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         if (msg_data_left(msg)) {
 933                 /* New message, alloc head skb */
 934                 head = alloc_skb(0, sk->sk_allocation);
 935                 while (!head) {
 936                         kcm_push(kcm);
 937                         err = sk_stream_wait_memory(sk, &timeo);
 938                         if (err)
 939                                 goto out_error;
 940
 941                         head = alloc_skb(0, sk->sk_allocation);
 942                 }
 943
 944                 skb = head;
 945
 946                 /* Set ip_summed to CHECKSUM_UNNECESSARY to avoid calling
 947                  * csum_and_copy_from_iter from skb_do_copy_data_nocache.
 948                  */
 949                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 950         }
 951
 952 start:
 953         while (msg_data_left(msg)) {
 954                 bool merge = true;
 955                 int i = skb_shinfo(skb)->nr_frags;
 956                 struct page_frag *pfrag = sk_page_frag(sk);
 957
 958                 if (!sk_page_frag_refill(sk, pfrag))
 959                         goto wait_for_memory;
 960
 961                 if (!skb_can_coalesce(skb, i, pfrag->page,
 962                                       pfrag->offset)) {
 963                         if (i == MAX_SKB_FRAGS) {
 964                                 struct sk_buff *tskb;
 965
 966                                 tskb = alloc_skb(0, sk->sk_allocation);
 967                                 if (!tskb)
 968                                         goto wait_for_memory;
 969
 970                                 if (head == skb)
 971                                         skb_shinfo(head)->frag_list = tskb;
 972                                 else
 973                                         skb->next = tskb;
 974
 975                                 skb = tskb;
 976                                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 977                                 continue;
 978                         }
 979                         merge = false;
 980                 }
 981
 982                 copy = min_t(int, msg_data_left(msg),
 983                              pfrag->size - pfrag->offset);
 984
 985                 if (!sk_wmem_schedule(sk, copy))
 986                         goto wait_for_memory;
 987
 988                 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
 989                                                pfrag->page,
 990                                                pfrag->offset,
 991                                                copy);
 992                 if (err)
 993                         goto out_error;
 994
 995                 /* Update the skb. */
 996                 if (merge) {
 997                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
 998                 } else {
 999                         skb_fill_page_desc(skb, i, pfrag->page,
1000                                            pfrag->offset, copy);
1001                         get_page(pfrag->page);
1002                 }
1003
1004                 pfrag->offset += copy;
1005                 copied += copy;
1006                 if (head != skb) {
1007                         head->len += copy;
1008                         head->data_len += copy;
1009                 }
1010
1011                 continue;
1012
1013 wait_for_memory:
1014                 kcm_push(kcm);
1015                 err = sk_stream_wait_memory(sk, &timeo);
1016                 if (err)
1017                         goto out_error;
1018         }
1019
1020         if (eor) {
1021                 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
1022
1023                 if (head) {
1024                         /* Message complete, queue it on send buffer */
1025                         __skb_queue_tail(&sk->sk_write_queue, head);
1026                         kcm->seq_skb = NULL;
1027                         KCM_STATS_INCR(kcm->stats.tx_msgs);
1028                 }
1029
1030                 if (msg->msg_flags & MSG_BATCH) {
1031                         kcm->tx_wait_more = true;
1032                 } else if (kcm->tx_wait_more || not_busy) {
1033                         err = kcm_write_msgs(kcm);
1034                         if (err < 0) {
1035                                 /* We got a hard error in write_msgs but have
1036                                  * already queued this message. Report an error
1037                                  * in the socket, but don't affect return value
1038                                  * from sendmsg
1039                                  */
1040                                 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
1041                                 report_csk_error(&kcm->sk, -err);
1042                         }
1043                 }
1044         } else {
1045                 /* Message not complete, save state */
1046 partial_message:
1047                 if (head) {
1048                         kcm->seq_skb = head;
1049                         kcm_tx_msg(head)->last_skb = skb;
1050                 }
1051         }
1052
1053         KCM_STATS_ADD(kcm->stats.tx_bytes, copied);
1054
1055         release_sock(sk);
1056         return copied;
1057
1058 out_error:
1059         kcm_push(kcm);
1060
1061         if (copied && sock->type == SOCK_SEQPACKET) {
1062                 /* Wrote some bytes before encountering an
1063                  * error, return partial success.
