net/kcm/kcmsock.c

   1 #include <linux/bpf.h>
   2 #include <linux/errno.h>
   3 #include <linux/errqueue.h>
   4 #include <linux/file.h>
   5 #include <linux/in.h>
   6 #include <linux/kernel.h>
   7 #include <linux/module.h>
   8 #include <linux/net.h>
   9 #include <linux/netdevice.h>
  10 #include <linux/poll.h>
  11 #include <linux/rculist.h>
  12 #include <linux/skbuff.h>
  13 #include <linux/socket.h>
  14 #include <linux/uaccess.h>
  15 #include <linux/workqueue.h>
  16 #include <net/kcm.h>
  17 #include <net/netns/generic.h>
  18 #include <net/sock.h>
  19 #include <net/tcp.h>
  20 #include <uapi/linux/kcm.h>
  21
  22 unsigned int kcm_net_id;
  23
  24 static struct kmem_cache *kcm_psockp __read_mostly;
  25 static struct kmem_cache *kcm_muxp __read_mostly;
  26 static struct workqueue_struct *kcm_wq;
  27
  28 static inline struct kcm_sock *kcm_sk(const struct sock *sk)
  29 {
  30         return (struct kcm_sock *)sk;
  31 }
  32
  33 static inline struct kcm_tx_msg *kcm_tx_msg(struct sk_buff *skb)
  34 {
  35         return (struct kcm_tx_msg *)skb->cb;
  36 }
  37
  38 static inline struct kcm_rx_msg *kcm_rx_msg(struct sk_buff *skb)
  39 {
  40         return (struct kcm_rx_msg *)((void *)skb->cb +
  41                                      offsetof(struct qdisc_skb_cb, data));
  42 }
  43
  44 static void report_csk_error(struct sock *csk, int err)
  45 {
  46         csk->sk_err = EPIPE;
  47         csk->sk_error_report(csk);
  48 }
  49
  50 /* Callback lock held */
  51 static void kcm_abort_rx_psock(struct kcm_psock *psock, int err,
  52                                struct sk_buff *skb)
  53 {
  54         struct sock *csk = psock->sk;
  55
  56         /* Unrecoverable error in receive */
  57
  58         if (psock->rx_stopped)
  59                 return;
  60
  61         psock->rx_stopped = 1;
  62         KCM_STATS_INCR(psock->stats.rx_aborts);
  63
  64         /* Report an error on the lower socket */
  65         report_csk_error(csk, err);
  66 }
  67
  68 static void kcm_abort_tx_psock(struct kcm_psock *psock, int err,
  69                                bool wakeup_kcm)
  70 {
  71         struct sock *csk = psock->sk;
  72         struct kcm_mux *mux = psock->mux;
  73
  74         /* Unrecoverable error in transmit */
  75
  76         spin_lock_bh(&mux->lock);
  77
  78         if (psock->tx_stopped) {
  79                 spin_unlock_bh(&mux->lock);
  80                 return;
  81         }
  82
  83         psock->tx_stopped = 1;
  84         KCM_STATS_INCR(psock->stats.tx_aborts);
  85
  86         if (!psock->tx_kcm) {
  87                 /* Take off psocks_avail list */
  88                 list_del(&psock->psock_avail_list);
  89         } else if (wakeup_kcm) {
  90                 /* In this case psock is being aborted while outside of
  91                  * write_msgs and psock is reserved. Schedule tx_work
  92                  * to handle the failure there. Need to commit tx_stopped
  93                  * before queuing work.
  94                  */
  95                 smp_mb();
  96
  97                 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
  98         }
  99
 100         spin_unlock_bh(&mux->lock);
 101
 102         /* Report error on lower socket */
 103         report_csk_error(csk, err);
 104 }
 105
 106 /* RX mux lock held. */
 107 static void kcm_update_rx_mux_stats(struct kcm_mux *mux,
 108                                     struct kcm_psock *psock)
 109 {
 110         KCM_STATS_ADD(mux->stats.rx_bytes,
 111                       psock->stats.rx_bytes - psock->saved_rx_bytes);
 112         mux->stats.rx_msgs +=
 113                 psock->stats.rx_msgs - psock->saved_rx_msgs;
 114         psock->saved_rx_msgs = psock->stats.rx_msgs;
 115         psock->saved_rx_bytes = psock->stats.rx_bytes;
 116 }
 117
 118 static void kcm_update_tx_mux_stats(struct kcm_mux *mux,
 119                                     struct kcm_psock *psock)
 120 {
 121         KCM_STATS_ADD(mux->stats.tx_bytes,
 122                       psock->stats.tx_bytes - psock->saved_tx_bytes);
 123         mux->stats.tx_msgs +=
 124                 psock->stats.tx_msgs - psock->saved_tx_msgs;
 125         psock->saved_tx_msgs = psock->stats.tx_msgs;
 126         psock->saved_tx_bytes = psock->stats.tx_bytes;
 127 }
 128
 129 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb);
 130
 131 /* KCM is ready to receive messages on its queue-- either the KCM is new or
 132  * has become unblocked after being blocked on full socket buffer. Queue any
 133  * pending ready messages on a psock. RX mux lock held.
 134  */
 135 static void kcm_rcv_ready(struct kcm_sock *kcm)
 136 {
 137         struct kcm_mux *mux = kcm->mux;
 138         struct kcm_psock *psock;
 139         struct sk_buff *skb;
 140
 141         if (unlikely(kcm->rx_wait || kcm->rx_psock || kcm->rx_disabled))
 142                 return;
 143
 144         while (unlikely((skb = __skb_dequeue(&mux->rx_hold_queue)))) {
 145                 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 146                         /* Assuming buffer limit has been reached */
 147                         skb_queue_head(&mux->rx_hold_queue, skb);
 148                         WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
 149                         return;
 150                 }
 151         }
 152
 153         while (!list_empty(&mux->psocks_ready)) {
 154                 psock = list_first_entry(&mux->psocks_ready, struct kcm_psock,
 155                                          psock_ready_list);
 156
 157                 if (kcm_queue_rcv_skb(&kcm->sk, psock->ready_rx_msg)) {
 158                         /* Assuming buffer limit has been reached */
 159                         WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
 160                         return;
 161                 }
 162
 163                 /* Consumed the ready message on the psock. Schedule rx_work to
 164                  * get more messages.
 165                  */
 166                 list_del(&psock->psock_ready_list);
 167                 psock->ready_rx_msg = NULL;
 168
 169                 /* Commit clearing of ready_rx_msg for queuing work */
 170                 smp_mb();
 171
 172                 queue_work(kcm_wq, &psock->rx_work);
 173         }
 174
 175         /* Buffer limit is okay now, add to ready list */
 176         list_add_tail(&kcm->wait_rx_list,
 177                       &kcm->mux->kcm_rx_waiters);
 178         kcm->rx_wait = true;
 179 }
 180
 181 static void kcm_rfree(struct sk_buff *skb)
 182 {
 183         struct sock *sk = skb->sk;
 184         struct kcm_sock *kcm = kcm_sk(sk);
 185         struct kcm_mux *mux = kcm->mux;
 186         unsigned int len = skb->truesize;
 187
 188         sk_mem_uncharge(sk, len);
 189         atomic_sub(len, &sk->sk_rmem_alloc);
 190
 191         /* For reading rx_wait and rx_psock without holding lock */
 192         smp_mb__after_atomic();
 193
 194         if (!kcm->rx_wait && !kcm->rx_psock &&
 195             sk_rmem_alloc_get(sk) < sk->sk_rcvlowat) {
 196                 spin_lock_bh(&mux->rx_lock);
 197                 kcm_rcv_ready(kcm);
 198                 spin_unlock_bh(&mux->rx_lock);
 199         }
 200 }
 201
 202 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb)
 203 {
 204         struct sk_buff_head *list = &sk->sk_receive_queue;
 205
 206         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
 207                 return -ENOMEM;
 208
 209         if (!sk_rmem_schedule(sk, skb, skb->truesize))
 210                 return -ENOBUFS;
 211
 212         skb->dev = NULL;
 213
 214         skb_orphan(skb);
 215         skb->sk = sk;
 216         skb->destructor = kcm_rfree;
 217         atomic_add(skb->truesize, &sk->sk_rmem_alloc);
 218         sk_mem_charge(sk, skb->truesize);
 219
 220         skb_queue_tail(list, skb);
 221
 222         if (!sock_flag(sk, SOCK_DEAD))
 223                 sk->sk_data_ready(sk);
 224
 225         return 0;
 226 }
 227
 228 /* Requeue received messages for a kcm socket to other kcm sockets. This is
 229  * called with a kcm socket is receive disabled.
 230  * RX mux lock held.
