net/kcm/kcmsock.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Kernel Connection Multiplexor
   4  *
   5  * Copyright (c) 2016 Tom Herbert <tom@herbertland.com>
   6  */
   7
   8 #include <linux/bpf.h>
   9 #include <linux/errno.h>
  10 #include <linux/errqueue.h>
  11 #include <linux/file.h>
  12 #include <linux/filter.h>
  13 #include <linux/in.h>
  14 #include <linux/kernel.h>
  15 #include <linux/module.h>
  16 #include <linux/net.h>
  17 #include <linux/netdevice.h>
  18 #include <linux/poll.h>
  19 #include <linux/rculist.h>
  20 #include <linux/skbuff.h>
  21 #include <linux/socket.h>
  22 #include <linux/uaccess.h>
  23 #include <linux/workqueue.h>
  24 #include <linux/syscalls.h>
  25 #include <linux/sched/signal.h>
  26
  27 #include <net/kcm.h>
  28 #include <net/netns/generic.h>
  29 #include <net/sock.h>
  30 #include <uapi/linux/kcm.h>
  31 #include <trace/events/sock.h>
  32
  33 unsigned int kcm_net_id;
  34
  35 static struct kmem_cache *kcm_psockp __read_mostly;
  36 static struct kmem_cache *kcm_muxp __read_mostly;
  37 static struct workqueue_struct *kcm_wq;
  38
  39 static inline struct kcm_sock *kcm_sk(const struct sock *sk)
  40 {
  41         return (struct kcm_sock *)sk;
  42 }
  43
  44 static inline struct kcm_tx_msg *kcm_tx_msg(struct sk_buff *skb)
  45 {
  46         return (struct kcm_tx_msg *)skb->cb;
  47 }
  48
  49 static void report_csk_error(struct sock *csk, int err)
  50 {
  51         csk->sk_err = EPIPE;
  52         sk_error_report(csk);
  53 }
  54
  55 static void kcm_abort_tx_psock(struct kcm_psock *psock, int err,
  56                                bool wakeup_kcm)
  57 {
  58         struct sock *csk = psock->sk;
  59         struct kcm_mux *mux = psock->mux;
  60
  61         /* Unrecoverable error in transmit */
  62
  63         spin_lock_bh(&mux->lock);
  64
  65         if (psock->tx_stopped) {
  66                 spin_unlock_bh(&mux->lock);
  67                 return;
  68         }
  69
  70         psock->tx_stopped = 1;
  71         KCM_STATS_INCR(psock->stats.tx_aborts);
  72
  73         if (!psock->tx_kcm) {
  74                 /* Take off psocks_avail list */
  75                 list_del(&psock->psock_avail_list);
  76         } else if (wakeup_kcm) {
  77                 /* In this case psock is being aborted while outside of
  78                  * write_msgs and psock is reserved. Schedule tx_work
  79                  * to handle the failure there. Need to commit tx_stopped
  80                  * before queuing work.
  81                  */
  82                 smp_mb();
  83
  84                 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
  85         }
  86
  87         spin_unlock_bh(&mux->lock);
  88
  89         /* Report error on lower socket */
  90         report_csk_error(csk, err);
  91 }
  92
  93 /* RX mux lock held. */
  94 static void kcm_update_rx_mux_stats(struct kcm_mux *mux,
  95                                     struct kcm_psock *psock)
  96 {
  97         STRP_STATS_ADD(mux->stats.rx_bytes,
  98                        psock->strp.stats.bytes -
  99                        psock->saved_rx_bytes);
 100         mux->stats.rx_msgs +=
 101                 psock->strp.stats.msgs - psock->saved_rx_msgs;
 102         psock->saved_rx_msgs = psock->strp.stats.msgs;
 103         psock->saved_rx_bytes = psock->strp.stats.bytes;
 104 }
 105
 106 static void kcm_update_tx_mux_stats(struct kcm_mux *mux,
 107                                     struct kcm_psock *psock)
 108 {
 109         KCM_STATS_ADD(mux->stats.tx_bytes,
 110                       psock->stats.tx_bytes - psock->saved_tx_bytes);
 111         mux->stats.tx_msgs +=
 112                 psock->stats.tx_msgs - psock->saved_tx_msgs;
 113         psock->saved_tx_msgs = psock->stats.tx_msgs;
 114         psock->saved_tx_bytes = psock->stats.tx_bytes;
 115 }
 116
 117 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb);
 118
 119 /* KCM is ready to receive messages on its queue-- either the KCM is new or
 120  * has become unblocked after being blocked on full socket buffer. Queue any
 121  * pending ready messages on a psock. RX mux lock held.
 122  */
 123 static void kcm_rcv_ready(struct kcm_sock *kcm)
 124 {
 125         struct kcm_mux *mux = kcm->mux;
 126         struct kcm_psock *psock;
 127         struct sk_buff *skb;
 128
 129         if (unlikely(kcm->rx_wait || kcm->rx_psock || kcm->rx_disabled))
 130                 return;
 131
 132         while (unlikely((skb = __skb_dequeue(&mux->rx_hold_queue)))) {
 133                 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 134                         /* Assuming buffer limit has been reached */
 135                         skb_queue_head(&mux->rx_hold_queue, skb);
 136                         WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
 137                         return;
 138                 }
 139         }
 140
 141         while (!list_empty(&mux->psocks_ready)) {
 142                 psock = list_first_entry(&mux->psocks_ready, struct kcm_psock,
 143                                          psock_ready_list);
 144
 145                 if (kcm_queue_rcv_skb(&kcm->sk, psock->ready_rx_msg)) {
 146                         /* Assuming buffer limit has been reached */
 147                         WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
 148                         return;
 149                 }
 150
 151                 /* Consumed the ready message on the psock. Schedule rx_work to
 152                  * get more messages.
 153                  */
 154                 list_del(&psock->psock_ready_list);
 155                 psock->ready_rx_msg = NULL;
 156                 /* Commit clearing of ready_rx_msg for queuing work */
 157                 smp_mb();
 158
 159                 strp_unpause(&psock->strp);
 160                 strp_check_rcv(&psock->strp);
 161         }
 162
 163         /* Buffer limit is okay now, add to ready list */
 164         list_add_tail(&kcm->wait_rx_list,
 165                       &kcm->mux->kcm_rx_waiters);
 166         /* paired with lockless reads in kcm_rfree() */
 167         WRITE_ONCE(kcm->rx_wait, true);
 168 }
 169
 170 static void kcm_rfree(struct sk_buff *skb)
 171 {
 172         struct sock *sk = skb->sk;
 173         struct kcm_sock *kcm = kcm_sk(sk);
 174         struct kcm_mux *mux = kcm->mux;
 175         unsigned int len = skb->truesize;
 176
 177         sk_mem_uncharge(sk, len);
 178         atomic_sub(len, &sk->sk_rmem_alloc);
 179
 180         /* For reading rx_wait and rx_psock without holding lock */
 181         smp_mb__after_atomic();
 182
 183         if (!READ_ONCE(kcm->rx_wait) && !READ_ONCE(kcm->rx_psock) &&
 184             sk_rmem_alloc_get(sk) < sk->sk_rcvlowat) {
 185                 spin_lock_bh(&mux->rx_lock);
 186                 kcm_rcv_ready(kcm);
 187                 spin_unlock_bh(&mux->rx_lock);
 188         }
 189 }
 190
 191 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb)
 192 {
 193         struct sk_buff_head *list = &sk->sk_receive_queue;
 194
 195         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
 196                 return -ENOMEM;
 197
 198         if (!sk_rmem_schedule(sk, skb, skb->truesize))
 199                 return -ENOBUFS;
 200
 201         skb->dev = NULL;
 202
 203         skb_orphan(skb);
 204         skb->sk = sk;
 205         skb->destructor = kcm_rfree;
 206         atomic_add(skb->truesize, &sk->sk_rmem_alloc);
 207         sk_mem_charge(sk, skb->truesize);
 208
 209         skb_queue_tail(list, skb);
 210
 211         if (!sock_flag(sk, SOCK_DEAD))
 212                 sk->sk_data_ready(sk);
 213
 214         return 0;
 215 }
 216
 217 /* Requeue received messages for a kcm socket to other kcm sockets. This is
 218  * called with a kcm socket is receive disabled.
 219  * RX mux lock held.
