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 <linux/sched/signal.h>
  28
  29 #include <net/kcm.h>
  30 #include <net/netns/generic.h>
  31 #include <net/sock.h>
  32 #include <uapi/linux/kcm.h>
  33
  34 unsigned int kcm_net_id;
  35
  36 static struct kmem_cache *kcm_psockp __read_mostly;
  37 static struct kmem_cache *kcm_muxp __read_mostly;
  38 static struct workqueue_struct *kcm_wq;
  39
  40 static inline struct kcm_sock *kcm_sk(const struct sock *sk)
  41 {
  42         return (struct kcm_sock *)sk;
  43 }
  44
  45 static inline struct kcm_tx_msg *kcm_tx_msg(struct sk_buff *skb)
  46 {
  47         return (struct kcm_tx_msg *)skb->cb;
  48 }
  49
  50 static void report_csk_error(struct sock *csk, int err)
  51 {
  52         csk->sk_err = EPIPE;
  53         csk->sk_error_report(csk);
  54 }
  55
  56 static void kcm_abort_tx_psock(struct kcm_psock *psock, int err,
  57                                bool wakeup_kcm)
  58 {
  59         struct sock *csk = psock->sk;
  60         struct kcm_mux *mux = psock->mux;
  61
  62         /* Unrecoverable error in transmit */
  63
  64         spin_lock_bh(&mux->lock);
  65
  66         if (psock->tx_stopped) {
  67                 spin_unlock_bh(&mux->lock);
  68                 return;
  69         }
  70
  71         psock->tx_stopped = 1;
  72         KCM_STATS_INCR(psock->stats.tx_aborts);
  73
  74         if (!psock->tx_kcm) {
  75                 /* Take off psocks_avail list */
  76                 list_del(&psock->psock_avail_list);
  77         } else if (wakeup_kcm) {
  78                 /* In this case psock is being aborted while outside of
  79                  * write_msgs and psock is reserved. Schedule tx_work
  80                  * to handle the failure there. Need to commit tx_stopped
  81                  * before queuing work.
  82                  */
  83                 smp_mb();
  84
  85                 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
  86         }
  87
  88         spin_unlock_bh(&mux->lock);
  89
  90         /* Report error on lower socket */
  91         report_csk_error(csk, err);
  92 }
  93
  94 /* RX mux lock held. */
  95 static void kcm_update_rx_mux_stats(struct kcm_mux *mux,
  96                                     struct kcm_psock *psock)
  97 {
  98         STRP_STATS_ADD(mux->stats.rx_bytes,
  99                        psock->strp.stats.bytes -
 100                        psock->saved_rx_bytes);
 101         mux->stats.rx_msgs +=
 102                 psock->strp.stats.msgs - psock->saved_rx_msgs;
 103         psock->saved_rx_msgs = psock->strp.stats.msgs;
 104         psock->saved_rx_bytes = psock->strp.stats.bytes;
 105 }
 106
 107 static void kcm_update_tx_mux_stats(struct kcm_mux *mux,
 108                                     struct kcm_psock *psock)
 109 {
 110         KCM_STATS_ADD(mux->stats.tx_bytes,
 111                       psock->stats.tx_bytes - psock->saved_tx_bytes);
 112         mux->stats.tx_msgs +=
 113                 psock->stats.tx_msgs - psock->saved_tx_msgs;
 114         psock->saved_tx_msgs = psock->stats.tx_msgs;
 115         psock->saved_tx_bytes = psock->stats.tx_bytes;
 116 }
 117
 118 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb);
 119
 120 /* KCM is ready to receive messages on its queue-- either the KCM is new or
 121  * has become unblocked after being blocked on full socket buffer. Queue any
 122  * pending ready messages on a psock. RX mux lock held.
 123  */
 124 static void kcm_rcv_ready(struct kcm_sock *kcm)
 125 {
 126         struct kcm_mux *mux = kcm->mux;
 127         struct kcm_psock *psock;
 128         struct sk_buff *skb;
 129
 130         if (unlikely(kcm->rx_wait || kcm->rx_psock || kcm->rx_disabled))
 131                 return;
 132
 133         while (unlikely((skb = __skb_dequeue(&mux->rx_hold_queue)))) {
 134                 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 135                         /* Assuming buffer limit has been reached */
 136                         skb_queue_head(&mux->rx_hold_queue, skb);
 137                         WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
 138                         return;
 139                 }
 140         }
 141
 142         while (!list_empty(&mux->psocks_ready)) {
 143                 psock = list_first_entry(&mux->psocks_ready, struct kcm_psock,
 144                                          psock_ready_list);
 145
 146                 if (kcm_queue_rcv_skb(&kcm->sk, psock->ready_rx_msg)) {
 147                         /* Assuming buffer limit has been reached */
 148                         WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
 149                         return;
 150                 }
 151
 152                 /* Consumed the ready message on the psock. Schedule rx_work to
 153                  * get more messages.
 154                  */
 155                 list_del(&psock->psock_ready_list);
 156                 psock->ready_rx_msg = NULL;
 157                 /* Commit clearing of ready_rx_msg for queuing work */
 158                 smp_mb();
 159
 160                 strp_unpause(&psock->strp);
 161                 strp_check_rcv(&psock->strp);
 162         }
 163
 164         /* Buffer limit is okay now, add to ready list */
 165         list_add_tail(&kcm->wait_rx_list,
 166                       &kcm->mux->kcm_rx_waiters);
 167         /* paired with lockless reads in kcm_rfree() */
 168         WRITE_ONCE(kcm->rx_wait, true);
 169 }
 170
 171 static void kcm_rfree(struct sk_buff *skb)
 172 {
 173         struct sock *sk = skb->sk;
 174         struct kcm_sock *kcm = kcm_sk(sk);
 175         struct kcm_mux *mux = kcm->mux;
 176         unsigned int len = skb->truesize;
 177
 178         sk_mem_uncharge(sk, len);
 179         atomic_sub(len, &sk->sk_rmem_alloc);
 180
 181         /* For reading rx_wait and rx_psock without holding lock */
 182         smp_mb__after_atomic();
 183
 184         if (!READ_ONCE(kcm->rx_wait) && !READ_ONCE(kcm->rx_psock) &&
 185             sk_rmem_alloc_get(sk) < sk->sk_rcvlowat) {
 186                 spin_lock_bh(&mux->rx_lock);
 187                 kcm_rcv_ready(kcm);
 188                 spin_unlock_bh(&mux->rx_lock);
 189         }
 190 }
 191
 192 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb)
 193 {
 194         struct sk_buff_head *list = &sk->sk_receive_queue;
 195
 196         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
 197                 return -ENOMEM;
 198
 199         if (!sk_rmem_schedule(sk, skb, skb->truesize))
 200                 return -ENOBUFS;
 201
 202         skb->dev = NULL;
 203
 204         skb_orphan(skb);
 205         skb->sk = sk;
 206         skb->destructor = kcm_rfree;
 207         atomic_add(skb->truesize, &sk->sk_rmem_alloc);
 208         sk_mem_charge(sk, skb->truesize);
 209
 210         skb_queue_tail(list, skb);
 211
 212         if (!sock_flag(sk, SOCK_DEAD))
 213                 sk->sk_data_ready(sk);
 214
 215         return 0;
 216 }
 217
 218 /* Requeue received messages for a kcm socket to other kcm sockets. This is
 219  * called with a kcm socket is receive disabled.
 220  * RX mux lock held.