 231  */
 232 static void requeue_rx_msgs(struct kcm_mux *mux, struct sk_buff_head *head)
 233 {
 234         struct sk_buff *skb;
 235         struct kcm_sock *kcm;
 236
 237         while ((skb = __skb_dequeue(head))) {
 238                 /* Reset destructor to avoid calling kcm_rcv_ready */
 239                 skb->destructor = sock_rfree;
 240                 skb_orphan(skb);
 241 try_again:
 242                 if (list_empty(&mux->kcm_rx_waiters)) {
 243                         skb_queue_tail(&mux->rx_hold_queue, skb);
 244                         continue;
 245                 }
 246
 247                 kcm = list_first_entry(&mux->kcm_rx_waiters,
 248                                        struct kcm_sock, wait_rx_list);
 249
 250                 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 251                         /* Should mean socket buffer full */
 252                         list_del(&kcm->wait_rx_list);
 253                         kcm->rx_wait = false;
 254
 255                         /* Commit rx_wait to read in kcm_free */
 256                         smp_wmb();
 257
 258                         goto try_again;
 259                 }
 260         }
 261 }
 262
 263 /* Lower sock lock held */
 264 static struct kcm_sock *reserve_rx_kcm(struct kcm_psock *psock,
 265                                        struct sk_buff *head)
 266 {
 267         struct kcm_mux *mux = psock->mux;
 268         struct kcm_sock *kcm;
 269
 270         WARN_ON(psock->ready_rx_msg);
 271
 272         if (psock->rx_kcm)
 273                 return psock->rx_kcm;
 274
 275         spin_lock_bh(&mux->rx_lock);
 276
 277         if (psock->rx_kcm) {
 278                 spin_unlock_bh(&mux->rx_lock);
 279                 return psock->rx_kcm;
 280         }
 281
 282         kcm_update_rx_mux_stats(mux, psock);
 283
 284         if (list_empty(&mux->kcm_rx_waiters)) {
 285                 psock->ready_rx_msg = head;
 286                 list_add_tail(&psock->psock_ready_list,
 287                               &mux->psocks_ready);
 288                 spin_unlock_bh(&mux->rx_lock);
 289                 return NULL;
 290         }
 291
 292         kcm = list_first_entry(&mux->kcm_rx_waiters,
 293                                struct kcm_sock, wait_rx_list);
 294         list_del(&kcm->wait_rx_list);
 295         kcm->rx_wait = false;
 296
 297         psock->rx_kcm = kcm;
 298         kcm->rx_psock = psock;
 299
 300         spin_unlock_bh(&mux->rx_lock);
 301
 302         return kcm;
 303 }
 304
 305 static void kcm_done(struct kcm_sock *kcm);
 306
 307 static void kcm_done_work(struct work_struct *w)
 308 {
 309         kcm_done(container_of(w, struct kcm_sock, done_work));
 310 }
 311
 312 /* Lower sock held */
 313 static void unreserve_rx_kcm(struct kcm_psock *psock,
 314                              bool rcv_ready)
 315 {
 316         struct kcm_sock *kcm = psock->rx_kcm;
 317         struct kcm_mux *mux = psock->mux;
 318
 319         if (!kcm)
 320                 return;
 321
 322         spin_lock_bh(&mux->rx_lock);
 323
 324         psock->rx_kcm = NULL;
 325         kcm->rx_psock = NULL;
 326
 327         /* Commit kcm->rx_psock before sk_rmem_alloc_get to sync with
 328          * kcm_rfree
 329          */
 330         smp_mb();
 331
 332         if (unlikely(kcm->done)) {
 333                 spin_unlock_bh(&mux->rx_lock);
 334
 335                 /* Need to run kcm_done in a task since we need to qcquire
 336                  * callback locks which may already be held here.
 337                  */
 338                 INIT_WORK(&kcm->done_work, kcm_done_work);
 339                 schedule_work(&kcm->done_work);
 340                 return;
 341         }
 342
 343         if (unlikely(kcm->rx_disabled)) {
 344                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
 345         } else if (rcv_ready || unlikely(!sk_rmem_alloc_get(&kcm->sk))) {
 346                 /* Check for degenerative race with rx_wait that all
 347                  * data was dequeued (accounted for in kcm_rfree).
 348                  */
 349                 kcm_rcv_ready(kcm);
 350         }
 351         spin_unlock_bh(&mux->rx_lock);
 352 }
 353
 354 /* Macro to invoke filter function. */
 355 #define KCM_RUN_FILTER(prog, ctx) \
 356         (*prog->bpf_func)(ctx, prog->insnsi)
 357
 358 /* Lower socket lock held */
 359 static int kcm_tcp_recv(read_descriptor_t *desc, struct sk_buff *orig_skb,
 360                         unsigned int orig_offset, size_t orig_len)
 361 {
 362         struct kcm_psock *psock = (struct kcm_psock *)desc->arg.data;
 363         struct kcm_rx_msg *rxm;
 364         struct kcm_sock *kcm;
 365         struct sk_buff *head, *skb;
 366         size_t eaten = 0, cand_len;
 367         ssize_t extra;
 368         int err;
 369         bool cloned_orig = false;
 370
 371         if (psock->ready_rx_msg)
 372                 return 0;
 373
 374         head = psock->rx_skb_head;
 375         if (head) {
 376                 /* Message already in progress */
 377
 378                 if (unlikely(orig_offset)) {
 379                         /* Getting data with a non-zero offset when a message is
 380                          * in progress is not expected. If it does happen, we
 381                          * need to clone and pull since we can't deal with
 382                          * offsets in the skbs for a message expect in the head.
 383                          */
 384                         orig_skb = skb_clone(orig_skb, GFP_ATOMIC);
 385                         if (!orig_skb) {
 386                                 KCM_STATS_INCR(psock->stats.rx_mem_fail);
 387                                 desc->error = -ENOMEM;
 388                                 return 0;
 389                         }
 390                         if (!pskb_pull(orig_skb, orig_offset)) {
 391                                 KCM_STATS_INCR(psock->stats.rx_mem_fail);
 392                                 kfree_skb(orig_skb);
 393                                 desc->error = -ENOMEM;
 394                                 return 0;
 395                         }
 396                         cloned_orig = true;
 397                         orig_offset = 0;
 398                 }
 399
 400                 if (!psock->rx_skb_nextp) {
 401                         /* We are going to append to the frags_list of head.
 402                          * Need to unshare the frag_list.
 403                          */
 404                         err = skb_unclone(head, GFP_ATOMIC);
 405                         if (err) {
 406                                 KCM_STATS_INCR(psock->stats.rx_mem_fail);
 407                                 desc->error = err;
 408                                 return 0;
 409                         }
 410
 411                         if (unlikely(skb_shinfo(head)->frag_list)) {
 412                                 /* We can't append to an sk_buff that already
 413                                  * has a frag_list. We create a new head, point
 414                                  * the frag_list of that to the old head, and
 415                                  * then are able to use the old head->next for
 416                                  * appending to the message.
 417                                  */
 418                                 if (WARN_ON(head->next)) {
 419                                         desc->error = -EINVAL;
 420                                         return 0;
 421                                 }
 422
 423                                 skb = alloc_skb(0, GFP_ATOMIC);
 424                                 if (!skb) {
 425                                         KCM_STATS_INCR(psock->stats.rx_mem_fail);
 426                                         desc->error = -ENOMEM;
 427                                         return 0;
 428                                 }
 429                                 skb->len = head->len;
 430                                 skb->data_len = head->len;
 431                                 skb->truesize = head->truesize;
 432                                 *kcm_rx_msg(skb) = *kcm_rx_msg(head);
 433                                 psock->rx_skb_nextp = &head->next;
 434                                 skb_shinfo(skb)->frag_list = head;
 435                                 psock->rx_skb_head = skb;
 436                                 head = skb;
 437                         } else {
 438                                 psock->rx_skb_nextp =
 439                                     &skb_shinfo(head)->frag_list;
 440                         }
 441                 }
 442         }
 443
 444         while (eaten < orig_len) {
 445                 /* Always clone since we will consume something */
 446                 skb = skb_clone(orig_skb, GFP_ATOMIC);
 447                 if (!skb) {
 448                         KCM_STATS_INCR(psock->stats.rx_mem_fail);
 449                         desc->error = -ENOMEM;
 450                         break;
 451                 }
 452
 453                 cand_len = orig_len - eaten;
 454
 455                 head = psock->rx_skb_head;
 456                 if (!head) {
 457                         head = skb;
 458                         psock->rx_skb_head = head;
 459                         /* Will set rx_skb_nextp on next packet if needed */
 460                         psock->rx_skb_nextp = NULL;
 461                         rxm = kcm_rx_msg(head);
 462                         memset(rxm, 0, sizeof(*rxm));
 463                         rxm->offset = orig_offset + eaten;
 464                 } else {
 465                         /* Unclone since we may be appending to an skb that we
 466                          * already share a frag_list with.