 220  */
 221 static void requeue_rx_msgs(struct kcm_mux *mux, struct sk_buff_head *head)
 222 {
 223         struct sk_buff *skb;
 224         struct kcm_sock *kcm;
 225
 226         while ((skb = skb_dequeue(head))) {
 227                 /* Reset destructor to avoid calling kcm_rcv_ready */
 228                 skb->destructor = sock_rfree;
 229                 skb_orphan(skb);
 230 try_again:
 231                 if (list_empty(&mux->kcm_rx_waiters)) {
 232                         skb_queue_tail(&mux->rx_hold_queue, skb);
 233                         continue;
 234                 }
 235
 236                 kcm = list_first_entry(&mux->kcm_rx_waiters,
 237                                        struct kcm_sock, wait_rx_list);
 238
 239                 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 240                         /* Should mean socket buffer full */
 241                         list_del(&kcm->wait_rx_list);
 242                         /* paired with lockless reads in kcm_rfree() */
 243                         WRITE_ONCE(kcm->rx_wait, false);
 244
 245                         /* Commit rx_wait to read in kcm_free */
 246                         smp_wmb();
 247
 248                         goto try_again;
 249                 }
 250         }
 251 }
 252
 253 /* Lower sock lock held */
 254 static struct kcm_sock *reserve_rx_kcm(struct kcm_psock *psock,
 255                                        struct sk_buff *head)
 256 {
 257         struct kcm_mux *mux = psock->mux;
 258         struct kcm_sock *kcm;
 259
 260         WARN_ON(psock->ready_rx_msg);
 261
 262         if (psock->rx_kcm)
 263                 return psock->rx_kcm;
 264
 265         spin_lock_bh(&mux->rx_lock);
 266
 267         if (psock->rx_kcm) {
 268                 spin_unlock_bh(&mux->rx_lock);
 269                 return psock->rx_kcm;
 270         }
 271
 272         kcm_update_rx_mux_stats(mux, psock);
 273
 274         if (list_empty(&mux->kcm_rx_waiters)) {
 275                 psock->ready_rx_msg = head;
 276                 strp_pause(&psock->strp);
 277                 list_add_tail(&psock->psock_ready_list,
 278                               &mux->psocks_ready);
 279                 spin_unlock_bh(&mux->rx_lock);
 280                 return NULL;
 281         }
 282
 283         kcm = list_first_entry(&mux->kcm_rx_waiters,
 284                                struct kcm_sock, wait_rx_list);
 285         list_del(&kcm->wait_rx_list);
 286         /* paired with lockless reads in kcm_rfree() */
 287         WRITE_ONCE(kcm->rx_wait, false);
 288
 289         psock->rx_kcm = kcm;
 290         /* paired with lockless reads in kcm_rfree() */
 291         WRITE_ONCE(kcm->rx_psock, psock);
 292
 293         spin_unlock_bh(&mux->rx_lock);
 294
 295         return kcm;
 296 }
 297
 298 static void kcm_done(struct kcm_sock *kcm);
 299
 300 static void kcm_done_work(struct work_struct *w)
 301 {
 302         kcm_done(container_of(w, struct kcm_sock, done_work));
 303 }
 304
 305 /* Lower sock held */
 306 static void unreserve_rx_kcm(struct kcm_psock *psock,
 307                              bool rcv_ready)
 308 {
 309         struct kcm_sock *kcm = psock->rx_kcm;
 310         struct kcm_mux *mux = psock->mux;
 311
 312         if (!kcm)
 313                 return;
 314
 315         spin_lock_bh(&mux->rx_lock);
 316
 317         psock->rx_kcm = NULL;
 318         /* paired with lockless reads in kcm_rfree() */
 319         WRITE_ONCE(kcm->rx_psock, NULL);
 320
 321         /* Commit kcm->rx_psock before sk_rmem_alloc_get to sync with
 322          * kcm_rfree
 323          */
 324         smp_mb();
 325
 326         if (unlikely(kcm->done)) {
 327                 spin_unlock_bh(&mux->rx_lock);
 328
 329                 /* Need to run kcm_done in a task since we need to qcquire
 330                  * callback locks which may already be held here.
 331                  */
 332                 INIT_WORK(&kcm->done_work, kcm_done_work);
 333                 schedule_work(&kcm->done_work);
 334                 return;
 335         }
 336
 337         if (unlikely(kcm->rx_disabled)) {
 338                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
 339         } else if (rcv_ready || unlikely(!sk_rmem_alloc_get(&kcm->sk))) {
 340                 /* Check for degenerative race with rx_wait that all
 341                  * data was dequeued (accounted for in kcm_rfree).
 342                  */
 343                 kcm_rcv_ready(kcm);
 344         }
 345         spin_unlock_bh(&mux->rx_lock);
 346 }
 347
 348 /* Lower sock lock held */
 349 static void psock_data_ready(struct sock *sk)
 350 {
 351         struct kcm_psock *psock;
 352
 353         trace_sk_data_ready(sk);
 354
 355         read_lock_bh(&sk->sk_callback_lock);
 356
 357         psock = (struct kcm_psock *)sk->sk_user_data;
 358         if (likely(psock))
 359                 strp_data_ready(&psock->strp);
 360
 361         read_unlock_bh(&sk->sk_callback_lock);
 362 }
 363
 364 /* Called with lower sock held */
 365 static void kcm_rcv_strparser(struct strparser *strp, struct sk_buff *skb)
 366 {
 367         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 368         struct kcm_sock *kcm;
 369
 370 try_queue:
 371         kcm = reserve_rx_kcm(psock, skb);
 372         if (!kcm) {
 373                  /* Unable to reserve a KCM, message is held in psock and strp
 374                   * is paused.
 375                   */
 376                 return;
 377         }
 378
 379         if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 380                 /* Should mean socket buffer full */
 381                 unreserve_rx_kcm(psock, false);
 382                 goto try_queue;
 383         }
 384 }
 385
 386 static int kcm_parse_func_strparser(struct strparser *strp, struct sk_buff *skb)
 387 {
 388         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 389         struct bpf_prog *prog = psock->bpf_prog;
 390         int res;
 391
 392         res = bpf_prog_run_pin_on_cpu(prog, skb);
 393         return res;
 394 }
 395
 396 static int kcm_read_sock_done(struct strparser *strp, int err)
 397 {
 398         struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
 399
 400         unreserve_rx_kcm(psock, true);
 401
 402         return err;
 403 }
 404
 405 static void psock_state_change(struct sock *sk)
 406 {
 407         /* TCP only does a EPOLLIN for a half close. Do a EPOLLHUP here
 408          * since application will normally not poll with EPOLLIN
 409          * on the TCP sockets.
 410          */
 411
 412         report_csk_error(sk, EPIPE);
 413 }
 414
 415 static void psock_write_space(struct sock *sk)
 416 {
 417         struct kcm_psock *psock;
 418         struct kcm_mux *mux;
 419         struct kcm_sock *kcm;
 420
 421         read_lock_bh(&sk->sk_callback_lock);
 422
 423         psock = (struct kcm_psock *)sk->sk_user_data;
 424         if (unlikely(!psock))
 425                 goto out;
 426         mux = psock->mux;
 427
 428         spin_lock_bh(&mux->lock);
 429
 430         /* Check if the socket is reserved so someone is waiting for sending. */
 431         kcm = psock->tx_kcm;
 432         if (kcm && !unlikely(kcm->tx_stopped))
 433                 queue_work(kcm_wq, &kcm->tx_work);
 434
 435         spin_unlock_bh(&mux->lock);
 436 out:
 437         read_unlock_bh(&sk->sk_callback_lock);
 438 }
 439
 440 static void unreserve_psock(struct kcm_sock *kcm);
 441
 442 /* kcm sock is locked. */
 443 static struct kcm_psock *reserve_psock(struct kcm_sock *kcm)
 444 {
 445         struct kcm_mux *mux = kcm->mux;
 446         struct kcm_psock *psock;
 447
 448         psock = kcm->tx_psock;
 449
 450         smp_rmb(); /* Must read tx_psock before tx_wait */
 451
 452         if (psock) {
 453                 WARN_ON(kcm->tx_wait);
 454                 if (unlikely(psock->tx_stopped))
 455                         unreserve_psock(kcm);
 456                 else
 457                         return kcm->tx_psock;
 458         }
 459
 460         spin_lock_bh(&mux->lock);
 461
 462         /* Check again under lock to see if psock was reserved for this
 463          * psock via psock_unreserve.
