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                 rxm = kcm_rx_msg(head);
 379                 if (unlikely(rxm->early_eaten)) {
 380                         /* Already some number of bytes on the receive sock
 381                          * data saved in rx_skb_head, just indicate they
 382                          * are consumed.
 383                          */
 384                         eaten = orig_len <= rxm->early_eaten ?
 385                                 orig_len : rxm->early_eaten;
 386                         rxm->early_eaten -= eaten;
 387
 388                         return eaten;
 389                 }
 390
 391                 if (unlikely(orig_offset)) {
 392                         /* Getting data with a non-zero offset when a message is
 393                          * in progress is not expected. If it does happen, we
 394                          * need to clone and pull since we can't deal with
 395                          * offsets in the skbs for a message expect in the head.
 396                          */
 397                         orig_skb = skb_clone(orig_skb, GFP_ATOMIC);
 398                         if (!orig_skb) {
 399                                 KCM_STATS_INCR(psock->stats.rx_mem_fail);
 400                                 desc->error = -ENOMEM;
 401                                 return 0;
 402                         }
 403                         if (!pskb_pull(orig_skb, orig_offset)) {
 404                                 KCM_STATS_INCR(psock->stats.rx_mem_fail);
 405                                 kfree_skb(orig_skb);
 406                                 desc->error = -ENOMEM;
 407                                 return 0;
 408                         }
 409                         cloned_orig = true;
 410                         orig_offset = 0;
 411                 }
 412
 413                 if (!psock->rx_skb_nextp) {
 414                         /* We are going to append to the frags_list of head.
 415                          * Need to unshare the frag_list.
 416                          */
 417                         err = skb_unclone(head, GFP_ATOMIC);
 418                         if (err) {
 419                                 KCM_STATS_INCR(psock->stats.rx_mem_fail);
 420                                 desc->error = err;
 421                                 return 0;
 422                         }
 423
 424                         if (unlikely(skb_shinfo(head)->frag_list)) {
 425                                 /* We can't append to an sk_buff that already
 426                                  * has a frag_list. We create a new head, point
 427                                  * the frag_list of that to the old head, and
 428                                  * then are able to use the old head->next for
 429                                  * appending to the message.
 430                                  */
 431                                 if (WARN_ON(head->next)) {
 432                                         desc->error = -EINVAL;
 433                                         return 0;
 434                                 }
 435
 436                                 skb = alloc_skb(0, GFP_ATOMIC);
 437                                 if (!skb) {
 438                                         KCM_STATS_INCR(psock->stats.rx_mem_fail);
 439                                         desc->error = -ENOMEM;
 440                                         return 0;
 441                                 }
 442                                 skb->len = head->len;
 443                                 skb->data_len = head->len;
 444                                 skb->truesize = head->truesize;
 445                                 *kcm_rx_msg(skb) = *kcm_rx_msg(head);
 446                                 psock->rx_skb_nextp = &head->next;
 447                                 skb_shinfo(skb)->frag_list = head;
 448                                 psock->rx_skb_head = skb;
 449                                 head = skb;
 450                         } else {
 451                                 psock->rx_skb_nextp =
 452                                     &skb_shinfo(head)->frag_list;
 453                         }
 454                 }
 455         }
 456
 457         while (eaten < orig_len) {
 458                 /* Always clone since we will consume something */
 459                 skb = skb_clone(orig_skb, GFP_ATOMIC);
 460                 if (!skb) {
 461                         KCM_STATS_INCR(psock->stats.rx_mem_fail);
 462                         desc->error = -ENOMEM;
 463                         break;
 464                 }
 465
 466                 cand_len = orig_len - eaten;
 467
 468                 head = psock->rx_skb_head;
 469                 if (!head) {
 470                         head = skb;
 471                         psock->rx_skb_head = head;
 472                         /* Will set rx_skb_nextp on next packet if needed */
 473                         psock->rx_skb_nextp = NULL;
 474                         rxm = kcm_rx_msg(head);
 475                         memset(rxm, 0, sizeof(*rxm));
 476                         rxm->offset = orig_offset + eaten;
 477                 } else {
 478                         /* Unclone since we may be appending to an skb that we
 479                          * already share a frag_list with.
 480                          */
 481                         err = skb_unclone(skb, GFP_ATOMIC);
 482                         if (err) {
 483                                 KCM_STATS_INCR(psock->stats.rx_mem_fail);
 484                                 desc->error = err;
 485                                 break;
 486                         }
 487
 488                         rxm = kcm_rx_msg(head);
 489                         *psock->rx_skb_nextp = skb;
 490                         psock->rx_skb_nextp = &skb->next;
 491                         head->data_len += skb->len;
 492                         head->len += skb->len;
 493                         head->truesize += skb->truesize;
 494                 }
 495
 496                 if (!rxm->full_len) {
 497                         ssize_t len;
 498
 499                         len = KCM_RUN_FILTER(psock->bpf_prog, head);
 500
 501                         if (!len) {
 502                                 /* Need more header to determine length */
 503                                 rxm->accum_len += cand_len;
 504                                 eaten += cand_len;
 505                                 KCM_STATS_INCR(psock->stats.rx_need_more_hdr);
 506                                 WARN_ON(eaten != orig_len);
 507                                 break;
 508                         } else if (len > psock->sk->sk_rcvbuf) {
 509                                 /* Message length exceeds maximum allowed */
 510                                 KCM_STATS_INCR(psock->stats.rx_msg_too_big);
 511                                 desc->error = -EMSGSIZE;
 512                                 psock->rx_skb_head = NULL;
 513                                 kcm_abort_rx_psock(psock, EMSGSIZE, head);
 514                                 break;
 515                         } else if (len <= (ssize_t)head->len -
 516                                           skb->len - rxm->offset) {
 517                                 /* Length must be into new skb (and also
 518                                  * greater than zero)
 519                                  */
 520                                 KCM_STATS_INCR(psock->stats.rx_bad_hdr_len);
 521                                 desc->error = -EPROTO;
 522                                 psock->rx_skb_head = NULL;
 523                                 kcm_abort_rx_psock(psock, EPROTO, head);
 524                                 break;
 525                         }
 526
 527                         rxm->full_len = len;
 528                 }
 529
 530                 extra = (ssize_t)(rxm->accum_len + cand_len) - rxm->full_len;
 531
 532                 if (extra < 0) {
 533                         /* Message not complete yet. */
 534                         if (rxm->full_len - rxm->accum_len >
 535                             tcp_inq(psock->sk)) {
 536                                 /* Don't have the whole messages in the socket
 537                                  * buffer. Set psock->rx_need_bytes to wait for
 538                                  * the rest of the message. Also, set "early
 539                                  * eaten" since we've already buffered the skb
 540                                  * but don't consume yet per tcp_read_sock.
