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 (copied && sock->type == SOCK_SEQPACKET) {
1069                 /* Wrote some bytes before encountering an
1070                  * error, return partial success.
1071                  */
1072                 goto partial_message;
1073         }
1074
1075         if (head != kcm->seq_skb)
1076                 kfree_skb(head);
1077
1078         err = sk_stream_error(sk, msg->msg_flags, err);
1079
1080         /* make sure we wake any epoll edge trigger waiter */
1081         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1082                 sk->sk_write_space(sk);
1083
1084         release_sock(sk);
1085         return err;
1086 }
1087
1088 static struct sk_buff *kcm_wait_data(struct sock *sk, int flags,
1089                                      long timeo, int *err)
1090 {
1091         struct sk_buff *skb;
1092
1093         while (!(skb = skb_peek(&sk->sk_receive_queue))) {
1094                 if (sk->sk_err) {
1095                         *err = sock_error(sk);
1096                         return NULL;
1097                 }
1098
1099                 if (sock_flag(sk, SOCK_DONE))
1100                         return NULL;
1101
1102                 if ((flags & MSG_DONTWAIT) || !timeo) {
1103                         *err = -EAGAIN;
1104                         return NULL;
1105                 }
1106
1107                 sk_wait_data(sk, &timeo, NULL);
1108
1109                 /* Handle signals */
1110                 if (signal_pending(current)) {
1111                         *err = sock_intr_errno(timeo);
1112                         return NULL;
1113                 }
1114         }
1115
1116         return skb;
1117 }
1118
1119 static int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
1120                        size_t len, int flags)
1121 {
1122         struct sock *sk = sock->sk;
1123         struct kcm_sock *kcm = kcm_sk(sk);
1124         int err = 0;
1125         long timeo;
1126         struct strp_msg *stm;
1127         int copied = 0;
1128         struct sk_buff *skb;
1129
1130         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1131
1132         lock_sock(sk);
1133
1134         skb = kcm_wait_data(sk, flags, timeo, &err);
1135         if (!skb)
1136                 goto out;
1137
1138         /* Okay, have a message on the receive queue */
1139
1140         stm = strp_msg(skb);
1141
1142         if (len > stm->full_len)
1143                 len = stm->full_len;
1144
1145         err = skb_copy_datagram_msg(skb, stm->offset, msg, len);
1146         if (err < 0)
1147                 goto out;
1148
1149         copied = len;
1150         if (likely(!(flags & MSG_PEEK))) {
1151                 KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1152                 if (copied < stm->full_len) {
1153                         if (sock->type == SOCK_DGRAM) {
1154                                 /* Truncated message */
1155                                 msg->msg_flags |= MSG_TRUNC;
1156                                 goto msg_finished;
1157                         }
1158                         stm->offset += copied;
1159                         stm->full_len -= copied;
1160                 } else {
1161 msg_finished:
1162                         /* Finished with message */
1163                         msg->msg_flags |= MSG_EOR;
1164                         KCM_STATS_INCR(kcm->stats.rx_msgs);
1165                         skb_unlink(skb, &sk->sk_receive_queue);
1166                         kfree_skb(skb);
1167                 }
1168         }
1169
1170 out:
1171         release_sock(sk);
1172
1173         return copied ? : err;
1174 }
1175
1176 static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1177                                struct pipe_inode_info *pipe, size_t len,
1178                                unsigned int flags)
1179 {
1180         struct sock *sk = sock->sk;
1181         struct kcm_sock *kcm = kcm_sk(sk);
1182         long timeo;
1183         struct strp_msg *stm;
1184         int err = 0;
1185         ssize_t copied;
1186         struct sk_buff *skb;
1187
1188         /* Only support splice for SOCKSEQPACKET */
1189
1190         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1191
1192         lock_sock(sk);
1193
1194         skb = kcm_wait_data(sk, flags, timeo, &err);
1195         if (!skb)
1196                 goto err_out;
1197
1198         /* Okay, have a message on the receive queue */
1199
1200         stm = strp_msg(skb);
1201
1202         if (len > stm->full_len)
1203                 len = stm->full_len;
1204
1205         copied = skb_splice_bits(skb, sk, stm->offset, pipe, len, flags);
1206         if (copied < 0) {
1207                 err = copied;
1208                 goto err_out;
1209         }
1210
1211         KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1212
1213         stm->offset += copied;
1214         stm->full_len -= copied;
1215
1216         /* We have no way to return MSG_EOR. If all the bytes have been
1217          * read we still leave the message in the receive socket buffer.