 467                          */
 468                         err = skb_unclone(skb, GFP_ATOMIC);
 469                         if (err) {
 470                                 KCM_STATS_INCR(psock->stats.rx_mem_fail);
 471                                 desc->error = err;
 472                                 break;
 473                         }
 474
 475                         rxm = kcm_rx_msg(head);
 476                         *psock->rx_skb_nextp = skb;
 477                         psock->rx_skb_nextp = &skb->next;
 478                         head->data_len += skb->len;
 479                         head->len += skb->len;
 480                         head->truesize += skb->truesize;
 481                 }
 482
 483                 if (!rxm->full_len) {
 484                         ssize_t len;
 485
 486                         len = KCM_RUN_FILTER(psock->bpf_prog, head);
 487
 488                         if (!len) {
 489                                 /* Need more header to determine length */
 490                                 rxm->accum_len += cand_len;
 491                                 eaten += cand_len;
 492                                 KCM_STATS_INCR(psock->stats.rx_need_more_hdr);
 493                                 WARN_ON(eaten != orig_len);
 494                                 break;
 495                         } else if (len <= (ssize_t)head->len -
 496                                           skb->len - rxm->offset) {
 497                                 /* Length must be into new skb (and also
 498                                  * greater than zero)
 499                                  */
 500                                 KCM_STATS_INCR(psock->stats.rx_bad_hdr_len);
 501                                 desc->error = -EPROTO;
 502                                 psock->rx_skb_head = NULL;
 503                                 kcm_abort_rx_psock(psock, EPROTO, head);
 504                                 break;
 505                         }
 506
 507                         rxm->full_len = len;
 508                 }
 509
 510                 extra = (ssize_t)(rxm->accum_len + cand_len) - rxm->full_len;
 511
 512                 if (extra < 0) {
 513                         /* Message not complete yet. */
 514                         rxm->accum_len += cand_len;
 515                         eaten += cand_len;
 516                         WARN_ON(eaten != orig_len);
 517                         break;
 518                 }
 519
 520                 /* Positive extra indicates ore bytes than needed for the
 521                  * message
 522                  */
 523
 524                 WARN_ON(extra > cand_len);
 525
 526                 eaten += (cand_len - extra);
 527
 528                 /* Hurray, we have a new message! */
 529                 psock->rx_skb_head = NULL;
 530                 KCM_STATS_INCR(psock->stats.rx_msgs);
 531
 532 try_queue:
 533                 kcm = reserve_rx_kcm(psock, head);
 534                 if (!kcm) {
 535                         /* Unable to reserve a KCM, message is held in psock. */
 536                         break;
 537                 }
 538
 539                 if (kcm_queue_rcv_skb(&kcm->sk, head)) {
 540                         /* Should mean socket buffer full */
 541                         unreserve_rx_kcm(psock, false);
 542                         goto try_queue;
 543                 }
 544         }
 545
 546         if (cloned_orig)
 547                 kfree_skb(orig_skb);
 548
 549         KCM_STATS_ADD(psock->stats.rx_bytes, eaten);
 550
 551         return eaten;
 552 }
 553
 554 /* Called with lock held on lower socket */
 555 static int psock_tcp_read_sock(struct kcm_psock *psock)
 556 {
 557         read_descriptor_t desc;
 558
 559         desc.arg.data = psock;
 560         desc.error = 0;
 561         desc.count = 1; /* give more than one skb per call */
 562
 563         /* sk should be locked here, so okay to do tcp_read_sock */
 564         tcp_read_sock(psock->sk, &desc, kcm_tcp_recv);
 565
 566         unreserve_rx_kcm(psock, true);
 567
 568         return desc.error;
 569 }
 570
 571 /* Lower sock lock held */
 572 static void psock_tcp_data_ready(struct sock *sk)
 573 {
 574         struct kcm_psock *psock;
 575
 576         read_lock_bh(&sk->sk_callback_lock);
 577
 578         psock = (struct kcm_psock *)sk->sk_user_data;
 579         if (unlikely(!psock || psock->rx_stopped))
 580                 goto out;
 581
 582         if (psock->ready_rx_msg)
 583                 goto out;
 584
 585         if (psock_tcp_read_sock(psock) == -ENOMEM)
 586                 queue_delayed_work(kcm_wq, &psock->rx_delayed_work, 0);
 587
 588 out:
 589         read_unlock_bh(&sk->sk_callback_lock);
 590 }
 591
 592 static void do_psock_rx_work(struct kcm_psock *psock)
 593 {
 594         read_descriptor_t rd_desc;
 595         struct sock *csk = psock->sk;
 596
 597         /* We need the read lock to synchronize with psock_tcp_data_ready. We
 598          * need the socket lock for calling tcp_read_sock.
 599          */
 600         lock_sock(csk);
 601         read_lock_bh(&csk->sk_callback_lock);
 602
 603         if (unlikely(csk->sk_user_data != psock))
 604                 goto out;
 605
 606         if (unlikely(psock->rx_stopped))
 607                 goto out;
 608
 609         if (psock->ready_rx_msg)
 610                 goto out;
 611
 612         rd_desc.arg.data = psock;
 613
 614         if (psock_tcp_read_sock(psock) == -ENOMEM)
 615                 queue_delayed_work(kcm_wq, &psock->rx_delayed_work, 0);
 616
 617 out:
 618         read_unlock_bh(&csk->sk_callback_lock);
 619         release_sock(csk);
 620 }
 621
 622 static void psock_rx_work(struct work_struct *w)
 623 {
 624         do_psock_rx_work(container_of(w, struct kcm_psock, rx_work));
 625 }
 626
 627 static void psock_rx_delayed_work(struct work_struct *w)
 628 {
 629         do_psock_rx_work(container_of(w, struct kcm_psock,
 630                                       rx_delayed_work.work));
 631 }
 632
 633 static void psock_tcp_state_change(struct sock *sk)
 634 {
 635         /* TCP only does a POLLIN for a half close. Do a POLLHUP here
 636          * since application will normally not poll with POLLIN
 637          * on the TCP sockets.
 638          */
 639
 640         report_csk_error(sk, EPIPE);
 641 }
 642
 643 static void psock_tcp_write_space(struct sock *sk)
 644 {
 645         struct kcm_psock *psock;
 646         struct kcm_mux *mux;
 647         struct kcm_sock *kcm;
 648
 649         read_lock_bh(&sk->sk_callback_lock);
 650
 651         psock = (struct kcm_psock *)sk->sk_user_data;
 652         if (unlikely(!psock))
 653                 goto out;
 654
 655         mux = psock->mux;
 656
 657         spin_lock_bh(&mux->lock);
 658
 659         /* Check if the socket is reserved so someone is waiting for sending. */
 660         kcm = psock->tx_kcm;
 661         if (kcm)
 662                 queue_work(kcm_wq, &kcm->tx_work);
 663
 664         spin_unlock_bh(&mux->lock);
 665 out:
 666         read_unlock_bh(&sk->sk_callback_lock);
 667 }
 668
 669 static void unreserve_psock(struct kcm_sock *kcm);
 670
 671 /* kcm sock is locked. */
 672 static struct kcm_psock *reserve_psock(struct kcm_sock *kcm)
 673 {
 674         struct kcm_mux *mux = kcm->mux;
 675         struct kcm_psock *psock;
 676
 677         psock = kcm->tx_psock;
 678
 679         smp_rmb(); /* Must read tx_psock before tx_wait */
 680
 681         if (psock) {
 682                 WARN_ON(kcm->tx_wait);
 683                 if (unlikely(psock->tx_stopped))
 684                         unreserve_psock(kcm);
 685                 else
 686                         return kcm->tx_psock;
 687         }
 688
 689         spin_lock_bh(&mux->lock);
 690
 691         /* Check again under lock to see if psock was reserved for this
 692          * psock via psock_unreserve.
 693          */
 694         psock = kcm->tx_psock;
 695         if (unlikely(psock)) {
 696                 WARN_ON(kcm->tx_wait);
 697                 spin_unlock_bh(&mux->lock);
 698                 return kcm->tx_psock;
 699         }
 700
 701         if (!list_empty(&mux->psocks_avail)) {
 702                 psock = list_first_entry(&mux->psocks_avail,
 703                                          struct kcm_psock,
 704                                          psock_avail_list);
 705                 list_del(&psock->psock_avail_list);
 706                 if (kcm->tx_wait) {
 707                         list_del(&kcm->wait_psock_list);
 708                         kcm->tx_wait = false;
 709                 }
 710                 kcm->tx_psock = psock;
 711                 psock->tx_kcm = kcm;
 712                 KCM_STATS_INCR(psock->stats.reserved);
 713         } else if (!kcm->tx_wait) {
 714                 list_add_tail(&kcm->wait_psock_list,
 715                               &mux->kcm_tx_waiters);
 716                 kcm->tx_wait = true;
 717         }
 718
 719         spin_unlock_bh(&mux->lock);
 720
 721         return psock;
 722 }
 723
 724 /* mux lock held */
 725 static void psock_now_avail(struct kcm_psock *psock)
 726 {
 727         struct kcm_mux *mux = psock->mux;
 728         struct kcm_sock *kcm;
 729
 730         if (list_empty(&mux->kcm_tx_waiters)) {
 731                 list_add_tail(&psock->psock_avail_list,
 732                               &mux->psocks_avail);
 733         } else {
 734                 kcm = list_first_entry(&mux->kcm_tx_waiters,
 735                                        struct kcm_sock,
 736                                        wait_psock_list);
 737                 list_del(&kcm->wait_psock_list);
 738                 kcm->tx_wait = false;
 739                 psock->tx_kcm = kcm;
 740
 741                 /* Commit before changing tx_psock since that is read in
 742                  * reserve_psock before queuing work.
 743                  */
 744                 smp_mb();
 745
 746                 kcm->tx_psock = psock;
 747                 KCM_STATS_INCR(psock->stats.reserved);
 748                 queue_work(kcm_wq, &kcm->tx_work);
 749         }
 750 }
 751
 752 /* kcm sock is locked. */
 753 static void unreserve_psock(struct kcm_sock *kcm)
 754 {
 755         struct kcm_psock *psock;
 756         struct kcm_mux *mux = kcm->mux;
 757
 758         spin_lock_bh(&mux->lock);
 759
 760         psock = kcm->tx_psock;
 761
 762         if (WARN_ON(!psock)) {
 763                 spin_unlock_bh(&mux->lock);
 764                 return;
 765         }
 766
 767         smp_rmb(); /* Read tx_psock before tx_wait */
 768
 769         kcm_update_tx_mux_stats(mux, psock);
 770
 771         WARN_ON(kcm->tx_wait);
 772
 773         kcm->tx_psock = NULL;
 774         psock->tx_kcm = NULL;
 775         KCM_STATS_INCR(psock->stats.unreserved);
 776
 777         if (unlikely(psock->tx_stopped)) {
 778                 if (psock->done) {
 779                         /* Deferred free */
 780                         list_del(&psock->psock_list);
 781                         mux->psocks_cnt--;
 782                         sock_put(psock->sk);
 783                         fput(psock->sk->sk_socket->file);
 784                         kmem_cache_free(kcm_psockp, psock);
 785                 }
 786
 787                 /* Don't put back on available list */
 788
 789                 spin_unlock_bh(&mux->lock);
 790
 791                 return;
 792         }
 793
 794         psock_now_avail(psock);
 795
 796         spin_unlock_bh(&mux->lock);
 797 }
 798
 799 static void kcm_report_tx_retry(struct kcm_sock *kcm)
 800 {
 801         struct kcm_mux *mux = kcm->mux;
 802
 803         spin_lock_bh(&mux->lock);
 804         KCM_STATS_INCR(mux->stats.tx_retries);
 805         spin_unlock_bh(&mux->lock);
 806 }
 807
 808 /* Write any messages ready on the kcm socket.  Called with kcm sock lock
 809  * held.  Return bytes actually sent or error.
