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         kcm->rx_wait = true;
 165 }
 166
 167 static void kcm_rfree(struct sk_buff *skb)
 168 {
 169         struct sock *sk = skb->sk;
 170         struct kcm_sock *kcm = kcm_sk(sk);
 171         struct kcm_mux *mux = kcm->mux;
 172         unsigned int len = skb->truesize;
 173
 174         sk_mem_uncharge(sk, len);
 175         atomic_sub(len, &sk->sk_rmem_alloc);
 176
 177         /* For reading rx_wait and rx_psock without holding lock */
 178         smp_mb__after_atomic();
 179
 180         if (!kcm->rx_wait && !kcm->rx_psock &&
 181             sk_rmem_alloc_get(sk) < sk->sk_rcvlowat) {
 182                 spin_lock_bh(&mux->rx_lock);
 183                 kcm_rcv_ready(kcm);
 184                 spin_unlock_bh(&mux->rx_lock);
 185         }
 186 }
 187
 188 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb)
 189 {
 190         struct sk_buff_head *list = &sk->sk_receive_queue;
 191
 192         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
 193                 return -ENOMEM;
 194
 195         if (!sk_rmem_schedule(sk, skb, skb->truesize))
 196                 return -ENOBUFS;
 197
 198         skb->dev = NULL;
 199
 200         skb_orphan(skb);
 201         skb->sk = sk;
 202         skb->destructor = kcm_rfree;
 203         atomic_add(skb->truesize, &sk->sk_rmem_alloc);
 204         sk_mem_charge(sk, skb->truesize);
 205
 206         skb_queue_tail(list, skb);
 207
 208         if (!sock_flag(sk, SOCK_DEAD))
 209                 sk->sk_data_ready(sk);
 210
 211         return 0;
 212 }
 213
 214 /* Requeue received messages for a kcm socket to other kcm sockets. This is
 215  * called with a kcm socket is receive disabled.
 216  * RX mux lock held.
 217  */
 218 static void requeue_rx_msgs(struct kcm_mux *mux, struct sk_buff_head *head)
 219 {
 220         struct sk_buff *skb;
 221         struct kcm_sock *kcm;
 222
 223         while ((skb = __skb_dequeue(head))) {
 224                 /* Reset destructor to avoid calling kcm_rcv_ready */
 225                 skb->destructor = sock_rfree;
 226                 skb_orphan(skb);
 227 try_again:
 228                 if (list_empty(&mux->kcm_rx_waiters)) {
 229                         skb_queue_tail(&mux->rx_hold_queue, skb);
 230                         continue;
 231                 }
 232
 233                 kcm = list_first_entry(&mux->kcm_rx_waiters,
 234                                        struct kcm_sock, wait_rx_list);
 235
 236                 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 237                         /* Should mean socket buffer full */
 238                         list_del(&kcm->wait_rx_list);
 239                         kcm->rx_wait = false;
 240
 241                         /* Commit rx_wait to read in kcm_free */
 242                         smp_wmb();
 243
 244                         goto try_again;
 245                 }
 246         }
 247 }
 248
 249 /* Lower sock lock held */
 250 static struct kcm_sock *reserve_rx_kcm(struct kcm_psock *psock,
 251                                        struct sk_buff *head)
 252 {
 253         struct kcm_mux *mux = psock->mux;
 254         struct kcm_sock *kcm;
 255
 256         WARN_ON(psock->ready_rx_msg);
 257
 258         if (psock->rx_kcm)
 259                 return psock->rx_kcm;
 260
 261         spin_lock_bh(&mux->rx_lock);
 262
 263         if (psock->rx_kcm) {
 264                 spin_unlock_bh(&mux->rx_lock);
 265                 return psock->rx_kcm;
 266         }
 267
 268         kcm_update_rx_mux_stats(mux, psock);
 269
 270         if (list_empty(&mux->kcm_rx_waiters)) {
 271                 psock->ready_rx_msg = head;
 272                 strp_pause(&psock->strp);
 273                 list_add_tail(&psock->psock_ready_list,
 274                               &mux->psocks_ready);
 275                 spin_unlock_bh(&mux->rx_lock);
 276                 return NULL;
 277         }
 278
 279         kcm = list_first_entry(&mux->kcm_rx_waiters,
 280                                struct kcm_sock, wait_rx_list);
 281         list_del(&kcm->wait_rx_list);
 282         kcm->rx_wait = false;
 283
 284         psock->rx_kcm = kcm;
 285         kcm->rx_psock = psock;
 286
 287         spin_unlock_bh(&mux->rx_lock);
 288
 289         return kcm;
 290 }
 291
 292 static void kcm_done(struct kcm_sock *kcm);
 293
 294 static void kcm_done_work(struct work_struct *w)
 295 {
 296         kcm_done(container_of(w, struct kcm_sock, done_work));
 297 }
 298
 299 /* Lower sock held */
 300 static void unreserve_rx_kcm(struct kcm_psock *psock,
 301                              bool rcv_ready)
 302 {
 303         struct kcm_sock *kcm = psock->rx_kcm;
 304         struct kcm_mux *mux = psock->mux;
 305
 306         if (!kcm)
 307                 return;
 308
 309         spin_lock_bh(&mux->rx_lock);
 310
 311         psock->rx_kcm = NULL;
 312         kcm->rx_psock = NULL;
 313
 314         /* Commit kcm->rx_psock before sk_rmem_alloc_get to sync with
 315          * kcm_rfree
 316          */
 317         smp_mb();
 318
 319         if (unlikely(kcm->done)) {
 320                 spin_unlock_bh(&mux->rx_lock);
 321
 322                 /* Need to run kcm_done in a task since we need to qcquire
 323                  * callback locks which may already be held here.
 324                  */
 325                 INIT_WORK(&kcm->done_work, kcm_done_work);
 326                 schedule_work(&kcm->done_work);
 327                 return;
 328         }
 329
 330         if (unlikely(kcm->rx_disabled)) {
 331                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
 332         } else if (rcv_ready || unlikely(!sk_rmem_alloc_get(&kcm->sk))) {
 333                 /* Check for degenerative race with rx_wait that all
 334                  * data was dequeued (accounted for in kcm_rfree).
 335                  */
 336                 kcm_rcv_ready(kcm);
 337         }
 338         spin_unlock_bh(&mux->rx_lock);
 339 }
 340
 341 /* Lower sock lock held */
 342 static void psock_data_ready(struct sock *sk)
 343 {
 344         struct kcm_psock *psock;
 345
 346         read_lock_bh(&sk->sk_callback_lock);
 347
 348         psock = (struct kcm_psock *)sk->sk_user_data;
 349         if (likely(psock))
 350                 strp_data_ready(&psock->strp);
 351
 352         read_unlock_bh(&sk->sk_callback_lock);
 353 }
 354
 355 /* Called with lower sock held */
 356 static void kcm_rcv_strparser(struct strparser *strp, struct sk_buff *skb)
 357 {
 358         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 359         struct kcm_sock *kcm;
 360
 361 try_queue:
 362         kcm = reserve_rx_kcm(psock, skb);
 363         if (!kcm) {
 364                  /* Unable to reserve a KCM, message is held in psock and strp
 365                   * is paused.
 366                   */
 367                 return;
 368         }
 369
 370         if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 371                 /* Should mean socket buffer full */
 372                 unreserve_rx_kcm(psock, false);
 373                 goto try_queue;
 374         }
 375 }
 376
 377 static int kcm_parse_func_strparser(struct strparser *strp, struct sk_buff *skb)
 378 {
 379         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 380         struct bpf_prog *prog = psock->bpf_prog;
 381
 382         return (*prog->bpf_func)(skb, prog->insnsi);
 383 }
 384
 385 static int kcm_read_sock_done(struct strparser *strp, int err)
 386 {
 387         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 388
 389         unreserve_rx_kcm(psock, true);
 390
 391         return err;
 392 }
 393
 394 static void psock_state_change(struct sock *sk)
 395 {
 396         /* TCP only does a EPOLLIN for a half close. Do a EPOLLHUP here
 397          * since application will normally not poll with EPOLLIN
 398          * on the TCP sockets.