 541                                  */
 542
 543                                 psock->rx_need_bytes = rxm->full_len -
 544                                                        rxm->accum_len;
 545                                 rxm->accum_len += cand_len;
 546                                 rxm->early_eaten = cand_len;
 547                                 KCM_STATS_ADD(psock->stats.rx_bytes, cand_len);
 548                                 desc->count = 0; /* Stop reading socket */
 549                                 break;
 550                         }
 551                         rxm->accum_len += cand_len;
 552                         eaten += cand_len;
 553                         WARN_ON(eaten != orig_len);
 554                         break;
 555                 }
 556
 557                 /* Positive extra indicates ore bytes than needed for the
 558                  * message
 559                  */
 560
 561                 WARN_ON(extra > cand_len);
 562
 563                 eaten += (cand_len - extra);
 564
 565                 /* Hurray, we have a new message! */
 566                 psock->rx_skb_head = NULL;
 567                 KCM_STATS_INCR(psock->stats.rx_msgs);
 568
 569 try_queue:
 570                 kcm = reserve_rx_kcm(psock, head);
 571                 if (!kcm) {
 572                         /* Unable to reserve a KCM, message is held in psock. */
 573                         break;
 574                 }
 575
 576                 if (kcm_queue_rcv_skb(&kcm->sk, head)) {
 577                         /* Should mean socket buffer full */
 578                         unreserve_rx_kcm(psock, false);
 579                         goto try_queue;
 580                 }
 581         }
 582
 583         if (cloned_orig)
 584                 kfree_skb(orig_skb);
 585
 586         KCM_STATS_ADD(psock->stats.rx_bytes, eaten);
 587
 588         return eaten;
 589 }
 590
 591 /* Called with lock held on lower socket */
 592 static int psock_tcp_read_sock(struct kcm_psock *psock)
 593 {
 594         read_descriptor_t desc;
 595
 596         desc.arg.data = psock;
 597         desc.error = 0;
 598         desc.count = 1; /* give more than one skb per call */
 599
 600         /* sk should be locked here, so okay to do tcp_read_sock */
 601         tcp_read_sock(psock->sk, &desc, kcm_tcp_recv);
 602
 603         unreserve_rx_kcm(psock, true);
 604
 605         return desc.error;
 606 }
 607
 608 /* Lower sock lock held */
 609 static void psock_tcp_data_ready(struct sock *sk)
 610 {
 611         struct kcm_psock *psock;
 612
 613         read_lock_bh(&sk->sk_callback_lock);
 614
 615         psock = (struct kcm_psock *)sk->sk_user_data;
 616         if (unlikely(!psock || psock->rx_stopped))
 617                 goto out;
 618
 619         if (psock->ready_rx_msg)
 620                 goto out;
 621
 622         if (psock->rx_need_bytes) {
 623                 if (tcp_inq(sk) >= psock->rx_need_bytes)
 624                         psock->rx_need_bytes = 0;
 625                 else
 626                         goto out;
 627         }
 628
 629         if (psock_tcp_read_sock(psock) == -ENOMEM)
 630                 queue_delayed_work(kcm_wq, &psock->rx_delayed_work, 0);
 631
 632 out:
 633         read_unlock_bh(&sk->sk_callback_lock);
 634 }
 635
 636 static void do_psock_rx_work(struct kcm_psock *psock)
 637 {
 638         read_descriptor_t rd_desc;
 639         struct sock *csk = psock->sk;
 640
 641         /* We need the read lock to synchronize with psock_tcp_data_ready. We
 642          * need the socket lock for calling tcp_read_sock.
 643          */
 644         lock_sock(csk);
 645         read_lock_bh(&csk->sk_callback_lock);
 646
 647         if (unlikely(csk->sk_user_data != psock))
 648                 goto out;
 649
 650         if (unlikely(psock->rx_stopped))
 651                 goto out;
 652
 653         if (psock->ready_rx_msg)
 654                 goto out;
 655
 656         rd_desc.arg.data = psock;
 657
 658         if (psock_tcp_read_sock(psock) == -ENOMEM)
 659                 queue_delayed_work(kcm_wq, &psock->rx_delayed_work, 0);
 660
 661 out:
 662         read_unlock_bh(&csk->sk_callback_lock);
 663         release_sock(csk);
 664 }
 665
 666 static void psock_rx_work(struct work_struct *w)
 667 {
 668         do_psock_rx_work(container_of(w, struct kcm_psock, rx_work));
 669 }
 670
 671 static void psock_rx_delayed_work(struct work_struct *w)
 672 {
 673         do_psock_rx_work(container_of(w, struct kcm_psock,
 674                                       rx_delayed_work.work));
 675 }
 676
 677 static void psock_tcp_state_change(struct sock *sk)
 678 {
 679         /* TCP only does a POLLIN for a half close. Do a POLLHUP here
 680          * since application will normally not poll with POLLIN
 681          * on the TCP sockets.
 682          */
 683
 684         report_csk_error(sk, EPIPE);
 685 }
 686
 687 static void psock_tcp_write_space(struct sock *sk)
 688 {
 689         struct kcm_psock *psock;
 690         struct kcm_mux *mux;
 691         struct kcm_sock *kcm;
 692
 693         read_lock_bh(&sk->sk_callback_lock);
 694
 695         psock = (struct kcm_psock *)sk->sk_user_data;
 696         if (unlikely(!psock))
 697                 goto out;
 698
 699         mux = psock->mux;
 700
 701         spin_lock_bh(&mux->lock);
 702
 703         /* Check if the socket is reserved so someone is waiting for sending. */
 704         kcm = psock->tx_kcm;
 705         if (kcm)
 706                 queue_work(kcm_wq, &kcm->tx_work);
 707
 708         spin_unlock_bh(&mux->lock);
 709 out:
 710         read_unlock_bh(&sk->sk_callback_lock);
 711 }
 712
 713 static void unreserve_psock(struct kcm_sock *kcm);
 714
 715 /* kcm sock is locked. */
 716 static struct kcm_psock *reserve_psock(struct kcm_sock *kcm)
 717 {
 718         struct kcm_mux *mux = kcm->mux;
 719         struct kcm_psock *psock;
 720
 721         psock = kcm->tx_psock;
 722
 723         smp_rmb(); /* Must read tx_psock before tx_wait */
 724
 725         if (psock) {
 726                 WARN_ON(kcm->tx_wait);
 727                 if (unlikely(psock->tx_stopped))
 728                         unreserve_psock(kcm);
 729                 else
 730                         return kcm->tx_psock;
 731         }
 732
 733         spin_lock_bh(&mux->lock);
 734
 735         /* Check again under lock to see if psock was reserved for this
 736          * psock via psock_unreserve.
 737          */
 738         psock = kcm->tx_psock;
 739         if (unlikely(psock)) {
 740                 WARN_ON(kcm->tx_wait);
 741                 spin_unlock_bh(&mux->lock);
 742                 return kcm->tx_psock;
 743         }
 744
 745         if (!list_empty(&mux->psocks_avail)) {
 746                 psock = list_first_entry(&mux->psocks_avail,
 747                                          struct kcm_psock,
 748                                          psock_avail_list);
 749                 list_del(&psock->psock_avail_list);
 750                 if (kcm->tx_wait) {
 751                         list_del(&kcm->wait_psock_list);
 752                         kcm->tx_wait = false;
 753                 }
 754                 kcm->tx_psock = psock;
 755                 psock->tx_kcm = kcm;
 756                 KCM_STATS_INCR(psock->stats.reserved);
 757         } else if (!kcm->tx_wait) {
 758                 list_add_tail(&kcm->wait_psock_list,
 759                               &mux->kcm_tx_waiters);
 760                 kcm->tx_wait = true;
 761         }
 762
 763         spin_unlock_bh(&mux->lock);
 764
 765         return psock;
 766 }
 767
 768 /* mux lock held */
 769 static void psock_now_avail(struct kcm_psock *psock)
 770 {
 771         struct kcm_mux *mux = psock->mux;
 772         struct kcm_sock *kcm;
 773
 774         if (list_empty(&mux->kcm_tx_waiters)) {
 775                 list_add_tail(&psock->psock_avail_list,
 776                               &mux->psocks_avail);
 777         } else {
 778                 kcm = list_first_entry(&mux->kcm_tx_waiters,
 779                                        struct kcm_sock,
 780                                        wait_psock_list);
 781                 list_del(&kcm->wait_psock_list);
 782                 kcm->tx_wait = false;
 783                 psock->tx_kcm = kcm;
 784
 785                 /* Commit before changing tx_psock since that is read in
 786                  * reserve_psock before queuing work.
 787                  */
 788                 smp_mb();
 789
 790                 kcm->tx_psock = psock;
 791                 KCM_STATS_INCR(psock->stats.reserved);
 792                 queue_work(kcm_wq, &kcm->tx_work);
 793         }
 794 }
 795
 796 /* kcm sock is locked. */
 797 static void unreserve_psock(struct kcm_sock *kcm)
 798 {
 799         struct kcm_psock *psock;
 800         struct kcm_mux *mux = kcm->mux;
 801
 802         spin_lock_bh(&mux->lock);
 803
 804         psock = kcm->tx_psock;
 805
 806         if (WARN_ON(!psock)) {
 807                 spin_unlock_bh(&mux->lock);
 808                 return;
 809         }
 810
 811         smp_rmb(); /* Read tx_psock before tx_wait */
 812
 813         kcm_update_tx_mux_stats(mux, psock);
 814
 815         WARN_ON(kcm->tx_wait);
 816
 817         kcm->tx_psock = NULL;
 818         psock->tx_kcm = NULL;
 819         KCM_STATS_INCR(psock->stats.unreserved);
 820
 821         if (unlikely(psock->tx_stopped)) {
 822                 if (psock->done) {
 823                         /* Deferred free */
 824                         list_del(&psock->psock_list);
 825                         mux->psocks_cnt--;
 826                         sock_put(psock->sk);
 827                         fput(psock->sk->sk_socket->file);
 828                         kmem_cache_free(kcm_psockp, psock);
 829                 }
 830
 831                 /* Don't put back on available list */
 832
 833                 spin_unlock_bh(&mux->lock);
 834
 835                 return;
 836         }
 837
 838         psock_now_avail(psock);
 839
 840         spin_unlock_bh(&mux->lock);
 841 }
 842
 843 static void kcm_report_tx_retry(struct kcm_sock *kcm)
 844 {
 845         struct kcm_mux *mux = kcm->mux;
 846
 847         spin_lock_bh(&mux->lock);
 848         KCM_STATS_INCR(mux->stats.tx_retries);
 849         spin_unlock_bh(&mux->lock);
 850 }
 851
 852 /* Write any messages ready on the kcm socket.  Called with kcm sock lock
 853  * held.  Return bytes actually sent or error.
