net/kcm/kcmsock.c

   1 /*
   2  * Kernel Connection Multiplexor
   3  *
   4  * Copyright (c) 2016 Tom Herbert <tom@herbertland.com>
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2
   8  * as published by the Free Software Foundation.
   9  */
  10
  11 #include <linux/bpf.h>
  12 #include <linux/errno.h>
  13 #include <linux/errqueue.h>
  14 #include <linux/file.h>
  15 #include <linux/in.h>
  16 #include <linux/kernel.h>
  17 #include <linux/module.h>
  18 #include <linux/net.h>
  19 #include <linux/netdevice.h>
  20 #include <linux/poll.h>
  21 #include <linux/rculist.h>
  22 #include <linux/skbuff.h>
  23 #include <linux/socket.h>
  24 #include <linux/uaccess.h>
  25 #include <linux/workqueue.h>
  26 #include <linux/syscalls.h>
  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         csk->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.rx_bytes -
  98                        psock->saved_rx_bytes);
  99         mux->stats.rx_msgs +=
 100                 psock->strp.stats.rx_msgs - psock->saved_rx_msgs;
 101         psock->saved_rx_msgs = psock->strp.stats.rx_msgs;
 102         psock->saved_rx_bytes = psock->strp.stats.rx_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         kcm->rx_wait = true;
 166 }
 167
 168 static void kcm_rfree(struct sk_buff *skb)
 169 {
 170         struct sock *sk = skb->sk;
 171         struct kcm_sock *kcm = kcm_sk(sk);
 172         struct kcm_mux *mux = kcm->mux;
 173         unsigned int len = skb->truesize;
 174
 175         sk_mem_uncharge(sk, len);
 176         atomic_sub(len, &sk->sk_rmem_alloc);
 177
 178         /* For reading rx_wait and rx_psock without holding lock */
 179         smp_mb__after_atomic();
 180
 181         if (!kcm->rx_wait && !kcm->rx_psock &&
 182             sk_rmem_alloc_get(sk) < sk->sk_rcvlowat) {
 183                 spin_lock_bh(&mux->rx_lock);
 184                 kcm_rcv_ready(kcm);
 185                 spin_unlock_bh(&mux->rx_lock);
 186         }
 187 }
 188
 189 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb)
 190 {
 191         struct sk_buff_head *list = &sk->sk_receive_queue;
 192
 193         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
 194                 return -ENOMEM;
 195
 196         if (!sk_rmem_schedule(sk, skb, skb->truesize))
 197                 return -ENOBUFS;
 198
 199         skb->dev = NULL;
 200
 201         skb_orphan(skb);
 202         skb->sk = sk;
 203         skb->destructor = kcm_rfree;
 204         atomic_add(skb->truesize, &sk->sk_rmem_alloc);
 205         sk_mem_charge(sk, skb->truesize);
 206
 207         skb_queue_tail(list, skb);
 208
 209         if (!sock_flag(sk, SOCK_DEAD))
 210                 sk->sk_data_ready(sk);
 211
 212         return 0;
 213 }
 214
 215 /* Requeue received messages for a kcm socket to other kcm sockets. This is
 216  * called with a kcm socket is receive disabled.
 217  * RX mux lock held.
 218  */
 219 static void requeue_rx_msgs(struct kcm_mux *mux, struct sk_buff_head *head)
 220 {
 221         struct sk_buff *skb;
 222         struct kcm_sock *kcm;
 223
 224         while ((skb = __skb_dequeue(head))) {
 225                 /* Reset destructor to avoid calling kcm_rcv_ready */
 226                 skb->destructor = sock_rfree;
 227                 skb_orphan(skb);
 228 try_again:
 229                 if (list_empty(&mux->kcm_rx_waiters)) {
 230                         skb_queue_tail(&mux->rx_hold_queue, skb);
 231                         continue;
 232                 }
 233
 234                 kcm = list_first_entry(&mux->kcm_rx_waiters,
 235                                        struct kcm_sock, wait_rx_list);
 236
 237                 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 238                         /* Should mean socket buffer full */
 239                         list_del(&kcm->wait_rx_list);
 240                         kcm->rx_wait = false;
 241
 242                         /* Commit rx_wait to read in kcm_free */
 243                         smp_wmb();
 244
 245                         goto try_again;
 246                 }
 247         }
 248 }
 249
 250 /* Lower sock lock held */
 251 static struct kcm_sock *reserve_rx_kcm(struct kcm_psock *psock,
 252                                        struct sk_buff *head)
 253 {
 254         struct kcm_mux *mux = psock->mux;
 255         struct kcm_sock *kcm;
 256
 257         WARN_ON(psock->ready_rx_msg);
 258
 259         if (psock->rx_kcm)
 260                 return psock->rx_kcm;
 261
 262         spin_lock_bh(&mux->rx_lock);
 263
 264         if (psock->rx_kcm) {
 265                 spin_unlock_bh(&mux->rx_lock);
 266                 return psock->rx_kcm;
 267         }
 268
 269         kcm_update_rx_mux_stats(mux, psock);
 270
 271         if (list_empty(&mux->kcm_rx_waiters)) {
 272                 psock->ready_rx_msg = head;
 273                 strp_pause(&psock->strp);
 274                 list_add_tail(&psock->psock_ready_list,
 275                               &mux->psocks_ready);
 276                 spin_unlock_bh(&mux->rx_lock);
 277                 return NULL;
 278         }
 279
 280         kcm = list_first_entry(&mux->kcm_rx_waiters,
 281                                struct kcm_sock, wait_rx_list);
 282         list_del(&kcm->wait_rx_list);
 283         kcm->rx_wait = false;
 284
 285         psock->rx_kcm = kcm;
 286         kcm->rx_psock = psock;
 287
 288         spin_unlock_bh(&mux->rx_lock);
 289
 290         return kcm;
 291 }
 292
 293 static void kcm_done(struct kcm_sock *kcm);
 294
 295 static void kcm_done_work(struct work_struct *w)
 296 {
 297         kcm_done(container_of(w, struct kcm_sock, done_work));
 298 }
 299
 300 /* Lower sock held */
 301 static void unreserve_rx_kcm(struct kcm_psock *psock,
 302                              bool rcv_ready)
 303 {
 304         struct kcm_sock *kcm = psock->rx_kcm;
 305         struct kcm_mux *mux = psock->mux;
 306
 307         if (!kcm)
 308                 return;
 309
 310         spin_lock_bh(&mux->rx_lock);
 311
 312         psock->rx_kcm = NULL;
 313         kcm->rx_psock = NULL;
 314
 315         /* Commit kcm->rx_psock before sk_rmem_alloc_get to sync with
 316          * kcm_rfree
 317          */
 318         smp_mb();
 319
 320         if (unlikely(kcm->done)) {
 321                 spin_unlock_bh(&mux->rx_lock);
 322
 323                 /* Need to run kcm_done in a task since we need to qcquire
 324                  * callback locks which may already be held here.
 325                  */
 326                 INIT_WORK(&kcm->done_work, kcm_done_work);
 327                 schedule_work(&kcm->done_work);
 328                 return;
 329         }
 330
 331         if (unlikely(kcm->rx_disabled)) {
 332                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
 333         } else if (rcv_ready || unlikely(!sk_rmem_alloc_get(&kcm->sk))) {
 334                 /* Check for degenerative race with rx_wait that all
 335                  * data was dequeued (accounted for in kcm_rfree).
 336                  */
 337                 kcm_rcv_ready(kcm);
 338         }
 339         spin_unlock_bh(&mux->rx_lock);
 340 }
 341
 342 /* Lower sock lock held */
 343 static void psock_data_ready(struct sock *sk)
 344 {
 345         struct kcm_psock *psock;
 346
 347         read_lock_bh(&sk->sk_callback_lock);
 348
 349         psock = (struct kcm_psock *)sk->sk_user_data;
 350         if (likely(psock))
 351                 strp_data_ready(&psock->strp);
 352
 353         read_unlock_bh(&sk->sk_callback_lock);
 354 }
 355
 356 /* Called with lower sock held */
 357 static void kcm_rcv_strparser(struct strparser *strp, struct sk_buff *skb)
 358 {
 359         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 360         struct kcm_sock *kcm;
 361
 362 try_queue:
 363         kcm = reserve_rx_kcm(psock, skb);
 364         if (!kcm) {
 365                  /* Unable to reserve a KCM, message is held in psock and strp
 366                   * is paused.
