net/kcm/kcmsock.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Kernel Connection Multiplexor
   4  *
   5  * Copyright (c) 2016 Tom Herbert <tom@herbertland.com>
   6  */
   7
   8 #include <linux/bpf.h>
   9 #include <linux/errno.h>
  10 #include <linux/errqueue.h>
  11 #include <linux/file.h>
  12 #include <linux/filter.h>
  13 #include <linux/in.h>
  14 #include <linux/kernel.h>
  15 #include <linux/module.h>
  16 #include <linux/net.h>
  17 #include <linux/netdevice.h>
  18 #include <linux/poll.h>
  19 #include <linux/rculist.h>
  20 #include <linux/skbuff.h>
  21 #include <linux/socket.h>
  22 #include <linux/uaccess.h>
  23 #include <linux/workqueue.h>
  24 #include <linux/syscalls.h>
  25 #include <linux/sched/signal.h>
  26
  27 #include <net/kcm.h>
  28 #include <net/netns/generic.h>
  29 #include <net/sock.h>
  30 #include <uapi/linux/kcm.h>
  31
  32 unsigned int kcm_net_id;
  33
  34 static struct kmem_cache *kcm_psockp __read_mostly;
  35 static struct kmem_cache *kcm_muxp __read_mostly;
  36 static struct workqueue_struct *kcm_wq;
  37
  38 static inline struct kcm_sock *kcm_sk(const struct sock *sk)
  39 {
  40         return (struct kcm_sock *)sk;
  41 }
  42
  43 static inline struct kcm_tx_msg *kcm_tx_msg(struct sk_buff *skb)
  44 {
  45         return (struct kcm_tx_msg *)skb->cb;
  46 }
  47
  48 static void report_csk_error(struct sock *csk, int err)
  49 {
  50         csk->sk_err = EPIPE;
  51         sk_error_report(csk);
  52 }
  53
  54 static void kcm_abort_tx_psock(struct kcm_psock *psock, int err,
  55                                bool wakeup_kcm)
  56 {
  57         struct sock *csk = psock->sk;
  58         struct kcm_mux *mux = psock->mux;
  59
  60         /* Unrecoverable error in transmit */
  61
  62         spin_lock_bh(&mux->lock);
  63
  64         if (psock->tx_stopped) {
  65                 spin_unlock_bh(&mux->lock);
  66                 return;
  67         }
  68
  69         psock->tx_stopped = 1;
  70         KCM_STATS_INCR(psock->stats.tx_aborts);
  71
  72         if (!psock->tx_kcm) {
  73                 /* Take off psocks_avail list */
  74                 list_del(&psock->psock_avail_list);
  75         } else if (wakeup_kcm) {
  76                 /* In this case psock is being aborted while outside of
  77                  * write_msgs and psock is reserved. Schedule tx_work
  78                  * to handle the failure there. Need to commit tx_stopped
  79                  * before queuing work.
  80                  */
  81                 smp_mb();
  82
  83                 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
  84         }
  85
  86         spin_unlock_bh(&mux->lock);
  87
  88         /* Report error on lower socket */
  89         report_csk_error(csk, err);
  90 }
  91
  92 /* RX mux lock held. */
  93 static void kcm_update_rx_mux_stats(struct kcm_mux *mux,
  94                                     struct kcm_psock *psock)
  95 {
  96         STRP_STATS_ADD(mux->stats.rx_bytes,
  97                        psock->strp.stats.bytes -
  98                        psock->saved_rx_bytes);
  99         mux->stats.rx_msgs +=
 100                 psock->strp.stats.msgs - psock->saved_rx_msgs;
 101         psock->saved_rx_msgs = psock->strp.stats.msgs;
 102         psock->saved_rx_bytes = psock->strp.stats.bytes;
 103 }
 104
 105 static void kcm_update_tx_mux_stats(struct kcm_mux *mux,
 106                                     struct kcm_psock *psock)
 107 {
 108         KCM_STATS_ADD(mux->stats.tx_bytes,
 109                       psock->stats.tx_bytes - psock->saved_tx_bytes);
 110         mux->stats.tx_msgs +=
 111                 psock->stats.tx_msgs - psock->saved_tx_msgs;
 112         psock->saved_tx_msgs = psock->stats.tx_msgs;
 113         psock->saved_tx_bytes = psock->stats.tx_bytes;
 114 }
 115
 116 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb);
 117
 118 /* KCM is ready to receive messages on its queue-- either the KCM is new or
 119  * has become unblocked after being blocked on full socket buffer. Queue any
 120  * pending ready messages on a psock. RX mux lock held.
 121  */
 122 static void kcm_rcv_ready(struct kcm_sock *kcm)
 123 {
 124         struct kcm_mux *mux = kcm->mux;
 125         struct kcm_psock *psock;
 126         struct sk_buff *skb;
 127
 128         if (unlikely(kcm->rx_wait || kcm->rx_psock || kcm->rx_disabled))
 129                 return;
 130
 131         while (unlikely((skb = __skb_dequeue(&mux->rx_hold_queue)))) {
 132                 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 133                         /* Assuming buffer limit has been reached */
 134                         skb_queue_head(&mux->rx_hold_queue, skb);
 135                         WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
 136                         return;
 137                 }
 138         }
 139
 140         while (!list_empty(&mux->psocks_ready)) {
 141                 psock = list_first_entry(&mux->psocks_ready, struct kcm_psock,
 142                                          psock_ready_list);
 143
 144                 if (kcm_queue_rcv_skb(&kcm->sk, psock->ready_rx_msg)) {
 145                         /* Assuming buffer limit has been reached */
 146                         WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
 147                         return;
 148                 }
 149
 150                 /* Consumed the ready message on the psock. Schedule rx_work to
 151                  * get more messages.
