net/ipv4/tcp.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the  BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              Implementation of the Transmission Control Protocol(TCP).
   7  *
   8  * Authors:     Ross Biro
   9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
  11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
  12  *              Florian La Roche, <flla@stud.uni-sb.de>
  13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
  15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
  16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
  17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  18  *              Jorge Cwik, <jorge@laser.satlink.net>
  19  *
  20  * Fixes:
  21  *              Alan Cox        :       Numerous verify_area() calls
  22  *              Alan Cox        :       Set the ACK bit on a reset
  23  *              Alan Cox        :       Stopped it crashing if it closed while
  24  *                                      sk->inuse=1 and was trying to connect
  25  *                                      (tcp_err()).
  26  *              Alan Cox        :       All icmp error handling was broken
  27  *                                      pointers passed where wrong and the
  28  *                                      socket was looked up backwards. Nobody
  29  *                                      tested any icmp error code obviously.
  30  *              Alan Cox        :       tcp_err() now handled properly. It
  31  *                                      wakes people on errors. poll
  32  *                                      behaves and the icmp error race
  33  *                                      has gone by moving it into sock.c
  34  *              Alan Cox        :       tcp_send_reset() fixed to work for
  35  *                                      everything not just packets for
  36  *                                      unknown sockets.
  37  *              Alan Cox        :       tcp option processing.
  38  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
  39  *                                      syn rule wrong]
  40  *              Herp Rosmanith  :       More reset fixes
  41  *              Alan Cox        :       No longer acks invalid rst frames.
  42  *                                      Acking any kind of RST is right out.
  43  *              Alan Cox        :       Sets an ignore me flag on an rst
  44  *                                      receive otherwise odd bits of prattle
  45  *                                      escape still
  46  *              Alan Cox        :       Fixed another acking RST frame bug.
  47  *                                      Should stop LAN workplace lockups.
  48  *              Alan Cox        :       Some tidyups using the new skb list
  49  *                                      facilities
  50  *              Alan Cox        :       sk->keepopen now seems to work
  51  *              Alan Cox        :       Pulls options out correctly on accepts
  52  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
  53  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
  54  *                                      bit to skb ops.
  55  *              Alan Cox        :       Tidied tcp_data to avoid a potential
  56  *                                      nasty.
  57  *              Alan Cox        :       Added some better commenting, as the
  58  *                                      tcp is hard to follow
  59  *              Alan Cox        :       Removed incorrect check for 20 * psh
  60  *      Michael O'Reilly        :       ack < copied bug fix.
  61  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
  62  *              Alan Cox        :       FIN with no memory -> CRASH
  63  *              Alan Cox        :       Added socket option proto entries.
  64  *                                      Also added awareness of them to accept.
  65  *              Alan Cox        :       Added TCP options (SOL_TCP)
  66  *              Alan Cox        :       Switched wakeup calls to callbacks,
  67  *                                      so the kernel can layer network
  68  *                                      sockets.
  69  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
  70  *              Alan Cox        :       Handle FIN (more) properly (we hope).
  71  *              Alan Cox        :       RST frames sent on unsynchronised
  72  *                                      state ack error.
  73  *              Alan Cox        :       Put in missing check for SYN bit.
  74  *              Alan Cox        :       Added tcp_select_window() aka NET2E
  75  *                                      window non shrink trick.
  76  *              Alan Cox        :       Added a couple of small NET2E timer
  77  *                                      fixes
  78  *              Charles Hedrick :       TCP fixes
  79  *              Toomas Tamm     :       TCP window fixes
  80  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
  81  *              Charles Hedrick :       Rewrote most of it to actually work
  82  *              Linus           :       Rewrote tcp_read() and URG handling
  83  *                                      completely
  84  *              Gerhard Koerting:       Fixed some missing timer handling
  85  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
  86  *              Gerhard Koerting:       PC/TCP workarounds
  87  *              Adam Caldwell   :       Assorted timer/timing errors
  88  *              Matthew Dillon  :       Fixed another RST bug
  89  *              Alan Cox        :       Move to kernel side addressing changes.