 367                   */
 368                 return;
 369         }
 370
 371         if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 372                 /* Should mean socket buffer full */
 373                 unreserve_rx_kcm(psock, false);
 374                 goto try_queue;
 375         }
 376 }
 377
 378 static int kcm_parse_func_strparser(struct strparser *strp, struct sk_buff *skb)
 379 {
 380         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 381         struct bpf_prog *prog = psock->bpf_prog;
 382
 383         return (*prog->bpf_func)(skb, prog->insnsi);
 384 }
 385
 386 static int kcm_read_sock_done(struct strparser *strp, int err)
 387 {
 388         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 389
 390         unreserve_rx_kcm(psock, true);
 391
 392         return err;
 393 }
 394
 395 static void psock_state_change(struct sock *sk)
 396 {
 397         /* TCP only does a POLLIN for a half close. Do a POLLHUP here
 398          * since application will normally not poll with POLLIN
 399          * on the TCP sockets.
 400          */
 401
 402         report_csk_error(sk, EPIPE);
 403 }
 404
 405 static void psock_write_space(struct sock *sk)
 406 {
 407         struct kcm_psock *psock;
 408         struct kcm_mux *mux;
 409         struct kcm_sock *kcm;
 410
 411         read_lock_bh(&sk->sk_callback_lock);
 412
 413         psock = (struct kcm_psock *)sk->sk_user_data;
 414         if (unlikely(!psock))
 415                 goto out;
 416         mux = psock->mux;
 417
 418         spin_lock_bh(&mux->lock);
 419
 420         /* Check if the socket is reserved so someone is waiting for sending. */
 421         kcm = psock->tx_kcm;
 422         if (kcm && !unlikely(kcm->tx_stopped))
 423                 queue_work(kcm_wq, &kcm->tx_work);
 424
 425         spin_unlock_bh(&mux->lock);
 426 out:
 427         read_unlock_bh(&sk->sk_callback_lock);
 428 }
 429
 430 static void unreserve_psock(struct kcm_sock *kcm);
 431
 432 /* kcm sock is locked. */
 433 static struct kcm_psock *reserve_psock(struct kcm_sock *kcm)
 434 {
 435         struct kcm_mux *mux = kcm->mux;
 436         struct kcm_psock *psock;
 437
 438         psock = kcm->tx_psock;
 439
 440         smp_rmb(); /* Must read tx_psock before tx_wait */
 441
 442         if (psock) {
 443                 WARN_ON(kcm->tx_wait);
 444                 if (unlikely(psock->tx_stopped))
 445                         unreserve_psock(kcm);
 446                 else
 447                         return kcm->tx_psock;
 448         }
 449
 450         spin_lock_bh(&mux->lock);
 451
 452         /* Check again under lock to see if psock was reserved for this
 453          * psock via psock_unreserve.
 454          */
 455         psock = kcm->tx_psock;
 456         if (unlikely(psock)) {
 457                 WARN_ON(kcm->tx_wait);
 458                 spin_unlock_bh(&mux->lock);
 459                 return kcm->tx_psock;
 460         }
 461
 462         if (!list_empty(&mux->psocks_avail)) {
 463                 psock = list_first_entry(&mux->psocks_avail,
 464                                          struct kcm_psock,
 465                                          psock_avail_list);
 466                 list_del(&psock->psock_avail_list);
 467                 if (kcm->tx_wait) {
 468                         list_del(&kcm->wait_psock_list);
 469                         kcm->tx_wait = false;
 470                 }
 471                 kcm->tx_psock = psock;
 472                 psock->tx_kcm = kcm;
 473                 KCM_STATS_INCR(psock->stats.reserved);
 474         } else if (!kcm->tx_wait) {
 475                 list_add_tail(&kcm->wait_psock_list,
 476                               &mux->kcm_tx_waiters);
 477                 kcm->tx_wait = true;
 478         }
 479
 480         spin_unlock_bh(&mux->lock);
 481
 482         return psock;
 483 }
 484
 485 /* mux lock held */
 486 static void psock_now_avail(struct kcm_psock *psock)
 487 {
 488         struct kcm_mux *mux = psock->mux;
 489         struct kcm_sock *kcm;
 490
 491         if (list_empty(&mux->kcm_tx_waiters)) {
 492                 list_add_tail(&psock->psock_avail_list,
 493                               &mux->psocks_avail);
 494         } else {
 495                 kcm = list_first_entry(&mux->kcm_tx_waiters,
 496                                        struct kcm_sock,
 497                                        wait_psock_list);
 498                 list_del(&kcm->wait_psock_list);
 499                 kcm->tx_wait = false;
 500                 psock->tx_kcm = kcm;
 501
 502                 /* Commit before changing tx_psock since that is read in
 503                  * reserve_psock before queuing work.
 504                  */
 505                 smp_mb();
 506
 507                 kcm->tx_psock = psock;
 508                 KCM_STATS_INCR(psock->stats.reserved);
 509                 queue_work(kcm_wq, &kcm->tx_work);
 510         }
 511 }
 512
 513 /* kcm sock is locked. */
 514 static void unreserve_psock(struct kcm_sock *kcm)
 515 {
 516         struct kcm_psock *psock;
 517         struct kcm_mux *mux = kcm->mux;
 518
 519         spin_lock_bh(&mux->lock);
 520
 521         psock = kcm->tx_psock;
 522
 523         if (WARN_ON(!psock)) {
 524                 spin_unlock_bh(&mux->lock);
 525                 return;
 526         }
 527
 528         smp_rmb(); /* Read tx_psock before tx_wait */
 529
 530         kcm_update_tx_mux_stats(mux, psock);
 531
 532         WARN_ON(kcm->tx_wait);
 533
 534         kcm->tx_psock = NULL;
 535         psock->tx_kcm = NULL;
 536         KCM_STATS_INCR(psock->stats.unreserved);
 537
 538         if (unlikely(psock->tx_stopped)) {
 539                 if (psock->done) {
 540                         /* Deferred free */
 541                         list_del(&psock->psock_list);
 542                         mux->psocks_cnt--;
 543                         sock_put(psock->sk);
 544                         fput(psock->sk->sk_socket->file);
 545                         kmem_cache_free(kcm_psockp, psock);
 546                 }
 547
 548                 /* Don't put back on available list */
 549
 550                 spin_unlock_bh(&mux->lock);
 551
 552                 return;
 553         }
 554
 555         psock_now_avail(psock);
 556
 557         spin_unlock_bh(&mux->lock);
 558 }
 559
 560 static void kcm_report_tx_retry(struct kcm_sock *kcm)
 561 {
 562         struct kcm_mux *mux = kcm->mux;
 563
 564         spin_lock_bh(&mux->lock);
 565         KCM_STATS_INCR(mux->stats.tx_retries);
 566         spin_unlock_bh(&mux->lock);
 567 }
 568
 569 /* Write any messages ready on the kcm socket.  Called with kcm sock lock
 570  * held.  Return bytes actually sent or error.
 571  */
 572 static int kcm_write_msgs(struct kcm_sock *kcm)
 573 {
 574         struct sock *sk = &kcm->sk;
 575         struct kcm_psock *psock;
 576         struct sk_buff *skb, *head;
 577         struct kcm_tx_msg *txm;
 578         unsigned short fragidx, frag_offset;
 579         unsigned int sent, total_sent = 0;
 580         int ret = 0;
 581
 582         kcm->tx_wait_more = false;
 583         psock = kcm->tx_psock;
 584         if (unlikely(psock && psock->tx_stopped)) {
 585                 /* A reserved psock was aborted asynchronously. Unreserve
 586                  * it and we'll retry the message.