 399          */
 400
 401         report_csk_error(sk, EPIPE);
 402 }
 403
 404 static void psock_write_space(struct sock *sk)
 405 {
 406         struct kcm_psock *psock;
 407         struct kcm_mux *mux;
 408         struct kcm_sock *kcm;
 409
 410         read_lock_bh(&sk->sk_callback_lock);
 411
 412         psock = (struct kcm_psock *)sk->sk_user_data;
 413         if (unlikely(!psock))
 414                 goto out;
 415         mux = psock->mux;
 416
 417         spin_lock_bh(&mux->lock);
 418
 419         /* Check if the socket is reserved so someone is waiting for sending. */
 420         kcm = psock->tx_kcm;
 421         if (kcm && !unlikely(kcm->tx_stopped))
 422                 queue_work(kcm_wq, &kcm->tx_work);
 423
 424         spin_unlock_bh(&mux->lock);
 425 out:
 426         read_unlock_bh(&sk->sk_callback_lock);
 427 }
 428
 429 static void unreserve_psock(struct kcm_sock *kcm);
 430
 431 /* kcm sock is locked. */
 432 static struct kcm_psock *reserve_psock(struct kcm_sock *kcm)
 433 {
 434         struct kcm_mux *mux = kcm->mux;
 435         struct kcm_psock *psock;
 436
 437         psock = kcm->tx_psock;
 438
 439         smp_rmb(); /* Must read tx_psock before tx_wait */
 440
 441         if (psock) {
 442                 WARN_ON(kcm->tx_wait);
 443                 if (unlikely(psock->tx_stopped))
 444                         unreserve_psock(kcm);
 445                 else
 446                         return kcm->tx_psock;
 447         }
 448
 449         spin_lock_bh(&mux->lock);
 450
 451         /* Check again under lock to see if psock was reserved for this
 452          * psock via psock_unreserve.
 453          */
 454         psock = kcm->tx_psock;
 455         if (unlikely(psock)) {
 456                 WARN_ON(kcm->tx_wait);
 457                 spin_unlock_bh(&mux->lock);
 458                 return kcm->tx_psock;
 459         }
 460
 461         if (!list_empty(&mux->psocks_avail)) {
 462                 psock = list_first_entry(&mux->psocks_avail,
 463                                          struct kcm_psock,
 464                                          psock_avail_list);
 465                 list_del(&psock->psock_avail_list);
 466                 if (kcm->tx_wait) {
 467                         list_del(&kcm->wait_psock_list);
 468                         kcm->tx_wait = false;
 469                 }
 470                 kcm->tx_psock = psock;
 471                 psock->tx_kcm = kcm;
 472                 KCM_STATS_INCR(psock->stats.reserved);
 473         } else if (!kcm->tx_wait) {
 474                 list_add_tail(&kcm->wait_psock_list,
 475                               &mux->kcm_tx_waiters);
 476                 kcm->tx_wait = true;
 477         }
 478
 479         spin_unlock_bh(&mux->lock);
 480
 481         return psock;
 482 }
 483
 484 /* mux lock held */
 485 static void psock_now_avail(struct kcm_psock *psock)
 486 {
 487         struct kcm_mux *mux = psock->mux;
 488         struct kcm_sock *kcm;
 489
 490         if (list_empty(&mux->kcm_tx_waiters)) {
 491                 list_add_tail(&psock->psock_avail_list,
 492                               &mux->psocks_avail);
 493         } else {
 494                 kcm = list_first_entry(&mux->kcm_tx_waiters,
 495                                        struct kcm_sock,
 496                                        wait_psock_list);
 497                 list_del(&kcm->wait_psock_list);
 498                 kcm->tx_wait = false;
 499                 psock->tx_kcm = kcm;
 500
 501                 /* Commit before changing tx_psock since that is read in
 502                  * reserve_psock before queuing work.
 503                  */
 504                 smp_mb();
 505
 506                 kcm->tx_psock = psock;
 507                 KCM_STATS_INCR(psock->stats.reserved);
 508                 queue_work(kcm_wq, &kcm->tx_work);
 509         }
 510 }
 511
 512 /* kcm sock is locked. */
 513 static void unreserve_psock(struct kcm_sock *kcm)
 514 {
 515         struct kcm_psock *psock;
 516         struct kcm_mux *mux = kcm->mux;
 517
 518         spin_lock_bh(&mux->lock);
 519
 520         psock = kcm->tx_psock;
 521
 522         if (WARN_ON(!psock)) {
 523                 spin_unlock_bh(&mux->lock);
 524                 return;
 525         }
 526
 527         smp_rmb(); /* Read tx_psock before tx_wait */
 528
 529         kcm_update_tx_mux_stats(mux, psock);
 530
 531         WARN_ON(kcm->tx_wait);
 532
 533         kcm->tx_psock = NULL;
 534         psock->tx_kcm = NULL;
 535         KCM_STATS_INCR(psock->stats.unreserved);
 536
 537         if (unlikely(psock->tx_stopped)) {
 538                 if (psock->done) {
 539                         /* Deferred free */
 540                         list_del(&psock->psock_list);
 541                         mux->psocks_cnt--;
 542                         sock_put(psock->sk);
 543                         fput(psock->sk->sk_socket->file);
 544                         kmem_cache_free(kcm_psockp, psock);
 545                 }
 546
 547                 /* Don't put back on available list */
 548
 549                 spin_unlock_bh(&mux->lock);
 550
 551                 return;
 552         }
 553
 554         psock_now_avail(psock);
 555
 556         spin_unlock_bh(&mux->lock);
 557 }
 558
 559 static void kcm_report_tx_retry(struct kcm_sock *kcm)
 560 {
 561         struct kcm_mux *mux = kcm->mux;
 562
 563         spin_lock_bh(&mux->lock);
 564         KCM_STATS_INCR(mux->stats.tx_retries);
 565         spin_unlock_bh(&mux->lock);
 566 }
 567
 568 /* Write any messages ready on the kcm socket.  Called with kcm sock lock
 569  * held.  Return bytes actually sent or error.
 570  */
 571 static int kcm_write_msgs(struct kcm_sock *kcm)
 572 {
 573         struct sock *sk = &kcm->sk;
 574         struct kcm_psock *psock;
 575         struct sk_buff *skb, *head;
 576         struct kcm_tx_msg *txm;
 577         unsigned short fragidx, frag_offset;
 578         unsigned int sent, total_sent = 0;
 579         int ret = 0;
 580
 581         kcm->tx_wait_more = false;
 582         psock = kcm->tx_psock;
 583         if (unlikely(psock && psock->tx_stopped)) {
 584                 /* A reserved psock was aborted asynchronously. Unreserve
 585                  * it and we'll retry the message.
