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         int res;
 382
 383         res = bpf_prog_run_pin_on_cpu(prog, skb);
 384         return res;
 385 }
 386
 387 static int kcm_read_sock_done(struct strparser *strp, int err)
 388 {
 389         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 390
 391         unreserve_rx_kcm(psock, true);
 392
 393         return err;
 394 }
 395
 396 static void psock_state_change(struct sock *sk)
 397 {
 398         /* TCP only does a EPOLLIN for a half close. Do a EPOLLHUP here
 399          * since application will normally not poll with EPOLLIN
 400          * on the TCP sockets.
 401          */
 402
 403         report_csk_error(sk, EPIPE);
 404 }
 405
 406 static void psock_write_space(struct sock *sk)
 407 {
 408         struct kcm_psock *psock;
 409         struct kcm_mux *mux;
 410         struct kcm_sock *kcm;
 411
 412         read_lock_bh(&sk->sk_callback_lock);
 413
 414         psock = (struct kcm_psock *)sk->sk_user_data;
 415         if (unlikely(!psock))
 416                 goto out;
 417         mux = psock->mux;
 418
 419         spin_lock_bh(&mux->lock);
 420
 421         /* Check if the socket is reserved so someone is waiting for sending. */
 422         kcm = psock->tx_kcm;
 423         if (kcm && !unlikely(kcm->tx_stopped))
 424                 queue_work(kcm_wq, &kcm->tx_work);
 425
 426         spin_unlock_bh(&mux->lock);
 427 out:
 428         read_unlock_bh(&sk->sk_callback_lock);
 429 }
 430
 431 static void unreserve_psock(struct kcm_sock *kcm);
 432
 433 /* kcm sock is locked. */
 434 static struct kcm_psock *reserve_psock(struct kcm_sock *kcm)
 435 {
 436         struct kcm_mux *mux = kcm->mux;
 437         struct kcm_psock *psock;
 438
 439         psock = kcm->tx_psock;
 440
 441         smp_rmb(); /* Must read tx_psock before tx_wait */
 442
 443         if (psock) {
 444                 WARN_ON(kcm->tx_wait);
 445                 if (unlikely(psock->tx_stopped))
 446                         unreserve_psock(kcm);
 447                 else
 448                         return kcm->tx_psock;
 449         }
 450
 451         spin_lock_bh(&mux->lock);
 452
 453         /* Check again under lock to see if psock was reserved for this
 454          * psock via psock_unreserve.
 455          */
 456         psock = kcm->tx_psock;
 457         if (unlikely(psock)) {
 458                 WARN_ON(kcm->tx_wait);
 459                 spin_unlock_bh(&mux->lock);
 460                 return kcm->tx_psock;
 461         }
 462
 463         if (!list_empty(&mux->psocks_avail)) {
 464                 psock = list_first_entry(&mux->psocks_avail,
 465                                          struct kcm_psock,
 466                                          psock_avail_list);
 467                 list_del(&psock->psock_avail_list);
 468                 if (kcm->tx_wait) {
 469                         list_del(&kcm->wait_psock_list);
 470                         kcm->tx_wait = false;
 471                 }
 472                 kcm->tx_psock = psock;
 473                 psock->tx_kcm = kcm;
 474                 KCM_STATS_INCR(psock->stats.reserved);
 475         } else if (!kcm->tx_wait) {
 476                 list_add_tail(&kcm->wait_psock_list,
 477                               &mux->kcm_tx_waiters);
 478                 kcm->tx_wait = true;
 479         }
 480
 481         spin_unlock_bh(&mux->lock);
 482
 483         return psock;
 484 }
 485
 486 /* mux lock held */
 487 static void psock_now_avail(struct kcm_psock *psock)
 488 {
 489         struct kcm_mux *mux = psock->mux;
 490         struct kcm_sock *kcm;
 491
 492         if (list_empty(&mux->kcm_tx_waiters)) {
 493                 list_add_tail(&psock->psock_avail_list,
 494                               &mux->psocks_avail);
 495         } else {
 496                 kcm = list_first_entry(&mux->kcm_tx_waiters,
 497                                        struct kcm_sock,
 498                                        wait_psock_list);
 499                 list_del(&kcm->wait_psock_list);
 500                 kcm->tx_wait = false;
 501                 psock->tx_kcm = kcm;
 502
 503                 /* Commit before changing tx_psock since that is read in
 504                  * reserve_psock before queuing work.
 505                  */
 506                 smp_mb();
 507
 508                 kcm->tx_psock = psock;
 509                 KCM_STATS_INCR(psock->stats.reserved);
 510                 queue_work(kcm_wq, &kcm->tx_work);
 511         }
 512 }
 513
 514 /* kcm sock is locked. */
 515 static void unreserve_psock(struct kcm_sock *kcm)
 516 {
 517         struct kcm_psock *psock;
 518         struct kcm_mux *mux = kcm->mux;
 519
 520         spin_lock_bh(&mux->lock);
 521
 522         psock = kcm->tx_psock;
 523
 524         if (WARN_ON(!psock)) {
 525                 spin_unlock_bh(&mux->lock);
 526                 return;
 527         }
 528
 529         smp_rmb(); /* Read tx_psock before tx_wait */
 530
 531         kcm_update_tx_mux_stats(mux, psock);
 532
 533         WARN_ON(kcm->tx_wait);
 534
 535         kcm->tx_psock = NULL;
 536         psock->tx_kcm = NULL;
 537         KCM_STATS_INCR(psock->stats.unreserved);
 538
 539         if (unlikely(psock->tx_stopped)) {
 540                 if (psock->done) {
 541                         /* Deferred free */
 542                         list_del(&psock->psock_list);
 543                         mux->psocks_cnt--;
 544                         sock_put(psock->sk);
 545                         fput(psock->sk->sk_socket->file);
 546                         kmem_cache_free(kcm_psockp, psock);
 547                 }
 548
 549                 /* Don't put back on available list */
 550
 551                 spin_unlock_bh(&mux->lock);
 552
 553                 return;
 554         }
 555
 556         psock_now_avail(psock);
 557
 558         spin_unlock_bh(&mux->lock);
 559 }
 560
 561 static void kcm_report_tx_retry(struct kcm_sock *kcm)
 562 {
 563         struct kcm_mux *mux = kcm->mux;
 564
 565         spin_lock_bh(&mux->lock);
 566         KCM_STATS_INCR(mux->stats.tx_retries);
 567         spin_unlock_bh(&mux->lock);
 568 }
 569
 570 /* Write any messages ready on the kcm socket.  Called with kcm sock lock
 571  * held.  Return bytes actually sent or error.
 572  */
 573 static int kcm_write_msgs(struct kcm_sock *kcm)
 574 {
 575         struct sock *sk = &kcm->sk;
 576         struct kcm_psock *psock;
 577         struct sk_buff *skb, *head;
 578         struct kcm_tx_msg *txm;
 579         unsigned short fragidx, frag_offset;
 580         unsigned int sent, total_sent = 0;
 581         int ret = 0;
 582
 583         kcm->tx_wait_more = false;
 584         psock = kcm->tx_psock;
 585         if (unlikely(psock && psock->tx_stopped)) {
 586                 /* A reserved psock was aborted asynchronously. Unreserve
 587                  * it and we'll retry the message.