 587                  */
 588                 unreserve_psock(kcm);
 589                 kcm_report_tx_retry(kcm);
 590                 if (skb_queue_empty(&sk->sk_write_queue))
 591                         return 0;
 592
 593                 kcm_tx_msg(skb_peek(&sk->sk_write_queue))->sent = 0;
 594
 595         } else if (skb_queue_empty(&sk->sk_write_queue)) {
 596                 return 0;
 597         }
 598
 599         head = skb_peek(&sk->sk_write_queue);
 600         txm = kcm_tx_msg(head);
 601
 602         if (txm->sent) {
 603                 /* Send of first skbuff in queue already in progress */
 604                 if (WARN_ON(!psock)) {
 605                         ret = -EINVAL;
 606                         goto out;
 607                 }
 608                 sent = txm->sent;
 609                 frag_offset = txm->frag_offset;
 610                 fragidx = txm->fragidx;
 611                 skb = txm->frag_skb;
 612
 613                 goto do_frag;
 614         }
 615
 616 try_again:
 617         psock = reserve_psock(kcm);
 618         if (!psock)
 619                 goto out;
 620
 621         do {
 622                 skb = head;
 623                 txm = kcm_tx_msg(head);
 624                 sent = 0;
 625
 626 do_frag_list:
 627                 if (WARN_ON(!skb_shinfo(skb)->nr_frags)) {
 628                         ret = -EINVAL;
 629                         goto out;
 630                 }
 631
 632                 for (fragidx = 0; fragidx < skb_shinfo(skb)->nr_frags;
 633                      fragidx++) {
 634                         skb_frag_t *frag;
 635
 636                         frag_offset = 0;
 637 do_frag:
 638                         frag = &skb_shinfo(skb)->frags[fragidx];
 639                         if (WARN_ON(!frag->size)) {
 640                                 ret = -EINVAL;
 641                                 goto out;
 642                         }
 643
 644                         ret = kernel_sendpage(psock->sk->sk_socket,
 645                                               frag->page.p,
 646                                               frag->page_offset + frag_offset,
 647                                               frag->size - frag_offset,
 648                                               MSG_DONTWAIT);
 649                         if (ret <= 0) {
 650                                 if (ret == -EAGAIN) {
 651                                         /* Save state to try again when there's
 652                                          * write space on the socket
 653                                          */
 654                                         txm->sent = sent;
 655                                         txm->frag_offset = frag_offset;
 656                                         txm->fragidx = fragidx;
 657                                         txm->frag_skb = skb;
 658
 659                                         ret = 0;
 660                                         goto out;
 661                                 }
 662
 663                                 /* Hard failure in sending message, abort this
 664                                  * psock since it has lost framing
 665                                  * synchonization and retry sending the
 666                                  * message from the beginning.
 667                                  */
 668                                 kcm_abort_tx_psock(psock, ret ? -ret : EPIPE,
 669                                                    true);
 670                                 unreserve_psock(kcm);
 671
 672                                 txm->sent = 0;
 673                                 kcm_report_tx_retry(kcm);
 674                                 ret = 0;
 675
 676                                 goto try_again;
 677                         }
 678
 679                         sent += ret;
 680                         frag_offset += ret;
 681                         KCM_STATS_ADD(psock->stats.tx_bytes, ret);
 682                         if (frag_offset < frag->size) {
 683                                 /* Not finished with this frag */
 684                                 goto do_frag;
 685                         }
 686                 }
 687
 688                 if (skb == head) {
 689                         if (skb_has_frag_list(skb)) {
 690                                 skb = skb_shinfo(skb)->frag_list;
 691                                 goto do_frag_list;
 692                         }
 693                 } else if (skb->next) {
 694                         skb = skb->next;
 695                         goto do_frag_list;
 696                 }
 697
 698                 /* Successfully sent the whole packet, account for it. */
 699                 skb_dequeue(&sk->sk_write_queue);
 700                 kfree_skb(head);
 701                 sk->sk_wmem_queued -= sent;
 702                 total_sent += sent;
 703                 KCM_STATS_INCR(psock->stats.tx_msgs);
 704         } while ((head = skb_peek(&sk->sk_write_queue)));
 705 out:
 706         if (!head) {
 707                 /* Done with all queued messages. */
 708                 WARN_ON(!skb_queue_empty(&sk->sk_write_queue));
 709                 unreserve_psock(kcm);
 710         }
 711
 712         /* Check if write space is available */
 713         sk->sk_write_space(sk);
 714
 715         return total_sent ? : ret;
 716 }
 717
 718 static void kcm_tx_work(struct work_struct *w)
 719 {
 720         struct kcm_sock *kcm = container_of(w, struct kcm_sock, tx_work);
 721         struct sock *sk = &kcm->sk;
 722         int err;
 723
 724         lock_sock(sk);
 725
 726         /* Primarily for SOCK_DGRAM sockets, also handle asynchronous tx
 727          * aborts
 728          */
 729         err = kcm_write_msgs(kcm);
 730         if (err < 0) {
 731                 /* Hard failure in write, report error on KCM socket */
 732                 pr_warn("KCM: Hard failure on kcm_write_msgs %d\n", err);
 733                 report_csk_error(&kcm->sk, -err);
 734                 goto out;
 735         }
 736
 737         /* Primarily for SOCK_SEQPACKET sockets */
 738         if (likely(sk->sk_socket) &&
 739             test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
 740                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 741                 sk->sk_write_space(sk);
 742         }
 743
 744 out:
 745         release_sock(sk);
 746 }
 747
 748 static void kcm_push(struct kcm_sock *kcm)
 749 {
 750         if (kcm->tx_wait_more)
 751                 kcm_write_msgs(kcm);
 752 }
 753
 754 static ssize_t kcm_sendpage(struct socket *sock, struct page *page,
 755                             int offset, size_t size, int flags)
 756
 757 {
 758         struct sock *sk = sock->sk;
 759         struct kcm_sock *kcm = kcm_sk(sk);
 760         struct sk_buff *skb = NULL, *head = NULL;
 761         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 762         bool eor;
 763         int err = 0;
 764         int i;
 765
 766         if (flags & MSG_SENDPAGE_NOTLAST)
 767                 flags |= MSG_MORE;
 768
 769         /* No MSG_EOR from splice, only look at MSG_MORE */
 770         eor = !(flags & MSG_MORE);
 771
 772         lock_sock(sk);
 773
 774         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 775
 776         err = -EPIPE;
 777         if (sk->sk_err)
 778                 goto out_error;
 779
 780         if (kcm->seq_skb) {
 781                 /* Previously opened message */
 782                 head = kcm->seq_skb;
 783                 skb = kcm_tx_msg(head)->last_skb;
 784                 i = skb_shinfo(skb)->nr_frags;
 785
 786                 if (skb_can_coalesce(skb, i, page, offset)) {
 787                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], size);
 788                         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
 789                         goto coalesced;
 790                 }
 791
 792                 if (i >= MAX_SKB_FRAGS) {
 793                         struct sk_buff *tskb;
 794
 795                         tskb = alloc_skb(0, sk->sk_allocation);
 796                         while (!tskb) {
 797                                 kcm_push(kcm);
 798                                 err = sk_stream_wait_memory(sk, &timeo);
 799                                 if (err)
 800                                         goto out_error;
 801                         }
 802
 803                         if (head == skb)
 804                                 skb_shinfo(head)->frag_list = tskb;
 805                         else
 806                                 skb->next = tskb;
 807
 808                         skb = tskb;
 809                         skb->ip_summed = CHECKSUM_UNNECESSARY;
 810                         i = 0;
 811                 }
 812         } else {
 813                 /* Call the sk_stream functions to manage the sndbuf mem. */
 814                 if (!sk_stream_memory_free(sk)) {
 815                         kcm_push(kcm);
 816                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 817                         err = sk_stream_wait_memory(sk, &timeo);
 818                         if (err)
 819                                 goto out_error;
 820                 }
 821
 822                 head = alloc_skb(0, sk->sk_allocation);
 823                 while (!head) {
 824                         kcm_push(kcm);
 825                         err = sk_stream_wait_memory(sk, &timeo);
 826                         if (err)
 827                                 goto out_error;
 828                 }
 829
 830                 skb = head;
 831                 i = 0;
 832         }
 833
 834         get_page(page);
 835         skb_fill_page_desc(skb, i, page, offset, size);
 836         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
 837
 838 coalesced:
 839         skb->len += size;
 840         skb->data_len += size;
 841         skb->truesize += size;
 842         sk->sk_wmem_queued += size;
 843         sk_mem_charge(sk, size);
 844
 845         if (head != skb) {
 846                 head->len += size;
 847                 head->data_len += size;
 848                 head->truesize += size;
 849         }
 850
 851         if (eor) {
 852                 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
 853
 854                 /* Message complete, queue it on send buffer */
 855                 __skb_queue_tail(&sk->sk_write_queue, head);
 856                 kcm->seq_skb = NULL;
 857                 KCM_STATS_INCR(kcm->stats.tx_msgs);
 858
 859                 if (flags & MSG_BATCH) {
 860                         kcm->tx_wait_more = true;
 861                 } else if (kcm->tx_wait_more || not_busy) {
 862                         err = kcm_write_msgs(kcm);
 863                         if (err < 0) {
 864                                 /* We got a hard error in write_msgs but have
 865                                  * already queued this message. Report an error
 866                                  * in the socket, but don't affect return value
 867                                  * from sendmsg
 868                                  */
 869                                 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
 870                                 report_csk_error(&kcm->sk, -err);
 871                         }
 872                 }
 873         } else {
 874                 /* Message not complete, save state */
 875                 kcm->seq_skb = head;
 876                 kcm_tx_msg(head)->last_skb = skb;
 877         }
 878
 879         KCM_STATS_ADD(kcm->stats.tx_bytes, size);
 880
 881         release_sock(sk);
 882         return size;
 883
 884 out_error:
 885         kcm_push(kcm);
 886
 887         err = sk_stream_error(sk, flags, err);
 888
 889         /* make sure we wake any epoll edge trigger waiter */
 890         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
 891                 sk->sk_write_space(sk);
 892
 893         release_sock(sk);
 894         return err;
 895 }
 896
 897 static int kcm_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
 898 {
 899         struct sock *sk = sock->sk;
 900         struct kcm_sock *kcm = kcm_sk(sk);
 901         struct sk_buff *skb = NULL, *head = NULL;
 902         size_t copy, copied = 0;
 903         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 904         int eor = (sock->type == SOCK_DGRAM) ?