 854  */
 855 static int kcm_write_msgs(struct kcm_sock *kcm)
 856 {
 857         struct sock *sk = &kcm->sk;
 858         struct kcm_psock *psock;
 859         struct sk_buff *skb, *head;
 860         struct kcm_tx_msg *txm;
 861         unsigned short fragidx, frag_offset;
 862         unsigned int sent, total_sent = 0;
 863         int ret = 0;
 864
 865         kcm->tx_wait_more = false;
 866         psock = kcm->tx_psock;
 867         if (unlikely(psock && psock->tx_stopped)) {
 868                 /* A reserved psock was aborted asynchronously. Unreserve
 869                  * it and we'll retry the message.
 870                  */
 871                 unreserve_psock(kcm);
 872                 kcm_report_tx_retry(kcm);
 873                 if (skb_queue_empty(&sk->sk_write_queue))
 874                         return 0;
 875
 876                 kcm_tx_msg(skb_peek(&sk->sk_write_queue))->sent = 0;
 877
 878         } else if (skb_queue_empty(&sk->sk_write_queue)) {
 879                 return 0;
 880         }
 881
 882         head = skb_peek(&sk->sk_write_queue);
 883         txm = kcm_tx_msg(head);
 884
 885         if (txm->sent) {
 886                 /* Send of first skbuff in queue already in progress */
 887                 if (WARN_ON(!psock)) {
 888                         ret = -EINVAL;
 889                         goto out;
 890                 }
 891                 sent = txm->sent;
 892                 frag_offset = txm->frag_offset;
 893                 fragidx = txm->fragidx;
 894                 skb = txm->frag_skb;
 895
 896                 goto do_frag;
 897         }
 898
 899 try_again:
 900         psock = reserve_psock(kcm);
 901         if (!psock)
 902                 goto out;
 903
 904         do {
 905                 skb = head;
 906                 txm = kcm_tx_msg(head);
 907                 sent = 0;
 908
 909 do_frag_list:
 910                 if (WARN_ON(!skb_shinfo(skb)->nr_frags)) {
 911                         ret = -EINVAL;
 912                         goto out;
 913                 }
 914
 915                 for (fragidx = 0; fragidx < skb_shinfo(skb)->nr_frags;
 916                      fragidx++) {
 917                         skb_frag_t *frag;
 918
 919                         frag_offset = 0;
 920 do_frag:
 921                         frag = &skb_shinfo(skb)->frags[fragidx];
 922                         if (WARN_ON(!frag->size)) {
 923                                 ret = -EINVAL;
 924                                 goto out;
 925                         }
 926
 927                         ret = kernel_sendpage(psock->sk->sk_socket,
 928                                               frag->page.p,
 929                                               frag->page_offset + frag_offset,
 930                                               frag->size - frag_offset,
 931                                               MSG_DONTWAIT);
 932                         if (ret <= 0) {
 933                                 if (ret == -EAGAIN) {
 934                                         /* Save state to try again when there's
 935                                          * write space on the socket
 936                                          */
 937                                         txm->sent = sent;
 938                                         txm->frag_offset = frag_offset;
 939                                         txm->fragidx = fragidx;
 940                                         txm->frag_skb = skb;
 941
 942                                         ret = 0;
 943                                         goto out;
 944                                 }
 945
 946                                 /* Hard failure in sending message, abort this
 947                                  * psock since it has lost framing
 948                                  * synchonization and retry sending the
 949                                  * message from the beginning.
 950                                  */
 951                                 kcm_abort_tx_psock(psock, ret ? -ret : EPIPE,
 952                                                    true);
 953                                 unreserve_psock(kcm);
 954
 955                                 txm->sent = 0;
 956                                 kcm_report_tx_retry(kcm);
 957                                 ret = 0;
 958
 959                                 goto try_again;
 960                         }
 961
 962                         sent += ret;
 963                         frag_offset += ret;
 964                         KCM_STATS_ADD(psock->stats.tx_bytes, ret);
 965                         if (frag_offset < frag->size) {
 966                                 /* Not finished with this frag */
 967                                 goto do_frag;
 968                         }
 969                 }
 970
 971                 if (skb == head) {
 972                         if (skb_has_frag_list(skb)) {
 973                                 skb = skb_shinfo(skb)->frag_list;
 974                                 goto do_frag_list;
 975                         }
 976                 } else if (skb->next) {
 977                         skb = skb->next;
 978                         goto do_frag_list;
 979                 }
 980
 981                 /* Successfully sent the whole packet, account for it. */
 982                 skb_dequeue(&sk->sk_write_queue);
 983                 kfree_skb(head);
 984                 sk->sk_wmem_queued -= sent;
 985                 total_sent += sent;
 986                 KCM_STATS_INCR(psock->stats.tx_msgs);
 987         } while ((head = skb_peek(&sk->sk_write_queue)));
 988 out:
 989         if (!head) {
 990                 /* Done with all queued messages. */
 991                 WARN_ON(!skb_queue_empty(&sk->sk_write_queue));
 992                 unreserve_psock(kcm);
 993         }
 994
 995         /* Check if write space is available */
 996         sk->sk_write_space(sk);
 997
 998         return total_sent ? : ret;
 999 }
1000
1001 static void kcm_tx_work(struct work_struct *w)
1002 {
1003         struct kcm_sock *kcm = container_of(w, struct kcm_sock, tx_work);
1004         struct sock *sk = &kcm->sk;
1005         int err;
1006
1007         lock_sock(sk);
1008
1009         /* Primarily for SOCK_DGRAM sockets, also handle asynchronous tx
1010          * aborts
1011          */
1012         err = kcm_write_msgs(kcm);
1013         if (err < 0) {
1014                 /* Hard failure in write, report error on KCM socket */
1015                 pr_warn("KCM: Hard failure on kcm_write_msgs %d\n", err);
1016                 report_csk_error(&kcm->sk, -err);
1017                 goto out;
1018         }
1019
1020         /* Primarily for SOCK_SEQPACKET sockets */
1021         if (likely(sk->sk_socket) &&
1022             test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1023                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1024                 sk->sk_write_space(sk);
1025         }
1026
1027 out:
1028         release_sock(sk);
1029 }
1030
1031 static void kcm_push(struct kcm_sock *kcm)
1032 {
1033         if (kcm->tx_wait_more)
1034                 kcm_write_msgs(kcm);
1035 }
1036
1037 static ssize_t kcm_sendpage(struct socket *sock, struct page *page,
1038                             int offset, size_t size, int flags)
1039
1040 {
1041         struct sock *sk = sock->sk;
1042         struct kcm_sock *kcm = kcm_sk(sk);
1043         struct sk_buff *skb = NULL, *head = NULL;
1044         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1045         bool eor;
1046         int err = 0;
1047         int i;
1048
1049         if (flags & MSG_SENDPAGE_NOTLAST)
1050                 flags |= MSG_MORE;
1051
1052         /* No MSG_EOR from splice, only look at MSG_MORE */
1053         eor = !(flags & MSG_MORE);
1054
1055         lock_sock(sk);
1056
1057         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1058
1059         err = -EPIPE;
1060         if (sk->sk_err)
1061                 goto out_error;
1062
1063         if (kcm->seq_skb) {
1064                 /* Previously opened message */
1065                 head = kcm->seq_skb;
1066                 skb = kcm_tx_msg(head)->last_skb;
1067                 i = skb_shinfo(skb)->nr_frags;
1068
1069                 if (skb_can_coalesce(skb, i, page, offset)) {
1070                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], size);
1071                         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1072                         goto coalesced;
1073                 }
1074
1075                 if (i >= MAX_SKB_FRAGS) {
1076                         struct sk_buff *tskb;
1077
1078                         tskb = alloc_skb(0, sk->sk_allocation);
1079                         while (!tskb) {
1080                                 kcm_push(kcm);
1081                                 err = sk_stream_wait_memory(sk, &timeo);
1082                                 if (err)
1083                                         goto out_error;
1084                         }
1085
1086                         if (head == skb)
1087                                 skb_shinfo(head)->frag_list = tskb;
1088                         else
1089                                 skb->next = tskb;
1090
1091                         skb = tskb;
1092                         skb->ip_summed = CHECKSUM_UNNECESSARY;
1093                         i = 0;
1094                 }
1095         } else {
1096                 /* Call the sk_stream functions to manage the sndbuf mem. */
1097                 if (!sk_stream_memory_free(sk)) {