 810  */
 811 static int kcm_write_msgs(struct kcm_sock *kcm)
 812 {
 813         struct sock *sk = &kcm->sk;
 814         struct kcm_psock *psock;
 815         struct sk_buff *skb, *head;
 816         struct kcm_tx_msg *txm;
 817         unsigned short fragidx, frag_offset;
 818         unsigned int sent, total_sent = 0;
 819         int ret = 0;
 820
 821         kcm->tx_wait_more = false;
 822         psock = kcm->tx_psock;
 823         if (unlikely(psock && psock->tx_stopped)) {
 824                 /* A reserved psock was aborted asynchronously. Unreserve
 825                  * it and we'll retry the message.
 826                  */
 827                 unreserve_psock(kcm);
 828                 kcm_report_tx_retry(kcm);
 829                 if (skb_queue_empty(&sk->sk_write_queue))
 830                         return 0;
 831
 832                 kcm_tx_msg(skb_peek(&sk->sk_write_queue))->sent = 0;
 833
 834         } else if (skb_queue_empty(&sk->sk_write_queue)) {
 835                 return 0;
 836         }
 837
 838         head = skb_peek(&sk->sk_write_queue);
 839         txm = kcm_tx_msg(head);
 840
 841         if (txm->sent) {
 842                 /* Send of first skbuff in queue already in progress */
 843                 if (WARN_ON(!psock)) {
 844                         ret = -EINVAL;
 845                         goto out;
 846                 }
 847                 sent = txm->sent;
 848                 frag_offset = txm->frag_offset;
 849                 fragidx = txm->fragidx;
 850                 skb = txm->frag_skb;
 851
 852                 goto do_frag;
 853         }
 854
 855 try_again:
 856         psock = reserve_psock(kcm);
 857         if (!psock)
 858                 goto out;
 859
 860         do {
 861                 skb = head;
 862                 txm = kcm_tx_msg(head);
 863                 sent = 0;
 864
 865 do_frag_list:
 866                 if (WARN_ON(!skb_shinfo(skb)->nr_frags)) {
 867                         ret = -EINVAL;
 868                         goto out;
 869                 }
 870
 871                 for (fragidx = 0; fragidx < skb_shinfo(skb)->nr_frags;
 872                      fragidx++) {
 873                         skb_frag_t *frag;
 874
 875                         frag_offset = 0;
 876 do_frag:
 877                         frag = &skb_shinfo(skb)->frags[fragidx];
 878                         if (WARN_ON(!frag->size)) {
 879                                 ret = -EINVAL;
 880                                 goto out;
 881                         }
 882
 883                         ret = kernel_sendpage(psock->sk->sk_socket,
 884                                               frag->page.p,
 885                                               frag->page_offset + frag_offset,
 886                                               frag->size - frag_offset,
 887                                               MSG_DONTWAIT);
 888                         if (ret <= 0) {
 889                                 if (ret == -EAGAIN) {
 890                                         /* Save state to try again when there's
 891                                          * write space on the socket
 892                                          */
 893                                         txm->sent = sent;
 894                                         txm->frag_offset = frag_offset;
 895                                         txm->fragidx = fragidx;
 896                                         txm->frag_skb = skb;
 897
 898                                         ret = 0;
 899                                         goto out;
 900                                 }
 901
 902                                 /* Hard failure in sending message, abort this
 903                                  * psock since it has lost framing
 904                                  * synchonization and retry sending the
 905                                  * message from the beginning.
 906                                  */
 907                                 kcm_abort_tx_psock(psock, ret ? -ret : EPIPE,
 908                                                    true);
 909                                 unreserve_psock(kcm);
 910
 911                                 txm->sent = 0;
 912                                 kcm_report_tx_retry(kcm);
 913                                 ret = 0;
 914
 915                                 goto try_again;
 916                         }
 917
 918                         sent += ret;
 919                         frag_offset += ret;
 920                         KCM_STATS_ADD(psock->stats.tx_bytes, ret);
 921                         if (frag_offset < frag->size) {
 922                                 /* Not finished with this frag */
 923                                 goto do_frag;
 924                         }
 925                 }
 926
 927                 if (skb == head) {
 928                         if (skb_has_frag_list(skb)) {
 929                                 skb = skb_shinfo(skb)->frag_list;
 930                                 goto do_frag_list;
 931                         }
 932                 } else if (skb->next) {
 933                         skb = skb->next;
 934                         goto do_frag_list;
 935                 }
 936
 937                 /* Successfully sent the whole packet, account for it. */
 938                 skb_dequeue(&sk->sk_write_queue);
 939                 kfree_skb(head);
 940                 sk->sk_wmem_queued -= sent;
 941                 total_sent += sent;
 942                 KCM_STATS_INCR(psock->stats.tx_msgs);
 943         } while ((head = skb_peek(&sk->sk_write_queue)));
 944 out:
 945         if (!head) {
 946                 /* Done with all queued messages. */
 947                 WARN_ON(!skb_queue_empty(&sk->sk_write_queue));
 948                 unreserve_psock(kcm);
 949         }
 950
 951         /* Check if write space is available */
 952         sk->sk_write_space(sk);
 953
 954         return total_sent ? : ret;
 955 }
 956
 957 static void kcm_tx_work(struct work_struct *w)
 958 {
 959         struct kcm_sock *kcm = container_of(w, struct kcm_sock, tx_work);
 960         struct sock *sk = &kcm->sk;
 961         int err;
 962
 963         lock_sock(sk);
 964
 965         /* Primarily for SOCK_DGRAM sockets, also handle asynchronous tx
 966          * aborts
 967          */
 968         err = kcm_write_msgs(kcm);
 969         if (err < 0) {
 970                 /* Hard failure in write, report error on KCM socket */
 971                 pr_warn("KCM: Hard failure on kcm_write_msgs %d\n", err);
 972                 report_csk_error(&kcm->sk, -err);
 973                 goto out;
 974         }
 975
 976         /* Primarily for SOCK_SEQPACKET sockets */
 977         if (likely(sk->sk_socket) &&
 978             test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
 979                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 980                 sk->sk_write_space(sk);
 981         }
 982
 983 out:
 984         release_sock(sk);
 985 }
 986
 987 static void kcm_push(struct kcm_sock *kcm)
 988 {
 989         if (kcm->tx_wait_more)
 990                 kcm_write_msgs(kcm);
 991 }
 992
 993 static ssize_t kcm_sendpage(struct socket *sock, struct page *page,
 994                             int offset, size_t size, int flags)
 995
 996 {
 997         struct sock *sk = sock->sk;
 998         struct kcm_sock *kcm = kcm_sk(sk);
 999         struct sk_buff *skb = NULL, *head = NULL;
1000         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1001         bool eor;
1002         int err = 0;
1003         int i;
1004
1005         if (flags & MSG_SENDPAGE_NOTLAST)
1006                 flags |= MSG_MORE;
1007
1008         /* No MSG_EOR from splice, only look at MSG_MORE */
1009         eor = !(flags & MSG_MORE);
1010
1011         lock_sock(sk);
1012
1013         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1014
1015         err = -EPIPE;
1016         if (sk->sk_err)
1017                 goto out_error;
1018
1019         if (kcm->seq_skb) {
1020                 /* Previously opened message */
1021                 head = kcm->seq_skb;
1022                 skb = kcm_tx_msg(head)->last_skb;
1023                 i = skb_shinfo(skb)->nr_frags;
1024
1025                 if (skb_can_coalesce(skb, i, page, offset)) {
1026                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], size);
1027                         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1028                         goto coalesced;
1029                 }
1030
1031                 if (i >= MAX_SKB_FRAGS) {
1032                         struct sk_buff *tskb;
1033
1034                         tskb = alloc_skb(0, sk->sk_allocation);
1035                         while (!tskb) {
1036                                 kcm_push(kcm);
1037                                 err = sk_stream_wait_memory(sk, &timeo);
1038                                 if (err)
1039                                         goto out_error;
1040                         }
1041
1042                         if (head == skb)
1043                                 skb_shinfo(head)->frag_list = tskb;
1044                         else
1045                                 skb->next = tskb;
1046
1047                         skb = tskb;
1048                         skb->ip_summed = CHECKSUM_UNNECESSARY;
1049                         i = 0;
1050                 }
1051         } else {
1052                 /* Call the sk_stream functions to manage the sndbuf mem. */
1053                 if (!sk_stream_memory_free(sk)) {
1054                         kcm_push(kcm);
1055                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1056                         err = sk_stream_wait_memory(sk, &timeo);
1057                         if (err)
1058                                 goto out_error;
1059                 }
1060
1061                 head = alloc_skb(0, sk->sk_allocation);
1062                 while (!head) {
1063                         kcm_push(kcm);
1064                         err = sk_stream_wait_memory(sk, &timeo);
1065                         if (err)
1066                                 goto out_error;
1067                 }
1068
1069                 skb = head;
1070                 i = 0;
1071         }
1072
1073         get_page(page);
1074         skb_fill_page_desc(skb, i, page, offset, size);
1075         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1076
1077 coalesced:
1078         skb->len += size;
1079         skb->data_len += size;
1080         skb->truesize += size;
1081         sk->sk_wmem_queued += size;
1082         sk_mem_charge(sk, size);
1083
1084         if (head != skb) {
1085                 head->len += size;
1086                 head->data_len += size;
1087                 head->truesize += size;
1088         }
1089
1090         if (eor) {
1091                 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
1092
1093                 /* Message complete, queue it on send buffer */
1094                 __skb_queue_tail(&sk->sk_write_queue, head);
1095                 kcm->seq_skb = NULL;
1096                 KCM_STATS_INCR(kcm->stats.tx_msgs);
1097
1098                 if (flags & MSG_BATCH) {
1099                         kcm->tx_wait_more = true;
1100                 } else if (kcm->tx_wait_more || not_busy) {
1101                         err = kcm_write_msgs(kcm);
1102                         if (err < 0) {
1103                                 /* We got a hard error in write_msgs but have
1104                                  * already queued this message. Report an error
1105                                  * in the socket, but don't affect return value
1106                                  * from sendmsg
1107                                  */
1108                                 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
1109                                 report_csk_error(&kcm->sk, -err);
1110                         }
1111                 }
1112         } else {
1113                 /* Message not complete, save state */
1114                 kcm->seq_skb = head;
1115                 kcm_tx_msg(head)->last_skb = skb;
1116         }
1117
1118         KCM_STATS_ADD(kcm->stats.tx_bytes, size);
1119
1120         release_sock(sk);
1121         return size;
1122
1123 out_error:
1124         kcm_push(kcm);
1125
1126         err = sk_stream_error(sk, flags, err);
1127
1128         /* make sure we wake any epoll edge trigger waiter */
1129         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1130                 sk->sk_write_space(sk);
1131
1132         release_sock(sk);
1133         return err;
1134 }
1135
1136 static int kcm_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1137 {
1138         struct sock *sk = sock->sk;
1139         struct kcm_sock *kcm = kcm_sk(sk);
1140         struct sk_buff *skb = NULL, *head = NULL;
1141         size_t copy, copied = 0;
1142         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1143         int eor = (sock->type == SOCK_DGRAM) ?