 221  */
 222 static void requeue_rx_msgs(struct kcm_mux *mux, struct sk_buff_head *head)
 223 {
 224         struct sk_buff *skb;
 225         struct kcm_sock *kcm;
 226
 227         while ((skb = skb_dequeue(head))) {
 228                 /* Reset destructor to avoid calling kcm_rcv_ready */
 229                 skb->destructor = sock_rfree;
 230                 skb_orphan(skb);
 231 try_again:
 232                 if (list_empty(&mux->kcm_rx_waiters)) {
 233                         skb_queue_tail(&mux->rx_hold_queue, skb);
 234                         continue;
 235                 }
 236
 237                 kcm = list_first_entry(&mux->kcm_rx_waiters,
 238                                        struct kcm_sock, wait_rx_list);
 239
 240                 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 241                         /* Should mean socket buffer full */
 242                         list_del(&kcm->wait_rx_list);
 243                         /* paired with lockless reads in kcm_rfree() */
 244                         WRITE_ONCE(kcm->rx_wait, false);
 245
 246                         /* Commit rx_wait to read in kcm_free */
 247                         smp_wmb();
 248
 249                         goto try_again;
 250                 }
 251         }
 252 }
 253
 254 /* Lower sock lock held */
 255 static struct kcm_sock *reserve_rx_kcm(struct kcm_psock *psock,
 256                                        struct sk_buff *head)
 257 {
 258         struct kcm_mux *mux = psock->mux;
 259         struct kcm_sock *kcm;
 260
 261         WARN_ON(psock->ready_rx_msg);
 262
 263         if (psock->rx_kcm)
 264                 return psock->rx_kcm;
 265
 266         spin_lock_bh(&mux->rx_lock);
 267
 268         if (psock->rx_kcm) {
 269                 spin_unlock_bh(&mux->rx_lock);
 270                 return psock->rx_kcm;
 271         }
 272
 273         kcm_update_rx_mux_stats(mux, psock);
 274
 275         if (list_empty(&mux->kcm_rx_waiters)) {
 276                 psock->ready_rx_msg = head;
 277                 strp_pause(&psock->strp);
 278                 list_add_tail(&psock->psock_ready_list,
 279                               &mux->psocks_ready);
 280                 spin_unlock_bh(&mux->rx_lock);
 281                 return NULL;
 282         }
 283
 284         kcm = list_first_entry(&mux->kcm_rx_waiters,
 285                                struct kcm_sock, wait_rx_list);
 286         list_del(&kcm->wait_rx_list);
 287         /* paired with lockless reads in kcm_rfree() */
 288         WRITE_ONCE(kcm->rx_wait, false);
 289
 290         psock->rx_kcm = kcm;
 291         /* paired with lockless reads in kcm_rfree() */
 292         WRITE_ONCE(kcm->rx_psock, psock);
 293
 294         spin_unlock_bh(&mux->rx_lock);
 295
 296         return kcm;
 297 }
 298
 299 static void kcm_done(struct kcm_sock *kcm);
 300
 301 static void kcm_done_work(struct work_struct *w)
 302 {
 303         kcm_done(container_of(w, struct kcm_sock, done_work));
 304 }
 305
 306 /* Lower sock held */
 307 static void unreserve_rx_kcm(struct kcm_psock *psock,
 308                              bool rcv_ready)
 309 {
 310         struct kcm_sock *kcm = psock->rx_kcm;
 311         struct kcm_mux *mux = psock->mux;
 312
 313         if (!kcm)
 314                 return;
 315
 316         spin_lock_bh(&mux->rx_lock);
 317
 318         psock->rx_kcm = NULL;
 319         /* paired with lockless reads in kcm_rfree() */
 320         WRITE_ONCE(kcm->rx_psock, NULL);
 321
 322         /* Commit kcm->rx_psock before sk_rmem_alloc_get to sync with
 323          * kcm_rfree
 324          */
 325         smp_mb();
 326
 327         if (unlikely(kcm->done)) {
 328                 spin_unlock_bh(&mux->rx_lock);
 329
 330                 /* Need to run kcm_done in a task since we need to qcquire
 331                  * callback locks which may already be held here.
 332                  */
 333                 INIT_WORK(&kcm->done_work, kcm_done_work);
 334                 schedule_work(&kcm->done_work);
 335                 return;
 336         }
 337
 338         if (unlikely(kcm->rx_disabled)) {
 339                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
 340         } else if (rcv_ready || unlikely(!sk_rmem_alloc_get(&kcm->sk))) {
 341                 /* Check for degenerative race with rx_wait that all
 342                  * data was dequeued (accounted for in kcm_rfree).
 343                  */
 344                 kcm_rcv_ready(kcm);
 345         }
 346         spin_unlock_bh(&mux->rx_lock);
 347 }
 348
 349 /* Lower sock lock held */
 350 static void psock_data_ready(struct sock *sk)
 351 {
 352         struct kcm_psock *psock;
 353
 354         read_lock_bh(&sk->sk_callback_lock);
 355
 356         psock = (struct kcm_psock *)sk->sk_user_data;
 357         if (likely(psock))
 358                 strp_data_ready(&psock->strp);
 359
 360         read_unlock_bh(&sk->sk_callback_lock);
 361 }
 362
 363 /* Called with lower sock held */
 364 static void kcm_rcv_strparser(struct strparser *strp, struct sk_buff *skb)
 365 {
 366         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 367         struct kcm_sock *kcm;
 368
 369 try_queue:
 370         kcm = reserve_rx_kcm(psock, skb);
 371         if (!kcm) {
 372                  /* Unable to reserve a KCM, message is held in psock and strp
 373                   * is paused.
 374                   */
 375                 return;
 376         }
 377
 378         if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 379                 /* Should mean socket buffer full */
 380                 unreserve_rx_kcm(psock, false);
 381                 goto try_queue;
 382         }
 383 }
 384
 385 static int kcm_parse_func_strparser(struct strparser *strp, struct sk_buff *skb)
 386 {
 387         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 388         struct bpf_prog *prog = psock->bpf_prog;
 389
 390         return (*prog->bpf_func)(skb, prog->insnsi);
 391 }
 392
 393 static int kcm_read_sock_done(struct strparser *strp, int err)
 394 {
 395         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 396
 397         unreserve_rx_kcm(psock, true);
 398
 399         return err;
 400 }
 401
 402 static void psock_state_change(struct sock *sk)
 403 {
 404         /* TCP only does a EPOLLIN for a half close. Do a EPOLLHUP here
 405          * since application will normally not poll with EPOLLIN
 406          * on the TCP sockets.
 407          */
 408
 409         report_csk_error(sk, EPIPE);
 410 }
 411
 412 static void psock_write_space(struct sock *sk)
 413 {
 414         struct kcm_psock *psock;
 415         struct kcm_mux *mux;
 416         struct kcm_sock *kcm;
 417
 418         read_lock_bh(&sk->sk_callback_lock);
 419
 420         psock = (struct kcm_psock *)sk->sk_user_data;
 421         if (unlikely(!psock))
 422                 goto out;
 423         mux = psock->mux;
 424
 425         spin_lock_bh(&mux->lock);
 426
 427         /* Check if the socket is reserved so someone is waiting for sending. */
 428         kcm = psock->tx_kcm;
 429         if (kcm && !unlikely(kcm->tx_stopped))
 430                 queue_work(kcm_wq, &kcm->tx_work);
 431
 432         spin_unlock_bh(&mux->lock);
 433 out:
 434         read_unlock_bh(&sk->sk_callback_lock);
 435 }
 436
 437 static void unreserve_psock(struct kcm_sock *kcm);
 438
 439 /* kcm sock is locked. */
 440 static struct kcm_psock *reserve_psock(struct kcm_sock *kcm)
 441 {
 442         struct kcm_mux *mux = kcm->mux;
 443         struct kcm_psock *psock;
 444
 445         psock = kcm->tx_psock;
 446
 447         smp_rmb(); /* Must read tx_psock before tx_wait */
 448
 449         if (psock) {
 450                 WARN_ON(kcm->tx_wait);
 451                 if (unlikely(psock->tx_stopped))
 452                         unreserve_psock(kcm);
 453                 else
 454                         return kcm->tx_psock;
 455         }
 456
 457         spin_lock_bh(&mux->lock);
 458
 459         /* Check again under lock to see if psock was reserved for this
 460          * psock via psock_unreserve.
 461          */
 462         psock = kcm->tx_psock;
 463         if (unlikely(psock)) {
 464                 WARN_ON(kcm->tx_wait);
 465                 spin_unlock_bh(&mux->lock);
 466                 return kcm->tx_psock;
 467         }
 468
 469         if (!list_empty(&mux->psocks_avail)) {
 470                 psock = list_first_entry(&mux->psocks_avail,
 471                                          struct kcm_psock,
 472                                          psock_avail_list);
 473                 list_del(&psock->psock_avail_list);
 474                 if (kcm->tx_wait) {
 475                         list_del(&kcm->wait_psock_list);
 476                         kcm->tx_wait = false;
 477                 }
 478                 kcm->tx_psock = psock;
 479                 psock->tx_kcm = kcm;
 480                 KCM_STATS_INCR(psock->stats.reserved);
 481         } else if (!kcm->tx_wait) {
 482                 list_add_tail(&kcm->wait_psock_list,
 483                               &mux->kcm_tx_waiters);
 484                 kcm->tx_wait = true;
 485         }
 486
 487         spin_unlock_bh(&mux->lock);
 488
 489         return psock;
 490 }
 491
 492 /* mux lock held */
 493 static void psock_now_avail(struct kcm_psock *psock)
 494 {
 495         struct kcm_mux *mux = psock->mux;
 496         struct kcm_sock *kcm;
 497
 498         if (list_empty(&mux->kcm_tx_waiters)) {
 499                 list_add_tail(&psock->psock_avail_list,
 500                               &mux->psocks_avail);
 501         } else {
 502                 kcm = list_first_entry(&mux->kcm_tx_waiters,
 503                                        struct kcm_sock,
 504                                        wait_psock_list);
 505                 list_del(&kcm->wait_psock_list);
 506                 kcm->tx_wait = false;
 507                 psock->tx_kcm = kcm;
 508
 509                 /* Commit before changing tx_psock since that is read in
 510                  * reserve_psock before queuing work.
