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         del_timer(&psock->rx_msg_timer);
  59
  60         if (psock->rx_stopped)
  61                 return;
  62
  63         psock->rx_stopped = 1;
  64         KCM_STATS_INCR(psock->stats.rx_aborts);
  65
  66         /* Report an error on the lower socket */
  67         report_csk_error(csk, err);
  68 }
  69
  70 static void kcm_abort_tx_psock(struct kcm_psock *psock, int err,
  71                                bool wakeup_kcm)
  72 {
  73         struct sock *csk = psock->sk;
  74         struct kcm_mux *mux = psock->mux;
  75
  76         /* Unrecoverable error in transmit */
  77
  78         spin_lock_bh(&mux->lock);
  79
  80         if (psock->tx_stopped) {
  81                 spin_unlock_bh(&mux->lock);
  82                 return;
  83         }
  84
  85         psock->tx_stopped = 1;
  86         KCM_STATS_INCR(psock->stats.tx_aborts);
  87
  88         if (!psock->tx_kcm) {
  89                 /* Take off psocks_avail list */
  90                 list_del(&psock->psock_avail_list);
  91         } else if (wakeup_kcm) {
  92                 /* In this case psock is being aborted while outside of
  93                  * write_msgs and psock is reserved. Schedule tx_work
  94                  * to handle the failure there. Need to commit tx_stopped
  95                  * before queuing work.
  96                  */
  97                 smp_mb();
  98
  99                 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
 100         }
 101
 102         spin_unlock_bh(&mux->lock);
 103
 104         /* Report error on lower socket */
 105         report_csk_error(csk, err);
 106 }
 107
 108 /* RX mux lock held. */
 109 static void kcm_update_rx_mux_stats(struct kcm_mux *mux,
 110                                     struct kcm_psock *psock)
 111 {
 112         KCM_STATS_ADD(mux->stats.rx_bytes,
 113                       psock->stats.rx_bytes - psock->saved_rx_bytes);
 114         mux->stats.rx_msgs +=
 115                 psock->stats.rx_msgs - psock->saved_rx_msgs;
 116         psock->saved_rx_msgs = psock->stats.rx_msgs;
 117         psock->saved_rx_bytes = psock->stats.rx_bytes;
 118 }
 119
 120 static void kcm_update_tx_mux_stats(struct kcm_mux *mux,
 121                                     struct kcm_psock *psock)
 122 {
 123         KCM_STATS_ADD(mux->stats.tx_bytes,
 124                       psock->stats.tx_bytes - psock->saved_tx_bytes);
 125         mux->stats.tx_msgs +=
 126                 psock->stats.tx_msgs - psock->saved_tx_msgs;
 127         psock->saved_tx_msgs = psock->stats.tx_msgs;
 128         psock->saved_tx_bytes = psock->stats.tx_bytes;
 129 }
 130
 131 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb);
 132
 133 /* KCM is ready to receive messages on its queue-- either the KCM is new or
 134  * has become unblocked after being blocked on full socket buffer. Queue any
 135  * pending ready messages on a psock. RX mux lock held.
 136  */
 137 static void kcm_rcv_ready(struct kcm_sock *kcm)
 138 {
 139         struct kcm_mux *mux = kcm->mux;
 140         struct kcm_psock *psock;
 141         struct sk_buff *skb;
 142
 143         if (unlikely(kcm->rx_wait || kcm->rx_psock || kcm->rx_disabled))
 144                 return;
 145
 146         while (unlikely((skb = __skb_dequeue(&mux->rx_hold_queue)))) {
 147                 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 148                         /* Assuming buffer limit has been reached */
 149                         skb_queue_head(&mux->rx_hold_queue, skb);
 150                         WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
 151                         return;
 152                 }
 153         }
 154
 155         while (!list_empty(&mux->psocks_ready)) {
 156                 psock = list_first_entry(&mux->psocks_ready, struct kcm_psock,
 157                                          psock_ready_list);
 158
 159                 if (kcm_queue_rcv_skb(&kcm->sk, psock->ready_rx_msg)) {
 160                         /* Assuming buffer limit has been reached */
 161                         WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
 162                         return;
 163                 }
 164
 165                 /* Consumed the ready message on the psock. Schedule rx_work to
 166                  * get more messages.
 167                  */
 168                 list_del(&psock->psock_ready_list);
 169                 psock->ready_rx_msg = NULL;
 170
 171                 /* Commit clearing of ready_rx_msg for queuing work */
 172                 smp_mb();
 173
 174                 queue_work(kcm_wq, &psock->rx_work);
 175         }
 176
 177         /* Buffer limit is okay now, add to ready list */
 178         list_add_tail(&kcm->wait_rx_list,
 179                       &kcm->mux->kcm_rx_waiters);
 180         kcm->rx_wait = true;
 181 }
 182
 183 static void kcm_rfree(struct sk_buff *skb)
 184 {
 185         struct sock *sk = skb->sk;
 186         struct kcm_sock *kcm = kcm_sk(sk);
 187         struct kcm_mux *mux = kcm->mux;
 188         unsigned int len = skb->truesize;
 189
 190         sk_mem_uncharge(sk, len);
 191         atomic_sub(len, &sk->sk_rmem_alloc);
 192
 193         /* For reading rx_wait and rx_psock without holding lock */
 194         smp_mb__after_atomic();
 195
 196         if (!kcm->rx_wait && !kcm->rx_psock &&
 197             sk_rmem_alloc_get(sk) < sk->sk_rcvlowat) {
 198                 spin_lock_bh(&mux->rx_lock);
 199                 kcm_rcv_ready(kcm);
 200                 spin_unlock_bh(&mux->rx_lock);
 201         }
 202 }
 203
 204 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb)
 205 {
 206         struct sk_buff_head *list = &sk->sk_receive_queue;
 207
 208         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
 209                 return -ENOMEM;
 210
 211         if (!sk_rmem_schedule(sk, skb, skb->truesize))
 212                 return -ENOBUFS;
 213
 214         skb->dev = NULL;
 215
 216         skb_orphan(skb);
 217         skb->sk = sk;
 218         skb->destructor = kcm_rfree;
 219         atomic_add(skb->truesize, &sk->sk_rmem_alloc);
 220         sk_mem_charge(sk, skb->truesize);
 221
 222         skb_queue_tail(list, skb);
 223
 224         if (!sock_flag(sk, SOCK_DEAD))
 225                 sk->sk_data_ready(sk);
 226
 227         return 0;
 228 }
 229
 230 /* Requeue received messages for a kcm socket to other kcm sockets. This is
 231  * called with a kcm socket is receive disabled.
 232  * RX mux lock held.
 233  */
 234 static void requeue_rx_msgs(struct kcm_mux *mux, struct sk_buff_head *head)
 235 {
 236         struct sk_buff *skb;
 237         struct kcm_sock *kcm;
 238
 239         while ((skb = __skb_dequeue(head))) {
 240                 /* Reset destructor to avoid calling kcm_rcv_ready */
 241                 skb->destructor = sock_rfree;
 242                 skb_orphan(skb);
 243 try_again:
 244                 if (list_empty(&mux->kcm_rx_waiters)) {
 245                         skb_queue_tail(&mux->rx_hold_queue, skb);
 246                         continue;
 247                 }
 248
 249                 kcm = list_first_entry(&mux->kcm_rx_waiters,
 250                                        struct kcm_sock, wait_rx_list);
 251
 252                 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
 253                         /* Should mean socket buffer full */
 254                         list_del(&kcm->wait_rx_list);
 255                         kcm->rx_wait = false;
 256
 257                         /* Commit rx_wait to read in kcm_free */
 258                         smp_wmb();
 259
 260                         goto try_again;
 261                 }
 262         }
 263 }
 264
 265 /* Lower sock lock held */
 266 static struct kcm_sock *reserve_rx_kcm(struct kcm_psock *psock,
 267                                        struct sk_buff *head)
 268 {
 269         struct kcm_mux *mux = psock->mux;
 270         struct kcm_sock *kcm;
 271
 272         WARN_ON(psock->ready_rx_msg);
 273
 274         if (psock->rx_kcm)
 275                 return psock->rx_kcm;
 276
 277         spin_lock_bh(&mux->rx_lock);
 278
 279         if (psock->rx_kcm) {
 280                 spin_unlock_bh(&mux->rx_lock);
 281                 return psock->rx_kcm;
 282         }
 283
 284         kcm_update_rx_mux_stats(mux, psock);
 285
 286         if (list_empty(&mux->kcm_rx_waiters)) {
 287                 psock->ready_rx_msg = head;
 288                 list_add_tail(&psock->psock_ready_list,
 289                               &mux->psocks_ready);
 290                 spin_unlock_bh(&mux->rx_lock);
 291                 return NULL;
 292         }
 293
 294         kcm = list_first_entry(&mux->kcm_rx_waiters,
 295                                struct kcm_sock, wait_rx_list);
 296         list_del(&kcm->wait_rx_list);
 297         kcm->rx_wait = false;
 298
 299         psock->rx_kcm = kcm;
 300         kcm->rx_psock = psock;
 301
 302         spin_unlock_bh(&mux->rx_lock);
 303
 304         return kcm;
 305 }
 306
 307 static void kcm_done(struct kcm_sock *kcm);
 308
 309 static void kcm_done_work(struct work_struct *w)
 310 {
 311         kcm_done(container_of(w, struct kcm_sock, done_work));
 312 }
 313
 314 /* Lower sock held */
 315 static void unreserve_rx_kcm(struct kcm_psock *psock,
 316                              bool rcv_ready)
 317 {
 318         struct kcm_sock *kcm = psock->rx_kcm;
 319         struct kcm_mux *mux = psock->mux;
 320
 321         if (!kcm)
 322                 return;
 323
 324         spin_lock_bh(&mux->rx_lock);
 325
 326         psock->rx_kcm = NULL;
 327         kcm->rx_psock = NULL;
 328
 329         /* Commit kcm->rx_psock before sk_rmem_alloc_get to sync with
 330          * kcm_rfree
 331          */
 332         smp_mb();
 333
 334         if (unlikely(kcm->done)) {
 335                 spin_unlock_bh(&mux->rx_lock);
 336
 337                 /* Need to run kcm_done in a task since we need to qcquire
 338                  * callback locks which may already be held here.