1064                  */
1065                 goto partial_message;
1066         }
1067
1068         if (head != kcm->seq_skb)
1069                 kfree_skb(head);
1070
1071         err = sk_stream_error(sk, msg->msg_flags, err);
1072
1073         /* make sure we wake any epoll edge trigger waiter */
1074         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1075                 sk->sk_write_space(sk);
1076
1077         release_sock(sk);
1078         return err;
1079 }
1080
1081 static struct sk_buff *kcm_wait_data(struct sock *sk, int flags,
1082                                      long timeo, int *err)
1083 {
1084         struct sk_buff *skb;
1085
1086         while (!(skb = skb_peek(&sk->sk_receive_queue))) {
1087                 if (sk->sk_err) {
1088                         *err = sock_error(sk);
1089                         return NULL;
1090                 }
1091
1092                 if (sock_flag(sk, SOCK_DONE))
1093                         return NULL;
1094
1095                 if ((flags & MSG_DONTWAIT) || !timeo) {
1096                         *err = -EAGAIN;
1097                         return NULL;
1098                 }
1099
1100                 sk_wait_data(sk, &timeo, NULL);
1101
1102                 /* Handle signals */
1103                 if (signal_pending(current)) {
1104                         *err = sock_intr_errno(timeo);
1105                         return NULL;
1106                 }
1107         }
1108
1109         return skb;
1110 }
1111
1112 static int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
1113                        size_t len, int flags)
1114 {
1115         struct sock *sk = sock->sk;
1116         struct kcm_sock *kcm = kcm_sk(sk);
1117         int err = 0;
1118         long timeo;
1119         struct strp_rx_msg *rxm;
1120         int copied = 0;
1121         struct sk_buff *skb;
1122
1123         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1124
1125         lock_sock(sk);
1126
1127         skb = kcm_wait_data(sk, flags, timeo, &err);
1128         if (!skb)
1129                 goto out;
1130
1131         /* Okay, have a message on the receive queue */
1132
1133         rxm = strp_rx_msg(skb);
1134
1135         if (len > rxm->full_len)
1136                 len = rxm->full_len;
1137
1138         err = skb_copy_datagram_msg(skb, rxm->offset, msg, len);
1139         if (err < 0)
1140                 goto out;
1141
1142         copied = len;
1143         if (likely(!(flags & MSG_PEEK))) {
1144                 KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1145                 if (copied < rxm->full_len) {
1146                         if (sock->type == SOCK_DGRAM) {
1147                                 /* Truncated message */
1148                                 msg->msg_flags |= MSG_TRUNC;
1149                                 goto msg_finished;
1150                         }
1151                         rxm->offset += copied;
1152                         rxm->full_len -= copied;
1153                 } else {
1154 msg_finished:
1155                         /* Finished with message */
1156                         msg->msg_flags |= MSG_EOR;
1157                         KCM_STATS_INCR(kcm->stats.rx_msgs);
1158                         skb_unlink(skb, &sk->sk_receive_queue);
1159                         kfree_skb(skb);
1160                 }
1161         }
1162
1163 out:
1164         release_sock(sk);
1165
1166         return copied ? : err;
1167 }
1168
1169 static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1170                                struct pipe_inode_info *pipe, size_t len,
1171                                unsigned int flags)
1172 {
1173         struct sock *sk = sock->sk;
1174         struct kcm_sock *kcm = kcm_sk(sk);
1175         long timeo;
1176         struct strp_rx_msg *rxm;
1177         int err = 0;
1178         ssize_t copied;
1179         struct sk_buff *skb;
1180
1181         /* Only support splice for SOCKSEQPACKET */
1182
1183         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1184
1185         lock_sock(sk);
1186
1187         skb = kcm_wait_data(sk, flags, timeo, &err);
1188         if (!skb)
1189                 goto err_out;
1190
1191         /* Okay, have a message on the receive queue */
1192
1193         rxm = strp_rx_msg(skb);
1194
1195         if (len > rxm->full_len)
1196                 len = rxm->full_len;
1197
1198         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, len, flags);
1199         if (copied < 0) {
1200                 err = copied;
1201                 goto err_out;
1202         }
1203
1204         KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1205
1206         rxm->offset += copied;
1207         rxm->full_len -= copied;
1208
1209         /* We have no way to return MSG_EOR. If all the bytes have been
1210          * read we still leave the message in the receive socket buffer.
1211          * A subsequent recvmsg needs to be done to return MSG_EOR and
1212          * finish reading the message.