 905                   !(msg->msg_flags & MSG_MORE) : !!(msg->msg_flags & MSG_EOR);
 906         int err = -EPIPE;
 907
 908         lock_sock(sk);
 909
 910         /* Per tcp_sendmsg this should be in poll */
 911         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 912
 913         if (sk->sk_err)
 914                 goto out_error;
 915
 916         if (kcm->seq_skb) {
 917                 /* Previously opened message */
 918                 head = kcm->seq_skb;
 919                 skb = kcm_tx_msg(head)->last_skb;
 920                 goto start;
 921         }
 922
 923         /* Call the sk_stream functions to manage the sndbuf mem. */
 924         if (!sk_stream_memory_free(sk)) {
 925                 kcm_push(kcm);
 926                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 927                 err = sk_stream_wait_memory(sk, &timeo);
 928                 if (err)
 929                         goto out_error;
 930         }
 931
 932         /* New message, alloc head skb */
 933         head = alloc_skb(0, sk->sk_allocation);
 934         while (!head) {
 935                 kcm_push(kcm);
 936                 err = sk_stream_wait_memory(sk, &timeo);
 937                 if (err)
 938                         goto out_error;
 939
 940                 head = alloc_skb(0, sk->sk_allocation);
 941         }
 942
 943         skb = head;
 944
 945         /* Set ip_summed to CHECKSUM_UNNECESSARY to avoid calling
 946          * csum_and_copy_from_iter from skb_do_copy_data_nocache.
 947          */
 948         skb->ip_summed = CHECKSUM_UNNECESSARY;
 949
 950 start:
 951         while (msg_data_left(msg)) {
 952                 bool merge = true;
 953                 int i = skb_shinfo(skb)->nr_frags;
 954                 struct page_frag *pfrag = sk_page_frag(sk);
 955
 956                 if (!sk_page_frag_refill(sk, pfrag))
 957                         goto wait_for_memory;
 958
 959                 if (!skb_can_coalesce(skb, i, pfrag->page,
 960                                       pfrag->offset)) {
 961                         if (i == MAX_SKB_FRAGS) {
 962                                 struct sk_buff *tskb;
 963
 964                                 tskb = alloc_skb(0, sk->sk_allocation);
 965                                 if (!tskb)
 966                                         goto wait_for_memory;
 967
 968                                 if (head == skb)
 969                                         skb_shinfo(head)->frag_list = tskb;
 970                                 else
 971                                         skb->next = tskb;
 972
 973                                 skb = tskb;
 974                                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 975                                 continue;
 976                         }
 977                         merge = false;
 978                 }
 979
 980                 copy = min_t(int, msg_data_left(msg),
 981                              pfrag->size - pfrag->offset);
 982
 983                 if (!sk_wmem_schedule(sk, copy))
 984                         goto wait_for_memory;
 985
 986                 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
 987                                                pfrag->page,
 988                                                pfrag->offset,
 989                                                copy);
 990                 if (err)
 991                         goto out_error;
 992
 993                 /* Update the skb. */
 994                 if (merge) {
 995                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
 996                 } else {
 997                         skb_fill_page_desc(skb, i, pfrag->page,
 998                                            pfrag->offset, copy);
 999                         get_page(pfrag->page);
1000                 }
1001
1002                 pfrag->offset += copy;
1003                 copied += copy;
1004                 if (head != skb) {
1005                         head->len += copy;
1006                         head->data_len += copy;
1007                 }
1008
1009                 continue;
1010
1011 wait_for_memory:
1012                 kcm_push(kcm);
1013                 err = sk_stream_wait_memory(sk, &timeo);
1014                 if (err)
1015                         goto out_error;
1016         }
1017
1018         if (eor) {
1019                 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
1020
1021                 /* Message complete, queue it on send buffer */
1022                 __skb_queue_tail(&sk->sk_write_queue, head);
1023                 kcm->seq_skb = NULL;
1024                 KCM_STATS_INCR(kcm->stats.tx_msgs);
1025
1026                 if (msg->msg_flags & MSG_BATCH) {
1027                         kcm->tx_wait_more = true;
1028                 } else if (kcm->tx_wait_more || not_busy) {
1029                         err = kcm_write_msgs(kcm);
1030                         if (err < 0) {
1031                                 /* We got a hard error in write_msgs but have
1032                                  * already queued this message. Report an error
1033                                  * in the socket, but don't affect return value
1034                                  * from sendmsg
1035                                  */
1036                                 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
1037                                 report_csk_error(&kcm->sk, -err);
1038                         }
1039                 }
1040         } else {
1041                 /* Message not complete, save state */
1042 partial_message:
1043                 kcm->seq_skb = head;
1044                 kcm_tx_msg(head)->last_skb = skb;
1045         }
1046
1047         KCM_STATS_ADD(kcm->stats.tx_bytes, copied);
1048
1049         release_sock(sk);
1050         return copied;
1051
1052 out_error:
1053         kcm_push(kcm);
1054
1055         if (copied && sock->type == SOCK_SEQPACKET) {
1056                 /* Wrote some bytes before encountering an
1057                  * error, return partial success.