  90  *              Alan Cox        :       Beginning work on TCP fastpathing
  91  *                                      (not yet usable)
  92  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
  93  *              Alan Cox        :       TCP fast path debugging
  94  *              Alan Cox        :       Window clamping
  95  *              Michael Riepe   :       Bug in tcp_check()
  96  *              Matt Dillon     :       More TCP improvements and RST bug fixes
  97  *              Matt Dillon     :       Yet more small nasties remove from the
  98  *                                      TCP code (Be very nice to this man if
  99  *                                      tcp finally works 100%) 8)
 100  *              Alan Cox        :       BSD accept semantics.
 101  *              Alan Cox        :       Reset on closedown bug.
 102  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
 103  *              Michael Pall    :       Handle poll() after URG properly in
 104  *                                      all cases.
 105  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
 106  *                                      (multi URG PUSH broke rlogin).
 107  *              Michael Pall    :       Fix the multi URG PUSH problem in
 108  *                                      tcp_readable(), poll() after URG
 109  *                                      works now.
 110  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
 111  *                                      BSD api.
 112  *              Alan Cox        :       Changed the semantics of sk->socket to
 113  *                                      fix a race and a signal problem with
 114  *                                      accept() and async I/O.
 115  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
 116  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
 117  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
 118  *                                      clients/servers which listen in on
 119  *                                      fixed ports.
 120  *              Alan Cox        :       Cleaned the above up and shrank it to
 121  *                                      a sensible code size.
 122  *              Alan Cox        :       Self connect lockup fix.
 123  *              Alan Cox        :       No connect to multicast.
 124  *              Ross Biro       :       Close unaccepted children on master
 125  *                                      socket close.
 126  *              Alan Cox        :       Reset tracing code.
 127  *              Alan Cox        :       Spurious resets on shutdown.
 128  *              Alan Cox        :       Giant 15 minute/60 second timer error
 129  *              Alan Cox        :       Small whoops in polling before an
 130  *                                      accept.
 131  *              Alan Cox        :       Kept the state trace facility since
 132  *                                      it's handy for debugging.
 133  *              Alan Cox        :       More reset handler fixes.
 134  *              Alan Cox        :       Started rewriting the code based on
 135  *                                      the RFC's for other useful protocol
 136  *                                      references see: Comer, KA9Q NOS, and
 137  *                                      for a reference on the difference
 138  *                                      between specifications and how BSD
 139  *                                      works see the 4.4lite source.
 140  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
 141  *                                      close.
 142  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
 143  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
 144  *              Alan Cox        :       Reimplemented timers as per the RFC
 145  *                                      and using multiple timers for sanity.
 146  *              Alan Cox        :       Small bug fixes, and a lot of new
 147  *                                      comments.
 148  *              Alan Cox        :       Fixed dual reader crash by locking
 149  *                                      the buffers (much like datagram.c)
 150  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
 151  *                                      now gets fed up of retrying without
 152  *                                      (even a no space) answer.
 153  *              Alan Cox        :       Extracted closing code better
 154  *              Alan Cox        :       Fixed the closing state machine to
 155  *                                      resemble the RFC.
 156  *              Alan Cox        :       More 'per spec' fixes.
 157  *              Jorge Cwik      :       Even faster checksumming.
 158  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
 159  *                                      only frames. At least one pc tcp stack
 160  *                                      generates them.
 161  *              Alan Cox        :       Cache last socket.
 162  *              Alan Cox        :       Per route irtt.
 163  *              Matt Day        :       poll()->select() match BSD precisely on error
 164  *              Alan Cox        :       New buffers
 165  *              Marc Tamsky     :       Various sk->prot->retransmits and
 166  *                                      sk->retransmits misupdating fixed.
 167  *                                      Fixed tcp_write_timeout: stuck close,
 168  *                                      and TCP syn retries gets used now.
 169  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
 170  *                                      ack if state is TCP_CLOSED.
 171  *              Alan Cox        :       Look up device on a retransmit - routes may
 172  *                                      change. Doesn't yet cope with MSS shrink right
 173  *                                      but it's a start!
 174  *              Marc Tamsky     :       Closing in closing fixes.
 175  *              Mike Shaver     :       RFC1122 verifications.
 176  *              Alan Cox        :       rcv_saddr errors.
 177  *              Alan Cox        :       Block double connect().
 178  *              Alan Cox        :       Small hooks for enSKIP.
 179  *              Alexey Kuznetsov:       Path MTU discovery.
 180  *              Alan Cox        :       Support soft errors.