 339                  */
 340                 INIT_WORK(&kcm->done_work, kcm_done_work);
 341                 schedule_work(&kcm->done_work);
 342                 return;
 343         }
 344
 345         if (unlikely(kcm->rx_disabled)) {
 346                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
 347         } else if (rcv_ready || unlikely(!sk_rmem_alloc_get(&kcm->sk))) {
 348                 /* Check for degenerative race with rx_wait that all
 349                  * data was dequeued (accounted for in kcm_rfree).
 350                  */
 351                 kcm_rcv_ready(kcm);
 352         }
 353         spin_unlock_bh(&mux->rx_lock);
 354 }
 355
 356 static void kcm_start_rx_timer(struct kcm_psock *psock)
 357 {
 358         if (psock->sk->sk_rcvtimeo)
 359                 mod_timer(&psock->rx_msg_timer, psock->sk->sk_rcvtimeo);
 360 }
 361
 362 /* Macro to invoke filter function. */
 363 #define KCM_RUN_FILTER(prog, ctx) \
 364         (*prog->bpf_func)(ctx, prog->insnsi)
 365
 366 /* Lower socket lock held */
 367 static int kcm_tcp_recv(read_descriptor_t *desc, struct sk_buff *orig_skb,
 368                         unsigned int orig_offset, size_t orig_len)
 369 {
 370         struct kcm_psock *psock = (struct kcm_psock *)desc->arg.data;
 371         struct kcm_rx_msg *rxm;
 372         struct kcm_sock *kcm;
 373         struct sk_buff *head, *skb;
 374         size_t eaten = 0, cand_len;
 375         ssize_t extra;
 376         int err;
 377         bool cloned_orig = false;
 378
 379         if (psock->ready_rx_msg)
 380                 return 0;
 381
 382         head = psock->rx_skb_head;
 383         if (head) {
 384                 /* Message already in progress */
 385
 386                 rxm = kcm_rx_msg(head);
 387                 if (unlikely(rxm->early_eaten)) {
 388                         /* Already some number of bytes on the receive sock
 389                          * data saved in rx_skb_head, just indicate they
 390                          * are consumed.
 391                          */
 392                         eaten = orig_len <= rxm->early_eaten ?
 393                                 orig_len : rxm->early_eaten;
 394                         rxm->early_eaten -= eaten;
 395
 396                         return eaten;
 397                 }
 398
 399                 if (unlikely(orig_offset)) {
 400                         /* Getting data with a non-zero offset when a message is
 401                          * in progress is not expected. If it does happen, we
 402                          * need to clone and pull since we can't deal with
 403                          * offsets in the skbs for a message expect in the head.
 404                          */
 405                         orig_skb = skb_clone(orig_skb, GFP_ATOMIC);
 406                         if (!orig_skb) {
 407                                 KCM_STATS_INCR(psock->stats.rx_mem_fail);
 408                                 desc->error = -ENOMEM;
 409                                 return 0;
 410                         }
 411                         if (!pskb_pull(orig_skb, orig_offset)) {
 412                                 KCM_STATS_INCR(psock->stats.rx_mem_fail);
 413                                 kfree_skb(orig_skb);
 414                                 desc->error = -ENOMEM;
 415                                 return 0;
 416                         }
 417                         cloned_orig = true;
 418                         orig_offset = 0;
 419                 }
 420
 421                 if (!psock->rx_skb_nextp) {
 422                         /* We are going to append to the frags_list of head.
 423                          * Need to unshare the frag_list.
 424                          */
 425                         err = skb_unclone(head, GFP_ATOMIC);
 426                         if (err) {
 427                                 KCM_STATS_INCR(psock->stats.rx_mem_fail);
 428                                 desc->error = err;
 429                                 return 0;
 430                         }
 431
 432                         if (unlikely(skb_shinfo(head)->frag_list)) {
 433                                 /* We can't append to an sk_buff that already
 434                                  * has a frag_list. We create a new head, point
 435                                  * the frag_list of that to the old head, and
 436                                  * then are able to use the old head->next for
 437                                  * appending to the message.
 438                                  */
 439                                 if (WARN_ON(head->next)) {
 440                                         desc->error = -EINVAL;
 441                                         return 0;
 442                                 }
 443
 444                                 skb = alloc_skb(0, GFP_ATOMIC);
 445                                 if (!skb) {
 446                                         KCM_STATS_INCR(psock->stats.rx_mem_fail);
 447                                         desc->error = -ENOMEM;
 448                                         return 0;
 449                                 }
 450                                 skb->len = head->len;
 451                                 skb->data_len = head->len;
 452                                 skb->truesize = head->truesize;
 453                                 *kcm_rx_msg(skb) = *kcm_rx_msg(head);
 454                                 psock->rx_skb_nextp = &head->next;
 455                                 skb_shinfo(skb)->frag_list = head;
 456                                 psock->rx_skb_head = skb;
 457                                 head = skb;
 458                         } else {
 459                                 psock->rx_skb_nextp =
 460                                     &skb_shinfo(head)->frag_list;
 461                         }
 462                 }
 463         }
 464
 465         while (eaten < orig_len) {
 466                 /* Always clone since we will consume something */
 467                 skb = skb_clone(orig_skb, GFP_ATOMIC);
 468                 if (!skb) {
 469                         KCM_STATS_INCR(psock->stats.rx_mem_fail);
 470                         desc->error = -ENOMEM;
 471                         break;
 472                 }
 473
 474                 cand_len = orig_len - eaten;
 475
 476                 head = psock->rx_skb_head;
 477                 if (!head) {
 478                         head = skb;
 479                         psock->rx_skb_head = head;
 480                         /* Will set rx_skb_nextp on next packet if needed */
 481                         psock->rx_skb_nextp = NULL;
 482                         rxm = kcm_rx_msg(head);
 483                         memset(rxm, 0, sizeof(*rxm));
 484                         rxm->offset = orig_offset + eaten;
 485                 } else {
 486                         /* Unclone since we may be appending to an skb that we
 487                          * already share a frag_list with.
 488                          */
 489                         err = skb_unclone(skb, GFP_ATOMIC);
 490                         if (err) {
 491                                 KCM_STATS_INCR(psock->stats.rx_mem_fail);
 492                                 desc->error = err;
 493                                 break;
 494                         }
 495
 496                         rxm = kcm_rx_msg(head);
 497                         *psock->rx_skb_nextp = skb;
 498                         psock->rx_skb_nextp = &skb->next;
 499                         head->data_len += skb->len;
 500                         head->len += skb->len;
 501                         head->truesize += skb->truesize;
 502                 }
 503
 504                 if (!rxm->full_len) {
 505                         ssize_t len;
 506
 507                         len = KCM_RUN_FILTER(psock->bpf_prog, head);
 508
 509                         if (!len) {
 510                                 /* Need more header to determine length */
 511                                 if (!rxm->accum_len) {
 512                                         /* Start RX timer for new message */
 513                                         kcm_start_rx_timer(psock);
 514                                 }
 515                                 rxm->accum_len += cand_len;
 516                                 eaten += cand_len;
 517                                 KCM_STATS_INCR(psock->stats.rx_need_more_hdr);
 518                                 WARN_ON(eaten != orig_len);
 519                                 break;
 520                         } else if (len > psock->sk->sk_rcvbuf) {
 521                                 /* Message length exceeds maximum allowed */
 522                                 KCM_STATS_INCR(psock->stats.rx_msg_too_big);
 523                                 desc->error = -EMSGSIZE;
 524                                 psock->rx_skb_head = NULL;
 525                                 kcm_abort_rx_psock(psock, EMSGSIZE, head);
 526                                 break;
 527                         } else if (len <= (ssize_t)head->len -
 528                                           skb->len - rxm->offset) {
 529                                 /* Length must be into new skb (and also
 530                                  * greater than zero)
 531                                  */
 532                                 KCM_STATS_INCR(psock->stats.rx_bad_hdr_len);
 533                                 desc->error = -EPROTO;
 534                                 psock->rx_skb_head = NULL;
 535                                 kcm_abort_rx_psock(psock, EPROTO, head);
 536                                 break;
 537                         }
 538
 539                         rxm->full_len = len;
 540                 }
 541
 542                 extra = (ssize_t)(rxm->accum_len + cand_len) - rxm->full_len;
 543
 544                 if (extra < 0) {
 545                         /* Message not complete yet. */
 546                         if (rxm->full_len - rxm->accum_len >
 547                             tcp_inq(psock->sk)) {
 548                                 /* Don't have the whole messages in the socket
 549                                  * buffer. Set psock->rx_need_bytes to wait for
 550                                  * the rest of the message. Also, set "early
 551                                  * eaten" since we've already buffered the skb
 552                                  * but don't consume yet per tcp_read_sock.