 586                  */
 587                 unreserve_psock(kcm);
 588                 kcm_report_tx_retry(kcm);
 589                 if (skb_queue_empty(&sk->sk_write_queue))
 590                         return 0;
 591
 592                 kcm_tx_msg(skb_peek(&sk->sk_write_queue))->sent = 0;
 593
 594         } else if (skb_queue_empty(&sk->sk_write_queue)) {
 595                 return 0;
 596         }
 597
 598         head = skb_peek(&sk->sk_write_queue);
 599         txm = kcm_tx_msg(head);
 600
 601         if (txm->sent) {
 602                 /* Send of first skbuff in queue already in progress */
 603                 if (WARN_ON(!psock)) {
 604                         ret = -EINVAL;
 605                         goto out;
 606                 }
 607                 sent = txm->sent;
 608                 frag_offset = txm->frag_offset;
 609                 fragidx = txm->fragidx;
 610                 skb = txm->frag_skb;
 611
 612                 goto do_frag;
 613         }
 614
 615 try_again:
 616         psock = reserve_psock(kcm);
 617         if (!psock)
 618                 goto out;
 619
 620         do {
 621                 skb = head;
 622                 txm = kcm_tx_msg(head);
 623                 sent = 0;
 624
 625 do_frag_list:
 626                 if (WARN_ON(!skb_shinfo(skb)->nr_frags)) {
 627                         ret = -EINVAL;
 628                         goto out;
 629                 }
 630
 631                 for (fragidx = 0; fragidx < skb_shinfo(skb)->nr_frags;
 632                      fragidx++) {
 633                         skb_frag_t *frag;
 634
 635                         frag_offset = 0;
 636 do_frag:
 637                         frag = &skb_shinfo(skb)->frags[fragidx];
 638                         if (WARN_ON(!frag->size)) {
 639                                 ret = -EINVAL;
 640                                 goto out;
 641                         }
 642
 643                         ret = kernel_sendpage(psock->sk->sk_socket,
 644                                               frag->page.p,
 645                                               frag->page_offset + frag_offset,
 646                                               frag->size - frag_offset,
 647                                               MSG_DONTWAIT);
 648                         if (ret <= 0) {
 649                                 if (ret == -EAGAIN) {
 650                                         /* Save state to try again when there's
 651                                          * write space on the socket
 652                                          */
 653                                         txm->sent = sent;
 654                                         txm->frag_offset = frag_offset;
 655                                         txm->fragidx = fragidx;
 656                                         txm->frag_skb = skb;
 657
 658                                         ret = 0;
 659                                         goto out;
 660                                 }
 661
 662                                 /* Hard failure in sending message, abort this
 663                                  * psock since it has lost framing
 664                                  * synchonization and retry sending the
 665                                  * message from the beginning.
 666                                  */
 667                                 kcm_abort_tx_psock(psock, ret ? -ret : EPIPE,
 668                                                    true);
 669                                 unreserve_psock(kcm);
 670
 671                                 txm->sent = 0;
 672                                 kcm_report_tx_retry(kcm);
 673                                 ret = 0;
 674
 675                                 goto try_again;
 676                         }
 677
 678                         sent += ret;
 679                         frag_offset += ret;
 680                         KCM_STATS_ADD(psock->stats.tx_bytes, ret);
 681                         if (frag_offset < frag->size) {
 682                                 /* Not finished with this frag */
 683                                 goto do_frag;
 684                         }
 685                 }
 686
 687                 if (skb == head) {
 688                         if (skb_has_frag_list(skb)) {
 689                                 skb = skb_shinfo(skb)->frag_list;
 690                                 goto do_frag_list;
 691                         }
 692                 } else if (skb->next) {
 693                         skb = skb->next;
 694                         goto do_frag_list;
 695                 }
 696
 697                 /* Successfully sent the whole packet, account for it. */
 698                 skb_dequeue(&sk->sk_write_queue);
 699                 kfree_skb(head);
 700                 sk->sk_wmem_queued -= sent;
 701                 total_sent += sent;
 702                 KCM_STATS_INCR(psock->stats.tx_msgs);
 703         } while ((head = skb_peek(&sk->sk_write_queue)));
 704 out:
 705         if (!head) {
 706                 /* Done with all queued messages. */
 707                 WARN_ON(!skb_queue_empty(&sk->sk_write_queue));
 708                 unreserve_psock(kcm);
 709         }
 710
 711         /* Check if write space is available */
 712         sk->sk_write_space(sk);
 713
 714         return total_sent ? : ret;
 715 }
 716
 717 static void kcm_tx_work(struct work_struct *w)
 718 {
 719         struct kcm_sock *kcm = container_of(w, struct kcm_sock, tx_work);
 720         struct sock *sk = &kcm->sk;
 721         int err;
 722
 723         lock_sock(sk);
 724
 725         /* Primarily for SOCK_DGRAM sockets, also handle asynchronous tx
 726          * aborts
 727          */
 728         err = kcm_write_msgs(kcm);
 729         if (err < 0) {
 730                 /* Hard failure in write, report error on KCM socket */
 731                 pr_warn("KCM: Hard failure on kcm_write_msgs %d\n", err);
 732                 report_csk_error(&kcm->sk, -err);
 733                 goto out;
 734         }
 735
 736         /* Primarily for SOCK_SEQPACKET sockets */
 737         if (likely(sk->sk_socket) &&
 738             test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
 739                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 740                 sk->sk_write_space(sk);
 741         }
 742
 743 out:
 744         release_sock(sk);
 745 }
 746
 747 static void kcm_push(struct kcm_sock *kcm)
 748 {
 749         if (kcm->tx_wait_more)
 750                 kcm_write_msgs(kcm);
 751 }
 752
 753 static ssize_t kcm_sendpage(struct socket *sock, struct page *page,
 754                             int offset, size_t size, int flags)
 755
 756 {
 757         struct sock *sk = sock->sk;
 758         struct kcm_sock *kcm = kcm_sk(sk);
 759         struct sk_buff *skb = NULL, *head = NULL;
 760         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 761         bool eor;
 762         int err = 0;
 763         int i;
 764
 765         if (flags & MSG_SENDPAGE_NOTLAST)
 766                 flags |= MSG_MORE;
 767
 768         /* No MSG_EOR from splice, only look at MSG_MORE */
 769         eor = !(flags & MSG_MORE);
 770
 771         lock_sock(sk);
 772
 773         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 774
 775         err = -EPIPE;
 776         if (sk->sk_err)
 777                 goto out_error;
 778
 779         if (kcm->seq_skb) {
 780                 /* Previously opened message */
 781                 head = kcm->seq_skb;
 782                 skb = kcm_tx_msg(head)->last_skb;
 783                 i = skb_shinfo(skb)->nr_frags;
 784
 785                 if (skb_can_coalesce(skb, i, page, offset)) {
 786                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], size);
 787                         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
 788                         goto coalesced;
 789                 }
 790
 791                 if (i >= MAX_SKB_FRAGS) {
 792                         struct sk_buff *tskb;
 793
 794                         tskb = alloc_skb(0, sk->sk_allocation);
 795                         while (!tskb) {
 796                                 kcm_push(kcm);
 797                                 err = sk_stream_wait_memory(sk, &timeo);
 798                                 if (err)
 799                                         goto out_error;
 800                         }
 801
 802                         if (head == skb)
 803                                 skb_shinfo(head)->frag_list = tskb;
 804                         else
 805                                 skb->next = tskb;
 806
 807                         skb = tskb;
 808                         skb->ip_summed = CHECKSUM_UNNECESSARY;
 809                         i = 0;
 810                 }
 811         } else {
 812                 /* Call the sk_stream functions to manage the sndbuf mem. */
 813                 if (!sk_stream_memory_free(sk)) {
 814                         kcm_push(kcm);
 815                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 816                         err = sk_stream_wait_memory(sk, &timeo);
 817                         if (err)
 818                                 goto out_error;
 819                 }
 820
 821                 head = alloc_skb(0, sk->sk_allocation);
 822                 while (!head) {
 823                         kcm_push(kcm);
 824                         err = sk_stream_wait_memory(sk, &timeo);
 825                         if (err)
 826                                 goto out_error;
 827                 }
 828
 829                 skb = head;
 830                 i = 0;
 831         }
 832
 833         get_page(page);
 834         skb_fill_page_desc(skb, i, page, offset, size);
 835         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
 836
 837 coalesced:
 838         skb->len += size;
 839         skb->data_len += size;
 840         skb->truesize += size;
 841         sk->sk_wmem_queued += size;
 842         sk_mem_charge(sk, size);
 843
 844         if (head != skb) {
 845                 head->len += size;
 846                 head->data_len += size;
 847                 head->truesize += size;
 848         }
 849
 850         if (eor) {
 851                 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
 852
 853                 /* Message complete, queue it on send buffer */
 854                 __skb_queue_tail(&sk->sk_write_queue, head);
 855                 kcm->seq_skb = NULL;
 856                 KCM_STATS_INCR(kcm->stats.tx_msgs);
 857
 858                 if (flags & MSG_BATCH) {
 859                         kcm->tx_wait_more = true;
 860                 } else if (kcm->tx_wait_more || not_busy) {
 861                         err = kcm_write_msgs(kcm);
 862                         if (err < 0) {
 863                                 /* We got a hard error in write_msgs but have
 864                                  * already queued this message. Report an error
 865                                  * in the socket, but don't affect return value
 866                                  * from sendmsg
 867                                  */
 868                                 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
 869                                 report_csk_error(&kcm->sk, -err);
 870                         }
 871                 }
 872         } else {
 873                 /* Message not complete, save state */
 874                 kcm->seq_skb = head;
 875                 kcm_tx_msg(head)->last_skb = skb;
 876         }
 877
 878         KCM_STATS_ADD(kcm->stats.tx_bytes, size);
 879
 880         release_sock(sk);
 881         return size;
 882
 883 out_error:
 884         kcm_push(kcm);
 885
 886         err = sk_stream_error(sk, flags, err);
 887
 888         /* make sure we wake any epoll edge trigger waiter */
 889         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
 890                 sk->sk_write_space(sk);
 891
 892         release_sock(sk);
 893         return err;
 894 }
 895
 896 static int kcm_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
 897 {
 898         struct sock *sk = sock->sk;
 899         struct kcm_sock *kcm = kcm_sk(sk);
 900         struct sk_buff *skb = NULL, *head = NULL;
 901         size_t copy, copied = 0;
 902         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 903         int eor = (sock->type == SOCK_DGRAM) ?