 152                  */
 153                 list_del(&psock->psock_ready_list);
 154                 psock->ready_rx_msg = NULL;
 155                 /* Commit clearing of ready_rx_msg for queuing work */
 156                 smp_mb();
 157
 158                 strp_unpause(&psock->strp);
 159                 strp_check_rcv(&psock->strp);
 160         }
 161
 162         /* Buffer limit is okay now, add to ready list */
 163         list_add_tail(&kcm->wait_rx_list,
 164                       &kcm->mux->kcm_rx_waiters);
 165         /* paired with lockless reads in kcm_rfree() */
 166         WRITE_ONCE(kcm->rx_wait, true);
 167 }
 168
 169 static void kcm_rfree(struct sk_buff *skb)
 170 {
 171         struct sock *sk = skb->sk;
 172         struct kcm_sock *kcm = kcm_sk(sk);
 173         struct kcm_mux *mux = kcm->mux;
 174         unsigned int len = skb->truesize;
 175
 176         sk_mem_uncharge(sk, len);
 177         atomic_sub(len, &sk->sk_rmem_alloc);
 178
 179         /* For reading rx_wait and rx_psock without holding lock */
 180         smp_mb__after_atomic();
 181
 182         if (!READ_ONCE(kcm->rx_wait) && !READ_ONCE(kcm->rx_psock) &&
 183             sk_rmem_alloc_get(sk) < sk->sk_rcvlowat) {
 184                 spin_lock_bh(&mux->rx_lock);
 185                 kcm_rcv_ready(kcm);
 186                 spin_unlock_bh(&mux->rx_lock);
 187         }
 188 }
 189
 190 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb)
 191 {
 192         struct sk_buff_head *list = &sk->sk_receive_queue;
 193
 194         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
 195                 return -ENOMEM;
 196
 197         if (!sk_rmem_schedule(sk, skb, skb->truesize))
 198                 return -ENOBUFS;
 199
 200         skb->dev = NULL;
 201
 202         skb_orphan(skb);
 203         skb->sk = sk;
 204         skb->destructor = kcm_rfree;
 205         atomic_add(skb->truesize, &sk->sk_rmem_alloc);
 206         sk_mem_charge(sk, skb->truesize);
 207
 208         skb_queue_tail(list, skb);
 209
 210         if (!sock_flag(sk, SOCK_DEAD))
 211                 sk->sk_data_ready(sk);
 212
 213         return 0;
 214 }
 215
 216 /* Requeue received messages for a kcm socket to other kcm sockets. This is
 217  * called with a kcm socket is receive disabled.
 218  * RX mux lock held.
 219  */
 220 static void requeue_rx_msgs(struct kcm_mux *mux, struct sk_buff_head *head)
 221 {
 222         struct sk_buff *skb;
 223         struct kcm_sock *kcm;
 224
 225         while ((skb = skb_dequeue(head))) {
 226                 /* Reset destructor to avoid calling kcm_rcv_ready */
 227                 skb->destructor = sock_rfree;
 228                 skb_orphan(skb);
 229 try_again:
 230                 if (list_empty(&mux->kcm_rx_waiters)) {
 231                         skb_queue_tail(&mux->rx_hold_queue, skb);
 232                         continue;
 233                 }
 234
 235                 kcm = list_first_entry(&mux->kcm_rx_waiters,
 236                                        struct kcm_sock, wait_rx_list);
 237
 238                 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 239                         /* Should mean socket buffer full */
 240                         list_del(&kcm->wait_rx_list);
 241                         /* paired with lockless reads in kcm_rfree() */
 242                         WRITE_ONCE(kcm->rx_wait, false);
 243
 244                         /* Commit rx_wait to read in kcm_free */
 245                         smp_wmb();
 246
 247                         goto try_again;
 248                 }
 249         }
 250 }
 251
 252 /* Lower sock lock held */
 253 static struct kcm_sock *reserve_rx_kcm(struct kcm_psock *psock,
 254                                        struct sk_buff *head)
 255 {
 256         struct kcm_mux *mux = psock->mux;
 257         struct kcm_sock *kcm;
 258
 259         WARN_ON(psock->ready_rx_msg);
 260
 261         if (psock->rx_kcm)
 262                 return psock->rx_kcm;
 263
 264         spin_lock_bh(&mux->rx_lock);
 265
 266         if (psock->rx_kcm) {
 267                 spin_unlock_bh(&mux->rx_lock);
 268                 return psock->rx_kcm;
 269         }
 270
 271         kcm_update_rx_mux_stats(mux, psock);
 272
 273         if (list_empty(&mux->kcm_rx_waiters)) {
 274                 psock->ready_rx_msg = head;
 275                 strp_pause(&psock->strp);
 276                 list_add_tail(&psock->psock_ready_list,
 277                               &mux->psocks_ready);
 278                 spin_unlock_bh(&mux->rx_lock);
 279                 return NULL;
 280         }
 281
 282         kcm = list_first_entry(&mux->kcm_rx_waiters,
 283                                struct kcm_sock, wait_rx_list);
 284         list_del(&kcm->wait_rx_list);
 285         /* paired with lockless reads in kcm_rfree() */
 286         WRITE_ONCE(kcm->rx_wait, false);
 287
 288         psock->rx_kcm = kcm;
 289         /* paired with lockless reads in kcm_rfree() */
 290         WRITE_ONCE(kcm->rx_psock, psock);
 291
 292         spin_unlock_bh(&mux->rx_lock);
 293
 294         return kcm;
 295 }
 296
 297 static void kcm_done(struct kcm_sock *kcm);
 298
 299 static void kcm_done_work(struct work_struct *w)
 300 {
 301         kcm_done(container_of(w, struct kcm_sock, done_work));
 302 }
 303
 304 /* Lower sock held */
 305 static void unreserve_rx_kcm(struct kcm_psock *psock,
 306                              bool rcv_ready)
 307 {
 308         struct kcm_sock *kcm = psock->rx_kcm;
 309         struct kcm_mux *mux = psock->mux;
 310
 311         if (!kcm)
 312                 return;
 313
 314         spin_lock_bh(&mux->rx_lock);
 315
 316         psock->rx_kcm = NULL;
 317         /* paired with lockless reads in kcm_rfree() */
 318         WRITE_ONCE(kcm->rx_psock, NULL);
 319
 320         /* Commit kcm->rx_psock before sk_rmem_alloc_get to sync with
 321          * kcm_rfree
 322          */
 323         smp_mb();
 324
 325         if (unlikely(kcm->done)) {
 326                 spin_unlock_bh(&mux->rx_lock);
 327
 328                 /* Need to run kcm_done in a task since we need to qcquire
 329                  * callback locks which may already be held here.
