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/in.h>
  13 #include <linux/kernel.h>
  14 #include <linux/module.h>
  15 #include <linux/net.h>
  16 #include <linux/netdevice.h>
  17 #include <linux/poll.h>
  18 #include <linux/rculist.h>
  19 #include <linux/skbuff.h>
  20 #include <linux/socket.h>
  21 #include <linux/uaccess.h>
  22 #include <linux/workqueue.h>
  23 #include <linux/syscalls.h>
  24 #include <linux/sched/signal.h>
  25
  26 #include <net/kcm.h>
  27 #include <net/netns/generic.h>
  28 #include <net/sock.h>
  29 #include <uapi/linux/kcm.h>
  30
  31 unsigned int kcm_net_id;
  32
  33 static struct kmem_cache *kcm_psockp __read_mostly;
  34 static struct kmem_cache *kcm_muxp __read_mostly;
  35 static struct workqueue_struct *kcm_wq;
  36
  37 static inline struct kcm_sock *kcm_sk(const struct sock *sk)
  38 {
  39         return (struct kcm_sock *)sk;
  40 }
  41
  42 static inline struct kcm_tx_msg *kcm_tx_msg(struct sk_buff *skb)
  43 {
  44         return (struct kcm_tx_msg *)skb->cb;
  45 }
  46
  47 static void report_csk_error(struct sock *csk, int err)
  48 {
  49         csk->sk_err = EPIPE;
  50         csk->sk_error_report(csk);
  51 }
  52
  53 static void kcm_abort_tx_psock(struct kcm_psock *psock, int err,
  54                                bool wakeup_kcm)
  55 {
  56         struct sock *csk = psock->sk;
  57         struct kcm_mux *mux = psock->mux;
  58
  59         /* Unrecoverable error in transmit */
  60
  61         spin_lock_bh(&mux->lock);
  62
  63         if (psock->tx_stopped) {
  64                 spin_unlock_bh(&mux->lock);
  65                 return;
  66         }
  67
  68         psock->tx_stopped = 1;
  69         KCM_STATS_INCR(psock->stats.tx_aborts);
  70
  71         if (!psock->tx_kcm) {
  72                 /* Take off psocks_avail list */
  73                 list_del(&psock->psock_avail_list);
  74         } else if (wakeup_kcm) {
  75                 /* In this case psock is being aborted while outside of
  76                  * write_msgs and psock is reserved. Schedule tx_work
  77                  * to handle the failure there. Need to commit tx_stopped
  78                  * before queuing work.
  79                  */
  80                 smp_mb();
  81
  82                 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
  83         }
  84
  85         spin_unlock_bh(&mux->lock);
  86
  87         /* Report error on lower socket */
  88         report_csk_error(csk, err);
  89 }
  90
  91 /* RX mux lock held. */
  92 static void kcm_update_rx_mux_stats(struct kcm_mux *mux,
  93                                     struct kcm_psock *psock)
  94 {
  95         STRP_STATS_ADD(mux->stats.rx_bytes,
  96                        psock->strp.stats.bytes -
  97                        psock->saved_rx_bytes);
  98         mux->stats.rx_msgs +=
  99                 psock->strp.stats.msgs - psock->saved_rx_msgs;
 100         psock->saved_rx_msgs = psock->strp.stats.msgs;
 101         psock->saved_rx_bytes = psock->strp.stats.bytes;
 102 }
 103
 104 static void kcm_update_tx_mux_stats(struct kcm_mux *mux,
 105                                     struct kcm_psock *psock)
 106 {
 107         KCM_STATS_ADD(mux->stats.tx_bytes,
 108                       psock->stats.tx_bytes - psock->saved_tx_bytes);
 109         mux->stats.tx_msgs +=
 110                 psock->stats.tx_msgs - psock->saved_tx_msgs;
 111         psock->saved_tx_msgs = psock->stats.tx_msgs;
 112         psock->saved_tx_bytes = psock->stats.tx_bytes;
 113 }
 114
 115 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb);
 116
 117 /* KCM is ready to receive messages on its queue-- either the KCM is new or
 118  * has become unblocked after being blocked on full socket buffer. Queue any
 119  * pending ready messages on a psock. RX mux lock held.
 120  */
 121 static void kcm_rcv_ready(struct kcm_sock *kcm)
 122 {
 123         struct kcm_mux *mux = kcm->mux;
 124         struct kcm_psock *psock;
 125         struct sk_buff *skb;
 126
 127         if (unlikely(kcm->rx_wait || kcm->rx_psock || kcm->rx_disabled))
 128                 return;
 129
 130         while (unlikely((skb = __skb_dequeue(&mux->rx_hold_queue)))) {
 131                 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 132                         /* Assuming buffer limit has been reached */
 133                         skb_queue_head(&mux->rx_hold_queue, skb);
 134                         WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
 135                         return;
 136                 }
 137         }
 138
 139         while (!list_empty(&mux->psocks_ready)) {
 140                 psock = list_first_entry(&mux->psocks_ready, struct kcm_psock,
 141                                          psock_ready_list);
 142
 143                 if (kcm_queue_rcv_skb(&kcm->sk, psock->ready_rx_msg)) {
 144                         /* Assuming buffer limit has been reached */
 145                         WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
 146                         return;
 147                 }
 148
 149                 /* Consumed the ready message on the psock. Schedule rx_work to
 150                  * get more messages.
 151                  */
 152                 list_del(&psock->psock_ready_list);
 153                 psock->ready_rx_msg = NULL;
 154                 /* Commit clearing of ready_rx_msg for queuing work */
 155                 smp_mb();
 156
 157                 strp_unpause(&psock->strp);
 158                 strp_check_rcv(&psock->strp);
 159         }
 160
 161         /* Buffer limit is okay now, add to ready list */
 162         list_add_tail(&kcm->wait_rx_list,
 163                       &kcm->mux->kcm_rx_waiters);
 164         /* paired with lockless reads in kcm_rfree() */
 165         WRITE_ONCE(kcm->rx_wait, true);
 166 }
 167
 168 static void kcm_rfree(struct sk_buff *skb)
 169 {
 170         struct sock *sk = skb->sk;
 171         struct kcm_sock *kcm = kcm_sk(sk);
 172         struct kcm_mux *mux = kcm->mux;
 173         unsigned int len = skb->truesize;
 174
 175         sk_mem_uncharge(sk, len);
 176         atomic_sub(len, &sk->sk_rmem_alloc);
 177
 178         /* For reading rx_wait and rx_psock without holding lock */
 179         smp_mb__after_atomic();
 180
 181         if (!READ_ONCE(kcm->rx_wait) && !READ_ONCE(kcm->rx_psock) &&
 182             sk_rmem_alloc_get(sk) < sk->sk_rcvlowat) {
 183                 spin_lock_bh(&mux->rx_lock);
 184                 kcm_rcv_ready(kcm);
 185                 spin_unlock_bh(&mux->rx_lock);
 186         }
 187 }
 188
 189 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb)
 190 {
 191         struct sk_buff_head *list = &sk->sk_receive_queue;
 192
 193         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
 194                 return -ENOMEM;
 195
 196         if (!sk_rmem_schedule(sk, skb, skb->truesize))
 197                 return -ENOBUFS;
 198
 199         skb->dev = NULL;
 200
 201         skb_orphan(skb);
 202         skb->sk = sk;
 203         skb->destructor = kcm_rfree;
 204         atomic_add(skb->truesize, &sk->sk_rmem_alloc);
 205         sk_mem_charge(sk, skb->truesize);
 206
 207         skb_queue_tail(list, skb);
 208
 209         if (!sock_flag(sk, SOCK_DEAD))
 210                 sk->sk_data_ready(sk);
 211
 212         return 0;
 213 }
 214
 215 /* Requeue received messages for a kcm socket to other kcm sockets. This is
 216  * called with a kcm socket is receive disabled.
 217  * RX mux lock held.