1218          * A subsequent recvmsg needs to be done to return MSG_EOR and
1219          * finish reading the message.
1220          */
1221
1222         release_sock(sk);
1223
1224         return copied;
1225
1226 err_out:
1227         release_sock(sk);
1228
1229         return err;
1230 }
1231
1232 /* kcm sock lock held */
1233 static void kcm_recv_disable(struct kcm_sock *kcm)
1234 {
1235         struct kcm_mux *mux = kcm->mux;
1236
1237         if (kcm->rx_disabled)
1238                 return;
1239
1240         spin_lock_bh(&mux->rx_lock);
1241
1242         kcm->rx_disabled = 1;
1243
1244         /* If a psock is reserved we'll do cleanup in unreserve */
1245         if (!kcm->rx_psock) {
1246                 if (kcm->rx_wait) {
1247                         list_del(&kcm->wait_rx_list);
1248                         /* paired with lockless reads in kcm_rfree() */
1249                         WRITE_ONCE(kcm->rx_wait, false);
1250                 }
1251
1252                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1253         }
1254
1255         spin_unlock_bh(&mux->rx_lock);
1256 }
1257
1258 /* kcm sock lock held */
1259 static void kcm_recv_enable(struct kcm_sock *kcm)
1260 {
1261         struct kcm_mux *mux = kcm->mux;
1262
1263         if (!kcm->rx_disabled)
1264                 return;
1265
1266         spin_lock_bh(&mux->rx_lock);
1267
1268         kcm->rx_disabled = 0;
1269         kcm_rcv_ready(kcm);
1270
1271         spin_unlock_bh(&mux->rx_lock);
1272 }
1273
1274 static int kcm_setsockopt(struct socket *sock, int level, int optname,
1275                           char __user *optval, unsigned int optlen)
1276 {
1277         struct kcm_sock *kcm = kcm_sk(sock->sk);
1278         int val, valbool;
1279         int err = 0;
1280
1281         if (level != SOL_KCM)
1282                 return -ENOPROTOOPT;
1283
1284         if (optlen < sizeof(int))
1285                 return -EINVAL;
1286
1287         if (get_user(val, (int __user *)optval))
1288                 return -EINVAL;
1289
1290         valbool = val ? 1 : 0;
1291
1292         switch (optname) {
1293         case KCM_RECV_DISABLE:
1294                 lock_sock(&kcm->sk);
1295                 if (valbool)
1296                         kcm_recv_disable(kcm);
1297                 else
1298                         kcm_recv_enable(kcm);
1299                 release_sock(&kcm->sk);
1300                 break;
1301         default:
1302                 err = -ENOPROTOOPT;
1303         }
1304
1305         return err;
1306 }
1307
1308 static int kcm_getsockopt(struct socket *sock, int level, int optname,
1309                           char __user *optval, int __user *optlen)
1310 {
1311         struct kcm_sock *kcm = kcm_sk(sock->sk);
1312         int val, len;
1313
1314         if (level != SOL_KCM)
1315                 return -ENOPROTOOPT;
1316
1317         if (get_user(len, optlen))
1318                 return -EFAULT;
1319
1320         len = min_t(unsigned int, len, sizeof(int));
1321         if (len < 0)
1322                 return -EINVAL;
1323
1324         switch (optname) {
1325         case KCM_RECV_DISABLE:
1326                 val = kcm->rx_disabled;
1327                 break;
1328         default:
1329                 return -ENOPROTOOPT;
1330         }
1331
1332         if (put_user(len, optlen))
1333                 return -EFAULT;
1334         if (copy_to_user(optval, &val, len))
1335                 return -EFAULT;
1336         return 0;
1337 }
1338
1339 static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1340 {
1341         struct kcm_sock *tkcm;
1342         struct list_head *head;
1343         int index = 0;
1344
1345         /* For SOCK_SEQPACKET sock type, datagram_poll checks the sk_state, so
1346          * we set sk_state, otherwise epoll_wait always returns right away with
1347          * EPOLLHUP
1348          */