 181  *              Alan Cox        :       Fix MTU discovery pathological case
 182  *                                      when the remote claims no mtu!
 183  *              Marc Tamsky     :       TCP_CLOSE fix.
 184  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
 185  *                                      window but wrong (fixes NT lpd problems)
 186  *              Pedro Roque     :       Better TCP window handling, delayed ack.
 187  *              Joerg Reuter    :       No modification of locked buffers in
 188  *                                      tcp_do_retransmit()
 189  *              Eric Schenk     :       Changed receiver side silly window
 190  *                                      avoidance algorithm to BSD style
 191  *                                      algorithm. This doubles throughput
 192  *                                      against machines running Solaris,
 193  *                                      and seems to result in general
 194  *                                      improvement.
 195  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
 196  *      Willy Konynenberg       :       Transparent proxying support.
 197  *      Mike McLagan            :       Routing by source
 198  *              Keith Owens     :       Do proper merging with partial SKB's in
 199  *                                      tcp_do_sendmsg to avoid burstiness.
 200  *              Eric Schenk     :       Fix fast close down bug with
 201  *                                      shutdown() followed by close().
 202  *              Andi Kleen      :       Make poll agree with SIGIO
 203  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
 204  *                                      lingertime == 0 (RFC 793 ABORT Call)
 205  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
 206  *                                      csum_and_copy_from_user() if possible.
 207  *
 208  *              This program is free software; you can redistribute it and/or
 209  *              modify it under the terms of the GNU General Public License
 210  *              as published by the Free Software Foundation; either version
 211  *              2 of the License, or(at your option) any later version.
 212  *
 213  * Description of States:
 214  *
 215  *      TCP_SYN_SENT            sent a connection request, waiting for ack
 216  *
 217  *      TCP_SYN_RECV            received a connection request, sent ack,
 218  *                              waiting for final ack in three-way handshake.
 219  *
 220  *      TCP_ESTABLISHED         connection established
 221  *
 222  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
 223  *                              transmission of remaining buffered data
 224  *
 225  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
 226  *                              to shutdown
 227  *
 228  *      TCP_CLOSING             both sides have shutdown but we still have
 229  *                              data we have to finish sending
 230  *
 231  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
 232  *                              closed, can only be entered from FIN_WAIT2
 233  *                              or CLOSING.  Required because the other end
 234  *                              may not have gotten our last ACK causing it
 235  *                              to retransmit the data packet (which we ignore)
 236  *
 237  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
 238  *                              us to finish writing our data and to shutdown
 239  *                              (we have to close() to move on to LAST_ACK)
 240  *
 241  *      TCP_LAST_ACK            out side has shutdown after remote has
 242  *                              shutdown.  There may still be data in our
 243  *                              buffer that we have to finish sending
 244  *
 245  *      TCP_CLOSE               socket is finished
 246  */
 247
 248 #include <linux/kernel.h>
 249 #include <linux/module.h>
 250 #include <linux/types.h>
 251 #include <linux/fcntl.h>
 252 #include <linux/poll.h>
 253 #include <linux/init.h>
 254 #include <linux/fs.h>
 255 #include <linux/skbuff.h>
 256 #include <linux/scatterlist.h>
 257 #include <linux/splice.h>
 258 #include <linux/net.h>
 259 #include <linux/socket.h>
 260 #include <linux/random.h>
 261 #include <linux/bootmem.h>
 262 #include <linux/highmem.h>
 263 #include <linux/swap.h>
 264 #include <linux/cache.h>
 265 #include <linux/err.h>
 266 #include <linux/crypto.h>
 267
 268 #include <net/icmp.h>
 269 #include <net/tcp.h>
 270 #include <net/xfrm.h>
 271 #include <net/ip.h>
 272 #include <net/netdma.h>
 273 #include <net/sock.h>
 274
 275 #include <asm/uaccess.h>
 276 #include <asm/ioctls.h>
 277
 278 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
 279
 280 atomic_t tcp_orphan_count = ATOMIC_INIT(0);
 281
 282 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 283
 284 int sysctl_tcp_mem[3] __read_mostly;
 285 int sysctl_tcp_wmem[3] __read_mostly;
 286 int sysctl_tcp_rmem[3] __read_mostly;
 287
 288 EXPORT_SYMBOL(sysctl_tcp_mem);
 289 EXPORT_SYMBOL(sysctl_tcp_rmem);
 290 EXPORT_SYMBOL(sysctl_tcp_wmem);
 291
 292 atomic_t tcp_memory_allocated;  /* Current allocated memory. */
 293 atomic_t tcp_sockets_allocated; /* Current number of TCP sockets. */
 294
 295 EXPORT_SYMBOL(tcp_memory_allocated);
 296 EXPORT_SYMBOL(tcp_sockets_allocated);
 297
 298 /*
 299  * TCP splice context
 300  */
 301 struct tcp_splice_state {
 302         struct pipe_inode_info *pipe;
 303         size_t len;
 304         unsigned int flags;
 305 };
 306
 307 /*
 308  * Pressure flag: try to collapse.