 330                  */
 331                 INIT_WORK(&kcm->done_work, kcm_done_work);
 332                 schedule_work(&kcm->done_work);
 333                 return;
 334         }
 335
 336         if (unlikely(kcm->rx_disabled)) {
 337                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
 338         } else if (rcv_ready || unlikely(!sk_rmem_alloc_get(&kcm->sk))) {
 339                 /* Check for degenerative race with rx_wait that all
 340                  * data was dequeued (accounted for in kcm_rfree).
 341                  */
 342                 kcm_rcv_ready(kcm);
 343         }
 344         spin_unlock_bh(&mux->rx_lock);
 345 }
 346
 347 /* Lower sock lock held */
 348 static void psock_data_ready(struct sock *sk)
 349 {
 350         struct kcm_psock *psock;
 351
 352         read_lock_bh(&sk->sk_callback_lock);
 353
 354         psock = (struct kcm_psock *)sk->sk_user_data;
 355         if (likely(psock))
 356                 strp_data_ready(&psock->strp);
 357
 358         read_unlock_bh(&sk->sk_callback_lock);
 359 }
 360
 361 /* Called with lower sock held */
 362 static void kcm_rcv_strparser(struct strparser *strp, struct sk_buff *skb)
 363 {
 364         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 365         struct kcm_sock *kcm;
 366
 367 try_queue:
 368         kcm = reserve_rx_kcm(psock, skb);
 369         if (!kcm) {
 370                  /* Unable to reserve a KCM, message is held in psock and strp
 371                   * is paused.
 372                   */
 373                 return;
 374         }
 375
 376         if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 377                 /* Should mean socket buffer full */
 378                 unreserve_rx_kcm(psock, false);
 379                 goto try_queue;
 380         }
 381 }
 382
 383 static int kcm_parse_func_strparser(struct strparser *strp, struct sk_buff *skb)
 384 {
 385         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 386         struct bpf_prog *prog = psock->bpf_prog;
 387         int res;
 388
 389         res = bpf_prog_run_pin_on_cpu(prog, skb);
 390         return res;
 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(!skb_frag_size(frag))) {
 647                                 ret = -EINVAL;
 648                                 goto out;
 649                         }
 650
 651                         ret = kernel_sendpage(psock->sk->sk_socket,
 652                                               skb_frag_page(frag),
 653                                               skb_frag_off(frag) + frag_offset,
 654                                               skb_frag_size(frag) - 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                                  * synchronization 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 < skb_frag_size(frag)) {
 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)->flags |= SKBFL_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_noacc(skb, i, page, offset, size);
 843         skb_shinfo(skb)->flags |= SKBFL_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 int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
1089                        size_t len, int flags)
1090 {
1091         struct sock *sk = sock->sk;
1092         struct kcm_sock *kcm = kcm_sk(sk);
1093         int err = 0;
1094         struct strp_msg *stm;
1095         int copied = 0;
1096         struct sk_buff *skb;
1097
1098         skb = skb_recv_datagram(sk, flags, &err);
1099         if (!skb)
1100                 goto out;
1101
1102         /* Okay, have a message on the receive queue */
1103
1104         stm = strp_msg(skb);
1105
1106         if (len > stm->full_len)
1107                 len = stm->full_len;
1108
1109         err = skb_copy_datagram_msg(skb, stm->offset, msg, len);
1110         if (err < 0)
1111                 goto out;
1112
1113         copied = len;
1114         if (likely(!(flags & MSG_PEEK))) {
1115                 KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1116                 if (copied < stm->full_len) {
1117                         if (sock->type == SOCK_DGRAM) {
1118                                 /* Truncated message */
1119                                 msg->msg_flags |= MSG_TRUNC;
1120                                 goto msg_finished;
1121                         }
1122                         stm->offset += copied;
1123                         stm->full_len -= copied;
1124                 } else {
1125 msg_finished:
1126                         /* Finished with message */
1127                         msg->msg_flags |= MSG_EOR;
1128                         KCM_STATS_INCR(kcm->stats.rx_msgs);
1129                 }
1130         }
1131
1132 out:
1133         skb_free_datagram(sk, skb);
1134         return copied ? : err;
1135 }
1136
1137 static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1138                                struct pipe_inode_info *pipe, size_t len,
1139                                unsigned int flags)
1140 {
1141         struct sock *sk = sock->sk;
1142         struct kcm_sock *kcm = kcm_sk(sk);
1143         struct strp_msg *stm;
1144         int err = 0;
1145         ssize_t copied;
1146         struct sk_buff *skb;
1147
1148         /* Only support splice for SOCKSEQPACKET */
1149
1150         skb = skb_recv_datagram(sk, flags, &err);
1151         if (!skb)
1152                 goto err_out;
1153
1154         /* Okay, have a message on the receive queue */
1155
1156         stm = strp_msg(skb);
1157
1158         if (len > stm->full_len)
1159                 len = stm->full_len;
1160
1161         copied = skb_splice_bits(skb, sk, stm->offset, pipe, len, flags);
1162         if (copied < 0) {
1163                 err = copied;
1164                 goto err_out;
1165         }
1166
1167         KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1168
1169         stm->offset += copied;
1170         stm->full_len -= copied;
1171
1172         /* We have no way to return MSG_EOR. If all the bytes have been
1173          * read we still leave the message in the receive socket buffer.
1174          * A subsequent recvmsg needs to be done to return MSG_EOR and
1175          * finish reading the message.