1349         kcm->sk.sk_state = TCP_ESTABLISHED;
1350
1351         /* Add to mux's kcm sockets list */
1352         kcm->mux = mux;
1353         spin_lock_bh(&mux->lock);
1354
1355         head = &mux->kcm_socks;
1356         list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1357                 if (tkcm->index != index)
1358                         break;
1359                 head = &tkcm->kcm_sock_list;
1360                 index++;
1361         }
1362
1363         list_add(&kcm->kcm_sock_list, head);
1364         kcm->index = index;
1365
1366         mux->kcm_socks_cnt++;
1367         spin_unlock_bh(&mux->lock);
1368
1369         INIT_WORK(&kcm->tx_work, kcm_tx_work);
1370
1371         spin_lock_bh(&mux->rx_lock);
1372         kcm_rcv_ready(kcm);
1373         spin_unlock_bh(&mux->rx_lock);
1374 }
1375
1376 static int kcm_attach(struct socket *sock, struct socket *csock,
1377                       struct bpf_prog *prog)
1378 {
1379         struct kcm_sock *kcm = kcm_sk(sock->sk);
1380         struct kcm_mux *mux = kcm->mux;
1381         struct sock *csk;
1382         struct kcm_psock *psock = NULL, *tpsock;
1383         struct list_head *head;
1384         int index = 0;
1385         static const struct strp_callbacks cb = {
1386                 .rcv_msg = kcm_rcv_strparser,
1387                 .parse_msg = kcm_parse_func_strparser,
1388                 .read_sock_done = kcm_read_sock_done,
1389         };
1390         int err = 0;
1391
1392         csk = csock->sk;
1393         if (!csk)
1394                 return -EINVAL;
1395
1396         lock_sock(csk);
1397
1398         /* Only allow TCP sockets to be attached for now */
1399         if ((csk->sk_family != AF_INET && csk->sk_family != AF_INET6) ||
1400             csk->sk_protocol != IPPROTO_TCP) {
1401                 err = -EOPNOTSUPP;
1402                 goto out;
1403         }
1404
1405         /* Don't allow listeners or closed sockets */
1406         if (csk->sk_state == TCP_LISTEN || csk->sk_state == TCP_CLOSE) {
1407                 err = -EOPNOTSUPP;
1408                 goto out;
1409         }
1410
1411         psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1412         if (!psock) {
1413                 err = -ENOMEM;
1414                 goto out;
1415         }
1416
1417         psock->mux = mux;
1418         psock->sk = csk;
1419         psock->bpf_prog = prog;
1420
1421         write_lock_bh(&csk->sk_callback_lock);
1422
1423         /* Check if sk_user_data is aready by KCM or someone else.
1424          * Must be done under lock to prevent race conditions.
1425          */
1426         if (csk->sk_user_data) {
1427                 write_unlock_bh(&csk->sk_callback_lock);
1428                 kmem_cache_free(kcm_psockp, psock);
1429                 err = -EALREADY;
1430                 goto out;
1431         }
1432
1433         err = strp_init(&psock->strp, csk, &cb);
1434         if (err) {
1435                 write_unlock_bh(&csk->sk_callback_lock);
1436                 kmem_cache_free(kcm_psockp, psock);
1437                 goto out;
1438         }
1439
1440         psock->save_data_ready = csk->sk_data_ready;
1441         psock->save_write_space = csk->sk_write_space;
1442         psock->save_state_change = csk->sk_state_change;
1443         csk->sk_user_data = psock;
1444         csk->sk_data_ready = psock_data_ready;
1445         csk->sk_write_space = psock_write_space;
1446         csk->sk_state_change = psock_state_change;
1447
1448         write_unlock_bh(&csk->sk_callback_lock);
1449
1450         sock_hold(csk);
1451
1452         /* Finished initialization, now add the psock to the MUX. */
1453         spin_lock_bh(&mux->lock);
1454         head = &mux->psocks;
1455         list_for_each_entry(tpsock, &mux->psocks, psock_list) {
1456                 if (tpsock->index != index)
1457                         break;