 904                   !(msg->msg_flags & MSG_MORE) : !!(msg->msg_flags & MSG_EOR);
 905         int err = -EPIPE;
 906
 907         lock_sock(sk);
 908
 909         /* Per tcp_sendmsg this should be in poll */
 910         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 911
 912         if (sk->sk_err)
 913                 goto out_error;
 914
 915         if (kcm->seq_skb) {
 916                 /* Previously opened message */
 917                 head = kcm->seq_skb;
 918                 skb = kcm_tx_msg(head)->last_skb;
 919                 goto start;
 920         }
 921
 922         /* Call the sk_stream functions to manage the sndbuf mem. */
 923         if (!sk_stream_memory_free(sk)) {
 924                 kcm_push(kcm);
 925                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 926                 err = sk_stream_wait_memory(sk, &timeo);
 927                 if (err)
 928                         goto out_error;
 929         }
 930
 931         if (msg_data_left(msg)) {
 932                 /* New message, alloc head skb */
 933                 head = alloc_skb(0, sk->sk_allocation);
 934                 while (!head) {
 935                         kcm_push(kcm);
 936                         err = sk_stream_wait_memory(sk, &timeo);
 937                         if (err)
 938                                 goto out_error;
 939
 940                         head = alloc_skb(0, sk->sk_allocation);
 941                 }
 942
 943                 skb = head;
 944
 945                 /* Set ip_summed to CHECKSUM_UNNECESSARY to avoid calling
 946                  * csum_and_copy_from_iter from skb_do_copy_data_nocache.
 947                  */
 948                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 949         }
 950
 951 start:
 952         while (msg_data_left(msg)) {
 953                 bool merge = true;
 954                 int i = skb_shinfo(skb)->nr_frags;
 955                 struct page_frag *pfrag = sk_page_frag(sk);
 956
 957                 if (!sk_page_frag_refill(sk, pfrag))
 958                         goto wait_for_memory;
 959
 960                 if (!skb_can_coalesce(skb, i, pfrag->page,
 961                                       pfrag->offset)) {
 962                         if (i == MAX_SKB_FRAGS) {
 963                                 struct sk_buff *tskb;
 964
 965                                 tskb = alloc_skb(0, sk->sk_allocation);
 966                                 if (!tskb)
 967                                         goto wait_for_memory;
 968
 969                                 if (head == skb)
 970                                         skb_shinfo(head)->frag_list = tskb;
 971                                 else
 972                                         skb->next = tskb;
 973
 974                                 skb = tskb;
 975                                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 976                                 continue;
 977                         }
 978                         merge = false;
 979                 }
 980
 981                 copy = min_t(int, msg_data_left(msg),
 982                              pfrag->size - pfrag->offset);
 983
 984                 if (!sk_wmem_schedule(sk, copy))
 985                         goto wait_for_memory;
 986
 987                 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
 988                                                pfrag->page,
 989                                                pfrag->offset,
 990                                                copy);
 991                 if (err)
 992                         goto out_error;
 993
 994                 /* Update the skb. */
 995                 if (merge) {
 996                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
 997                 } else {
 998                         skb_fill_page_desc(skb, i, pfrag->page,
 999                                            pfrag->offset, copy);
1000                         get_page(pfrag->page);
1001                 }
1002
1003                 pfrag->offset += copy;
1004                 copied += copy;
1005                 if (head != skb) {
1006                         head->len += copy;
1007                         head->data_len += copy;
1008                 }
1009
1010                 continue;
1011
1012 wait_for_memory:
1013                 kcm_push(kcm);
1014                 err = sk_stream_wait_memory(sk, &timeo);
1015                 if (err)
1016                         goto out_error;
1017         }
1018
1019         if (eor) {
1020                 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
1021
1022                 if (head) {
1023                         /* Message complete, queue it on send buffer */
1024                         __skb_queue_tail(&sk->sk_write_queue, head);
1025                         kcm->seq_skb = NULL;
1026                         KCM_STATS_INCR(kcm->stats.tx_msgs);
1027                 }
1028
1029                 if (msg->msg_flags & MSG_BATCH) {
1030                         kcm->tx_wait_more = true;
1031                 } else if (kcm->tx_wait_more || not_busy) {
1032                         err = kcm_write_msgs(kcm);
1033                         if (err < 0) {
1034                                 /* We got a hard error in write_msgs but have
1035                                  * already queued this message. Report an error
1036                                  * in the socket, but don't affect return value
1037                                  * from sendmsg
1038                                  */
1039                                 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
1040                                 report_csk_error(&kcm->sk, -err);
1041                         }
1042                 }
1043         } else {
1044                 /* Message not complete, save state */
1045 partial_message:
1046                 if (head) {
1047                         kcm->seq_skb = head;
1048                         kcm_tx_msg(head)->last_skb = skb;
1049                 }
1050         }
1051
1052         KCM_STATS_ADD(kcm->stats.tx_bytes, copied);
1053
1054         release_sock(sk);
1055         return copied;
1056
1057 out_error:
1058         kcm_push(kcm);
1059
1060         if (copied && sock->type == SOCK_SEQPACKET) {
1061                 /* Wrote some bytes before encountering an
1062                  * error, return partial success.