 553                                  */
 554
 555                                 if (!rxm->accum_len) {
 556                                         /* Start RX timer for new message */
 557                                         kcm_start_rx_timer(psock);
 558                                 }
 559
 560                                 psock->rx_need_bytes = rxm->full_len -
 561                                                        rxm->accum_len;
 562                                 rxm->accum_len += cand_len;
 563                                 rxm->early_eaten = cand_len;
 564                                 KCM_STATS_ADD(psock->stats.rx_bytes, cand_len);
 565                                 desc->count = 0; /* Stop reading socket */
 566                                 break;
 567                         }
 568                         rxm->accum_len += cand_len;
 569                         eaten += cand_len;
 570                         WARN_ON(eaten != orig_len);
 571                         break;
 572                 }
 573
 574                 /* Positive extra indicates ore bytes than needed for the
 575                  * message
 576                  */
 577
 578                 WARN_ON(extra > cand_len);
 579
 580                 eaten += (cand_len - extra);
 581
 582                 /* Hurray, we have a new message! */
 583                 del_timer(&psock->rx_msg_timer);
 584                 psock->rx_skb_head = NULL;
 585                 KCM_STATS_INCR(psock->stats.rx_msgs);
 586
 587 try_queue:
 588                 kcm = reserve_rx_kcm(psock, head);
 589                 if (!kcm) {
 590                         /* Unable to reserve a KCM, message is held in psock. */
 591                         break;
 592                 }
 593
 594                 if (kcm_queue_rcv_skb(&kcm->sk, head)) {
 595                         /* Should mean socket buffer full */
 596                         unreserve_rx_kcm(psock, false);
 597                         goto try_queue;
 598                 }
 599         }
 600
 601         if (cloned_orig)
 602                 kfree_skb(orig_skb);
 603
 604         KCM_STATS_ADD(psock->stats.rx_bytes, eaten);
 605
 606         return eaten;
 607 }
 608
 609 /* Called with lock held on lower socket */
 610 static int psock_tcp_read_sock(struct kcm_psock *psock)
 611 {
 612         read_descriptor_t desc;
 613
 614         desc.arg.data = psock;
 615         desc.error = 0;
 616         desc.count = 1; /* give more than one skb per call */
 617
 618         /* sk should be locked here, so okay to do tcp_read_sock */
 619         tcp_read_sock(psock->sk, &desc, kcm_tcp_recv);
 620
 621         unreserve_rx_kcm(psock, true);
 622
 623         return desc.error;
 624 }
 625
 626 /* Lower sock lock held */
 627 static void psock_tcp_data_ready(struct sock *sk)
 628 {
 629         struct kcm_psock *psock;
 630
 631         read_lock_bh(&sk->sk_callback_lock);
 632
 633         psock = (struct kcm_psock *)sk->sk_user_data;
 634         if (unlikely(!psock || psock->rx_stopped))
 635                 goto out;
 636
 637         if (psock->ready_rx_msg)
 638                 goto out;
 639
 640         if (psock->rx_need_bytes) {
 641                 if (tcp_inq(sk) >= psock->rx_need_bytes)
 642                         psock->rx_need_bytes = 0;
 643                 else
 644                         goto out;
 645         }
 646
 647         if (psock_tcp_read_sock(psock) == -ENOMEM)
 648                 queue_delayed_work(kcm_wq, &psock->rx_delayed_work, 0);
 649
 650 out:
 651         read_unlock_bh(&sk->sk_callback_lock);
 652 }
 653
 654 static void do_psock_rx_work(struct kcm_psock *psock)
 655 {
 656         read_descriptor_t rd_desc;
 657         struct sock *csk = psock->sk;
 658
 659         /* We need the read lock to synchronize with psock_tcp_data_ready. We
 660          * need the socket lock for calling tcp_read_sock.
 661          */
 662         lock_sock(csk);
 663         read_lock_bh(&csk->sk_callback_lock);
 664
 665         if (unlikely(csk->sk_user_data != psock))
 666                 goto out;
 667
 668         if (unlikely(psock->rx_stopped))
 669                 goto out;
 670
 671         if (psock->ready_rx_msg)
 672                 goto out;
 673
 674         rd_desc.arg.data = psock;
 675
 676         if (psock_tcp_read_sock(psock) == -ENOMEM)
 677                 queue_delayed_work(kcm_wq, &psock->rx_delayed_work, 0);
 678
 679 out:
 680         read_unlock_bh(&csk->sk_callback_lock);
 681         release_sock(csk);
 682 }
 683
 684 static void psock_rx_work(struct work_struct *w)
 685 {
 686         do_psock_rx_work(container_of(w, struct kcm_psock, rx_work));
 687 }
 688
 689 static void psock_rx_delayed_work(struct work_struct *w)
 690 {
 691         do_psock_rx_work(container_of(w, struct kcm_psock,
 692                                       rx_delayed_work.work));
 693 }
 694
 695 static void psock_tcp_state_change(struct sock *sk)
 696 {
 697         /* TCP only does a POLLIN for a half close. Do a POLLHUP here
 698          * since application will normally not poll with POLLIN
 699          * on the TCP sockets.
 700          */
 701
 702         report_csk_error(sk, EPIPE);
 703 }
 704
 705 static void psock_tcp_write_space(struct sock *sk)
 706 {
 707         struct kcm_psock *psock;
 708         struct kcm_mux *mux;
 709         struct kcm_sock *kcm;
 710
 711         read_lock_bh(&sk->sk_callback_lock);
 712
 713         psock = (struct kcm_psock *)sk->sk_user_data;
 714         if (unlikely(!psock))
 715                 goto out;
 716
 717         mux = psock->mux;
 718
 719         spin_lock_bh(&mux->lock);
 720
 721         /* Check if the socket is reserved so someone is waiting for sending. */
 722         kcm = psock->tx_kcm;
 723         if (kcm)
 724                 queue_work(kcm_wq, &kcm->tx_work);
 725
 726         spin_unlock_bh(&mux->lock);
 727 out:
 728         read_unlock_bh(&sk->sk_callback_lock);
 729 }
 730
 731 static void unreserve_psock(struct kcm_sock *kcm);
 732
 733 /* kcm sock is locked. */
 734 static struct kcm_psock *reserve_psock(struct kcm_sock *kcm)
 735 {
 736         struct kcm_mux *mux = kcm->mux;
 737         struct kcm_psock *psock;
 738
 739         psock = kcm->tx_psock;
 740
 741         smp_rmb(); /* Must read tx_psock before tx_wait */
 742
 743         if (psock) {
 744                 WARN_ON(kcm->tx_wait);
 745                 if (unlikely(psock->tx_stopped))
 746                         unreserve_psock(kcm);
 747                 else
 748                         return kcm->tx_psock;
 749         }
 750
 751         spin_lock_bh(&mux->lock);
 752
 753         /* Check again under lock to see if psock was reserved for this
 754          * psock via psock_unreserve.
 755          */
 756         psock = kcm->tx_psock;
 757         if (unlikely(psock)) {
 758                 WARN_ON(kcm->tx_wait);
 759                 spin_unlock_bh(&mux->lock);
 760                 return kcm->tx_psock;
 761         }
 762
 763         if (!list_empty(&mux->psocks_avail)) {
 764                 psock = list_first_entry(&mux->psocks_avail,
 765                                          struct kcm_psock,
 766                                          psock_avail_list);
 767                 list_del(&psock->psock_avail_list);
 768                 if (kcm->tx_wait) {
 769                         list_del(&kcm->wait_psock_list);
 770                         kcm->tx_wait = false;
 771                 }
 772                 kcm->tx_psock = psock;
 773                 psock->tx_kcm = kcm;
 774                 KCM_STATS_INCR(psock->stats.reserved);
 775         } else if (!kcm->tx_wait) {
 776                 list_add_tail(&kcm->wait_psock_list,
 777                               &mux->kcm_tx_waiters);
 778                 kcm->tx_wait = true;
 779         }
 780
 781         spin_unlock_bh(&mux->lock);
 782
 783         return psock;
 784 }
 785
 786 /* mux lock held */
 787 static void psock_now_avail(struct kcm_psock *psock)
 788 {
 789         struct kcm_mux *mux = psock->mux;
 790         struct kcm_sock *kcm;
 791
 792         if (list_empty(&mux->kcm_tx_waiters)) {
 793                 list_add_tail(&psock->psock_avail_list,
 794                               &mux->psocks_avail);
 795         } else {
 796                 kcm = list_first_entry(&mux->kcm_tx_waiters,
 797                                        struct kcm_sock,
 798                                        wait_psock_list);
 799                 list_del(&kcm->wait_psock_list);
 800                 kcm->tx_wait = false;
 801                 psock->tx_kcm = kcm;
 802
 803                 /* Commit before changing tx_psock since that is read in
 804                  * reserve_psock before queuing work.
 805                  */
 806                 smp_mb();
 807
 808                 kcm->tx_psock = psock;
 809                 KCM_STATS_INCR(psock->stats.reserved);
 810                 queue_work(kcm_wq, &kcm->tx_work);
 811         }
 812 }
 813
 814 /* kcm sock is locked. */
 815 static void unreserve_psock(struct kcm_sock *kcm)
 816 {
 817         struct kcm_psock *psock;
 818         struct kcm_mux *mux = kcm->mux;
 819
 820         spin_lock_bh(&mux->lock);
 821
 822         psock = kcm->tx_psock;
 823
 824         if (WARN_ON(!psock)) {
 825                 spin_unlock_bh(&mux->lock);
 826                 return;
 827         }
 828
 829         smp_rmb(); /* Read tx_psock before tx_wait */
 830
 831         kcm_update_tx_mux_stats(mux, psock);
 832
 833         WARN_ON(kcm->tx_wait);
 834
 835         kcm->tx_psock = NULL;
 836         psock->tx_kcm = NULL;
 837         KCM_STATS_INCR(psock->stats.unreserved);
 838
 839         if (unlikely(psock->tx_stopped)) {
 840                 if (psock->done) {
 841                         /* Deferred free */
 842                         list_del(&psock->psock_list);
 843                         mux->psocks_cnt--;
 844                         sock_put(psock->sk);
 845                         fput(psock->sk->sk_socket->file);
 846                         kmem_cache_free(kcm_psockp, psock);
 847                 }
 848
 849                 /* Don't put back on available list */
 850
 851                 spin_unlock_bh(&mux->lock);
 852
 853                 return;
 854         }
 855
 856         psock_now_avail(psock);
 857
 858         spin_unlock_bh(&mux->lock);
 859 }
 860
 861 static void kcm_report_tx_retry(struct kcm_sock *kcm)
 862 {
 863         struct kcm_mux *mux = kcm->mux;
 864
 865         spin_lock_bh(&mux->lock);
 866         KCM_STATS_INCR(mux->stats.tx_retries);
 867         spin_unlock_bh(&mux->lock);
 868 }
 869
 870 /* Write any messages ready on the kcm socket.  Called with kcm sock lock
 871  * held.  Return bytes actually sent or error.