 218  */
 219 static void requeue_rx_msgs(struct kcm_mux *mux, struct sk_buff_head *head)
 220 {
 221         struct sk_buff *skb;
 222         struct kcm_sock *kcm;
 223
 224         while ((skb = skb_dequeue(head))) {
 225                 /* Reset destructor to avoid calling kcm_rcv_ready */
 226                 skb->destructor = sock_rfree;
 227                 skb_orphan(skb);
 228 try_again:
 229                 if (list_empty(&mux->kcm_rx_waiters)) {
 230                         skb_queue_tail(&mux->rx_hold_queue, skb);
 231                         continue;
 232                 }
 233
 234                 kcm = list_first_entry(&mux->kcm_rx_waiters,
 235                                        struct kcm_sock, wait_rx_list);
 236
 237                 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 238                         /* Should mean socket buffer full */
 239                         list_del(&kcm->wait_rx_list);
 240                         /* paired with lockless reads in kcm_rfree() */
 241                         WRITE_ONCE(kcm->rx_wait, false);
 242
 243                         /* Commit rx_wait to read in kcm_free */
 244                         smp_wmb();
 245
 246                         goto try_again;
 247                 }
 248         }
 249 }
 250
 251 /* Lower sock lock held */
 252 static struct kcm_sock *reserve_rx_kcm(struct kcm_psock *psock,
 253                                        struct sk_buff *head)
 254 {
 255         struct kcm_mux *mux = psock->mux;
 256         struct kcm_sock *kcm;
 257
 258         WARN_ON(psock->ready_rx_msg);
 259
 260         if (psock->rx_kcm)
 261                 return psock->rx_kcm;
 262
 263         spin_lock_bh(&mux->rx_lock);
 264
 265         if (psock->rx_kcm) {
 266                 spin_unlock_bh(&mux->rx_lock);
 267                 return psock->rx_kcm;
 268         }
 269
 270         kcm_update_rx_mux_stats(mux, psock);
 271
 272         if (list_empty(&mux->kcm_rx_waiters)) {
 273                 psock->ready_rx_msg = head;
 274                 strp_pause(&psock->strp);
 275                 list_add_tail(&psock->psock_ready_list,
 276                               &mux->psocks_ready);
 277                 spin_unlock_bh(&mux->rx_lock);
 278                 return NULL;
 279         }
 280
 281         kcm = list_first_entry(&mux->kcm_rx_waiters,
 282                                struct kcm_sock, wait_rx_list);
 283         list_del(&kcm->wait_rx_list);
 284         /* paired with lockless reads in kcm_rfree() */
 285         WRITE_ONCE(kcm->rx_wait, false);
 286
 287         psock->rx_kcm = kcm;
 288         /* paired with lockless reads in kcm_rfree() */
 289         WRITE_ONCE(kcm->rx_psock, psock);
 290
 291         spin_unlock_bh(&mux->rx_lock);
 292
 293         return kcm;
 294 }
 295
 296 static void kcm_done(struct kcm_sock *kcm);
 297
 298 static void kcm_done_work(struct work_struct *w)
 299 {
 300         kcm_done(container_of(w, struct kcm_sock, done_work));
 301 }
 302
 303 /* Lower sock held */
 304 static void unreserve_rx_kcm(struct kcm_psock *psock,
 305                              bool rcv_ready)
 306 {
 307         struct kcm_sock *kcm = psock->rx_kcm;
 308         struct kcm_mux *mux = psock->mux;
 309
 310         if (!kcm)
 311                 return;
 312
 313         spin_lock_bh(&mux->rx_lock);
 314
 315         psock->rx_kcm = NULL;
 316         /* paired with lockless reads in kcm_rfree() */
 317         WRITE_ONCE(kcm->rx_psock, NULL);
 318
 319         /* Commit kcm->rx_psock before sk_rmem_alloc_get to sync with
 320          * kcm_rfree
 321          */
 322         smp_mb();
 323
 324         if (unlikely(kcm->done)) {
 325                 spin_unlock_bh(&mux->rx_lock);
 326
 327                 /* Need to run kcm_done in a task since we need to qcquire
 328                  * callback locks which may already be held here.
 329                  */
 330                 INIT_WORK(&kcm->done_work, kcm_done_work);
 331                 schedule_work(&kcm->done_work);
 332                 return;
 333         }
 334
 335         if (unlikely(kcm->rx_disabled)) {
 336                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
 337         } else if (rcv_ready || unlikely(!sk_rmem_alloc_get(&kcm->sk))) {
 338                 /* Check for degenerative race with rx_wait that all
 339                  * data was dequeued (accounted for in kcm_rfree).
 340                  */
 341                 kcm_rcv_ready(kcm);
 342         }
 343         spin_unlock_bh(&mux->rx_lock);
 344 }
 345
 346 /* Lower sock lock held */
 347 static void psock_data_ready(struct sock *sk)
 348 {
 349         struct kcm_psock *psock;
 350
 351         read_lock_bh(&sk->sk_callback_lock);
 352
 353         psock = (struct kcm_psock *)sk->sk_user_data;
 354         if (likely(psock))
 355                 strp_data_ready(&psock->strp);
 356
 357         read_unlock_bh(&sk->sk_callback_lock);
 358 }
 359
 360 /* Called with lower sock held */
 361 static void kcm_rcv_strparser(struct strparser *strp, struct sk_buff *skb)
 362 {
 363         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 364         struct kcm_sock *kcm;
 365
 366 try_queue:
 367         kcm = reserve_rx_kcm(psock, skb);
 368         if (!kcm) {
 369                  /* Unable to reserve a KCM, message is held in psock and strp
 370                   * is paused.
 371                   */
 372                 return;
 373         }
 374
 375         if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 376                 /* Should mean socket buffer full */
 377                 unreserve_rx_kcm(psock, false);
 378                 goto try_queue;
 379         }
 380 }
 381
 382 static int kcm_parse_func_strparser(struct strparser *strp, struct sk_buff *skb)
 383 {
 384         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 385         struct bpf_prog *prog = psock->bpf_prog;
 386         int res;
 387
 388         res = bpf_prog_run_pin_on_cpu(prog, skb);
 389         return res;
 390 }
 391
 392 static int kcm_read_sock_done(struct strparser *strp, int err)
 393 {
 394         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 395
 396         unreserve_rx_kcm(psock, true);
 397
 398         return err;
 399 }
 400
 401 static void psock_state_change(struct sock *sk)
 402 {
 403         /* TCP only does a EPOLLIN for a half close. Do a EPOLLHUP here
 404          * since application will normally not poll with EPOLLIN
 405          * on the TCP sockets.
 406          */
 407
 408         report_csk_error(sk, EPIPE);
 409 }
 410
 411 static void psock_write_space(struct sock *sk)
 412 {
 413         struct kcm_psock *psock;
 414         struct kcm_mux *mux;
 415         struct kcm_sock *kcm;
 416
 417         read_lock_bh(&sk->sk_callback_lock);
 418
 419         psock = (struct kcm_psock *)sk->sk_user_data;
 420         if (unlikely(!psock))
 421                 goto out;
 422         mux = psock->mux;
 423
 424         spin_lock_bh(&mux->lock);
 425
 426         /* Check if the socket is reserved so someone is waiting for sending. */
 427         kcm = psock->tx_kcm;
 428         if (kcm && !unlikely(kcm->tx_stopped))
 429                 queue_work(kcm_wq, &kcm->tx_work);
 430
 431         spin_unlock_bh(&mux->lock);
 432 out:
 433         read_unlock_bh(&sk->sk_callback_lock);
 434 }
 435
 436 static void unreserve_psock(struct kcm_sock *kcm);
 437
 438 /* kcm sock is locked. */
 439 static struct kcm_psock *reserve_psock(struct kcm_sock *kcm)
 440 {
 441         struct kcm_mux *mux = kcm->mux;
 442         struct kcm_psock *psock;
 443
 444         psock = kcm->tx_psock;
 445
 446         smp_rmb(); /* Must read tx_psock before tx_wait */
 447
 448         if (psock) {
 449                 WARN_ON(kcm->tx_wait);
 450                 if (unlikely(psock->tx_stopped))
 451                         unreserve_psock(kcm);
 452                 else
 453                         return kcm->tx_psock;
 454         }
 455
 456         spin_lock_bh(&mux->lock);
 457
 458         /* Check again under lock to see if psock was reserved for this
 459          * psock via psock_unreserve.
 460          */
 461         psock = kcm->tx_psock;
 462         if (unlikely(psock)) {
 463                 WARN_ON(kcm->tx_wait);
 464                 spin_unlock_bh(&mux->lock);
 465                 return kcm->tx_psock;
 466         }
 467
 468         if (!list_empty(&mux->psocks_avail)) {
 469                 psock = list_first_entry(&mux->psocks_avail,
 470                                          struct kcm_psock,
 471                                          psock_avail_list);
 472                 list_del(&psock->psock_avail_list);
 473                 if (kcm->tx_wait) {
 474                         list_del(&kcm->wait_psock_list);
 475                         kcm->tx_wait = false;
 476                 }
 477                 kcm->tx_psock = psock;
 478                 psock->tx_kcm = kcm;
 479                 KCM_STATS_INCR(psock->stats.reserved);
 480         } else if (!kcm->tx_wait) {
 481                 list_add_tail(&kcm->wait_psock_list,
 482                               &mux->kcm_tx_waiters);
 483                 kcm->tx_wait = true;
 484         }
 485
 486         spin_unlock_bh(&mux->lock);
 487
 488         return psock;
 489 }
 490
 491 /* mux lock held */
 492 static void psock_now_avail(struct kcm_psock *psock)
 493 {
 494         struct kcm_mux *mux = psock->mux;
 495         struct kcm_sock *kcm;
 496
 497         if (list_empty(&mux->kcm_tx_waiters)) {
 498                 list_add_tail(&psock->psock_avail_list,
 499                               &mux->psocks_avail);
 500         } else {
 501                 kcm = list_first_entry(&mux->kcm_tx_waiters,
 502                                        struct kcm_sock,
 503                                        wait_psock_list);
 504                 list_del(&kcm->wait_psock_list);
 505                 kcm->tx_wait = false;
 506                 psock->tx_kcm = kcm;
 507
 508                 /* Commit before changing tx_psock since that is read in
 509                  * reserve_psock before queuing work.