 511                  */
 512                 smp_mb();
 513
 514                 kcm->tx_psock = psock;
 515                 KCM_STATS_INCR(psock->stats.reserved);
 516                 queue_work(kcm_wq, &kcm->tx_work);
 517         }
 518 }
 519
 520 /* kcm sock is locked. */
 521 static void unreserve_psock(struct kcm_sock *kcm)
 522 {
 523         struct kcm_psock *psock;
 524         struct kcm_mux *mux = kcm->mux;
 525
 526         spin_lock_bh(&mux->lock);
 527
 528         psock = kcm->tx_psock;
 529
 530         if (WARN_ON(!psock)) {
 531                 spin_unlock_bh(&mux->lock);
 532                 return;
 533         }
 534
 535         smp_rmb(); /* Read tx_psock before tx_wait */
 536
 537         kcm_update_tx_mux_stats(mux, psock);
 538
 539         WARN_ON(kcm->tx_wait);
 540
 541         kcm->tx_psock = NULL;
 542         psock->tx_kcm = NULL;
 543         KCM_STATS_INCR(psock->stats.unreserved);
 544
 545         if (unlikely(psock->tx_stopped)) {
 546                 if (psock->done) {
 547                         /* Deferred free */
 548                         list_del(&psock->psock_list);
 549                         mux->psocks_cnt--;
 550                         sock_put(psock->sk);
 551                         fput(psock->sk->sk_socket->file);
 552                         kmem_cache_free(kcm_psockp, psock);
 553                 }
 554
 555                 /* Don't put back on available list */
 556
 557                 spin_unlock_bh(&mux->lock);
 558
 559                 return;
 560         }
 561
 562         psock_now_avail(psock);
 563
 564         spin_unlock_bh(&mux->lock);
 565 }
 566
 567 static void kcm_report_tx_retry(struct kcm_sock *kcm)
 568 {
 569         struct kcm_mux *mux = kcm->mux;
 570
 571         spin_lock_bh(&mux->lock);
 572         KCM_STATS_INCR(mux->stats.tx_retries);
 573         spin_unlock_bh(&mux->lock);
 574 }
 575
 576 /* Write any messages ready on the kcm socket.  Called with kcm sock lock
 577  * held.  Return bytes actually sent or error.
 578  */
 579 static int kcm_write_msgs(struct kcm_sock *kcm)
 580 {
 581         struct sock *sk = &kcm->sk;
 582         struct kcm_psock *psock;
 583         struct sk_buff *skb, *head;
 584         struct kcm_tx_msg *txm;
 585         unsigned short fragidx, frag_offset;
 586         unsigned int sent, total_sent = 0;
 587         int ret = 0;
 588
 589         kcm->tx_wait_more = false;
 590         psock = kcm->tx_psock;
 591         if (unlikely(psock && psock->tx_stopped)) {
 592                 /* A reserved psock was aborted asynchronously. Unreserve
 593                  * it and we'll retry the message.
 594                  */
 595                 unreserve_psock(kcm);
 596                 kcm_report_tx_retry(kcm);
 597                 if (skb_queue_empty(&sk->sk_write_queue))
 598                         return 0;
 599
 600                 kcm_tx_msg(skb_peek(&sk->sk_write_queue))->sent = 0;
 601
 602         } else if (skb_queue_empty(&sk->sk_write_queue)) {
 603                 return 0;
 604         }
 605
 606         head = skb_peek(&sk->sk_write_queue);
 607         txm = kcm_tx_msg(head);
 608
 609         if (txm->sent) {
 610                 /* Send of first skbuff in queue already in progress */
 611                 if (WARN_ON(!psock)) {
 612                         ret = -EINVAL;
 613                         goto out;
 614                 }
 615                 sent = txm->sent;
 616                 frag_offset = txm->frag_offset;
 617                 fragidx = txm->fragidx;
 618                 skb = txm->frag_skb;
 619
 620                 goto do_frag;
 621         }
 622
 623 try_again:
 624         psock = reserve_psock(kcm);
 625         if (!psock)
 626                 goto out;
 627
 628         do {
 629                 skb = head;
 630                 txm = kcm_tx_msg(head);
 631                 sent = 0;
 632
 633 do_frag_list:
 634                 if (WARN_ON(!skb_shinfo(skb)->nr_frags)) {
 635                         ret = -EINVAL;
 636                         goto out;
 637                 }
 638
 639                 for (fragidx = 0; fragidx < skb_shinfo(skb)->nr_frags;
 640                      fragidx++) {
 641                         skb_frag_t *frag;
 642
 643                         frag_offset = 0;
 644 do_frag:
 645                         frag = &skb_shinfo(skb)->frags[fragidx];
 646                         if (WARN_ON(!frag->size)) {
 647                                 ret = -EINVAL;
 648                                 goto out;
 649                         }
 650
 651                         ret = kernel_sendpage(psock->sk->sk_socket,
 652                                               frag->page.p,
 653                                               frag->page_offset + frag_offset,
 654                                               frag->size - frag_offset,
 655                                               MSG_DONTWAIT);
 656                         if (ret <= 0) {
 657                                 if (ret == -EAGAIN) {
 658                                         /* Save state to try again when there's
 659                                          * write space on the socket
 660                                          */
 661                                         txm->sent = sent;
 662                                         txm->frag_offset = frag_offset;
 663                                         txm->fragidx = fragidx;
 664                                         txm->frag_skb = skb;
 665
 666                                         ret = 0;
 667                                         goto out;
 668                                 }
 669
 670                                 /* Hard failure in sending message, abort this
 671                                  * psock since it has lost framing
 672                                  * synchonization and retry sending the
 673                                  * message from the beginning.
 674                                  */
 675                                 kcm_abort_tx_psock(psock, ret ? -ret : EPIPE,
 676                                                    true);
 677                                 unreserve_psock(kcm);
 678
 679                                 txm->sent = 0;
 680                                 kcm_report_tx_retry(kcm);
 681                                 ret = 0;
 682
 683                                 goto try_again;
 684                         }
 685
 686                         sent += ret;
 687                         frag_offset += ret;
 688                         KCM_STATS_ADD(psock->stats.tx_bytes, ret);
 689                         if (frag_offset < frag->size) {
 690                                 /* Not finished with this frag */
 691                                 goto do_frag;
 692                         }
 693                 }
 694
 695                 if (skb == head) {
 696                         if (skb_has_frag_list(skb)) {
 697                                 skb = skb_shinfo(skb)->frag_list;
 698                                 goto do_frag_list;
 699                         }
 700                 } else if (skb->next) {
 701                         skb = skb->next;
 702                         goto do_frag_list;
 703                 }
 704
 705                 /* Successfully sent the whole packet, account for it. */
 706                 skb_dequeue(&sk->sk_write_queue);
 707                 kfree_skb(head);
 708                 sk->sk_wmem_queued -= sent;
 709                 total_sent += sent;
 710                 KCM_STATS_INCR(psock->stats.tx_msgs);
 711         } while ((head = skb_peek(&sk->sk_write_queue)));
 712 out:
 713         if (!head) {
 714                 /* Done with all queued messages. */
 715                 WARN_ON(!skb_queue_empty(&sk->sk_write_queue));
 716                 unreserve_psock(kcm);
 717         }
 718
 719         /* Check if write space is available */
 720         sk->sk_write_space(sk);
 721
 722         return total_sent ? : ret;
 723 }
 724
 725 static void kcm_tx_work(struct work_struct *w)
 726 {
 727         struct kcm_sock *kcm = container_of(w, struct kcm_sock, tx_work);
 728         struct sock *sk = &kcm->sk;
 729         int err;
 730
 731         lock_sock(sk);
 732
 733         /* Primarily for SOCK_DGRAM sockets, also handle asynchronous tx
 734          * aborts
 735          */
 736         err = kcm_write_msgs(kcm);
 737         if (err < 0) {
 738                 /* Hard failure in write, report error on KCM socket */
 739                 pr_warn("KCM: Hard failure on kcm_write_msgs %d\n", err);
 740                 report_csk_error(&kcm->sk, -err);
 741                 goto out;
 742         }
 743
 744         /* Primarily for SOCK_SEQPACKET sockets */
 745         if (likely(sk->sk_socket) &&
 746             test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
 747                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 748                 sk->sk_write_space(sk);
 749         }
 750
 751 out:
 752         release_sock(sk);
 753 }
 754
 755 static void kcm_push(struct kcm_sock *kcm)
 756 {
 757         if (kcm->tx_wait_more)
 758                 kcm_write_msgs(kcm);
 759 }
 760
 761 static ssize_t kcm_sendpage(struct socket *sock, struct page *page,
 762                             int offset, size_t size, int flags)
 763
 764 {
 765         struct sock *sk = sock->sk;
 766         struct kcm_sock *kcm = kcm_sk(sk);
 767         struct sk_buff *skb = NULL, *head = NULL;
 768         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 769         bool eor;
 770         int err = 0;
 771         int i;
 772
 773         if (flags & MSG_SENDPAGE_NOTLAST)
 774                 flags |= MSG_MORE;
 775
 776         /* No MSG_EOR from splice, only look at MSG_MORE */
 777         eor = !(flags & MSG_MORE);
 778
 779         lock_sock(sk);
 780
 781         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 782
 783         err = -EPIPE;
 784         if (sk->sk_err)
 785                 goto out_error;
 786
 787         if (kcm->seq_skb) {
 788                 /* Previously opened message */
 789                 head = kcm->seq_skb;
 790                 skb = kcm_tx_msg(head)->last_skb;
 791                 i = skb_shinfo(skb)->nr_frags;
 792
 793                 if (skb_can_coalesce(skb, i, page, offset)) {
 794                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], size);
 795                         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
 796                         goto coalesced;
 797                 }
 798
 799                 if (i >= MAX_SKB_FRAGS) {
 800                         struct sk_buff *tskb;
 801
 802                         tskb = alloc_skb(0, sk->sk_allocation);