1058                  */
1059                 goto partial_message;
1060         }
1061
1062         if (head != kcm->seq_skb)
1063                 kfree_skb(head);
1064
1065         err = sk_stream_error(sk, msg->msg_flags, err);
1066
1067         /* make sure we wake any epoll edge trigger waiter */
1068         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1069                 sk->sk_write_space(sk);
1070
1071         release_sock(sk);
1072         return err;
1073 }
1074
1075 static struct sk_buff *kcm_wait_data(struct sock *sk, int flags,
1076                                      long timeo, int *err)
1077 {
1078         struct sk_buff *skb;
1079
1080         while (!(skb = skb_peek(&sk->sk_receive_queue))) {
1081                 if (sk->sk_err) {
1082                         *err = sock_error(sk);
1083                         return NULL;
1084                 }
1085
1086                 if (sock_flag(sk, SOCK_DONE))
1087                         return NULL;
1088
1089                 if ((flags & MSG_DONTWAIT) || !timeo) {
1090                         *err = -EAGAIN;
1091                         return NULL;
1092                 }
1093
1094                 sk_wait_data(sk, &timeo, NULL);
1095
1096                 /* Handle signals */
1097                 if (signal_pending(current)) {
1098                         *err = sock_intr_errno(timeo);
1099                         return NULL;
1100                 }
1101         }
1102
1103         return skb;
1104 }
1105
1106 static int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
1107                        size_t len, int flags)
1108 {
1109         struct sock *sk = sock->sk;
1110         struct kcm_sock *kcm = kcm_sk(sk);
1111         int err = 0;
1112         long timeo;
1113         struct strp_rx_msg *rxm;
1114         int copied = 0;
1115         struct sk_buff *skb;
1116
1117         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1118
1119         lock_sock(sk);
1120
1121         skb = kcm_wait_data(sk, flags, timeo, &err);
1122         if (!skb)
1123                 goto out;
1124
1125         /* Okay, have a message on the receive queue */
1126
1127         rxm = strp_rx_msg(skb);
1128
1129         if (len > rxm->full_len)
1130                 len = rxm->full_len;
1131
1132         err = skb_copy_datagram_msg(skb, rxm->offset, msg, len);
1133         if (err < 0)
1134                 goto out;
1135
1136         copied = len;
1137         if (likely(!(flags & MSG_PEEK))) {
1138                 KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1139                 if (copied < rxm->full_len) {
1140                         if (sock->type == SOCK_DGRAM) {
1141                                 /* Truncated message */
1142                                 msg->msg_flags |= MSG_TRUNC;
1143                                 goto msg_finished;
1144                         }
1145                         rxm->offset += copied;
1146                         rxm->full_len -= copied;
1147                 } else {
1148 msg_finished:
1149                         /* Finished with message */
1150                         msg->msg_flags |= MSG_EOR;
1151                         KCM_STATS_INCR(kcm->stats.rx_msgs);
1152                         skb_unlink(skb, &sk->sk_receive_queue);
1153                         kfree_skb(skb);
1154                 }
1155         }
1156
1157 out:
1158         release_sock(sk);
1159
1160         return copied ? : err;
1161 }
1162
1163 static ssize_t kcm_sock_splice(struct sock *sk,
1164                                struct pipe_inode_info *pipe,
1165                                struct splice_pipe_desc *spd)
1166 {
1167         int ret;
1168
1169         release_sock(sk);
1170         ret = splice_to_pipe(pipe, spd);
1171         lock_sock(sk);
1172
1173         return ret;
1174 }
1175
1176 static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1177                                struct pipe_inode_info *pipe, size_t len,
1178                                unsigned int flags)
1179 {
1180         struct sock *sk = sock->sk;
1181         struct kcm_sock *kcm = kcm_sk(sk);
1182         long timeo;
1183         struct strp_rx_msg *rxm;
1184         int err = 0;
1185         ssize_t copied;
1186         struct sk_buff *skb;
1187
1188         /* Only support splice for SOCKSEQPACKET */
1189
1190         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1191
1192         lock_sock(sk);
1193
1194         skb = kcm_wait_data(sk, flags, timeo, &err);
1195         if (!skb)
1196                 goto err_out;
1197
1198         /* Okay, have a message on the receive queue */
1199
1200         rxm = strp_rx_msg(skb);
1201
1202         if (len > rxm->full_len)
1203                 len = rxm->full_len;
1204
1205         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, len, flags,
1206                                  kcm_sock_splice);
1207         if (copied < 0) {
1208                 err = copied;
1209                 goto err_out;
1210         }
1211
1212         KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1213
1214         rxm->offset += copied;
1215         rxm->full_len -= copied;
1216
1217         /* We have no way to return MSG_EOR. If all the bytes have been
1218          * read we still leave the message in the receive socket buffer.
1219          * A subsequent recvmsg needs to be done to return MSG_EOR and
1220          * finish reading the message.
1221          */
1222
1223         release_sock(sk);
1224
1225         return copied;
1226
1227 err_out:
1228         release_sock(sk);
1229
1230         return err;
1231 }
1232
1233 /* kcm sock lock held */
1234 static void kcm_recv_disable(struct kcm_sock *kcm)
1235 {
1236         struct kcm_mux *mux = kcm->mux;
1237
1238         if (kcm->rx_disabled)
1239                 return;
1240
1241         spin_lock_bh(&mux->rx_lock);
1242
1243         kcm->rx_disabled = 1;
1244
1245         /* If a psock is reserved we'll do cleanup in unreserve */
1246         if (!kcm->rx_psock) {
1247                 if (kcm->rx_wait) {
1248                         list_del(&kcm->wait_rx_list);
1249                         kcm->rx_wait = false;
1250                 }
1251
1252                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1253         }
1254
1255         spin_unlock_bh(&mux->rx_lock);
1256 }
1257
1258 /* kcm sock lock held */
1259 static void kcm_recv_enable(struct kcm_sock *kcm)
1260 {
1261         struct kcm_mux *mux = kcm->mux;
1262
1263         if (!kcm->rx_disabled)
1264                 return;
1265
1266         spin_lock_bh(&mux->rx_lock);
1267
1268         kcm->rx_disabled = 0;
1269         kcm_rcv_ready(kcm);
1270
1271         spin_unlock_bh(&mux->rx_lock);
1272 }
1273
1274 static int kcm_setsockopt(struct socket *sock, int level, int optname,
1275                           char __user *optval, unsigned int optlen)
1276 {
1277         struct kcm_sock *kcm = kcm_sk(sock->sk);
1278         int val, valbool;
1279         int err = 0;
1280
1281         if (level != SOL_KCM)
1282                 return -ENOPROTOOPT;
1283
1284         if (optlen < sizeof(int))
1285                 return -EINVAL;
1286
1287         if (get_user(val, (int __user *)optval))
1288                 return -EINVAL;
1289
1290         valbool = val ? 1 : 0;
1291
1292         switch (optname) {
1293         case KCM_RECV_DISABLE:
1294                 lock_sock(&kcm->sk);
1295                 if (valbool)
1296                         kcm_recv_disable(kcm);
1297                 else
1298                         kcm_recv_enable(kcm);
1299                 release_sock(&kcm->sk);
1300                 break;
1301         default:
1302                 err = -ENOPROTOOPT;
1303         }
1304
1305         return err;
1306 }
1307
1308 static int kcm_getsockopt(struct socket *sock, int level, int optname,