 510                  */
 511                 smp_mb();
 512
 513                 kcm->tx_psock = psock;
 514                 KCM_STATS_INCR(psock->stats.reserved);
 515                 queue_work(kcm_wq, &kcm->tx_work);
 516         }
 517 }
 518
 519 /* kcm sock is locked. */
 520 static void unreserve_psock(struct kcm_sock *kcm)
 521 {
 522         struct kcm_psock *psock;
 523         struct kcm_mux *mux = kcm->mux;
 524
 525         spin_lock_bh(&mux->lock);
 526
 527         psock = kcm->tx_psock;
 528
 529         if (WARN_ON(!psock)) {
 530                 spin_unlock_bh(&mux->lock);
 531                 return;
 532         }
 533
 534         smp_rmb(); /* Read tx_psock before tx_wait */
 535
 536         kcm_update_tx_mux_stats(mux, psock);
 537
 538         WARN_ON(kcm->tx_wait);
 539
 540         kcm->tx_psock = NULL;
 541         psock->tx_kcm = NULL;
 542         KCM_STATS_INCR(psock->stats.unreserved);
 543
 544         if (unlikely(psock->tx_stopped)) {
 545                 if (psock->done) {
 546                         /* Deferred free */
 547                         list_del(&psock->psock_list);
 548                         mux->psocks_cnt--;
 549                         sock_put(psock->sk);
 550                         fput(psock->sk->sk_socket->file);
 551                         kmem_cache_free(kcm_psockp, psock);
 552                 }
 553
 554                 /* Don't put back on available list */
 555
 556                 spin_unlock_bh(&mux->lock);
 557
 558                 return;
 559         }
 560
 561         psock_now_avail(psock);
 562
 563         spin_unlock_bh(&mux->lock);
 564 }
 565
 566 static void kcm_report_tx_retry(struct kcm_sock *kcm)
 567 {
 568         struct kcm_mux *mux = kcm->mux;
 569
 570         spin_lock_bh(&mux->lock);
 571         KCM_STATS_INCR(mux->stats.tx_retries);
 572         spin_unlock_bh(&mux->lock);
 573 }
 574
 575 /* Write any messages ready on the kcm socket.  Called with kcm sock lock
 576  * held.  Return bytes actually sent or error.
 577  */
 578 static int kcm_write_msgs(struct kcm_sock *kcm)
 579 {
 580         struct sock *sk = &kcm->sk;
 581         struct kcm_psock *psock;
 582         struct sk_buff *skb, *head;
 583         struct kcm_tx_msg *txm;
 584         unsigned short fragidx, frag_offset;
 585         unsigned int sent, total_sent = 0;
 586         int ret = 0;
 587
 588         kcm->tx_wait_more = false;
 589         psock = kcm->tx_psock;
 590         if (unlikely(psock && psock->tx_stopped)) {
 591                 /* A reserved psock was aborted asynchronously. Unreserve
 592                  * it and we'll retry the message.
 593                  */
 594                 unreserve_psock(kcm);
 595                 kcm_report_tx_retry(kcm);
 596                 if (skb_queue_empty(&sk->sk_write_queue))
 597                         return 0;
 598
 599                 kcm_tx_msg(skb_peek(&sk->sk_write_queue))->sent = 0;
 600
 601         } else if (skb_queue_empty(&sk->sk_write_queue)) {
 602                 return 0;
 603         }
 604
 605         head = skb_peek(&sk->sk_write_queue);
 606         txm = kcm_tx_msg(head);
 607
 608         if (txm->sent) {
 609                 /* Send of first skbuff in queue already in progress */
 610                 if (WARN_ON(!psock)) {
 611                         ret = -EINVAL;
 612                         goto out;
 613                 }
 614                 sent = txm->sent;
 615                 frag_offset = txm->frag_offset;
 616                 fragidx = txm->fragidx;
 617                 skb = txm->frag_skb;
 618
 619                 goto do_frag;
 620         }
 621
 622 try_again:
 623         psock = reserve_psock(kcm);
 624         if (!psock)
 625                 goto out;
 626
 627         do {
 628                 skb = head;
 629                 txm = kcm_tx_msg(head);
 630                 sent = 0;
 631
 632 do_frag_list:
 633                 if (WARN_ON(!skb_shinfo(skb)->nr_frags)) {
 634                         ret = -EINVAL;
 635                         goto out;
 636                 }
 637
 638                 for (fragidx = 0; fragidx < skb_shinfo(skb)->nr_frags;
 639                      fragidx++) {
 640                         skb_frag_t *frag;
 641
 642                         frag_offset = 0;
 643 do_frag:
 644                         frag = &skb_shinfo(skb)->frags[fragidx];
 645                         if (WARN_ON(!skb_frag_size(frag))) {
 646                                 ret = -EINVAL;
 647                                 goto out;
 648                         }
 649
 650                         ret = kernel_sendpage(psock->sk->sk_socket,
 651                                               skb_frag_page(frag),
 652                                               skb_frag_off(frag) + frag_offset,
 653                                               skb_frag_size(frag) - frag_offset,
 654                                               MSG_DONTWAIT);
 655                         if (ret <= 0) {
 656                                 if (ret == -EAGAIN) {
 657                                         /* Save state to try again when there's
 658                                          * write space on the socket
 659                                          */
 660                                         txm->sent = sent;
 661                                         txm->frag_offset = frag_offset;
 662                                         txm->fragidx = fragidx;
 663                                         txm->frag_skb = skb;
 664
 665                                         ret = 0;
 666                                         goto out;
 667                                 }
 668
 669                                 /* Hard failure in sending message, abort this
 670                                  * psock since it has lost framing
 671                                  * synchonization and retry sending the
 672                                  * message from the beginning.