1098                         kcm_push(kcm);
1099                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1100                         err = sk_stream_wait_memory(sk, &timeo);
1101                         if (err)
1102                                 goto out_error;
1103                 }
1104
1105                 head = alloc_skb(0, sk->sk_allocation);
1106                 while (!head) {
1107                         kcm_push(kcm);
1108                         err = sk_stream_wait_memory(sk, &timeo);
1109                         if (err)
1110                                 goto out_error;
1111                 }
1112
1113                 skb = head;
1114                 i = 0;
1115         }
1116
1117         get_page(page);
1118         skb_fill_page_desc(skb, i, page, offset, size);
1119         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1120
1121 coalesced:
1122         skb->len += size;
1123         skb->data_len += size;
1124         skb->truesize += size;
1125         sk->sk_wmem_queued += size;
1126         sk_mem_charge(sk, size);
1127
1128         if (head != skb) {
1129                 head->len += size;
1130                 head->data_len += size;
1131                 head->truesize += size;
1132         }
1133
1134         if (eor) {
1135                 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
1136
1137                 /* Message complete, queue it on send buffer */
1138                 __skb_queue_tail(&sk->sk_write_queue, head);
1139                 kcm->seq_skb = NULL;
1140                 KCM_STATS_INCR(kcm->stats.tx_msgs);
1141
1142                 if (flags & MSG_BATCH) {
1143                         kcm->tx_wait_more = true;
1144                 } else if (kcm->tx_wait_more || not_busy) {
1145                         err = kcm_write_msgs(kcm);
1146                         if (err < 0) {
1147                                 /* We got a hard error in write_msgs but have
1148                                  * already queued this message. Report an error
1149                                  * in the socket, but don't affect return value
1150                                  * from sendmsg
1151                                  */
1152                                 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
1153                                 report_csk_error(&kcm->sk, -err);
1154                         }
1155                 }
1156         } else {
1157                 /* Message not complete, save state */
1158                 kcm->seq_skb = head;
1159                 kcm_tx_msg(head)->last_skb = skb;
1160         }
1161
1162         KCM_STATS_ADD(kcm->stats.tx_bytes, size);
1163
1164         release_sock(sk);
1165         return size;
1166
1167 out_error:
1168         kcm_push(kcm);
1169
1170         err = sk_stream_error(sk, flags, err);
1171
1172         /* make sure we wake any epoll edge trigger waiter */
1173         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1174                 sk->sk_write_space(sk);
1175
1176         release_sock(sk);
1177         return err;
1178 }
1179
1180 static int kcm_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1181 {
1182         struct sock *sk = sock->sk;
1183         struct kcm_sock *kcm = kcm_sk(sk);
1184         struct sk_buff *skb = NULL, *head = NULL;
1185         size_t copy, copied = 0;
1186         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1187         int eor = (sock->type == SOCK_DGRAM) ?
1188                   !(msg->msg_flags & MSG_MORE) : !!(msg->msg_flags & MSG_EOR);
1189         int err = -EPIPE;
1190
1191         lock_sock(sk);
1192
1193         /* Per tcp_sendmsg this should be in poll */
1194         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1195
1196         if (sk->sk_err)
1197                 goto out_error;
1198
1199         if (kcm->seq_skb) {
1200                 /* Previously opened message */
1201                 head = kcm->seq_skb;
1202                 skb = kcm_tx_msg(head)->last_skb;
1203                 goto start;
1204         }
1205
1206         /* Call the sk_stream functions to manage the sndbuf mem. */
1207         if (!sk_stream_memory_free(sk)) {
1208                 kcm_push(kcm);
1209                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1210                 err = sk_stream_wait_memory(sk, &timeo);
1211                 if (err)
1212                         goto out_error;
1213         }
1214
1215         /* New message, alloc head skb */
1216         head = alloc_skb(0, sk->sk_allocation);
1217         while (!head) {
1218                 kcm_push(kcm);
1219                 err = sk_stream_wait_memory(sk, &timeo);
1220                 if (err)
1221                         goto out_error;
1222
1223                 head = alloc_skb(0, sk->sk_allocation);
1224         }
1225
1226         skb = head;
1227
1228         /* Set ip_summed to CHECKSUM_UNNECESSARY to avoid calling
1229          * csum_and_copy_from_iter from skb_do_copy_data_nocache.
1230          */
1231         skb->ip_summed = CHECKSUM_UNNECESSARY;
1232
1233 start:
1234         while (msg_data_left(msg)) {
1235                 bool merge = true;
1236                 int i = skb_shinfo(skb)->nr_frags;
1237                 struct page_frag *pfrag = sk_page_frag(sk);
1238
1239                 if (!sk_page_frag_refill(sk, pfrag))
1240                         goto wait_for_memory;
1241
1242                 if (!skb_can_coalesce(skb, i, pfrag->page,
1243                                       pfrag->offset)) {
1244                         if (i == MAX_SKB_FRAGS) {
1245                                 struct sk_buff *tskb;
1246
1247                                 tskb = alloc_skb(0, sk->sk_allocation);
1248                                 if (!tskb)
1249                                         goto wait_for_memory;
1250
1251                                 if (head == skb)
1252                                         skb_shinfo(head)->frag_list = tskb;
1253                                 else
1254                                         skb->next = tskb;
1255
1256                                 skb = tskb;
1257                                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1258                                 continue;
1259                         }
1260                         merge = false;
1261                 }
1262
1263                 copy = min_t(int, msg_data_left(msg),
1264                              pfrag->size - pfrag->offset);
1265
1266                 if (!sk_wmem_schedule(sk, copy))
1267                         goto wait_for_memory;
1268
1269                 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1270                                                pfrag->page,
1271                                                pfrag->offset,
1272                                                copy);
1273                 if (err)
1274                         goto out_error;
1275
1276                 /* Update the skb. */
1277                 if (merge) {
1278                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1279                 } else {
1280                         skb_fill_page_desc(skb, i, pfrag->page,
1281                                            pfrag->offset, copy);
1282                         get_page(pfrag->page);
1283                 }
1284
1285                 pfrag->offset += copy;
1286                 copied += copy;
1287                 if (head != skb) {
1288                         head->len += copy;
1289                         head->data_len += copy;
1290                 }
1291
1292                 continue;
1293
1294 wait_for_memory:
1295                 kcm_push(kcm);
1296                 err = sk_stream_wait_memory(sk, &timeo);
1297                 if (err)
1298                         goto out_error;
1299         }
1300
1301         if (eor) {
1302                 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
1303
1304                 /* Message complete, queue it on send buffer */
1305                 __skb_queue_tail(&sk->sk_write_queue, head);
1306                 kcm->seq_skb = NULL;
1307                 KCM_STATS_INCR(kcm->stats.tx_msgs);
1308
1309                 if (msg->msg_flags & MSG_BATCH) {
1310                         kcm->tx_wait_more = true;
1311                 } else if (kcm->tx_wait_more || not_busy) {
1312                         err = kcm_write_msgs(kcm);
1313                         if (err < 0) {
1314                                 /* We got a hard error in write_msgs but have
1315                                  * already queued this message. Report an error
1316                                  * in the socket, but don't affect return value
1317                                  * from sendmsg
1318                                  */
1319                                 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
1320                                 report_csk_error(&kcm->sk, -err);
1321                         }
1322                 }
1323         } else {
1324                 /* Message not complete, save state */
1325 partial_message:
1326                 kcm->seq_skb = head;
1327                 kcm_tx_msg(head)->last_skb = skb;
1328         }
1329
1330         KCM_STATS_ADD(kcm->stats.tx_bytes, copied);
1331
1332         release_sock(sk);
1333         return copied;
1334
1335 out_error:
1336         kcm_push(kcm);
1337
1338         if (copied && sock->type == SOCK_SEQPACKET) {
1339                 /* Wrote some bytes before encountering an
1340                  * error, return partial success.