 309  * Technical note: it is used by multiple contexts non atomically.
 310  * All the __sk_mem_schedule() is of this nature: accounting
 311  * is strict, actions are advisory and have some latency.
 312  */
 313 int tcp_memory_pressure __read_mostly;
 314
 315 EXPORT_SYMBOL(tcp_memory_pressure);
 316
 317 void tcp_enter_memory_pressure(struct sock *sk)
 318 {
 319         if (!tcp_memory_pressure) {
 320                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
 321                 tcp_memory_pressure = 1;
 322         }
 323 }
 324
 325 EXPORT_SYMBOL(tcp_enter_memory_pressure);
 326
 327 /*
 328  *      Wait for a TCP event.
 329  *
 330  *      Note that we don't need to lock the socket, as the upper poll layers
 331  *      take care of normal races (between the test and the event) and we don't
 332  *      go look at any of the socket buffers directly.
 333  */
 334 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 335 {
 336         unsigned int mask;
 337         struct sock *sk = sock->sk;
 338         struct tcp_sock *tp = tcp_sk(sk);
 339
 340         poll_wait(file, sk->sk_sleep, wait);
 341         if (sk->sk_state == TCP_LISTEN)
 342                 return inet_csk_listen_poll(sk);
 343
 344         /* Socket is not locked. We are protected from async events
 345          * by poll logic and correct handling of state changes
 346          * made by other threads is impossible in any case.
 347          */
 348
 349         mask = 0;
 350         if (sk->sk_err)
 351                 mask = POLLERR;
 352
 353         /*
 354          * POLLHUP is certainly not done right. But poll() doesn't
 355          * have a notion of HUP in just one direction, and for a
 356          * socket the read side is more interesting.
 357          *
 358          * Some poll() documentation says that POLLHUP is incompatible
 359          * with the POLLOUT/POLLWR flags, so somebody should check this
 360          * all. But careful, it tends to be safer to return too many
 361          * bits than too few, and you can easily break real applications
 362          * if you don't tell them that something has hung up!
 363          *
 364          * Check-me.
 365          *
 366          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
 367          * our fs/select.c). It means that after we received EOF,
 368          * poll always returns immediately, making impossible poll() on write()
 369          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
 370          * if and only if shutdown has been made in both directions.
 371          * Actually, it is interesting to look how Solaris and DUX
 372          * solve this dilemma. I would prefer, if POLLHUP were maskable,
 373          * then we could set it on SND_SHUTDOWN. BTW examples given
 374          * in Stevens' books assume exactly this behaviour, it explains
 375          * why POLLHUP is incompatible with POLLOUT.    --ANK
 376          *
 377          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 378          * blocking on fresh not-connected or disconnected socket. --ANK
 379          */
 380         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
 381                 mask |= POLLHUP;
 382         if (sk->sk_shutdown & RCV_SHUTDOWN)
 383                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
 384
 385         /* Connected? */
 386         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
 387                 /* Potential race condition. If read of tp below will
 388                  * escape above sk->sk_state, we can be illegally awaken
 389                  * in SYN_* states. */
 390                 if ((tp->rcv_nxt != tp->copied_seq) &&
 391                     (tp->urg_seq != tp->copied_seq ||
 392                      tp->rcv_nxt != tp->copied_seq + 1 ||
 393                      sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data))
 394                         mask |= POLLIN | POLLRDNORM;
 395
 396                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 397                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
 398                                 mask |= POLLOUT | POLLWRNORM;
 399                         } else {  /* send SIGIO later */
 400                                 set_bit(SOCK_ASYNC_NOSPACE,
 401                                         &sk->sk_socket->flags);
 402                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 403
 404                                 /* Race breaker. If space is freed after
 405                                  * wspace test but before the flags are set,
 406                                  * IO signal will be lost.