 588                  */
 589                 unreserve_psock(kcm);
 590                 kcm_report_tx_retry(kcm);
 591                 if (skb_queue_empty(&sk->sk_write_queue))
 592                         return 0;
 593
 594                 kcm_tx_msg(skb_peek(&sk->sk_write_queue))->sent = 0;
 595
 596         } else if (skb_queue_empty(&sk->sk_write_queue)) {
 597                 return 0;
 598         }
 599
 600         head = skb_peek(&sk->sk_write_queue);
 601         txm = kcm_tx_msg(head);
 602
 603         if (txm->sent) {
 604                 /* Send of first skbuff in queue already in progress */
 605                 if (WARN_ON(!psock)) {
 606                         ret = -EINVAL;
 607                         goto out;
 608                 }
 609                 sent = txm->sent;
 610                 frag_offset = txm->frag_offset;
 611                 fragidx = txm->fragidx;
 612                 skb = txm->frag_skb;
 613
 614                 goto do_frag;
 615         }
 616
 617 try_again:
 618         psock = reserve_psock(kcm);
 619         if (!psock)
 620                 goto out;
 621
 622         do {
 623                 skb = head;
 624                 txm = kcm_tx_msg(head);
 625                 sent = 0;
 626
 627 do_frag_list:
 628                 if (WARN_ON(!skb_shinfo(skb)->nr_frags)) {
 629                         ret = -EINVAL;
 630                         goto out;
 631                 }
 632
 633                 for (fragidx = 0; fragidx < skb_shinfo(skb)->nr_frags;
 634                      fragidx++) {
 635                         skb_frag_t *frag;
 636
 637                         frag_offset = 0;
 638 do_frag:
 639                         frag = &skb_shinfo(skb)->frags[fragidx];
 640                         if (WARN_ON(!skb_frag_size(frag))) {
 641                                 ret = -EINVAL;
 642                                 goto out;
 643                         }
 644
 645                         ret = kernel_sendpage(psock->sk->sk_socket,
 646                                               skb_frag_page(frag),
 647                                               skb_frag_off(frag) + frag_offset,
 648                                               skb_frag_size(frag) - frag_offset,
 649                                               MSG_DONTWAIT);
 650                         if (ret <= 0) {
 651                                 if (ret == -EAGAIN) {
 652                                         /* Save state to try again when there's
 653                                          * write space on the socket
 654                                          */
 655                                         txm->sent = sent;
 656                                         txm->frag_offset = frag_offset;
 657                                         txm->fragidx = fragidx;
 658                                         txm->frag_skb = skb;
 659
 660                                         ret = 0;
 661                                         goto out;
 662                                 }
 663
 664                                 /* Hard failure in sending message, abort this
 665                                  * psock since it has lost framing
 666                                  * synchronization and retry sending the
 667                                  * message from the beginning.
 668                                  */
 669                                 kcm_abort_tx_psock(psock, ret ? -ret : EPIPE,
 670                                                    true);
 671                                 unreserve_psock(kcm);
 672
 673                                 txm->sent = 0;
 674                                 kcm_report_tx_retry(kcm);
 675                                 ret = 0;
 676
 677                                 goto try_again;
 678                         }
 679
 680                         sent += ret;
 681                         frag_offset += ret;
 682                         KCM_STATS_ADD(psock->stats.tx_bytes, ret);
 683                         if (frag_offset < skb_frag_size(frag)) {
 684                                 /* Not finished with this frag */
 685                                 goto do_frag;
 686                         }
 687                 }
 688
 689                 if (skb == head) {
 690                         if (skb_has_frag_list(skb)) {
 691                                 skb = skb_shinfo(skb)->frag_list;
 692                                 goto do_frag_list;
 693                         }
 694                 } else if (skb->next) {
 695                         skb = skb->next;
 696                         goto do_frag_list;
 697                 }
 698
 699                 /* Successfully sent the whole packet, account for it. */
 700                 skb_dequeue(&sk->sk_write_queue);
 701                 kfree_skb(head);
 702                 sk->sk_wmem_queued -= sent;
 703                 total_sent += sent;
 704                 KCM_STATS_INCR(psock->stats.tx_msgs);
 705         } while ((head = skb_peek(&sk->sk_write_queue)));
 706 out:
 707         if (!head) {
 708                 /* Done with all queued messages. */
 709                 WARN_ON(!skb_queue_empty(&sk->sk_write_queue));
 710                 unreserve_psock(kcm);
 711         }
 712
 713         /* Check if write space is available */
 714         sk->sk_write_space(sk);
 715
 716         return total_sent ? : ret;
 717 }
 718
 719 static void kcm_tx_work(struct work_struct *w)
 720 {
 721         struct kcm_sock *kcm = container_of(w, struct kcm_sock, tx_work);
 722         struct sock *sk = &kcm->sk;
 723         int err;
 724
 725         lock_sock(sk);
 726
 727         /* Primarily for SOCK_DGRAM sockets, also handle asynchronous tx
 728          * aborts
 729          */
 730         err = kcm_write_msgs(kcm);
 731         if (err < 0) {
 732                 /* Hard failure in write, report error on KCM socket */
 733                 pr_warn("KCM: Hard failure on kcm_write_msgs %d\n", err);
 734                 report_csk_error(&kcm->sk, -err);
 735                 goto out;
 736         }
 737
 738         /* Primarily for SOCK_SEQPACKET sockets */
 739         if (likely(sk->sk_socket) &&
 740             test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
 741                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 742                 sk->sk_write_space(sk);
 743         }
 744
 745 out:
 746         release_sock(sk);
 747 }
 748
 749 static void kcm_push(struct kcm_sock *kcm)
 750 {
 751         if (kcm->tx_wait_more)
 752                 kcm_write_msgs(kcm);
 753 }
 754
 755 static ssize_t kcm_sendpage(struct socket *sock, struct page *page,
 756                             int offset, size_t size, int flags)
 757
 758 {
 759         struct sock *sk = sock->sk;
 760         struct kcm_sock *kcm = kcm_sk(sk);
 761         struct sk_buff *skb = NULL, *head = NULL;
 762         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 763         bool eor;
 764         int err = 0;
 765         int i;
 766
 767         if (flags & MSG_SENDPAGE_NOTLAST)
 768                 flags |= MSG_MORE;
 769
 770         /* No MSG_EOR from splice, only look at MSG_MORE */
 771         eor = !(flags & MSG_MORE);
 772
 773         lock_sock(sk);
 774
 775         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 776
 777         err = -EPIPE;
 778         if (sk->sk_err)
 779                 goto out_error;
 780
 781         if (kcm->seq_skb) {
 782                 /* Previously opened message */
 783                 head = kcm->seq_skb;
 784                 skb = kcm_tx_msg(head)->last_skb;
 785                 i = skb_shinfo(skb)->nr_frags;
 786
 787                 if (skb_can_coalesce(skb, i, page, offset)) {
 788                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], size);
 789                         skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
 790                         goto coalesced;
 791                 }
 792
 793                 if (i >= MAX_SKB_FRAGS) {
 794                         struct sk_buff *tskb;
 795
 796                         tskb = alloc_skb(0, sk->sk_allocation);
 797                         while (!tskb) {
 798                                 kcm_push(kcm);
 799                                 err = sk_stream_wait_memory(sk, &timeo);
 800                                 if (err)
 801                                         goto out_error;
 802                         }
 803
 804                         if (head == skb)
 805                                 skb_shinfo(head)->frag_list = tskb;
 806                         else
 807                                 skb->next = tskb;
 808
 809                         skb = tskb;
 810                         skb->ip_summed = CHECKSUM_UNNECESSARY;
 811                         i = 0;
 812                 }
 813         } else {
 814                 /* Call the sk_stream functions to manage the sndbuf mem. */
 815                 if (!sk_stream_memory_free(sk)) {
 816                         kcm_push(kcm);
 817                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 818                         err = sk_stream_wait_memory(sk, &timeo);
 819                         if (err)
 820                                 goto out_error;
 821                 }
 822
 823                 head = alloc_skb(0, sk->sk_allocation);
 824                 while (!head) {
 825                         kcm_push(kcm);