 803                         while (!tskb) {
 804                                 kcm_push(kcm);
 805                                 err = sk_stream_wait_memory(sk, &timeo);
 806                                 if (err)
 807                                         goto out_error;
 808                         }
 809
 810                         if (head == skb)
 811                                 skb_shinfo(head)->frag_list = tskb;
 812                         else
 813                                 skb->next = tskb;
 814
 815                         skb = tskb;
 816                         skb->ip_summed = CHECKSUM_UNNECESSARY;
 817                         i = 0;
 818                 }
 819         } else {
 820                 /* Call the sk_stream functions to manage the sndbuf mem. */
 821                 if (!sk_stream_memory_free(sk)) {
 822                         kcm_push(kcm);
 823                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 824                         err = sk_stream_wait_memory(sk, &timeo);
 825                         if (err)
 826                                 goto out_error;
 827                 }
 828
 829                 head = alloc_skb(0, sk->sk_allocation);
 830                 while (!head) {
 831                         kcm_push(kcm);
 832                         err = sk_stream_wait_memory(sk, &timeo);
 833                         if (err)
 834                                 goto out_error;
 835                 }
 836
 837                 skb = head;
 838                 i = 0;
 839         }
 840
 841         get_page(page);
 842         skb_fill_page_desc(skb, i, page, offset, size);
 843         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
 844
 845 coalesced:
 846         skb->len += size;
 847         skb->data_len += size;
 848         skb->truesize += size;
 849         sk->sk_wmem_queued += size;
 850         sk_mem_charge(sk, size);
 851
 852         if (head != skb) {
 853                 head->len += size;
 854                 head->data_len += size;
 855                 head->truesize += size;
 856         }
 857
 858         if (eor) {
 859                 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
 860
 861                 /* Message complete, queue it on send buffer */
 862                 __skb_queue_tail(&sk->sk_write_queue, head);
 863                 kcm->seq_skb = NULL;
 864                 KCM_STATS_INCR(kcm->stats.tx_msgs);
 865
 866                 if (flags & MSG_BATCH) {
 867                         kcm->tx_wait_more = true;
 868                 } else if (kcm->tx_wait_more || not_busy) {
 869                         err = kcm_write_msgs(kcm);
 870                         if (err < 0) {
 871                                 /* We got a hard error in write_msgs but have
 872                                  * already queued this message. Report an error
 873                                  * in the socket, but don't affect return value
 874                                  * from sendmsg
 875                                  */
 876                                 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
 877                                 report_csk_error(&kcm->sk, -err);
 878                         }
 879                 }
 880         } else {
 881                 /* Message not complete, save state */
 882                 kcm->seq_skb = head;
 883                 kcm_tx_msg(head)->last_skb = skb;
 884         }
 885
 886         KCM_STATS_ADD(kcm->stats.tx_bytes, size);
 887
 888         release_sock(sk);
 889         return size;
 890
 891 out_error:
 892         kcm_push(kcm);
 893
 894         err = sk_stream_error(sk, flags, err);
 895
 896         /* make sure we wake any epoll edge trigger waiter */
 897         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
 898                 sk->sk_write_space(sk);
 899
 900         release_sock(sk);
 901         return err;
 902 }
 903
 904 static int kcm_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
 905 {
 906         struct sock *sk = sock->sk;
 907         struct kcm_sock *kcm = kcm_sk(sk);
 908         struct sk_buff *skb = NULL, *head = NULL;
 909         size_t copy, copied = 0;
 910         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 911         int eor = (sock->type == SOCK_DGRAM) ?
 912                   !(msg->msg_flags & MSG_MORE) : !!(msg->msg_flags & MSG_EOR);
 913         int err = -EPIPE;
 914
 915         lock_sock(sk);
 916
 917         /* Per tcp_sendmsg this should be in poll */
 918         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 919
 920         if (sk->sk_err)
 921                 goto out_error;
 922
 923         if (kcm->seq_skb) {
 924                 /* Previously opened message */
 925                 head = kcm->seq_skb;
 926                 skb = kcm_tx_msg(head)->last_skb;
 927                 goto start;
 928         }
 929
 930         /* Call the sk_stream functions to manage the sndbuf mem. */
 931         if (!sk_stream_memory_free(sk)) {
 932                 kcm_push(kcm);
 933                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 934                 err = sk_stream_wait_memory(sk, &timeo);
 935                 if (err)
 936                         goto out_error;
 937         }
 938
 939         if (msg_data_left(msg)) {
 940                 /* New message, alloc head skb */
 941                 head = alloc_skb(0, sk->sk_allocation);
 942                 while (!head) {
 943                         kcm_push(kcm);
 944                         err = sk_stream_wait_memory(sk, &timeo);
 945                         if (err)
 946                                 goto out_error;
 947
 948                         head = alloc_skb(0, sk->sk_allocation);
 949                 }
 950
 951                 skb = head;
 952
 953                 /* Set ip_summed to CHECKSUM_UNNECESSARY to avoid calling
 954                  * csum_and_copy_from_iter from skb_do_copy_data_nocache.
 955                  */
 956                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 957         }
 958
 959 start:
 960         while (msg_data_left(msg)) {
 961                 bool merge = true;
 962                 int i = skb_shinfo(skb)->nr_frags;
 963                 struct page_frag *pfrag = sk_page_frag(sk);
 964
 965                 if (!sk_page_frag_refill(sk, pfrag))
 966                         goto wait_for_memory;
 967
 968                 if (!skb_can_coalesce(skb, i, pfrag->page,
 969                                       pfrag->offset)) {
 970                         if (i == MAX_SKB_FRAGS) {
 971                                 struct sk_buff *tskb;
 972
 973                                 tskb = alloc_skb(0, sk->sk_allocation);
 974                                 if (!tskb)
 975                                         goto wait_for_memory;
 976
 977                                 if (head == skb)
 978                                         skb_shinfo(head)->frag_list = tskb;
 979                                 else
 980                                         skb->next = tskb;
 981
 982                                 skb = tskb;
 983                                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 984                                 continue;
 985                         }
 986                         merge = false;
 987                 }
 988
 989                 copy = min_t(int, msg_data_left(msg),
 990                              pfrag->size - pfrag->offset);
 991
 992                 if (!sk_wmem_schedule(sk, copy))
 993                         goto wait_for_memory;
 994
 995                 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
 996                                                pfrag->page,
 997                                                pfrag->offset,
 998                                                copy);
 999                 if (err)
1000                         goto out_error;
1001
1002                 /* Update the skb. */
1003                 if (merge) {
1004                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1005                 } else {
1006                         skb_fill_page_desc(skb, i, pfrag->page,
1007                                            pfrag->offset, copy);
1008                         get_page(pfrag->page);
1009                 }
1010
1011                 pfrag->offset += copy;
1012                 copied += copy;
1013                 if (head != skb) {
1014                         head->len += copy;
1015                         head->data_len += copy;
1016                 }
1017
1018                 continue;
1019
1020 wait_for_memory:
1021                 kcm_push(kcm);
1022                 err = sk_stream_wait_memory(sk, &timeo);
1023                 if (err)
1024                         goto out_error;
1025         }
1026
1027         if (eor) {
1028                 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
1029
1030                 if (head) {
1031                         /* Message complete, queue it on send buffer */
1032                         __skb_queue_tail(&sk->sk_write_queue, head);
1033                         kcm->seq_skb = NULL;
1034                         KCM_STATS_INCR(kcm->stats.tx_msgs);
1035                 }
1036
1037                 if (msg->msg_flags & MSG_BATCH) {
1038                         kcm->tx_wait_more = true;
1039                 } else if (kcm->tx_wait_more || not_busy) {
1040                         err = kcm_write_msgs(kcm);
1041                         if (err < 0) {
1042                                 /* We got a hard error in write_msgs but have
1043                                  * already queued this message. Report an error
1044                                  * in the socket, but don't affect return value
1045                                  * from sendmsg
1046                                  */
1047                                 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
1048                                 report_csk_error(&kcm->sk, -err);
1049                         }
1050                 }
1051         } else {
1052                 /* Message not complete, save state */
1053 partial_message:
1054                 if (head) {
1055                         kcm->seq_skb = head;
1056                         kcm_tx_msg(head)->last_skb = skb;
1057                 }
1058         }
1059
1060         KCM_STATS_ADD(kcm->stats.tx_bytes, copied);
1061
1062         release_sock(sk);
1063         return copied;
1064
1065 out_error:
1066         kcm_push(kcm);
1067
1068         if (sock->type == SOCK_SEQPACKET) {
1069                 /* Wrote some bytes before encountering an
1070                  * error, return partial success.