 872  */
 873 static int kcm_write_msgs(struct kcm_sock *kcm)
 874 {
 875         struct sock *sk = &kcm->sk;
 876         struct kcm_psock *psock;
 877         struct sk_buff *skb, *head;
 878         struct kcm_tx_msg *txm;
 879         unsigned short fragidx, frag_offset;
 880         unsigned int sent, total_sent = 0;
 881         int ret = 0;
 882
 883         kcm->tx_wait_more = false;
 884         psock = kcm->tx_psock;
 885         if (unlikely(psock && psock->tx_stopped)) {
 886                 /* A reserved psock was aborted asynchronously. Unreserve
 887                  * it and we'll retry the message.
 888                  */
 889                 unreserve_psock(kcm);
 890                 kcm_report_tx_retry(kcm);
 891                 if (skb_queue_empty(&sk->sk_write_queue))
 892                         return 0;
 893
 894                 kcm_tx_msg(skb_peek(&sk->sk_write_queue))->sent = 0;
 895
 896         } else if (skb_queue_empty(&sk->sk_write_queue)) {
 897                 return 0;
 898         }
 899
 900         head = skb_peek(&sk->sk_write_queue);
 901         txm = kcm_tx_msg(head);
 902
 903         if (txm->sent) {
 904                 /* Send of first skbuff in queue already in progress */
 905                 if (WARN_ON(!psock)) {
 906                         ret = -EINVAL;
 907                         goto out;
 908                 }
 909                 sent = txm->sent;
 910                 frag_offset = txm->frag_offset;
 911                 fragidx = txm->fragidx;
 912                 skb = txm->frag_skb;
 913
 914                 goto do_frag;
 915         }
 916
 917 try_again:
 918         psock = reserve_psock(kcm);
 919         if (!psock)
 920                 goto out;
 921
 922         do {
 923                 skb = head;
 924                 txm = kcm_tx_msg(head);
 925                 sent = 0;
 926
 927 do_frag_list:
 928                 if (WARN_ON(!skb_shinfo(skb)->nr_frags)) {
 929                         ret = -EINVAL;
 930                         goto out;
 931                 }
 932
 933                 for (fragidx = 0; fragidx < skb_shinfo(skb)->nr_frags;
 934                      fragidx++) {
 935                         skb_frag_t *frag;
 936
 937                         frag_offset = 0;
 938 do_frag:
 939                         frag = &skb_shinfo(skb)->frags[fragidx];
 940                         if (WARN_ON(!frag->size)) {
 941                                 ret = -EINVAL;
 942                                 goto out;
 943                         }
 944
 945                         ret = kernel_sendpage(psock->sk->sk_socket,
 946                                               frag->page.p,
 947                                               frag->page_offset + frag_offset,
 948                                               frag->size - frag_offset,
 949                                               MSG_DONTWAIT);
 950                         if (ret <= 0) {
 951                                 if (ret == -EAGAIN) {
 952                                         /* Save state to try again when there's
 953                                          * write space on the socket
 954                                          */
 955                                         txm->sent = sent;
 956                                         txm->frag_offset = frag_offset;
 957                                         txm->fragidx = fragidx;
 958                                         txm->frag_skb = skb;
 959
 960                                         ret = 0;
 961                                         goto out;
 962                                 }
 963
 964                                 /* Hard failure in sending message, abort this
 965                                  * psock since it has lost framing
 966                                  * synchonization and retry sending the
 967                                  * message from the beginning.
 968                                  */
 969                                 kcm_abort_tx_psock(psock, ret ? -ret : EPIPE,
 970                                                    true);
 971                                 unreserve_psock(kcm);
 972
 973                                 txm->sent = 0;
 974                                 kcm_report_tx_retry(kcm);
 975                                 ret = 0;
 976
 977                                 goto try_again;
 978                         }
 979
 980                         sent += ret;
 981                         frag_offset += ret;
 982                         KCM_STATS_ADD(psock->stats.tx_bytes, ret);
 983                         if (frag_offset < frag->size) {
 984                                 /* Not finished with this frag */
 985                                 goto do_frag;
 986                         }
 987                 }
 988
 989                 if (skb == head) {
 990                         if (skb_has_frag_list(skb)) {
 991                                 skb = skb_shinfo(skb)->frag_list;
 992                                 goto do_frag_list;
 993                         }
 994                 } else if (skb->next) {
 995                         skb = skb->next;
 996                         goto do_frag_list;
 997                 }
 998
 999                 /* Successfully sent the whole packet, account for it. */
1000                 skb_dequeue(&sk->sk_write_queue);
1001                 kfree_skb(head);
1002                 sk->sk_wmem_queued -= sent;
1003                 total_sent += sent;
1004                 KCM_STATS_INCR(psock->stats.tx_msgs);
1005         } while ((head = skb_peek(&sk->sk_write_queue)));
1006 out:
1007         if (!head) {
1008                 /* Done with all queued messages. */
1009                 WARN_ON(!skb_queue_empty(&sk->sk_write_queue));
1010                 unreserve_psock(kcm);
1011         }
1012
1013         /* Check if write space is available */
1014         sk->sk_write_space(sk);
1015
1016         return total_sent ? : ret;
1017 }
1018
1019 static void kcm_tx_work(struct work_struct *w)
1020 {
1021         struct kcm_sock *kcm = container_of(w, struct kcm_sock, tx_work);
1022         struct sock *sk = &kcm->sk;
1023         int err;
1024
1025         lock_sock(sk);
1026
1027         /* Primarily for SOCK_DGRAM sockets, also handle asynchronous tx
1028          * aborts
1029          */
1030         err = kcm_write_msgs(kcm);
1031         if (err < 0) {
1032                 /* Hard failure in write, report error on KCM socket */
1033                 pr_warn("KCM: Hard failure on kcm_write_msgs %d\n", err);
1034                 report_csk_error(&kcm->sk, -err);
1035                 goto out;
1036         }
1037
1038         /* Primarily for SOCK_SEQPACKET sockets */
1039         if (likely(sk->sk_socket) &&
1040             test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1041                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1042                 sk->sk_write_space(sk);
1043         }
1044
1045 out:
1046         release_sock(sk);
1047 }
1048
1049 static void kcm_push(struct kcm_sock *kcm)
1050 {
1051         if (kcm->tx_wait_more)
1052                 kcm_write_msgs(kcm);
1053 }
1054
1055 static ssize_t kcm_sendpage(struct socket *sock, struct page *page,
1056                             int offset, size_t size, int flags)
1057
1058 {
1059         struct sock *sk = sock->sk;
1060         struct kcm_sock *kcm = kcm_sk(sk);
1061         struct sk_buff *skb = NULL, *head = NULL;
1062         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1063         bool eor;
1064         int err = 0;
1065         int i;
1066
1067         if (flags & MSG_SENDPAGE_NOTLAST)
1068                 flags |= MSG_MORE;
1069
1070         /* No MSG_EOR from splice, only look at MSG_MORE */
1071         eor = !(flags & MSG_MORE);
1072
1073         lock_sock(sk);
1074
1075         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1076
1077         err = -EPIPE;
1078         if (sk->sk_err)
1079                 goto out_error;
1080
1081         if (kcm->seq_skb) {
1082                 /* Previously opened message */
1083                 head = kcm->seq_skb;
1084                 skb = kcm_tx_msg(head)->last_skb;
1085                 i = skb_shinfo(skb)->nr_frags;
1086
1087                 if (skb_can_coalesce(skb, i, page, offset)) {
1088                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], size);
1089                         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1090                         goto coalesced;
1091                 }
1092
1093                 if (i >= MAX_SKB_FRAGS) {
1094                         struct sk_buff *tskb;
1095
1096                         tskb = alloc_skb(0, sk->sk_allocation);
1097                         while (!tskb) {
1098                                 kcm_push(kcm);
1099                                 err = sk_stream_wait_memory(sk, &timeo);
1100                                 if (err)
1101                                         goto out_error;
1102                         }
1103
1104                         if (head == skb)
1105                                 skb_shinfo(head)->frag_list = tskb;
1106                         else
1107                                 skb->next = tskb;
1108
1109                         skb = tskb;
1110                         skb->ip_summed = CHECKSUM_UNNECESSARY;
1111                         i = 0;
1112                 }
1113         } else {
1114                 /* Call the sk_stream functions to manage the sndbuf mem. */
1115                 if (!sk_stream_memory_free(sk)) {
1116                         kcm_push(kcm);
1117                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1118                         err = sk_stream_wait_memory(sk, &timeo);
1119                         if (err)
1120                                 goto out_error;
1121                 }
1122
1123                 head = alloc_skb(0, sk->sk_allocation);
1124                 while (!head) {
1125                         kcm_push(kcm);
1126                         err = sk_stream_wait_memory(sk, &timeo);
1127                         if (err)
1128                                 goto out_error;
1129                 }
1130
1131                 skb = head;
1132                 i = 0;
1133         }
1134
1135         get_page(page);
1136         skb_fill_page_desc(skb, i, page, offset, size);
1137         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1138
1139 coalesced:
1140         skb->len += size;
1141         skb->data_len += size;
1142         skb->truesize += size;
1143         sk->sk_wmem_queued += size;
1144         sk_mem_charge(sk, size);
1145
1146         if (head != skb) {
1147                 head->len += size;
1148                 head->data_len += size;
1149                 head->truesize += size;
1150         }
1151
1152         if (eor) {
1153                 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
1154
1155                 /* Message complete, queue it on send buffer */
1156                 __skb_queue_tail(&sk->sk_write_queue, head);
1157                 kcm->seq_skb = NULL;
1158                 KCM_STATS_INCR(kcm->stats.tx_msgs);
1159
1160                 if (flags & MSG_BATCH) {
1161                         kcm->tx_wait_more = true;
1162                 } else if (kcm->tx_wait_more || not_busy) {
1163                         err = kcm_write_msgs(kcm);
1164                         if (err < 0) {
1165                                 /* We got a hard error in write_msgs but have
1166                                  * already queued this message. Report an error
1167                                  * in the socket, but don't affect return value
1168                                  * from sendmsg
1169                                  */
1170                                 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
1171                                 report_csk_error(&kcm->sk, -err);
1172                         }
1173                 }
1174         } else {
1175                 /* Message not complete, save state */
1176                 kcm->seq_skb = head;
1177                 kcm_tx_msg(head)->last_skb = skb;
1178         }
1179
1180         KCM_STATS_ADD(kcm->stats.tx_bytes, size);
1181
1182         release_sock(sk);
1183         return size;
1184
1185 out_error:
1186         kcm_push(kcm);
1187
1188         err = sk_stream_error(sk, flags, err);
1189
1190         /* make sure we wake any epoll edge trigger waiter */
1191         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1192                 sk->sk_write_space(sk);
1193
1194         release_sock(sk);
1195         return err;
1196 }
1197
1198 static int kcm_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1199 {
1200         struct sock *sk = sock->sk;
1201         struct kcm_sock *kcm = kcm_sk(sk);
1202         struct sk_buff *skb = NULL, *head = NULL;
1203         size_t copy, copied = 0;
1204         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1205         int eor = (sock->type == SOCK_DGRAM) ?