 464          */
 465         psock = kcm->tx_psock;
 466         if (unlikely(psock)) {
 467                 WARN_ON(kcm->tx_wait);
 468                 spin_unlock_bh(&mux->lock);
 469                 return kcm->tx_psock;
 470         }
 471
 472         if (!list_empty(&mux->psocks_avail)) {
 473                 psock = list_first_entry(&mux->psocks_avail,
 474                                          struct kcm_psock,
 475                                          psock_avail_list);
 476                 list_del(&psock->psock_avail_list);
 477                 if (kcm->tx_wait) {
 478                         list_del(&kcm->wait_psock_list);
 479                         kcm->tx_wait = false;
 480                 }
 481                 kcm->tx_psock = psock;
 482                 psock->tx_kcm = kcm;
 483                 KCM_STATS_INCR(psock->stats.reserved);
 484         } else if (!kcm->tx_wait) {
 485                 list_add_tail(&kcm->wait_psock_list,
 486                               &mux->kcm_tx_waiters);
 487                 kcm->tx_wait = true;
 488         }
 489
 490         spin_unlock_bh(&mux->lock);
 491
 492         return psock;
 493 }
 494
 495 /* mux lock held */
 496 static void psock_now_avail(struct kcm_psock *psock)
 497 {
 498         struct kcm_mux *mux = psock->mux;
 499         struct kcm_sock *kcm;
 500
 501         if (list_empty(&mux->kcm_tx_waiters)) {
 502                 list_add_tail(&psock->psock_avail_list,
 503                               &mux->psocks_avail);
 504         } else {
 505                 kcm = list_first_entry(&mux->kcm_tx_waiters,
 506                                        struct kcm_sock,
 507                                        wait_psock_list);
 508                 list_del(&kcm->wait_psock_list);
 509                 kcm->tx_wait = false;
 510                 psock->tx_kcm = kcm;
 511
 512                 /* Commit before changing tx_psock since that is read in
 513                  * reserve_psock before queuing work.
 514                  */
 515                 smp_mb();
 516
 517                 kcm->tx_psock = psock;
 518                 KCM_STATS_INCR(psock->stats.reserved);
 519                 queue_work(kcm_wq, &kcm->tx_work);
 520         }
 521 }
 522
 523 /* kcm sock is locked. */
 524 static void unreserve_psock(struct kcm_sock *kcm)
 525 {
 526         struct kcm_psock *psock;
 527         struct kcm_mux *mux = kcm->mux;
 528
 529         spin_lock_bh(&mux->lock);
 530
 531         psock = kcm->tx_psock;
 532
 533         if (WARN_ON(!psock)) {
 534                 spin_unlock_bh(&mux->lock);
 535                 return;
 536         }
 537
 538         smp_rmb(); /* Read tx_psock before tx_wait */
 539
 540         kcm_update_tx_mux_stats(mux, psock);
 541
 542         WARN_ON(kcm->tx_wait);
 543
 544         kcm->tx_psock = NULL;
 545         psock->tx_kcm = NULL;
 546         KCM_STATS_INCR(psock->stats.unreserved);
 547
 548         if (unlikely(psock->tx_stopped)) {
 549                 if (psock->done) {
 550                         /* Deferred free */
 551                         list_del(&psock->psock_list);
 552                         mux->psocks_cnt--;
 553                         sock_put(psock->sk);
 554                         fput(psock->sk->sk_socket->file);
 555                         kmem_cache_free(kcm_psockp, psock);
 556                 }
 557
 558                 /* Don't put back on available list */
 559
 560                 spin_unlock_bh(&mux->lock);
 561
 562                 return;
 563         }
 564
 565         psock_now_avail(psock);
 566
 567         spin_unlock_bh(&mux->lock);
 568 }
 569
 570 static void kcm_report_tx_retry(struct kcm_sock *kcm)
 571 {
 572         struct kcm_mux *mux = kcm->mux;
 573
 574         spin_lock_bh(&mux->lock);
 575         KCM_STATS_INCR(mux->stats.tx_retries);
 576         spin_unlock_bh(&mux->lock);
 577 }
 578
 579 /* Write any messages ready on the kcm socket.  Called with kcm sock lock
 580  * held.  Return bytes actually sent or error.
 581  */
 582 static int kcm_write_msgs(struct kcm_sock *kcm)
 583 {
 584         unsigned int total_sent = 0;
 585         struct sock *sk = &kcm->sk;
 586         struct kcm_psock *psock;
 587         struct sk_buff *head;
 588         int ret = 0;
 589
 590         kcm->tx_wait_more = false;
 591         psock = kcm->tx_psock;
 592         if (unlikely(psock && psock->tx_stopped)) {
 593                 /* A reserved psock was aborted asynchronously. Unreserve
 594                  * it and we'll retry the message.
 595                  */
 596                 unreserve_psock(kcm);
 597                 kcm_report_tx_retry(kcm);
 598                 if (skb_queue_empty(&sk->sk_write_queue))
 599                         return 0;
 600
 601                 kcm_tx_msg(skb_peek(&sk->sk_write_queue))->started_tx = false;
 602         }
 603
 604 retry:
 605         while ((head = skb_peek(&sk->sk_write_queue))) {
 606                 struct msghdr msg = {
 607                         .msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES,
 608                 };
 609                 struct kcm_tx_msg *txm = kcm_tx_msg(head);
 610                 struct sk_buff *skb;
 611                 unsigned int msize;
 612                 int i;
 613
 614                 if (!txm->started_tx) {
 615                         psock = reserve_psock(kcm);
 616                         if (!psock)
 617                                 goto out;
 618                         skb = head;
 619                         txm->frag_offset = 0;
 620                         txm->sent = 0;
 621                         txm->started_tx = true;
 622                 } else {
 623                         if (WARN_ON(!psock)) {
 624                                 ret = -EINVAL;
 625                                 goto out;
 626                         }
 627                         skb = txm->frag_skb;
 628                 }
 629
 630                 if (WARN_ON(!skb_shinfo(skb)->nr_frags)) {
 631                         ret = -EINVAL;
 632                         goto out;
 633                 }
 634
 635                 msize = 0;
 636                 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
 637                         msize += skb_frag_size(&skb_shinfo(skb)->frags[i]);
 638
 639                 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE,
 640                               skb_shinfo(skb)->frags, skb_shinfo(skb)->nr_frags,
 641                               msize);
 642                 iov_iter_advance(&msg.msg_iter, txm->frag_offset);
 643
 644                 do {
 645                         ret = sock_sendmsg(psock->sk->sk_socket, &msg);
 646                         if (ret <= 0) {
 647                                 if (ret == -EAGAIN) {
 648                                         /* Save state to try again when there's
 649                                          * write space on the socket
 650                                          */
 651                                         txm->frag_skb = skb;
 652                                         ret = 0;
 653                                         goto out;
 654                                 }
 655
 656                                 /* Hard failure in sending message, abort this
 657                                  * psock since it has lost framing
 658                                  * synchronization and retry sending the
 659                                  * message from the beginning.