1458                 head = &tpsock->psock_list;
1459                 index++;
1460         }
1461
1462         list_add(&psock->psock_list, head);
1463         psock->index = index;
1464
1465         KCM_STATS_INCR(mux->stats.psock_attach);
1466         mux->psocks_cnt++;
1467         psock_now_avail(psock);
1468         spin_unlock_bh(&mux->lock);
1469
1470         /* Schedule RX work in case there are already bytes queued */
1471         strp_check_rcv(&psock->strp);
1472
1473 out:
1474         release_sock(csk);
1475
1476         return err;
1477 }
1478
1479 static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1480 {
1481         struct socket *csock;
1482         struct bpf_prog *prog;
1483         int err;
1484
1485         csock = sockfd_lookup(info->fd, &err);
1486         if (!csock)
1487                 return -ENOENT;
1488
1489         prog = bpf_prog_get_type(info->bpf_fd, BPF_PROG_TYPE_SOCKET_FILTER);
1490         if (IS_ERR(prog)) {
1491                 err = PTR_ERR(prog);
1492                 goto out;
1493         }
1494
1495         err = kcm_attach(sock, csock, prog);
1496         if (err) {
1497                 bpf_prog_put(prog);
1498                 goto out;
1499         }
1500
1501         /* Keep reference on file also */
1502
1503         return 0;
1504 out:
1505         fput(csock->file);
1506         return err;
1507 }
1508
1509 static void kcm_unattach(struct kcm_psock *psock)
1510 {
1511         struct sock *csk = psock->sk;
1512         struct kcm_mux *mux = psock->mux;
1513
1514         lock_sock(csk);
1515
1516         /* Stop getting callbacks from TCP socket. After this there should
1517          * be no way to reserve a kcm for this psock.
1518          */
1519         write_lock_bh(&csk->sk_callback_lock);
1520         csk->sk_user_data = NULL;
1521         csk->sk_data_ready = psock->save_data_ready;
1522         csk->sk_write_space = psock->save_write_space;
1523         csk->sk_state_change = psock->save_state_change;
1524         strp_stop(&psock->strp);
1525
1526         if (WARN_ON(psock->rx_kcm)) {
1527                 write_unlock_bh(&csk->sk_callback_lock);
1528                 release_sock(csk);
1529                 return;
1530         }
1531
1532         spin_lock_bh(&mux->rx_lock);
1533
1534         /* Stop receiver activities. After this point psock should not be
1535          * able to get onto ready list either through callbacks or work.
1536          */
1537         if (psock->ready_rx_msg) {
1538                 list_del(&psock->psock_ready_list);
1539                 kfree_skb(psock->ready_rx_msg);
1540                 psock->ready_rx_msg = NULL;
1541                 KCM_STATS_INCR(mux->stats.rx_ready_drops);
1542         }
1543
1544         spin_unlock_bh(&mux->rx_lock);
1545
1546         write_unlock_bh(&csk->sk_callback_lock);
1547
1548         /* Call strp_done without sock lock */
1549         release_sock(csk);
1550         strp_done(&psock->strp);
1551         lock_sock(csk);
1552
1553         bpf_prog_put(psock->bpf_prog);
1554
1555         spin_lock_bh(&mux->lock);
1556
1557         aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1558         save_strp_stats(&psock->strp, &mux->aggregate_strp_stats);
1559
1560         KCM_STATS_INCR(mux->stats.psock_unattach);
1561
1562         if (psock->tx_kcm) {
1563                 /* psock was reserved.  Just mark it finished and we will clean
1564                  * up in the kcm paths, we need kcm lock which can not be
1565                  * acquired here.
1566                  */
1567                 KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1568                 spin_unlock_bh(&mux->lock);
1569
1570                 /* We are unattaching a socket that is reserved. Abort the
1571                  * socket since we may be out of sync in sending on it. We need
1572                  * to do this without the mux lock.