1063                  */
1064                 goto partial_message;
1065         }
1066
1067         if (head != kcm->seq_skb)
1068                 kfree_skb(head);
1069
1070         err = sk_stream_error(sk, msg->msg_flags, err);
1071
1072         /* make sure we wake any epoll edge trigger waiter */
1073         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1074                 sk->sk_write_space(sk);
1075
1076         release_sock(sk);
1077         return err;
1078 }
1079
1080 static struct sk_buff *kcm_wait_data(struct sock *sk, int flags,
1081                                      long timeo, int *err)
1082 {
1083         struct sk_buff *skb;
1084
1085         while (!(skb = skb_peek(&sk->sk_receive_queue))) {
1086                 if (sk->sk_err) {
1087                         *err = sock_error(sk);
1088                         return NULL;
1089                 }
1090
1091                 if (sock_flag(sk, SOCK_DONE))
1092                         return NULL;
1093
1094                 if ((flags & MSG_DONTWAIT) || !timeo) {
1095                         *err = -EAGAIN;
1096                         return NULL;
1097                 }
1098
1099                 sk_wait_data(sk, &timeo, NULL);
1100
1101                 /* Handle signals */
1102                 if (signal_pending(current)) {
1103                         *err = sock_intr_errno(timeo);
1104                         return NULL;
1105                 }
1106         }
1107
1108         return skb;
1109 }
1110
1111 static int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
1112                        size_t len, int flags)
1113 {
1114         struct sock *sk = sock->sk;
1115         struct kcm_sock *kcm = kcm_sk(sk);
1116         int err = 0;
1117         long timeo;
1118         struct strp_msg *stm;
1119         int copied = 0;
1120         struct sk_buff *skb;
1121
1122         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1123
1124         lock_sock(sk);
1125
1126         skb = kcm_wait_data(sk, flags, timeo, &err);
1127         if (!skb)
1128                 goto out;
1129
1130         /* Okay, have a message on the receive queue */
1131
1132         stm = strp_msg(skb);
1133
1134         if (len > stm->full_len)
1135                 len = stm->full_len;
1136
1137         err = skb_copy_datagram_msg(skb, stm->offset, msg, len);
1138         if (err < 0)
1139                 goto out;
1140
1141         copied = len;
1142         if (likely(!(flags & MSG_PEEK))) {
1143                 KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1144                 if (copied < stm->full_len) {
1145                         if (sock->type == SOCK_DGRAM) {
1146                                 /* Truncated message */
1147                                 msg->msg_flags |= MSG_TRUNC;
1148                                 goto msg_finished;
1149                         }
1150                         stm->offset += copied;
1151                         stm->full_len -= copied;
1152                 } else {
1153 msg_finished:
1154                         /* Finished with message */
1155                         msg->msg_flags |= MSG_EOR;
1156                         KCM_STATS_INCR(kcm->stats.rx_msgs);
1157                         skb_unlink(skb, &sk->sk_receive_queue);
1158                         kfree_skb(skb);
1159                 }
1160         }
1161
1162 out:
1163         release_sock(sk);
1164
1165         return copied ? : err;
1166 }
1167
1168 static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1169                                struct pipe_inode_info *pipe, size_t len,
1170                                unsigned int flags)
1171 {
1172         struct sock *sk = sock->sk;
1173         struct kcm_sock *kcm = kcm_sk(sk);
1174         long timeo;
1175         struct strp_msg *stm;
1176         int err = 0;
1177         ssize_t copied;
1178         struct sk_buff *skb;
1179
1180         /* Only support splice for SOCKSEQPACKET */
1181
1182         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1183
1184         lock_sock(sk);
1185
1186         skb = kcm_wait_data(sk, flags, timeo, &err);
1187         if (!skb)
1188                 goto err_out;
1189
1190         /* Okay, have a message on the receive queue */
1191
1192         stm = strp_msg(skb);
1193
1194         if (len > stm->full_len)
1195                 len = stm->full_len;
1196
1197         copied = skb_splice_bits(skb, sk, stm->offset, pipe, len, flags);
1198         if (copied < 0) {
1199                 err = copied;
1200                 goto err_out;
1201         }
1202
1203         KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1204
1205         stm->offset += copied;
1206         stm->full_len -= copied;
1207
1208         /* We have no way to return MSG_EOR. If all the bytes have been
1209          * read we still leave the message in the receive socket buffer.
1210          * A subsequent recvmsg needs to be done to return MSG_EOR and
1211          * finish reading the message.
1212          */
1213
1214         release_sock(sk);
1215
1216         return copied;
1217
1218 err_out:
1219         release_sock(sk);
1220
1221         return err;
1222 }
1223
1224 /* kcm sock lock held */
1225 static void kcm_recv_disable(struct kcm_sock *kcm)
1226 {
1227         struct kcm_mux *mux = kcm->mux;
1228
1229         if (kcm->rx_disabled)
1230                 return;
1231
1232         spin_lock_bh(&mux->rx_lock);
1233
1234         kcm->rx_disabled = 1;
1235
1236         /* If a psock is reserved we'll do cleanup in unreserve */
1237         if (!kcm->rx_psock) {
1238                 if (kcm->rx_wait) {
1239                         list_del(&kcm->wait_rx_list);
1240                         kcm->rx_wait = false;
1241                 }
1242
1243                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1244         }
1245
1246         spin_unlock_bh(&mux->rx_lock);
1247 }
1248
1249 /* kcm sock lock held */
1250 static void kcm_recv_enable(struct kcm_sock *kcm)
1251 {
1252         struct kcm_mux *mux = kcm->mux;
1253
1254         if (!kcm->rx_disabled)
1255                 return;
1256
1257         spin_lock_bh(&mux->rx_lock);
1258
1259         kcm->rx_disabled = 0;
1260         kcm_rcv_ready(kcm);
1261
1262         spin_unlock_bh(&mux->rx_lock);
1263 }
1264
1265 static int kcm_setsockopt(struct socket *sock, int level, int optname,
1266                           char __user *optval, unsigned int optlen)
1267 {
1268         struct kcm_sock *kcm = kcm_sk(sock->sk);
1269         int val, valbool;
1270         int err = 0;
1271
1272         if (level != SOL_KCM)
1273                 return -ENOPROTOOPT;
1274
1275         if (optlen < sizeof(int))
1276                 return -EINVAL;
1277
1278         if (get_user(val, (int __user *)optval))
1279                 return -EINVAL;
1280
1281         valbool = val ? 1 : 0;
1282
1283         switch (optname) {
1284         case KCM_RECV_DISABLE:
1285                 lock_sock(&kcm->sk);
1286                 if (valbool)
1287                         kcm_recv_disable(kcm);
1288                 else
1289                         kcm_recv_enable(kcm);
1290                 release_sock(&kcm->sk);
1291                 break;
1292         default:
1293                 err = -ENOPROTOOPT;
1294         }
1295
1296         return err;
1297 }
1298
1299 static int kcm_getsockopt(struct socket *sock, int level, int optname,
1300                           char __user *optval, int __user *optlen)
1301 {
1302         struct kcm_sock *kcm = kcm_sk(sock->sk);
1303         int val, len;
1304
1305         if (level != SOL_KCM)
1306                 return -ENOPROTOOPT;
1307
1308         if (get_user(len, optlen))
1309                 return -EFAULT;
1310
1311         len = min_t(unsigned int, len, sizeof(int));
1312         if (len < 0)
1313                 return -EINVAL;
1314
1315         switch (optname) {
1316         case KCM_RECV_DISABLE:
1317                 val = kcm->rx_disabled;
1318                 break;
1319         default:
1320                 return -ENOPROTOOPT;
1321         }
1322
1323         if (put_user(len, optlen))
1324                 return -EFAULT;
1325         if (copy_to_user(optval, &val, len))
1326                 return -EFAULT;
1327         return 0;
1328 }
1329
1330 static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1331 {
1332         struct kcm_sock *tkcm;
1333         struct list_head *head;
1334         int index = 0;
1335
1336         /* For SOCK_SEQPACKET sock type, datagram_poll checks the sk_state, so
1337          * we set sk_state, otherwise epoll_wait always returns right away with
1338          * EPOLLHUP
1339          */
1340         kcm->sk.sk_state = TCP_ESTABLISHED;
1341
1342         /* Add to mux's kcm sockets list */
1343         kcm->mux = mux;
1344         spin_lock_bh(&mux->lock);
1345
1346         head = &mux->kcm_socks;
1347         list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1348                 if (tkcm->index != index)
1349                         break;
1350                 head = &tkcm->kcm_sock_list;
1351                 index++;
1352         }
1353
1354         list_add(&kcm->kcm_sock_list, head);
1355         kcm->index = index;
1356
1357         mux->kcm_socks_cnt++;
1358         spin_unlock_bh(&mux->lock);
1359
1360         INIT_WORK(&kcm->tx_work, kcm_tx_work);
1361
1362         spin_lock_bh(&mux->rx_lock);
1363         kcm_rcv_ready(kcm);
1364         spin_unlock_bh(&mux->rx_lock);
1365 }
1366
1367 static int kcm_attach(struct socket *sock, struct socket *csock,
1368                       struct bpf_prog *prog)
1369 {
1370         struct kcm_sock *kcm = kcm_sk(sock->sk);
1371         struct kcm_mux *mux = kcm->mux;
1372         struct sock *csk;
1373         struct kcm_psock *psock = NULL, *tpsock;
1374         struct list_head *head;
1375         int index = 0;
1376         static const struct strp_callbacks cb = {
1377                 .rcv_msg = kcm_rcv_strparser,
1378                 .parse_msg = kcm_parse_func_strparser,
1379                 .read_sock_done = kcm_read_sock_done,
1380         };
1381         int err = 0;
1382
1383         csk = csock->sk;
1384         if (!csk)
1385                 return -EINVAL;
1386
1387         lock_sock(csk);
1388
1389         /* Only allow TCP sockets to be attached for now */
1390         if ((csk->sk_family != AF_INET && csk->sk_family != AF_INET6) ||
1391             csk->sk_protocol != IPPROTO_TCP) {
1392                 err = -EOPNOTSUPP;
1393                 goto out;
1394         }
1395
1396         /* Don't allow listeners or closed sockets */
1397         if (csk->sk_state == TCP_LISTEN || csk->sk_state == TCP_CLOSE) {
1398                 err = -EOPNOTSUPP;
1399                 goto out;
1400         }
1401
1402         psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1403         if (!psock) {
1404                 err = -ENOMEM;
1405                 goto out;
1406         }
1407
1408         psock->mux = mux;
1409         psock->sk = csk;
1410         psock->bpf_prog = prog;
1411
1412         err = strp_init(&psock->strp, csk, &cb);
1413         if (err) {
1414                 kmem_cache_free(kcm_psockp, psock);
1415                 goto out;
1416         }
1417
1418         write_lock_bh(&csk->sk_callback_lock);
1419
1420         /* Check if sk_user_data is aready by KCM or someone else.