 673                                  */
 674                                 kcm_abort_tx_psock(psock, ret ? -ret : EPIPE,
 675                                                    true);
 676                                 unreserve_psock(kcm);
 677
 678                                 txm->sent = 0;
 679                                 kcm_report_tx_retry(kcm);
 680                                 ret = 0;
 681
 682                                 goto try_again;
 683                         }
 684
 685                         sent += ret;
 686                         frag_offset += ret;
 687                         KCM_STATS_ADD(psock->stats.tx_bytes, ret);
 688                         if (frag_offset < skb_frag_size(frag)) {
 689                                 /* Not finished with this frag */
 690                                 goto do_frag;
 691                         }
 692                 }
 693
 694                 if (skb == head) {
 695                         if (skb_has_frag_list(skb)) {
 696                                 skb = skb_shinfo(skb)->frag_list;
 697                                 goto do_frag_list;
 698                         }
 699                 } else if (skb->next) {
 700                         skb = skb->next;
 701                         goto do_frag_list;
 702                 }
 703
 704                 /* Successfully sent the whole packet, account for it. */
 705                 skb_dequeue(&sk->sk_write_queue);
 706                 kfree_skb(head);
 707                 sk->sk_wmem_queued -= sent;
 708                 total_sent += sent;
 709                 KCM_STATS_INCR(psock->stats.tx_msgs);
 710         } while ((head = skb_peek(&sk->sk_write_queue)));
 711 out:
 712         if (!head) {
 713                 /* Done with all queued messages. */
 714                 WARN_ON(!skb_queue_empty(&sk->sk_write_queue));
 715                 unreserve_psock(kcm);
 716         }
 717
 718         /* Check if write space is available */
 719         sk->sk_write_space(sk);
 720
 721         return total_sent ? : ret;
 722 }
 723
 724 static void kcm_tx_work(struct work_struct *w)
 725 {
 726         struct kcm_sock *kcm = container_of(w, struct kcm_sock, tx_work);
 727         struct sock *sk = &kcm->sk;
 728         int err;
 729
 730         lock_sock(sk);
 731
 732         /* Primarily for SOCK_DGRAM sockets, also handle asynchronous tx
 733          * aborts
 734          */
 735         err = kcm_write_msgs(kcm);
 736         if (err < 0) {
 737                 /* Hard failure in write, report error on KCM socket */
 738                 pr_warn("KCM: Hard failure on kcm_write_msgs %d\n", err);
 739                 report_csk_error(&kcm->sk, -err);
 740                 goto out;
 741         }
 742
 743         /* Primarily for SOCK_SEQPACKET sockets */
 744         if (likely(sk->sk_socket) &&
 745             test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
 746                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 747                 sk->sk_write_space(sk);
 748         }
 749
 750 out:
 751         release_sock(sk);
 752 }
 753
 754 static void kcm_push(struct kcm_sock *kcm)
 755 {
 756         if (kcm->tx_wait_more)
 757                 kcm_write_msgs(kcm);
 758 }
 759
 760 static ssize_t kcm_sendpage(struct socket *sock, struct page *page,
 761                             int offset, size_t size, int flags)
 762
 763 {
 764         struct sock *sk = sock->sk;
 765         struct kcm_sock *kcm = kcm_sk(sk);
 766         struct sk_buff *skb = NULL, *head = NULL;
 767         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 768         bool eor;
 769         int err = 0;
 770         int i;
 771
 772         if (flags & MSG_SENDPAGE_NOTLAST)
 773                 flags |= MSG_MORE;
 774
 775         /* No MSG_EOR from splice, only look at MSG_MORE */
 776         eor = !(flags & MSG_MORE);
 777
 778         lock_sock(sk);
 779
 780         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 781
 782         err = -EPIPE;
 783         if (sk->sk_err)
 784                 goto out_error;
 785
 786         if (kcm->seq_skb) {
 787                 /* Previously opened message */
 788                 head = kcm->seq_skb;
 789                 skb = kcm_tx_msg(head)->last_skb;
 790                 i = skb_shinfo(skb)->nr_frags;
 791
 792                 if (skb_can_coalesce(skb, i, page, offset)) {
 793                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], size);
 794                         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
 795                         goto coalesced;
 796                 }
 797
 798                 if (i >= MAX_SKB_FRAGS) {
 799                         struct sk_buff *tskb;
 800
 801                         tskb = alloc_skb(0, sk->sk_allocation);
 802                         while (!tskb) {
 803                                 kcm_push(kcm);
 804                                 err = sk_stream_wait_memory(sk, &timeo);
 805                                 if (err)
 806                                         goto out_error;
 807                         }
 808
 809                         if (head == skb)
 810                                 skb_shinfo(head)->frag_list = tskb;
 811                         else
 812                                 skb->next = tskb;
 813
 814                         skb = tskb;
 815                         skb->ip_summed = CHECKSUM_UNNECESSARY;
 816                         i = 0;
 817                 }
 818         } else {
 819                 /* Call the sk_stream functions to manage the sndbuf mem. */
 820                 if (!sk_stream_memory_free(sk)) {
 821                         kcm_push(kcm);
 822                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 823                         err = sk_stream_wait_memory(sk, &timeo);
 824                         if (err)
 825                                 goto out_error;
 826                 }
 827
 828                 head = alloc_skb(0, sk->sk_allocation);
 829                 while (!head) {
 830                         kcm_push(kcm);
 831                         err = sk_stream_wait_memory(sk, &timeo);
 832                         if (err)
 833                                 goto out_error;
 834                 }
 835
 836                 skb = head;
 837                 i = 0;
 838         }
 839
 840         get_page(page);
 841         skb_fill_page_desc(skb, i, page, offset, size);
 842         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
 843
 844 coalesced:
 845         skb->len += size;
 846         skb->data_len += size;
 847         skb->truesize += size;
 848         sk->sk_wmem_queued += size;
 849         sk_mem_charge(sk, size);
 850
 851         if (head != skb) {
 852                 head->len += size;
 853                 head->data_len += size;
 854                 head->truesize += size;
 855         }
 856
 857         if (eor) {
 858                 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
 859
 860                 /* Message complete, queue it on send buffer */
 861                 __skb_queue_tail(&sk->sk_write_queue, head);
 862                 kcm->seq_skb = NULL;
 863                 KCM_STATS_INCR(kcm->stats.tx_msgs);
 864
 865                 if (flags & MSG_BATCH) {
 866                         kcm->tx_wait_more = true;
 867                 } else if (kcm->tx_wait_more || not_busy) {
 868                         err = kcm_write_msgs(kcm);
 869                         if (err < 0) {
 870                                 /* We got a hard error in write_msgs but have
 871                                  * already queued this message. Report an error
 872                                  * in the socket, but don't affect return value
 873                                  * from sendmsg
 874                                  */
 875                                 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
 876                                 report_csk_error(&kcm->sk, -err);
 877                         }
 878                 }
 879         } else {
 880                 /* Message not complete, save state */
 881                 kcm->seq_skb = head;
 882                 kcm_tx_msg(head)->last_skb = skb;
 883         }
 884
 885         KCM_STATS_ADD(kcm->stats.tx_bytes, size);
 886
 887         release_sock(sk);
 888         return size;
 889
 890 out_error:
 891         kcm_push(kcm);
 892
 893         err = sk_stream_error(sk, flags, err);
 894
 895         /* make sure we wake any epoll edge trigger waiter */
 896         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
 897                 sk->sk_write_space(sk);
 898
 899         release_sock(sk);
 900         return err;
 901 }
 902
 903 static int kcm_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
 904 {
 905         struct sock *sk = sock->sk;
 906         struct kcm_sock *kcm = kcm_sk(sk);
 907         struct sk_buff *skb = NULL, *head = NULL;
 908         size_t copy, copied = 0;
 909         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 910         int eor = (sock->type == SOCK_DGRAM) ?
 911                   !(msg->msg_flags & MSG_MORE) : !!(msg->msg_flags & MSG_EOR);
 912         int err = -EPIPE;
 913
 914         lock_sock(sk);
 915
 916         /* Per tcp_sendmsg this should be in poll */
 917         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 918
 919         if (sk->sk_err)
 920                 goto out_error;
 921
 922         if (kcm->seq_skb) {
 923                 /* Previously opened message */
 924                 head = kcm->seq_skb;
 925                 skb = kcm_tx_msg(head)->last_skb;
 926                 goto start;
 927         }
 928
 929         /* Call the sk_stream functions to manage the sndbuf mem. */
 930         if (!sk_stream_memory_free(sk)) {
 931                 kcm_push(kcm);
 932                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 933                 err = sk_stream_wait_memory(sk, &timeo);
 934                 if (err)
 935                         goto out_error;
 936         }
 937
 938         if (msg_data_left(msg)) {
 939                 /* New message, alloc head skb */
 940                 head = alloc_skb(0, sk->sk_allocation);
 941                 while (!head) {
 942                         kcm_push(kcm);
 943                         err = sk_stream_wait_memory(sk, &timeo);
 944                         if (err)
 945                                 goto out_error;
 946
 947                         head = alloc_skb(0, sk->sk_allocation);
 948                 }
 949
 950                 skb = head;
 951
 952                 /* Set ip_summed to CHECKSUM_UNNECESSARY to avoid calling
 953                  * csum_and_copy_from_iter from skb_do_copy_data_nocache.