 660                                  */
 661                                 kcm_abort_tx_psock(psock, ret ? -ret : EPIPE,
 662                                                    true);
 663                                 unreserve_psock(kcm);
 664                                 psock = NULL;
 665
 666                                 txm->started_tx = false;
 667                                 kcm_report_tx_retry(kcm);
 668                                 ret = 0;
 669                                 goto retry;
 670                         }
 671
 672                         txm->sent += ret;
 673                         txm->frag_offset += ret;
 674                         KCM_STATS_ADD(psock->stats.tx_bytes, ret);
 675                 } while (msg.msg_iter.count > 0);
 676
 677                 if (skb == head) {
 678                         if (skb_has_frag_list(skb)) {
 679                                 txm->frag_skb = skb_shinfo(skb)->frag_list;
 680                                 txm->frag_offset = 0;
 681                                 continue;
 682                         }
 683                 } else if (skb->next) {
 684                         txm->frag_skb = skb->next;
 685                         txm->frag_offset = 0;
 686                         continue;
 687                 }
 688
 689                 /* Successfully sent the whole packet, account for it. */
 690                 sk->sk_wmem_queued -= txm->sent;
 691                 total_sent += txm->sent;
 692                 skb_dequeue(&sk->sk_write_queue);
 693                 kfree_skb(head);
 694                 KCM_STATS_INCR(psock->stats.tx_msgs);
 695         }
 696 out:
 697         if (!head) {
 698                 /* Done with all queued messages. */
 699                 WARN_ON(!skb_queue_empty(&sk->sk_write_queue));
 700                 if (psock)
 701                         unreserve_psock(kcm);
 702         }
 703
 704         /* Check if write space is available */
 705         sk->sk_write_space(sk);
 706
 707         return total_sent ? : ret;
 708 }
 709
 710 static void kcm_tx_work(struct work_struct *w)
 711 {
 712         struct kcm_sock *kcm = container_of(w, struct kcm_sock, tx_work);
 713         struct sock *sk = &kcm->sk;
 714         int err;
 715
 716         lock_sock(sk);
 717
 718         /* Primarily for SOCK_DGRAM sockets, also handle asynchronous tx
 719          * aborts
 720          */
 721         err = kcm_write_msgs(kcm);
 722         if (err < 0) {
 723                 /* Hard failure in write, report error on KCM socket */
 724                 pr_warn("KCM: Hard failure on kcm_write_msgs %d\n", err);
 725                 report_csk_error(&kcm->sk, -err);
 726                 goto out;
 727         }
 728
 729         /* Primarily for SOCK_SEQPACKET sockets */
 730         if (likely(sk->sk_socket) &&
 731             test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
 732                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 733                 sk->sk_write_space(sk);
 734         }
 735
 736 out:
 737         release_sock(sk);
 738 }
 739
 740 static void kcm_push(struct kcm_sock *kcm)
 741 {
 742         if (kcm->tx_wait_more)
 743                 kcm_write_msgs(kcm);
 744 }
 745
 746 static int kcm_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
 747 {
 748         struct sock *sk = sock->sk;
 749         struct kcm_sock *kcm = kcm_sk(sk);
 750         struct sk_buff *skb = NULL, *head = NULL;
 751         size_t copy, copied = 0;
 752         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 753         int eor = (sock->type == SOCK_DGRAM) ?
 754                   !(msg->msg_flags & MSG_MORE) : !!(msg->msg_flags & MSG_EOR);
 755         int err = -EPIPE;
 756
 757         lock_sock(sk);
 758
 759         /* Per tcp_sendmsg this should be in poll */
 760         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 761
 762         if (sk->sk_err)
 763                 goto out_error;
 764
 765         if (kcm->seq_skb) {
 766                 /* Previously opened message */
 767                 head = kcm->seq_skb;
 768                 skb = kcm_tx_msg(head)->last_skb;
 769                 goto start;
 770         }
 771
 772         /* Call the sk_stream functions to manage the sndbuf mem. */
 773         if (!sk_stream_memory_free(sk)) {
 774                 kcm_push(kcm);
 775                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 776                 err = sk_stream_wait_memory(sk, &timeo);
 777                 if (err)
 778                         goto out_error;
 779         }
 780
 781         if (msg_data_left(msg)) {
 782                 /* New message, alloc head skb */
 783                 head = alloc_skb(0, sk->sk_allocation);
 784                 while (!head) {
 785                         kcm_push(kcm);
 786                         err = sk_stream_wait_memory(sk, &timeo);
 787                         if (err)
 788                                 goto out_error;
 789
 790                         head = alloc_skb(0, sk->sk_allocation);
 791                 }
 792
 793                 skb = head;
 794
 795                 /* Set ip_summed to CHECKSUM_UNNECESSARY to avoid calling
 796                  * csum_and_copy_from_iter from skb_do_copy_data_nocache.
 797                  */
 798                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 799         }
 800
 801 start:
 802         while (msg_data_left(msg)) {
 803                 bool merge = true;
 804                 int i = skb_shinfo(skb)->nr_frags;
 805                 struct page_frag *pfrag = sk_page_frag(sk);
 806
 807                 if (!sk_page_frag_refill(sk, pfrag))
 808                         goto wait_for_memory;
 809
 810                 if (!skb_can_coalesce(skb, i, pfrag->page,
 811                                       pfrag->offset)) {
 812                         if (i == MAX_SKB_FRAGS) {
 813                                 struct sk_buff *tskb;
 814
 815                                 tskb = alloc_skb(0, sk->sk_allocation);
 816                                 if (!tskb)
 817                                         goto wait_for_memory;
 818
 819                                 if (head == skb)
 820                                         skb_shinfo(head)->frag_list = tskb;
 821                                 else
 822                                         skb->next = tskb;
 823
 824                                 skb = tskb;
 825                                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 826                                 continue;
 827                         }
 828                         merge = false;
 829                 }
 830
 831                 if (msg->msg_flags & MSG_SPLICE_PAGES) {
 832                         copy = msg_data_left(msg);
 833                         if (!sk_wmem_schedule(sk, copy))
 834                                 goto wait_for_memory;
 835
 836                         err = skb_splice_from_iter(skb, &msg->msg_iter, copy,
 837                                                    sk->sk_allocation);
 838                         if (err < 0) {
 839                                 if (err == -EMSGSIZE)
 840                                         goto wait_for_memory;
 841                                 goto out_error;
 842                         }
 843
 844                         copy = err;
 845                         skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
 846                         sk_wmem_queued_add(sk, copy);
 847                         sk_mem_charge(sk, copy);
 848
 849                         if (head != skb)
 850                                 head->truesize += copy;
 851                 } else {
 852                         copy = min_t(int, msg_data_left(msg),
 853                                      pfrag->size - pfrag->offset);
 854                         if (!sk_wmem_schedule(sk, copy))
 855                                 goto wait_for_memory;
 856
 857                         err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
 858                                                        pfrag->page,
 859                                                        pfrag->offset,
 860                                                        copy);
 861                         if (err)
 862                                 goto out_error;
 863
 864                         /* Update the skb. */
 865                         if (merge) {
 866                                 skb_frag_size_add(
 867                                         &skb_shinfo(skb)->frags[i - 1], copy);
 868                         } else {
 869                                 skb_fill_page_desc(skb, i, pfrag->page,
 870                                                    pfrag->offset, copy);
 871                                 get_page(pfrag->page);
 872                         }
 873
 874                         pfrag->offset += copy;
 875                 }
 876
 877                 copied += copy;
 878                 if (head != skb) {
 879                         head->len += copy;
 880                         head->data_len += copy;
 881                 }
 882
 883                 continue;
 884
 885 wait_for_memory:
 886                 kcm_push(kcm);
 887                 err = sk_stream_wait_memory(sk, &timeo);
 888                 if (err)
 889                         goto out_error;
 890         }
 891
 892         if (eor) {
 893                 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
 894
 895                 if (head) {
 896                         /* Message complete, queue it on send buffer */
 897                         __skb_queue_tail(&sk->sk_write_queue, head);
 898                         kcm->seq_skb = NULL;
 899                         KCM_STATS_INCR(kcm->stats.tx_msgs);
 900                 }
 901
 902                 if (msg->msg_flags & MSG_BATCH) {
 903                         kcm->tx_wait_more = true;
 904                 } else if (kcm->tx_wait_more || not_busy) {
 905                         err = kcm_write_msgs(kcm);
 906                         if (err < 0) {
 907                                 /* We got a hard error in write_msgs but have
 908                                  * already queued this message. Report an error
 909                                  * in the socket, but don't affect return value
 910                                  * from sendmsg
 911                                  */
 912                                 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
 913                                 report_csk_error(&kcm->sk, -err);
 914                         }
 915                 }
 916         } else {
 917                 /* Message not complete, save state */
 918 partial_message:
 919                 if (head) {
 920                         kcm->seq_skb = head;
 921                         kcm_tx_msg(head)->last_skb = skb;
 922                 }
 923         }
 924
 925         KCM_STATS_ADD(kcm->stats.tx_bytes, copied);
 926
 927         release_sock(sk);
 928         return copied;
 929
 930 out_error:
 931         kcm_push(kcm);
 932
 933         if (sock->type == SOCK_SEQPACKET) {
 934                 /* Wrote some bytes before encountering an
 935                  * error, return partial success.