 407                                  */
 408                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
 409                                         mask |= POLLOUT | POLLWRNORM;
 410                         }
 411                 }
 412
 413                 if (tp->urg_data & TCP_URG_VALID)
 414                         mask |= POLLPRI;
 415         }
 416         return mask;
 417 }
 418
 419 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 420 {
 421         struct tcp_sock *tp = tcp_sk(sk);
 422         int answ;
 423
 424         switch (cmd) {
 425         case SIOCINQ:
 426                 if (sk->sk_state == TCP_LISTEN)
 427                         return -EINVAL;
 428
 429                 lock_sock(sk);
 430                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 431                         answ = 0;
 432                 else if (sock_flag(sk, SOCK_URGINLINE) ||
 433                          !tp->urg_data ||
 434                          before(tp->urg_seq, tp->copied_seq) ||
 435                          !before(tp->urg_seq, tp->rcv_nxt)) {
 436                         answ = tp->rcv_nxt - tp->copied_seq;
 437
 438                         /* Subtract 1, if FIN is in queue. */
 439                         if (answ && !skb_queue_empty(&sk->sk_receive_queue))
 440                                 answ -=
 441                        tcp_hdr((struct sk_buff *)sk->sk_receive_queue.prev)->fin;
 442                 } else
 443                         answ = tp->urg_seq - tp->copied_seq;
 444                 release_sock(sk);
 445                 break;
 446         case SIOCATMARK:
 447                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 448                 break;
 449         case SIOCOUTQ:
 450                 if (sk->sk_state == TCP_LISTEN)
 451                         return -EINVAL;
 452
 453                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 454                         answ = 0;
 455                 else
 456                         answ = tp->write_seq - tp->snd_una;
 457                 break;
 458         default:
 459                 return -ENOIOCTLCMD;
 460         }
 461
 462         return put_user(answ, (int __user *)arg);
 463 }
 464
 465 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 466 {
 467         TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
 468         tp->pushed_seq = tp->write_seq;
 469 }
 470
 471 static inline int forced_push(struct tcp_sock *tp)
 472 {
 473         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 474 }
 475
 476 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
 477 {
 478         struct tcp_sock *tp = tcp_sk(sk);
 479         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 480
 481         skb->csum    = 0;
 482         tcb->seq     = tcb->end_seq = tp->write_seq;
 483         tcb->flags   = TCPCB_FLAG_ACK;
 484         tcb->sacked  = 0;
 485         skb_header_release(skb);
 486         tcp_add_write_queue_tail(sk, skb);
 487         sk->sk_wmem_queued += skb->truesize;
 488         sk_mem_charge(sk, skb->truesize);
 489         if (tp->nonagle & TCP_NAGLE_PUSH)
 490                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 491 }
 492
 493 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
 494                                 struct sk_buff *skb)
 495 {
 496         if (flags & MSG_OOB) {
 497                 tp->urg_mode = 1;
 498                 tp->snd_up = tp->write_seq;
 499         }
 500 }
 501
 502 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
 503                             int nonagle)
 504 {
 505         struct tcp_sock *tp = tcp_sk(sk);
 506
 507         if (tcp_send_head(sk)) {
 508                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 509                 if (!(flags & MSG_MORE) || forced_push(tp))
 510                         tcp_mark_push(tp, skb);
 511                 tcp_mark_urg(tp, flags, skb);
 512                 __tcp_push_pending_frames(sk, mss_now,
 513                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
 514         }
 515 }
 516
 517 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 518                                 unsigned int offset, size_t len)
 519 {
 520         struct tcp_splice_state *tss = rd_desc->arg.data;
 521         int ret;
 522
 523         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
 524                               tss->flags);
 525         if (ret > 0)
 526                 rd_desc->count -= ret;
 527         return ret;
 528 }
 529
 530 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 531 {
 532         /* Store TCP splice context information in read_descriptor_t. */
 533         read_descriptor_t rd_desc = {
 534                 .arg.data = tss,
 535                 .count    = tss->len,
 536         };
 537
 538         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 539 }
 540
 541 /**
 542  *  tcp_splice_read - splice data from TCP socket to a pipe
 543  * @sock:       socket to splice from
 544  * @ppos:       position (not valid)
 545  * @pipe:       pipe to splice to
 546  * @len:        number of bytes to splice
 547  * @flags:      splice modifier flags
 548  *
 549  * Description:
 550  *    Will read pages from given socket and fill them into a pipe.