1176          */
1177
1178         skb_free_datagram(sk, skb);
1179         return copied;
1180
1181 err_out:
1182         skb_free_datagram(sk, skb);
1183         return err;
1184 }
1185
1186 /* kcm sock lock held */
1187 static void kcm_recv_disable(struct kcm_sock *kcm)
1188 {
1189         struct kcm_mux *mux = kcm->mux;
1190
1191         if (kcm->rx_disabled)
1192                 return;
1193
1194         spin_lock_bh(&mux->rx_lock);
1195
1196         kcm->rx_disabled = 1;
1197
1198         /* If a psock is reserved we'll do cleanup in unreserve */
1199         if (!kcm->rx_psock) {
1200                 if (kcm->rx_wait) {
1201                         list_del(&kcm->wait_rx_list);
1202                         /* paired with lockless reads in kcm_rfree() */
1203                         WRITE_ONCE(kcm->rx_wait, false);
1204                 }
1205
1206                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1207         }
1208
1209         spin_unlock_bh(&mux->rx_lock);
1210 }
1211
1212 /* kcm sock lock held */
1213 static void kcm_recv_enable(struct kcm_sock *kcm)
1214 {
1215         struct kcm_mux *mux = kcm->mux;
1216
1217         if (!kcm->rx_disabled)
1218                 return;
1219
1220         spin_lock_bh(&mux->rx_lock);
1221
1222         kcm->rx_disabled = 0;
1223         kcm_rcv_ready(kcm);
1224
1225         spin_unlock_bh(&mux->rx_lock);
1226 }
1227
1228 static int kcm_setsockopt(struct socket *sock, int level, int optname,
1229                           sockptr_t optval, unsigned int optlen)
1230 {
1231         struct kcm_sock *kcm = kcm_sk(sock->sk);
1232         int val, valbool;
1233         int err = 0;
1234
1235         if (level != SOL_KCM)
1236                 return -ENOPROTOOPT;
1237
1238         if (optlen < sizeof(int))
1239                 return -EINVAL;
1240
1241         if (copy_from_sockptr(&val, optval, sizeof(int)))
1242                 return -EFAULT;
1243
1244         valbool = val ? 1 : 0;
1245
1246         switch (optname) {
1247         case KCM_RECV_DISABLE:
1248                 lock_sock(&kcm->sk);
1249                 if (valbool)
1250                         kcm_recv_disable(kcm);
1251                 else
1252                         kcm_recv_enable(kcm);
1253                 release_sock(&kcm->sk);
1254                 break;
1255         default:
1256                 err = -ENOPROTOOPT;
1257         }
1258
1259         return err;
1260 }
1261
1262 static int kcm_getsockopt(struct socket *sock, int level, int optname,
1263                           char __user *optval, int __user *optlen)
1264 {
1265         struct kcm_sock *kcm = kcm_sk(sock->sk);
1266         int val, len;
1267
1268         if (level != SOL_KCM)
1269                 return -ENOPROTOOPT;
1270
1271         if (get_user(len, optlen))
1272                 return -EFAULT;
1273
1274         len = min_t(unsigned int, len, sizeof(int));
1275         if (len < 0)
1276                 return -EINVAL;
1277
1278         switch (optname) {
1279         case KCM_RECV_DISABLE:
1280                 val = kcm->rx_disabled;
1281                 break;
1282         default:
1283                 return -ENOPROTOOPT;
1284         }
1285
1286         if (put_user(len, optlen))
1287                 return -EFAULT;
1288         if (copy_to_user(optval, &val, len))
1289                 return -EFAULT;
1290         return 0;
1291 }
1292
1293 static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1294 {
1295         struct kcm_sock *tkcm;
1296         struct list_head *head;
1297         int index = 0;
1298
1299         /* For SOCK_SEQPACKET sock type, datagram_poll checks the sk_state, so
1300          * we set sk_state, otherwise epoll_wait always returns right away with
1301          * EPOLLHUP
1302          */
1303         kcm->sk.sk_state = TCP_ESTABLISHED;
1304
1305         /* Add to mux's kcm sockets list */
1306         kcm->mux = mux;
1307         spin_lock_bh(&mux->lock);
1308
1309         head = &mux->kcm_socks;
1310         list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1311                 if (tkcm->index != index)
1312                         break;
1313                 head = &tkcm->kcm_sock_list;
1314                 index++;
1315         }
1316
1317         list_add(&kcm->kcm_sock_list, head);
1318         kcm->index = index;
1319
1320         mux->kcm_socks_cnt++;
1321         spin_unlock_bh(&mux->lock);
1322
1323         INIT_WORK(&kcm->tx_work, kcm_tx_work);
1324
1325         spin_lock_bh(&mux->rx_lock);
1326         kcm_rcv_ready(kcm);
1327         spin_unlock_bh(&mux->rx_lock);
1328 }
1329
1330 static int kcm_attach(struct socket *sock, struct socket *csock,
1331                       struct bpf_prog *prog)
1332 {
1333         struct kcm_sock *kcm = kcm_sk(sock->sk);
1334         struct kcm_mux *mux = kcm->mux;
1335         struct sock *csk;
1336         struct kcm_psock *psock = NULL, *tpsock;
1337         struct list_head *head;
1338         int index = 0;
1339         static const struct strp_callbacks cb = {
1340                 .rcv_msg = kcm_rcv_strparser,
1341                 .parse_msg = kcm_parse_func_strparser,
1342                 .read_sock_done = kcm_read_sock_done,
1343         };
1344         int err = 0;
1345
1346         csk = csock->sk;
1347         if (!csk)
1348                 return -EINVAL;
1349
1350         lock_sock(csk);
1351
1352         /* Only allow TCP sockets to be attached for now */
1353         if ((csk->sk_family != AF_INET && csk->sk_family != AF_INET6) ||
1354             csk->sk_protocol != IPPROTO_TCP) {
1355                 err = -EOPNOTSUPP;
1356                 goto out;
1357         }
1358
1359         /* Don't allow listeners or closed sockets */
1360         if (csk->sk_state == TCP_LISTEN || csk->sk_state == TCP_CLOSE) {
1361                 err = -EOPNOTSUPP;
1362                 goto out;
1363         }
1364
1365         psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1366         if (!psock) {
1367                 err = -ENOMEM;
1368                 goto out;
1369         }
1370
1371         psock->mux = mux;
1372         psock->sk = csk;
1373         psock->bpf_prog = prog;
1374
1375         write_lock_bh(&csk->sk_callback_lock);
1376
1377         /* Check if sk_user_data is already by KCM or someone else.