1421          * Must be done under lock to prevent race conditions.
1422          */
1423         if (csk->sk_user_data) {
1424                 write_unlock_bh(&csk->sk_callback_lock);
1425                 strp_stop(&psock->strp);
1426                 strp_done(&psock->strp);
1427                 kmem_cache_free(kcm_psockp, psock);
1428                 err = -EALREADY;
1429                 goto out;
1430         }
1431
1432         psock->save_data_ready = csk->sk_data_ready;
1433         psock->save_write_space = csk->sk_write_space;
1434         psock->save_state_change = csk->sk_state_change;
1435         csk->sk_user_data = psock;
1436         csk->sk_data_ready = psock_data_ready;
1437         csk->sk_write_space = psock_write_space;
1438         csk->sk_state_change = psock_state_change;
1439
1440         write_unlock_bh(&csk->sk_callback_lock);
1441
1442         sock_hold(csk);
1443
1444         /* Finished initialization, now add the psock to the MUX. */
1445         spin_lock_bh(&mux->lock);
1446         head = &mux->psocks;
1447         list_for_each_entry(tpsock, &mux->psocks, psock_list) {
1448                 if (tpsock->index != index)
1449                         break;
1450                 head = &tpsock->psock_list;
1451                 index++;
1452         }
1453
1454         list_add(&psock->psock_list, head);
1455         psock->index = index;
1456
1457         KCM_STATS_INCR(mux->stats.psock_attach);
1458         mux->psocks_cnt++;
1459         psock_now_avail(psock);
1460         spin_unlock_bh(&mux->lock);
1461
1462         /* Schedule RX work in case there are already bytes queued */
1463         strp_check_rcv(&psock->strp);
1464
1465 out:
1466         release_sock(csk);
1467
1468         return err;
1469 }
1470
1471 static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1472 {
1473         struct socket *csock;
1474         struct bpf_prog *prog;
1475         int err;
1476
1477         csock = sockfd_lookup(info->fd, &err);
1478         if (!csock)
1479                 return -ENOENT;
1480
1481         prog = bpf_prog_get_type(info->bpf_fd, BPF_PROG_TYPE_SOCKET_FILTER);
1482         if (IS_ERR(prog)) {
1483                 err = PTR_ERR(prog);
1484                 goto out;
1485         }
1486
1487         err = kcm_attach(sock, csock, prog);
1488         if (err) {
1489                 bpf_prog_put(prog);
1490                 goto out;
1491         }
1492
1493         /* Keep reference on file also */
1494
1495         return 0;
1496 out:
1497         fput(csock->file);
1498         return err;
1499 }
1500
1501 static void kcm_unattach(struct kcm_psock *psock)
1502 {
1503         struct sock *csk = psock->sk;
1504         struct kcm_mux *mux = psock->mux;
1505
1506         lock_sock(csk);
1507
1508         /* Stop getting callbacks from TCP socket. After this there should
1509          * be no way to reserve a kcm for this psock.
1510          */
1511         write_lock_bh(&csk->sk_callback_lock);
1512         csk->sk_user_data = NULL;
1513         csk->sk_data_ready = psock->save_data_ready;
1514         csk->sk_write_space = psock->save_write_space;
1515         csk->sk_state_change = psock->save_state_change;
1516         strp_stop(&psock->strp);
1517
1518         if (WARN_ON(psock->rx_kcm)) {
1519                 write_unlock_bh(&csk->sk_callback_lock);
1520                 release_sock(csk);
1521                 return;
1522         }
1523
1524         spin_lock_bh(&mux->rx_lock);
1525
1526         /* Stop receiver activities. After this point psock should not be
1527          * able to get onto ready list either through callbacks or work.
1528          */
1529         if (psock->ready_rx_msg) {
1530                 list_del(&psock->psock_ready_list);
1531                 kfree_skb(psock->ready_rx_msg);
1532                 psock->ready_rx_msg = NULL;
1533                 KCM_STATS_INCR(mux->stats.rx_ready_drops);
1534         }
1535
1536         spin_unlock_bh(&mux->rx_lock);
1537
1538         write_unlock_bh(&csk->sk_callback_lock);
1539
1540         /* Call strp_done without sock lock */
1541         release_sock(csk);
1542         strp_done(&psock->strp);
1543         lock_sock(csk);
1544
1545         bpf_prog_put(psock->bpf_prog);
1546
1547         spin_lock_bh(&mux->lock);
1548
1549         aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1550         save_strp_stats(&psock->strp, &mux->aggregate_strp_stats);
1551
1552         KCM_STATS_INCR(mux->stats.psock_unattach);
1553
1554         if (psock->tx_kcm) {
1555                 /* psock was reserved.  Just mark it finished and we will clean
1556                  * up in the kcm paths, we need kcm lock which can not be
1557                  * acquired here.
1558                  */
1559                 KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1560                 spin_unlock_bh(&mux->lock);
1561
1562                 /* We are unattaching a socket that is reserved. Abort the
1563                  * socket since we may be out of sync in sending on it. We need
1564                  * to do this without the mux lock.