1206                   !(msg->msg_flags & MSG_MORE) : !!(msg->msg_flags & MSG_EOR);
1207         int err = -EPIPE;
1208
1209         lock_sock(sk);
1210
1211         /* Per tcp_sendmsg this should be in poll */
1212         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1213
1214         if (sk->sk_err)
1215                 goto out_error;
1216
1217         if (kcm->seq_skb) {
1218                 /* Previously opened message */
1219                 head = kcm->seq_skb;
1220                 skb = kcm_tx_msg(head)->last_skb;
1221                 goto start;
1222         }
1223
1224         /* Call the sk_stream functions to manage the sndbuf mem. */
1225         if (!sk_stream_memory_free(sk)) {
1226                 kcm_push(kcm);
1227                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1228                 err = sk_stream_wait_memory(sk, &timeo);
1229                 if (err)
1230                         goto out_error;
1231         }
1232
1233         /* New message, alloc head skb */
1234         head = alloc_skb(0, sk->sk_allocation);
1235         while (!head) {
1236                 kcm_push(kcm);
1237                 err = sk_stream_wait_memory(sk, &timeo);
1238                 if (err)
1239                         goto out_error;
1240
1241                 head = alloc_skb(0, sk->sk_allocation);
1242         }
1243
1244         skb = head;
1245
1246         /* Set ip_summed to CHECKSUM_UNNECESSARY to avoid calling
1247          * csum_and_copy_from_iter from skb_do_copy_data_nocache.
1248          */
1249         skb->ip_summed = CHECKSUM_UNNECESSARY;
1250
1251 start:
1252         while (msg_data_left(msg)) {
1253                 bool merge = true;
1254                 int i = skb_shinfo(skb)->nr_frags;
1255                 struct page_frag *pfrag = sk_page_frag(sk);
1256
1257                 if (!sk_page_frag_refill(sk, pfrag))
1258                         goto wait_for_memory;
1259
1260                 if (!skb_can_coalesce(skb, i, pfrag->page,
1261                                       pfrag->offset)) {
1262                         if (i == MAX_SKB_FRAGS) {
1263                                 struct sk_buff *tskb;
1264
1265                                 tskb = alloc_skb(0, sk->sk_allocation);
1266                                 if (!tskb)
1267                                         goto wait_for_memory;
1268
1269                                 if (head == skb)
1270                                         skb_shinfo(head)->frag_list = tskb;
1271                                 else
1272                                         skb->next = tskb;
1273
1274                                 skb = tskb;
1275                                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1276                                 continue;
1277                         }
1278                         merge = false;
1279                 }
1280
1281                 copy = min_t(int, msg_data_left(msg),
1282                              pfrag->size - pfrag->offset);
1283
1284                 if (!sk_wmem_schedule(sk, copy))
1285                         goto wait_for_memory;
1286
1287                 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1288                                                pfrag->page,
1289                                                pfrag->offset,
1290                                                copy);
1291                 if (err)
1292                         goto out_error;
1293
1294                 /* Update the skb. */
1295                 if (merge) {
1296                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1297                 } else {
1298                         skb_fill_page_desc(skb, i, pfrag->page,
1299                                            pfrag->offset, copy);
1300                         get_page(pfrag->page);
1301                 }
1302
1303                 pfrag->offset += copy;
1304                 copied += copy;
1305                 if (head != skb) {
1306                         head->len += copy;
1307                         head->data_len += copy;
1308                 }
1309
1310                 continue;
1311
1312 wait_for_memory:
1313                 kcm_push(kcm);
1314                 err = sk_stream_wait_memory(sk, &timeo);
1315                 if (err)
1316                         goto out_error;
1317         }
1318
1319         if (eor) {
1320                 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
1321
1322                 /* Message complete, queue it on send buffer */
1323                 __skb_queue_tail(&sk->sk_write_queue, head);
1324                 kcm->seq_skb = NULL;
1325                 KCM_STATS_INCR(kcm->stats.tx_msgs);
1326
1327                 if (msg->msg_flags & MSG_BATCH) {
1328                         kcm->tx_wait_more = true;
1329                 } else if (kcm->tx_wait_more || not_busy) {
1330                         err = kcm_write_msgs(kcm);
1331                         if (err < 0) {
1332                                 /* We got a hard error in write_msgs but have
1333                                  * already queued this message. Report an error
1334                                  * in the socket, but don't affect return value
1335                                  * from sendmsg
1336                                  */
1337                                 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
1338                                 report_csk_error(&kcm->sk, -err);
1339                         }
1340                 }
1341         } else {
1342                 /* Message not complete, save state */
1343 partial_message:
1344                 kcm->seq_skb = head;
1345                 kcm_tx_msg(head)->last_skb = skb;
1346         }
1347
1348         KCM_STATS_ADD(kcm->stats.tx_bytes, copied);
1349
1350         release_sock(sk);
1351         return copied;
1352
1353 out_error:
1354         kcm_push(kcm);
1355
1356         if (copied && sock->type == SOCK_SEQPACKET) {
1357                 /* Wrote some bytes before encountering an
1358                  * error, return partial success.
1359                  */
1360                 goto partial_message;
1361         }
1362
1363         if (head != kcm->seq_skb)
1364                 kfree_skb(head);
1365
1366         err = sk_stream_error(sk, msg->msg_flags, err);
1367
1368         /* make sure we wake any epoll edge trigger waiter */
1369         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1370                 sk->sk_write_space(sk);
1371
1372         release_sock(sk);
1373         return err;
1374 }
1375
1376 static struct sk_buff *kcm_wait_data(struct sock *sk, int flags,
1377                                      long timeo, int *err)
1378 {
1379         struct sk_buff *skb;
1380
1381         while (!(skb = skb_peek(&sk->sk_receive_queue))) {
1382                 if (sk->sk_err) {
1383                         *err = sock_error(sk);
1384                         return NULL;
1385                 }
1386
1387                 if (sock_flag(sk, SOCK_DONE))
1388                         return NULL;
1389
1390                 if ((flags & MSG_DONTWAIT) || !timeo) {
1391                         *err = -EAGAIN;
1392                         return NULL;
1393                 }
1394
1395                 sk_wait_data(sk, &timeo, NULL);
1396
1397                 /* Handle signals */
1398                 if (signal_pending(current)) {
1399                         *err = sock_intr_errno(timeo);
1400                         return NULL;
1401                 }
1402         }
1403
1404         return skb;
1405 }
1406
1407 static int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
1408                        size_t len, int flags)
1409 {
1410         struct sock *sk = sock->sk;
1411         struct kcm_sock *kcm = kcm_sk(sk);
1412         int err = 0;
1413         long timeo;
1414         struct kcm_rx_msg *rxm;
1415         int copied = 0;
1416         struct sk_buff *skb;
1417
1418         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1419
1420         lock_sock(sk);
1421
1422         skb = kcm_wait_data(sk, flags, timeo, &err);
1423         if (!skb)
1424                 goto out;
1425
1426         /* Okay, have a message on the receive queue */
1427
1428         rxm = kcm_rx_msg(skb);
1429
1430         if (len > rxm->full_len)
1431                 len = rxm->full_len;
1432
1433         err = skb_copy_datagram_msg(skb, rxm->offset, msg, len);
1434         if (err < 0)
1435                 goto out;
1436
1437         copied = len;
1438         if (likely(!(flags & MSG_PEEK))) {
1439                 KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1440                 if (copied < rxm->full_len) {
1441                         if (sock->type == SOCK_DGRAM) {
1442                                 /* Truncated message */
1443                                 msg->msg_flags |= MSG_TRUNC;
1444                                 goto msg_finished;
1445                         }
1446                         rxm->offset += copied;
1447                         rxm->full_len -= copied;
1448                 } else {
1449 msg_finished:
1450                         /* Finished with message */
1451                         msg->msg_flags |= MSG_EOR;
1452                         KCM_STATS_INCR(kcm->stats.rx_msgs);
1453                         skb_unlink(skb, &sk->sk_receive_queue);
1454                         kfree_skb(skb);
1455                 }
1456         }
1457
1458 out:
1459         release_sock(sk);
1460
1461         return copied ? : err;
1462 }
1463
1464 static ssize_t kcm_sock_splice(struct sock *sk,
1465                                struct pipe_inode_info *pipe,
1466                                struct splice_pipe_desc *spd)
1467 {
1468         int ret;
1469
1470         release_sock(sk);
1471         ret = splice_to_pipe(pipe, spd);
1472         lock_sock(sk);
1473
1474         return ret;
1475 }
1476
1477 static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1478                                struct pipe_inode_info *pipe, size_t len,
1479                                unsigned int flags)
1480 {
1481         struct sock *sk = sock->sk;
1482         struct kcm_sock *kcm = kcm_sk(sk);
1483         long timeo;
1484         struct kcm_rx_msg *rxm;
1485         int err = 0;
1486         ssize_t copied;
1487         struct sk_buff *skb;
1488
1489         /* Only support splice for SOCKSEQPACKET */
1490
1491         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1492
1493         lock_sock(sk);
1494
1495         skb = kcm_wait_data(sk, flags, timeo, &err);
1496         if (!skb)
1497                 goto err_out;
1498
1499         /* Okay, have a message on the receive queue */
1500
1501         rxm = kcm_rx_msg(skb);
1502
1503         if (len > rxm->full_len)
1504                 len = rxm->full_len;
1505
1506         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, len, flags,
1507                                  kcm_sock_splice);
1508         if (copied < 0) {
1509                 err = copied;
1510                 goto err_out;
1511         }
1512
1513         KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1514
1515         rxm->offset += copied;
1516         rxm->full_len -= copied;
1517
1518         /* We have no way to return MSG_EOR. If all the bytes have been
1519          * read we still leave the message in the receive socket buffer.