1573                  */
1574                 kcm_abort_tx_psock(psock, EPIPE, false);
1575
1576                 spin_lock_bh(&mux->lock);
1577                 if (!psock->tx_kcm) {
1578                         /* psock now unreserved in window mux was unlocked */
1579                         goto no_reserved;
1580                 }
1581                 psock->done = 1;
1582
1583                 /* Commit done before queuing work to process it */
1584                 smp_mb();
1585
1586                 /* Queue tx work to make sure psock->done is handled */
1587                 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
1588                 spin_unlock_bh(&mux->lock);
1589         } else {
1590 no_reserved:
1591                 if (!psock->tx_stopped)
1592                         list_del(&psock->psock_avail_list);
1593                 list_del(&psock->psock_list);
1594                 mux->psocks_cnt--;
1595                 spin_unlock_bh(&mux->lock);
1596
1597                 sock_put(csk);
1598                 fput(csk->sk_socket->file);
1599                 kmem_cache_free(kcm_psockp, psock);
1600         }
1601
1602         release_sock(csk);
1603 }
1604
1605 static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1606 {
1607         struct kcm_sock *kcm = kcm_sk(sock->sk);
1608         struct kcm_mux *mux = kcm->mux;
1609         struct kcm_psock *psock;
1610         struct socket *csock;
1611         struct sock *csk;
1612         int err;
1613
1614         csock = sockfd_lookup(info->fd, &err);
1615         if (!csock)
1616                 return -ENOENT;
1617
1618         csk = csock->sk;
1619         if (!csk) {
1620                 err = -EINVAL;
1621                 goto out;
1622         }
1623
1624         err = -ENOENT;
1625
1626         spin_lock_bh(&mux->lock);
1627
1628         list_for_each_entry(psock, &mux->psocks, psock_list) {
1629                 if (psock->sk != csk)
1630                         continue;
1631
1632                 /* Found the matching psock */
1633
1634                 if (psock->unattaching || WARN_ON(psock->done)) {
1635                         err = -EALREADY;
1636                         break;
1637                 }
1638
1639                 psock->unattaching = 1;
1640
1641                 spin_unlock_bh(&mux->lock);
1642
1643                 /* Lower socket lock should already be held */
1644                 kcm_unattach(psock);
1645
1646                 err = 0;
1647                 goto out;
1648         }
1649
1650         spin_unlock_bh(&mux->lock);
1651
1652 out:
1653         fput(csock->file);
1654         return err;
1655 }
1656
1657 static struct proto kcm_proto = {
1658         .name   = "KCM",
1659         .owner  = THIS_MODULE,
1660         .obj_size = sizeof(struct kcm_sock),
1661 };
1662
1663 /* Clone a kcm socket. */
1664 static struct file *kcm_clone(struct socket *osock)
1665 {
1666         struct socket *newsock;
1667         struct sock *newsk;
1668
1669         newsock = sock_alloc();
1670         if (!newsock)
1671                 return ERR_PTR(-ENFILE);
1672
1673         newsock->type = osock->type;
1674         newsock->ops = osock->ops;
1675
1676         __module_get(newsock->ops->owner);
1677
1678         newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1679                          &kcm_proto, false);
1680         if (!newsk) {
1681                 sock_release(newsock);
1682                 return ERR_PTR(-ENOMEM);
1683         }
1684         sock_init_data(newsock, newsk);
1685         init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1686
1687         return sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1688 }
1689
1690 static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
1691 {
1692         int err;
1693
1694         switch (cmd) {
1695         case SIOCKCMATTACH: {
1696                 struct kcm_attach info;
1697
1698                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1699                         return -EFAULT;
1700
1701                 err = kcm_attach_ioctl(sock, &info);
1702
1703                 break;
1704         }
1705         case SIOCKCMUNATTACH: {
1706                 struct kcm_unattach info;
1707
1708                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1709                         return -EFAULT;
1710
1711                 err = kcm_unattach_ioctl(sock, &info);
1712
1713                 break;
1714         }
1715         case SIOCKCMCLONE: {
1716                 struct kcm_clone info;
1717                 struct file *file;
1718
1719                 info.fd = get_unused_fd_flags(0);
1720                 if (unlikely(info.fd < 0))
1721                         return info.fd;
1722
1723                 file = kcm_clone(sock);
1724                 if (IS_ERR(file)) {
1725                         put_unused_fd(info.fd);