1341                  */
1342                 goto partial_message;
1343         }
1344
1345         if (head != kcm->seq_skb)
1346                 kfree_skb(head);
1347
1348         err = sk_stream_error(sk, msg->msg_flags, err);
1349
1350         /* make sure we wake any epoll edge trigger waiter */
1351         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1352                 sk->sk_write_space(sk);
1353
1354         release_sock(sk);
1355         return err;
1356 }
1357
1358 static struct sk_buff *kcm_wait_data(struct sock *sk, int flags,
1359                                      long timeo, int *err)
1360 {
1361         struct sk_buff *skb;
1362
1363         while (!(skb = skb_peek(&sk->sk_receive_queue))) {
1364                 if (sk->sk_err) {
1365                         *err = sock_error(sk);
1366                         return NULL;
1367                 }
1368
1369                 if (sock_flag(sk, SOCK_DONE))
1370                         return NULL;
1371
1372                 if ((flags & MSG_DONTWAIT) || !timeo) {
1373                         *err = -EAGAIN;
1374                         return NULL;
1375                 }
1376
1377                 sk_wait_data(sk, &timeo, NULL);
1378
1379                 /* Handle signals */
1380                 if (signal_pending(current)) {
1381                         *err = sock_intr_errno(timeo);
1382                         return NULL;
1383                 }
1384         }
1385
1386         return skb;
1387 }
1388
1389 static int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
1390                        size_t len, int flags)
1391 {
1392         struct sock *sk = sock->sk;
1393         struct kcm_sock *kcm = kcm_sk(sk);
1394         int err = 0;
1395         long timeo;
1396         struct kcm_rx_msg *rxm;
1397         int copied = 0;
1398         struct sk_buff *skb;
1399
1400         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1401
1402         lock_sock(sk);
1403
1404         skb = kcm_wait_data(sk, flags, timeo, &err);
1405         if (!skb)
1406                 goto out;
1407
1408         /* Okay, have a message on the receive queue */
1409
1410         rxm = kcm_rx_msg(skb);
1411
1412         if (len > rxm->full_len)
1413                 len = rxm->full_len;
1414
1415         err = skb_copy_datagram_msg(skb, rxm->offset, msg, len);
1416         if (err < 0)
1417                 goto out;
1418
1419         copied = len;
1420         if (likely(!(flags & MSG_PEEK))) {
1421                 KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1422                 if (copied < rxm->full_len) {
1423                         if (sock->type == SOCK_DGRAM) {
1424                                 /* Truncated message */
1425                                 msg->msg_flags |= MSG_TRUNC;
1426                                 goto msg_finished;
1427                         }
1428                         rxm->offset += copied;
1429                         rxm->full_len -= copied;
1430                 } else {
1431 msg_finished:
1432                         /* Finished with message */
1433                         msg->msg_flags |= MSG_EOR;
1434                         KCM_STATS_INCR(kcm->stats.rx_msgs);
1435                         skb_unlink(skb, &sk->sk_receive_queue);
1436                         kfree_skb(skb);
1437                 }
1438         }
1439
1440 out:
1441         release_sock(sk);
1442
1443         return copied ? : err;
1444 }
1445
1446 static ssize_t kcm_sock_splice(struct sock *sk,
1447                                struct pipe_inode_info *pipe,
1448                                struct splice_pipe_desc *spd)
1449 {
1450         int ret;
1451
1452         release_sock(sk);
1453         ret = splice_to_pipe(pipe, spd);
1454         lock_sock(sk);
1455
1456         return ret;
1457 }
1458
1459 static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1460                                struct pipe_inode_info *pipe, size_t len,
1461                                unsigned int flags)
1462 {
1463         struct sock *sk = sock->sk;
1464         struct kcm_sock *kcm = kcm_sk(sk);
1465         long timeo;
1466         struct kcm_rx_msg *rxm;
1467         int err = 0;
1468         size_t copied;
1469         struct sk_buff *skb;
1470
1471         /* Only support splice for SOCKSEQPACKET */
1472
1473         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1474
1475         lock_sock(sk);
1476
1477         skb = kcm_wait_data(sk, flags, timeo, &err);
1478         if (!skb)
1479                 goto err_out;
1480
1481         /* Okay, have a message on the receive queue */
1482
1483         rxm = kcm_rx_msg(skb);
1484
1485         if (len > rxm->full_len)
1486                 len = rxm->full_len;
1487
1488         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, len, flags,
1489                                  kcm_sock_splice);
1490         if (copied < 0) {
1491                 err = copied;
1492                 goto err_out;
1493         }
1494
1495         KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1496
1497         rxm->offset += copied;
1498         rxm->full_len -= copied;
1499
1500         /* We have no way to return MSG_EOR. If all the bytes have been
1501          * read we still leave the message in the receive socket buffer.
1502          * A subsequent recvmsg needs to be done to return MSG_EOR and
1503          * finish reading the message.
1504          */
1505
1506         release_sock(sk);
1507
1508         return copied;
1509
1510 err_out:
1511         release_sock(sk);
1512
1513         return err;
1514 }
1515
1516 /* kcm sock lock held */
1517 static void kcm_recv_disable(struct kcm_sock *kcm)
1518 {
1519         struct kcm_mux *mux = kcm->mux;
1520
1521         if (kcm->rx_disabled)
1522                 return;
1523
1524         spin_lock_bh(&mux->rx_lock);
1525
1526         kcm->rx_disabled = 1;
1527
1528         /* If a psock is reserved we'll do cleanup in unreserve */
1529         if (!kcm->rx_psock) {
1530                 if (kcm->rx_wait) {
1531                         list_del(&kcm->wait_rx_list);
1532                         kcm->rx_wait = false;
1533                 }
1534
1535                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1536         }
1537
1538         spin_unlock_bh(&mux->rx_lock);
1539 }
1540
1541 /* kcm sock lock held */
1542 static void kcm_recv_enable(struct kcm_sock *kcm)
1543 {
1544         struct kcm_mux *mux = kcm->mux;
1545
1546         if (!kcm->rx_disabled)
1547                 return;
1548
1549         spin_lock_bh(&mux->rx_lock);
1550
1551         kcm->rx_disabled = 0;
1552         kcm_rcv_ready(kcm);
1553
1554         spin_unlock_bh(&mux->rx_lock);
1555 }
1556
1557 static int kcm_setsockopt(struct socket *sock, int level, int optname,
1558                           char __user *optval, unsigned int optlen)
1559 {
1560         struct kcm_sock *kcm = kcm_sk(sock->sk);
1561         int val, valbool;
1562         int err = 0;
1563
1564         if (level != SOL_KCM)
1565                 return -ENOPROTOOPT;
1566
1567         if (optlen < sizeof(int))
1568                 return -EINVAL;
1569
1570         if (get_user(val, (int __user *)optval))
1571                 return -EINVAL;
1572
1573         valbool = val ? 1 : 0;
1574
1575         switch (optname) {
1576         case KCM_RECV_DISABLE:
1577                 lock_sock(&kcm->sk);
1578                 if (valbool)
1579                         kcm_recv_disable(kcm);
1580                 else
1581                         kcm_recv_enable(kcm);
1582                 release_sock(&kcm->sk);
1583                 break;
1584         default:
1585                 err = -ENOPROTOOPT;
1586         }
1587
1588         return err;
1589 }
1590
1591 static int kcm_getsockopt(struct socket *sock, int level, int optname,
1592                           char __user *optval, int __user *optlen)
