net/kcm/kcmsock.c

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