1213          */
1214
1215         release_sock(sk);
1216
1217         return copied;
1218
1219 err_out:
1220         release_sock(sk);
1221
1222         return err;
1223 }
1224
1225 /* kcm sock lock held */
1226 static void kcm_recv_disable(struct kcm_sock *kcm)
1227 {
1228         struct kcm_mux *mux = kcm->mux;
1229
1230         if (kcm->rx_disabled)
1231                 return;
1232
1233         spin_lock_bh(&mux->rx_lock);
1234
1235         kcm->rx_disabled = 1;
1236
1237         /* If a psock is reserved we'll do cleanup in unreserve */
1238         if (!kcm->rx_psock) {
1239                 if (kcm->rx_wait) {
1240                         list_del(&kcm->wait_rx_list);
1241                         kcm->rx_wait = false;
1242                 }
1243
1244                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1245         }
1246
1247         spin_unlock_bh(&mux->rx_lock);
1248 }
1249
1250 /* kcm sock lock held */
1251 static void kcm_recv_enable(struct kcm_sock *kcm)
1252 {
1253         struct kcm_mux *mux = kcm->mux;
1254
1255         if (!kcm->rx_disabled)
1256                 return;
1257
1258         spin_lock_bh(&mux->rx_lock);
1259
1260         kcm->rx_disabled = 0;
1261         kcm_rcv_ready(kcm);
1262
1263         spin_unlock_bh(&mux->rx_lock);
1264 }
1265
1266 static int kcm_setsockopt(struct socket *sock, int level, int optname,
1267                           char __user *optval, unsigned int optlen)
1268 {
1269         struct kcm_sock *kcm = kcm_sk(sock->sk);
1270         int val, valbool;
1271         int err = 0;
1272
1273         if (level != SOL_KCM)
1274                 return -ENOPROTOOPT;
1275
1276         if (optlen < sizeof(int))
1277                 return -EINVAL;
1278
1279         if (get_user(val, (int __user *)optval))
1280                 return -EINVAL;
1281
1282         valbool = val ? 1 : 0;
1283
1284         switch (optname) {
1285         case KCM_RECV_DISABLE:
1286                 lock_sock(&kcm->sk);
1287                 if (valbool)
1288                         kcm_recv_disable(kcm);
1289                 else
1290                         kcm_recv_enable(kcm);
1291                 release_sock(&kcm->sk);
1292                 break;
1293         default:
1294                 err = -ENOPROTOOPT;
1295         }
1296
1297         return err;
1298 }
1299
1300 static int kcm_getsockopt(struct socket *sock, int level, int optname,
1301                           char __user *optval, int __user *optlen)
1302 {
1303         struct kcm_sock *kcm = kcm_sk(sock->sk);
1304         int val, len;
1305
1306         if (level != SOL_KCM)
1307                 return -ENOPROTOOPT;
1308
1309         if (get_user(len, optlen))
1310                 return -EFAULT;
1311
1312         len = min_t(unsigned int, len, sizeof(int));
1313         if (len < 0)
1314                 return -EINVAL;
1315
1316         switch (optname) {
1317         case KCM_RECV_DISABLE:
1318                 val = kcm->rx_disabled;
1319                 break;
1320         default:
1321                 return -ENOPROTOOPT;
1322         }
1323
1324         if (put_user(len, optlen))
1325                 return -EFAULT;
1326         if (copy_to_user(optval, &val, len))
1327                 return -EFAULT;
1328         return 0;
1329 }
1330
1331 static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1332 {
1333         struct kcm_sock *tkcm;
1334         struct list_head *head;
1335         int index = 0;
1336
1337         /* For SOCK_SEQPACKET sock type, datagram_poll checks the sk_state, so
1338          * we set sk_state, otherwise epoll_wait always returns right away with
1339          * POLLHUP
1340          */
1341         kcm->sk.sk_state = TCP_ESTABLISHED;
1342
1343         /* Add to mux's kcm sockets list */
1344         kcm->mux = mux;
1345         spin_lock_bh(&mux->lock);
1346
1347         head = &mux->kcm_socks;
1348         list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1349                 if (tkcm->index != index)
1350                         break;
1351                 head = &tkcm->kcm_sock_list;
1352                 index++;
1353         }
1354
1355         list_add(&kcm->kcm_sock_list, head);
1356         kcm->index = index;
1357
1358         mux->kcm_socks_cnt++;
1359         spin_unlock_bh(&mux->lock);
1360
1361         INIT_WORK(&kcm->tx_work, kcm_tx_work);
1362
1363         spin_lock_bh(&mux->rx_lock);
1364         kcm_rcv_ready(kcm);
1365         spin_unlock_bh(&mux->rx_lock);
1366 }
1367
1368 static int kcm_attach(struct socket *sock, struct socket *csock,
1369                       struct bpf_prog *prog)
1370 {
1371         struct kcm_sock *kcm = kcm_sk(sock->sk);
1372         struct kcm_mux *mux = kcm->mux;
1373         struct sock *csk;
1374         struct kcm_psock *psock = NULL, *tpsock;
1375         struct list_head *head;
1376         int index = 0;
1377         struct strp_callbacks cb;
1378         int err = 0;
1379
1380         csk = csock->sk;
1381         if (!csk)
1382                 return -EINVAL;
1383
1384         lock_sock(csk);
1385
1386         /* Only allow TCP sockets to be attached for now */
1387         if ((csk->sk_family != AF_INET && csk->sk_family != AF_INET6) ||
1388             csk->sk_protocol != IPPROTO_TCP) {
1389                 err = -EOPNOTSUPP;
1390                 goto out;
1391         }
1392
1393         /* Don't allow listeners or closed sockets */
1394         if (csk->sk_state == TCP_LISTEN || csk->sk_state == TCP_CLOSE) {
1395                 err = -EOPNOTSUPP;
1396                 goto out;
1397         }
1398
1399         psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1400         if (!psock) {
1401                 err = -ENOMEM;
1402                 goto out;
1403         }
1404
1405         psock->mux = mux;
1406         psock->sk = csk;
1407         psock->bpf_prog = prog;
1408
1409         cb.rcv_msg = kcm_rcv_strparser;
1410         cb.abort_parser = NULL;
1411         cb.parse_msg = kcm_parse_func_strparser;
1412         cb.read_sock_done = kcm_read_sock_done;
1413
1414         err = strp_init(&psock->strp, csk, &cb);
1415         if (err) {
1416                 kmem_cache_free(kcm_psockp, psock);
1417                 goto out;
1418         }
1419
1420         write_lock_bh(&csk->sk_callback_lock);
1421
1422         /* Check if sk_user_data is aready by KCM or someone else.