 936                  */
 937                 if (copied)
 938                         goto partial_message;
 939                 if (head != kcm->seq_skb)
 940                         kfree_skb(head);
 941         } else {
 942                 kfree_skb(head);
 943                 kcm->seq_skb = NULL;
 944         }
 945
 946         err = sk_stream_error(sk, msg->msg_flags, err);
 947
 948         /* make sure we wake any epoll edge trigger waiter */
 949         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
 950                 sk->sk_write_space(sk);
 951
 952         release_sock(sk);
 953         return err;
 954 }
 955
 956 static void kcm_splice_eof(struct socket *sock)
 957 {
 958         struct sock *sk = sock->sk;
 959         struct kcm_sock *kcm = kcm_sk(sk);
 960
 961         if (skb_queue_empty_lockless(&sk->sk_write_queue))
 962                 return;
 963
 964         lock_sock(sk);
 965         kcm_write_msgs(kcm);
 966         release_sock(sk);
 967 }
 968
 969 static int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
 970                        size_t len, int flags)
 971 {
 972         struct sock *sk = sock->sk;
 973         struct kcm_sock *kcm = kcm_sk(sk);
 974         int err = 0;
 975         struct strp_msg *stm;
 976         int copied = 0;
 977         struct sk_buff *skb;
 978
 979         skb = skb_recv_datagram(sk, flags, &err);
 980         if (!skb)
 981                 goto out;
 982
 983         /* Okay, have a message on the receive queue */
 984
 985         stm = strp_msg(skb);
 986
 987         if (len > stm->full_len)
 988                 len = stm->full_len;
 989
 990         err = skb_copy_datagram_msg(skb, stm->offset, msg, len);
 991         if (err < 0)
 992                 goto out;
 993
 994         copied = len;
 995         if (likely(!(flags & MSG_PEEK))) {
 996                 KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
 997                 if (copied < stm->full_len) {
 998                         if (sock->type == SOCK_DGRAM) {
 999                                 /* Truncated message */
1000                                 msg->msg_flags |= MSG_TRUNC;
1001                                 goto msg_finished;
1002                         }
1003                         stm->offset += copied;
1004                         stm->full_len -= copied;
1005                 } else {
1006 msg_finished:
1007                         /* Finished with message */
1008                         msg->msg_flags |= MSG_EOR;
1009                         KCM_STATS_INCR(kcm->stats.rx_msgs);
1010                 }
1011         }
1012
1013 out:
1014         skb_free_datagram(sk, skb);
1015         return copied ? : err;
1016 }
1017
1018 static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1019                                struct pipe_inode_info *pipe, size_t len,
1020                                unsigned int flags)
1021 {
1022         struct sock *sk = sock->sk;
1023         struct kcm_sock *kcm = kcm_sk(sk);
1024         struct strp_msg *stm;
1025         int err = 0;
1026         ssize_t copied;
1027         struct sk_buff *skb;
1028
1029         /* Only support splice for SOCKSEQPACKET */
1030
1031         skb = skb_recv_datagram(sk, flags, &err);
1032         if (!skb)
1033                 goto err_out;
1034
1035         /* Okay, have a message on the receive queue */
1036
1037         stm = strp_msg(skb);
1038
1039         if (len > stm->full_len)
1040                 len = stm->full_len;
1041
1042         copied = skb_splice_bits(skb, sk, stm->offset, pipe, len, flags);
1043         if (copied < 0) {
1044                 err = copied;
1045                 goto err_out;
1046         }
1047
1048         KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1049
1050         stm->offset += copied;
1051         stm->full_len -= copied;
1052
1053         /* We have no way to return MSG_EOR. If all the bytes have been
1054          * read we still leave the message in the receive socket buffer.
1055          * A subsequent recvmsg needs to be done to return MSG_EOR and
1056          * finish reading the message.
1057          */
1058
1059         skb_free_datagram(sk, skb);
1060         return copied;
1061
1062 err_out:
1063         skb_free_datagram(sk, skb);
1064         return err;
1065 }
1066
1067 /* kcm sock lock held */
1068 static void kcm_recv_disable(struct kcm_sock *kcm)
1069 {
1070         struct kcm_mux *mux = kcm->mux;
1071
1072         if (kcm->rx_disabled)
1073                 return;
1074
1075         spin_lock_bh(&mux->rx_lock);
1076
1077         kcm->rx_disabled = 1;
1078
1079         /* If a psock is reserved we'll do cleanup in unreserve */
1080         if (!kcm->rx_psock) {
1081                 if (kcm->rx_wait) {
1082                         list_del(&kcm->wait_rx_list);
1083                         /* paired with lockless reads in kcm_rfree() */
1084                         WRITE_ONCE(kcm->rx_wait, false);
1085                 }
1086
1087                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1088         }
1089
1090         spin_unlock_bh(&mux->rx_lock);
1091 }
1092
1093 /* kcm sock lock held */
1094 static void kcm_recv_enable(struct kcm_sock *kcm)
1095 {
1096         struct kcm_mux *mux = kcm->mux;
1097
1098         if (!kcm->rx_disabled)
1099                 return;
1100
1101         spin_lock_bh(&mux->rx_lock);
1102
1103         kcm->rx_disabled = 0;
1104         kcm_rcv_ready(kcm);
1105
1106         spin_unlock_bh(&mux->rx_lock);
1107 }
1108
1109 static int kcm_setsockopt(struct socket *sock, int level, int optname,
1110                           sockptr_t optval, unsigned int optlen)
1111 {
1112         struct kcm_sock *kcm = kcm_sk(sock->sk);
1113         int val, valbool;
1114         int err = 0;
1115
1116         if (level != SOL_KCM)
1117                 return -ENOPROTOOPT;
1118
1119         if (optlen < sizeof(int))
1120                 return -EINVAL;
1121
1122         if (copy_from_sockptr(&val, optval, sizeof(int)))
1123                 return -EFAULT;
1124
1125         valbool = val ? 1 : 0;
1126
1127         switch (optname) {
1128         case KCM_RECV_DISABLE:
1129                 lock_sock(&kcm->sk);
1130                 if (valbool)
1131                         kcm_recv_disable(kcm);
1132                 else
1133                         kcm_recv_enable(kcm);
1134                 release_sock(&kcm->sk);
1135                 break;
1136         default:
1137                 err = -ENOPROTOOPT;
1138         }
1139
1140         return err;
1141 }
1142
1143 static int kcm_getsockopt(struct socket *sock, int level, int optname,
1144                           char __user *optval, int __user *optlen)
1145 {
1146         struct kcm_sock *kcm = kcm_sk(sock->sk);
1147         int val, len;
1148
1149         if (level != SOL_KCM)
1150                 return -ENOPROTOOPT;
1151
1152         if (get_user(len, optlen))
1153                 return -EFAULT;
1154
1155         if (len < 0)
1156                 return -EINVAL;
1157
1158         len = min_t(unsigned int, len, sizeof(int));
1159
1160         switch (optname) {
1161         case KCM_RECV_DISABLE:
1162                 val = kcm->rx_disabled;
1163                 break;
1164         default:
1165                 return -ENOPROTOOPT;
1166         }
1167
1168         if (put_user(len, optlen))
1169                 return -EFAULT;
1170         if (copy_to_user(optval, &val, len))
1171                 return -EFAULT;
1172         return 0;
1173 }
1174
1175 static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1176 {
1177         struct kcm_sock *tkcm;
1178         struct list_head *head;
1179         int index = 0;
1180
1181         /* For SOCK_SEQPACKET sock type, datagram_poll checks the sk_state, so
1182          * we set sk_state, otherwise epoll_wait always returns right away with
1183          * EPOLLHUP
1184          */
1185         kcm->sk.sk_state = TCP_ESTABLISHED;
1186
1187         /* Add to mux's kcm sockets list */
1188         kcm->mux = mux;
1189         spin_lock_bh(&mux->lock);
1190
1191         head = &mux->kcm_socks;
1192         list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1193                 if (tkcm->index != index)
1194                         break;
1195                 head = &tkcm->kcm_sock_list;
1196                 index++;
1197         }
1198
1199         list_add(&kcm->kcm_sock_list, head);
1200         kcm->index = index;
1201
1202         mux->kcm_socks_cnt++;
1203         spin_unlock_bh(&mux->lock);
1204
1205         INIT_WORK(&kcm->tx_work, kcm_tx_work);
1206
1207         spin_lock_bh(&mux->rx_lock);
1208         kcm_rcv_ready(kcm);
1209         spin_unlock_bh(&mux->rx_lock);
1210 }
1211
1212 static int kcm_attach(struct socket *sock, struct socket *csock,
1213                       struct bpf_prog *prog)
1214 {
1215         struct kcm_sock *kcm = kcm_sk(sock->sk);
1216         struct kcm_mux *mux = kcm->mux;
1217         struct sock *csk;
1218         struct kcm_psock *psock = NULL, *tpsock;
1219         struct list_head *head;
1220         int index = 0;
1221         static const struct strp_callbacks cb = {
1222                 .rcv_msg = kcm_rcv_strparser,
1223                 .parse_msg = kcm_parse_func_strparser,
1224                 .read_sock_done = kcm_read_sock_done,
1225         };
1226         int err = 0;
1227
1228         csk = csock->sk;
1229         if (!csk)
1230                 return -EINVAL;
1231
1232         lock_sock(csk);
1233
1234         /* Only allow TCP sockets to be attached for now */
1235         if ((csk->sk_family != AF_INET && csk->sk_family != AF_INET6) ||
1236             csk->sk_protocol != IPPROTO_TCP) {
1237                 err = -EOPNOTSUPP;
1238                 goto out;
1239         }
1240
1241         /* Don't allow listeners or closed sockets */
1242         if (csk->sk_state == TCP_LISTEN || csk->sk_state == TCP_CLOSE) {
1243                 err = -EOPNOTSUPP;
1244                 goto out;
1245         }
1246
1247         psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1248         if (!psock) {
1249                 err = -ENOMEM;
1250                 goto out;
1251         }
1252
1253         psock->mux = mux;
1254         psock->sk = csk;
1255         psock->bpf_prog = prog;
1256
1257         write_lock_bh(&csk->sk_callback_lock);
1258
1259         /* Check if sk_user_data is already by KCM or someone else.