1520          * A subsequent recvmsg needs to be done to return MSG_EOR and
1521          * finish reading the message.
1522          */
1523
1524         release_sock(sk);
1525
1526         return copied;
1527
1528 err_out:
1529         release_sock(sk);
1530
1531         return err;
1532 }
1533
1534 /* kcm sock lock held */
1535 static void kcm_recv_disable(struct kcm_sock *kcm)
1536 {
1537         struct kcm_mux *mux = kcm->mux;
1538
1539         if (kcm->rx_disabled)
1540                 return;
1541
1542         spin_lock_bh(&mux->rx_lock);
1543
1544         kcm->rx_disabled = 1;
1545
1546         /* If a psock is reserved we'll do cleanup in unreserve */
1547         if (!kcm->rx_psock) {
1548                 if (kcm->rx_wait) {
1549                         list_del(&kcm->wait_rx_list);
1550                         kcm->rx_wait = false;
1551                 }
1552
1553                 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1554         }
1555
1556         spin_unlock_bh(&mux->rx_lock);
1557 }
1558
1559 /* kcm sock lock held */
1560 static void kcm_recv_enable(struct kcm_sock *kcm)
1561 {
1562         struct kcm_mux *mux = kcm->mux;
1563
1564         if (!kcm->rx_disabled)
1565                 return;
1566
1567         spin_lock_bh(&mux->rx_lock);
1568
1569         kcm->rx_disabled = 0;
1570         kcm_rcv_ready(kcm);
1571
1572         spin_unlock_bh(&mux->rx_lock);
1573 }
1574
1575 static int kcm_setsockopt(struct socket *sock, int level, int optname,
1576                           char __user *optval, unsigned int optlen)
1577 {
1578         struct kcm_sock *kcm = kcm_sk(sock->sk);
1579         int val, valbool;
1580         int err = 0;
1581
1582         if (level != SOL_KCM)
1583                 return -ENOPROTOOPT;
1584
1585         if (optlen < sizeof(int))
1586                 return -EINVAL;
1587
1588         if (get_user(val, (int __user *)optval))
1589                 return -EINVAL;
1590
1591         valbool = val ? 1 : 0;
1592
1593         switch (optname) {
1594         case KCM_RECV_DISABLE:
1595                 lock_sock(&kcm->sk);
1596                 if (valbool)
1597                         kcm_recv_disable(kcm);
1598                 else
1599                         kcm_recv_enable(kcm);
1600                 release_sock(&kcm->sk);
1601                 break;
1602         default:
1603                 err = -ENOPROTOOPT;
1604         }
1605
1606         return err;
1607 }
1608
1609 static int kcm_getsockopt(struct socket *sock, int level, int optname,
1610                           char __user *optval, int __user *optlen)
1611 {
1612         struct kcm_sock *kcm = kcm_sk(sock->sk);
1613         int val, len;
1614
1615         if (level != SOL_KCM)
1616                 return -ENOPROTOOPT;
1617
1618         if (get_user(len, optlen))
1619                 return -EFAULT;
1620
1621         len = min_t(unsigned int, len, sizeof(int));
1622         if (len < 0)
1623                 return -EINVAL;
1624
1625         switch (optname) {
1626         case KCM_RECV_DISABLE:
1627                 val = kcm->rx_disabled;
1628                 break;
1629         default:
1630                 return -ENOPROTOOPT;
1631         }
1632
1633         if (put_user(len, optlen))
1634                 return -EFAULT;
1635         if (copy_to_user(optval, &val, len))
1636                 return -EFAULT;
1637         return 0;
1638 }
1639
1640 static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1641 {
1642         struct kcm_sock *tkcm;
1643         struct list_head *head;
1644         int index = 0;
1645
1646         /* For SOCK_SEQPACKET sock type, datagram_poll checks the sk_state, so
1647          * we set sk_state, otherwise epoll_wait always returns right away with
1648          * POLLHUP
1649          */
1650         kcm->sk.sk_state = TCP_ESTABLISHED;
1651
1652         /* Add to mux's kcm sockets list */
1653         kcm->mux = mux;
1654         spin_lock_bh(&mux->lock);
1655
1656         head = &mux->kcm_socks;
1657         list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1658                 if (tkcm->index != index)
1659                         break;
1660                 head = &tkcm->kcm_sock_list;
1661                 index++;
1662         }
1663
1664         list_add(&kcm->kcm_sock_list, head);
1665         kcm->index = index;
1666
1667         mux->kcm_socks_cnt++;
1668         spin_unlock_bh(&mux->lock);
1669
1670         INIT_WORK(&kcm->tx_work, kcm_tx_work);
1671
1672         spin_lock_bh(&mux->rx_lock);
1673         kcm_rcv_ready(kcm);
1674         spin_unlock_bh(&mux->rx_lock);
1675 }
1676
1677 static void kcm_rx_msg_timeout(unsigned long arg)
1678 {
1679         struct kcm_psock *psock = (struct kcm_psock *)arg;
1680
1681         /* Message assembly timed out */
1682         KCM_STATS_INCR(psock->stats.rx_msg_timeouts);
1683         kcm_abort_rx_psock(psock, ETIMEDOUT, NULL);
1684 }
1685
1686 static int kcm_attach(struct socket *sock, struct socket *csock,
1687                       struct bpf_prog *prog)
1688 {
1689         struct kcm_sock *kcm = kcm_sk(sock->sk);
1690         struct kcm_mux *mux = kcm->mux;
1691         struct sock *csk;
1692         struct kcm_psock *psock = NULL, *tpsock;
1693         struct list_head *head;
1694         int index = 0;
1695
1696         if (csock->ops->family != PF_INET &&
1697             csock->ops->family != PF_INET6)
1698                 return -EINVAL;
1699
1700         csk = csock->sk;
1701         if (!csk)
1702                 return -EINVAL;
1703
1704         /* Only support TCP for now */
1705         if (csk->sk_protocol != IPPROTO_TCP)
1706                 return -EINVAL;
1707
1708         psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1709         if (!psock)
1710                 return -ENOMEM;
1711
1712         psock->mux = mux;
1713         psock->sk = csk;
1714         psock->bpf_prog = prog;
1715
1716         setup_timer(&psock->rx_msg_timer, kcm_rx_msg_timeout,
1717                     (unsigned long)psock);
1718
1719         INIT_WORK(&psock->rx_work, psock_rx_work);
1720         INIT_DELAYED_WORK(&psock->rx_delayed_work, psock_rx_delayed_work);
1721
1722         sock_hold(csk);
1723
1724         write_lock_bh(&csk->sk_callback_lock);
1725         psock->save_data_ready = csk->sk_data_ready;
1726         psock->save_write_space = csk->sk_write_space;
1727         psock->save_state_change = csk->sk_state_change;
1728         csk->sk_user_data = psock;
1729         csk->sk_data_ready = psock_tcp_data_ready;
1730         csk->sk_write_space = psock_tcp_write_space;
1731         csk->sk_state_change = psock_tcp_state_change;
1732         write_unlock_bh(&csk->sk_callback_lock);
1733
1734         /* Finished initialization, now add the psock to the MUX. */
1735         spin_lock_bh(&mux->lock);
1736         head = &mux->psocks;
1737         list_for_each_entry(tpsock, &mux->psocks, psock_list) {
1738                 if (tpsock->index != index)
1739                         break;
1740                 head = &tpsock->psock_list;
1741                 index++;
1742         }
1743
1744         list_add(&psock->psock_list, head);
1745         psock->index = index;
1746
1747         KCM_STATS_INCR(mux->stats.psock_attach);
1748         mux->psocks_cnt++;
1749         psock_now_avail(psock);
1750         spin_unlock_bh(&mux->lock);
1751
1752         /* Schedule RX work in case there are already bytes queued */
1753         queue_work(kcm_wq, &psock->rx_work);
1754
1755         return 0;
1756 }
1757
1758 static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1759 {
1760         struct socket *csock;
1761         struct bpf_prog *prog;
1762         int err;
1763
1764         csock = sockfd_lookup(info->fd, &err);
1765         if (!csock)
1766                 return -ENOENT;
1767
1768         prog = bpf_prog_get_type(info->bpf_fd, BPF_PROG_TYPE_SOCKET_FILTER);
1769         if (IS_ERR(prog)) {
1770                 err = PTR_ERR(prog);
1771                 goto out;
1772         }
1773
1774         err = kcm_attach(sock, csock, prog);
1775         if (err) {
1776                 bpf_prog_put(prog);
1777                 goto out;
1778         }
1779
1780         /* Keep reference on file also */
1781
1782         return 0;
1783 out:
1784         fput(csock->file);
1785         return err;
1786 }
1787
1788 static void kcm_unattach(struct kcm_psock *psock)
1789 {
1790         struct sock *csk = psock->sk;
1791         struct kcm_mux *mux = psock->mux;
1792
1793         /* Stop getting callbacks from TCP socket. After this there should
1794          * be no way to reserve a kcm for this psock.