 826                         err = sk_stream_wait_memory(sk, &timeo);
 827                         if (err)
 828                                 goto out_error;
 829                 }
 830
 831                 skb = head;
 832                 i = 0;
 833         }
 834
 835         get_page(page);
 836         skb_fill_page_desc(skb, i, page, offset, size);
 837         skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
 838
 839 coalesced:
 840         skb->len += size;
 841         skb->data_len += size;
 842         skb->truesize += size;
 843         sk->sk_wmem_queued += size;
 844         sk_mem_charge(sk, size);
 845
 846         if (head != skb) {
 847                 head->len += size;
 848                 head->data_len += size;
 849                 head->truesize += size;
 850         }
 851
 852         if (eor) {
 853                 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
 854
 855                 /* Message complete, queue it on send buffer */
 856                 __skb_queue_tail(&sk->sk_write_queue, head);
 857                 kcm->seq_skb = NULL;
 858                 KCM_STATS_INCR(kcm->stats.tx_msgs);
 859
 860                 if (flags & MSG_BATCH) {
 861                         kcm->tx_wait_more = true;
 862                 } else if (kcm->tx_wait_more || not_busy) {
 863                         err = kcm_write_msgs(kcm);
 864                         if (err < 0) {
 865                                 /* We got a hard error in write_msgs but have
 866                                  * already queued this message. Report an error
 867                                  * in the socket, but don't affect return value
 868                                  * from sendmsg
 869                                  */
 870                                 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
 871                                 report_csk_error(&kcm->sk, -err);
 872                         }
 873                 }
 874         } else {
 875                 /* Message not complete, save state */
 876                 kcm->seq_skb = head;
 877                 kcm_tx_msg(head)->last_skb = skb;
 878         }
 879
 880         KCM_STATS_ADD(kcm->stats.tx_bytes, size);
 881
 882         release_sock(sk);
 883         return size;
 884
 885 out_error:
 886         kcm_push(kcm);
 887
 888         err = sk_stream_error(sk, flags, err);
 889
 890         /* make sure we wake any epoll edge trigger waiter */
 891         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
 892                 sk->sk_write_space(sk);
 893
 894         release_sock(sk);
 895         return err;
 896 }
 897
 898 static int kcm_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
 899 {
 900         struct sock *sk = sock->sk;
 901         struct kcm_sock *kcm = kcm_sk(sk);
 902         struct sk_buff *skb = NULL, *head = NULL;
 903         size_t copy, copied = 0;
 904         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 905         int eor = (sock->type == SOCK_DGRAM) ?
 906                   !(msg->msg_flags & MSG_MORE) : !!(msg->msg_flags & MSG_EOR);
 907         int err = -EPIPE;
 908
 909         lock_sock(sk);
 910
 911         /* Per tcp_sendmsg this should be in poll */
 912         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 913
 914         if (sk->sk_err)
 915                 goto out_error;
 916
 917         if (kcm->seq_skb) {
 918                 /* Previously opened message */
 919                 head = kcm->seq_skb;
 920                 skb = kcm_tx_msg(head)->last_skb;
 921                 goto start;
 922         }
 923
 924         /* Call the sk_stream functions to manage the sndbuf mem. */
 925         if (!sk_stream_memory_free(sk)) {
 926                 kcm_push(kcm);
 927                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 928                 err = sk_stream_wait_memory(sk, &timeo);
 929                 if (err)
 930                         goto out_error;
 931         }
 932
 933         if (msg_data_left(msg)) {
 934                 /* New message, alloc head skb */
 935                 head = alloc_skb(0, sk->sk_allocation);
 936                 while (!head) {
 937                         kcm_push(kcm);
 938                         err = sk_stream_wait_memory(sk, &timeo);
 939                         if (err)
 940                                 goto out_error;
 941
 942                         head = alloc_skb(0, sk->sk_allocation);
 943                 }
 944
 945                 skb = head;
 946
 947                 /* Set ip_summed to CHECKSUM_UNNECESSARY to avoid calling
 948                  * csum_and_copy_from_iter from skb_do_copy_data_nocache.
 949                  */
 950                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 951         }
 952
 953 start:
 954         while (msg_data_left(msg)) {
 955                 bool merge = true;
 956                 int i = skb_shinfo(skb)->nr_frags;
 957                 struct page_frag *pfrag = sk_page_frag(sk);
 958
 959                 if (!sk_page_frag_refill(sk, pfrag))
 960                         goto wait_for_memory;
 961
 962                 if (!skb_can_coalesce(skb, i, pfrag->page,
 963                                       pfrag->offset)) {
 964                         if (i == MAX_SKB_FRAGS) {
 965                                 struct sk_buff *tskb;
 966
 967                                 tskb = alloc_skb(0, sk->sk_allocation);
 968                                 if (!tskb)
 969                                         goto wait_for_memory;
 970
 971                                 if (head == skb)
 972                                         skb_shinfo(head)->frag_list = tskb;
 973                                 else
 974                                         skb->next = tskb;
 975
 976                                 skb = tskb;
 977                                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 978                                 continue;
 979                         }
 980                         merge = false;
 981                 }
 982
 983                 copy = min_t(int, msg_data_left(msg),
 984                              pfrag->size - pfrag->offset);
 985
 986                 if (!sk_wmem_schedule(sk, copy))
 987                         goto wait_for_memory;
 988
 989                 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
 990                                                pfrag->page,
 991                                                pfrag->offset,
 992                                                copy);
 993                 if (err)
 994                         goto out_error;
 995
 996                 /* Update the skb. */
 997                 if (merge) {
 998                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
 999                 } else {
1000                         skb_fill_page_desc(skb, i, pfrag->page,
1001                                            pfrag->offset, copy);
1002                         get_page(pfrag->page);
1003                 }
1004
1005                 pfrag->offset += copy;
1006                 copied += copy;
1007                 if (head != skb) {
1008                         head->len += copy;
1009                         head->data_len += copy;
1010                 }
1011
1012                 continue;
1013
1014 wait_for_memory:
1015                 kcm_push(kcm);
1016                 err = sk_stream_wait_memory(sk, &timeo);
1017                 if (err)
1018                         goto out_error;
1019         }
1020
1021         if (eor) {
1022                 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
1023
1024                 if (head) {
1025                         /* Message complete, queue it on send buffer */
1026                         __skb_queue_tail(&sk->sk_write_queue, head);
1027                         kcm->seq_skb = NULL;
1028                         KCM_STATS_INCR(kcm->stats.tx_msgs);
1029                 }
1030
1031                 if (msg->msg_flags & MSG_BATCH) {
1032                         kcm->tx_wait_more = true;
1033                 } else if (kcm->tx_wait_more || not_busy) {
1034                         err = kcm_write_msgs(kcm);
1035                         if (err < 0) {
1036                                 /* We got a hard error in write_msgs but have
1037                                  * already queued this message. Report an error
1038                                  * in the socket, but don't affect return value
1039                                  * from sendmsg
1040                                  */
1041                                 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
1042                                 report_csk_error(&kcm->sk, -err);
1043                         }
1044                 }
1045         } else {
1046                 /* Message not complete, save state */
1047 partial_message:
1048                 if (head) {
1049                         kcm->seq_skb = head;
1050                         kcm_tx_msg(head)->last_skb = skb;
1051                 }
1052         }
1053
1054         KCM_STATS_ADD(kcm->stats.tx_bytes, copied);
1055
1056         release_sock(sk);
1057         return copied;
1058
1059 out_error:
1060         kcm_push(kcm);
1061
1062         if (copied && sock->type == SOCK_SEQPACKET) {
1063                 /* Wrote some bytes before encountering an
1064                  * error, return partial success.