1071                  */
1072                 if (copied)
1073                         goto partial_message;
1074                 if (head != kcm->seq_skb)
1075                         kfree_skb(head);
1076         } else {
1077                 kfree_skb(head);
1078                 kcm->seq_skb = NULL;
1079         }
1080
1081         err = sk_stream_error(sk, msg->msg_flags, err);
1082
1083         /* make sure we wake any epoll edge trigger waiter */
1084         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1085                 sk->sk_write_space(sk);
1086
1087         release_sock(sk);
1088         return err;
1089 }
1090
1091 static int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
1092                        size_t len, int flags)
1093 {
1094         int noblock = flags & MSG_DONTWAIT;
1095         struct sock *sk = sock->sk;
1096         struct kcm_sock *kcm = kcm_sk(sk);
1097         int err = 0;
1098         struct strp_msg *stm;
1099         int copied = 0;
1100         struct sk_buff *skb;
1101
1102         skb = skb_recv_datagram(sk, flags, noblock, &err);
1103         if (!skb)
1104                 goto out;
1105
1106         /* Okay, have a message on the receive queue */
1107
1108         stm = strp_msg(skb);
1109
1110         if (len > stm->full_len)
1111                 len = stm->full_len;
1112
1113         err = skb_copy_datagram_msg(skb, stm->offset, msg, len);
1114         if (err < 0)
1115                 goto out;
1116
1117         copied = len;
1118         if (likely(!(flags & MSG_PEEK))) {
1119                 KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1120                 if (copied < stm->full_len) {
1121                         if (sock->type == SOCK_DGRAM) {
1122                                 /* Truncated message */
1123                                 msg->msg_flags |= MSG_TRUNC;
1124                                 goto msg_finished;
1125                         }
1126                         stm->offset += copied;
1127                         stm->full_len -= copied;
1128                 } else {
1129 msg_finished:
1130                         /* Finished with message */
1131                         msg->msg_flags |= MSG_EOR;
1132                         KCM_STATS_INCR(kcm->stats.rx_msgs);
1133                 }
1134         }
1135
1136 out:
1137         skb_free_datagram(sk, skb);
1138         return copied ? : err;
1139 }
1140
1141 static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1142                                struct pipe_inode_info *pipe, size_t len,
1143                                unsigned int flags)
1144 {
1145         int noblock = flags & MSG_DONTWAIT;
1146         struct sock *sk = sock->sk;
1147         struct kcm_sock *kcm = kcm_sk(sk);
1148         struct strp_msg *stm;
1149         int err = 0;
1150         ssize_t copied;
1151         struct sk_buff *skb;
1152
1153         /* Only support splice for SOCKSEQPACKET */
1154
1155         skb = skb_recv_datagram(sk, flags, noblock, &err);
1156         if (!skb)
1157                 goto err_out;
1158
1159         /* Okay, have a message on the receive queue */
1160
1161         stm = strp_msg(skb);
1162
1163         if (len > stm->full_len)
1164                 len = stm->full_len;
1165
1166         copied = skb_splice_bits(skb, sk, stm->offset, pipe, len, flags);
1167         if (copied < 0) {
1168                 err = copied;
1169                 goto err_out;
1170         }
1171
1172         KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1173
1174         stm->offset += copied;
1175         stm->full_len -= copied;
1176
1177         /* We have no way to return MSG_EOR. If all the bytes have been
1178          * read we still leave the message in the receive socket buffer.
1179          * A subsequent recvmsg needs to be done to return MSG_EOR and
1180          * finish reading the message.
1181          */
1182
1183         skb_free_datagram(sk, skb);
1184         return copied;
1185
1186 err_out:
1187         skb_free_datagram(sk, skb);
1188         return err;
1189 }
1190
1191 /* kcm sock lock held */
1192 static void kcm_recv_disable(struct kcm_sock *kcm)
1193 {
1194         struct kcm_mux *mux = kcm->mux;
1195
1196         if (kcm->rx_disabled)
1197                 return;
1198
1199         spin_lock_bh(&mux->rx_lock);
1200
1201         kcm->rx_disabled = 1;
1202
1203         /* If a psock is reserved we'll do cleanup in unreserve */
1204         if (!kcm->rx_psock) {
1205                 if (kcm->rx_wait) {
1206                         list_del(&kcm->wait_rx_list);
1207                         /* paired with lockless reads in kcm_rfree() */
1208                         WRITE_ONCE(kcm->rx_wait, false);
1209                 }
1210
1211                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1212         }
1213
1214         spin_unlock_bh(&mux->rx_lock);
1215 }
1216
1217 /* kcm sock lock held */
1218 static void kcm_recv_enable(struct kcm_sock *kcm)
1219 {
1220         struct kcm_mux *mux = kcm->mux;
1221
1222         if (!kcm->rx_disabled)
1223                 return;
1224
1225         spin_lock_bh(&mux->rx_lock);
1226
1227         kcm->rx_disabled = 0;
1228         kcm_rcv_ready(kcm);
1229
1230         spin_unlock_bh(&mux->rx_lock);
1231 }
1232
1233 static int kcm_setsockopt(struct socket *sock, int level, int optname,
1234                           char __user *optval, unsigned int optlen)
1235 {
1236         struct kcm_sock *kcm = kcm_sk(sock->sk);
1237         int val, valbool;
1238         int err = 0;
1239
1240         if (level != SOL_KCM)
1241                 return -ENOPROTOOPT;
1242
1243         if (optlen < sizeof(int))
1244                 return -EINVAL;
1245
1246         if (get_user(val, (int __user *)optval))
1247                 return -EINVAL;
1248
1249         valbool = val ? 1 : 0;
1250
1251         switch (optname) {
1252         case KCM_RECV_DISABLE:
1253                 lock_sock(&kcm->sk);
1254                 if (valbool)
1255                         kcm_recv_disable(kcm);
1256                 else
1257                         kcm_recv_enable(kcm);
1258                 release_sock(&kcm->sk);
1259                 break;
1260         default:
1261                 err = -ENOPROTOOPT;
1262         }
1263
1264         return err;
1265 }
1266
1267 static int kcm_getsockopt(struct socket *sock, int level, int optname,
1268                           char __user *optval, int __user *optlen)
1269 {
1270         struct kcm_sock *kcm = kcm_sk(sock->sk);
1271         int val, len;
1272
1273         if (level != SOL_KCM)
1274                 return -ENOPROTOOPT;
1275
1276         if (get_user(len, optlen))
1277                 return -EFAULT;
1278
1279         len = min_t(unsigned int, len, sizeof(int));
1280         if (len < 0)
1281                 return -EINVAL;
1282
1283         switch (optname) {
1284         case KCM_RECV_DISABLE:
1285                 val = kcm->rx_disabled;
1286                 break;
1287         default:
1288                 return -ENOPROTOOPT;
1289         }
1290
1291         if (put_user(len, optlen))
1292                 return -EFAULT;
1293         if (copy_to_user(optval, &val, len))
1294                 return -EFAULT;
1295         return 0;
1296 }
1297
1298 static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1299 {
1300         struct kcm_sock *tkcm;
1301         struct list_head *head;
1302         int index = 0;
1303
1304         /* For SOCK_SEQPACKET sock type, datagram_poll checks the sk_state, so
1305          * we set sk_state, otherwise epoll_wait always returns right away with
1306          * EPOLLHUP
1307          */
1308         kcm->sk.sk_state = TCP_ESTABLISHED;
1309
