);
2069 module_exit(kcm_exit);
2070
2071 MODULE_LICENSE("GPL");
2072 MODULE_ALIAS_NETPROTO(PF_KCM);