1065                  */
1066                 goto partial_message;
1067         }
1068
1069         if (skb_has_frag_list(head)) {
1070                 kfree_skb_list(skb_shinfo(head)->frag_list);
1071                 skb_shinfo(head)->frag_list = NULL;
1072         }
1073
1074         if (head != kcm->seq_skb)
1075                 kfree_skb(head);
1076
1077         err = sk_stream_error(sk, msg->msg_flags, err);
1078
1079         /* make sure we wake any epoll edge trigger waiter */
1080         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1081                 sk->sk_write_space(sk);
1082
1083         release_sock(sk);
1084         return err;
1085 }
1086
1087 static struct sk_buff *kcm_wait_data(struct sock *sk, int flags,
1088                                      long timeo, int *err)
1089 {
1090         struct sk_buff *skb;
1091
1092         while (!(skb = skb_peek(&sk->sk_receive_queue))) {
1093                 if (sk->sk_err) {
1094                         *err = sock_error(sk);
1095                         return NULL;
1096                 }
1097
1098                 if (sock_flag(sk, SOCK_DONE))
1099                         return NULL;
1100
1101                 if ((flags & MSG_DONTWAIT) || !timeo) {
1102                         *err = -EAGAIN;
1103                         return NULL;
1104                 }
1105
1106                 sk_wait_data(sk, &timeo, NULL);
1107
1108                 /* Handle signals */
1109                 if (signal_pending(current)) {
1110                         *err = sock_intr_errno(timeo);
1111                         return NULL;
1112                 }
1113         }
1114
1115         return skb;
1116 }
1117
1118 static int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
1119                        size_t len, int flags)
1120 {
1121         struct sock *sk = sock->sk;
1122         struct kcm_sock *kcm = kcm_sk(sk);
1123         int err = 0;
1124         long timeo;
1125         struct strp_msg *stm;
1126         int copied = 0;
1127         struct sk_buff *skb;
1128
1129         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1130
1131         lock_sock(sk);
1132
1133         skb = kcm_wait_data(sk, flags, timeo, &err);
1134         if (!skb)
1135                 goto out;
1136
1137         /* Okay, have a message on the receive queue */
1138
1139         stm = strp_msg(skb);
1140
1141         if (len > stm->full_len)
1142                 len = stm->full_len;
1143
1144         err = skb_copy_datagram_msg(skb, stm->offset, msg, len);
1145         if (err < 0)
1146                 goto out;
1147
1148         copied = len;
1149         if (likely(!(flags & MSG_PEEK))) {
1150                 KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1151                 if (copied < stm->full_len) {
1152                         if (sock->type == SOCK_DGRAM) {
1153                                 /* Truncated message */
1154                                 msg->msg_flags |= MSG_TRUNC;
1155                                 goto msg_finished;
1156                         }
1157                         stm->offset += copied;
1158                         stm->full_len -= copied;
1159                 } else {
1160 msg_finished:
1161                         /* Finished with message */
1162                         msg->msg_flags |= MSG_EOR;
1163                         KCM_STATS_INCR(kcm->stats.rx_msgs);
1164                         skb_unlink(skb, &sk->sk_receive_queue);
1165                         kfree_skb(skb);
1166                 }
1167         }
1168
1169 out:
1170         release_sock(sk);
1171
1172         return copied ? : err;
1173 }
1174
1175 static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1176                                struct pipe_inode_info *pipe, size_t len,
1177                                unsigned int flags)
1178 {
1179         struct sock *sk = sock->sk;
1180         struct kcm_sock *kcm = kcm_sk(sk);
1181         long timeo;
1182         struct strp_msg *stm;
1183         int err = 0;
1184         ssize_t copied;
1185         struct sk_buff *skb;
1186
1187         /* Only support splice for SOCKSEQPACKET */
1188
1189         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1190
1191         lock_sock(sk);
1192
1193         skb = kcm_wait_data(sk, flags, timeo, &err);
1194         if (!skb)
1195                 goto err_out;
1196
1197         /* Okay, have a message on the receive queue */
1198
1199         stm = strp_msg(skb);
1200
1201         if (len > stm->full_len)
1202                 len = stm->full_len;
1203
1204         copied = skb_splice_bits(skb, sk, stm->offset, pipe, len, flags);
1205         if (copied < 0) {
1206                 err = copied;
1207                 goto err_out;
1208         }
1209
1210         KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1211
1212         stm->offset += copied;
1213         stm->full_len -= copied;
1214
1215         /* We have no way to return MSG_EOR. If all the bytes have been
1216          * read we still leave the message in the receive socket buffer.
1217          * A subsequent recvmsg needs to be done to return MSG_EOR and
1218          * finish reading the message.