1310         /* Add to mux's kcm sockets list */
1311         kcm->mux = mux;
1312         spin_lock_bh(&mux->lock);
1313
1314         head = &mux->kcm_socks;
1315         list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1316                 if (tkcm->index != index)
1317                         break;
1318                 head = &tkcm->kcm_sock_list;
1319                 index++;
1320         }
1321
1322         list_add(&kcm->kcm_sock_list, head);
1323         kcm->index = index;
1324
1325         mux->kcm_socks_cnt++;
1326         spin_unlock_bh(&mux->lock);
1327
1328         INIT_WORK(&kcm->tx_work, kcm_tx_work);
1329
1330         spin_lock_bh(&mux->rx_lock);
1331         kcm_rcv_ready(kcm);
1332         spin_unlock_bh(&mux->rx_lock);
1333 }
1334
1335 static int kcm_attach(struct socket *sock, struct socket *csock,
1336                       struct bpf_prog *prog)
1337 {
1338         struct kcm_sock *kcm = kcm_sk(sock->sk);
1339         struct kcm_mux *mux = kcm->mux;
1340         struct sock *csk;
1341         struct kcm_psock *psock = NULL, *tpsock;
1342         struct list_head *head;
1343         int index = 0;
1344         static const struct strp_callbacks cb = {
1345                 .rcv_msg = kcm_rcv_strparser,
1346                 .parse_msg = kcm_parse_func_strparser,
1347                 .read_sock_done = kcm_read_sock_done,
1348         };
1349         int err = 0;
1350
1351         csk = csock->sk;
1352         if (!csk)
1353                 return -EINVAL;
1354
1355         lock_sock(csk);
1356
1357         /* Only allow TCP sockets to be attached for now */
1358         if ((csk->sk_family != AF_INET && csk->sk_family != AF_INET6) ||
1359             csk->sk_protocol != IPPROTO_TCP) {
1360                 err = -EOPNOTSUPP;
1361                 goto out;
1362         }
1363
1364         /* Don't allow listeners or closed sockets */
1365         if (csk->sk_state == TCP_LISTEN || csk->sk_state == TCP_CLOSE) {
1366                 err = -EOPNOTSUPP;
1367                 goto out;
1368         }
1369
1370         psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1371         if (!psock) {
1372                 err = -ENOMEM;
1373                 goto out;
1374         }
1375
1376         psock->mux = mux;
1377         psock->sk = csk;
1378         psock->bpf_prog = prog;
1379
1380         write_lock_bh(&csk->sk_callback_lock);
1381
1382         /* Check if sk_user_data is aready by KCM or someone else.
1383          * Must be done under lock to prevent race conditions.
1384          */
1385         if (csk->sk_user_data) {
1386                 write_unlock_bh(&csk->sk_callback_lock);
1387                 kmem_cache_free(kcm_psockp, psock);
1388                 err = -EALREADY;
1389                 goto out;
1390         }
1391
1392         err = strp_init(&psock->strp, csk, &cb);
1393         if (err) {
1394                 write_unlock_bh(&csk->sk_callback_lock);
1395                 kmem_cache_free(kcm_psockp, psock);
1396                 goto out;
1397         }
1398
1399         psock->save_data_ready = csk->sk_data_ready;
1400         psock->save_write_space = csk->sk_write_space;
1401         psock->save_state_change = csk->sk_state_change;
1402         csk->sk_user_data = psock;
1403         csk->sk_data_ready = psock_data_ready;
1404         csk->sk_write_space = psock_write_space;
1405         csk->sk_state_change = psock_state_change;
1406
1407         write_unlock_bh(&csk->sk_callback_lock);
1408
1409         sock_hold(csk);
1410
1411         /* Finished initialization, now add the psock to the MUX. */
1412         spin_lock_bh(&mux->lock);
1413         head = &mux->psocks;
1414         list_for_each_entry(tpsock, &mux->psocks, psock_list) {
1415                 if (tpsock->index != index)
1416                         break;
1417                 head = &tpsock->psock_list;
1418                 index++;
1419         }
1420
1421         list_add(&psock->psock_list, head);
1422         psock->index = index;
1423
1424         KCM_STATS_INCR(mux->stats.psock_attach);
1425         mux->psocks_cnt++;
1426         psock_now_avail(psock);
1427         spin_unlock_bh(&mux->lock);
1428
1429         /* Schedule RX work in case there are already bytes queued */
1430         strp_check_rcv(&psock->strp);
1431
1432 out:
1433         release_sock(csk);
1434
1435         return err;
1436 }
1437
1438 static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1439 {
1440         struct socket *csock;
1441         struct bpf_prog *prog;
1442         int err;
1443
1444         csock = sockfd_lookup(info->fd, &err);
1445         if (!csock)
1446                 return -ENOENT;
1447
1448         prog = bpf_prog_get_type(info->bpf_fd, BPF_PROG_TYPE_SOCKET_FILTER);
1449         if (IS_ERR(prog)) {
1450                 err = PTR_ERR(prog);
1451                 goto out;
1452         }
1453
1454         err = kcm_attach(sock, csock, prog);
1455         if (err) {
1456                 bpf_prog_put(prog);
1457                 goto out;
1458         }
1459
1460         /* Keep reference on file also */
1461
1462         return 0;
1463 out:
1464         fput(csock->file);
1465         return err;
1466 }
1467
1468 static void kcm_unattach(struct kcm_psock *psock)
1469 {
1470         struct sock *csk = psock->sk;
1471         struct kcm_mux *mux = psock->mux;
1472
1473         lock_sock(csk);
1474
1475         /* Stop getting callbacks from TCP socket. After this there should
1476          * be no way to reserve a kcm for this psock.
1477          */
1478         write_lock_bh(&csk->sk_callback_lock);
1479         csk->sk_user_data = NULL;
1480         csk->sk_data_ready = psock->save_data_ready;
1481         csk->sk_write_space = psock->save_write_space;
1482         csk->sk_state_change = psock->save_state_change;
1483         strp_stop(&psock->strp);
1484
1485         if (WARN_ON(psock->rx_kcm)) {
1486                 write_unlock_bh(&csk->sk_callback_lock);
1487                 release_sock(csk);
1488                 return;
1489         }
1490
1491         spin_lock_bh(&mux->rx_lock);
1492
1493         /* Stop receiver activities. After this point psock should not be
1494          * able to get onto ready list either through callbacks or work.
1495          */
1496         if (psock->ready_rx_msg) {
1497                 list_del(&psock->psock_ready_list);
1498                 kfree_skb(psock->ready_rx_msg);
1499                 psock->ready_rx_msg = NULL;
1500                 KCM_STATS_INCR(mux->stats.rx_ready_drops);
1501         }
1502
1503         spin_unlock_bh(&mux->rx_lock);
1504
1505         write_unlock_bh(&csk->sk_callback_lock);
1506
1507         /* Call strp_done without sock lock */
1508         release_sock(csk);
1509         strp_done(&psock->strp);
1510         lock_sock(csk);
1511
1512         bpf_prog_put(psock->bpf_prog);
1513
1514         spin_lock_bh(&mux->lock);
1515
1516         aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1517         save_strp_stats(&psock->strp, &mux->aggregate_strp_stats);
1518
1519         KCM_STATS_INCR(mux->stats.psock_unattach);
1520
1521         if (psock->tx_kcm) {
1522                 /* psock was reserved.  Just mark it finished and we will clean
1523                  * up in the kcm paths, we need kcm lock which can not be
1524                  * acquired here.
1525                  */
1526                 KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1527                 spin_unlock_bh(&mux->lock);
1528
1529                 /* We are unattaching a socket that is reserved. Abort the
1530                  * socket since we may be out of sync in sending on it. We need
1531                  * to do this without the mux lock.