1795          */
1796         write_lock_bh(&csk->sk_callback_lock);
1797         csk->sk_user_data = NULL;
1798         csk->sk_data_ready = psock->save_data_ready;
1799         csk->sk_write_space = psock->save_write_space;
1800         csk->sk_state_change = psock->save_state_change;
1801         psock->rx_stopped = 1;
1802
1803         if (WARN_ON(psock->rx_kcm)) {
1804                 write_unlock_bh(&csk->sk_callback_lock);
1805                 return;
1806         }
1807
1808         spin_lock_bh(&mux->rx_lock);
1809
1810         /* Stop receiver activities. After this point psock should not be
1811          * able to get onto ready list either through callbacks or work.
1812          */
1813         if (psock->ready_rx_msg) {
1814                 list_del(&psock->psock_ready_list);
1815                 kfree_skb(psock->ready_rx_msg);
1816                 psock->ready_rx_msg = NULL;
1817                 KCM_STATS_INCR(mux->stats.rx_ready_drops);
1818         }
1819
1820         spin_unlock_bh(&mux->rx_lock);
1821
1822         write_unlock_bh(&csk->sk_callback_lock);
1823
1824         del_timer_sync(&psock->rx_msg_timer);
1825         cancel_work_sync(&psock->rx_work);
1826         cancel_delayed_work_sync(&psock->rx_delayed_work);
1827
1828         bpf_prog_put(psock->bpf_prog);
1829
1830         kfree_skb(psock->rx_skb_head);
1831         psock->rx_skb_head = NULL;
1832
1833         spin_lock_bh(&mux->lock);
1834
1835         aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1836
1837         KCM_STATS_INCR(mux->stats.psock_unattach);
1838
1839         if (psock->tx_kcm) {
1840                 /* psock was reserved.  Just mark it finished and we will clean
1841                  * up in the kcm paths, we need kcm lock which can not be
1842                  * acquired here.
1843                  */
1844                 KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1845                 spin_unlock_bh(&mux->lock);
1846
1847                 /* We are unattaching a socket that is reserved. Abort the
1848                  * socket since we may be out of sync in sending on it. We need
1849                  * to do this without the mux lock.
1850                  */
1851                 kcm_abort_tx_psock(psock, EPIPE, false);
1852
1853                 spin_lock_bh(&mux->lock);
1854                 if (!psock->tx_kcm) {
1855                         /* psock now unreserved in window mux was unlocked */
1856                         goto no_reserved;
1857                 }
1858                 psock->done = 1;
1859
1860                 /* Commit done before queuing work to process it */
1861                 smp_mb();
1862
1863                 /* Queue tx work to make sure psock->done is handled */
1864                 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
1865                 spin_unlock_bh(&mux->lock);
1866         } else {
1867 no_reserved:
1868                 if (!psock->tx_stopped)
1869                         list_del(&psock->psock_avail_list);
1870                 list_del(&psock->psock_list);
1871                 mux->psocks_cnt--;
1872                 spin_unlock_bh(&mux->lock);
1873
1874                 sock_put(csk);
1875                 fput(csk->sk_socket->file);
1876                 kmem_cache_free(kcm_psockp, psock);
1877         }
1878 }
1879
1880 static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1881 {
1882         struct kcm_sock *kcm = kcm_sk(sock->sk);
1883         struct kcm_mux *mux = kcm->mux;
1884         struct kcm_psock *psock;
1885         struct socket *csock;
1886         struct sock *csk;
1887         int err;
1888
1889         csock = sockfd_lookup(info->fd, &err);
1890         if (!csock)
1891                 return -ENOENT;
1892
1893         csk = csock->sk;
1894         if (!csk) {
1895                 err = -EINVAL;
1896                 goto out;
1897         }
1898
1899         err = -ENOENT;
1900
1901         spin_lock_bh(&mux->lock);
1902
1903         list_for_each_entry(psock, &mux->psocks, psock_list) {
1904                 if (psock->sk != csk)
1905                         continue;
1906
1907                 /* Found the matching psock */
1908
1909                 if (psock->unattaching || WARN_ON(psock->done)) {
1910                         err = -EALREADY;
1911                         break;
1912                 }
1913
1914                 psock->unattaching = 1;
1915
1916                 spin_unlock_bh(&mux->lock);
1917
1918                 kcm_unattach(psock);
1919
1920                 err = 0;
1921                 goto out;
1922         }
1923
1924         spin_unlock_bh(&mux->lock);
1925
1926 out:
1927         fput(csock->file);
1928         return err;
1929 }
1930
1931 static struct proto kcm_proto = {
1932         .name   = "KCM",
1933         .owner  = THIS_MODULE,
1934         .obj_size = sizeof(struct kcm_sock),
1935 };
1936
1937 /* Clone a kcm socket. */
1938 static int kcm_clone(struct socket *osock, struct kcm_clone *info,
1939                      struct socket **newsockp)
1940 {
1941         struct socket *newsock;
1942         struct sock *newsk;
1943         struct file *newfile;
1944         int err, newfd;
1945
1946         err = -ENFILE;
1947         newsock = sock_alloc();
1948         if (!newsock)
1949                 goto out;
1950
1951         newsock->type = osock->type;
1952         newsock->ops = osock->ops;
1953
1954         __module_get(newsock->ops->owner);
1955
1956         newfd = get_unused_fd_flags(0);
1957         if (unlikely(newfd < 0)) {
1958                 err = newfd;
1959                 goto out_fd_fail;
1960         }
1961
1962         newfile = sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1963         if (unlikely(IS_ERR(newfile))) {
1964                 err = PTR_ERR(newfile);
1965                 goto out_sock_alloc_fail;
1966         }
1967
1968         newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1969                          &kcm_proto, true);
1970         if (!newsk) {
1971                 err = -ENOMEM;
1972                 goto out_sk_alloc_fail;
1973         }
1974
1975         sock_init_data(newsock, newsk);
1976         init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1977
1978         fd_install(newfd, newfile);
1979         *newsockp = newsock;
1980         info->fd = newfd;
1981
1982         return 0;
1983
1984 out_sk_alloc_fail:
1985         fput(newfile);
1986 out_sock_alloc_fail:
1987         put_unused_fd(newfd);
1988 out_fd_fail:
1989         sock_release(newsock);
1990 out:
1991         return err;
1992 }
1993
1994 static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
1995 {
1996         int err;
1997
1998         switch (cmd) {
1999         case SIOCKCMATTACH: {
2000                 struct kcm_attach info;
2001
2002                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
2003                         err = -EFAULT;
2004
2005                 err = kcm_attach_ioctl(sock, &info);
2006
2007                 break;
2008         }
2009         case SIOCKCMUNATTACH: {
2010                 struct kcm_unattach info;
2011
2012                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
2013                         err = -EFAULT;
2014
2015                 err = kcm_unattach_ioctl(sock, &info);
2016
2017                 break;
2018         }
2019         case SIOCKCMCLONE: {
2020                 struct kcm_clone info;
2021                 struct socket *newsock = NULL;
2022
2023                 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
2024                         err = -EFAULT;
2025
2026                 err = kcm_clone(sock, &info, &newsock);
2027
2028                 if (!err) {
2029                         if (copy_to_user((void __user *)arg, &info,
2030                                          sizeof(info))) {
2031                                 err = -EFAULT;
2032                                 sock_release(newsock);
2033                         }
2034                 }
2035
2036                 break;
2037         }
2038         default:
2039                 err = -ENOIOCTLCMD;
2040                 break;
2041         }
2042
2043         return err;
2044 }
2045
2046 static void free_mux(struct rcu_head *rcu)
2047 {
2048         struct kcm_mux *mux = container_of(rcu,
2049             struct kcm_mux, rcu);
2050
2051         kmem_cache_free(kcm_muxp, mux);
2052 }
2053
2054 static void release_mux(struct kcm_mux *mux)
2055 {
2056         struct kcm_net *knet = mux->knet;
2057         struct kcm_psock *psock, *tmp_psock;
2058
2059         /* Release psocks */
2060         list_for_each_entry_safe(psock, tmp_psock,
2061                                  &mux->psocks, psock_list) {
2062                 if (!WARN_ON(psock->unattaching))
2063                         kcm_unattach(psock);
2064         }
2065
2066         if (WARN_ON(mux->psocks_cnt))
2067                 return;
2068
2069         __skb_queue_purge(&mux->rx_hold_queue);
2070
2071         mutex_lock(&knet->mutex);
2072         aggregate_mux_stats(&mux->stats, &knet->aggregate_mux_stats);
2073         aggregate_psock_stats(&mux->aggregate_psock_stats,
2074                               &knet->aggregate_psock_stats);
2075         list_del_rcu(&mux->kcm_mux_list);
2076         knet->count--;
2077         mutex_unlock(&knet->mutex);
2078
2079         call_rcu(&mux->rcu, free_mux);
2080 }
2081
2082 static void kcm_done(struct kcm_sock *kcm)
2083 {
2084         struct kcm_mux *mux = kcm->mux;
2085         struct sock *sk = &kcm->sk;
2086         int socks_cnt;
2087
2088         spin_lock_bh(&mux->rx_lock);
2089         if (kcm->rx_psock) {
2090                 /* Cleanup in unreserve_rx_kcm */
2091                 WARN_ON(kcm->done);
2092                 kcm->rx_disabled = 1;
2093                 kcm->done = 1;
2094                 spin_unlock_bh(&mux->rx_lock);
2095                 return;
2096         }
2097
2098         if (kcm->rx_wait) {
2099                 list_del(&kcm->wait_rx_list);
2100                 kcm->rx_wait = false;
2101         }
2102         /* Move any pending receive messages to other kcm sockets */
2103         requeue_rx_msgs(mux, &sk->sk_receive_queue);
2104
2105         spin_unlock_bh(&mux->rx_lock);
2106
2107         if (WARN_ON(sk_rmem_alloc_get(sk)))
2108                 return;
2109
2110         /* Detach from MUX */
2111         spin_lock_bh(&mux->lock);
2112
2113         list_del(&kcm->kcm_sock_list);
2114         mux->kcm_socks_cnt--;
2115         socks_cnt = mux->kcm_socks_cnt;
2116
2117         spin_unlock_bh(&mux->lock);
2118
2119         if (!socks_cnt) {
2120                 /* We are done with the mux now. */
2121                 release_mux(mux);
2122         }
2123
2124         WARN_ON(kcm->rx_wait);
2125
2126         sock_put(&kcm->sk);
2127 }
2128
2129 /* Called by kcm_release to close a KCM socket.