 954                  */
 955                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 956         }
 957
 958 start:
 959         while (msg_data_left(msg)) {
 960                 bool merge = true;
 961                 int i = skb_shinfo(skb)->nr_frags;
 962                 struct page_frag *pfrag = sk_page_frag(sk);
 963
 964                 if (!sk_page_frag_refill(sk, pfrag))
 965                         goto wait_for_memory;
 966
 967                 if (!skb_can_coalesce(skb, i, pfrag->page,
 968                                       pfrag->offset)) {
 969                         if (i == MAX_SKB_FRAGS) {
 970                                 struct sk_buff *tskb;
 971
 972                                 tskb = alloc_skb(0, sk->sk_allocation);
 973                                 if (!tskb)
 974                                         goto wait_for_memory;
 975
 976                                 if (head == skb)
 977                                         skb_shinfo(head)->frag_list = tskb;
 978                                 else
 979                                         skb->next = tskb;
 980
 981                                 skb = tskb;
 982                                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 983                                 continue;
 984                         }
 985                         merge = false;
 986                 }
 987
 988                 copy = min_t(int, msg_data_left(msg),
 989                              pfrag->size - pfrag->offset);
 990
 991                 if (!sk_wmem_schedule(sk, copy))
 992                         goto wait_for_memory;
 993
 994                 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
 995                                                pfrag->page,
 996                                                pfrag->offset,
 997                                                copy);
 998                 if (err)
 999                         goto out_error;
1000
1001                 /* Update the skb. */
1002                 if (merge) {
1003                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1004                 } else {
1005                         skb_fill_page_desc(skb, i, pfrag->page,
1006                                            pfrag->offset, copy);
1007                         get_page(pfrag->page);
1008                 }
1009
1010                 pfrag->offset += copy;
1011                 copied += copy;
1012                 if (head != skb) {
1013                         head->len += copy;
1014                         head->data_len += copy;
1015                 }
1016
1017                 continue;
1018
1019 wait_for_memory:
1020                 kcm_push(kcm);
1021                 err = sk_stream_wait_memory(sk, &timeo);
1022                 if (err)
1023                         goto out_error;
1024         }
1025
1026         if (eor) {
1027                 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
1028
1029                 if (head) {
1030                         /* Message complete, queue it on send buffer */
1031                         __skb_queue_tail(&sk->sk_write_queue, head);
1032                         kcm->seq_skb = NULL;
1033                         KCM_STATS_INCR(kcm->stats.tx_msgs);
1034                 }
1035
1036                 if (msg->msg_flags & MSG_BATCH) {
1037                         kcm->tx_wait_more = true;
1038                 } else if (kcm->tx_wait_more || not_busy) {
1039                         err = kcm_write_msgs(kcm);
1040                         if (err < 0) {
1041                                 /* We got a hard error in write_msgs but have
1042                                  * already queued this message. Report an error
1043                                  * in the socket, but don't affect return value
1044                                  * from sendmsg
1045                                  */
1046                                 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
1047                                 report_csk_error(&kcm->sk, -err);
1048                         }
1049                 }
1050         } else {
1051                 /* Message not complete, save state */
1052 partial_message:
1053                 if (head) {
1054                         kcm->seq_skb = head;
1055                         kcm_tx_msg(head)->last_skb = skb;
1056                 }
1057         }
1058
1059         KCM_STATS_ADD(kcm->stats.tx_bytes, copied);
1060
1061         release_sock(sk);
1062         return copied;
1063
1064 out_error:
1065         kcm_push(kcm);
1066
1067         if (sock->type == SOCK_SEQPACKET) {
1068                 /* Wrote some bytes before encountering an
1069                  * error, return partial success.
1070                  */
1071                 if (copied)
1072                         goto partial_message;
1073                 if (head != kcm->seq_skb)
1074                         kfree_skb(head);
1075         } else {
1076                 kfree_skb(head);
1077                 kcm->seq_skb = NULL;
1078         }
1079
1080         err = sk_stream_error(sk, msg->msg_flags, err);
1081
1082         /* make sure we wake any epoll edge trigger waiter */
1083         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1084                 sk->sk_write_space(sk);
1085
1086         release_sock(sk);
1087         return err;
1088 }
1089
1090 static int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
1091                        size_t len, int flags)
1092 {
1093         int noblock = flags & MSG_DONTWAIT;
1094         struct sock *sk = sock->sk;
1095         struct kcm_sock *kcm = kcm_sk(sk);
1096         int err = 0;
1097         struct strp_msg *stm;
1098         int copied = 0;
1099         struct sk_buff *skb;
1100
1101         skb = skb_recv_datagram(sk, flags, noblock, &err);
1102         if (!skb)
1103                 goto out;
1104
1105         /* Okay, have a message on the receive queue */
1106
1107         stm = strp_msg(skb);
1108
1109         if (len > stm->full_len)
1110                 len = stm->full_len;
1111
1112         err = skb_copy_datagram_msg(skb, stm->offset, msg, len);
1113         if (err < 0)
1114                 goto out;
1115
1116         copied = len;
1117         if (likely(!(flags & MSG_PEEK))) {
1118                 KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1119                 if (copied < stm->full_len) {
1120                         if (sock->type == SOCK_DGRAM) {
1121                                 /* Truncated message */
1122                                 msg->msg_flags |= MSG_TRUNC;
1123                                 goto msg_finished;
1124                         }
1125                         stm->offset += copied;
1126                         stm->full_len -= copied;
1127                 } else {
1128 msg_finished:
1129                         /* Finished with message */
1130                         msg->msg_flags |= MSG_EOR;
1131                         KCM_STATS_INCR(kcm->stats.rx_msgs);
1132                 }
1133         }
1134
1135 out:
1136         skb_free_datagram(sk, skb);
1137         return copied ? : err;
1138 }
1139
1140 static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1141                                struct pipe_inode_info *pipe, size_t len,
1142                                unsigned int flags)
1143 {
1144         int noblock = flags & MSG_DONTWAIT;
1145         struct sock *sk = sock->sk;
1146         struct kcm_sock *kcm = kcm_sk(sk);
1147         struct strp_msg *stm;
1148         int err = 0;
1149         ssize_t copied;
1150         struct sk_buff *skb;
1151
1152         /* Only support splice for SOCKSEQPACKET */
1153
1154         skb = skb_recv_datagram(sk, flags, noblock, &err);
1155         if (!skb)
1156                 goto err_out;
1157
1158         /* Okay, have a message on the receive queue */
1159
1160         stm = strp_msg(skb);
1161
1162         if (len > stm->full_len)
1163                 len = stm->full_len;
1164
1165         copied = skb_splice_bits(skb, sk, stm->offset, pipe, len, flags);
1166         if (copied < 0) {
1167                 err = copied;
1168                 goto err_out;
1169         }
1170
1171         KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1172
1173         stm->offset += copied;
1174         stm->full_len -= copied;
1175
1176         /* We have no way to return MSG_EOR. If all the bytes have been
1177          * read we still leave the message in the receive socket buffer.
1178          * A subsequent recvmsg needs to be done to return MSG_EOR and
1179          * finish reading the message.
1180          */
1181
1182         skb_free_datagram(sk, skb);
1183         return copied;
1184
1185 err_out:
1186         skb_free_datagram(sk, skb);
1187         return err;
1188 }
1189
1190 /* kcm sock lock held */
1191 static void kcm_recv_disable(struct kcm_sock *kcm)
1192 {
1193         struct kcm_mux *mux = kcm->mux;
1194
1195         if (kcm->rx_disabled)
1196                 return;
1197
1198         spin_lock_bh(&mux->rx_lock);
1199
1200         kcm->rx_disabled = 1;
1201
1202         /* If a psock is reserved we'll do cleanup in unreserve */
1203         if (!kcm->rx_psock) {
1204                 if (kcm->rx_wait) {
1205                         list_del(&kcm->wait_rx_list);
1206                         /* paired with lockless reads in kcm_rfree() */
1207                         WRITE_ONCE(kcm->rx_wait, false);
1208                 }
1209
1210                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1211         }
1212
1213         spin_unlock_bh(&mux->rx_lock);
1214 }
1215
1216 /* kcm sock lock held */
1217 static void kcm_recv_enable(struct kcm_sock *kcm)
1218 {
1219         struct kcm_mux *mux = kcm->mux;
1220
1221         if (!kcm->rx_disabled)
1222                 return;
1223
1224         spin_lock_bh(&mux->rx_lock);
1225
1226         kcm->rx_disabled = 0;
1227         kcm_rcv_ready(kcm);
1228
1229         spin_unlock_bh(&mux->rx_lock);
1230 }
1231