1219          */
1220
1221         release_sock(sk);
1222
1223         return copied;
1224
1225 err_out:
1226         release_sock(sk);
1227
1228         return err;
1229 }
1230
1231 /* kcm sock lock held */
1232 static void kcm_recv_disable(struct kcm_sock *kcm)
1233 {
1234         struct kcm_mux *mux = kcm->mux;
1235
1236         if (kcm->rx_disabled)
1237                 return;
1238
1239         spin_lock_bh(&mux->rx_lock);
1240
1241         kcm->rx_disabled = 1;
1242
1243         /* If a psock is reserved we'll do cleanup in unreserve */
1244         if (!kcm->rx_psock) {
1245                 if (kcm->rx_wait) {
1246                         list_del(&kcm->wait_rx_list);
1247                         kcm->rx_wait = false;
1248                 }
1249
1250                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1251         }
1252
1253         spin_unlock_bh(&mux->rx_lock);
1254 }
1255
1256 /* kcm sock lock held */
1257 static void kcm_recv_enable(struct kcm_sock *kcm)
1258 {
1259         struct kcm_mux *mux = kcm->mux;
1260
1261         if (!kcm->rx_disabled)
1262                 return;
1263
1264         spin_lock_bh(&mux->rx_lock);
1265
1266         kcm->rx_disabled = 0;
1267         kcm_rcv_ready(kcm);
1268
1269         spin_unlock_bh(&mux->rx_lock);
1270 }
1271
1272 static int kcm_setsockopt(struct socket *sock, int level, int optname,
1273                           sockptr_t optval, unsigned int optlen)
1274 {
1275         struct kcm_sock *kcm = kcm_sk(sock->sk);
1276         int val, valbool;
1277         int err = 0;
1278
1279         if (level != SOL_KCM)
1280                 return -ENOPROTOOPT;
1281
1282         if (optlen < sizeof(int))
1283                 return -EINVAL;
1284
1285         if (copy_from_sockptr(&val, optval, sizeof(int)))
1286                 return -EFAULT;
1287
1288         valbool = val ? 1 : 0;
1289
1290         switch (optname) {
1291         case KCM_RECV_DISABLE:
1292                 lock_sock(&kcm->sk);
1293                 if (valbool)
1294                         kcm_recv_disable(kcm);
1295                 else
1296                         kcm_recv_enable(kcm);
1297                 release_sock(&kcm->sk);
1298                 break;
1299         default:
1300                 err = -ENOPROTOOPT;
1301         }
1302
1303         return err;
1304 }
1305
1306 static int kcm_getsockopt(struct socket *sock, int level, int optname,
1307                           char __user *optval, int __user *optlen)
1308 {
1309         struct kcm_sock *kcm = kcm_sk(sock->sk);
1310         int val, len;
1311
1312         if (level != SOL_KCM)
1313                 return -ENOPROTOOPT;
1314
1315         if (get_user(len, optlen))
1316                 return -EFAULT;
1317
1318         len = min_t(unsigned int, len, sizeof(int));
1319         if (len < 0)
1320                 return -EINVAL;
1321
1322         switch (optname) {
1323         case KCM_RECV_DISABLE:
1324                 val = kcm->rx_disabled;
1325                 break;
1326         default:
1327                 return -ENOPROTOOPT;
1328         }
1329
1330         if (put_user(len, optlen))
1331                 return -EFAULT;
1332         if (copy_to_user(optval, &val, len))
1333                 return -EFAULT;
1334         return 0;
1335 }
1336
1337 static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1338 {
1339         struct kcm_sock *tkcm;
1340         struct list_head *head;
1341         int index = 0;
1342
1343         /* For SOCK_SEQPACKET sock type, datagram_poll checks the sk_state, so
1344          * we set sk_state, otherwise epoll_wait always returns right away with
1345          * EPOLLHUP
1346          */
1347         kcm->sk.sk_state = TCP_ESTABLISHED;
1348
1349         /* Add to mux's kcm sockets list */
1350         kcm->mux = mux;
1351         spin_lock_bh(&mux->lock);
1352
1353         head = &mux->kcm_socks;
1354         list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1355                 if (tkcm->index != index)
1356                         break;
1357                 head = &tkcm->kcm_sock_list;
1358                 index++;
1359         }
1360
1361         list_add(&kcm->kcm_sock_list, head);
1362         kcm->index = index;
1363
1364         mux->kcm_socks_cnt++;
1365         spin_unlock_bh(&mux->lock);
1366
1367         INIT_WORK(&kcm->tx_work, kcm_tx_work);
1368
1369         spin_lock_bh(&mux->rx_lock);
1370         kcm_rcv_ready(kcm);
1371         spin_unlock_bh(&mux->rx_lock);
1372 }
1373
1374 static int kcm_attach(struct socket *sock, struct socket *csock,
1375                       struct bpf_prog *prog)
1376 {
1377         struct kcm_sock *kcm = kcm_sk(sock->sk);
1378         struct kcm_mux *mux = kcm->mux;
1379         struct sock *csk;
1380         struct kcm_psock *psock = NULL, *tpsock;
1381         struct list_head *head;
1382         int index = 0;
1383         static const struct strp_callbacks cb = {
1384                 .rcv_msg = kcm_rcv_strparser,
1385                 .parse_msg = kcm_parse_func_strparser,
1386                 .read_sock_done = kcm_read_sock_done,
1387         };
1388         int err = 0;
1389
1390         csk = csock->sk;
1391         if (!csk)
1392                 return -EINVAL;
1393
1394         lock_sock(csk);
1395
1396         /* Only allow TCP sockets to be attached for now */
1397         if ((csk->sk_family != AF_INET && csk->sk_family != AF_INET6) ||
1398             csk->sk_protocol != IPPROTO_TCP) {
1399                 err = -EOPNOTSUPP;
1400                 goto out;
1401         }
1402
1403         /* Don't allow listeners or closed sockets */
1404         if (csk->sk_state == TCP_LISTEN || csk->sk_state == TCP_CLOSE) {
1405                 err = -EOPNOTSUPP;
1406                 goto out;
1407         }
1408
1409         psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1410         if (!psock) {
1411                 err = -ENOMEM;
1412                 goto out;
1413         }
1414
1415         psock->mux = mux;
1416         psock->sk = csk;
1417         psock->bpf_prog = prog;
1418
1419         err = strp_init(&psock->strp, csk, &cb);
1420         if (err) {
1421                 kmem_cache_free(kcm_psockp, psock);
1422                 goto out;
1423         }
1424
1425         write_lock_bh(&csk->sk_callback_lock);
1426
1427         /* Check if sk_user_data is already by KCM or someone else.
1428          * Must be done under lock to prevent race conditions.