1144                   !(msg->msg_flags & MSG_MORE) : !!(msg->msg_flags & MSG_EOR);
1145         int err = -EPIPE;
1146
1147         lock_sock(sk);
1148
1149         /* Per tcp_sendmsg this should be in poll */
1150         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1151
1152         if (sk->sk_err)
1153                 goto out_error;
1154
1155         if (kcm->seq_skb) {
1156                 /* Previously opened message */
1157                 head = kcm->seq_skb;
1158                 skb = kcm_tx_msg(head)->last_skb;
1159                 goto start;
1160         }
1161
1162         /* Call the sk_stream functions to manage the sndbuf mem. */
1163         if (!sk_stream_memory_free(sk)) {
1164                 kcm_push(kcm);
1165                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1166                 err = sk_stream_wait_memory(sk, &timeo);
1167                 if (err)
1168                         goto out_error;
1169         }
1170
1171         /* New message, alloc head skb */
1172         head = alloc_skb(0, sk->sk_allocation);
1173         while (!head) {
1174                 kcm_push(kcm);
1175                 err = sk_stream_wait_memory(sk, &timeo);
1176                 if (err)
1177                         goto out_error;
1178
1179                 head = alloc_skb(0, sk->sk_allocation);
1180         }
1181
1182         skb = head;
1183
1184         /* Set ip_summed to CHECKSUM_UNNECESSARY to avoid calling
1185          * csum_and_copy_from_iter from skb_do_copy_data_nocache.
1186          */
1187         skb->ip_summed = CHECKSUM_UNNECESSARY;
1188
1189 start:
1190         while (msg_data_left(msg)) {
1191                 bool merge = true;
1192                 int i = skb_shinfo(skb)->nr_frags;
1193                 struct page_frag *pfrag = sk_page_frag(sk);
1194
1195                 if (!sk_page_frag_refill(sk, pfrag))
1196                         goto wait_for_memory;
1197
1198                 if (!skb_can_coalesce(skb, i, pfrag->page,
1199                                       pfrag->offset)) {
1200                         if (i == MAX_SKB_FRAGS) {
1201                                 struct sk_buff *tskb;
1202
1203                                 tskb = alloc_skb(0, sk->sk_allocation);
1204                                 if (!tskb)
1205                                         goto wait_for_memory;
1206
1207                                 if (head == skb)
1208                                         skb_shinfo(head)->frag_list = tskb;
1209                                 else
1210                                         skb->next = tskb;
1211
1212                                 skb = tskb;
1213                                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1214                                 continue;
1215                         }
1216                         merge = false;
1217                 }
1218
1219                 copy = min_t(int, msg_data_left(msg),
1220                              pfrag->size - pfrag->offset);
1221
1222                 if (!sk_wmem_schedule(sk, copy))
1223                         goto wait_for_memory;
1224
1225                 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1226                                                pfrag->page,
1227                                                pfrag->offset,
1228                                                copy);
1229                 if (err)
1230                         goto out_error;
1231
1232                 /* Update the skb. */
1233                 if (merge) {
1234                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1235                 } else {
1236                         skb_fill_page_desc(skb, i, pfrag->page,
1237                                            pfrag->offset, copy);
1238                         get_page(pfrag->page);
1239                 }
1240
1241                 pfrag->offset += copy;
1242                 copied += copy;
1243                 if (head != skb) {
1244                         head->len += copy;
1245                         head->data_len += copy;
1246                 }
1247
1248                 continue;
1249
1250 wait_for_memory:
1251                 kcm_push(kcm);
1252                 err = sk_stream_wait_memory(sk, &timeo);
1253                 if (err)
1254                         goto out_error;
1255         }
1256
1257         if (eor) {
1258                 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
1259
1260                 /* Message complete, queue it on send buffer */
1261                 __skb_queue_tail(&sk->sk_write_queue, head);
1262                 kcm->seq_skb = NULL;
1263                 KCM_STATS_INCR(kcm->stats.tx_msgs);
1264
1265                 if (msg->msg_flags & MSG_BATCH) {
1266                         kcm->tx_wait_more = true;
1267                 } else if (kcm->tx_wait_more || not_busy) {
1268                         err = kcm_write_msgs(kcm);
1269                         if (err < 0) {
1270                                 /* We got a hard error in write_msgs but have
1271                                  * already queued this message. Report an error
1272                                  * in the socket, but don't affect return value
1273                                  * from sendmsg
1274                                  */
1275                                 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
1276                                 report_csk_error(&kcm->sk, -err);
1277                         }
1278                 }
1279         } else {
1280                 /* Message not complete, save state */
1281 partial_message:
1282                 kcm->seq_skb = head;
1283                 kcm_tx_msg(head)->last_skb = skb;
1284         }
1285
1286         KCM_STATS_ADD(kcm->stats.tx_bytes, copied);
1287
1288         release_sock(sk);
1289         return copied;
1290
1291 out_error:
1292         kcm_push(kcm);
1293
1294         if (copied && sock->type == SOCK_SEQPACKET) {
1295                 /* Wrote some bytes before encountering an
1296                  * error, return partial success.
1297                  */
1298                 goto partial_message;
1299         }
1300
1301         if (head != kcm->seq_skb)
1302                 kfree_skb(head);
1303
1304         err = sk_stream_error(sk, msg->msg_flags, err);
1305
1306         /* make sure we wake any epoll edge trigger waiter */
1307         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1308                 sk->sk_write_space(sk);
1309
1310         release_sock(sk);
1311         return err;
1312 }
1313
1314 static struct sk_buff *kcm_wait_data(struct sock *sk, int flags,
1315                                      long timeo, int *err)
1316 {
1317         struct sk_buff *skb;
1318
1319         while (!(skb = skb_peek(&sk->sk_receive_queue))) {
1320                 if (sk->sk_err) {
1321                         *err = sock_error(sk);
1322                         return NULL;
1323                 }
1324
1325                 if (sock_flag(sk, SOCK_DONE))
1326                         return NULL;
1327
1328                 if ((flags & MSG_DONTWAIT) || !timeo) {
1329                         *err = -EAGAIN;
1330                         return NULL;
1331                 }
1332
1333                 sk_wait_data(sk, &timeo, NULL);
1334
1335                 /* Handle signals */
1336                 if (signal_pending(current)) {
1337                         *err = sock_intr_errno(timeo);
1338                         return NULL;
1339                 }
1340         }
1341
1342         return skb;
1343 }
1344
1345 static int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
1346                        size_t len, int flags)
1347 {
1348         struct sock *sk = sock->sk;
1349         struct kcm_sock *kcm = kcm_sk(sk);
1350         int err = 0;
1351         long timeo;
1352         struct kcm_rx_msg *rxm;
1353         int copied = 0;
1354         struct sk_buff *skb;
1355
1356         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1357
1358         lock_sock(sk);
1359
1360         skb = kcm_wait_data(sk, flags, timeo, &err);
1361         if (!skb)
1362                 goto out;
1363
1364         /* Okay, have a message on the receive queue */
1365
1366         rxm = kcm_rx_msg(skb);
1367
1368         if (len > rxm->full_len)
1369                 len = rxm->full_len;
1370
1371         err = skb_copy_datagram_msg(skb, rxm->offset, msg, len);
1372         if (err < 0)
1373                 goto out;
1374
1375         copied = len;
1376         if (likely(!(flags & MSG_PEEK))) {
1377                 KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1378                 if (copied < rxm->full_len) {
1379                         if (sock->type == SOCK_DGRAM) {
1380                                 /* Truncated message */
1381                                 msg->msg_flags |= MSG_TRUNC;
1382                                 goto msg_finished;
1383                         }
1384                         rxm->offset += copied;
1385                         rxm->full_len -= copied;
1386                 } else {
1387 msg_finished:
1388                         /* Finished with message */
1389                         msg->msg_flags |= MSG_EOR;
1390                         KCM_STATS_INCR(kcm->stats.rx_msgs);
1391                         skb_unlink(skb, &sk->sk_receive_queue);
1392                         kfree_skb(skb);
1393                 }
1394         }
1395
1396 out:
1397         release_sock(sk);
1398
1399         return copied ? : err;
1400 }
1401
1402 static ssize_t kcm_sock_splice(struct sock *sk,
1403                                struct pipe_inode_info *pipe,
1404                                struct splice_pipe_desc *spd)
1405 {
1406         int ret;
1407
1408         release_sock(sk);
1409         ret = splice_to_pipe(pipe, spd);
1410         lock_sock(sk);
1411
1412         return ret;
1413 }
1414
1415 static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1416                                struct pipe_inode_info *pipe, size_t len,
1417                                unsigned int flags)
1418 {
1419         struct sock *sk = sock->sk;
1420         struct kcm_sock *kcm = kcm_sk(sk);
1421         long timeo;
1422         struct kcm_rx_msg *rxm;
1423         int err = 0;
1424         size_t copied;
1425         struct sk_buff *skb;
1426
1427         /* Only support splice for SOCKSEQPACKET */
1428
1429         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1430
1431         lock_sock(sk);
1432
1433         skb = kcm_wait_data(sk, flags, timeo, &err);
1434         if (!skb)
1435                 goto err_out;
1436
1437         /* Okay, have a message on the receive queue */
1438
1439         rxm = kcm_rx_msg(skb);
1440
1441         if (len > rxm->full_len)
1442                 len = rxm->full_len;
1443
1444         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, len, flags,
1445                                  kcm_sock_splice);
1446         if (copied < 0) {
1447                 err = copied;
1448                 goto err_out;
1449         }
1450
1451         KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1452
1453         rxm->offset += copied;
1454         rxm->full_len -= copied;
1455
1456         /* We have no way to return MSG_EOR. If all the bytes have been
1457          * read we still leave the message in the receive socket buffer.