1726                         return PTR_ERR(file);
1727                 }
1728                 if (copy_to_user((void __user *)arg, &info,
1729                                  sizeof(info))) {
1730                         put_unused_fd(info.fd);
1731                         fput(file);
1732                         return -EFAULT;
1733                 }
1734                 fd_install(info.fd, file);
1735                 err = 0;
1736                 break;
1737         }
1738         default:
1739                 err = -ENOIOCTLCMD;
1740                 break;
1741         }
1742
1743         return err;
1744 }
1745
1746 static void free_mux(struct rcu_head *rcu)
1747 {
1748         struct kcm_mux *mux = container_of(rcu,
1749             struct kcm_mux, rcu);
1750
1751         kmem_cache_free(kcm_muxp, mux);
1752 }
1753
1754 static void release_mux(struct kcm_mux *mux)
1755 {
1756         struct kcm_net *knet = mux->knet;
1757         struct kcm_psock *psock, *tmp_psock;
1758
1759         /* Release psocks */
1760         list_for_each_entry_safe(psock, tmp_psock,
1761                                  &mux->psocks, psock_list) {
1762                 if (!WARN_ON(psock->unattaching))
1763                         kcm_unattach(psock);
1764         }
1765
1766         if (WARN_ON(mux->psocks_cnt))
1767                 return;
1768
1769         __skb_queue_purge(&mux->rx_hold_queue);
1770
1771         mutex_lock(&knet->mutex);
1772         aggregate_mux_stats(&mux->stats, &knet->aggregate_mux_stats);
1773         aggregate_psock_stats(&mux->aggregate_psock_stats,
1774                               &knet->aggregate_psock_stats);
1775         aggregate_strp_stats(&mux->aggregate_strp_stats,
1776                              &knet->aggregate_strp_stats);
1777         list_del_rcu(&mux->kcm_mux_list);
1778         knet->count--;
1779         mutex_unlock(&knet->mutex);
1780
1781         call_rcu(&mux->rcu, free_mux);
1782 }
1783
1784 static void kcm_done(struct kcm_sock *kcm)
1785 {
1786         struct kcm_mux *mux = kcm->mux;
1787         struct sock *sk = &kcm->sk;
1788         int socks_cnt;
1789
1790         spin_lock_bh(&mux->rx_lock);
1791         if (kcm->rx_psock) {
1792                 /* Cleanup in unreserve_rx_kcm */
1793                 WARN_ON(kcm->done);
1794                 kcm->rx_disabled = 1;
1795                 kcm->done = 1;
1796                 spin_unlock_bh(&mux->rx_lock);
1797                 return;
1798         }
1799
1800         if (kcm->rx_wait) {
1801                 list_del(&kcm->wait_rx_list);
1802                 /* paired with lockless reads in kcm_rfree() */
1803                 WRITE_ONCE(kcm->rx_wait, false);
1804         }
1805         /* Move any pending receive messages to other kcm sockets */
1806         requeue_rx_msgs(mux, &sk->sk_receive_queue);
1807
1808         spin_unlock_bh(&mux->rx_lock);
1809
1810         if (WARN_ON(sk_rmem_alloc_get(sk)))
1811                 return;
1812
1813         /* Detach from MUX */
1814         spin_lock_bh(&mux->lock);
1815
1816         list_del(&kcm->kcm_sock_list);
1817         mux->kcm_socks_cnt--;
1818         socks_cnt = mux->kcm_socks_cnt;
1819
1820         spin_unlock_bh(&mux->lock);
1821
1822         if (!socks_cnt) {
1823                 /* We are done with the mux now. */
1824                 release_mux(mux);
1825         }
1826
1827         WARN_ON(kcm->rx_wait);
1828
1829         sock_put(&kcm->sk);
1830 }
1831
1832 /* Called by kcm_release to close a KCM socket.
1833  * If this is the last KCM socket on the MUX, destroy the MUX.
1834  */
1835 static int kcm_release(struct socket *sock)
1836 {
1837         struct sock *sk = sock->sk;
1838         struct kcm_sock *kcm;
1839         struct kcm_mux *mux;
1840         struct kcm_psock *psock;
1841
1842         if (!sk)
1843                 return 0;
1844
1845         kcm = kcm_sk(sk);
1846         mux = kcm->mux;
1847
1848         sock_orphan(sk);
1849         kfree_skb(kcm->seq_skb);
1850
1851         lock_sock(sk);
1852         /* Purge queue under lock to avoid race condition with tx_work trying
1853          * to act when queue is nonempty. If tx_work runs after this point
1854          * it will just return.
1855          */
1856         __skb_queue_purge(&sk->sk_write_queue);
1857
1858         /* Set tx_stopped. This is checked when psock is bound to a kcm and we
1859          * get a writespace callback. This prevents further work being queued
1860          * from the callback (unbinding the psock occurs after canceling work.
1861          */
1862         kcm->tx_stopped = 1;
1863
1864         release_sock(sk);
1865
1866         spin_lock_bh(&mux->lock);
1867         if (kcm->tx_wait) {
1868                 /* Take of tx_wait list, after this point there should be no way
1869                  * that a psock will be assigned to this kcm.