1378          * Must be done under lock to prevent race conditions.
1379          */
1380         if (csk->sk_user_data) {
1381                 write_unlock_bh(&csk->sk_callback_lock);
1382                 kmem_cache_free(kcm_psockp, psock);
1383                 err = -EALREADY;
1384                 goto out;
1385         }
1386
1387         err = strp_init(&psock->strp, csk, &cb);
1388         if (err) {
1389                 write_unlock_bh(&csk->sk_callback_lock);
1390                 kmem_cache_free(kcm_psockp, psock);
1391                 goto out;
1392         }
1393
1394         psock->save_data_ready = csk->sk_data_ready;
1395         psock->save_write_space = csk->sk_write_space;
1396         psock->save_state_change = csk->sk_state_change;
1397         csk->sk_user_data = psock;
1398         csk->sk_data_ready = psock_data_ready;
1399         csk->sk_write_space = psock_write_space;
1400         csk->sk_state_change = psock_state_change;
1401
1402         write_unlock_bh(&csk->sk_callback_lock);
1403
1404         sock_hold(csk);
1405
1406         /* Finished initialization, now add the psock to the MUX. */
1407         spin_lock_bh(&mux->lock);
1408         head = &mux->psocks;
1409         list_for_each_entry(tpsock, &mux->psocks, psock_list) {
1410                 if (tpsock->index != index)
1411                         break;
1412                 head = &tpsock->psock_list;
1413                 index++;
1414         }
1415
1416         list_add(&psock->psock_list, head);
1417         psock->index = index;
1418
1419         KCM_STATS_INCR(mux->stats.psock_attach);
1420         mux->psocks_cnt++;
1421         psock_now_avail(psock);
1422         spin_unlock_bh(&mux->lock);
1423
1424         /* Schedule RX work in case there are already bytes queued */
1425         strp_check_rcv(&psock->strp);
1426
1427 out:
1428         release_sock(csk);
1429
1430         return err;
1431 }
1432
1433 static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1434 {
1435         struct socket *csock;
1436         struct bpf_prog *prog;
1437         int err;
1438
1439         csock = sockfd_lookup(info->fd, &err);
1440         if (!csock)
1441                 return -ENOENT;
1442
1443         prog = bpf_prog_get_type(info->bpf_fd, BPF_PROG_TYPE_SOCKET_FILTER);
1444         if (IS_ERR(prog)) {
1445                 err = PTR_ERR(prog);
1446                 goto out;
1447         }
1448
1449         err = kcm_attach(sock, csock, prog);
1450         if (err) {
1451                 bpf_prog_put(prog);
1452                 goto out;
1453         }
1454
1455         /* Keep reference on file also */
1456
1457         return 0;
1458 out:
1459         sockfd_put(csock);
1460         return err;
1461 }
1462
1463 static void kcm_unattach(struct kcm_psock *psock)
1464 {
1465         struct sock *csk = psock->sk;
1466         struct kcm_mux *mux = psock->mux;
1467
1468         lock_sock(csk);
1469
1470         /* Stop getting callbacks from TCP socket. After this there should
1471          * be no way to reserve a kcm for this psock.
1472          */
1473         write_lock_bh(&csk->sk_callback_lock);
1474         csk->sk_user_data = NULL;
1475         csk->sk_data_ready = psock->save_data_ready;
1476         csk->sk_write_space = psock->save_write_space;
1477         csk->sk_state_change = psock->save_state_change;
1478         strp_stop(&psock->strp);
1479
1480         if (WARN_ON(psock->rx_kcm)) {
1481                 write_unlock_bh(&csk->sk_callback_lock);
1482                 release_sock(csk);
1483                 return;
1484         }
1485
1486         spin_lock_bh(&mux->rx_lock);
1487
1488         /* Stop receiver activities. After this point psock should not be
1489          * able to get onto ready list either through callbacks or work.
1490          */
1491         if (psock->ready_rx_msg) {
1492                 list_del(&psock->psock_ready_list);
1493                 kfree_skb(psock->ready_rx_msg);
1494                 psock->ready_rx_msg = NULL;
1495                 KCM_STATS_INCR(mux->stats.rx_ready_drops);
1496         }
1497
1498         spin_unlock_bh(&mux->rx_lock);
1499
1500         write_unlock_bh(&csk->sk_callback_lock);
1501
1502         /* Call strp_done without sock lock */
1503         release_sock(csk);
1504         strp_done(&psock->strp);
1505         lock_sock(csk);
1506
1507         bpf_prog_put(psock->bpf_prog);
1508
1509         spin_lock_bh(&mux->lock);
1510
1511         aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1512         save_strp_stats(&psock->strp, &mux->aggregate_strp_stats);
1513
1514         KCM_STATS_INCR(mux->stats.psock_unattach);
1515
1516         if (psock->tx_kcm) {
1517                 /* psock was reserved.  Just mark it finished and we will clean
1518                  * up in the kcm paths, we need kcm lock which can not be
1519                  * acquired here.
1520                  */
1521                 KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1522                 spin_unlock_bh(&mux->lock);
1523
1524                 /* We are unattaching a socket that is reserved. Abort the
1525                  * socket since we may be out of sync in sending on it. We need
1526                  * to do this without the mux lock.