1593 {
1594         struct kcm_sock *kcm = kcm_sk(sock->sk);
1595         int val, len;
1596
1597         if (level != SOL_KCM)
1598                 return -ENOPROTOOPT;
1599
1600         if (get_user(len, optlen))
1601                 return -EFAULT;
1602
1603         len = min_t(unsigned int, len, sizeof(int));
1604         if (len < 0)
1605                 return -EINVAL;
1606
1607         switch (optname) {
1608         case KCM_RECV_DISABLE:
1609                 val = kcm->rx_disabled;
1610                 break;
1611         default:
1612                 return -ENOPROTOOPT;
1613         }
1614
1615         if (put_user(len, optlen))
1616                 return -EFAULT;
1617         if (copy_to_user(optval, &val, len))
1618                 return -EFAULT;
1619         return 0;
1620 }
1621
1622 static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1623 {
1624         struct kcm_sock *tkcm;
1625         struct list_head *head;
1626         int index = 0;
1627
1628         /* For SOCK_SEQPACKET sock type, datagram_poll checks the sk_state, so
1629          * we set sk_state, otherwise epoll_wait always returns right away with
1630          * POLLHUP
1631          */
1632         kcm->sk.sk_state = TCP_ESTABLISHED;
1633
1634         /* Add to mux's kcm sockets list */
1635         kcm->mux = mux;
1636         spin_lock_bh(&mux->lock);
1637
1638         head = &mux->kcm_socks;
1639         list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1640                 if (tkcm->index != index)
1641                         break;
1642                 head = &tkcm->kcm_sock_list;
1643                 index++;
1644         }
1645
1646         list_add(&kcm->kcm_sock_list, head);
1647         kcm->index = index;
1648
1649         mux->kcm_socks_cnt++;
1650         spin_unlock_bh(&mux->lock);
1651
1652         INIT_WORK(&kcm->tx_work, kcm_tx_work);
1653
1654         spin_lock_bh(&mux->rx_lock);
1655         kcm_rcv_ready(kcm);
1656         spin_unlock_bh(&mux->rx_lock);
1657 }
1658
1659 static int kcm_attach(struct socket *sock, struct socket *csock,
1660                       struct bpf_prog *prog)
1661 {
1662         struct kcm_sock *kcm = kcm_sk(sock->sk);
1663         struct kcm_mux *mux = kcm->mux;
1664         struct sock *csk;
1665         struct kcm_psock *psock = NULL, *tpsock;
1666         struct list_head *head;
1667         int index = 0;
1668
1669         if (csock->ops->family != PF_INET &&
1670             csock->ops->family != PF_INET6)
1671                 return -EINVAL;
1672
1673         csk = csock->sk;
1674         if (!csk)
1675                 return -EINVAL;
1676
1677         /* Only support TCP for now */
1678         if (csk->sk_protocol != IPPROTO_TCP)
1679                 return -EINVAL;
1680
1681         psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1682         if (!psock)
1683                 return -ENOMEM;
1684
1685         psock->mux = mux;
1686         psock->sk = csk;
1687         psock->bpf_prog = prog;
1688         INIT_WORK(&psock->rx_work, psock_rx_work);
1689         INIT_DELAYED_WORK(&psock->rx_delayed_work, psock_rx_delayed_work);
1690
1691         sock_hold(csk);
1692
1693         write_lock_bh(&csk->sk_callback_lock);
1694         psock->save_data_ready = csk->sk_data_ready;
1695         psock->save_write_space = csk->sk_write_space;
1696         psock->save_state_change = csk->sk_state_change;
1697         csk->sk_user_data = psock;
1698         csk->sk_data_ready = psock_tcp_data_ready;
1699         csk->sk_write_space = psock_tcp_write_space;
1700         csk->sk_state_change = psock_tcp_state_change;
1701         write_unlock_bh(&csk->sk_callback_lock);
1702
1703         /* Finished initialization, now add the psock to the MUX. */
1704         spin_lock_bh(&mux->lock);
1705         head = &mux->psocks;
1706         list_for_each_entry(tpsock, &mux->psocks, psock_list) {
1707                 if (tpsock->index != index)
1708                         break;
1709                 head = &tpsock->psock_list;
1710                 index++;
1711         }
1712
1713         list_add(&psock->psock_list, head);
1714         psock->index = index;
1715
1716         KCM_STATS_INCR(mux->stats.psock_attach);
1717         mux->psocks_cnt++;
1718         psock_now_avail(psock);
1719         spin_unlock_bh(&mux->lock);
1720
1721         /* Schedule RX work in case there are already bytes queued */
1722         queue_work(kcm_wq, &psock->rx_work);
1723
1724         return 0;
1725 }
1726
1727 static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1728 {
1729         struct socket *csock;
1730         struct bpf_prog *prog;
1731         int err;
1732
1733         csock = sockfd_lookup(info->fd, &err);
1734         if (!csock)
1735                 return -ENOENT;
1736
1737         prog = bpf_prog_get(info->bpf_fd);
1738         if (IS_ERR(prog)) {
1739                 err = PTR_ERR(prog);
1740                 goto out;
1741         }
1742
1743         if (prog->type != BPF_PROG_TYPE_SOCKET_FILTER) {
1744                 bpf_prog_put(prog);
1745                 err = -EINVAL;
1746                 goto out;
1747         }
1748
1749         err = kcm_attach(sock, csock, prog);
1750         if (err) {
1751                 bpf_prog_put(prog);
1752                 goto out;
1753         }
1754
1755         /* Keep reference on file also */
1756
1757         return 0;
1758 out:
1759         fput(csock->file);
1760         return err;
1761 }
1762
1763 static void kcm_unattach(struct kcm_psock *psock)
1764 {
1765         struct sock *csk = psock->sk;
1766         struct kcm_mux *mux = psock->mux;
1767
1768         /* Stop getting callbacks from TCP socket. After this there should
1769          * be no way to reserve a kcm for this psock.
1770          */
1771         write_lock_bh(&csk->sk_callback_lock);
1772         csk->sk_user_data = NULL;
1773         csk->sk_data_ready = psock->save_data_ready;
1774         csk->sk_write_space = psock->save_write_space;
1775         csk->sk_state_change = psock->save_state_change;
1776         psock->rx_stopped = 1;
1777
1778         if (WARN_ON(psock->rx_kcm)) {
1779                 write_unlock_bh(&csk->sk_callback_lock);
1780                 return;
1781         }
1782
1783         spin_lock_bh(&mux->rx_lock);
1784
1785         /* Stop receiver activities. After this point psock should not be
1786          * able to get onto ready list either through callbacks or work.
1787          */
1788         if (psock->ready_rx_msg) {
1789                 list_del(&psock->psock_ready_list);
1790                 kfree_skb(psock->ready_rx_msg);
1791                 psock->ready_rx_msg = NULL;
1792                 KCM_STATS_INCR(mux->stats.rx_ready_drops);
1793         }
1794
1795         spin_unlock_bh(&mux->rx_lock);
1796
1797         write_unlock_bh(&csk->sk_callback_lock);
1798
1799         cancel_work_sync(&psock->rx_work);
1800         cancel_delayed_work_sync(&psock->rx_delayed_work);
1801
1802         bpf_prog_put(psock->bpf_prog);
1803
1804         kfree_skb(psock->rx_skb_head);
1805         psock->rx_skb_head = NULL;
1806
1807         spin_lock_bh(&mux->lock);
1808
1809         aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1810
1811         KCM_STATS_INCR(mux->stats.psock_unattach);
1812
1813         if (psock->tx_kcm) {
1814                 /* psock was reserved.  Just mark it finished and we will clean
1815                  * up in the kcm paths, we need kcm lock which can not be
1816                  * acquired here.
1817                  */
1818                 KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1819                 spin_unlock_bh(&mux->lock);
1820
1821                 /* We are unattaching a socket that is reserved. Abort the
1822                  * socket since we may be out of sync in sending on it. We need
1823                  * to do this without the mux lock.