1423          * Must be done under lock to prevent race conditions.
1424          */
1425         if (csk->sk_user_data) {
1426                 write_unlock_bh(&csk->sk_callback_lock);
1427                 strp_stop(&psock->strp);
1428                 strp_done(&psock->strp);
1429                 kmem_cache_free(kcm_psockp, psock);
1430                 err = -EALREADY;
1431                 goto out;
1432         }
1433
1434         psock->save_data_ready = csk->sk_data_ready;
1435         psock->save_write_space = csk->sk_write_space;
1436         psock->save_state_change = csk->sk_state_change;
1437         csk->sk_user_data = psock;
1438         csk->sk_data_ready = psock_data_ready;
1439         csk->sk_write_space = psock_write_space;
1440         csk->sk_state_change = psock_state_change;
1441
1442         write_unlock_bh(&csk->sk_callback_lock);
1443
1444         sock_hold(csk);
1445
1446         /* Finished initialization, now add the psock to the MUX. */
1447         spin_lock_bh(&mux->lock);
1448         head = &mux->psocks;
1449         list_for_each_entry(tpsock, &mux->psocks, psock_list) {
1450                 if (tpsock->index != index)
1451                         break;
1452                 head = &tpsock->psock_list;
1453                 index++;
1454         }
1455
1456         list_add(&psock->psock_list, head);
1457         psock->index = index;
1458
1459         KCM_STATS_INCR(mux->stats.psock_attach);
1460         mux->psocks_cnt++;
1461         psock_now_avail(psock);
1462         spin_unlock_bh(&mux->lock);
1463
1464         /* Schedule RX work in case there are already bytes queued */
1465         strp_check_rcv(&psock->strp);
1466
1467 out:
1468         release_sock(csk);
1469
1470         return err;
1471 }
1472
1473 static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1474 {
1475         struct socket *csock;
1476         struct bpf_prog *prog;
1477         int err;
1478
1479         csock = sockfd_lookup(info->fd, &err);
1480         if (!csock)
1481                 return -ENOENT;
1482
1483         prog = bpf_prog_get_type(info->bpf_fd, BPF_PROG_TYPE_SOCKET_FILTER);
1484         if (IS_ERR(prog)) {
1485                 err = PTR_ERR(prog);
1486                 goto out;
1487         }
1488
1489         err = kcm_attach(sock, csock, prog);
1490         if (err) {
1491                 bpf_prog_put(prog);
1492                 goto out;
1493         }
1494
1495         /* Keep reference on file also */
1496
1497         return 0;
1498 out:
1499         fput(csock->file);
1500         return err;
1501 }
1502
1503 static void kcm_unattach(struct kcm_psock *psock)
1504 {
1505         struct sock *csk = psock->sk;
1506         struct kcm_mux *mux = psock->mux;
1507
1508         lock_sock(csk);
1509
1510         /* Stop getting callbacks from TCP socket. After this there should
1511          * be no way to reserve a kcm for this psock.
1512          */
1513         write_lock_bh(&csk->sk_callback_lock);
1514         csk->sk_user_data = NULL;
1515         csk->sk_data_ready = psock->save_data_ready;
1516         csk->sk_write_space = psock->save_write_space;
1517         csk->sk_state_change = psock->save_state_change;
1518         strp_stop(&psock->strp);
1519
1520         if (WARN_ON(psock->rx_kcm)) {
1521                 write_unlock_bh(&csk->sk_callback_lock);
1522                 release_sock(csk);
1523                 return;
1524         }
1525
1526         spin_lock_bh(&mux->rx_lock);
1527
1528         /* Stop receiver activities. After this point psock should not be
1529          * able to get onto ready list either through callbacks or work.