1527                  */
1528                 kcm_abort_tx_psock(psock, EPIPE, false);
1529
1530                 spin_lock_bh(&mux->lock);
1531                 if (!psock->tx_kcm) {
1532                         /* psock now unreserved in window mux was unlocked */
1533                         goto no_reserved;
1534                 }
1535                 psock->done = 1;
1536
1537                 /* Commit done before queuing work to process it */
1538                 smp_mb();
1539
1540                 /* Queue tx work to make sure psock->done is handled */
1541                 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
1542                 spin_unlock_bh(&mux->lock);
1543         } else {
1544 no_reserved:
1545                 if (!psock->tx_stopped)
1546                         list_del(&psock->psock_avail_list);
1547                 list_del(&psock->psock_list);
1548                 mux->psocks_cnt--;
1549                 spin_unlock_bh(&mux->lock);
1550
1551                 sock_put(csk);
1552                 fput(csk->sk_socket->file);
1553                 kmem_cache_free(kcm_psockp, psock);
1554         }
1555
1556         release_sock(csk);
1557 }
1558
1559 static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1560 {
1561         struct kcm_sock *kcm = kcm_sk(sock->sk);
1562         struct kcm_mux *mux = kcm->mux;
1563         struct kcm_psock *psock;
1564         struct socket *csock;
1565         struct sock *csk;
1566         int err;
1567
1568         csock = sockfd_lookup(info->fd, &err);
1569         if (!csock)
1570                 return -ENOENT;
1571
1572         csk = csock->sk;
1573         if (!csk) {
1574                 err = -EINVAL;
1575                 goto out;
1576         }
1577
1578         err = -ENOENT;
1579
1580         spin_lock_bh(&mux->lock);
1581
1582         list_for_each_entry(psock, &mux->psocks, psock_list) {
1583                 if (psock->sk != csk)
1584                         continue;
1585
1586                 /* Found the matching psock */
1587
1588                 if (psock->unattaching || WARN_ON(psock->done)) {
1589                         err = -EALREADY;
1590                         break;
1591                 }
1592
1593                 psock->unattaching = 1;
1594
1595                 spin_unlock_bh(&mux->lock);
1596
1597                 /* Lower socket lock should already be held */
1598                 kcm_unattach(psock);
1599
1600                 err = 0;
1601                 goto out;
1602         }
1603
1604         spin_unlock_bh(&mux->lock);
1605
1606 out:
1607         sockfd_put(csock);
1608         return err;
1609 }
1610
1611 static struct proto kcm_proto = {
1612         .name   = "KCM",
1613         .owner  = THIS_MODULE,
1614         .obj_size = sizeof(struct kcm_sock),
1615 };
1616
1617 /* Clone a kcm socket. */
1618 static struct file *kcm_clone(struct socket *osock)
1619 {
1620         struct socket *newsock;
1621         struct sock *newsk;
1622
1623         newsock = sock_alloc();
1624         if (!newsock)
1625                 return ERR_PTR(-ENFILE);
1626
1627         newsock->type = osock->type;
1628         newsock->ops = osock->ops;
1629
1630         __module_get(newsock->ops->owner);
1631
1632         newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1633                          &kcm_proto, false);
1634         if (!newsk) {
1635                 sock_release(newsock);
1636                 return ERR_PTR(-ENOMEM);
1637         }
1638         sock_init_data(newsock, newsk);
1639         init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1640
1641         return sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1642 }
1643
1644 static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
1645 {
1646         int err;
1647
1648         switch (cmd) {
1649         case SIOCKCMATTACH: {
1650                 struct kcm_attach info;
1651
1652                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1653                         return -EFAULT;
1654
1655                 err = kcm_attach_ioctl(sock, &info);
1656
1657                 break;
1658         }
1659         case SIOCKCMUNATTACH: {
1660                 struct kcm_unattach info;
1661
1662                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1663                         return -EFAULT;
1664
1665                 err = kcm_unattach_ioctl(sock, &info);
1666
1667                 break;
1668         }
1669         case SIOCKCMCLONE: {
1670                 struct kcm_clone info;
1671                 struct file *file;
1672
1673                 info.fd = get_unused_fd_flags(0);
1674                 if (unlikely(info.fd < 0))
1675                         return info.fd;
1676
1677                 file = kcm_clone(sock);
1678                 if (IS_ERR(file)) {
1679                         put_unused_fd(info.fd);
1680                         return PTR_ERR(file);
1681                 }
1682                 if (copy_to_user((void __user *)arg, &info,
1683                                  sizeof(info))) {
1684                         put_unused_fd(info.fd);
1685                         fput(file);
1686                         return -EFAULT;
1687                 }
1688                 fd_install(info.fd, file);
1689                 err = 0;
1690                 break;
1691         }
1692         default:
1693                 err = -ENOIOCTLCMD;
1694                 break;
1695         }
1696
1697         return err;
1698 }
1699
1700 static void free_mux(struct rcu_head *rcu)
1701 {
1702         struct kcm_mux *mux = container_of(rcu,
1703             struct kcm_mux, rcu);
1704
1705         kmem_cache_free(kcm_muxp, mux);
1706 }
1707
1708 static void release_mux(struct kcm_mux *mux)
1709 {
1710         struct kcm_net *knet = mux->knet;
1711         struct kcm_psock *psock, *tmp_psock;
1712
1713         /* Release psocks */