1309                           char __user *optval, int __user *optlen)
1310 {
1311         struct kcm_sock *kcm = kcm_sk(sock->sk);
1312         int val, len;
1313
1314         if (level != SOL_KCM)
1315                 return -ENOPROTOOPT;
1316
1317         if (get_user(len, optlen))
1318                 return -EFAULT;
1319
1320         len = min_t(unsigned int, len, sizeof(int));
1321         if (len < 0)
1322                 return -EINVAL;
1323
1324         switch (optname) {
1325         case KCM_RECV_DISABLE:
1326                 val = kcm->rx_disabled;
1327                 break;
1328         default:
1329                 return -ENOPROTOOPT;
1330         }
1331
1332         if (put_user(len, optlen))
1333                 return -EFAULT;
1334         if (copy_to_user(optval, &val, len))
1335                 return -EFAULT;
1336         return 0;
1337 }
1338
1339 static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1340 {
1341         struct kcm_sock *tkcm;
1342         struct list_head *head;
1343         int index = 0;
1344
1345         /* For SOCK_SEQPACKET sock type, datagram_poll checks the sk_state, so
1346          * we set sk_state, otherwise epoll_wait always returns right away with
1347          * POLLHUP
1348          */
1349         kcm->sk.sk_state = TCP_ESTABLISHED;
1350
1351         /* Add to mux's kcm sockets list */
1352         kcm->mux = mux;
1353         spin_lock_bh(&mux->lock);
1354
1355         head = &mux->kcm_socks;
1356         list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1357                 if (tkcm->index != index)
1358                         break;
1359                 head = &tkcm->kcm_sock_list;
1360                 index++;
1361         }
1362
1363         list_add(&kcm->kcm_sock_list, head);
1364         kcm->index = index;
1365
1366         mux->kcm_socks_cnt++;
1367         spin_unlock_bh(&mux->lock);
1368
1369         INIT_WORK(&kcm->tx_work, kcm_tx_work);
1370
1371         spin_lock_bh(&mux->rx_lock);
1372         kcm_rcv_ready(kcm);
1373         spin_unlock_bh(&mux->rx_lock);
1374 }
1375
1376 static int kcm_attach(struct socket *sock, struct socket *csock,
1377                       struct bpf_prog *prog)
1378 {
1379         struct kcm_sock *kcm = kcm_sk(sock->sk);
1380         struct kcm_mux *mux = kcm->mux;
1381         struct sock *csk;
1382         struct kcm_psock *psock = NULL, *tpsock;
1383         struct list_head *head;
1384         int index = 0;
1385         struct strp_callbacks cb;
1386         int err;
1387
1388         csk = csock->sk;
1389         if (!csk)
1390                 return -EINVAL;
1391
1392         psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1393         if (!psock)
1394                 return -ENOMEM;
1395
1396         psock->mux = mux;
1397         psock->sk = csk;
1398         psock->bpf_prog = prog;
1399
1400         cb.rcv_msg = kcm_rcv_strparser;
1401         cb.abort_parser = NULL;
1402         cb.parse_msg = kcm_parse_func_strparser;
1403         cb.read_sock_done = kcm_read_sock_done;
1404
1405         err = strp_init(&psock->strp, csk, &cb);
1406         if (err) {
1407                 kmem_cache_free(kcm_psockp, psock);
1408                 return err;
1409         }
1410
1411         sock_hold(csk);
1412
1413         write_lock_bh(&csk->sk_callback_lock);
1414         psock->save_data_ready = csk->sk_data_ready;
1415         psock->save_write_space = csk->sk_write_space;
1416         psock->save_state_change = csk->sk_state_change;
1417         csk->sk_user_data = psock;
1418         csk->sk_data_ready = psock_data_ready;
1419         csk->sk_write_space = psock_write_space;
1420         csk->sk_state_change = psock_state_change;
1421         write_unlock_bh(&csk->sk_callback_lock);
1422
1423         /* Finished initialization, now add the psock to the MUX. */
1424         spin_lock_bh(&mux->lock);
1425         head = &mux->psocks;
1426         list_for_each_entry(tpsock, &mux->psocks, psock_list) {
1427                 if (tpsock->index != index)
1428                         break;
1429                 head = &tpsock->psock_list;
1430                 index++;
1431         }
1432
1433         list_add(&psock->psock_list, head);
1434         psock->index = index;
1435
1436         KCM_STATS_INCR(mux->stats.psock_attach);
1437         mux->psocks_cnt++;
1438         psock_now_avail(psock);
1439         spin_unlock_bh(&mux->lock);
1440
1441         /* Schedule RX work in case there are already bytes queued */
1442         strp_check_rcv(&psock->strp);
1443
1444         return 0;
1445 }
1446
1447 static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1448 {
1449         struct socket *csock;
1450         struct bpf_prog *prog;
1451         int err;
1452
1453         csock = sockfd_lookup(info->fd, &err);
1454         if (!csock)
1455                 return -ENOENT;
1456
1457         prog = bpf_prog_get_type(info->bpf_fd, BPF_PROG_TYPE_SOCKET_FILTER);
1458         if (IS_ERR(prog)) {
1459                 err = PTR_ERR(prog);
1460                 goto out;
1461         }
1462
1463         err = kcm_attach(sock, csock, prog);
1464         if (err) {
1465                 bpf_prog_put(prog);
1466                 goto out;
1467         }
1468
1469         /* Keep reference on file also */
1470
1471         return 0;
1472 out:
1473         fput(csock->file);
1474         return err;
1475 }
1476
1477 static void kcm_unattach(struct kcm_psock *psock)
1478 {
1479         struct sock *csk = psock->sk;
1480         struct kcm_mux *mux = psock->mux;
1481
1482         lock_sock(csk);
1483
1484         /* Stop getting callbacks from TCP socket. After this there should
1485          * be no way to reserve a kcm for this psock.
1486          */
1487         write_lock_bh(&csk->sk_callback_lock);
1488         csk->sk_user_data = NULL;
1489         csk->sk_data_ready = psock->save_data_ready;
1490         csk->sk_write_space = psock->save_write_space;
1491         csk->sk_state_change = psock->save_state_change;
1492         strp_stop(&psock->strp);
1493
1494         if (WARN_ON(psock->rx_kcm)) {
1495                 write_unlock_bh(&csk->sk_callback_lock);
1496                 return;
1497         }
1498
1499         spin_lock_bh(&mux->rx_lock);
1500
1501         /* Stop receiver activities. After this point psock should not be
1502          * able to get onto ready list either through callbacks or work.
1503          */
1504         if (psock->ready_rx_msg) {
1505                 list_del(&psock->psock_ready_list);
1506                 kfree_skb(psock->ready_rx_msg);
1507                 psock->ready_rx_msg = NULL;
1508                 KCM_STATS_INCR(mux->stats.rx_ready_drops);
1509         }
1510
1511         spin_unlock_bh(&mux->rx_lock);
1512
1513         write_unlock_bh(&csk->sk_callback_lock);
1514
1515         /* Call strp_done without sock lock */
1516         release_sock(csk);
1517         strp_done(&psock->strp);
1518         lock_sock(csk);
1519
1520         bpf_prog_put(psock->bpf_prog);
1521
1522         spin_lock_bh(&mux->lock);
1523
1524         aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1525         save_strp_stats(&psock->strp, &mux->aggregate_strp_stats);
1526
1527         KCM_STATS_INCR(mux->stats.psock_unattach);
1528
1529         if (psock->tx_kcm) {
1530                 /* psock was reserved.  Just mark it finished and we will clean
1531                  * up in the kcm paths, we need kcm lock which can not be
1532                  * acquired here.
1533                  */
1534                 KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1535                 spin_unlock_bh(&mux->lock);
1536
1537                 /* We are unattaching a socket that is reserved. Abort the
1538                  * socket since we may be out of sync in sending on it. We need
1539                  * to do this without the mux lock.
