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