1232 static int kcm_setsockopt(struct socket *sock, int level, int optname,
1233                           sockptr_t optval, unsigned int optlen)
1234 {
1235         struct kcm_sock *kcm = kcm_sk(sock->sk);
1236         int val, valbool;
1237         int err = 0;
1238
1239         if (level != SOL_KCM)
1240                 return -ENOPROTOOPT;
1241
1242         if (optlen < sizeof(int))
1243                 return -EINVAL;
1244
1245         if (copy_from_sockptr(&val, optval, sizeof(int)))
1246                 return -EFAULT;
1247
1248         valbool = val ? 1 : 0;
1249
1250         switch (optname) {
1251         case KCM_RECV_DISABLE:
1252                 lock_sock(&kcm->sk);
1253                 if (valbool)
1254                         kcm_recv_disable(kcm);
1255                 else
1256                         kcm_recv_enable(kcm);
1257                 release_sock(&kcm->sk);
1258                 break;
1259         default:
1260                 err = -ENOPROTOOPT;
1261         }
1262
1263         return err;
1264 }
1265
1266 static int kcm_getsockopt(struct socket *sock, int level, int optname,
1267                           char __user *optval, int __user *optlen)
1268 {
1269         struct kcm_sock *kcm = kcm_sk(sock->sk);
1270         int val, len;
1271
1272         if (level != SOL_KCM)
1273                 return -ENOPROTOOPT;
1274
1275         if (get_user(len, optlen))
1276                 return -EFAULT;
1277
1278         if (len < 0)
1279                 return -EINVAL;
1280
1281         len = min_t(unsigned int, len, sizeof(int));
1282
1283         switch (optname) {
1284         case KCM_RECV_DISABLE:
1285                 val = kcm->rx_disabled;
1286                 break;
1287         default:
1288                 return -ENOPROTOOPT;
1289         }
1290
1291         if (put_user(len, optlen))
1292                 return -EFAULT;
1293         if (copy_to_user(optval, &val, len))
1294                 return -EFAULT;
1295         return 0;
1296 }
1297
1298 static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1299 {
1300         struct kcm_sock *tkcm;
1301         struct list_head *head;
1302         int index = 0;
1303
1304         /* For SOCK_SEQPACKET sock type, datagram_poll checks the sk_state, so
1305          * we set sk_state, otherwise epoll_wait always returns right away with
1306          * EPOLLHUP
1307          */
1308         kcm->sk.sk_state = TCP_ESTABLISHED;
1309
1310         /* Add to mux's kcm sockets list */
1311         kcm->mux = mux;
1312         spin_lock_bh(&mux->lock);
1313
1314         head = &mux->kcm_socks;
1315         list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1316                 if (tkcm->index != index)
1317                         break;
1318                 head = &tkcm->kcm_sock_list;
1319                 index++;
1320         }
1321
1322         list_add(&kcm->kcm_sock_list, head);
1323         kcm->index = index;
1324
1325         mux->kcm_socks_cnt++;
1326         spin_unlock_bh(&mux->lock);
1327
1328         INIT_WORK(&kcm->tx_work, kcm_tx_work);
1329
1330         spin_lock_bh(&mux->rx_lock);
1331         kcm_rcv_ready(kcm);
1332         spin_unlock_bh(&mux->rx_lock);
1333 }
1334
1335 static int kcm_attach(struct socket *sock, struct socket *csock,
1336                       struct bpf_prog *prog)
1337 {
1338         struct kcm_sock *kcm = kcm_sk(sock->sk);
1339         struct kcm_mux *mux = kcm->mux;
1340         struct sock *csk;
1341         struct kcm_psock *psock = NULL, *tpsock;
1342         struct list_head *head;
1343         int index = 0;
1344         static const struct strp_callbacks cb = {
1345                 .rcv_msg = kcm_rcv_strparser,
1346                 .parse_msg = kcm_parse_func_strparser,
1347                 .read_sock_done = kcm_read_sock_done,
1348         };
1349         int err = 0;
1350
1351         csk = csock->sk;
1352         if (!csk)
1353                 return -EINVAL;
1354
1355         lock_sock(csk);
1356
1357         /* Only allow TCP sockets to be attached for now */
1358         if ((csk->sk_family != AF_INET && csk->sk_family != AF_INET6) ||
1359             csk->sk_protocol != IPPROTO_TCP) {
1360                 err = -EOPNOTSUPP;
1361                 goto out;
1362         }
1363
1364         /* Don't allow listeners or closed sockets */
1365         if (csk->sk_state == TCP_LISTEN || csk->sk_state == TCP_CLOSE) {
1366                 err = -EOPNOTSUPP;
1367                 goto out;
1368         }
1369
1370         psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1371         if (!psock) {
1372                 err = -ENOMEM;
1373                 goto out;
1374         }
1375
1376         psock->mux = mux;
1377         psock->sk = csk;
1378         psock->bpf_prog = prog;
1379
1380         write_lock_bh(&csk->sk_callback_lock);
1381
1382         /* Check if sk_user_data is aready by KCM or someone else.
1383          * Must be done under lock to prevent race conditions.
1384          */
1385         if (csk->sk_user_data) {
1386                 write_unlock_bh(&csk->sk_callback_lock);
1387                 kmem_cache_free(kcm_psockp, psock);
1388                 err = -EALREADY;
1389                 goto out;
1390         }
1391
1392         err = strp_init(&psock->strp, csk, &cb);
1393         if (err) {
1394                 write_unlock_bh(&csk->sk_callback_lock);
1395                 kmem_cache_free(kcm_psockp, psock);
1396                 goto out;
1397         }
1398
1399         psock->save_data_ready = csk->sk_data_ready;
1400         psock->save_write_space = csk->sk_write_space;
1401         psock->save_state_change = csk->sk_state_change;
1402         csk->sk_user_data = psock;
1403         csk->sk_data_ready = psock_data_ready;
1404         csk->sk_write_space = psock_write_space;
1405         csk->sk_state_change = psock_state_change;
1406
1407         write_unlock_bh(&csk->sk_callback_lock);
1408
1409         sock_hold(csk);
1410
1411         /* Finished initialization, now add the psock to the MUX. */
1412         spin_lock_bh(&mux->lock);
1413         head = &mux->psocks;
1414         list_for_each_entry(tpsock, &mux->psocks, psock_list) {
1415                 if (tpsock->index != index)
1416                         break;
1417                 head = &tpsock->psock_list;
1418                 index++;
1419         }
1420
1421         list_add(&psock->psock_list, head);
1422         psock->index = index;
1423
1424         KCM_STATS_INCR(mux->stats.psock_attach);
1425         mux->psocks_cnt++;
1426         psock_now_avail(psock);
1427         spin_unlock_bh(&mux->lock);
1428
1429         /* Schedule RX work in case there are already bytes queued */
1430         strp_check_rcv(&psock->strp);
1431
1432 out:
1433         release_sock(csk);
1434
1435         return err;
1436 }
1437
1438 static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1439 {
1440         struct socket *csock;
1441         struct bpf_prog *prog;
1442         int err;
1443
1444         csock = sockfd_lookup(info->fd, &err);
1445         if (!csock)
1446                 return -ENOENT;
1447
1448         prog = bpf_prog_get_type(info->bpf_fd, BPF_PROG_TYPE_SOCKET_FILTER);
1449         if (IS_ERR(prog)) {
1450                 err = PTR_ERR(prog);
1451                 goto out;
1452         }
1453
1454         err = kcm_attach(sock, csock, prog);
1455         if (err) {
1456                 bpf_prog_put(prog);
1457                 goto out;
1458         }
1459
1460         /* Keep reference on file also */
1461
1462         return 0;
1463 out:
1464         fput(csock->file);
1465         return err;
1466 }
1467
1468 static void kcm_unattach(struct kcm_psock *psock)
1469 {
1470         struct sock *csk = psock->sk;
1471         struct kcm_mux *mux = psock->mux;
1472
1473         lock_sock(csk);
1474
1475         /* Stop getting callbacks from TCP socket. After this there should
1476          * be no way to reserve a kcm for this psock.
1477          */
1478         write_lock_bh(&csk->sk_callback_lock);
1479         csk->sk_user_data = NULL;
1480         csk->sk_data_ready = psock->save_data_ready;
1481         csk->sk_write_space = psock->save_write_space;
1482         csk->sk_state_change = psock->save_state_change;
1483         strp_stop(&psock->strp);
1484
1485         if (WARN_ON(psock->rx_kcm)) {
1486                 write_unlock_bh(&csk->sk_callback_lock);
1487                 release_sock(csk);
1488                 return;
1489         }
1490
1491         spin_lock_bh(&mux->rx_lock);
1492
1493         /* Stop receiver activities. After this point psock should not be
1494          * able to get onto ready list either through callbacks or work.
1495          */
1496         if (psock->ready_rx_msg) {
1497                 list_del(&psock->psock_ready_list);
1498                 kfree_skb(psock->ready_rx_msg);
1499                 psock->ready_rx_msg = NULL;
1500                 KCM_STATS_INCR(mux->stats.rx_ready_drops);
1501         }
1502
1503         spin_unlock_bh(&mux->rx_lock);
1504
1505         write_unlock_bh(&csk->sk_callback_lock);
1506
1507         /* Call strp_done without sock lock */
1508         release_sock(csk);
1509         strp_done(&psock->strp);
1510         lock_sock(csk);
1511
1512         bpf_prog_put(psock->bpf_prog);
1513
1514         spin_lock_bh(&mux->lock);
1515
1516         aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1517         save_strp_stats(&psock->strp, &mux->aggregate_strp_stats);
1518
1519         KCM_STATS_INCR(mux->stats.psock_unattach);
1520
1521         if (psock->tx_kcm) {
1522                 /* psock was reserved.  Just mark it finished and we will clean
1523                  * up in the kcm paths, we need kcm lock which can not be
1524                  * acquired here.
1525                  */
1526                 KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1527                 spin_unlock_bh(&mux->lock);
1528
1529                 /* We are unattaching a socket that is reserved. Abort the
1530                  * socket since we may be out of sync in sending on it. We need
1531                  * to do this without the mux lock.