1714         list_for_each_entry_safe(psock, tmp_psock,
1715                                  &mux->psocks, psock_list) {
1716                 if (!WARN_ON(psock->unattaching))
1717                         kcm_unattach(psock);
1718         }
1719
1720         if (WARN_ON(mux->psocks_cnt))
1721                 return;
1722
1723         __skb_queue_purge(&mux->rx_hold_queue);
1724
1725         mutex_lock(&knet->mutex);
1726         aggregate_mux_stats(&mux->stats, &knet->aggregate_mux_stats);
1727         aggregate_psock_stats(&mux->aggregate_psock_stats,
1728                               &knet->aggregate_psock_stats);
1729         aggregate_strp_stats(&mux->aggregate_strp_stats,
1730                              &knet->aggregate_strp_stats);
1731         list_del_rcu(&mux->kcm_mux_list);
1732         knet->count--;
1733         mutex_unlock(&knet->mutex);
1734
1735         call_rcu(&mux->rcu, free_mux);
1736 }
1737
1738 static void kcm_done(struct kcm_sock *kcm)
1739 {
1740         struct kcm_mux *mux = kcm->mux;
1741         struct sock *sk = &kcm->sk;
1742         int socks_cnt;
1743
1744         spin_lock_bh(&mux->rx_lock);
1745         if (kcm->rx_psock) {
1746                 /* Cleanup in unreserve_rx_kcm */
1747                 WARN_ON(kcm->done);
1748                 kcm->rx_disabled = 1;
1749                 kcm->done = 1;
1750                 spin_unlock_bh(&mux->rx_lock);
1751                 return;
1752         }
1753
1754         if (kcm->rx_wait) {
1755                 list_del(&kcm->wait_rx_list);
1756                 /* paired with lockless reads in kcm_rfree() */
1757                 WRITE_ONCE(kcm->rx_wait, false);
1758         }
1759         /* Move any pending receive messages to other kcm sockets */
1760         requeue_rx_msgs(mux, &sk->sk_receive_queue);
1761
1762         spin_unlock_bh(&mux->rx_lock);
1763
1764         if (WARN_ON(sk_rmem_alloc_get(sk)))
1765                 return;
1766
1767         /* Detach from MUX */
1768         spin_lock_bh(&mux->lock);
1769
1770         list_del(&kcm->kcm_sock_list);
1771         mux->kcm_socks_cnt--;
1772         socks_cnt = mux->kcm_socks_cnt;
1773
1774         spin_unlock_bh(&mux->lock);
1775
1776         if (!socks_cnt) {
1777                 /* We are done with the mux now. */
1778                 release_mux(mux);
1779         }
1780
1781         WARN_ON(kcm->rx_wait);
1782
1783         sock_put(&kcm->sk);
1784 }
1785
1786 /* Called by kcm_release to close a KCM socket.
1787  * If this is the last KCM socket on the MUX, destroy the MUX.
1788  */
1789 static int kcm_release(struct socket *sock)
1790 {
1791         struct sock *sk = sock->sk;
1792         struct kcm_sock *kcm;
1793         struct kcm_mux *mux;
1794         struct kcm_psock *psock;
1795
1796         if (!sk)
1797                 return 0;
1798
1799         kcm = kcm_sk(sk);
1800         mux = kcm->mux;
1801
1802         lock_sock(sk);
1803         sock_orphan(sk);
1804         kfree_skb(kcm->seq_skb);
1805
1806         /* Purge queue under lock to avoid race condition with tx_work trying
1807          * to act when queue is nonempty. If tx_work runs after this point
1808          * it will just return.
1809          */
1810         __skb_queue_purge(&sk->sk_write_queue);
1811
1812         /* Set tx_stopped. This is checked when psock is bound to a kcm and we
1813          * get a writespace callback. This prevents further work being queued
1814          * from the callback (unbinding the psock occurs after canceling work.
1815          */
1816         kcm->tx_stopped = 1;
1817
1818         release_sock(sk);
1819
1820         spin_lock_bh(&mux->lock);
1821         if (kcm->tx_wait) {
1822                 /* Take of tx_wait list, after this point there should be no way
1823                  * that a psock will be assigned to this kcm.
1824                  */
1825                 list_del(&kcm->wait_psock_list);
1826                 kcm->tx_wait = false;
1827         }
1828         spin_unlock_bh(&mux->lock);
1829
1830         /* Cancel work. After this point there should be no outside references
1831          * to the kcm socket.
1832          */
1833         cancel_work_sync(&kcm->tx_work);
1834
1835         lock_sock(sk);
1836         psock = kcm->tx_psock;
1837         if (psock) {
1838                 /* A psock was reserved, so we need to kill it since it
1839                  * may already have some bytes queued from a message. We
1840                  * need to do this after removing kcm from tx_wait list.
1841                  */
1842                 kcm_abort_tx_psock(psock, EPIPE, false);
1843                 unreserve_psock(kcm);
1844         }
1845         release_sock(sk);
1846
1847         WARN_ON(kcm->tx_wait);
1848         WARN_ON(kcm->tx_psock);
1849
1850         sock->sk = NULL;
1851
1852         kcm_done(kcm);
1853
1854         return 0;
1855 }
1856
1857 static const struct proto_ops kcm_dgram_ops = {
1858         .family =       PF_KCM,
1859         .owner =        THIS_MODULE,
1860         .release =      kcm_release,
1861         .bind =         sock_no_bind,
1862         .connect =      sock_no_connect,
1863         .socketpair =   sock_no_socketpair,
1864         .accept =       sock_no_accept,
1865         .getname =      sock_no_getname,
1866         .poll =         datagram_poll,
1867         .ioctl =        kcm_ioctl,
1868         .listen =       sock_no_listen,
1869         .shutdown =     sock_no_shutdown,
1870         .setsockopt =   kcm_setsockopt,
1871         .getsockopt =   kcm_getsockopt,
1872         .sendmsg =      kcm_sendmsg,
1873         .recvmsg =      kcm_recvmsg,
1874         .mmap =         sock_no_mmap,
1875         .sendpage =     kcm_sendpage,
1876 };
1877
1878 static const struct proto_ops kcm_seqpacket_ops = {
1879         .family =       PF_KCM,
1880         .owner =        THIS_MODULE,
1881         .release =      kcm_release,