1429          */
1430         if (csk->sk_user_data) {
1431                 write_unlock_bh(&csk->sk_callback_lock);
1432                 strp_stop(&psock->strp);
1433                 strp_done(&psock->strp);
1434                 kmem_cache_free(kcm_psockp, psock);
1435                 err = -EALREADY;
1436                 goto out;
1437         }
1438
1439         psock->save_data_ready = csk->sk_data_ready;
1440         psock->save_write_space = csk->sk_write_space;
1441         psock->save_state_change = csk->sk_state_change;
1442         csk->sk_user_data = psock;
1443         csk->sk_data_ready = psock_data_ready;
1444         csk->sk_write_space = psock_write_space;
1445         csk->sk_state_change = psock_state_change;
1446
1447         write_unlock_bh(&csk->sk_callback_lock);
1448
1449         sock_hold(csk);
1450
1451         /* Finished initialization, now add the psock to the MUX. */
1452         spin_lock_bh(&mux->lock);
1453         head = &mux->psocks;
1454         list_for_each_entry(tpsock, &mux->psocks, psock_list) {
1455                 if (tpsock->index != index)
1456                         break;
1457                 head = &tpsock->psock_list;
1458                 index++;
1459         }
1460
1461         list_add(&psock->psock_list, head);
1462         psock->index = index;
1463
1464         KCM_STATS_INCR(mux->stats.psock_attach);
1465         mux->psocks_cnt++;
1466         psock_now_avail(psock);
1467         spin_unlock_bh(&mux->lock);
1468
1469         /* Schedule RX work in case there are already bytes queued */
1470         strp_check_rcv(&psock->strp);
1471
1472 out:
1473         release_sock(csk);
1474
1475         return err;
1476 }
1477
1478 static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1479 {
1480         struct socket *csock;
1481         struct bpf_prog *prog;
1482         int err;
1483
1484         csock = sockfd_lookup(info->fd, &err);
1485         if (!csock)
1486                 return -ENOENT;
1487
1488         prog = bpf_prog_get_type(info->bpf_fd, BPF_PROG_TYPE_SOCKET_FILTER);
1489         if (IS_ERR(prog)) {
1490                 err = PTR_ERR(prog);
1491                 goto out;
1492         }
1493
1494         err = kcm_attach(sock, csock, prog);
1495         if (err) {
1496                 bpf_prog_put(prog);
1497                 goto out;
1498         }
1499
1500         /* Keep reference on file also */
1501
1502         return 0;
1503 out:
1504         sockfd_put(csock);
1505         return err;
1506 }
1507
1508 static void kcm_unattach(struct kcm_psock *psock)
1509 {
1510         struct sock *csk = psock->sk;
1511         struct kcm_mux *mux = psock->mux;
1512
1513         lock_sock(csk);
1514
1515         /* Stop getting callbacks from TCP socket. After this there should
1516          * be no way to reserve a kcm for this psock.
1517          */
1518         write_lock_bh(&csk->sk_callback_lock);
1519         csk->sk_user_data = NULL;
1520         csk->sk_data_ready = psock->save_data_ready;
1521         csk->sk_write_space = psock->save_write_space;
1522         csk->sk_state_change = psock->save_state_change;
1523         strp_stop(&psock->strp);
1524
1525         if (WARN_ON(psock->rx_kcm)) {
1526                 write_unlock_bh(&csk->sk_callback_lock);
1527                 release_sock(csk);
1528                 return;
1529         }
1530
1531         spin_lock_bh(&mux->rx_lock);
1532
1533         /* Stop receiver activities. After this point psock should not be
1534          * able to get onto ready list either through callbacks or work.
1535          */
1536         if (psock->ready_rx_msg) {
1537                 list_del(&psock->psock_ready_list);
1538                 kfree_skb(psock->ready_rx_msg);
1539                 psock->ready_rx_msg = NULL;
1540                 KCM_STATS_INCR(mux->stats.rx_ready_drops);
1541         }
1542
1543         spin_unlock_bh(&mux->rx_lock);
1544
1545         write_unlock_bh(&csk->sk_callback_lock);
1546
1547         /* Call strp_done without sock lock */
1548         release_sock(csk);
1549         strp_done(&psock->strp);
1550         lock_sock(csk);
1551
1552         bpf_prog_put(psock->bpf_prog);
1553
1554         spin_lock_bh(&mux->lock);
1555
1556         aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1557         save_strp_stats(&psock->strp, &mux->aggregate_strp_stats);
1558
1559         KCM_STATS_INCR(mux->stats.psock_unattach);
1560
1561         if (psock->tx_kcm) {
1562                 /* psock was reserved.  Just mark it finished and we will clean
1563                  * up in the kcm paths, we need kcm lock which can not be
1564                  * acquired here.
1565                  */
1566                 KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1567                 spin_unlock_bh(&mux->lock);
1568
1569                 /* We are unattaching a socket that is reserved. Abort the
1570                  * socket since we may be out of sync in sending on it. We need
1571                  * to do this without the mux lock.
1572                  */
1573                 kcm_abort_tx_psock(psock, EPIPE, false);
1574
1575                 spin_lock_bh(&mux->lock);
1576                 if (!psock->tx_kcm) {
1577                         /* psock now unreserved in window mux was unlocked */
1578                         goto no_reserved;
1579                 }
1580                 psock->done = 1;
1581
1582                 /* Commit done before queuing work to process it */
1583                 smp_mb();
1584
1585                 /* Queue tx work to make sure psock->done is handled */
1586                 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
1587                 spin_unlock_bh(&mux->lock);
1588         } else {
1589 no_reserved:
1590                 if (!psock->tx_stopped)
1591                         list_del(&psock->psock_avail_list);
1592                 list_del(&psock->psock_list);
1593                 mux->psocks_cnt--;
1594                 spin_unlock_bh(&mux->lock);
1595
1596                 sock_put(csk);
1597                 fput(csk->sk_socket->file);
1598                 kmem_cache_free(kcm_psockp, psock);
1599         }
1600
1601         release_sock(csk);
1602 }
1603
1604 static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1605 {
1606         struct kcm_sock *kcm = kcm_sk(sock->sk);
1607         struct kcm_mux *mux = kcm->mux;
1608         struct kcm_psock *psock;
1609         struct socket *csock;
1610         struct sock *csk;
1611         int err;
1612
1613         csock = sockfd_lookup(info->fd, &err);
1614         if (!csock)
1615                 return -ENOENT;
1616
1617         csk = csock->sk;
1618         if (!csk) {
1619                 err = -EINVAL;
1620                 goto out;
1621         }
1622
1623         err = -ENOENT;
1624
1625         spin_lock_bh(&mux->lock);
1626
1627         list_for_each_entry(psock, &mux->psocks, psock_list) {
1628                 if (psock->sk != csk)
1629                         continue;
1630
1631                 /* Found the matching psock */
1632
1633                 if (psock->unattaching || WARN_ON(psock->done)) {
1634                         err = -EALREADY;
1635                         break;
1636                 }
1637
1638                 psock->unattaching = 1;
1639
1640                 spin_unlock_bh(&mux->lock);
1641
1642                 /* Lower socket lock should already be held */
1643                 kcm_unattach(psock);
1644
1645                 err = 0;
1646                 goto out;
1647         }
1648
1649         spin_unlock_bh(&mux->lock);
1650
1651 out:
1652         sockfd_put(csock);
1653         return err;
1654 }
1655
