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