1532                  */
1533                 kcm_abort_tx_psock(psock, EPIPE, false);
1534
1535                 spin_lock_bh(&mux->lock);
1536                 if (!psock->tx_kcm) {
1537                         /* psock now unreserved in window mux was unlocked */
1538                         goto no_reserved;
1539                 }
1540                 psock->done = 1;
1541
1542                 /* Commit done before queuing work to process it */
1543                 smp_mb();
1544
1545                 /* Queue tx work to make sure psock->done is handled */
1546                 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
1547                 spin_unlock_bh(&mux->lock);
1548         } else {
1549 no_reserved:
1550                 if (!psock->tx_stopped)
1551                         list_del(&psock->psock_avail_list);
1552                 list_del(&psock->psock_list);
1553                 mux->psocks_cnt--;
1554                 spin_unlock_bh(&mux->lock);
1555
1556                 sock_put(csk);
1557                 fput(csk->sk_socket->file);
1558                 kmem_cache_free(kcm_psockp, psock);
1559         }
1560
1561         release_sock(csk);
1562 }
1563
1564 static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1565 {
1566         struct kcm_sock *kcm = kcm_sk(sock->sk);
1567         struct kcm_mux *mux = kcm->mux;
1568         struct kcm_psock *psock;
1569         struct socket *csock;
1570         struct sock *csk;
1571         int err;
1572
1573         csock = sockfd_lookup(info->fd, &err);
1574         if (!csock)
1575                 return -ENOENT;
1576
1577         csk = csock->sk;
1578         if (!csk) {
1579                 err = -EINVAL;
1580                 goto out;
1581         }
1582
1583         err = -ENOENT;
1584
1585         spin_lock_bh(&mux->lock);
1586
1587         list_for_each_entry(psock, &mux->psocks, psock_list) {
1588                 if (psock->sk != csk)
1589                         continue;
1590
1591                 /* Found the matching psock */
1592
1593                 if (psock->unattaching || WARN_ON(psock->done)) {
1594                         err = -EALREADY;
1595                         break;
1596                 }
1597
1598                 psock->unattaching = 1;
1599
1600                 spin_unlock_bh(&mux->lock);
1601
1602                 /* Lower socket lock should already be held */
1603                 kcm_unattach(psock);
1604
1605                 err = 0;
1606                 goto out;
1607         }
1608
1609         spin_unlock_bh(&mux->lock);
1610
1611 out:
1612         fput(csock->file);
1613         return err;
1614 }
1615
1616 static struct proto kcm_proto = {
1617         .name   = "KCM",
1618         .owner  = THIS_MODULE,
1619         .obj_size = sizeof(struct kcm_sock),
1620 };
1621
1622 /* Clone a kcm socket. */
1623 static struct file *kcm_clone(struct socket *osock)
1624 {
1625         struct socket *newsock;
1626         struct sock *newsk;
1627
1628         newsock = sock_alloc();
1629         if (!newsock)
1630                 return ERR_PTR(-ENFILE);
1631
1632         newsock->type = osock->type;
1633         newsock->ops = osock->ops;
1634
1635         __module_get(newsock->ops->owner);
1636
1637         newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1638                          &kcm_proto, false);
1639         if (!newsk) {
1640                 sock_release(newsock);
1641                 return ERR_PTR(-ENOMEM);
1642         }
1643         sock_init_data(newsock, newsk);
1644         init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1645
1646         return sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1647 }
1648
1649 static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
1650 {
1651         int err;
1652
1653         switch (cmd) {
1654         case SIOCKCMATTACH: {
1655                 struct kcm_attach info;
1656
1657                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1658                         return -EFAULT;
1659
1660                 err = kcm_attach_ioctl(sock, &info);
1661
1662                 break;
1663         }
1664         case SIOCKCMUNATTACH: {
1665                 struct kcm_unattach info;
1666
1667                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1668                         return -EFAULT;
1669
1670                 err = kcm_unattach_ioctl(sock, &info);
1671
1672                 break;
1673         }
1674         case SIOCKCMCLONE: {
1675                 struct kcm_clone info;
1676                 struct file *file;
1677
1678                 info.fd = get_unused_fd_flags(0);
1679                 if (unlikely(info.fd < 0))
1680                         return info.fd;
1681
1682                 file = kcm_clone(sock);
1683                 if (IS_ERR(file)) {
1684                         put_unused_fd(info.fd);
1685                         return PTR_ERR(file);
1686                 }
1687                 if (copy_to_user((void __user *)arg, &info,
1688                                  sizeof(info))) {
1689                         put_unused_fd(info.fd);
1690                         fput(file);
1691                         return -EFAULT;
1692                 }
1693                 fd_install(info.fd, file);
1694                 err = 0;
1695                 break;
1696         }
1697         default:
1698                 err = -ENOIOCTLCMD;
1699                 break;
1700         }
1701
1702         return err;
1703 }
1704
1705 static void free_mux(struct rcu_head *rcu)
1706 {
1707         struct kcm_mux *mux = container_of(rcu,
1708             struct kcm_mux, rcu);
1709
1710         kmem_cache_free(kcm_muxp, mux);
1711 }
1712
1713 static void release_mux(struct kcm_mux *mux)
1714 {
1715         struct kcm_net *knet = mux->knet;
1716         struct kcm_psock *psock, *tmp_psock;
1717
1718         /* Release psocks */
1719         list_for_each_entry_safe(psock, tmp_psock,
1720                                  &mux->psocks, psock_list) {
1721                 if (!WARN_ON(psock->unattaching))
1722                         kcm_unattach(psock);
1723         }
1724
1725         if (WARN_ON(mux->psocks_cnt))
1726                 return;
1727
1728         __skb_queue_purge(&mux->rx_hold_queue);
1729
1730         mutex_lock(&knet->mutex);
1731         aggregate_mux_stats(&mux->stats, &knet->aggregate_mux_stats);
1732         aggregate_psock_stats(&mux->aggregate_psock_stats,
1733                               &knet->aggregate_psock_stats);
1734         aggregate_strp_stats(&mux->aggregate_strp_stats,
1735                              &knet->aggregate_strp_stats);
1736         list_del_rcu(&mux->kcm_mux_list);
1737         knet->count--;
1738         mutex_unlock(&knet->mutex);
1739
1740         call_rcu(&mux->rcu, free_mux);
1741 }
1742
1743 static void kcm_done(struct kcm_sock *kcm)
1744 {
1745         struct kcm_mux *mux = kcm->mux;
1746         struct sock *sk = &kcm->sk;
1747         int socks_cnt;
1748
1749         spin_lock_bh(&mux->rx_lock);
1750         if (kcm->rx_psock) {
1751                 /* Cleanup in unreserve_rx_kcm */
1752                 WARN_ON(kcm->done);
1753                 kcm->rx_disabled = 1;
1754                 kcm->done = 1;
1755                 spin_unlock_bh(&mux->rx_lock);
1756                 return;
1757         }
1758
1759         if (kcm->rx_wait) {
1760                 list_del(&kcm->wait_rx_list);
1761                 /* paired with lockless reads in kcm_rfree() */
1762                 WRITE_ONCE(kcm->rx_wait, false);
1763         }
1764         /* Move any pending receive messages to other kcm sockets */
1765         requeue_rx_msgs(mux, &sk->sk_receive_queue);
1766
1767         spin_unlock_bh(&mux->rx_lock);
1768
1769         if (WARN_ON(sk_rmem_alloc_get(sk)))
1770                 return;
1771
1772         /* Detach from MUX */
1773         spin_lock_bh(&mux->lock);
1774
1775         list_del(&kcm->kcm_sock_list);
1776         mux->kcm_socks_cnt--;
1777         socks_cnt = mux->kcm_socks_cnt;
1778
1779         spin_unlock_bh(&mux->lock);
1780
1781         if (!socks_cnt) {
1782                 /* We are done with the mux now. */
1783                 release_mux(mux);
1784         }
1785
1786         WARN_ON(kcm->rx_wait);
1787
1788         sock_put(&kcm->sk);
1789 }
1790
1791 /* Called by kcm_release to close a KCM socket.
1792  * If this is the last KCM socket on the MUX, destroy the MUX.
1793  */
1794 static int kcm_release(struct socket *sock)
1795 {
1796         struct sock *sk = sock->sk;
1797         struct kcm_sock *kcm;
1798         struct kcm_mux *mux;
1799         struct kcm_psock *psock;
1800
1801         if (!sk)
1802                 return 0;
1803
1804         kcm = kcm_sk(sk);
1805         mux = kcm->mux;
1806
1807         lock_sock(sk);
1808         sock_orphan(sk);
1809         kfree_skb(kcm->seq_skb);
1810
1811         /* Purge queue under lock to avoid race condition with tx_work trying
1812          * to act when queue is nonempty. If tx_work runs after this point
1813          * it will just return.
1814          */
1815         __skb_queue_purge(&sk->sk_write_queue);
1816
1817         /* Set tx_stopped. This is checked when psock is bound to a kcm and we
1818          * get a writespace callback. This prevents further work being queued
1819          * from the callback (unbinding the psock occurs after canceling work.