 551  *
 552  **/
 553 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 554                         struct pipe_inode_info *pipe, size_t len,
 555                         unsigned int flags)
 556 {
 557         struct sock *sk = sock->sk;
 558         struct tcp_splice_state tss = {
 559                 .pipe = pipe,
 560                 .len = len,
 561                 .flags = flags,
 562         };
 563         long timeo;
 564         ssize_t spliced;
 565         int ret;
 566
 567         /*
 568          * We can't seek on a socket input
 569          */
 570         if (unlikely(*ppos))
 571                 return -ESPIPE;
 572
 573         ret = spliced = 0;
 574
 575         lock_sock(sk);
 576
 577         timeo = sock_rcvtimeo(sk, flags & SPLICE_F_NONBLOCK);
 578         while (tss.len) {
 579                 ret = __tcp_splice_read(sk, &tss);
 580                 if (ret < 0)
 581                         break;
 582                 else if (!ret) {
 583                         if (spliced)
 584                                 break;
 585                         if (sock_flag(sk, SOCK_DONE))
 586                                 break;
 587                         if (sk->sk_err) {
 588                                 ret = sock_error(sk);
 589                                 break;
 590                         }
 591                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 592                                 break;
 593                         if (sk->sk_state == TCP_CLOSE) {
 594                                 /*
 595                                  * This occurs when user tries to read
 596                                  * from never connected socket.
 597                                  */
 598                                 if (!sock_flag(sk, SOCK_DONE))
 599                                         ret = -ENOTCONN;
 600                                 break;
 601                         }
 602                         if (flags & SPLICE_F_NONBLOCK) {
 603                                 ret = -EAGAIN;
 604                                 break;
 605                         }
 606                         if (!timeo) {
 607                                 ret = -EAGAIN;
 608                                 break;
 609                         }
 610                         sk_wait_data(sk, &timeo);
 611                         if (signal_pending(current)) {
 612                                 ret = sock_intr_errno(timeo);
 613                                 break;
 614                         }
 615                         continue;
 616                 }
 617                 tss.len -= ret;
 618                 spliced += ret;
 619
 620                 if (!timeo)
 621                         break;
 622                 release_sock(sk);
 623                 lock_sock(sk);
 624
 625                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 626                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 627                     signal_pending(current))
 628                         break;
 629         }
 630
 631         release_sock(sk);
 632
 633         if (spliced)
 634                 return spliced;
 635
 636         return ret;
 637 }
 638
 639 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
 640 {
 641         struct sk_buff *skb;
 642
 643         /* The TCP header must be at least 32-bit aligned.  */
 644         size = ALIGN(size, 4);
 645
 646         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 647         if (skb) {
 648                 if (sk_wmem_schedule(sk, skb->truesize)) {
 649                         /*
 650                          * Make sure that we have exactly size bytes
 651                          * available to the caller, no more, no less.