1458          * A subsequent recvmsg needs to be done to return MSG_EOR and
1459          * finish reading the message.
1460          */
1461
1462         release_sock(sk);
1463
1464         return copied;
1465
1466 err_out:
1467         release_sock(sk);
1468
1469         return err;
1470 }
1471
1472 /* kcm sock lock held */
1473 static void kcm_recv_disable(struct kcm_sock *kcm)
1474 {
1475         struct kcm_mux *mux = kcm->mux;
1476
1477         if (kcm->rx_disabled)
1478                 return;
1479
1480         spin_lock_bh(&mux->rx_lock);
1481
1482         kcm->rx_disabled = 1;
1483
1484         /* If a psock is reserved we'll do cleanup in unreserve */
1485         if (!kcm->rx_psock) {
1486                 if (kcm->rx_wait) {
1487                         list_del(&kcm->wait_rx_list);
1488                         kcm->rx_wait = false;
1489                 }
1490
1491                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1492         }
1493
1494         spin_unlock_bh(&mux->rx_lock);
1495 }
1496
1497 /* kcm sock lock held */
1498 static void kcm_recv_enable(struct kcm_sock *kcm)
1499 {
1500         struct kcm_mux *mux = kcm->mux;
1501
1502         if (!kcm->rx_disabled)
1503                 return;
1504
1505         spin_lock_bh(&mux->rx_lock);
1506
1507         kcm->rx_disabled = 0;
1508         kcm_rcv_ready(kcm);
1509
1510         spin_unlock_bh(&mux->rx_lock);
1511 }
1512
1513 static int kcm_setsockopt(struct socket *sock, int level, int optname,
1514                           char __user *optval, unsigned int optlen)
1515 {
1516         struct kcm_sock *kcm = kcm_sk(sock->sk);
1517         int val, valbool;
1518         int err = 0;
1519
1520         if (level != SOL_KCM)
1521                 return -ENOPROTOOPT;
1522
1523         if (optlen < sizeof(int))
1524                 return -EINVAL;
1525
1526         if (get_user(val, (int __user *)optval))
1527                 return -EINVAL;
1528
1529         valbool = val ? 1 : 0;
1530
1531         switch (optname) {
1532         case KCM_RECV_DISABLE:
1533                 lock_sock(&kcm->sk);
1534                 if (valbool)
1535                         kcm_recv_disable(kcm);
1536                 else
1537                         kcm_recv_enable(kcm);
1538                 release_sock(&kcm->sk);
1539                 break;
1540         default:
1541                 err = -ENOPROTOOPT;
1542         }
1543
1544         return err;
1545 }
1546
1547 static int kcm_getsockopt(struct socket *sock, int level, int optname,
1548                           char __user *optval, int __user *optlen)
1549 {
1550         struct kcm_sock *kcm = kcm_sk(sock->sk);
1551         int val, len;
1552
1553         if (level != SOL_KCM)
1554                 return -ENOPROTOOPT;
1555
1556         if (get_user(len, optlen))
1557                 return -EFAULT;
1558
1559         len = min_t(unsigned int, len, sizeof(int));
1560         if (len < 0)
1561                 return -EINVAL;
1562
1563         switch (optname) {
1564         case KCM_RECV_DISABLE:
1565                 val = kcm->rx_disabled;
1566                 break;
1567         default:
1568                 return -ENOPROTOOPT;
1569         }
1570
1571         if (put_user(len, optlen))
1572                 return -EFAULT;
1573         if (copy_to_user(optval, &val, len))
1574                 return -EFAULT;
1575         return 0;
1576 }
1577
1578 static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1579 {
1580         struct kcm_sock *tkcm;
1581         struct list_head *head;
1582         int index = 0;
1583
1584         /* For SOCK_SEQPACKET sock type, datagram_poll checks the sk_state, so
1585          * we set sk_state, otherwise epoll_wait always returns right away with
1586          * POLLHUP
1587          */
1588         kcm->sk.sk_state = TCP_ESTABLISHED;
1589
1590         /* Add to mux's kcm sockets list */
1591         kcm->mux = mux;
1592         spin_lock_bh(&mux->lock);
1593
1594         head = &mux->kcm_socks;
1595         list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1596                 if (tkcm->index != index)
1597                         break;
1598                 head = &tkcm->kcm_sock_list;
1599                 index++;
1600         }
1601
1602         list_add(&kcm->kcm_sock_list, head);
1603         kcm->index = index;
1604
1605         mux->kcm_socks_cnt++;
1606         spin_unlock_bh(&mux->lock);
1607
1608         INIT_WORK(&kcm->tx_work, kcm_tx_work);
1609
1610         spin_lock_bh(&mux->rx_lock);
1611         kcm_rcv_ready(kcm);
1612         spin_unlock_bh(&mux->rx_lock);
1613 }
1614
1615 static int kcm_attach(struct socket *sock, struct socket *csock,
1616                       struct bpf_prog *prog)
1617 {
1618         struct kcm_sock *kcm = kcm_sk(sock->sk);
1619         struct kcm_mux *mux = kcm->mux;
1620         struct sock *csk;
1621         struct kcm_psock *psock = NULL, *tpsock;
1622         struct list_head *head;
1623         int index = 0;
1624
1625         if (csock->ops->family != PF_INET &&
1626             csock->ops->family != PF_INET6)
1627                 return -EINVAL;
1628
1629         csk = csock->sk;
1630         if (!csk)
1631                 return -EINVAL;
1632
1633         /* Only support TCP for now */
1634         if (csk->sk_protocol != IPPROTO_TCP)
1635                 return -EINVAL;
1636
1637         psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1638         if (!psock)
1639                 return -ENOMEM;
1640
1641         psock->mux = mux;
1642         psock->sk = csk;
1643         psock->bpf_prog = prog;
1644         INIT_WORK(&psock->rx_work, psock_rx_work);
1645         INIT_DELAYED_WORK(&psock->rx_delayed_work, psock_rx_delayed_work);
1646
1647         sock_hold(csk);
1648
1649         write_lock_bh(&csk->sk_callback_lock);
1650         psock->save_data_ready = csk->sk_data_ready;
1651         psock->save_write_space = csk->sk_write_space;
1652         psock->save_state_change = csk->sk_state_change;
1653         csk->sk_user_data = psock;
1654         csk->sk_data_ready = psock_tcp_data_ready;
1655         csk->sk_write_space = psock_tcp_write_space;
1656         csk->sk_state_change = psock_tcp_state_change;
1657         write_unlock_bh(&csk->sk_callback_lock);
1658
1659         /* Finished initialization, now add the psock to the MUX. */
1660         spin_lock_bh(&mux->lock);
1661         head = &mux->psocks;
1662         list_for_each_entry(tpsock, &mux->psocks, psock_list) {
1663                 if (tpsock->index != index)
1664                         break;
1665                 head = &tpsock->psock_list;
1666                 index++;
1667         }
1668
1669         list_add(&psock->psock_list, head);
1670         psock->index = index;
1671
1672         KCM_STATS_INCR(mux->stats.psock_attach);
1673         mux->psocks_cnt++;
1674         psock_now_avail(psock);
1675         spin_unlock_bh(&mux->lock);
1676
1677         /* Schedule RX work in case there are already bytes queued */
1678         queue_work(kcm_wq, &psock->rx_work);
1679
1680         return 0;
1681 }
1682
1683 static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1684 {
1685         struct socket *csock;
1686         struct bpf_prog *prog;
1687         int err;
1688
1689         csock = sockfd_lookup(info->fd, &err);
1690         if (!csock)
1691                 return -ENOENT;
1692
1693         prog = bpf_prog_get(info->bpf_fd);
1694         if (IS_ERR(prog)) {
1695                 err = PTR_ERR(prog);
1696                 goto out;
1697         }
1698
1699         if (prog->type != BPF_PROG_TYPE_SOCKET_FILTER) {
1700                 bpf_prog_put(prog);
1701                 err = -EINVAL;
1702                 goto out;
1703         }
1704
1705         err = kcm_attach(sock, csock, prog);
1706         if (err) {
1707                 bpf_prog_put(prog);
1708                 goto out;
1709         }
1710
1711         /* Keep reference on file also */
1712
1713         return 0;
1714 out:
1715         fput(csock->file);
1716         return err;
1717 }
1718
1719 static void kcm_unattach(struct kcm_psock *psock)
1720 {
1721         struct sock *csk = psock->sk;
1722         struct kcm_mux *mux = psock->mux;
1723
1724         /* Stop getting callbacks from TCP socket. After this there should
1725          * be no way to reserve a kcm for this psock.