1870                  */
1871                 list_del(&kcm->wait_psock_list);
1872                 kcm->tx_wait = false;
1873         }
1874         spin_unlock_bh(&mux->lock);
1875
1876         /* Cancel work. After this point there should be no outside references
1877          * to the kcm socket.
1878          */
1879         cancel_work_sync(&kcm->tx_work);
1880
1881         lock_sock(sk);
1882         psock = kcm->tx_psock;
1883         if (psock) {
1884                 /* A psock was reserved, so we need to kill it since it
1885                  * may already have some bytes queued from a message. We
1886                  * need to do this after removing kcm from tx_wait list.
1887                  */
1888                 kcm_abort_tx_psock(psock, EPIPE, false);
1889                 unreserve_psock(kcm);
1890         }
1891         release_sock(sk);
1892
1893         WARN_ON(kcm->tx_wait);
1894         WARN_ON(kcm->tx_psock);
1895
1896         sock->sk = NULL;
1897
1898         kcm_done(kcm);
1899
1900         return 0;
1901 }
1902
1903 static const struct proto_ops kcm_dgram_ops = {
1904         .family =       PF_KCM,
1905         .owner =        THIS_MODULE,
1906         .release =      kcm_release,
1907         .bind =         sock_no_bind,
1908         .connect =      sock_no_connect,
1909         .socketpair =   sock_no_socketpair,
1910         .accept =       sock_no_accept,
1911         .getname =      sock_no_getname,
1912         .poll =         datagram_poll,
1913         .ioctl =        kcm_ioctl,
1914         .listen =       sock_no_listen,
1915         .shutdown =     sock_no_shutdown,
1916         .setsockopt =   kcm_setsockopt,
1917         .getsockopt =   kcm_getsockopt,
1918         .sendmsg =      kcm_sendmsg,
1919         .recvmsg =      kcm_recvmsg,
1920         .mmap =         sock_no_mmap,
1921         .sendpage =     kcm_sendpage,
1922 };
1923
1924 static const struct proto_ops kcm_seqpacket_ops = {
1925         .family =       PF_KCM,
1926         .owner =        THIS_MODULE,
1927         .release =      kcm_release,
1928         .bind =         sock_no_bind,
1929         .connect =      sock_no_connect,
1930         .socketpair =   sock_no_socketpair,
1931         .accept =       sock_no_accept,
1932         .getname =      sock_no_getname,
1933         .poll =         datagram_poll,
1934         .ioctl =        kcm_ioctl,
1935         .listen =       sock_no_listen,
1936         .shutdown =     sock_no_shutdown,
1937         .setsockopt =   kcm_setsockopt,
1938         .getsockopt =   kcm_getsockopt,
1939         .sendmsg =      kcm_sendmsg,
1940         .recvmsg =      kcm_recvmsg,
1941         .mmap =         sock_no_mmap,
1942         .sendpage =     kcm_sendpage,
1943         .splice_read =  kcm_splice_read,
1944 };
1945
1946 /* Create proto operation for kcm sockets */
1947 static int kcm_create(struct net *net, struct socket *sock,
1948                       int protocol, int kern)
1949 {
1950         struct kcm_net *knet = net_generic(net, kcm_net_id);
1951         struct sock *sk;
1952         struct kcm_mux *mux;
1953
1954         switch (sock->type) {
1955         case SOCK_DGRAM:
1956                 sock->ops = &kcm_dgram_ops;
1957                 break;
1958         case SOCK_SEQPACKET:
1959                 sock->ops = &kcm_seqpacket_ops;
1960                 break;
1961         default:
1962                 return -ESOCKTNOSUPPORT;
1963         }
1964
1965         if (protocol != KCMPROTO_CONNECTED)
1966                 return -EPROTONOSUPPORT;
1967
1968         sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
1969         if (!sk)
1970                 return -ENOMEM;
1971
1972         /* Allocate a kcm mux, shared between KCM sockets */
1973         mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
1974         if (!mux) {
1975                 sk_free(sk);
1976                 return -ENOMEM;
1977         }
1978
1979         spin_lock_init(&mux->lock);
1980         spin_lock_init(&mux->rx_lock);