 652                          */
 653                         skb_reserve(skb, skb_tailroom(skb) - size);
 654                         return skb;
 655                 }
 656                 __kfree_skb(skb);
 657         } else {
 658                 sk->sk_prot->enter_memory_pressure(sk);
 659                 sk_stream_moderate_sndbuf(sk);
 660         }
 661         return NULL;
 662 }
 663
 664 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
 665                          size_t psize, int flags)
 666 {
 667         struct tcp_sock *tp = tcp_sk(sk);
 668         int mss_now, size_goal;
 669         int err;
 670         ssize_t copied;
 671         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 672
 673         /* Wait for a connection to finish. */
 674         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 675                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 676                         goto out_err;
 677
 678         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 679
 680         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 681         size_goal = tp->xmit_size_goal;
 682         copied = 0;
 683
 684         err = -EPIPE;
 685         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 686                 goto do_error;
 687
 688         while (psize > 0) {
 689                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 690                 struct page *page = pages[poffset / PAGE_SIZE];
 691                 int copy, i, can_coalesce;
 692                 int offset = poffset % PAGE_SIZE;
 693                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
 694
 695                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
 696 new_segment:
 697                         if (!sk_stream_memory_free(sk))
 698                                 goto wait_for_sndbuf;
 699
 700                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
 701                         if (!skb)
 702                                 goto wait_for_memory;
 703
 704                         skb_entail(sk, skb);
 705                         copy = size_goal;
 706                 }
 707
 708                 if (copy > size)
 709                         copy = size;
 710
 711                 i = skb_shinfo(skb)->nr_frags;
 712                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 713                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
 714                         tcp_mark_push(tp, skb);
 715                         goto new_segment;
 716                 }
 717                 if (!sk_wmem_schedule(sk, copy))
 718                         goto wait_for_memory;
 719
 720                 if (can_coalesce) {
 721                         skb_shinfo(skb)->frags[i - 1].size += copy;
 722                 } else {
 723                         get_page(page);
 724                         skb_fill_page_desc(skb, i, page, offset, copy);
 725                 }
 726
 727                 skb->len += copy;
 728                 skb->data_len += copy;
 729                 skb->truesize += copy;
 730                 sk->sk_wmem_queued += copy;
 731                 sk_mem_charge(sk, copy);
 732                 skb->ip_summed = CHECKSUM_PARTIAL;
 733                 tp->write_seq += copy;
 734                 TCP_SKB_CB(skb)->end_seq += copy;
 735                 skb_shinfo(skb)->gso_segs = 0;
 736
 737                 if (!copied)
 738                         TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
 739
 740                 copied += copy;
 741                 poffset += copy;
 742                 if (!(psize -= copy))
 743                         goto out;
 744
 745                 if (skb->len < size_goal || (flags & MSG_OOB))
 746                         continue;
 747
 748                 if (forced_push(tp)) {
 749                         tcp_mark_push(tp, skb);
 750                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
 751                 } else if (skb == tcp_send_head(sk))
 752                         tcp_push_one(sk, mss_now);
 753                 continue;
 754
 755 wait_for_sndbuf:
 756                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 757 wait_for_memory:
 758                 if (copied)
 759                         tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
 760
 761                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 762                         goto do_error;
 763
 764                 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 765                 size_goal = tp->xmit_size_goal;
 766         }
 767
 768 out:
 769         if (copied)
 770                 tcp_push(sk, flags, mss_now, tp->nonagle);
 771         return copied;
 772
 773 do_error:
 774         if (copied)
 775                 goto out;
 776 out_err:
 777         return sk_stream_error(sk, flags, err);
 778 }
 779
 780 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
 781                      size_t size, int flags)
 782 {
 783         ssize_t res;
 784         struct sock *sk = sock->sk;
 785
 786         if (!(sk->sk_route_caps & NETIF_F_SG) ||
 787             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
 788                 return sock_no_sendpage(sock, page, offset, size, flags);
 789
 790         lock_sock(sk);
 791         TCP_CHECK_TIMER(sk);
 792         res = do_tcp_sendpages(sk, &page, offset, size, flags);
 793         TCP_CHECK_TIMER(sk);
 794         release_sock(sk);
 795         return res;
 796 }
 797
 798 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
 799 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
 800
 801 static inline int select_size(struct sock *sk)
 802 {
 803         struct tcp_sock *tp = tcp_sk(sk);
 804         int tmp = tp->mss_cache;
 805
 806         if (sk->sk_route_caps & NETIF_F_SG) {
 807                 if (sk_can_gso(sk))
 808                         tmp = 0;
 809                 else {
 810                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
 811
 812                         if (tmp >= pgbreak &&
 813                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
 814                                 tmp = pgbreak;
 815                 }
 816         }
 817
 818         return tmp;
 819 }
 820
 821 int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
 822                 size_t size)
 823 {
 824         struct sock *sk = sock->sk;
 825         struct iovec *iov;
 826         struct tcp_sock *tp = tcp_sk(sk);
 827         struct sk_buff *skb;
 828         int iovlen, flags;
 829         int mss_now, size_goal;
 830         int err, copied;
 831         long timeo;
 832
 833         lock_sock(sk);
 834         TCP_CHECK_TIMER(sk);
 835
 836         flags = msg->msg_flags;
 837         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 838
 839         /* Wait for a connection to finish. */
 840         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 841                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 842                         goto out_err;
 843
 844         /* This should be in poll */
 845         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 846
 847         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 848         size_goal = tp->xmit_size_goal;
 849
 850         /* Ok commence sending. */
 851         iovlen = msg->msg_iovlen;
 852         iov = msg->msg_iov;
 853         copied = 0;
 854
 855         err = -EPIPE;
 856         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 857                 goto do_error;
 858
 859         while (--iovlen >= 0) {
 860                 int seglen = iov->iov_len;
 861                 unsigned char __user *from = iov->iov_base;
 862
 863                 iov++;
 864
 865                 while (seglen > 0) {
 866                         int copy;
 867
 868                         skb = tcp_write_queue_tail(sk);
 869
 870                         if (!tcp_send_head(sk) ||
 871                             (copy = size_goal - skb->len) <= 0) {
 872
 873 new_segment:
 874                                 /* Allocate new segment. If the interface is SG,
 875                                  * allocate skb fitting to single page.