1530          */
1531         if (psock->ready_rx_msg) {
1532                 list_del(&psock->psock_ready_list);
1533                 kfree_skb(psock->ready_rx_msg);
1534                 psock->ready_rx_msg = NULL;
1535                 KCM_STATS_INCR(mux->stats.rx_ready_drops);
1536         }
1537
1538         spin_unlock_bh(&mux->rx_lock);
1539
1540         write_unlock_bh(&csk->sk_callback_lock);
1541
1542         /* Call strp_done without sock lock */
1543         release_sock(csk);
1544         strp_done(&psock->strp);
1545         lock_sock(csk);
1546
1547         bpf_prog_put(psock->bpf_prog);
1548
1549         spin_lock_bh(&mux->lock);
1550
1551         aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1552         save_strp_stats(&psock->strp, &mux->aggregate_strp_stats);
1553
1554         KCM_STATS_INCR(mux->stats.psock_unattach);
1555
1556         if (psock->tx_kcm) {
1557                 /* psock was reserved.  Just mark it finished and we will clean
1558                  * up in the kcm paths, we need kcm lock which can not be
1559                  * acquired here.
1560                  */
1561                 KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1562                 spin_unlock_bh(&mux->lock);
1563
1564                 /* We are unattaching a socket that is reserved. Abort the
1565                  * socket since we may be out of sync in sending on it. We need
1566                  * to do this without the mux lock.
1567                  */
1568                 kcm_abort_tx_psock(psock, EPIPE, false);
1569
1570                 spin_lock_bh(&mux->lock);
1571                 if (!psock->tx_kcm) {
1572                         /* psock now unreserved in window mux was unlocked */
1573                         goto no_reserved;
1574                 }
1575                 psock->done = 1;
1576
1577                 /* Commit done before queuing work to process it */
1578                 smp_mb();
1579
1580                 /* Queue tx work to make sure psock->done is handled */
1581                 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
1582                 spin_unlock_bh(&mux->lock);
1583         } else {
1584 no_reserved:
1585                 if (!psock->tx_stopped)
1586                         list_del(&psock->psock_avail_list);
1587                 list_del(&psock->psock_list);
1588                 mux->psocks_cnt--;
1589                 spin_unlock_bh(&mux->lock);
1590
1591                 sock_put(csk);
1592                 fput(csk->sk_socket->file);
1593                 kmem_cache_free(kcm_psockp, psock);
1594         }
1595
1596         release_sock(csk);
1597 }
1598
1599 static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1600 {
1601         struct kcm_sock *kcm = kcm_sk(sock->sk);
1602         struct kcm_mux *mux = kcm->mux;
1603         struct kcm_psock *psock;
1604         struct socket *csock;
1605         struct sock *csk;
1606         int err;
1607
1608         csock = sockfd_lookup(info->fd, &err);
1609         if (!csock)
1610                 return -ENOENT;
1611
1612         csk = csock->sk;
1613         if (!csk) {
1614                 err = -EINVAL;
1615                 goto out;
1616         }
1617
1618         err = -ENOENT;
1619
1620         spin_lock_bh(&mux->lock);
1621
1622         list_for_each_entry(psock, &mux->psocks, psock_list) {
1623                 if (psock->sk != csk)
1624                         continue;
1625
1626                 /* Found the matching psock */
1627
1628                 if (psock->unattaching || WARN_ON(psock->done)) {
1629                         err = -EALREADY;
1630                         break;
1631                 }
1632
1633                 psock->unattaching = 1;
1634
1635                 spin_unlock_bh(&mux->lock);
1636
1637                 /* Lower socket lock should already be held */
1638                 kcm_unattach(psock);
1639
1640                 err = 0;
1641                 goto out;
1642         }
1643
1644         spin_unlock_bh(&mux->lock);
1645
1646 out:
1647         fput(csock->file);
1648         return err;
1649 }
1650
1651 static struct proto kcm_proto = {
1652         .name   = "KCM",
1653         .owner  = THIS_MODULE,
1654         .obj_size = sizeof(struct kcm_sock),
1655 };
1656
1657 /* Clone a kcm socket. */
1658 static struct file *kcm_clone(struct socket *osock)
1659 {
1660         struct socket *newsock;
1661         struct sock *newsk;
1662         struct file *file;
1663
1664         newsock = sock_alloc();
1665         if (!newsock)
1666                 return ERR_PTR(-ENFILE);
1667
1668         newsock->type = osock->type;
1669         newsock->ops = osock->ops;
1670
1671         __module_get(newsock->ops->owner);
1672
1673         newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1674                          &kcm_proto, false);
1675         if (!newsk) {
1676                 sock_release(newsock);
1677                 return ERR_PTR(-ENOMEM);
1678         }