1726          */
1727         write_lock_bh(&csk->sk_callback_lock);
1728         csk->sk_user_data = NULL;
1729         csk->sk_data_ready = psock->save_data_ready;
1730         csk->sk_write_space = psock->save_write_space;
1731         csk->sk_state_change = psock->save_state_change;
1732         psock->rx_stopped = 1;
1733
1734         if (WARN_ON(psock->rx_kcm)) {
1735                 write_unlock_bh(&csk->sk_callback_lock);
1736                 return;
1737         }
1738
1739         spin_lock_bh(&mux->rx_lock);
1740
1741         /* Stop receiver activities. After this point psock should not be
1742          * able to get onto ready list either through callbacks or work.
1743          */
1744         if (psock->ready_rx_msg) {
1745                 list_del(&psock->psock_ready_list);
1746                 kfree_skb(psock->ready_rx_msg);
1747                 psock->ready_rx_msg = NULL;
1748                 KCM_STATS_INCR(mux->stats.rx_ready_drops);
1749         }
1750
1751         spin_unlock_bh(&mux->rx_lock);
1752
1753         write_unlock_bh(&csk->sk_callback_lock);
1754
1755         cancel_work_sync(&psock->rx_work);
1756         cancel_delayed_work_sync(&psock->rx_delayed_work);
1757
1758         bpf_prog_put(psock->bpf_prog);
1759
1760         kfree_skb(psock->rx_skb_head);
1761         psock->rx_skb_head = NULL;
1762
1763         spin_lock_bh(&mux->lock);
1764
1765         aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1766
1767         KCM_STATS_INCR(mux->stats.psock_unattach);
1768
1769         if (psock->tx_kcm) {
1770                 /* psock was reserved.  Just mark it finished and we will clean
1771                  * up in the kcm paths, we need kcm lock which can not be
1772                  * acquired here.
1773                  */
1774                 KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1775                 spin_unlock_bh(&mux->lock);
1776
1777                 /* We are unattaching a socket that is reserved. Abort the
1778                  * socket since we may be out of sync in sending on it. We need
1779                  * to do this without the mux lock.
1780                  */
1781                 kcm_abort_tx_psock(psock, EPIPE, false);
1782
1783                 spin_lock_bh(&mux->lock);
1784                 if (!psock->tx_kcm) {
1785                         /* psock now unreserved in window mux was unlocked */
1786                         goto no_reserved;
1787                 }
1788                 psock->done = 1;
1789
1790                 /* Commit done before queuing work to process it */
1791                 smp_mb();
1792
1793                 /* Queue tx work to make sure psock->done is handled */
1794                 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
1795                 spin_unlock_bh(&mux->lock);
1796         } else {
1797 no_reserved:
1798                 if (!psock->tx_stopped)
1799                         list_del(&psock->psock_avail_list);
1800                 list_del(&psock->psock_list);
1801                 mux->psocks_cnt--;
1802                 spin_unlock_bh(&mux->lock);
1803
1804                 sock_put(csk);
1805                 fput(csk->sk_socket->file);
1806                 kmem_cache_free(kcm_psockp, psock);
1807         }
1808 }
1809
1810 static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1811 {
1812         struct kcm_sock *kcm = kcm_sk(sock->sk);
1813         struct kcm_mux *mux = kcm->mux;
1814         struct kcm_psock *psock;
1815         struct socket *csock;
1816         struct sock *csk;
1817         int err;
1818
1819         csock = sockfd_lookup(info->fd, &err);
1820         if (!csock)
1821                 return -ENOENT;
1822
1823         csk = csock->sk;
1824         if (!csk) {
1825                 err = -EINVAL;
1826                 goto out;
1827         }
1828
1829         err = -ENOENT;
1830
1831         spin_lock_bh(&mux->lock);
1832
1833         list_for_each_entry(psock, &mux->psocks, psock_list) {
1834                 if (psock->sk != csk)
1835                         continue;
1836
1837                 /* Found the matching psock */
1838
1839                 if (psock->unattaching || WARN_ON(psock->done)) {
1840                         err = -EALREADY;
1841                         break;
1842                 }
1843
1844                 psock->unattaching = 1;
1845
1846                 spin_unlock_bh(&mux->lock);
1847
1848                 kcm_unattach(psock);
1849
1850                 err = 0;
1851                 goto out;
1852         }
1853
1854         spin_unlock_bh(&mux->lock);
1855
1856 out:
1857         fput(csock->file);
1858         return err;
1859 }
1860
1861 static struct proto kcm_proto = {
1862         .name   = "KCM",
1863         .owner  = THIS_MODULE,
1864         .obj_size = sizeof(struct kcm_sock),
1865 };
1866
1867 /* Clone a kcm socket. */
1868 static int kcm_clone(struct socket *osock, struct kcm_clone *info,
1869                      struct socket **newsockp)
1870 {
1871         struct socket *newsock;
1872         struct sock *newsk;
1873         struct file *newfile;
1874         int err, newfd;
1875
1876         err = -ENFILE;
1877         newsock = sock_alloc();
1878         if (!newsock)
1879                 goto out;
1880
1881         newsock->type = osock->type;
1882         newsock->ops = osock->ops;
1883
1884         __module_get(newsock->ops->owner);
1885
1886         newfd = get_unused_fd_flags(0);
1887         if (unlikely(newfd < 0)) {
1888                 err = newfd;
1889                 goto out_fd_fail;
1890         }
1891
1892         newfile = sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1893         if (unlikely(IS_ERR(newfile))) {
1894                 err = PTR_ERR(newfile);
1895                 goto out_sock_alloc_fail;
1896         }
1897
1898         newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1899                          &kcm_proto, true);
1900         if (!newsk) {
1901                 err = -ENOMEM;
1902                 goto out_sk_alloc_fail;
1903         }
1904
1905         sock_init_data(newsock, newsk);
1906         init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1907
1908         fd_install(newfd, newfile);
1909         *newsockp = newsock;
1910         info->fd = newfd;
1911
1912         return 0;
1913
1914 out_sk_alloc_fail:
1915         fput(newfile);
1916 out_sock_alloc_fail:
1917         put_unused_fd(newfd);
1918 out_fd_fail:
1919         sock_release(newsock);
1920 out:
1921         return err;
1922 }
1923
1924 static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
1925 {
1926         int err;
1927
1928         switch (cmd) {
1929         case SIOCKCMATTACH: {
1930                 struct kcm_attach info;
1931
1932                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1933                         err = -EFAULT;
1934
1935                 err = kcm_attach_ioctl(sock, &info);
1936
1937                 break;
1938         }
1939         case SIOCKCMUNATTACH: {
1940                 struct kcm_unattach info;
1941
1942                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1943                         err = -EFAULT;
1944
1945                 err = kcm_unattach_ioctl(sock, &info);
1946
1947                 break;
1948         }
1949         case SIOCKCMCLONE: {
1950                 struct kcm_clone info;
1951                 struct socket *newsock = NULL;
1952
1953                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1954                         err = -EFAULT;
1955
1956                 err = kcm_clone(sock, &info, &newsock);
1957
1958                 if (!err) {
1959                         if (copy_to_user((void __user *)arg, &info,
1960                                          sizeof(info))) {
1961                                 err = -EFAULT;
1962                                 sock_release(newsock);
1963                         }
1964                 }
1965
1966                 break;
1967         }
1968         default:
1969                 err = -ENOIOCTLCMD;
1970                 break;
1971         }
1972
1973         return err;
1974 }
1975
1976 static void free_mux(struct rcu_head *rcu)
1977 {
1978         struct kcm_mux *mux = container_of(rcu,
1979             struct kcm_mux, rcu);
1980
1981         kmem_cache_free(kcm_muxp, mux);
1982 }
1983
1984 static void release_mux(struct kcm_mux *mux)
1985 {
1986         struct kcm_net *knet = mux->knet;
1987         struct kcm_psock *psock, *tmp_psock;
1988
1989         /* Release psocks */
1990         list_for_each_entry_safe(psock, tmp_psock,
1991                                  &mux->psocks, psock_list) {
1992                 if (!WARN_ON(psock->unattaching))
1993                         kcm_unattach(psock);
1994         }
1995
1996         if (WARN_ON(mux->psocks_cnt))
1997                 return;
1998
1999         __skb_queue_purge(&mux->rx_hold_queue);
2000
2001         mutex_lock(&knet->mutex);
2002         aggregate_mux_stats(&mux->stats, &knet->aggregate_mux_stats);
2003         aggregate_psock_stats(&mux->aggregate_psock_stats,
2004                               &knet->aggregate_psock_stats);
2005         list_del_rcu(&mux->kcm_mux_list);
2006         knet->count--;
2007         mutex_unlock(&knet->mutex);
2008
2009         call_rcu(&mux->rcu, free_mux);
2010 }
2011
2012 static void kcm_done(struct kcm_sock *kcm)
2013 {
2014         struct kcm_mux *mux = kcm->mux;
2015         struct sock *sk = &kcm->sk;
2016         int socks_cnt;
2017
2018         spin_lock_bh(&mux->rx_lock);
2019         if (kcm->rx_psock) {
2020                 /* Cleanup in unreserve_rx_kcm */
2021                 WARN_ON(kcm->done);
2022                 kcm->rx_disabled = 1;
2023                 kcm->done = 1;
2024                 spin_unlock_bh(&mux->rx_lock);
2025                 return;
2026         }
2027
2028         if (kcm->rx_wait) {
2029                 list_del(&kcm->wait_rx_list);
2030                 kcm->rx_wait = false;
2031         }
2032         /* Move any pending receive messages to other kcm sockets */
2033         requeue_rx_msgs(mux, &sk->sk_receive_queue);
2034
2035         spin_unlock_bh(&mux->rx_lock);
2036
2037         if (WARN_ON(sk_rmem_alloc_get(sk)))
2038                 return;
2039
2040         /* Detach from MUX */
2041         spin_lock_bh(&mux->lock);
2042
2043         list_del(&kcm->kcm_sock_list);
2044         mux->kcm_socks_cnt--;
2045         socks_cnt = mux->kcm_socks_cnt;
2046
2047         spin_unlock_bh(&mux->lock);
2048
2049         if (!socks_cnt) {
2050                 /* We are done with the mux now. */
2051                 release_mux(mux);
2052         }
2053
2054         WARN_ON(kcm->rx_wait);
2055
2056         sock_put(&kcm->sk);
2057 }
2058
2059 /* Called by kcm_release to close a KCM socket.