1656 static struct proto kcm_proto = {
1657         .name   = "KCM",
1658         .owner  = THIS_MODULE,
1659         .obj_size = sizeof(struct kcm_sock),
1660 };
1661
1662 /* Clone a kcm socket. */
1663 static struct file *kcm_clone(struct socket *osock)
1664 {
1665         struct socket *newsock;
1666         struct sock *newsk;
1667
1668         newsock = sock_alloc();
1669         if (!newsock)
1670                 return ERR_PTR(-ENFILE);
1671
1672         newsock->type = osock->type;
1673         newsock->ops = osock->ops;
1674
1675         __module_get(newsock->ops->owner);
1676
1677         newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1678                          &kcm_proto, false);
1679         if (!newsk) {
1680                 sock_release(newsock);
1681                 return ERR_PTR(-ENOMEM);
1682         }
1683         sock_init_data(newsock, newsk);
1684         init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1685
1686         return sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1687 }
1688
1689 static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
1690 {
1691         int err;
1692
1693         switch (cmd) {
1694         case SIOCKCMATTACH: {
1695                 struct kcm_attach info;
1696
1697                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1698                         return -EFAULT;
1699
1700                 err = kcm_attach_ioctl(sock, &info);
1701
1702                 break;
1703         }
1704         case SIOCKCMUNATTACH: {
1705                 struct kcm_unattach info;
1706
1707                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1708                         return -EFAULT;
1709
1710                 err = kcm_unattach_ioctl(sock, &info);
1711
1712                 break;
1713         }
1714         case SIOCKCMCLONE: {
1715                 struct kcm_clone info;
1716                 struct file *file;
1717
1718                 info.fd = get_unused_fd_flags(0);
1719                 if (unlikely(info.fd < 0))
1720                         return info.fd;
1721
1722                 file = kcm_clone(sock);
1723                 if (IS_ERR(file)) {
1724                         put_unused_fd(info.fd);
1725                         return PTR_ERR(file);
1726                 }
1727                 if (copy_to_user((void __user *)arg, &info,
1728                                  sizeof(info))) {
1729                         put_unused_fd(info.fd);
1730                         fput(file);
1731                         return -EFAULT;
1732                 }
1733                 fd_install(info.fd, file);
1734                 err = 0;
1735                 break;
1736         }
1737         default:
1738                 err = -ENOIOCTLCMD;
1739                 break;
1740         }
1741
1742         return err;
1743 }
1744
1745 static void free_mux(struct rcu_head *rcu)
1746 {
1747         struct kcm_mux *mux = container_of(rcu,
1748             struct kcm_mux, rcu);
1749
1750         kmem_cache_free(kcm_muxp, mux);
1751 }
1752
1753 static void release_mux(struct kcm_mux *mux)
1754 {
1755         struct kcm_net *knet = mux->knet;
1756         struct kcm_psock *psock, *tmp_psock;
1757
1758         /* Release psocks */
1759         list_for_each_entry_safe(psock, tmp_psock,
1760                                  &mux->psocks, psock_list) {
1761                 if (!WARN_ON(psock->unattaching))
1762                         kcm_unattach(psock);
1763         }
1764
1765         if (WARN_ON(mux->psocks_cnt))
1766                 return;
1767
1768         __skb_queue_purge(&mux->rx_hold_queue);
1769
1770         mutex_lock(&knet->mutex);
1771         aggregate_mux_stats(&mux->stats, &knet->aggregate_mux_stats);
1772         aggregate_psock_stats(&mux->aggregate_psock_stats,
1773                               &knet->aggregate_psock_stats);
1774         aggregate_strp_stats(&mux->aggregate_strp_stats,
1775                              &knet->aggregate_strp_stats);
1776         list_del_rcu(&mux->kcm_mux_list);
1777         knet->count--;
1778         mutex_unlock(&knet->mutex);
1779
1780         call_rcu(&mux->rcu, free_mux);
1781 }
1782
1783 static void kcm_done(struct kcm_sock *kcm)
1784 {
1785         struct kcm_mux *mux = kcm->mux;
1786         struct sock *sk = &kcm->sk;
1787         int socks_cnt;
1788
1789         spin_lock_bh(&mux->rx_lock);
1790         if (kcm->rx_psock) {
1791                 /* Cleanup in unreserve_rx_kcm */
1792                 WARN_ON(kcm->done);
1793                 kcm->rx_disabled = 1;
1794                 kcm->done = 1;
1795                 spin_unlock_bh(&mux->rx_lock);
1796                 return;
1797         }
1798
1799         if (kcm->rx_wait) {
1800                 list_del(&kcm->wait_rx_list);
1801                 kcm->rx_wait = false;
1802         }
1803         /* Move any pending receive messages to other kcm sockets */
1804         requeue_rx_msgs(mux, &sk->sk_receive_queue);
1805
1806         spin_unlock_bh(&mux->rx_lock);
1807
1808         if (WARN_ON(sk_rmem_alloc_get(sk)))
1809                 return;
1810
1811         /* Detach from MUX */
1812         spin_lock_bh(&mux->lock);
1813
1814         list_del(&kcm->kcm_sock_list);
1815         mux->kcm_socks_cnt--;
1816         socks_cnt = mux->kcm_socks_cnt;
1817
1818         spin_unlock_bh(&mux->lock);
1819
1820         if (!socks_cnt) {
1821                 /* We are done with the mux now. */
1822                 release_mux(mux);
1823         }
1824
1825         WARN_ON(kcm->rx_wait);
1826
1827         sock_put(&kcm->sk);
1828 }
1829
1830 /* Called by kcm_release to close a KCM socket.
1831  * If this is the last KCM socket on the MUX, destroy the MUX.
1832  */
1833 static int kcm_release(struct socket *sock)
1834 {
1835         struct sock *sk = sock->sk;
1836         struct kcm_sock *kcm;
1837         struct kcm_mux *mux;
1838         struct kcm_psock *psock;
1839
1840         if (!sk)
1841                 return 0;
1842
1843         kcm = kcm_sk(sk);
1844         mux = kcm->mux;
1845
1846         sock_orphan(sk);
1847         kfree_skb(kcm->seq_skb);
1848
1849         lock_sock(sk);
1850         /* Purge queue under lock to avoid race condition with tx_work trying
1851          * to act when queue is nonempty. If tx_work runs after this point
1852          * it will just return.
1853          */
1854         __skb_queue_purge(&sk->sk_write_queue);
1855
1856         /* Set tx_stopped. This is checked when psock is bound to a kcm and we
1857          * get a writespace callback. This prevents further work being queued
1858          * from the callback (unbinding the psock occurs after canceling work.
1859          */
1860         kcm->tx_stopped = 1;
1861
1862         release_sock(sk);
1863
1864         spin_lock_bh(&mux->lock);
1865         if (kcm->tx_wait) {
1866                 /* Take of tx_wait list, after this point there should be no way
1867                  * that a psock will be assigned to this kcm.
1868                  */
1869                 list_del(&kcm->wait_psock_list);
1870                 kcm->tx_wait = false;
1871         }
1872         spin_unlock_bh(&mux->lock);
1873
1874         /* Cancel work. After this point there should be no outside references
1875          * to the kcm socket.
1876          */
1877         cancel_work_sync(&kcm->tx_work);
1878
1879         lock_sock(sk);
1880         psock = kcm->tx_psock;
1881         if (psock) {
1882                 /* A psock was reserved, so we need to kill it since it
1883                  * may already have some bytes queued from a message. We
1884                  * need to do this after removing kcm from tx_wait list.