1679         sock_init_data(newsock, newsk);
1680         init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1681
1682         file = sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1683         if (IS_ERR(file))
1684                 sock_release(newsock);
1685
1686         return file;
1687 }
1688
1689 static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
1690 {
1691         int err;
1692
1693         switch (cmd) {
1694         case SIOCKCMATTACH: {
1695                 struct kcm_attach info;
1696
1697                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1698                         return -EFAULT;
1699
1700                 err = kcm_attach_ioctl(sock, &info);
1701
1702                 break;
1703         }
1704         case SIOCKCMUNATTACH: {
1705                 struct kcm_unattach info;
1706
1707                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1708                         return -EFAULT;
1709
1710                 err = kcm_unattach_ioctl(sock, &info);
1711
1712                 break;
1713         }
1714         case SIOCKCMCLONE: {
1715                 struct kcm_clone info;
1716                 struct file *file;
1717
1718                 info.fd = get_unused_fd_flags(0);
1719                 if (unlikely(info.fd < 0))
1720                         return info.fd;
1721
1722                 file = kcm_clone(sock);
1723                 if (IS_ERR(file)) {
1724                         put_unused_fd(info.fd);
1725                         return PTR_ERR(file);
1726                 }
1727                 if (copy_to_user((void __user *)arg, &info,
1728                                  sizeof(info))) {
1729                         put_unused_fd(info.fd);
1730                         fput(file);
1731                         return -EFAULT;
1732                 }
1733                 fd_install(info.fd, file);
1734                 err = 0;
1735                 break;
1736         }
1737         default:
1738                 err = -ENOIOCTLCMD;
1739                 break;
1740         }
1741
1742         return err;
1743 }
1744
1745 static void free_mux(struct rcu_head *rcu)
1746 {
1747         struct kcm_mux *mux = container_of(rcu,
1748             struct kcm_mux, rcu);
1749
1750         kmem_cache_free(kcm_muxp, mux);
1751 }
1752
1753 static void release_mux(struct kcm_mux *mux)
1754 {
1755         struct kcm_net *knet = mux->knet;
1756         struct kcm_psock *psock, *tmp_psock;
1757
1758         /* Release psocks */
1759         list_for_each_entry_safe(psock, tmp_psock,
1760                                  &mux->psocks, psock_list) {
1761                 if (!WARN_ON(psock->unattaching))
1762                         kcm_unattach(psock);
1763         }
1764
1765         if (WARN_ON(mux->psocks_cnt))
1766                 return;
1767
1768         __skb_queue_purge(&mux->rx_hold_queue);
1769
1770         mutex_lock(&knet->mutex);
1771         aggregate_mux_stats(&mux->stats, &knet->aggregate_mux_stats);
1772         aggregate_psock_stats(&mux->aggregate_psock_stats,
1773                               &knet->aggregate_psock_stats);
1774         aggregate_strp_stats(&mux->aggregate_strp_stats,
1775                              &knet->aggregate_strp_stats);
1776         list_del_rcu(&mux->kcm_mux_list);
1777         knet->count--;
1778         mutex_unlock(&knet->mutex);
1779
1780         call_rcu(&mux->rcu, free_mux);
1781 }
1782
1783 static void kcm_done(struct kcm_sock *kcm)
1784 {
1785         struct kcm_mux *mux = kcm->mux;
1786         struct sock *sk = &kcm->sk;
1787         int socks_cnt;
1788
1789         spin_lock_bh(&mux->rx_lock);
1790         if (kcm->rx_psock) {
1791                 /* Cleanup in unreserve_rx_kcm */
1792                 WARN_ON(kcm->done);
1793                 kcm->rx_disabled = 1;
1794                 kcm->done = 1;
1795                 spin_unlock_bh(&mux->rx_lock);
1796                 return;
1797         }
1798
1799         if (kcm->rx_wait) {
1800                 list_del(&kcm->wait_rx_list);
1801                 kcm->rx_wait = false;
1802         }
1803         /* Move any pending receive messages to other kcm sockets */
1804         requeue_rx_msgs(mux, &sk->sk_receive_queue);
1805
1806         spin_unlock_bh(&mux->rx_lock);
1807
1808         if (WARN_ON(sk_rmem_alloc_get(sk)))
1809                 return;
1810
1811         /* Detach from MUX */
1812         spin_lock_bh(&mux->lock);
1813
1814         list_del(&kcm->kcm_sock_list);
1815         mux->kcm_socks_cnt--;
1816         socks_cnt = mux->kcm_socks_cnt;
1817
1818         spin_unlock_bh(&mux->lock);
1819
1820         if (!socks_cnt) {
1821                 /* We are done with the mux now. */
1822                 release_mux(mux);
1823         }
1824
1825         WARN_ON(kcm->rx_wait);
1826
1827         sock_put(&kcm->sk);
1828 }
1829
1830 /* Called by kcm_release to close a KCM socket.
1831  * If this is the last KCM socket on the MUX, destroy the MUX.