2060  * If this is the last KCM socket on the MUX, destroy the MUX.
2061  */
2062 static int kcm_release(struct socket *sock)
2063 {
2064         struct sock *sk = sock->sk;
2065         struct kcm_sock *kcm;
2066         struct kcm_mux *mux;
2067         struct kcm_psock *psock;
2068
2069         if (!sk)
2070                 return 0;
2071
2072         kcm = kcm_sk(sk);
2073         mux = kcm->mux;
2074
2075         sock_orphan(sk);
2076         kfree_skb(kcm->seq_skb);
2077
2078         lock_sock(sk);
2079         /* Purge queue under lock to avoid race condition with tx_work trying
2080          * to act when queue is nonempty. If tx_work runs after this point
2081          * it will just return.
2082          */
2083         __skb_queue_purge(&sk->sk_write_queue);
2084         release_sock(sk);
2085
2086         spin_lock_bh(&mux->lock);
2087         if (kcm->tx_wait) {
2088                 /* Take of tx_wait list, after this point there should be no way
2089                  * that a psock will be assigned to this kcm.
2090                  */
2091                 list_del(&kcm->wait_psock_list);
2092                 kcm->tx_wait = false;
2093         }
2094         spin_unlock_bh(&mux->lock);
2095
2096         /* Cancel work. After this point there should be no outside references
2097          * to the kcm socket.
2098          */
2099         cancel_work_sync(&kcm->tx_work);
2100
2101         lock_sock(sk);
2102         psock = kcm->tx_psock;
2103         if (psock) {
2104                 /* A psock was reserved, so we need to kill it since it
2105                  * may already have some bytes queued from a message. We
2106                  * need to do this after removing kcm from tx_wait list.
2107                  */
2108                 kcm_abort_tx_psock(psock, EPIPE, false);
2109                 unreserve_psock(kcm);
2110         }
2111         release_sock(sk);
2112
2113         WARN_ON(kcm->tx_wait);
2114         WARN_ON(kcm->tx_psock);
2115
2116         sock->sk = NULL;
2117
2118         kcm_done(kcm);
2119
2120         return 0;
2121 }
2122
2123 static const struct proto_ops kcm_dgram_ops = {
2124         .family =       PF_KCM,
2125         .owner =        THIS_MODULE,
2126         .release =      kcm_release,
2127         .bind =         sock_no_bind,
2128         .connect =      sock_no_connect,
2129         .socketpair =   sock_no_socketpair,
2130         .accept =       sock_no_accept,
2131         .getname =      sock_no_getname,
2132         .poll =         datagram_poll,
2133         .ioctl =        kcm_ioctl,
2134         .listen =       sock_no_listen,
2135         .shutdown =     sock_no_shutdown,
2136         .setsockopt =   kcm_setsockopt,
2137         .getsockopt =   kcm_getsockopt,
2138         .sendmsg =      kcm_sendmsg,
2139         .recvmsg =      kcm_recvmsg,
2140         .mmap =         sock_no_mmap,
2141         .sendpage =     kcm_sendpage,
2142 };
2143
2144 static const struct proto_ops kcm_seqpacket_ops = {
2145         .family =       PF_KCM,
2146         .owner =        THIS_MODULE,
2147         .release =      kcm_release,
2148         .bind =         sock_no_bind,
2149         .connect =      sock_no_connect,
2150         .socketpair =   sock_no_socketpair,
2151         .accept =       sock_no_accept,
2152         .getname =      sock_no_getname,
2153         .poll =         datagram_poll,
2154         .ioctl =        kcm_ioctl,
2155         .listen =       sock_no_listen,
2156         .shutdown =     sock_no_shutdown,
2157         .setsockopt =   kcm_setsockopt,
2158         .getsockopt =   kcm_getsockopt,
2159         .sendmsg =      kcm_sendmsg,
2160         .recvmsg =      kcm_recvmsg,
2161         .mmap =         sock_no_mmap,
2162         .sendpage =     kcm_sendpage,
2163         .splice_read =  kcm_splice_read,
2164 };
2165
2166 /* Create proto operation for kcm sockets */
2167 static int kcm_create(struct net *net, struct socket *sock,
2168                       int protocol, int kern)
2169 {
2170         struct kcm_net *knet = net_generic(net, kcm_net_id);
2171         struct sock *sk;
2172         struct kcm_mux *mux;
2173
2174         switch (sock->type) {
2175         case SOCK_DGRAM:
2176                 sock->ops = &kcm_dgram_ops;
2177                 break;
2178         case SOCK_SEQPACKET:
2179                 sock->ops = &kcm_seqpacket_ops;
2180                 break;
2181         default:
2182                 return -ESOCKTNOSUPPORT;
2183         }
2184
2185         if (protocol != KCMPROTO_CONNECTED)
2186                 return -EPROTONOSUPPORT;
2187
2188         sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
2189         if (!sk)
2190                 return -ENOMEM;
2191
2192         /* Allocate a kcm mux, shared between KCM sockets */
2193         mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
2194         if (!mux) {
2195                 sk_free(sk);
2196                 return -ENOMEM;
2197         }
2198
2199         spin_lock_init(&mux->lock);
2200         spin_lock_init(&mux->rx_lock);
2201         INIT_LIST_HEAD(&mux->kcm_socks);
2202         INIT_LIST_HEAD(&mux->kcm_rx_waiters);
2203         INIT_LIST_HEAD(&mux->kcm_tx_waiters);
2204
2205         INIT_LIST_HEAD(&mux->psocks);
2206         INIT_LIST_HEAD(&mux->psocks_ready);
2207         INIT_LIST_HEAD(&mux->psocks_avail);
2208
2209         mux->knet = knet;
2210
2211         /* Add new MUX to list */
2212         mutex_lock(&knet->mutex);
2213         list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
2214         knet->count++;
2215         mutex_unlock(&knet->mutex);
2216
2217         skb_queue_head_init(&mux->rx_hold_queue);
2218
2219         /* Init KCM socket */
2220         sock_init_data(sock, sk);
2221         init_kcm_sock(kcm_sk(sk), mux);
2222
2223         return 0;
2224 }
2225
2226 static struct net_proto_family kcm_family_ops = {
2227         .family = PF_KCM,
2228         .create = kcm_create,
2229         .owner  = THIS_MODULE,
2230 };
2231
2232 static __net_init int kcm_init_net(struct net *net)
2233 {
2234         struct kcm_net *knet = net_generic(net, kcm_net_id);
2235
2236         INIT_LIST_HEAD_RCU(&knet->mux_list);
2237         mutex_init(&knet->mutex);
2238
2239         return 0;
2240 }
2241
2242 static __net_exit void kcm_exit_net(struct net *net)
2243 {
2244         struct kcm_net *knet = net_generic(net, kcm_net_id);
2245
2246         /* All KCM sockets should be closed at this point, which should mean
2247          * that all multiplexors and psocks have been destroyed.
2248          */
2249         WARN_ON(!list_empty(&knet->mux_list));
2250 }
2251
2252 static struct pernet_operations kcm_net_ops = {
2253         .init = kcm_init_net,
2254         .exit = kcm_exit_net,
2255         .id   = &kcm_net_id,
2256         .size = sizeof(struct kcm_net),
2257 };
2258
2259 static int __init kcm_init(void)
2260 {
2261         int err = -ENOMEM;
2262
2263         kcm_muxp = kmem_cache_create("kcm_mux_cache",
2264                                      sizeof(struct kcm_mux), 0,
2265                                      SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2266         if (!kcm_muxp)
2267                 goto fail;
2268
2269         kcm_psockp = kmem_cache_create("kcm_psock_cache",
2270                                        sizeof(struct kcm_psock), 0,
2271                                         SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2272         if (!kcm_psockp)
2273                 goto fail;
2274
2275         kcm_wq = create_singlethread_workqueue("kkcmd");
2276         if (!kcm_wq)
2277                 goto fail;
2278
2279         err = proto_register(&kcm_proto, 1);
2280         if (err)
2281                 goto fail;
2282
2283         err = sock_register(&kcm_family_ops);
2284         if (err)
2285                 goto sock_register_fail;
2286
2287         err = register_pernet_device(&kcm_net_ops);
2288         if (err)
2289                 goto net_ops_fail;
2290
2291         err = kcm_proc_init();
2292         if (err)
2293                 goto proc_init_fail;
2294
2295         return 0;
2296
2297 proc_init_fail:
2298         unregister_pernet_device(&kcm_net_ops);
2299
2300 net_ops_fail:
2301         sock_unregister(PF_KCM);
2302
2303 sock_register_fail:
2304         proto_unregister(&kcm_proto);
2305
2306 fail:
2307         kmem_cache_destroy(kcm_muxp);
2308         kmem_cache_destroy(kcm_psockp);
2309
2310         if (kcm_wq)
2311                 destroy_workqueue(kcm_wq);
2312
2313         return err;
2314 }
2315
2316 static void __exit kcm_exit(void)
2317 {
2318         kcm_proc_exit();
2319         unregister_pernet_device(&kcm_net_ops);
2320         sock_unregister(PF_KCM);
2321         proto_unregister(&kcm_proto);
2322         destroy_workqueue(kcm_wq);
2323
2324         kmem_cache_destroy(kcm_muxp);
2325         kmem_cache_destroy(kcm_psockp);
2326 }
2327
2328 module_init(kcm_init);
2329 module_exit(kcm_exit);
2330
2331 MODULE_LICENSE("GPL");
2332 MODULE_ALIAS_NETPROTO(PF_KCM);
2333