1532                  */
1533                 kcm_abort_tx_psock(psock, EPIPE, false);
1534
1535                 spin_lock_bh(&mux->lock);
1536                 if (!psock->tx_kcm) {
1537                         /* psock now unreserved in window mux was unlocked */
1538                         goto no_reserved;
1539                 }
1540                 psock->done = 1;
1541
1542                 /* Commit done before queuing work to process it */
1543                 smp_mb();
1544
1545                 /* Queue tx work to make sure psock->done is handled */
1546                 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
1547                 spin_unlock_bh(&mux->lock);
1548         } else {
1549 no_reserved:
1550                 if (!psock->tx_stopped)
1551                         list_del(&psock->psock_avail_list);
1552                 list_del(&psock->psock_list);
1553                 mux->psocks_cnt--;
1554                 spin_unlock_bh(&mux->lock);
1555
1556                 sock_put(csk);
1557                 fput(csk->sk_socket->file);
1558                 kmem_cache_free(kcm_psockp, psock);
1559         }
1560
1561         release_sock(csk);
1562 }
1563
1564 static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1565 {
1566         struct kcm_sock *kcm = kcm_sk(sock->sk);
1567         struct kcm_mux *mux = kcm->mux;
1568         struct kcm_psock *psock;
1569         struct socket *csock;
1570         struct sock *csk;
1571         int err;
1572
1573         csock = sockfd_lookup(info->fd, &err);
1574         if (!csock)
1575                 return -ENOENT;
1576
1577         csk = csock->sk;
1578         if (!csk) {
1579                 err = -EINVAL;
1580                 goto out;
1581         }
1582
1583         err = -ENOENT;
1584
1585         spin_lock_bh(&mux->lock);
1586
1587         list_for_each_entry(psock, &mux->psocks, psock_list) {
1588                 if (psock->sk != csk)
1589                         continue;
1590
1591                 /* Found the matching psock */
1592
1593                 if (psock->unattaching || WARN_ON(psock->done)) {
1594                         err = -EALREADY;
1595                         break;
1596                 }
1597
1598                 psock->unattaching = 1;
1599
1600                 spin_unlock_bh(&mux->lock);
1601
1602                 /* Lower socket lock should already be held */
1603                 kcm_unattach(psock);
1604
1605                 err = 0;
1606                 goto out;
1607         }
1608
1609         spin_unlock_bh(&mux->lock);
1610
1611 out:
1612         fput(csock->file);
1613         return err;
1614 }
1615
1616 static struct proto kcm_proto = {
1617         .name   = "KCM",
1618         .owner  = THIS_MODULE,
1619         .obj_size = sizeof(struct kcm_sock),
1620 };
1621
1622 /* Clone a kcm socket. */
1623 static struct file *kcm_clone(struct socket *osock)
1624 {
1625         struct socket *newsock;
1626         struct sock *newsk;
1627
1628         newsock = sock_alloc();
1629         if (!newsock)
1630                 return ERR_PTR(-ENFILE);
1631
1632         newsock->type = osock->type;
1633         newsock->ops = osock->ops;
1634
1635         __module_get(newsock->ops->owner);
1636
1637         newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1638                          &kcm_proto, false);
1639         if (!newsk) {
1640                 sock_release(newsock);
1641                 return ERR_PTR(-ENOMEM);
1642         }
1643         sock_init_data(newsock, newsk);
1644         init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1645
1646         return sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1647 }
1648
1649 static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
1650 {
1651         int err;
1652
1653         switch (cmd) {
1654         case SIOCKCMATTACH: {
1655                 struct kcm_attach info;
1656
1657                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1658                         return -EFAULT;
1659
1660                 err = kcm_attach_ioctl(sock, &info);
1661
1662                 break;
1663         }
1664         case SIOCKCMUNATTACH: {
1665                 struct kcm_unattach info;
1666
1667                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1668                         return -EFAULT;
1669
1670                 err = kcm_unattach_ioctl(sock, &info);
1671
1672                 break;
1673         }
1674         case SIOCKCMCLONE: {
1675                 struct kcm_clone info;
1676                 struct file *file;
1677
1678                 info.fd = get_unused_fd_flags(0);
1679                 if (unlikely(info.fd < 0))
1680                         return info.fd;
1681
1682                 file = kcm_clone(sock);
1683                 if (IS_ERR(file)) {
1684                         put_unused_fd(info.fd);
1685                         return PTR_ERR(file);
1686                 }
1687                 if (copy_to_user((void __user *)arg, &info,
1688                                  sizeof(info))) {
1689                         put_unused_fd(info.fd);
1690                         fput(file);
1691                         return -EFAULT;
1692                 }
1693                 fd_install(info.fd, file);
1694                 err = 0;
1695                 break;
1696         }
1697         default:
1698                 err = -ENOIOCTLCMD;
1699                 break;
1700         }
1701
1702         return err;
1703 }
1704
1705 static void free_mux(struct rcu_head *rcu)
1706 {
1707         struct kcm_mux *mux = container_of(rcu,
1708             struct kcm_mux, rcu);
1709
1710         kmem_cache_free(kcm_muxp, mux);
1711 }
1712
1713 static void release_mux(struct kcm_mux *mux)
1714 {
1715         struct kcm_net *knet = mux->knet;
1716         struct kcm_psock *psock, *tmp_psock;
1717
1718         /* Release psocks */
1719         list_for_each_entry_safe(psock, tmp_psock,
1720                                  &mux->psocks, psock_list) {
1721                 if (!WARN_ON(psock->unattaching))
1722                         kcm_unattach(psock);
1723         }
1724
1725         if (WARN_ON(mux->psocks_cnt))
1726                 return;
1727
1728         __skb_queue_purge(&mux->rx_hold_queue);
1729
1730         mutex_lock(&knet->mutex);
1731         aggregate_mux_stats(&mux->stats, &knet->aggregate_mux_stats);
1732         aggregate_psock_stats(&mux->aggregate_psock_stats,
1733                               &knet->aggregate_psock_stats);
1734         aggregate_strp_stats(&mux->aggregate_strp_stats,
1735                              &knet->aggregate_strp_stats);
1736         list_del_rcu(&mux->kcm_mux_list);
1737         knet->count--;
1738         mutex_unlock(&knet->mutex);
1739
1740         call_rcu(&mux->rcu, free_mux);
1741 }
1742
1743 static void kcm_done(struct kcm_sock *kcm)
1744 {
1745         struct kcm_mux *mux = kcm->mux;
1746         struct sock *sk = &kcm->sk;
1747         int socks_cnt;
1748
1749         spin_lock_bh(&mux->rx_lock);
1750         if (kcm->rx_psock) {
1751                 /* Cleanup in unreserve_rx_kcm */
1752                 WARN_ON(kcm->done);
1753                 kcm->rx_disabled = 1;
1754                 kcm->done = 1;
1755                 spin_unlock_bh(&mux->rx_lock);
1756                 return;
1757         }
1758
1759         if (kcm->rx_wait) {
1760                 list_del(&kcm->wait_rx_list);
1761                 /* paired with lockless reads in kcm_rfree() */
1762                 WRITE_ONCE(kcm->rx_wait, false);
1763         }
1764         /* Move any pending receive messages to other kcm sockets */
1765         requeue_rx_msgs(mux, &sk->sk_receive_queue);
1766
1767         spin_unlock_bh(&mux->rx_lock);
1768
1769         if (WARN_ON(sk_rmem_alloc_get(sk)))
1770                 return;
1771
1772         /* Detach from MUX */
1773         spin_lock_bh(&mux->lock);
1774
1775         list_del(&kcm->kcm_sock_list);
1776         mux->kcm_socks_cnt--;
1777         socks_cnt = mux->kcm_socks_cnt;
1778
1779         spin_unlock_bh(&mux->lock);
1780
1781         if (!socks_cnt) {
1782                 /* We are done with the mux now. */
1783                 release_mux(mux);
1784         }
1785
1786         WARN_ON(kcm->rx_wait);
1787
1788         sock_put(&kcm->sk);
1789 }
1790
1791 /* Called by kcm_release to close a KCM socket.
1792  * If this is the last KCM socket on the MUX, destroy the MUX.
1793  */
1794 static int kcm_release(struct socket *sock)
1795 {
1796         struct sock *sk = sock->sk;
1797         struct kcm_sock *kcm;
1798         struct kcm_mux *mux;
1799         struct kcm_psock *psock;
1800
1801         if (!sk)
1802                 return 0;
1803
1804         kcm = kcm_sk(sk);
1805         mux = kcm->mux;
1806
1807         lock_sock(sk);
1808         sock_orphan(sk);
1809         kfree_skb(kcm->seq_skb);
1810
1811         /* Purge queue under lock to avoid race condition with tx_work trying
1812          * to act when queue is nonempty. If tx_work runs after this point
1813          * it will just return.
1814          */
1815         __skb_queue_purge(&sk->sk_write_queue);
1816
1817         /* Set tx_stopped. This is checked when psock is bound to a kcm and we
1818          * get a writespace callback. This prevents further work being queued
1819          * from the callback (unbinding the psock occurs after canceling work.
1820          */
1821         kcm->tx_stopped = 1;
1822
1823         release_sock(sk);
1824
1825         spin_lock_bh(&mux->lock);
1826         if (kcm->tx_wait) {
1827                 /* Take of tx_wait list, after this point there should be no way
1828                  * that a psock will be assigned to this kcm.