1824                  */
1825                 kcm_abort_tx_psock(psock, EPIPE, false);
1826
1827                 spin_lock_bh(&mux->lock);
1828                 if (!psock->tx_kcm) {
1829                         /* psock now unreserved in window mux was unlocked */
1830                         goto no_reserved;
1831                 }
1832                 psock->done = 1;
1833
1834                 /* Commit done before queuing work to process it */
1835                 smp_mb();
1836
1837                 /* Queue tx work to make sure psock->done is handled */
1838                 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
1839                 spin_unlock_bh(&mux->lock);
1840         } else {
1841 no_reserved:
1842                 if (!psock->tx_stopped)
1843                         list_del(&psock->psock_avail_list);
1844                 list_del(&psock->psock_list);
1845                 mux->psocks_cnt--;
1846                 spin_unlock_bh(&mux->lock);
1847
1848                 sock_put(csk);
1849                 fput(csk->sk_socket->file);
1850                 kmem_cache_free(kcm_psockp, psock);
1851         }
1852 }
1853
1854 static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1855 {
1856         struct kcm_sock *kcm = kcm_sk(sock->sk);
1857         struct kcm_mux *mux = kcm->mux;
1858         struct kcm_psock *psock;
1859         struct socket *csock;
1860         struct sock *csk;
1861         int err;
1862
1863         csock = sockfd_lookup(info->fd, &err);
1864         if (!csock)
1865                 return -ENOENT;
1866
1867         csk = csock->sk;
1868         if (!csk) {
1869                 err = -EINVAL;
1870                 goto out;
1871         }
1872
1873         err = -ENOENT;
1874
1875         spin_lock_bh(&mux->lock);
1876
1877         list_for_each_entry(psock, &mux->psocks, psock_list) {
1878                 if (psock->sk != csk)
1879                         continue;
1880
1881                 /* Found the matching psock */
1882
1883                 if (psock->unattaching || WARN_ON(psock->done)) {
1884                         err = -EALREADY;
1885                         break;
1886                 }
1887
1888                 psock->unattaching = 1;
1889
1890                 spin_unlock_bh(&mux->lock);
1891
1892                 kcm_unattach(psock);
1893
1894                 err = 0;
1895                 goto out;
1896         }
1897
1898         spin_unlock_bh(&mux->lock);
1899
1900 out:
1901         fput(csock->file);
1902         return err;
1903 }
1904
1905 static struct proto kcm_proto = {
1906         .name   = "KCM",
1907         .owner  = THIS_MODULE,
1908         .obj_size = sizeof(struct kcm_sock),
1909 };
1910
1911 /* Clone a kcm socket. */
1912 static int kcm_clone(struct socket *osock, struct kcm_clone *info,
1913                      struct socket **newsockp)
1914 {
1915         struct socket *newsock;
1916         struct sock *newsk;
1917         struct file *newfile;
1918         int err, newfd;
1919
1920         err = -ENFILE;
1921         newsock = sock_alloc();
1922         if (!newsock)
1923                 goto out;
1924
1925         newsock->type = osock->type;
1926         newsock->ops = osock->ops;
1927
1928         __module_get(newsock->ops->owner);
1929
1930         newfd = get_unused_fd_flags(0);
1931         if (unlikely(newfd < 0)) {
1932                 err = newfd;
1933                 goto out_fd_fail;
1934         }
1935
1936         newfile = sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1937         if (unlikely(IS_ERR(newfile))) {
1938                 err = PTR_ERR(newfile);
1939                 goto out_sock_alloc_fail;
1940         }
1941
1942         newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1943                          &kcm_proto, true);
1944         if (!newsk) {
1945                 err = -ENOMEM;
1946                 goto out_sk_alloc_fail;
1947         }
1948
1949         sock_init_data(newsock, newsk);
1950         init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1951
1952         fd_install(newfd, newfile);
1953         *newsockp = newsock;
1954         info->fd = newfd;
1955
1956         return 0;
1957
1958 out_sk_alloc_fail:
1959         fput(newfile);
1960 out_sock_alloc_fail:
1961         put_unused_fd(newfd);
1962 out_fd_fail:
1963         sock_release(newsock);
1964 out:
1965         return err;
1966 }
1967
1968 static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
1969 {
1970         int err;
1971
1972         switch (cmd) {
1973         case SIOCKCMATTACH: {
1974                 struct kcm_attach info;
1975
1976                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1977                         err = -EFAULT;
1978
1979                 err = kcm_attach_ioctl(sock, &info);
1980
1981                 break;
1982         }
1983         case SIOCKCMUNATTACH: {
1984                 struct kcm_unattach info;
1985
1986                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1987                         err = -EFAULT;
1988
1989                 err = kcm_unattach_ioctl(sock, &info);
1990
1991                 break;
1992         }
1993         case SIOCKCMCLONE: {
1994                 struct kcm_clone info;
1995                 struct socket *newsock = NULL;
1996
1997                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1998                         err = -EFAULT;
1999
2000                 err = kcm_clone(sock, &info, &newsock);
2001
2002                 if (!err) {
2003                         if (copy_to_user((void __user *)arg, &info,
2004                                          sizeof(info))) {
2005                                 err = -EFAULT;
2006                                 sock_release(newsock);
2007                         }
2008                 }
2009
2010                 break;
2011         }
2012         default:
2013                 err = -ENOIOCTLCMD;
2014                 break;
2015         }
2016
2017         return err;
2018 }
2019
2020 static void free_mux(struct rcu_head *rcu)
2021 {
2022         struct kcm_mux *mux = container_of(rcu,
2023             struct kcm_mux, rcu);
2024
2025         kmem_cache_free(kcm_muxp, mux);
2026 }
2027
2028 static void release_mux(struct kcm_mux *mux)
2029 {
2030         struct kcm_net *knet = mux->knet;
2031         struct kcm_psock *psock, *tmp_psock;
2032
2033         /* Release psocks */
2034         list_for_each_entry_safe(psock, tmp_psock,
2035                                  &mux->psocks, psock_list) {
2036                 if (!WARN_ON(psock->unattaching))
2037                         kcm_unattach(psock);
2038         }
2039
2040         if (WARN_ON(mux->psocks_cnt))
2041                 return;
2042
2043         __skb_queue_purge(&mux->rx_hold_queue);
2044
2045         mutex_lock(&knet->mutex);
2046         aggregate_mux_stats(&mux->stats, &knet->aggregate_mux_stats);
2047         aggregate_psock_stats(&mux->aggregate_psock_stats,
2048                               &knet->aggregate_psock_stats);
2049         list_del_rcu(&mux->kcm_mux_list);
2050         knet->count--;
2051         mutex_unlock(&knet->mutex);
2052
2053         call_rcu(&mux->rcu, free_mux);
2054 }
2055
2056 static void kcm_done(struct kcm_sock *kcm)
2057 {
2058         struct kcm_mux *mux = kcm->mux;
2059         struct sock *sk = &kcm->sk;
2060         int socks_cnt;
2061
2062         spin_lock_bh(&mux->rx_lock);
2063         if (kcm->rx_psock) {
2064                 /* Cleanup in unreserve_rx_kcm */
2065                 WARN_ON(kcm->done);
2066                 kcm->rx_disabled = 1;
2067                 kcm->done = 1;
2068                 spin_unlock_bh(&mux->rx_lock);
2069                 return;
2070         }
2071
2072         if (kcm->rx_wait) {
2073                 list_del(&kcm->wait_rx_list);
2074                 kcm->rx_wait = false;
2075         }
2076         /* Move any pending receive messages to other kcm sockets */
2077         requeue_rx_msgs(mux, &sk->sk_receive_queue);
2078
2079         spin_unlock_bh(&mux->rx_lock);
2080
2081         if (WARN_ON(sk_rmem_alloc_get(sk)))
2082                 return;
2083
2084         /* Detach from MUX */
2085         spin_lock_bh(&mux->lock);
2086
2087         list_del(&kcm->kcm_sock_list);
2088         mux->kcm_socks_cnt--;
2089         socks_cnt = mux->kcm_socks_cnt;
2090
2091         spin_unlock_bh(&mux->lock);
2092
2093         if (!socks_cnt) {
2094                 /* We are done with the mux now. */
2095                 release_mux(mux);
2096         }
2097
2098         WARN_ON(kcm->rx_wait);
2099
2100         sock_put(&kcm->sk);
2101 }
2102
2103 /* Called by kcm_release to close a KCM socket.
2104  * If this is the last KCM socket on the MUX, destroy the MUX.
2105  */
2106 static int kcm_release(struct socket *sock)
2107 {
2108         struct sock *sk = sock->sk;
2109         struct kcm_sock *kcm;
2110         struct kcm_mux *mux;
2111         struct kcm_psock *psock;
2112
2113         if (!sk)
2114                 return 0;
2115
2116         kcm = kcm_sk(sk);
2117         mux = kcm->mux;
2118
2119         sock_orphan(sk);
2120         kfree_skb(kcm->seq_skb);
2121
2122         lock_sock(sk);
2123         /* Purge queue under lock to avoid race condition with tx_work trying
2124          * to act when queue is nonempty. If tx_work runs after this point
2125          * it will just return.
2126          */
2127         __skb_queue_purge(&sk->sk_write_queue);
2128         release_sock(sk);
2129
2130         spin_lock_bh(&mux->lock);
2131         if (kcm->tx_wait) {
2132                 /* Take of tx_wait list, after this point there should be no way
2133                  * that a psock will be assigned to this kcm.