1260          * Must be done under lock to prevent race conditions.
1261          */
1262         if (csk->sk_user_data) {
1263                 write_unlock_bh(&csk->sk_callback_lock);
1264                 kmem_cache_free(kcm_psockp, psock);
1265                 err = -EALREADY;
1266                 goto out;
1267         }
1268
1269         err = strp_init(&psock->strp, csk, &cb);
1270         if (err) {
1271                 write_unlock_bh(&csk->sk_callback_lock);
1272                 kmem_cache_free(kcm_psockp, psock);
1273                 goto out;
1274         }
1275
1276         psock->save_data_ready = csk->sk_data_ready;
1277         psock->save_write_space = csk->sk_write_space;
1278         psock->save_state_change = csk->sk_state_change;
1279         csk->sk_user_data = psock;
1280         csk->sk_data_ready = psock_data_ready;
1281         csk->sk_write_space = psock_write_space;
1282         csk->sk_state_change = psock_state_change;
1283
1284         write_unlock_bh(&csk->sk_callback_lock);
1285
1286         sock_hold(csk);
1287
1288         /* Finished initialization, now add the psock to the MUX. */
1289         spin_lock_bh(&mux->lock);
1290         head = &mux->psocks;
1291         list_for_each_entry(tpsock, &mux->psocks, psock_list) {
1292                 if (tpsock->index != index)
1293                         break;
1294                 head = &tpsock->psock_list;
1295                 index++;
1296         }
1297
1298         list_add(&psock->psock_list, head);
1299         psock->index = index;
1300
1301         KCM_STATS_INCR(mux->stats.psock_attach);
1302         mux->psocks_cnt++;
1303         psock_now_avail(psock);
1304         spin_unlock_bh(&mux->lock);
1305
1306         /* Schedule RX work in case there are already bytes queued */
1307         strp_check_rcv(&psock->strp);
1308
1309 out:
1310         release_sock(csk);
1311
1312         return err;
1313 }
1314
1315 static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1316 {
1317         struct socket *csock;
1318         struct bpf_prog *prog;
1319         int err;
1320
1321         csock = sockfd_lookup(info->fd, &err);
1322         if (!csock)
1323                 return -ENOENT;
1324
1325         prog = bpf_prog_get_type(info->bpf_fd, BPF_PROG_TYPE_SOCKET_FILTER);
1326         if (IS_ERR(prog)) {
1327                 err = PTR_ERR(prog);
1328                 goto out;
1329         }
1330
1331         err = kcm_attach(sock, csock, prog);
1332         if (err) {
1333                 bpf_prog_put(prog);
1334                 goto out;
1335         }
1336
1337         /* Keep reference on file also */
1338
1339         return 0;
1340 out:
1341         sockfd_put(csock);
1342         return err;
1343 }
1344
1345 static void kcm_unattach(struct kcm_psock *psock)
1346 {
1347         struct sock *csk = psock->sk;
1348         struct kcm_mux *mux = psock->mux;
1349
1350         lock_sock(csk);
1351
1352         /* Stop getting callbacks from TCP socket. After this there should
1353          * be no way to reserve a kcm for this psock.
1354          */
1355         write_lock_bh(&csk->sk_callback_lock);
1356         csk->sk_user_data = NULL;
1357         csk->sk_data_ready = psock->save_data_ready;
1358         csk->sk_write_space = psock->save_write_space;
1359         csk->sk_state_change = psock->save_state_change;
1360         strp_stop(&psock->strp);
1361
1362         if (WARN_ON(psock->rx_kcm)) {
1363                 write_unlock_bh(&csk->sk_callback_lock);
1364                 release_sock(csk);
1365                 return;
1366         }
1367
1368         spin_lock_bh(&mux->rx_lock);
1369
1370         /* Stop receiver activities. After this point psock should not be
1371          * able to get onto ready list either through callbacks or work.
1372          */
1373         if (psock->ready_rx_msg) {
1374                 list_del(&psock->psock_ready_list);
1375                 kfree_skb(psock->ready_rx_msg);
1376                 psock->ready_rx_msg = NULL;
1377                 KCM_STATS_INCR(mux->stats.rx_ready_drops);
1378         }
1379
1380         spin_unlock_bh(&mux->rx_lock);
1381
1382         write_unlock_bh(&csk->sk_callback_lock);
1383
1384         /* Call strp_done without sock lock */
1385         release_sock(csk);
1386         strp_done(&psock->strp);
1387         lock_sock(csk);
1388
1389         bpf_prog_put(psock->bpf_prog);
1390
1391         spin_lock_bh(&mux->lock);
1392
1393         aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1394         save_strp_stats(&psock->strp, &mux->aggregate_strp_stats);
1395
1396         KCM_STATS_INCR(mux->stats.psock_unattach);
1397
1398         if (psock->tx_kcm) {
1399                 /* psock was reserved.  Just mark it finished and we will clean
1400                  * up in the kcm paths, we need kcm lock which can not be
1401                  * acquired here.
1402                  */
1403                 KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1404                 spin_unlock_bh(&mux->lock);
1405
1406                 /* We are unattaching a socket that is reserved. Abort the
1407                  * socket since we may be out of sync in sending on it. We need
1408                  * to do this without the mux lock.