1981         INIT_LIST_HEAD(&mux->kcm_socks);
1982         INIT_LIST_HEAD(&mux->kcm_rx_waiters);
1983         INIT_LIST_HEAD(&mux->kcm_tx_waiters);
1984
1985         INIT_LIST_HEAD(&mux->psocks);
1986         INIT_LIST_HEAD(&mux->psocks_ready);
1987         INIT_LIST_HEAD(&mux->psocks_avail);
1988
1989         mux->knet = knet;
1990
1991         /* Add new MUX to list */
1992         mutex_lock(&knet->mutex);
1993         list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
1994         knet->count++;
1995         mutex_unlock(&knet->mutex);
1996
1997         skb_queue_head_init(&mux->rx_hold_queue);
1998
1999         /* Init KCM socket */
2000         sock_init_data(sock, sk);
2001         init_kcm_sock(kcm_sk(sk), mux);
2002
2003         return 0;
2004 }
2005
2006 static const struct net_proto_family kcm_family_ops = {
2007         .family = PF_KCM,
2008         .create = kcm_create,
2009         .owner  = THIS_MODULE,
2010 };
2011
2012 static __net_init int kcm_init_net(struct net *net)
2013 {
2014         struct kcm_net *knet = net_generic(net, kcm_net_id);
2015
2016         INIT_LIST_HEAD_RCU(&knet->mux_list);
2017         mutex_init(&knet->mutex);
2018
2019         return 0;
2020 }
2021
2022 static __net_exit void kcm_exit_net(struct net *net)
2023 {
2024         struct kcm_net *knet = net_generic(net, kcm_net_id);
2025
2026         /* All KCM sockets should be closed at this point, which should mean
2027          * that all multiplexors and psocks have been destroyed.
2028          */
2029         WARN_ON(!list_empty(&knet->mux_list));
2030 }
2031
2032 static struct pernet_operations kcm_net_ops = {
2033         .init = kcm_init_net,
2034         .exit = kcm_exit_net,
2035         .id   = &kcm_net_id,
2036         .size = sizeof(struct kcm_net),
2037 };
2038
2039 static int __init kcm_init(void)
2040 {
2041         int err = -ENOMEM;
2042
2043         kcm_muxp = kmem_cache_create("kcm_mux_cache",
2044                                      sizeof(struct kcm_mux), 0,
2045                                      SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2046         if (!kcm_muxp)
2047                 goto fail;
2048
2049         kcm_psockp = kmem_cache_create("kcm_psock_cache",
2050                                        sizeof(struct kcm_psock), 0,
2051                                         SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2052         if (!kcm_psockp)
2053                 goto fail;
2054
2055         kcm_wq = create_singlethread_workqueue("kkcmd");
2056         if (!kcm_wq)
2057                 goto fail;
2058
2059         err = proto_register(&kcm_proto, 1);
2060         if (err)
2061                 goto fail;
2062
2063         err = register_pernet_device(&kcm_net_ops);
2064         if (err)
2065                 goto net_ops_fail;
2066
2067         err = sock_register(&kcm_family_ops);
2068         if (err)
2069                 goto sock_register_fail;
2070
2071         err = kcm_proc_init();
2072         if (err)
2073                 goto proc_init_fail;
2074
2075         return 0;
2076
2077 proc_init_fail:
2078         sock_unregister(PF_KCM);
2079
2080 sock_register_fail:
2081         unregister_pernet_device(&kcm_net_ops);
2082
2083 net_ops_fail:
2084         proto_unregister(&kcm_proto);
2085
2086 fail:
2087         kmem_cache_destroy(kcm_muxp);
2088         kmem_cache_destroy(kcm_psockp);
2089
2090         if (kcm_wq)
2091                 destroy_workqueue(kcm_wq);
2092
2093         return err;
2094 }
2095
2096 static void __exit kcm_exit(void)
2097 {
2098         kcm_proc_exit();
2099         sock_unregister(PF_KCM);
2100         unregister_pernet_device(&kcm_net_ops);
2101         proto_unregister(&kcm_proto);
2102         destroy_workqueue(kcm_wq);
2103
2104         kmem_cache_destroy(kcm_muxp);
2105         kmem_cache_destroy(kcm_psockp);
2106 }
2107
2108 module_init(kcm_init);
2109 module_exit(kcm_exit);
2110
2111 MODULE_LICENSE("GPL");
2112 MODULE_ALIAS_NETPROTO(PF_KCM);