2134                  */
2135                 list_del(&kcm->wait_psock_list);
2136                 kcm->tx_wait = false;
2137         }
2138         spin_unlock_bh(&mux->lock);
2139
2140         /* Cancel work. After this point there should be no outside references
2141          * to the kcm socket.
2142          */
2143         cancel_work_sync(&kcm->tx_work);
2144
2145         lock_sock(sk);
2146         psock = kcm->tx_psock;
2147         if (psock) {
2148                 /* A psock was reserved, so we need to kill it since it
2149                  * may already have some bytes queued from a message. We
2150                  * need to do this after removing kcm from tx_wait list.
2151                  */
2152                 kcm_abort_tx_psock(psock, EPIPE, false);
2153                 unreserve_psock(kcm);
2154         }
2155         release_sock(sk);
2156
2157         WARN_ON(kcm->tx_wait);
2158         WARN_ON(kcm->tx_psock);
2159
2160         sock->sk = NULL;
2161
2162         kcm_done(kcm);
2163
2164         return 0;
2165 }
2166
2167 static const struct proto_ops kcm_dgram_ops = {
2168         .family =       PF_KCM,
2169         .owner =        THIS_MODULE,
2170         .release =      kcm_release,
2171         .bind =         sock_no_bind,
2172         .connect =      sock_no_connect,
2173         .socketpair =   sock_no_socketpair,
2174         .accept =       sock_no_accept,
2175         .getname =      sock_no_getname,
2176         .poll =         datagram_poll,
2177         .ioctl =        kcm_ioctl,
2178         .listen =       sock_no_listen,
2179         .shutdown =     sock_no_shutdown,
2180         .setsockopt =   kcm_setsockopt,
2181         .getsockopt =   kcm_getsockopt,
2182         .sendmsg =      kcm_sendmsg,
2183         .recvmsg =      kcm_recvmsg,
2184         .mmap =         sock_no_mmap,
2185         .sendpage =     kcm_sendpage,
2186 };
2187
2188 static const struct proto_ops kcm_seqpacket_ops = {
2189         .family =       PF_KCM,
2190         .owner =        THIS_MODULE,
2191         .release =      kcm_release,
2192         .bind =         sock_no_bind,
2193         .connect =      sock_no_connect,
2194         .socketpair =   sock_no_socketpair,
2195         .accept =       sock_no_accept,
2196         .getname =      sock_no_getname,
2197         .poll =         datagram_poll,
2198         .ioctl =        kcm_ioctl,
2199         .listen =       sock_no_listen,
2200         .shutdown =     sock_no_shutdown,
2201         .setsockopt =   kcm_setsockopt,
2202         .getsockopt =   kcm_getsockopt,
2203         .sendmsg =      kcm_sendmsg,
2204         .recvmsg =      kcm_recvmsg,
2205         .mmap =         sock_no_mmap,
2206         .sendpage =     kcm_sendpage,
2207         .splice_read =  kcm_splice_read,
2208 };
2209
2210 /* Create proto operation for kcm sockets */
2211 static int kcm_create(struct net *net, struct socket *sock,
2212                       int protocol, int kern)
2213 {
2214         struct kcm_net *knet = net_generic(net, kcm_net_id);
2215         struct sock *sk;
2216         struct kcm_mux *mux;
2217
2218         switch (sock->type) {
2219         case SOCK_DGRAM:
2220                 sock->ops = &kcm_dgram_ops;
2221                 break;
2222         case SOCK_SEQPACKET:
2223                 sock->ops = &kcm_seqpacket_ops;
2224                 break;
2225         default:
2226                 return -ESOCKTNOSUPPORT;
2227         }
2228
2229         if (protocol != KCMPROTO_CONNECTED)
2230                 return -EPROTONOSUPPORT;
2231
2232         sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
2233         if (!sk)
2234                 return -ENOMEM;
2235
2236         /* Allocate a kcm mux, shared between KCM sockets */
2237         mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
2238         if (!mux) {
2239                 sk_free(sk);
2240                 return -ENOMEM;
2241         }
2242
2243         spin_lock_init(&mux->lock);
2244         spin_lock_init(&mux->rx_lock);
2245         INIT_LIST_HEAD(&mux->kcm_socks);
2246         INIT_LIST_HEAD(&mux->kcm_rx_waiters);
2247         INIT_LIST_HEAD(&mux->kcm_tx_waiters);
2248
2249         INIT_LIST_HEAD(&mux->psocks);
2250         INIT_LIST_HEAD(&mux->psocks_ready);
2251         INIT_LIST_HEAD(&mux->psocks_avail);
2252
2253         mux->knet = knet;
2254
2255         /* Add new MUX to list */
2256         mutex_lock(&knet->mutex);
2257         list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
2258         knet->count++;
2259         mutex_unlock(&knet->mutex);
2260
2261         skb_queue_head_init(&mux->rx_hold_queue);
2262
2263         /* Init KCM socket */
2264         sock_init_data(sock, sk);
2265         init_kcm_sock(kcm_sk(sk), mux);
2266
2267         return 0;
2268 }
2269
2270 static struct net_proto_family kcm_family_ops = {
2271         .family = PF_KCM,
2272         .create = kcm_create,
2273         .owner  = THIS_MODULE,
2274 };
2275
2276 static __net_init int kcm_init_net(struct net *net)
2277 {
2278         struct kcm_net *knet = net_generic(net, kcm_net_id);
2279
2280         INIT_LIST_HEAD_RCU(&knet->mux_list);
2281         mutex_init(&knet->mutex);
2282
2283         return 0;
2284 }
2285
2286 static __net_exit void kcm_exit_net(struct net *net)
2287 {
2288         struct kcm_net *knet = net_generic(net, kcm_net_id);
2289
2290         /* All KCM sockets should be closed at this point, which should mean
2291          * that all multiplexors and psocks have been destroyed.
2292          */
2293         WARN_ON(!list_empty(&knet->mux_list));
2294 }
2295
2296 static struct pernet_operations kcm_net_ops = {
2297         .init = kcm_init_net,
2298         .exit = kcm_exit_net,
2299         .id   = &kcm_net_id,
2300         .size = sizeof(struct kcm_net),
2301 };
2302
2303 static int __init kcm_init(void)
2304 {
2305         int err = -ENOMEM;
2306
2307         kcm_muxp = kmem_cache_create("kcm_mux_cache",
2308                                      sizeof(struct kcm_mux), 0,
2309                                      SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2310         if (!kcm_muxp)
2311                 goto fail;
2312
2313         kcm_psockp = kmem_cache_create("kcm_psock_cache",
2314                                        sizeof(struct kcm_psock), 0,
2315                                         SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2316         if (!kcm_psockp)
2317                 goto fail;
2318
2319         kcm_wq = create_singlethread_workqueue("kkcmd");
2320         if (!kcm_wq)
2321                 goto fail;
2322
2323         err = proto_register(&kcm_proto, 1);
2324         if (err)
2325                 goto fail;
2326
2327         err = sock_register(&kcm_family_ops);
2328         if (err)
2329                 goto sock_register_fail;
2330
2331         err = register_pernet_device(&kcm_net_ops);
2332         if (err)
2333                 goto net_ops_fail;
2334
2335         err = kcm_proc_init();
2336         if (err)
2337                 goto proc_init_fail;
2338
2339         return 0;
2340
2341 proc_init_fail:
2342         unregister_pernet_device(&kcm_net_ops);
2343
2344 net_ops_fail:
2345         sock_unregister(PF_KCM);
2346
2347 sock_register_fail:
2348         proto_unregister(&kcm_proto);
2349
2350 fail:
2351         kmem_cache_destroy(kcm_muxp);
2352         kmem_cache_destroy(kcm_psockp);
2353
2354         if (kcm_wq)
2355                 destroy_workqueue(kcm_wq);
2356
2357         return err;
2358 }
2359
2360 static void __exit kcm_exit(void)
2361 {
2362         kcm_proc_exit();
2363         unregister_pernet_device(&kcm_net_ops);
2364         sock_unregister(PF_KCM);
2365         proto_unregister(&kcm_proto);
2366         destroy_workqueue(kcm_wq);
2367
2368         kmem_cache_destroy(kcm_muxp);
2369         kmem_cache_destroy(kcm_psockp);
2370 }
2371
2372 module_init(kcm_init);
2373 module_exit(kcm_exit);
2374
2375 MODULE_LICENSE("GPL");
2376 MODULE_ALIAS_NETPROTO(PF_KCM);
2377