1409                  */
1410                 kcm_abort_tx_psock(psock, EPIPE, false);
1411
1412                 spin_lock_bh(&mux->lock);
1413                 if (!psock->tx_kcm) {
1414                         /* psock now unreserved in window mux was unlocked */
1415                         goto no_reserved;
1416                 }
1417                 psock->done = 1;
1418
1419                 /* Commit done before queuing work to process it */
1420                 smp_mb();
1421
1422                 /* Queue tx work to make sure psock->done is handled */
1423                 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
1424                 spin_unlock_bh(&mux->lock);
1425         } else {
1426 no_reserved:
1427                 if (!psock->tx_stopped)
1428                         list_del(&psock->psock_avail_list);
1429                 list_del(&psock->psock_list);
1430                 mux->psocks_cnt--;
1431                 spin_unlock_bh(&mux->lock);
1432
1433                 sock_put(csk);
1434                 fput(csk->sk_socket->file);
1435                 kmem_cache_free(kcm_psockp, psock);
1436         }
1437
1438         release_sock(csk);
1439 }
1440
1441 static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1442 {
1443         struct kcm_sock *kcm = kcm_sk(sock->sk);
1444         struct kcm_mux *mux = kcm->mux;
1445         struct kcm_psock *psock;
1446         struct socket *csock;
1447         struct sock *csk;
1448         int err;
1449
1450         csock = sockfd_lookup(info->fd, &err);
1451         if (!csock)
1452                 return -ENOENT;
1453
1454         csk = csock->sk;
1455         if (!csk) {
1456                 err = -EINVAL;
1457                 goto out;
1458         }
1459
1460         err = -ENOENT;
1461
1462         spin_lock_bh(&mux->lock);
1463
1464         list_for_each_entry(psock, &mux->psocks, psock_list) {
1465                 if (psock->sk != csk)
1466                         continue;
1467
1468                 /* Found the matching psock */
1469
1470                 if (psock->unattaching || WARN_ON(psock->done)) {
1471                         err = -EALREADY;
1472                         break;
1473                 }
1474
1475                 psock->unattaching = 1;
1476
1477                 spin_unlock_bh(&mux->lock);
1478
1479                 /* Lower socket lock should already be held */
1480                 kcm_unattach(psock);
1481
1482                 err = 0;
1483                 goto out;
1484         }
1485
1486         spin_unlock_bh(&mux->lock);
1487
1488 out:
1489         sockfd_put(csock);
1490         return err;
1491 }
1492
1493 static struct proto kcm_proto = {
1494         .name   = "KCM",
1495         .owner  = THIS_MODULE,
1496         .obj_size = sizeof(struct kcm_sock),
1497 };
1498
1499 /* Clone a kcm socket. */
1500 static struct file *kcm_clone(struct socket *osock)
1501 {
1502         struct socket *newsock;
1503         struct sock *newsk;
1504
1505         newsock = sock_alloc();
1506         if (!newsock)
1507                 return ERR_PTR(-ENFILE);
1508
1509         newsock->type = osock->type;
1510         newsock->ops = osock->ops;
1511
1512         __module_get(newsock->ops->owner);
1513
1514         newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1515                          &kcm_proto, false);
1516         if (!newsk) {
1517                 sock_release(newsock);
1518                 return ERR_PTR(-ENOMEM);
1519         }
1520         sock_init_data(newsock, newsk);
1521         init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1522
1523         return sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1524 }
1525
1526 static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
1527 {
1528         int err;
1529
1530         switch (cmd) {
1531         case SIOCKCMATTACH: {
1532                 struct kcm_attach info;
1533
1534                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1535                         return -EFAULT;
1536
1537                 err = kcm_attach_ioctl(sock, &info);
1538
1539                 break;
1540         }
1541         case SIOCKCMUNATTACH: {
1542                 struct kcm_unattach info;
1543
1544                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1545                         return -EFAULT;
1546
1547                 err = kcm_unattach_ioctl(sock, &info);
1548
1549                 break;
1550         }
1551         case SIOCKCMCLONE: {
1552                 struct kcm_clone info;
1553                 struct file *file;
1554
1555                 info.fd = get_unused_fd_flags(0);
1556                 if (unlikely(info.fd < 0))
1557                         return info.fd;
1558
1559                 file = kcm_clone(sock);
1560                 if (IS_ERR(file)) {
1561                         put_unused_fd(info.fd);
1562                         return PTR_ERR(file);
1563                 }
1564                 if (copy_to_user((void __user *)arg, &info,
1565                                  sizeof(info))) {
1566                         put_unused_fd(info.fd);
1567                         fput(file);
1568                         return -EFAULT;
1569                 }
1570                 fd_install(info.fd, file);
1571                 err = 0;
1572                 break;
1573         }
1574         default:
1575                 err = -ENOIOCTLCMD;
1576                 break;
1577         }
1578
1579         return err;
1580 }
1581
1582 static void free_mux(struct rcu_head *rcu)
1583 {
1584         struct kcm_mux *mux = container_of(rcu,
1585             struct kcm_mux, rcu);
1586
1587         kmem_cache_free(kcm_muxp, mux);
1588 }
1589
1590 static void release_mux(struct kcm_mux *mux)
1591 {
1592         struct kcm_net *knet = mux->knet;
1593         struct kcm_psock *psock, *tmp_psock;
1594
1595         /* Release psocks */
1596         list_for_each_entry_safe(psock, tmp_psock,
1597                                  &mux->psocks, psock_list) {
1598                 if (!WARN_ON(psock->unattaching))
1599                         kcm_unattach(psock);
1600         }
1601
1602         if (WARN_ON(mux->psocks_cnt))
1603                 return;
1604
1605         __skb_queue_purge(&mux->rx_hold_queue);
1606
1607         mutex_lock(&knet->mutex);
1608         aggregate_mux_stats(&mux->stats, &knet->aggregate_mux_stats);
1609         aggregate_psock_stats(&mux->aggregate_psock_stats,
1610                               &knet->aggregate_psock_stats);
1611         aggregate_strp_stats(&mux->aggregate_strp_stats,
1612                              &knet->aggregate_strp_stats);
1613         list_del_rcu(&mux->kcm_mux_list);
1614         knet->count--;
1615         mutex_unlock(&knet->mutex);
1616
1617         call_rcu(&mux->rcu, free_mux);
1618 }
1619
1620 static void kcm_done(struct kcm_sock *kcm)
1621 {
1622         struct kcm_mux *mux = kcm->mux;
1623         struct sock *sk = &kcm->sk;
1624         int socks_cnt;
1625
1626         spin_lock_bh(&mux->rx_lock);
1627         if (kcm->rx_psock) {
1628                 /* Cleanup in unreserve_rx_kcm */
1629                 WARN_ON(kcm->done);
1630                 kcm->rx_disabled = 1;
1631                 kcm->done = 1;
1632                 spin_unlock_bh(&mux->rx_lock);
1633                 return;
1634         }
1635
1636         if (kcm->rx_wait) {
1637                 list_del(&kcm->wait_rx_list);
1638                 /* paired with lockless reads in kcm_rfree() */
1639                 WRITE_ONCE(kcm->rx_wait, false);
1640         }
1641         /* Move any pending receive messages to other kcm sockets */
1642         requeue_rx_msgs(mux, &sk->sk_receive_queue);
1643
1644         spin_unlock_bh(&mux->rx_lock);
1645
1646         if (WARN_ON(sk_rmem_alloc_get(sk)))
1647                 return;
1648
1649         /* Detach from MUX */
1650         spin_lock_bh(&mux->lock);
1651
1652         list_del(&kcm->kcm_sock_list);
1653         mux->kcm_socks_cnt--;
1654         socks_cnt = mux->kcm_socks_cnt;
1655
1656         spin_unlock_bh(&mux->lock);
1657
1658         if (!socks_cnt) {
1659                 /* We are done with the mux now. */
1660                 release_mux(mux);
1661         }
1662
1663         WARN_ON(kcm->rx_wait);
1664
1665         sock_put(&kcm->sk);
1666 }
1667
1668 /* Called by kcm_release to close a KCM socket.
1669  * If this is the last KCM socket on the MUX, destroy the MUX.
1670  */
1671 static int kcm_release(struct socket *sock)
1672 {
1673         struct sock *sk = sock->sk;
1674         struct kcm_sock *kcm;
1675         struct kcm_mux *mux;
1676         struct kcm_psock *psock;
1677
1678         if (!sk)
1679                 return 0;
1680
1681         kcm = kcm_sk(sk);
1682         mux = kcm->mux;
1683
1684         lock_sock(sk);
1685         sock_orphan(sk);
1686         kfree_skb(kcm->seq_skb);
1687
1688         /* Purge queue under lock to avoid race condition with tx_work trying
1689          * to act when queue is nonempty. If tx_work runs after this point
1690          * it will just return.
1691          */
1692         __skb_queue_purge(&sk->sk_write_queue);
1693
1694         /* Set tx_stopped. This is checked when psock is bound to a kcm and we
1695          * get a writespace callback. This prevents further work being queued
1696          * from the callback (unbinding the psock occurs after canceling work.
1697          */
1698         kcm->tx_stopped = 1;
1699
1700         release_sock(sk);
1701
1702         spin_lock_bh(&mux->lock);
1703         if (kcm->tx_wait) {
1704                 /* Take of tx_wait list, after this point there should be no way
1705                  * that a psock will be assigned to this kcm.
1706                  */
1707                 list_del(&kcm->wait_psock_list);
1708                 kcm->tx_wait = false;
1709         }
1710         spin_unlock_bh(&mux->lock);
1711
1712         /* Cancel work. After this point there should be no outside references
1713          * to the kcm socket.