1565                  */
1566                 kcm_abort_tx_psock(psock, EPIPE, false);
1567
1568                 spin_lock_bh(&mux->lock);
1569                 if (!psock->tx_kcm) {
1570                         /* psock now unreserved in window mux was unlocked */
1571                         goto no_reserved;
1572                 }
1573                 psock->done = 1;
1574
1575                 /* Commit done before queuing work to process it */
1576                 smp_mb();
1577
1578                 /* Queue tx work to make sure psock->done is handled */
1579                 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
1580                 spin_unlock_bh(&mux->lock);
1581         } else {
1582 no_reserved:
1583                 if (!psock->tx_stopped)
1584                         list_del(&psock->psock_avail_list);
1585                 list_del(&psock->psock_list);
1586                 mux->psocks_cnt--;
1587                 spin_unlock_bh(&mux->lock);
1588
1589                 sock_put(csk);
1590                 fput(csk->sk_socket->file);
1591                 kmem_cache_free(kcm_psockp, psock);
1592         }
1593
1594         release_sock(csk);
1595 }
1596
1597 static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1598 {
1599         struct kcm_sock *kcm = kcm_sk(sock->sk);
1600         struct kcm_mux *mux = kcm->mux;
1601         struct kcm_psock *psock;
1602         struct socket *csock;
1603         struct sock *csk;
1604         int err;
1605
1606         csock = sockfd_lookup(info->fd, &err);
1607         if (!csock)
1608                 return -ENOENT;
1609
1610         csk = csock->sk;
1611         if (!csk) {
1612                 err = -EINVAL;
1613                 goto out;
1614         }
1615
1616         err = -ENOENT;
1617
1618         spin_lock_bh(&mux->lock);
1619
1620         list_for_each_entry(psock, &mux->psocks, psock_list) {
1621                 if (psock->sk != csk)
1622                         continue;
1623
1624                 /* Found the matching psock */
1625
1626                 if (psock->unattaching || WARN_ON(psock->done)) {
1627                         err = -EALREADY;
1628                         break;
1629                 }
1630
1631                 psock->unattaching = 1;
1632
1633                 spin_unlock_bh(&mux->lock);
1634
1635                 /* Lower socket lock should already be held */
1636                 kcm_unattach(psock);
1637
1638                 err = 0;
1639                 goto out;
1640         }
1641
1642         spin_unlock_bh(&mux->lock);
1643
1644 out:
1645         fput(csock->file);
1646         return err;
1647 }
1648
1649 static struct proto kcm_proto = {
1650         .name   = "KCM",
1651         .owner  = THIS_MODULE,
1652         .obj_size = sizeof(struct kcm_sock),
1653 };
1654
1655 /* Clone a kcm socket. */
1656 static struct file *kcm_clone(struct socket *osock)
1657 {
1658         struct socket *newsock;
1659         struct sock *newsk;
1660
1661         newsock = sock_alloc();
1662         if (!newsock)
1663                 return ERR_PTR(-ENFILE);
1664
1665         newsock->type = osock->type;
1666         newsock->ops = osock->ops;
1667
1668         __module_get(newsock->ops->owner);
1669
1670         newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1671                          &kcm_proto, false);
1672         if (!newsk) {
1673                 sock_release(newsock);
1674                 return ERR_PTR(-ENOMEM);
1675         }
1676         sock_init_data(newsock, newsk);
1677         init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1678
1679         return sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1680 }
1681
1682 static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
1683 {
1684         int err;
1685
1686         switch (cmd) {
1687         case SIOCKCMATTACH: {
1688                 struct kcm_attach info;
1689
1690                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1691                         return -EFAULT;
1692
1693                 err = kcm_attach_ioctl(sock, &info);
1694
1695                 break;
1696         }
1697         case SIOCKCMUNATTACH: {
1698                 struct kcm_unattach info;
1699
1700                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1701                         return -EFAULT;
1702
1703                 err = kcm_unattach_ioctl(sock, &info);
1704
1705                 break;
1706         }
1707         case SIOCKCMCLONE: {
1708                 struct kcm_clone info;
1709                 struct file *file;
1710
1711                 info.fd = get_unused_fd_flags(0);
1712                 if (unlikely(info.fd < 0))
1713                         return info.fd;
1714
1715                 file = kcm_clone(sock);
1716                 if (IS_ERR(file)) {
1717                         put_unused_fd(info.fd);
1718                         return PTR_ERR(file);
1719                 }
1720                 if (copy_to_user((void __user *)arg, &info,
1721                                  sizeof(info))) {
1722                         put_unused_fd(info.fd);
1723                         fput(file);
1724                         return -EFAULT;
1725                 }
1726                 fd_install(info.fd, file);
1727                 err = 0;
1728                 break;
1729         }
1730         default:
1731                 err = -ENOIOCTLCMD;
1732                 break;
1733         }
1734
1735         return err;
1736 }
1737
1738 static void free_mux(struct rcu_head *rcu)
1739 {
1740         struct kcm_mux *mux = container_of(rcu,
1741             struct kcm_mux, rcu);
1742
1743         kmem_cache_free(kcm_muxp, mux);
1744 }
1745
1746 static void release_mux(struct kcm_mux *mux)
1747 {
1748         struct kcm_net *knet = mux->knet;
1749         struct kcm_psock *psock, *tmp_psock;
1750
1751         /* Release psocks */
1752         list_for_each_entry_safe(psock, tmp_psock,
1753                                  &mux->psocks, psock_list) {
1754                 if (!WARN_ON(psock->unattaching))
1755                         kcm_unattach(psock);
1756         }
1757
1758         if (WARN_ON(mux->psocks_cnt))
1759                 return;
1760
1761         __skb_queue_purge(&mux->rx_hold_queue);
1762
1763         mutex_lock(&knet->mutex);
1764         aggregate_mux_stats(&mux->stats, &knet->aggregate_mux_stats);
1765         aggregate_psock_stats(&mux->aggregate_psock_stats,
1766                               &knet->aggregate_psock_stats);
1767         aggregate_strp_stats(&mux->aggregate_strp_stats,
1768                              &knet->aggregate_strp_stats);
1769         list_del_rcu(&mux->kcm_mux_list);
1770         knet->count--;
1771         mutex_unlock(&knet->mutex);
1772
1773         call_rcu(&mux->rcu, free_mux);
1774 }
1775
1776 static void kcm_done(struct kcm_sock *kcm)
1777 {
1778         struct kcm_mux *mux = kcm->mux;
1779         struct sock *sk = &kcm->sk;
1780         int socks_cnt;
1781
1782         spin_lock_bh(&mux->rx_lock);
1783         if (kcm->rx_psock) {
1784                 /* Cleanup in unreserve_rx_kcm */
1785                 WARN_ON(kcm->done);
1786                 kcm->rx_disabled = 1;
1787                 kcm->done = 1;
1788                 spin_unlock_bh(&mux->rx_lock);
1789                 return;
1790         }
1791
1792         if (kcm->rx_wait) {
1793                 list_del(&kcm->wait_rx_list);
1794                 kcm->rx_wait = false;
1795         }
1796         /* Move any pending receive messages to other kcm sockets */
1797         requeue_rx_msgs(mux, &sk->sk_receive_queue);
1798
1799         spin_unlock_bh(&mux->rx_lock);
1800
1801         if (WARN_ON(sk_rmem_alloc_get(sk)))
1802                 return;
1803
1804         /* Detach from MUX */
1805         spin_lock_bh(&mux->lock);
1806
1807         list_del(&kcm->kcm_sock_list);
1808         mux->kcm_socks_cnt--;
1809         socks_cnt = mux->kcm_socks_cnt;
1810
1811         spin_unlock_bh(&mux->lock);
1812
1813         if (!socks_cnt) {
1814                 /* We are done with the mux now. */
1815                 release_mux(mux);
1816         }
1817
1818         WARN_ON(kcm->rx_wait);
1819
1820         sock_put(&kcm->sk);
1821 }
1822
1823 /* Called by kcm_release to close a KCM socket.
1824  * If this is the last KCM socket on the MUX, destroy the MUX.
1825  */
1826 static int kcm_release(struct socket *sock)
1827 {
1828         struct sock *sk = sock->sk;
1829         struct kcm_sock *kcm;
1830         struct kcm_mux *mux;
1831         struct kcm_psock *psock;
1832
1833         if (!sk)
1834                 return 0;
1835
1836         kcm = kcm_sk(sk);
1837         mux = kcm->mux;
1838
1839         sock_orphan(sk);
1840         kfree_skb(kcm->seq_skb);
1841
1842         lock_sock(sk);
1843         /* Purge queue under lock to avoid race condition with tx_work trying
1844          * to act when queue is nonempty. If tx_work runs after this point
1845          * it will just return.
1846          */
1847         __skb_queue_purge(&sk->sk_write_queue);
1848
1849         /* Set tx_stopped. This is checked when psock is bound to a kcm and we
1850          * get a writespace callback. This prevents further work being queued
1851          * from the callback (unbinding the psock occurs after canceling work.
1852          */
1853         kcm->tx_stopped = 1;
1854
1855         release_sock(sk);
1856
1857         spin_lock_bh(&mux->lock);
1858         if (kcm->tx_wait) {
1859                 /* Take of tx_wait list, after this point there should be no way
1860                  * that a psock will be assigned to this kcm.
1861                  */
1862                 list_del(&kcm->wait_psock_list);
1863                 kcm->tx_wait = false;
1864         }
1865         spin_unlock_bh(&mux->lock);
1866
1867         /* Cancel work. After this point there should be no outside references
1868          * to the kcm socket.
1869          */
1870         cancel_work_sync(&kcm->tx_work);
1871
1872         lock_sock(sk);
1873         psock = kcm->tx_psock;
1874         if (psock) {
1875                 /* A psock was reserved, so we need to kill it since it
1876                  * may already have some bytes queued from a message. We
1877                  * need to do this after removing kcm from tx_wait list.