1540                  */
1541                 kcm_abort_tx_psock(psock, EPIPE, false);
1542
1543                 spin_lock_bh(&mux->lock);
1544                 if (!psock->tx_kcm) {
1545                         /* psock now unreserved in window mux was unlocked */
1546                         goto no_reserved;
1547                 }
1548                 psock->done = 1;
1549
1550                 /* Commit done before queuing work to process it */
1551                 smp_mb();
1552
1553                 /* Queue tx work to make sure psock->done is handled */
1554                 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
1555                 spin_unlock_bh(&mux->lock);
1556         } else {
1557 no_reserved:
1558                 if (!psock->tx_stopped)
1559                         list_del(&psock->psock_avail_list);
1560                 list_del(&psock->psock_list);
1561                 mux->psocks_cnt--;
1562                 spin_unlock_bh(&mux->lock);
1563
1564                 sock_put(csk);
1565                 fput(csk->sk_socket->file);
1566                 kmem_cache_free(kcm_psockp, psock);
1567         }
1568
1569         release_sock(csk);
1570 }
1571
1572 static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1573 {
1574         struct kcm_sock *kcm = kcm_sk(sock->sk);
1575         struct kcm_mux *mux = kcm->mux;
1576         struct kcm_psock *psock;
1577         struct socket *csock;
1578         struct sock *csk;
1579         int err;
1580
1581         csock = sockfd_lookup(info->fd, &err);
1582         if (!csock)
1583                 return -ENOENT;
1584
1585         csk = csock->sk;
1586         if (!csk) {
1587                 err = -EINVAL;
1588                 goto out;
1589         }
1590
1591         err = -ENOENT;
1592
1593         spin_lock_bh(&mux->lock);
1594
1595         list_for_each_entry(psock, &mux->psocks, psock_list) {
1596                 if (psock->sk != csk)
1597                         continue;
1598
1599                 /* Found the matching psock */
1600
1601                 if (psock->unattaching || WARN_ON(psock->done)) {
1602                         err = -EALREADY;
1603                         break;
1604                 }
1605
1606                 psock->unattaching = 1;
1607
1608                 spin_unlock_bh(&mux->lock);
1609
1610                 /* Lower socket lock should already be held */
1611                 kcm_unattach(psock);
1612
1613                 err = 0;
1614                 goto out;
1615         }
1616
1617         spin_unlock_bh(&mux->lock);
1618
1619 out:
1620         fput(csock->file);
1621         return err;
1622 }
1623
1624 static struct proto kcm_proto = {
1625         .name   = "KCM",
1626         .owner  = THIS_MODULE,
1627         .obj_size = sizeof(struct kcm_sock),
1628 };
1629
1630 /* Clone a kcm socket. */
1631 static int kcm_clone(struct socket *osock, struct kcm_clone *info,
1632                      struct socket **newsockp)
1633 {
1634         struct socket *newsock;
1635         struct sock *newsk;
1636         struct file *newfile;
1637         int err, newfd;
1638
1639         err = -ENFILE;
1640         newsock = sock_alloc();
1641         if (!newsock)
1642                 goto out;
1643
1644         newsock->type = osock->type;
1645         newsock->ops = osock->ops;
1646
1647         __module_get(newsock->ops->owner);
1648
1649         newfd = get_unused_fd_flags(0);
1650         if (unlikely(newfd < 0)) {
1651                 err = newfd;
1652                 goto out_fd_fail;
1653         }
1654
1655         newfile = sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1656         if (unlikely(IS_ERR(newfile))) {
1657                 err = PTR_ERR(newfile);
1658                 goto out_sock_alloc_fail;
1659         }
1660
1661         newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1662                          &kcm_proto, true);
1663         if (!newsk) {
1664                 err = -ENOMEM;
1665                 goto out_sk_alloc_fail;
1666         }
1667
1668         sock_init_data(newsock, newsk);
1669         init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1670
1671         fd_install(newfd, newfile);
1672         *newsockp = newsock;
1673         info->fd = newfd;
1674
1675         return 0;
1676
1677 out_sk_alloc_fail:
1678         fput(newfile);
1679 out_sock_alloc_fail:
1680         put_unused_fd(newfd);
1681 out_fd_fail:
1682         sock_release(newsock);
1683 out:
1684         return err;
1685 }
1686
1687 static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
1688 {
1689         int err;
1690
1691         switch (cmd) {
1692         case SIOCKCMATTACH: {
1693                 struct kcm_attach info;
1694
1695                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1696                         err = -EFAULT;
1697
1698                 err = kcm_attach_ioctl(sock, &info);
1699
1700                 break;
1701         }
1702         case SIOCKCMUNATTACH: {
1703                 struct kcm_unattach info;
1704
1705                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1706                         err = -EFAULT;
1707
1708                 err = kcm_unattach_ioctl(sock, &info);
1709
1710                 break;
1711         }
1712         case SIOCKCMCLONE: {
1713                 struct kcm_clone info;
1714                 struct socket *newsock = NULL;
1715
1716                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1717                         err = -EFAULT;
1718
1719                 err = kcm_clone(sock, &info, &newsock);
1720
1721                 if (!err) {
1722                         if (copy_to_user((void __user *)arg, &info,
1723                                          sizeof(info))) {
1724                                 err = -EFAULT;
1725                                 sys_close(info.fd);
1726                         }
1727                 }
1728
1729                 break;
1730         }
1731         default:
1732                 err = -ENOIOCTLCMD;
1733                 break;
1734         }
1735
1736         return err;
1737 }
1738
1739 static void free_mux(struct rcu_head *rcu)
1740 {
1741         struct kcm_mux *mux = container_of(rcu,
1742             struct kcm_mux, rcu);
1743
1744         kmem_cache_free(kcm_muxp, mux);
1745 }
1746
1747 static void release_mux(struct kcm_mux *mux)
1748 {
1749         struct kcm_net *knet = mux->knet;
1750         struct kcm_psock *psock, *tmp_psock;
1751
1752         /* Release psocks */
1753         list_for_each_entry_safe(psock, tmp_psock,
1754                                  &mux->psocks, psock_list) {
1755                 if (!WARN_ON(psock->unattaching))
1756                         kcm_unattach(psock);
1757         }
1758
1759         if (WARN_ON(mux->psocks_cnt))
1760                 return;
1761
1762         __skb_queue_purge(&mux->rx_hold_queue);
1763
1764         mutex_lock(&knet->mutex);
1765         aggregate_mux_stats(&mux->stats, &knet->aggregate_mux_stats);
1766         aggregate_psock_stats(&mux->aggregate_psock_stats,
1767                               &knet->aggregate_psock_stats);
1768         aggregate_strp_stats(&mux->aggregate_strp_stats,
1769                              &knet->aggregate_strp_stats);
1770         list_del_rcu(&mux->kcm_mux_list);
1771         knet->count--;
1772         mutex_unlock(&knet->mutex);
1773
1774         call_rcu(&mux->rcu, free_mux);
1775 }
1776
1777 static void kcm_done(struct kcm_sock *kcm)
1778 {
1779         struct kcm_mux *mux = kcm->mux;
1780         struct sock *sk = &kcm->sk;
1781         int socks_cnt;
1782
1783         spin_lock_bh(&mux->rx_lock);
1784         if (kcm->rx_psock) {
1785                 /* Cleanup in unreserve_rx_kcm */
1786                 WARN_ON(kcm->done);
1787                 kcm->rx_disabled = 1;
1788                 kcm->done = 1;
1789                 spin_unlock_bh(&mux->rx_lock);
1790                 return;
1791         }
1792
1793         if (kcm->rx_wait) {
1794                 list_del(&kcm->wait_rx_list);
1795                 kcm->rx_wait = false;
1796         }
1797         /* Move any pending receive messages to other kcm sockets */
1798         requeue_rx_msgs(mux, &sk->sk_receive_queue);
1799
1800         spin_unlock_bh(&mux->rx_lock);
1801
1802         if (WARN_ON(sk_rmem_alloc_get(sk)))
1803                 return;
1804
1805         /* Detach from MUX */
1806         spin_lock_bh(&mux->lock);
1807
1808         list_del(&kcm->kcm_sock_list);
1809         mux->kcm_socks_cnt--;
1810         socks_cnt = mux->kcm_socks_cnt;
1811
1812         spin_unlock_bh(&mux->lock);
1813
1814         if (!socks_cnt) {
1815                 /* We are done with the mux now. */
1816                 release_mux(mux);
1817         }
1818
1819         WARN_ON(kcm->rx_wait);
1820
1821         sock_put(&kcm->sk);
1822 }
1823
1824 /* Called by kcm_release to close a KCM socket.
1825  * If this is the last KCM socket on the MUX, destroy the MUX.
1826  */
1827 static int kcm_release(struct socket *sock)
1828 {
1829         struct sock *sk = sock->sk;
1830         struct kcm_sock *kcm;
1831         struct kcm_mux *mux;
1832         struct kcm_psock *psock;
1833
1834         if (!sk)
1835                 return 0;
1836
1837         kcm = kcm_sk(sk);
1838         mux = kcm->mux;
1839
1840         sock_orphan(sk);
1841         kfree_skb(kcm->seq_skb);
1842
1843         lock_sock(sk);
1844         /* Purge queue under lock to avoid race condition with tx_work trying
1845          * to act when queue is nonempty. If tx_work runs after this point
1846          * it will just return.
1847          */
1848         __skb_queue_purge(&sk->sk_write_queue);
1849
1850         /* Set tx_stopped. This is checked when psock is bound to a kcm and we
1851          * get a writespace callback. This prevents further work being queued
1852          * from the callback (unbinding the psock occurs after canceling work.
1853          */
1854         kcm->tx_stopped = 1;
1855
1856         release_sock(sk);
1857
1858         spin_lock_bh(&mux->lock);
1859         if (kcm->tx_wait) {
1860                 /* Take of tx_wait list, after this point there should be no way
1861                  * that a psock will be assigned to this kcm.
1862                  */
1863                 list_del(&kcm->wait_psock_list);
1864                 kcm->tx_wait = false;
1865         }
1866         spin_unlock_bh(&mux->lock);
1867
1868         /* Cancel work. After this point there should be no outside references
1869          * to the kcm socket.