1829                  */
1830                 list_del(&kcm->wait_psock_list);
1831                 kcm->tx_wait = false;
1832         }
1833         spin_unlock_bh(&mux->lock);
1834
1835         /* Cancel work. After this point there should be no outside references
1836          * to the kcm socket.
1837          */
1838         cancel_work_sync(&kcm->tx_work);
1839
1840         lock_sock(sk);
1841         psock = kcm->tx_psock;
1842         if (psock) {
1843                 /* A psock was reserved, so we need to kill it since it
1844                  * may already have some bytes queued from a message. We
1845                  * need to do this after removing kcm from tx_wait list.
1846                  */
1847                 kcm_abort_tx_psock(psock, EPIPE, false);
1848                 unreserve_psock(kcm);
1849         }
1850         release_sock(sk);
1851
1852         WARN_ON(kcm->tx_wait);
1853         WARN_ON(kcm->tx_psock);
1854
1855         sock->sk = NULL;
1856
1857         kcm_done(kcm);
1858
1859         return 0;
1860 }
1861
1862 static const struct proto_ops kcm_dgram_ops = {
1863         .family =       PF_KCM,
1864         .owner =        THIS_MODULE,
1865         .release =      kcm_release,
1866         .bind =         sock_no_bind,
1867         .connect =      sock_no_connect,
1868         .socketpair =   sock_no_socketpair,
1869         .accept =       sock_no_accept,
1870         .getname =      sock_no_getname,
1871         .poll =         datagram_poll,
1872         .ioctl =        kcm_ioctl,
1873         .listen =       sock_no_listen,
1874         .shutdown =     sock_no_shutdown,
1875         .setsockopt =   kcm_setsockopt,
1876         .getsockopt =   kcm_getsockopt,
1877         .sendmsg =      kcm_sendmsg,
1878         .recvmsg =      kcm_recvmsg,
1879         .mmap =         sock_no_mmap,
1880         .sendpage =     kcm_sendpage,
1881 };
1882
1883 static const struct proto_ops kcm_seqpacket_ops = {
1884         .family =       PF_KCM,
1885         .owner =        THIS_MODULE,
1886         .release =      kcm_release,
1887         .bind =         sock_no_bind,
1888         .connect =      sock_no_connect,
1889         .socketpair =   sock_no_socketpair,
1890         .accept =       sock_no_accept,
1891         .getname =      sock_no_getname,
1892         .poll =         datagram_poll,
1893         .ioctl =        kcm_ioctl,
1894         .listen =       sock_no_listen,
1895         .shutdown =     sock_no_shutdown,
1896         .setsockopt =   kcm_setsockopt,
1897         .getsockopt =   kcm_getsockopt,
1898         .sendmsg =      kcm_sendmsg,
1899         .recvmsg =      kcm_recvmsg,
1900         .mmap =         sock_no_mmap,
1901         .sendpage =     kcm_sendpage,
1902         .splice_read =  kcm_splice_read,
1903 };
1904
1905 /* Create proto operation for kcm sockets */
1906 static int kcm_create(struct net *net, struct socket *sock,
1907                       int protocol, int kern)
1908 {
1909         struct kcm_net *knet = net_generic(net, kcm_net_id);
1910         struct sock *sk;
1911         struct kcm_mux *mux;
1912
1913         switch (sock->type) {
1914         case SOCK_DGRAM:
1915                 sock->ops = &kcm_dgram_ops;
1916                 break;
1917         case SOCK_SEQPACKET:
1918                 sock->ops = &kcm_seqpacket_ops;
1919                 break;
1920         default:
1921                 return -ESOCKTNOSUPPORT;
1922         }
1923
1924         if (protocol != KCMPROTO_CONNECTED)
1925                 return -EPROTONOSUPPORT;
1926
1927         sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
1928         if (!sk)
1929                 return -ENOMEM;
1930
1931         /* Allocate a kcm mux, shared between KCM sockets */
1932         mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
1933         if (!mux) {
1934                 sk_free(sk);
1935                 return -ENOMEM;
1936         }
1937
1938         spin_lock_init(&mux->lock);
1939         spin_lock_init(&mux->rx_lock);
1940         INIT_LIST_HEAD(&mux->kcm_socks);
1941         INIT_LIST_HEAD(&mux->kcm_rx_waiters);
1942         INIT_LIST_HEAD(&mux->kcm_tx_waiters);
1943
1944         INIT_LIST_HEAD(&mux->psocks);
1945         INIT_LIST_HEAD(&mux->psocks_ready);
1946         INIT_LIST_HEAD(&mux->psocks_avail);
1947
1948         mux->knet = knet;
1949
1950         /* Add new MUX to list */
1951         mutex_lock(&knet->mutex);
1952         list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
1953         knet->count++;
1954         mutex_unlock(&knet->mutex);
1955
1956         skb_queue_head_init(&mux->rx_hold_queue);
1957
1958         /* Init KCM socket */
1959         sock_init_data(sock, sk);
1960         init_kcm_sock(kcm_sk(sk), mux);
1961
1962         return 0;
1963 }
1964
1965 static const struct net_proto_family kcm_family_ops = {
1966         .family = PF_KCM,
1967         .create = kcm_create,
1968         .owner  = THIS_MODULE,
1969 };
1970
1971 static __net_init int kcm_init_net(struct net *net)
1972 {
1973         struct kcm_net *knet = net_generic(net, kcm_net_id);
1974
1975         INIT_LIST_HEAD_RCU(&knet->mux_list);
1976         mutex_init(&knet->mutex);
1977
1978         return 0;
1979 }
1980
1981 static __net_exit void kcm_exit_net(struct net *net)
1982 {
1983         struct kcm_net *knet = net_generic(net, kcm_net_id);
1984
1985         /* All KCM sockets should be closed at this point, which should mean
1986          * that all multiplexors and psocks have been destroyed.
1987          */
1988         WARN_ON(!list_empty(&knet->mux_list));
1989
1990         mutex_destroy(&knet->mutex);
1991 }
1992
1993 static struct pernet_operations kcm_net_ops = {
1994         .init = kcm_init_net,
1995         .exit = kcm_exit_net,
1996         .id   = &kcm_net_id,
1997         .size = sizeof(struct kcm_net),
1998 };
1999
2000 static int __init kcm_init(void)
2001 {
2002         int err = -ENOMEM;
2003
2004         kcm_muxp = kmem_cache_create("kcm_mux_cache",
2005                                      sizeof(struct kcm_mux), 0,
2006                                      SLAB_HWCACHE_ALIGN, NULL);
2007         if (!kcm_muxp)
2008                 goto fail;
2009
2010         kcm_psockp = kmem_cache_create("kcm_psock_cache",
2011                                        sizeof(struct kcm_psock), 0,
2012                                         SLAB_HWCACHE_ALIGN, NULL);
2013         if (!kcm_psockp)
2014                 goto fail;
2015
2016         kcm_wq = create_singlethread_workqueue("kkcmd");
2017         if (!kcm_wq)
2018                 goto fail;
2019
2020         err = proto_register(&kcm_proto, 1);
2021         if (err)
2022                 goto fail;
2023
2024         err = register_pernet_device(&kcm_net_ops);
2025         if (err)
2026                 goto net_ops_fail;
2027
2028         err = sock_register(&kcm_family_ops);
2029         if (err)
2030                 goto sock_register_fail;
2031
2032         err = kcm_proc_init();
2033         if (err)
2034                 goto proc_init_fail;
2035
2036         return 0;
2037
2038 proc_init_fail:
2039         sock_unregister(PF_KCM);
2040
2041 sock_register_fail:
2042         unregister_pernet_device(&kcm_net_ops);
2043
2044 net_ops_fail:
2045         proto_unregister(&kcm_proto);
2046
2047 fail:
2048         kmem_cache_destroy(kcm_muxp);
2049         kmem_cache_destroy(kcm_psockp);
2050
2051         if (kcm_wq)
2052                 destroy_workqueue(kcm_wq);
2053
2054         return err;
2055 }
2056
2057 static void __exit kcm_exit(void)
2058 {
2059         kcm_proc_exit();
2060         sock_unregister(PF_KCM);
2061         unregister_pernet_device(&kcm_net_ops);
2062         proto_unregister(&kcm_proto);
2063         destroy_workqueue(kcm_wq);
2064
2065         kmem_cache_destroy(kcm_muxp);
2066         kmem_cache_destroy(kcm_psockp);
2067 }
2068
2069 module_init(kcm_init);
2070 module_exit(kcm_exit);
2071
2072 MODULE_LICENSE("GPL");
2073 MODULE_ALIAS_NETPROTO(PF_KCM);