1885                  */
1886                 kcm_abort_tx_psock(psock, EPIPE, false);
1887                 unreserve_psock(kcm);
1888         }
1889         release_sock(sk);
1890
1891         WARN_ON(kcm->tx_wait);
1892         WARN_ON(kcm->tx_psock);
1893
1894         sock->sk = NULL;
1895
1896         kcm_done(kcm);
1897
1898         return 0;
1899 }
1900
1901 static const struct proto_ops kcm_dgram_ops = {
1902         .family =       PF_KCM,
1903         .owner =        THIS_MODULE,
1904         .release =      kcm_release,
1905         .bind =         sock_no_bind,
1906         .connect =      sock_no_connect,
1907         .socketpair =   sock_no_socketpair,
1908         .accept =       sock_no_accept,
1909         .getname =      sock_no_getname,
1910         .poll =         datagram_poll,
1911         .ioctl =        kcm_ioctl,
1912         .listen =       sock_no_listen,
1913         .shutdown =     sock_no_shutdown,
1914         .setsockopt =   kcm_setsockopt,
1915         .getsockopt =   kcm_getsockopt,
1916         .sendmsg =      kcm_sendmsg,
1917         .recvmsg =      kcm_recvmsg,
1918         .mmap =         sock_no_mmap,
1919         .sendpage =     kcm_sendpage,
1920 };
1921
1922 static const struct proto_ops kcm_seqpacket_ops = {
1923         .family =       PF_KCM,
1924         .owner =        THIS_MODULE,
1925         .release =      kcm_release,
1926         .bind =         sock_no_bind,
1927         .connect =      sock_no_connect,
1928         .socketpair =   sock_no_socketpair,
1929         .accept =       sock_no_accept,
1930         .getname =      sock_no_getname,
1931         .poll =         datagram_poll,
1932         .ioctl =        kcm_ioctl,
1933         .listen =       sock_no_listen,
1934         .shutdown =     sock_no_shutdown,
1935         .setsockopt =   kcm_setsockopt,
1936         .getsockopt =   kcm_getsockopt,
1937         .sendmsg =      kcm_sendmsg,
1938         .recvmsg =      kcm_recvmsg,
1939         .mmap =         sock_no_mmap,
1940         .sendpage =     kcm_sendpage,
1941         .splice_read =  kcm_splice_read,
1942 };
1943
1944 /* Create proto operation for kcm sockets */
1945 static int kcm_create(struct net *net, struct socket *sock,
1946                       int protocol, int kern)
1947 {
1948         struct kcm_net *knet = net_generic(net, kcm_net_id);
1949         struct sock *sk;
1950         struct kcm_mux *mux;
1951
1952         switch (sock->type) {
1953         case SOCK_DGRAM:
1954                 sock->ops = &kcm_dgram_ops;
1955                 break;
1956         case SOCK_SEQPACKET:
1957                 sock->ops = &kcm_seqpacket_ops;
1958                 break;
1959         default:
1960                 return -ESOCKTNOSUPPORT;
1961         }
1962
1963         if (protocol != KCMPROTO_CONNECTED)
1964                 return -EPROTONOSUPPORT;
1965
1966         sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
1967         if (!sk)
1968                 return -ENOMEM;
1969
1970         /* Allocate a kcm mux, shared between KCM sockets */
1971         mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
1972         if (!mux) {
1973                 sk_free(sk);
1974                 return -ENOMEM;
1975         }
1976
1977         spin_lock_init(&mux->lock);
1978         spin_lock_init(&mux->rx_lock);
1979         INIT_LIST_HEAD(&mux->kcm_socks);
1980         INIT_LIST_HEAD(&mux->kcm_rx_waiters);
1981         INIT_LIST_HEAD(&mux->kcm_tx_waiters);
1982
1983         INIT_LIST_HEAD(&mux->psocks);
1984         INIT_LIST_HEAD(&mux->psocks_ready);
1985         INIT_LIST_HEAD(&mux->psocks_avail);
1986
1987         mux->knet = knet;
1988
1989         /* Add new MUX to list */
1990         mutex_lock(&knet->mutex);
1991         list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
1992         knet->count++;
1993         mutex_unlock(&knet->mutex);
1994
1995         skb_queue_head_init(&mux->rx_hold_queue);
1996
1997         /* Init KCM socket */
1998         sock_init_data(sock, sk);
1999         init_kcm_sock(kcm_sk(sk), mux);
2000
2001         return 0;
2002 }
2003
2004 static const struct net_proto_family kcm_family_ops = {
2005         .family = PF_KCM,
2006         .create = kcm_create,
2007         .owner  = THIS_MODULE,
2008 };
2009
2010 static __net_init int kcm_init_net(struct net *net)
2011 {
2012         struct kcm_net *knet = net_generic(net, kcm_net_id);
2013
2014         INIT_LIST_HEAD_RCU(&knet->mux_list);
2015         mutex_init(&knet->mutex);
2016
2017         return 0;
2018 }
2019
2020 static __net_exit void kcm_exit_net(struct net *net)
2021 {
2022         struct kcm_net *knet = net_generic(net, kcm_net_id);
2023
2024         /* All KCM sockets should be closed at this point, which should mean
2025          * that all multiplexors and psocks have been destroyed.
2026          */
2027         WARN_ON(!list_empty(&knet->mux_list));
2028 }
2029
2030 static struct pernet_operations kcm_net_ops = {
2031         .init = kcm_init_net,
2032         .exit = kcm_exit_net,
2033         .id   = &kcm_net_id,
2034         .size = sizeof(struct kcm_net),
2035 };
2036
2037 static int __init kcm_init(void)
2038 {
2039         int err = -ENOMEM;
2040
2041         kcm_muxp = kmem_cache_create("kcm_mux_cache",
2042                                      sizeof(struct kcm_mux), 0,
2043                                      SLAB_HWCACHE_ALIGN, NULL);
2044         if (!kcm_muxp)
2045                 goto fail;
2046
2047         kcm_psockp = kmem_cache_create("kcm_psock_cache",
2048                                        sizeof(struct kcm_psock), 0,
2049                                         SLAB_HWCACHE_ALIGN, NULL);
2050         if (!kcm_psockp)
2051                 goto fail;
2052
2053         kcm_wq = create_singlethread_workqueue("kkcmd");
2054         if (!kcm_wq)
2055                 goto fail;
2056
2057         err = proto_register(&kcm_proto, 1);
2058         if (err)
2059                 goto fail;
2060
2061         err = register_pernet_device(&kcm_net_ops);
2062         if (err)
2063                 goto net_ops_fail;
2064
2065         err = sock_register(&kcm_family_ops);
2066         if (err)
2067                 goto sock_register_fail;
2068
2069         err = kcm_proc_init();
2070         if (err)
2071                 goto proc_init_fail;
2072
2073         return 0;
2074
2075 proc_init_fail:
2076         sock_unregister(PF_KCM);
2077
2078 sock_register_fail:
2079         unregister_pernet_device(&kcm_net_ops);
2080
2081 net_ops_fail:
2082         proto_unregister(&kcm_proto);
2083
2084 fail:
2085         kmem_cache_destroy(kcm_muxp);
2086         kmem_cache_destroy(kcm_psockp);
2087
2088         if (kcm_wq)
2089                 destroy_workqueue(kcm_wq);
2090
2091         return err;
2092 }
2093
2094 static void __exit kcm_exit(void)
2095 {
2096         kcm_proc_exit();
2097         sock_unregister(PF_KCM);
2098         unregister_pernet_device(&kcm_net_ops);
2099         proto_unregister(&kcm_proto);
2100         destroy_workqueue(kcm_wq);
2101
2102         kmem_cache_destroy(kcm_muxp);
2103         kmem_cache_destroy(kcm_psockp);
2104 }
2105
2106 module_init(kcm_init);
2107 module_exit(kcm_exit);
2108
2109 MODULE_LICENSE("GPL");
2110 MODULE_ALIAS_NETPROTO(PF_KCM);