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