2130  * If this is the last KCM socket on the MUX, destroy the MUX.
2131  */
2132 static int kcm_release(struct socket *sock)
2133 {
2134         struct sock *sk = sock->sk;
2135         struct kcm_sock *kcm;
2136         struct kcm_mux *mux;
2137         struct kcm_psock *psock;
2138
2139         if (!sk)
2140                 return 0;
2141
2142         kcm = kcm_sk(sk);
2143         mux = kcm->mux;
2144
2145         sock_orphan(sk);
2146         kfree_skb(kcm->seq_skb);
2147
2148         lock_sock(sk);
2149         /* Purge queue under lock to avoid race condition with tx_work trying
2150          * to act when queue is nonempty. If tx_work runs after this point
2151          * it will just return.
2152          */
2153         __skb_queue_purge(&sk->sk_write_queue);
2154         release_sock(sk);
2155
2156         spin_lock_bh(&mux->lock);
2157         if (kcm->tx_wait) {
2158                 /* Take of tx_wait list, after this point there should be no way
2159                  * that a psock will be assigned to this kcm.
2160                  */
2161                 list_del(&kcm->wait_psock_list);
2162                 kcm->tx_wait = false;
2163         }
2164         spin_unlock_bh(&mux->lock);
2165
2166         /* Cancel work. After this point there should be no outside references
2167          * to the kcm socket.
2168          */
2169         cancel_work_sync(&kcm->tx_work);
2170
2171         lock_sock(sk);
2172         psock = kcm->tx_psock;
2173         if (psock) {
2174                 /* A psock was reserved, so we need to kill it since it
2175                  * may already have some bytes queued from a message. We
2176                  * need to do this after removing kcm from tx_wait list.
2177                  */
2178                 kcm_abort_tx_psock(psock, EPIPE, false);
2179                 unreserve_psock(kcm);
2180         }
2181         release_sock(sk);
2182
2183         WARN_ON(kcm->tx_wait);
2184         WARN_ON(kcm->tx_psock);
2185
2186         sock->sk = NULL;
2187
2188         kcm_done(kcm);
2189
2190         return 0;
2191 }
2192
2193 static const struct proto_ops kcm_dgram_ops = {
2194         .family =       PF_KCM,
2195         .owner =        THIS_MODULE,
2196         .release =      kcm_release,
2197         .bind =         sock_no_bind,
2198         .connect =      sock_no_connect,
2199         .socketpair =   sock_no_socketpair,
2200         .accept =       sock_no_accept,
2201         .getname =      sock_no_getname,
2202         .poll =         datagram_poll,
2203         .ioctl =        kcm_ioctl,
2204         .listen =       sock_no_listen,
2205         .shutdown =     sock_no_shutdown,
2206         .setsockopt =   kcm_setsockopt,
2207         .getsockopt =   kcm_getsockopt,
2208         .sendmsg =      kcm_sendmsg,
2209         .recvmsg =      kcm_recvmsg,
2210         .mmap =         sock_no_mmap,
2211         .sendpage =     kcm_sendpage,
2212 };
2213
2214 static const struct proto_ops kcm_seqpacket_ops = {
2215         .family =       PF_KCM,
2216         .owner =        THIS_MODULE,
2217         .release =      kcm_release,
2218         .bind =         sock_no_bind,
2219         .connect =      sock_no_connect,
2220         .socketpair =   sock_no_socketpair,
2221         .accept =       sock_no_accept,
2222         .getname =      sock_no_getname,
2223         .poll =         datagram_poll,
2224         .ioctl =        kcm_ioctl,
2225         .listen =       sock_no_listen,
2226         .shutdown =     sock_no_shutdown,
2227         .setsockopt =   kcm_setsockopt,
2228         .getsockopt =   kcm_getsockopt,
2229         .sendmsg =      kcm_sendmsg,
2230         .recvmsg =      kcm_recvmsg,
2231         .mmap =         sock_no_mmap,
2232         .sendpage =     kcm_sendpage,
2233         .splice_read =  kcm_splice_read,
2234 };
2235
2236 /* Create proto operation for kcm sockets */
2237 static int kcm_create(struct net *net, struct socket *sock,
2238                       int protocol, int kern)
2239 {
2240         struct kcm_net *knet = net_generic(net, kcm_net_id);
2241         struct sock *sk;
2242         struct kcm_mux *mux;
2243
2244         switch (sock->type) {
2245         case SOCK_DGRAM:
2246                 sock->ops = &kcm_dgram_ops;
2247                 break;
2248         case SOCK_SEQPACKET:
2249                 sock->ops = &kcm_seqpacket_ops;
2250                 break;
2251         default:
2252                 return -ESOCKTNOSUPPORT;
2253         }
2254
2255         if (protocol != KCMPROTO_CONNECTED)
2256                 return -EPROTONOSUPPORT;
2257
2258         sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
2259         if (!sk)
2260                 return -ENOMEM;
2261
2262         /* Allocate a kcm mux, shared between KCM sockets */
2263         mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
2264         if (!mux) {
2265                 sk_free(sk);
2266                 return -ENOMEM;
2267         }
2268
2269         spin_lock_init(&mux->lock);
2270         spin_lock_init(&mux->rx_lock);
2271         INIT_LIST_HEAD(&mux->kcm_socks);
2272         INIT_LIST_HEAD(&mux->kcm_rx_waiters);
2273         INIT_LIST_HEAD(&mux->kcm_tx_waiters);
2274
2275         INIT_LIST_HEAD(&mux->psocks);
2276         INIT_LIST_HEAD(&mux->psocks_ready);
2277         INIT_LIST_HEAD(&mux->psocks_avail);
2278
2279         mux->knet = knet;
2280
2281         /* Add new MUX to list */
2282         mutex_lock(&knet->mutex);
2283         list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
2284         knet->count++;
2285         mutex_unlock(&knet->mutex);
2286
2287         skb_queue_head_init(&mux->rx_hold_queue);
2288
2289         /* Init KCM socket */
2290         sock_init_data(sock, sk);
2291         init_kcm_sock(kcm_sk(sk), mux);
2292
2293         return 0;
2294 }
2295
2296 static struct net_proto_family kcm_family_ops = {
2297         .family = PF_KCM,
2298         .create = kcm_create,
2299         .owner  = THIS_MODULE,
2300 };
2301
2302 static __net_init int kcm_init_net(struct net *net)
2303 {
2304         struct kcm_net *knet = net_generic(net, kcm_net_id);
2305
2306         INIT_LIST_HEAD_RCU(&knet->mux_list);
2307         mutex_init(&knet->mutex);
2308
2309         return 0;
2310 }
2311
2312 static __net_exit void kcm_exit_net(struct net *net)
2313 {
2314         struct kcm_net *knet = net_generic(net, kcm_net_id);
2315
2316         /* All KCM sockets should be closed at this point, which should mean
2317          * that all multiplexors and psocks have been destroyed.
2318          */
2319         WARN_ON(!list_empty(&knet->mux_list));
2320 }
2321
2322 static struct pernet_operations kcm_net_ops = {
2323         .init = kcm_init_net,
2324         .exit = kcm_exit_net,
2325         .id   = &kcm_net_id,
2326         .size = sizeof(struct kcm_net),
2327 };
2328
2329 static int __init kcm_init(void)
2330 {
2331         int err = -ENOMEM;
2332
2333         kcm_muxp = kmem_cache_create("kcm_mux_cache",
2334                                      sizeof(struct kcm_mux), 0,
2335                                      SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2336         if (!kcm_muxp)
2337                 goto fail;
2338
2339         kcm_psockp = kmem_cache_create("kcm_psock_cache",
2340                                        sizeof(struct kcm_psock), 0,
2341                                         SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2342         if (!kcm_psockp)
2343                 goto fail;
2344
2345         kcm_wq = create_singlethread_workqueue("kkcmd");
2346         if (!kcm_wq)
2347                 goto fail;
2348
2349         err = proto_register(&kcm_proto, 1);
2350         if (err)
2351                 goto fail;
2352
2353         err = sock_register(&kcm_family_ops);
2354         if (err)
2355                 goto sock_register_fail;
2356
2357         err = register_pernet_device(&kcm_net_ops);
2358         if (err)
2359                 goto net_ops_fail;
2360
2361         err = kcm_proc_init();
2362         if (err)
2363                 goto proc_init_fail;
2364
2365         return 0;
2366
2367 proc_init_fail:
2368         unregister_pernet_device(&kcm_net_ops);
2369
2370 net_ops_fail:
2371         sock_unregister(PF_KCM);
2372
2373 sock_register_fail:
2374         proto_unregister(&kcm_proto);
2375
2376 fail:
2377         kmem_cache_destroy(kcm_muxp);
2378         kmem_cache_destroy(kcm_psockp);
2379
2380         if (kcm_wq)
2381                 destroy_workqueue(kcm_wq);
2382
2383         return err;
2384 }
2385
2386 static void __exit kcm_exit(void)
2387 {
2388         kcm_proc_exit();
2389         unregister_pernet_device(&kcm_net_ops);
2390         sock_unregister(PF_KCM);
2391         proto_unregister(&kcm_proto);
2392         destroy_workqueue(kcm_wq);
2393
2394         kmem_cache_destroy(kcm_muxp);
2395         kmem_cache_destroy(kcm_psockp);
2396 }
2397
2398 module_init(kcm_init);
2399 module_exit(kcm_exit);
2400
2401 MODULE_LICENSE("GPL");
2402 MODULE_ALIAS_NETPROTO(PF_KCM);
2403