1878                  */
1879                 kcm_abort_tx_psock(psock, EPIPE, false);
1880                 unreserve_psock(kcm);
1881         }
1882         release_sock(sk);
1883
1884         WARN_ON(kcm->tx_wait);
1885         WARN_ON(kcm->tx_psock);
1886
1887         sock->sk = NULL;
1888
1889         kcm_done(kcm);
1890
1891         return 0;
1892 }
1893
1894 static const struct proto_ops kcm_dgram_ops = {
1895         .family =       PF_KCM,
1896         .owner =        THIS_MODULE,
1897         .release =      kcm_release,
1898         .bind =         sock_no_bind,
1899         .connect =      sock_no_connect,
1900         .socketpair =   sock_no_socketpair,
1901         .accept =       sock_no_accept,
1902         .getname =      sock_no_getname,
1903         .poll =         datagram_poll,
1904         .ioctl =        kcm_ioctl,
1905         .listen =       sock_no_listen,
1906         .shutdown =     sock_no_shutdown,
1907         .setsockopt =   kcm_setsockopt,
1908         .getsockopt =   kcm_getsockopt,
1909         .sendmsg =      kcm_sendmsg,
1910         .recvmsg =      kcm_recvmsg,
1911         .mmap =         sock_no_mmap,
1912         .sendpage =     kcm_sendpage,
1913 };
1914
1915 static const struct proto_ops kcm_seqpacket_ops = {
1916         .family =       PF_KCM,
1917         .owner =        THIS_MODULE,
1918         .release =      kcm_release,
1919         .bind =         sock_no_bind,
1920         .connect =      sock_no_connect,
1921         .socketpair =   sock_no_socketpair,
1922         .accept =       sock_no_accept,
1923         .getname =      sock_no_getname,
1924         .poll =         datagram_poll,
1925         .ioctl =        kcm_ioctl,
1926         .listen =       sock_no_listen,
1927         .shutdown =     sock_no_shutdown,
1928         .setsockopt =   kcm_setsockopt,
1929         .getsockopt =   kcm_getsockopt,
1930         .sendmsg =      kcm_sendmsg,
1931         .recvmsg =      kcm_recvmsg,
1932         .mmap =         sock_no_mmap,
1933         .sendpage =     kcm_sendpage,
1934         .splice_read =  kcm_splice_read,
1935 };
1936
1937 /* Create proto operation for kcm sockets */
1938 static int kcm_create(struct net *net, struct socket *sock,
1939                       int protocol, int kern)
1940 {
1941         struct kcm_net *knet = net_generic(net, kcm_net_id);
1942         struct sock *sk;
1943         struct kcm_mux *mux;
1944
1945         switch (sock->type) {
1946         case SOCK_DGRAM:
1947                 sock->ops = &kcm_dgram_ops;
1948                 break;
1949         case SOCK_SEQPACKET:
1950                 sock->ops = &kcm_seqpacket_ops;
1951                 break;
1952         default:
1953                 return -ESOCKTNOSUPPORT;
1954         }
1955
1956         if (protocol != KCMPROTO_CONNECTED)
1957                 return -EPROTONOSUPPORT;
1958
1959         sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
1960         if (!sk)
1961                 return -ENOMEM;
1962
1963         /* Allocate a kcm mux, shared between KCM sockets */
1964         mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
1965         if (!mux) {
1966                 sk_free(sk);
1967                 return -ENOMEM;
1968         }
1969
1970         spin_lock_init(&mux->lock);
1971         spin_lock_init(&mux->rx_lock);
1972         INIT_LIST_HEAD(&mux->kcm_socks);
1973         INIT_LIST_HEAD(&mux->kcm_rx_waiters);
1974         INIT_LIST_HEAD(&mux->kcm_tx_waiters);
1975
1976         INIT_LIST_HEAD(&mux->psocks);
1977         INIT_LIST_HEAD(&mux->psocks_ready);
1978         INIT_LIST_HEAD(&mux->psocks_avail);
1979
1980         mux->knet = knet;
1981
1982         /* Add new MUX to list */
1983         mutex_lock(&knet->mutex);
1984         list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
1985         knet->count++;
1986         mutex_unlock(&knet->mutex);
1987
1988         skb_queue_head_init(&mux->rx_hold_queue);
1989
1990         /* Init KCM socket */
1991         sock_init_data(sock, sk);
1992         init_kcm_sock(kcm_sk(sk), mux);
1993
1994         return 0;
1995 }
1996
1997 static const struct net_proto_family kcm_family_ops = {
1998         .family = PF_KCM,
1999         .create = kcm_create,
2000         .owner  = THIS_MODULE,
2001 };
2002
2003 static __net_init int kcm_init_net(struct net *net)
2004 {
2005         struct kcm_net *knet = net_generic(net, kcm_net_id);
2006
2007         INIT_LIST_HEAD_RCU(&knet->mux_list);
2008         mutex_init(&knet->mutex);
2009
2010         return 0;
2011 }
2012
2013 static __net_exit void kcm_exit_net(struct net *net)
2014 {
2015         struct kcm_net *knet = net_generic(net, kcm_net_id);
2016
2017         /* All KCM sockets should be closed at this point, which should mean
2018          * that all multiplexors and psocks have been destroyed.
2019          */
2020         WARN_ON(!list_empty(&knet->mux_list));
2021 }
2022
2023 static struct pernet_operations kcm_net_ops = {
2024         .init = kcm_init_net,
2025         .exit = kcm_exit_net,
2026         .id   = &kcm_net_id,
2027         .size = sizeof(struct kcm_net),
2028 };
2029
2030 static int __init kcm_init(void)
2031 {
2032         int err = -ENOMEM;
2033
2034         kcm_muxp = kmem_cache_create("kcm_mux_cache",
2035                                      sizeof(struct kcm_mux), 0,
2036                                      SLAB_HWCACHE_ALIGN, NULL);
2037         if (!kcm_muxp)
2038                 goto fail;
2039
2040         kcm_psockp = kmem_cache_create("kcm_psock_cache",
2041                                        sizeof(struct kcm_psock), 0,
2042                                         SLAB_HWCACHE_ALIGN, NULL);
2043         if (!kcm_psockp)
2044                 goto fail;
2045
2046         kcm_wq = create_singlethread_workqueue("kkcmd");
2047         if (!kcm_wq)
2048                 goto fail;
2049
2050         err = proto_register(&kcm_proto, 1);
2051         if (err)
2052                 goto fail;
2053
2054         err = register_pernet_device(&kcm_net_ops);
2055         if (err)
2056                 goto net_ops_fail;
2057
2058         err = sock_register(&kcm_family_ops);
2059         if (err)
2060                 goto sock_register_fail;
2061
2062         err = kcm_proc_init();
2063         if (err)
2064                 goto proc_init_fail;
2065
2066         return 0;
2067
2068 proc_init_fail:
2069         sock_unregister(PF_KCM);
2070
2071 sock_register_fail:
2072         unregister_pernet_device(&kcm_net_ops);
2073
2074 net_ops_fail:
2075         proto_unregister(&kcm_proto);
2076
2077 fail:
2078         kmem_cache_destroy(kcm_muxp);
2079         kmem_cache_destroy(kcm_psockp);
2080
2081         if (kcm_wq)
2082                 destroy_workqueue(kcm_wq);
2083
2084         return err;
2085 }
2086
2087 static void __exit kcm_exit(void)
2088 {
2089         kcm_proc_exit();
2090         sock_unregister(PF_KCM);
2091         unregister_pernet_device(&kcm_net_ops);
2092         proto_unregister(&kcm_proto);
2093         destroy_workqueue(kcm_wq);
2094
2095         kmem_cache_destroy(kcm_muxp);
2096         kmem_cache_destroy(kcm_psockp);
2097 }
2098
2099 module_init(kcm_init);
2100 module_exit(kcm_exit);
2101
2102 MODULE_LICENSE("GPL");
2103 MODULE_ALIAS_NETPROTO(PF_KCM);