1832  */
1833 static int kcm_release(struct socket *sock)
1834 {
1835         struct sock *sk = sock->sk;
1836         struct kcm_sock *kcm;
1837         struct kcm_mux *mux;
1838         struct kcm_psock *psock;
1839
1840         if (!sk)
1841                 return 0;
1842
1843         kcm = kcm_sk(sk);
1844         mux = kcm->mux;
1845
1846         sock_orphan(sk);
1847         kfree_skb(kcm->seq_skb);
1848
1849         lock_sock(sk);
1850         /* Purge queue under lock to avoid race condition with tx_work trying
1851          * to act when queue is nonempty. If tx_work runs after this point
1852          * it will just return.
1853          */
1854         __skb_queue_purge(&sk->sk_write_queue);
1855
1856         /* Set tx_stopped. This is checked when psock is bound to a kcm and we
1857          * get a writespace callback. This prevents further work being queued
1858          * from the callback (unbinding the psock occurs after canceling work.
1859          */
1860         kcm->tx_stopped = 1;
1861
1862         release_sock(sk);
1863
1864         spin_lock_bh(&mux->lock);
1865         if (kcm->tx_wait) {
1866                 /* Take of tx_wait list, after this point there should be no way
1867                  * that a psock will be assigned to this kcm.
1868                  */
1869                 list_del(&kcm->wait_psock_list);
1870                 kcm->tx_wait = false;
1871         }
1872         spin_unlock_bh(&mux->lock);
1873
1874         /* Cancel work. After this point there should be no outside references
1875          * to the kcm socket.
1876          */
1877         cancel_work_sync(&kcm->tx_work);
1878
1879         lock_sock(sk);
1880         psock = kcm->tx_psock;
1881         if (psock) {
1882                 /* A psock was reserved, so we need to kill it since it
1883                  * may already have some bytes queued from a message. We
1884                  * need to do this after removing kcm from tx_wait list.
1885                  */
1886                 kcm_abort_tx_psock(psock, EPIPE, false);
1887                 unreserve_psock(kcm);
1888         }
1889         release_sock(sk);
1890
1891         WARN_ON(kcm->tx_wait);
1892         WARN_ON(kcm->tx_psock);
1893
1894         sock->sk = NULL;
1895
1896         kcm_done(kcm);
1897
1898         return 0;
1899 }
1900
1901 static const struct proto_ops kcm_dgram_ops = {
1902         .family =       PF_KCM,
1903         .owner =        THIS_MODULE,
1904         .release =      kcm_release,
1905         .bind =         sock_no_bind,
1906         .connect =      sock_no_connect,
1907         .socketpair =   sock_no_socketpair,
1908         .accept =       sock_no_accept,
1909         .getname =      sock_no_getname,
1910         .poll =         datagram_poll,
1911         .ioctl =        kcm_ioctl,
1912         .listen =       sock_no_listen,
1913         .shutdown =     sock_no_shutdown,
1914         .setsockopt =   kcm_setsockopt,
1915         .getsockopt =   kcm_getsockopt,
1916         .sendmsg =      kcm_sendmsg,
1917         .recvmsg =      kcm_recvmsg,
1918         .mmap =         sock_no_mmap,
1919         .sendpage =     kcm_sendpage,
1920 };
1921
1922 static const struct proto_ops kcm_seqpacket_ops = {
1923         .family =       PF_KCM,
1924         .owner =        THIS_MODULE,
1925         .release =      kcm_release,
1926         .bind =         sock_no_bind,
1927         .connect =      sock_no_connect,
1928         .socketpair =   sock_no_socketpair,
1929         .accept =       sock_no_accept,
1930         .getname =      sock_no_getname,
1931         .poll =         datagram_poll,
1932         .ioctl =        kcm_ioctl,
1933         .listen =       sock_no_listen,
1934         .shutdown =     sock_no_shutdown,
1935         .setsockopt =   kcm_setsockopt,
1936         .getsockopt =   kcm_getsockopt,
1937         .sendmsg =      kcm_sendmsg,
1938         .recvmsg =      kcm_recvmsg,
1939         .mmap =         sock_no_mmap,
1940         .sendpage =     kcm_sendpage,
1941         .splice_read =  kcm_splice_read,
1942 };
1943
1944 /* Create proto operation for kcm sockets */
1945 static int kcm_create(struct net *net, struct socket *sock,
1946                       int protocol, int kern)
1947 {
1948         struct kcm_net *knet = net_generic(net, kcm_net_id);
1949         struct sock *sk;
1950         struct kcm_mux *mux;
1951
1952         switch (sock->type) {
1953         case SOCK_DGRAM:
1954                 sock->ops = &kcm_dgram_ops;
1955                 break;
1956         case SOCK_SEQPACKET:
1957                 sock->ops = &kcm_seqpacket_ops;
1958                 break;