1714          */
1715         cancel_work_sync(&kcm->tx_work);
1716
1717         lock_sock(sk);
1718         psock = kcm->tx_psock;
1719         if (psock) {
1720                 /* A psock was reserved, so we need to kill it since it
1721                  * may already have some bytes queued from a message. We
1722                  * need to do this after removing kcm from tx_wait list.
1723                  */
1724                 kcm_abort_tx_psock(psock, EPIPE, false);
1725                 unreserve_psock(kcm);
1726         }
1727         release_sock(sk);
1728
1729         WARN_ON(kcm->tx_wait);
1730         WARN_ON(kcm->tx_psock);
1731
1732         sock->sk = NULL;
1733
1734         kcm_done(kcm);
1735
1736         return 0;
1737 }
1738
1739 static const struct proto_ops kcm_dgram_ops = {
1740         .family =       PF_KCM,
1741         .owner =        THIS_MODULE,
1742         .release =      kcm_release,
1743         .bind =         sock_no_bind,
1744         .connect =      sock_no_connect,
1745         .socketpair =   sock_no_socketpair,
1746         .accept =       sock_no_accept,
1747         .getname =      sock_no_getname,
1748         .poll =         datagram_poll,
1749         .ioctl =        kcm_ioctl,
1750         .listen =       sock_no_listen,
1751         .shutdown =     sock_no_shutdown,
1752         .setsockopt =   kcm_setsockopt,
1753         .getsockopt =   kcm_getsockopt,
1754         .sendmsg =      kcm_sendmsg,
1755         .recvmsg =      kcm_recvmsg,
1756         .mmap =         sock_no_mmap,
1757         .splice_eof =   kcm_splice_eof,
1758 };
1759
1760 static const struct proto_ops kcm_seqpacket_ops = {
1761         .family =       PF_KCM,
1762         .owner =        THIS_MODULE,
1763         .release =      kcm_release,
1764         .bind =         sock_no_bind,
1765         .connect =      sock_no_connect,
1766         .socketpair =   sock_no_socketpair,
1767         .accept =       sock_no_accept,
1768         .getname =      sock_no_getname,
1769         .poll =         datagram_poll,
1770         .ioctl =        kcm_ioctl,
1771         .listen =       sock_no_listen,
1772         .shutdown =     sock_no_shutdown,
1773         .setsockopt =   kcm_setsockopt,
1774         .getsockopt =   kcm_getsockopt,
1775         .sendmsg =      kcm_sendmsg,
1776         .recvmsg =      kcm_recvmsg,
1777         .mmap =         sock_no_mmap,
1778         .splice_eof =   kcm_splice_eof,
1779         .splice_read =  kcm_splice_read,
1780 };
1781
1782 /* Create proto operation for kcm sockets */
1783 static int kcm_create(struct net *net, struct socket *sock,
1784                       int protocol, int kern)
1785 {
1786         struct kcm_net *knet = net_generic(net, kcm_net_id);
1787         struct sock *sk;
1788         struct kcm_mux *mux;
1789
1790         switch (sock->type) {
1791         case SOCK_DGRAM:
1792                 sock->ops = &kcm_dgram_ops;
1793                 break;
1794         case SOCK_SEQPACKET:
1795                 sock->ops = &kcm_seqpacket_ops;
1796                 break;
1797         default:
1798                 return -ESOCKTNOSUPPORT;
1799         }
1800
1801         if (protocol != KCMPROTO_CONNECTED)
1802                 return -EPROTONOSUPPORT;
1803
1804         sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
1805         if (!sk)
1806                 return -ENOMEM;
1807
1808         /* Allocate a kcm mux, shared between KCM sockets */
1809         mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
1810         if (!mux) {
1811                 sk_free(sk);
1812                 return -ENOMEM;
1813         }
1814
1815         spin_lock_init(&mux->lock);
1816         spin_lock_init(&mux->rx_lock);
1817         INIT_LIST_HEAD(&mux->kcm_socks);
1818         INIT_LIST_HEAD(&mux->kcm_rx_waiters);
1819         INIT_LIST_HEAD(&mux->kcm_tx_waiters);
1820
1821         INIT_LIST_HEAD(&mux->psocks);
1822         INIT_LIST_HEAD(&mux->psocks_ready);
1823         INIT_LIST_HEAD(&mux->psocks_avail);
1824
1825         mux->knet = knet;
1826
1827         /* Add new MUX to list */
1828         mutex_lock(&knet->mutex);
1829         list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
1830         knet->count++;
1831         mutex_unlock(&knet->mutex);
1832
1833         skb_queue_head_init(&mux->rx_hold_queue);
1834
1835         /* Init KCM socket */
1836         sock_init_data(sock, sk);
1837         init_kcm_sock(kcm_sk(sk), mux);
1838
1839         return 0;
1840 }
1841
1842 static const struct net_proto_family kcm_family_ops = {
1843         .family = PF_KCM,
1844         .create = kcm_create,
1845         .owner  = THIS_MODULE,
1846 };
1847
1848 static __net_init int kcm_init_net(struct net *net)
1849 {
1850         struct kcm_net *knet = net_generic(net, kcm_net_id);
1851
1852         INIT_LIST_HEAD_RCU(&knet->mux_list);
1853         mutex_init(&knet->mutex);
1854
1855         return 0;
1856 }
1857
1858 static __net_exit void kcm_exit_net(struct net *net)
1859 {
1860         struct kcm_net *knet = net_generic(net, kcm_net_id);
1861
1862         /* All KCM sockets should be closed at this point, which should mean
1863          * that all multiplexors and psocks have been destroyed.
1864          */
1865         WARN_ON(!list_empty(&knet->mux_list));
1866
1867         mutex_destroy(&knet->mutex);
1868 }
1869
1870 static struct pernet_operations kcm_net_ops = {
1871         .init = kcm_init_net,
1872         .exit = kcm_exit_net,
1873         .id   = &kcm_net_id,
1874         .size = sizeof(struct kcm_net),
1875 };
1876
1877 static int __init kcm_init(void)
1878 {
1879         int err = -ENOMEM;
1880
1881         kcm_muxp = kmem_cache_create("kcm_mux_cache",
1882                                      sizeof(struct kcm_mux), 0,
1883                                      SLAB_HWCACHE_ALIGN, NULL);
1884         if (!kcm_muxp)
1885                 goto fail;
1886
1887         kcm_psockp = kmem_cache_create("kcm_psock_cache",
1888                                        sizeof(struct kcm_psock), 0,
1889                                         SLAB_HWCACHE_ALIGN, NULL);
1890         if (!kcm_psockp)
1891                 goto fail;
1892
1893         kcm_wq = create_singlethread_workqueue("kkcmd");
1894         if (!kcm_wq)
1895                 goto fail;
1896
1897         err = proto_register(&kcm_proto, 1);
1898         if (err)
1899                 goto fail;
1900
1901         err = register_pernet_device(&kcm_net_ops);
1902         if (err)
1903                 goto net_ops_fail;
1904
1905         err = sock_register(&kcm_family_ops);
1906         if (err)
1907                 goto sock_register_fail;
1908
1909         err = kcm_proc_init();
1910         if (err)
1911                 goto proc_init_fail;
1912
1913         return 0;
1914
1915 proc_init_fail:
1916         sock_unregister(PF_KCM);
1917
1918 sock_register_fail:
1919         unregister_pernet_device(&kcm_net_ops);
1920
1921 net_ops_fail:
1922         proto_unregister(&kcm_proto);
1923
1924 fail:
1925         kmem_cache_destroy(kcm_muxp);
1926         kmem_cache_destroy(kcm_psockp);
1927
1928         if (kcm_wq)
1929                 destroy_workqueue(kcm_wq);
1930
1931         return err;
1932 }
1933
1934 static void __exit kcm_exit(void)
1935 {
1936         kcm_proc_exit();
1937         sock_unregister(PF_KCM);
1938         unregister_pernet_device(&kcm_net_ops);
1939         proto_unregister(&kcm_proto);
1940         destroy_workqueue(kcm_wq);
1941
1942         kmem_cache_destroy(kcm_muxp);
1943         kmem_cache_destroy(kcm_psockp);
1944 }
1945
1946 module_init(kcm_init);
1947 module_exit(kcm_exit);
1948
1949 MODULE_LICENSE("GPL");
1950 MODULE_DESCRIPTION("KCM (Kernel Connection Multiplexor) sockets");
1951 MODULE_ALIAS_NETPROTO(PF_KCM);