1882         .bind =         sock_no_bind,
1883         .connect =      sock_no_connect,
1884         .socketpair =   sock_no_socketpair,
1885         .accept =       sock_no_accept,
1886         .getname =      sock_no_getname,
1887         .poll =         datagram_poll,
1888         .ioctl =        kcm_ioctl,
1889         .listen =       sock_no_listen,
1890         .shutdown =     sock_no_shutdown,
1891         .setsockopt =   kcm_setsockopt,
1892         .getsockopt =   kcm_getsockopt,
1893         .sendmsg =      kcm_sendmsg,
1894         .recvmsg =      kcm_recvmsg,
1895         .mmap =         sock_no_mmap,
1896         .sendpage =     kcm_sendpage,
1897         .splice_read =  kcm_splice_read,
1898 };
1899
1900 /* Create proto operation for kcm sockets */
1901 static int kcm_create(struct net *net, struct socket *sock,
1902                       int protocol, int kern)
1903 {
1904         struct kcm_net *knet = net_generic(net, kcm_net_id);
1905         struct sock *sk;
1906         struct kcm_mux *mux;
1907
1908         switch (sock->type) {
1909         case SOCK_DGRAM:
1910                 sock->ops = &kcm_dgram_ops;
1911                 break;
1912         case SOCK_SEQPACKET:
1913                 sock->ops = &kcm_seqpacket_ops;
1914                 break;
1915         default:
1916                 return -ESOCKTNOSUPPORT;
1917         }
1918
1919         if (protocol != KCMPROTO_CONNECTED)
1920                 return -EPROTONOSUPPORT;
1921
1922         sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
1923         if (!sk)
1924                 return -ENOMEM;
1925
1926         /* Allocate a kcm mux, shared between KCM sockets */
1927         mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
1928         if (!mux) {
1929                 sk_free(sk);
1930                 return -ENOMEM;
1931         }
1932
1933         spin_lock_init(&mux->lock);
1934         spin_lock_init(&mux->rx_lock);
1935         INIT_LIST_HEAD(&mux->kcm_socks);
1936         INIT_LIST_HEAD(&mux->kcm_rx_waiters);
1937         INIT_LIST_HEAD(&mux->kcm_tx_waiters);
1938
1939         INIT_LIST_HEAD(&mux->psocks);
1940         INIT_LIST_HEAD(&mux->psocks_ready);
1941         INIT_LIST_HEAD(&mux->psocks_avail);
1942
1943         mux->knet = knet;
1944
1945         /* Add new MUX to list */
1946         mutex_lock(&knet->mutex);
1947         list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
1948         knet->count++;
1949         mutex_unlock(&knet->mutex);
1950
1951         skb_queue_head_init(&mux->rx_hold_queue);
1952
1953         /* Init KCM socket */
1954         sock_init_data(sock, sk);
1955         init_kcm_sock(kcm_sk(sk), mux);
1956
1957         return 0;
1958 }
1959
1960 static const struct net_proto_family kcm_family_ops = {
1961         .family = PF_KCM,
1962         .create = kcm_create,
1963         .owner  = THIS_MODULE,
1964 };
1965
1966 static __net_init int kcm_init_net(struct net *net)
1967 {
1968         struct kcm_net *knet = net_generic(net, kcm_net_id);
1969
1970         INIT_LIST_HEAD_RCU(&knet->mux_list);
1971         mutex_init(&knet->mutex);
1972
1973         return 0;
1974 }
1975
1976 static __net_exit void kcm_exit_net(struct net *net)
1977 {
1978         struct kcm_net *knet = net_generic(net, kcm_net_id);
1979
1980         /* All KCM sockets should be closed at this point, which should mean
1981          * that all multiplexors and psocks have been destroyed.
1982          */
1983         WARN_ON(!list_empty(&knet->mux_list));
1984 }
1985
1986 static struct pernet_operations kcm_net_ops = {
1987         .init = kcm_init_net,
1988         .exit = kcm_exit_net,
1989         .id   = &kcm_net_id,
1990         .size = sizeof(struct kcm_net),
1991 };
1992
1993 static int __init kcm_init(void)
1994 {
1995         int err = -ENOMEM;
1996
1997         kcm_muxp = kmem_cache_create("kcm_mux_cache",
1998                                      sizeof(struct kcm_mux), 0,
1999                                      SLAB_HWCACHE_ALIGN, NULL);
2000         if (!kcm_muxp)
2001                 goto fail;
2002
2003         kcm_psockp = kmem_cache_create("kcm_psock_cache",
2004                                        sizeof(struct kcm_psock), 0,
2005                                         SLAB_HWCACHE_ALIGN, NULL);
2006         if (!kcm_psockp)
2007                 goto fail;
2008
2009         kcm_wq = create_singlethread_workqueue("kkcmd");
2010         if (!kcm_wq)
2011                 goto fail;
2012
2013         err = proto_register(&kcm_proto, 1);
2014         if (err)
2015                 goto fail;
2016
2017         err = register_pernet_device(&kcm_net_ops);
2018         if (err)
2019                 goto net_ops_fail;
2020
2021         err = sock_register(&kcm_family_ops);
2022         if (err)
2023                 goto sock_register_fail;
2024
2025         err = kcm_proc_init();
2026         if (err)
2027                 goto proc_init_fail;
2028
2029         return 0;
2030
2031 proc_init_fail:
2032         sock_unregister(PF_KCM);
2033
2034 sock_register_fail:
2035         unregister_pernet_device(&kcm_net_ops);
2036
2037 net_ops_fail:
2038         proto_unregister(&kcm_proto);
2039
2040 fail:
2041         kmem_cache_destroy(kcm_muxp);
2042         kmem_cache_destroy(kcm_psockp);
2043
2044         if (kcm_wq)
2045                 destroy_workqueue(kcm_wq);
2046
2047         return err;
2048 }
2049
2050 static void __exit kcm_exit(void)
2051 {
2052         kcm_proc_exit();
2053         sock_unregister(PF_KCM);
2054         unregister_pernet_device(&kcm_net_ops);
2055         proto_unregister(&kcm_proto);
2056         destroy_workqueue(kcm_wq);
2057
2058         kmem_cache_destroy(kcm_muxp);
2059         kmem_cache_destroy(kcm_psockp);
2060 }
2061
2062 module_init(kcm_init);
2063 module_exit(kcm_exit);
2064
2065 MODULE_LICENSE("GPL");
2066 MODULE_ALIAS_NETPROTO(PF_KCM);