1959         default:
1960                 return -ESOCKTNOSUPPORT;
1961         }
1962
1963         if (protocol != KCMPROTO_CONNECTED)
1964                 return -EPROTONOSUPPORT;
1965
1966         sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
1967         if (!sk)
1968                 return -ENOMEM;
1969
1970         /* Allocate a kcm mux, shared between KCM sockets */
1971         mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
1972         if (!mux) {
1973                 sk_free(sk);
1974                 return -ENOMEM;
1975         }
1976
1977         spin_lock_init(&mux->lock);
1978         spin_lock_init(&mux->rx_lock);
1979         INIT_LIST_HEAD(&mux->kcm_socks);
1980         INIT_LIST_HEAD(&mux->kcm_rx_waiters);
1981         INIT_LIST_HEAD(&mux->kcm_tx_waiters);
1982
1983         INIT_LIST_HEAD(&mux->psocks);
1984         INIT_LIST_HEAD(&mux->psocks_ready);
1985         INIT_LIST_HEAD(&mux->psocks_avail);
1986
1987         mux->knet = knet;
1988
1989         /* Add new MUX to list */
1990         mutex_lock(&knet->mutex);
1991         list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
1992         knet->count++;
1993         mutex_unlock(&knet->mutex);
1994
1995         skb_queue_head_init(&mux->rx_hold_queue);
1996
1997         /* Init KCM socket */
1998         sock_init_data(sock, sk);
1999         init_kcm_sock(kcm_sk(sk), mux);
2000
2001         return 0;
2002 }
2003
2004 static struct net_proto_family kcm_family_ops = {
2005         .family = PF_KCM,
2006         .create = kcm_create,
2007         .owner  = THIS_MODULE,
2008 };
2009
2010 static __net_init int kcm_init_net(struct net *net)
2011 {
2012         struct kcm_net *knet = net_generic(net, kcm_net_id);
2013
2014         INIT_LIST_HEAD_RCU(&knet->mux_list);
2015         mutex_init(&knet->mutex);
2016
2017         return 0;
2018 }
2019
2020 static __net_exit void kcm_exit_net(struct net *net)
2021 {
2022         struct kcm_net *knet = net_generic(net, kcm_net_id);
2023
2024         /* All KCM sockets should be closed at this point, which should mean
2025          * that all multiplexors and psocks have been destroyed.
2026          */
2027         WARN_ON(!list_empty(&knet->mux_list));
2028 }
2029
2030 static struct pernet_operations kcm_net_ops = {
2031         .init = kcm_init_net,
2032         .exit = kcm_exit_net,
2033         .id   = &kcm_net_id,
2034         .size = sizeof(struct kcm_net),
2035 };
2036
2037 static int __init kcm_init(void)
2038 {
2039         int err = -ENOMEM;
2040
2041         kcm_muxp = kmem_cache_create("kcm_mux_cache",
2042                                      sizeof(struct kcm_mux), 0,
2043                                      SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2044         if (!kcm_muxp)
2045                 goto fail;
2046
2047         kcm_psockp = kmem_cache_create("kcm_psock_cache",
2048                                        sizeof(struct kcm_psock), 0,
2049                                         SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2050         if (!kcm_psockp)
2051                 goto fail;
2052
2053         kcm_wq = create_singlethread_workqueue("kkcmd");
2054         if (!kcm_wq)
2055                 goto fail;
2056
2057         err = proto_register(&kcm_proto, 1);
2058         if (err)
2059                 goto fail;
2060
2061         err = register_pernet_device(&kcm_net_ops);
2062         if (err)
2063                 goto net_ops_fail;
2064
2065         err = sock_register(&kcm_family_ops);
2066         if (err)
2067                 goto sock_register_fail;
2068
2069         err = kcm_proc_init();
2070         if (err)
2071                 goto proc_init_fail;
2072
2073         return 0;
2074
2075 proc_init_fail:
2076         sock_unregister(PF_KCM);
2077
2078 sock_register_fail:
2079         unregister_pernet_device(&kcm_net_ops);
2080
2081 net_ops_fail:
2082         proto_unregister(&kcm_proto);
2083
2084 fail:
2085         kmem_cache_destroy(kcm_muxp);
2086         kmem_cache_destroy(kcm_psockp);
2087
2088         if (kcm_wq)
2089                 destroy_workqueue(kcm_wq);
2090
2091         return err;
2092 }
2093
2094 static void __exit kcm_exit(void)
2095 {
2096         kcm_proc_exit();
2097         sock_unregister(PF_KCM);
2098         unregister_pernet_device(&kcm_net_ops);
2099         proto_unregister(&kcm_proto);
2100         destroy_workqueue(kcm_wq);
2101
2102         kmem_cache_destroy(kcm_muxp);
2103         kmem_cache_destroy(kcm_psockp);
2104 }
2105
2106 module_init(kcm_init);
2107 module_exit(kcm_exit);
2108
2109 MODULE_LICENSE("GPL");
2110 MODULE_ALIAS_NETPROTO(PF_KCM);
2111