1870          */
1871         cancel_work_sync(&kcm->tx_work);
1872
1873         lock_sock(sk);
1874         psock = kcm->tx_psock;
1875         if (psock) {
1876                 /* A psock was reserved, so we need to kill it since it
1877                  * may already have some bytes queued from a message. We
1878                  * need to do this after removing kcm from tx_wait list.
1879                  */
1880                 kcm_abort_tx_psock(psock, EPIPE, false);
1881                 unreserve_psock(kcm);
1882         }
1883         release_sock(sk);
1884
1885         WARN_ON(kcm->tx_wait);
1886         WARN_ON(kcm->tx_psock);
1887
1888         sock->sk = NULL;
1889
1890         kcm_done(kcm);
1891
1892         return 0;
1893 }
1894
1895 static const struct proto_ops kcm_dgram_ops = {
1896         .family =       PF_KCM,
1897         .owner =        THIS_MODULE,
1898         .release =      kcm_release,
1899         .bind =         sock_no_bind,
1900         .connect =      sock_no_connect,
1901         .socketpair =   sock_no_socketpair,
1902         .accept =       sock_no_accept,
1903         .getname =      sock_no_getname,
1904         .poll =         datagram_poll,
1905         .ioctl =        kcm_ioctl,
1906         .listen =       sock_no_listen,
1907         .shutdown =     sock_no_shutdown,
1908         .setsockopt =   kcm_setsockopt,
1909         .getsockopt =   kcm_getsockopt,
1910         .sendmsg =      kcm_sendmsg,
1911         .recvmsg =      kcm_recvmsg,
1912         .mmap =         sock_no_mmap,
1913         .sendpage =     kcm_sendpage,
1914 };
1915
1916 static const struct proto_ops kcm_seqpacket_ops = {
1917         .family =       PF_KCM,
1918         .owner =        THIS_MODULE,
1919         .release =      kcm_release,
1920         .bind =         sock_no_bind,
1921         .connect =      sock_no_connect,
1922         .socketpair =   sock_no_socketpair,
1923         .accept =       sock_no_accept,
1924         .getname =      sock_no_getname,
1925         .poll =         datagram_poll,
1926         .ioctl =        kcm_ioctl,
1927         .listen =       sock_no_listen,
1928         .shutdown =     sock_no_shutdown,
1929         .setsockopt =   kcm_setsockopt,
1930         .getsockopt =   kcm_getsockopt,
1931         .sendmsg =      kcm_sendmsg,
1932         .recvmsg =      kcm_recvmsg,
1933         .mmap =         sock_no_mmap,
1934         .sendpage =     kcm_sendpage,
1935         .splice_read =  kcm_splice_read,
1936 };
1937
1938 /* Create proto operation for kcm sockets */
1939 static int kcm_create(struct net *net, struct socket *sock,
1940                       int protocol, int kern)
1941 {
1942         struct kcm_net *knet = net_generic(net, kcm_net_id);
1943         struct sock *sk;
1944         struct kcm_mux *mux;
1945
1946         switch (sock->type) {
1947         case SOCK_DGRAM:
1948                 sock->ops = &kcm_dgram_ops;
1949                 break;
1950         case SOCK_SEQPACKET:
1951                 sock->ops = &kcm_seqpacket_ops;
1952                 break;
1953         default:
1954                 return -ESOCKTNOSUPPORT;
1955         }
1956
1957         if (protocol != KCMPROTO_CONNECTED)
1958                 return -EPROTONOSUPPORT;
1959
1960         sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
1961         if (!sk)
1962                 return -ENOMEM;
1963
1964         /* Allocate a kcm mux, shared between KCM sockets */
1965         mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
1966         if (!mux) {
1967                 sk_free(sk);
1968                 return -ENOMEM;
1969         }
1970
1971         spin_lock_init(&mux->lock);
1972         spin_lock_init(&mux->rx_lock);
1973         INIT_LIST_HEAD(&mux->kcm_socks);
1974         INIT_LIST_HEAD(&mux->kcm_rx_waiters);
1975         INIT_LIST_HEAD(&mux->kcm_tx_waiters);
1976
1977         INIT_LIST_HEAD(&mux->psocks);
1978         INIT_LIST_HEAD(&mux->psocks_ready);
1979         INIT_LIST_HEAD(&mux->psocks_avail);
1980
1981         mux->knet = knet;
1982
1983         /* Add new MUX to list */
1984         mutex_lock(&knet->mutex);
1985         list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
1986         knet->count++;
1987         mutex_unlock(&knet->mutex);
1988
1989         skb_queue_head_init(&mux->rx_hold_queue);
1990
1991         /* Init KCM socket */
1992         sock_init_data(sock, sk);
1993         init_kcm_sock(kcm_sk(sk), mux);
1994
1995         return 0;
1996 }
1997
1998 static struct net_proto_family kcm_family_ops = {
1999         .family = PF_KCM,
2000         .create = kcm_create,
2001         .owner  = THIS_MODULE,
2002 };
2003
2004 static __net_init int kcm_init_net(struct net *net)
2005 {
2006         struct kcm_net *knet = net_generic(net, kcm_net_id);
2007
2008         INIT_LIST_HEAD_RCU(&knet->mux_list);
2009         mutex_init(&knet->mutex);
2010
2011         return 0;
2012 }
2013
2014 static __net_exit void kcm_exit_net(struct net *net)
2015 {
2016         struct kcm_net *knet = net_generic(net, kcm_net_id);
2017
2018         /* All KCM sockets should be closed at this point, which should mean
2019          * that all multiplexors and psocks have been destroyed.
2020          */
2021         WARN_ON(!list_empty(&knet->mux_list));
2022 }
2023
2024 static struct pernet_operations kcm_net_ops = {
2025         .init = kcm_init_net,
2026         .exit = kcm_exit_net,
2027         .id   = &kcm_net_id,
2028         .size = sizeof(struct kcm_net),
2029 };
2030
2031 static int __init kcm_init(void)
2032 {
2033         int err = -ENOMEM;
2034
2035         kcm_muxp = kmem_cache_create("kcm_mux_cache",
2036                                      sizeof(struct kcm_mux), 0,
2037                                      SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2038         if (!kcm_muxp)
2039                 goto fail;
2040
2041         kcm_psockp = kmem_cache_create("kcm_psock_cache",
2042                                        sizeof(struct kcm_psock), 0,
2043                                         SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2044         if (!kcm_psockp)
2045                 goto fail;
2046
2047         kcm_wq = create_singlethread_workqueue("kkcmd");
2048         if (!kcm_wq)
2049                 goto fail;
2050
2051         err = proto_register(&kcm_proto, 1);
2052         if (err)
2053                 goto fail;
2054
2055         err = sock_register(&kcm_family_ops);
2056         if (err)
2057                 goto sock_register_fail;
2058
2059         err = register_pernet_device(&kcm_net_ops);
2060         if (err)
2061                 goto net_ops_fail;
2062
2063         err = kcm_proc_init();
2064         if (err)
2065                 goto proc_init_fail;
2066
2067         return 0;
2068
2069 proc_init_fail:
2070         unregister_pernet_device(&kcm_net_ops);
2071
2072 net_ops_fail:
2073         sock_unregister(PF_KCM);
2074
2075 sock_register_fail:
2076         proto_unregister(&kcm_proto);
2077
2078 fail:
2079         kmem_cache_destroy(kcm_muxp);
2080         kmem_cache_destroy(kcm_psockp);
2081
2082         if (kcm_wq)
2083                 destroy_workqueue(kcm_wq);
2084
2085         return err;
2086 }
2087
2088 static void __exit kcm_exit(void)
2089 {
2090         kcm_proc_exit();
2091         unregister_pernet_device(&kcm_net_ops);
2092         sock_unregister(PF_KCM);
2093         proto_unregister(&kcm_proto);
2094         destroy_workqueue(kcm_wq);
2095
2096         kmem_cache_destroy(kcm_muxp);
2097         kmem_cache_destroy(kcm_psockp);
2098 }
2099
2100 module_init(kcm_init);
2101 module_exit(kcm_exit);
2102
2103 MODULE_LICENSE("GPL");
2104 MODULE_ALIAS_NETPROTO(PF_KCM);
2105