1820          */
1821         kcm->tx_stopped = 1;
1822
1823         release_sock(sk);
1824
1825         spin_lock_bh(&mux->lock);
1826         if (kcm->tx_wait) {
1827                 /* Take of tx_wait list, after this point there should be no way
1828                  * that a psock will be assigned to this kcm.
1829                  */
1830                 list_del(&kcm->wait_psock_list);
1831                 kcm->tx_wait = false;
1832         }
1833         spin_unlock_bh(&mux->lock);
1834
1835         /* Cancel work. After this point there should be no outside references
1836          * to the kcm socket.
1837          */
1838         cancel_work_sync(&kcm->tx_work);
1839
1840         lock_sock(sk);
1841         psock = kcm->tx_psock;
1842         if (psock) {
1843                 /* A psock was reserved, so we need to kill it since it
1844                  * may already have some bytes queued from a message. We
1845                  * need to do this after removing kcm from tx_wait list.
1846                  */
1847                 kcm_abort_tx_psock(psock, EPIPE, false);
1848                 unreserve_psock(kcm);
1849         }
1850         release_sock(sk);
1851
1852         WARN_ON(kcm->tx_wait);
1853         WARN_ON(kcm->tx_psock);
1854
1855         sock->sk = NULL;
1856
1857         kcm_done(kcm);
1858
1859         return 0;
1860 }
1861
1862 static const struct proto_ops kcm_dgram_ops = {
1863         .family =       PF_KCM,
1864         .owner =        THIS_MODULE,
1865         .release =      kcm_release,
1866         .bind =         sock_no_bind,
1867         .connect =      sock_no_connect,
1868         .socketpair =   sock_no_socketpair,
1869         .accept =       sock_no_accept,
1870         .getname =      sock_no_getname,
1871         .poll =         datagram_poll,
1872         .ioctl =        kcm_ioctl,
1873         .listen =       sock_no_listen,
1874         .shutdown =     sock_no_shutdown,
1875         .setsockopt =   kcm_setsockopt,
1876         .getsockopt =   kcm_getsockopt,
1877         .sendmsg =      kcm_sendmsg,
1878         .recvmsg =      kcm_recvmsg,
1879         .mmap =         sock_no_mmap,
1880         .sendpage =     kcm_sendpage,
1881 };
1882
1883 static const struct proto_ops kcm_seqpacket_ops = {
1884         .family =       PF_KCM,
1885         .owner =        THIS_MODULE,
1886         .release =      kcm_release,
1887         .bind =         sock_no_bind,
1888         .connect =      sock_no_connect,
1889         .socketpair =   sock_no_socketpair,
1890         .accept =       sock_no_accept,
1891         .getname =      sock_no_getname,
1892         .poll =         datagram_poll,
1893         .ioctl =        kcm_ioctl,
1894         .listen =       sock_no_listen,
1895         .shutdown =     sock_no_shutdown,
1896         .setsockopt =   kcm_setsockopt,
1897         .getsockopt =   kcm_getsockopt,
1898         .sendmsg =      kcm_sendmsg,
1899         .recvmsg =      kcm_recvmsg,
1900         .mmap =         sock_no_mmap,
1901         .sendpage =     kcm_sendpage,
1902         .splice_read =  kcm_splice_read,
1903 };
1904
1905 /* Create proto operation for kcm sockets */
1906 static int kcm_create(struct net *net, struct socket *sock,
1907                       int protocol, int kern)
1908 {
1909         struct kcm_net *knet = net_generic(net, kcm_net_id);
1910         struct sock *sk;
1911         struct kcm_mux *mux;
1912
1913         switch (sock->type) {
1914         case SOCK_DGRAM:
1915                 sock->ops = &kcm_dgram_ops;
1916                 break;
1917         case SOCK_SEQPACKET:
1918                 sock->ops = &kcm_seqpacket_ops;
1919                 break;
1920         default:
1921                 return -ESOCKTNOSUPPORT;
1922         }
1923
1924         if (protocol != KCMPROTO_CONNECTED)
1925                 return -EPROTONOSUPPORT;
1926
1927         sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
1928         if (!sk)
1929                 return -ENOMEM;
1930
1931         /* Allocate a kcm mux, shared between KCM sockets */
1932         mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
1933         if (!mux) {
1934                 sk_free(sk);
1935                 return -ENOMEM;
1936         }
1937
1938         spin_lock_init(&mux->lock);
1939         spin_lock_init(&mux->rx_lock);
1940         INIT_LIST_HEAD(&mux->kcm_socks);
1941         INIT_LIST_HEAD(&mux->kcm_rx_waiters);
1942         INIT_LIST_HEAD(&mux->kcm_tx_waiters);
1943
1944         INIT_LIST_HEAD(&mux->psocks);
1945         INIT_LIST_HEAD(&mux->psocks_ready);
1946         INIT_LIST_HEAD(&mux->psocks_avail);
1947
1948         mux->knet = knet;
1949
1950         /* Add new MUX to list */
1951         mutex_lock(&knet->mutex);
1952         list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
1953         knet->count++;
1954         mutex_unlock(&knet->mutex);
1955
1956         skb_queue_head_init(&mux->rx_hold_queue);
1957
1958         /* Init KCM socket */
1959         sock_init_data(sock, sk);
1960         init_kcm_sock(kcm_sk(sk), mux);
1961
1962         return 0;
1963 }
1964
1965 static const struct net_proto_family kcm_family_ops = {
1966         .family = PF_KCM,
1967         .create = kcm_create,
1968         .owner  = THIS_MODULE,
1969 };
1970
1971 static __net_init int kcm_init_net(struct net *net)
1972 {
1973         struct kcm_net *knet = net_generic(net, kcm_net_id);
1974
1975         INIT_LIST_HEAD_RCU(&knet->mux_list);
1976         mutex_init(&knet->mutex);
1977
1978         return 0;
1979 }
1980
1981 static __net_exit void kcm_exit_net(struct net *net)
1982 {
1983         struct kcm_net *knet = net_generic(net, kcm_net_id);
1984
1985         /* All KCM sockets should be closed at this point, which should mean
1986          * that all multiplexors and psocks have been destroyed.
1987          */
1988         WARN_ON(!list_empty(&knet->mux_list));
1989
1990         mutex_destroy(&knet->mutex);
1991 }
1992
1993 static struct pernet_operations kcm_net_ops = {
1994         .init = kcm_init_net,
1995         .exit = kcm_exit_net,
1996         .id   = &kcm_net_id,
1997         .size = sizeof(struct kcm_net),
1998 };
1999
2000 static int __init kcm_init(void)
2001 {
2002         int err = -ENOMEM;
2003
2004         kcm_muxp = kmem_cache_create("kcm_mux_cache",
2005                                      sizeof(struct kcm_mux), 0,
2006                                      SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2007         if (!kcm_muxp)
2008                 goto fail;
2009
2010         kcm_psockp = kmem_cache_create("kcm_psock_cache",
2011                                        sizeof(struct kcm_psock), 0,
2012                                         SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2013         if (!kcm_psockp)
2014                 goto fail;
2015
2016         kcm_wq = create_singlethread_workqueue("kkcmd");
2017         if (!kcm_wq)
2018                 goto fail;
2019
2020         err = proto_register(&kcm_proto, 1);
2021         if (err)
2022                 goto fail;
2023
2024         err = register_pernet_device(&kcm_net_ops);
2025         if (err)
2026                 goto net_ops_fail;
2027
2028         err = sock_register(&kcm_family_ops);
2029         if (err)
2030                 goto sock_register_fail;
2031
2032         err = kcm_proc_init();
2033         if (err)
2034                 goto proc_init_fail;
2035
2036         return 0;
2037
2038 proc_init_fail:
2039         sock_unregister(PF_KCM);
2040
2041 sock_register_fail:
2042         unregister_pernet_device(&kcm_net_ops);
2043
2044 net_ops_fail:
2045         proto_unregister(&kcm_proto);
2046
2047 fail:
2048         kmem_cache_destroy(kcm_muxp);
2049         kmem_cache_destroy(kcm_psockp);
2050
2051         if (kcm_wq)
2052                 destroy_workqueue(kcm_wq);
2053
2054         return err;
2055 }
2056
2057 static void __exit kcm_exit(void)
2058 {
2059         kcm_proc_exit();
2060         sock_unregister(PF_KCM);
2061         unregister_pernet_device(&kcm_net_ops);
2062         proto_unregister(&kcm_proto);
2063         destroy_workqueue(kcm_wq);
2064
2065         kmem_cache_destroy(kcm_muxp);
2066         kmem_cache_destroy(kcm_psockp);
2067 }
2068
2069 module_init(kcm_init);
2070 module_exit(kcm_exit);
2071
2072 MODULE_LICENSE("GPL");
2073 MODULE_ALIAS_NETPROTO(PF_KCM);