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