 876                                  */
 877                                 if (!sk_stream_memory_free(sk))
 878                                         goto wait_for_sndbuf;
 879
 880                                 skb = sk_stream_alloc_skb(sk, select_size(sk),
 881                                                 sk->sk_allocation);
 882                                 if (!skb)
 883                                         goto wait_for_memory;
 884
 885                                 /*
 886                                  * Check whether we can use HW checksum.
 887                                  */
 888                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
 889                                         skb->ip_summed = CHECKSUM_PARTIAL;
 890
 891                                 skb_entail(sk, skb);
 892                                 copy = size_goal;
 893                         }
 894
 895                         /* Try to append data to the end of skb. */
 896                         if (copy > seglen)
 897                                 copy = seglen;
 898
 899                         /* Where to copy to? */
 900                         if (skb_tailroom(skb) > 0) {
 901                                 /* We have some space in skb head. Superb! */
 902                                 if (copy > skb_tailroom(skb))
 903                                         copy = skb_tailroom(skb);
 904                                 if ((err = skb_add_data(skb, from, copy)) != 0)
 905                                         goto do_fault;
 906                         } else {
 907                                 int merge = 0;
 908                                 int i = skb_shinfo(skb)->nr_frags;
 909                                 struct page *page = TCP_PAGE(sk);
 910                                 int off = TCP_OFF(sk);
 911
 912                                 if (skb_can_coalesce(skb, i, page, off) &&
 913                                     off != PAGE_SIZE) {
 914                                         /* We can extend the last page
 915                                          * fragment. */
 916                                         merge = 1;
 917                                 } else if (i == MAX_SKB_FRAGS ||
 918                                            (!i &&
 919                                            !(sk->sk_route_caps & NETIF_F_SG))) {
 920                                         /* Need to add new fragment and cannot
 921                                          * do this because interface is non-SG,
 922                                          * or because all the page slots are
 923                                          * busy. */
 924                                         tcp_mark_push(tp, skb);
 925                                         goto new_segment;
 926                                 } else if (page) {
 927                                         if (off == PAGE_SIZE) {
 928                                                 put_page(page);
 929                                                 TCP_PAGE(sk) = page = NULL;
 930                                                 off = 0;
 931                                         }
 932                                 } else
 933                                         off = 0;
 934
 935                                 if (copy > PAGE_SIZE - off)
 936                                         copy = PAGE_SIZE - off;
 937
 938                                 if (!sk_wmem_schedule(sk, copy))
 939                                         goto wait_for_memory;
 940
 941                                 if (!page) {
 942                                         /* Allocate new cache page. */
 943                                         if (!(page = sk_stream_alloc_page(sk)))
 944                                                 goto wait_for_memory;
 945                                 }
 946
 947                                 /* Time to copy data. We are close to
 948                                  * the end! */
 949                                 err = skb_copy_to_page(sk, from, skb, page,
 950                                                        off, copy);
 951                                 if (err) {
 952                                         /* If this page was new, give it to the
 953                                          * socket so it does not get leaked.
 954                                          */
 955                                         if (!TCP_PAGE(sk)) {
 956                                                 TCP_PAGE(sk) = page;
 957                                                 TCP_OFF(sk) = 0;
 958                                         }
 959                                         goto do_error;
 960                                 }
 961
 962                                 /* Update the skb. */
 963                                 if (merge) {
 964                                         skb_shinfo(skb)->frags[i - 1].size +=
 965                                                                         copy;
 966                                 } else {
 967                                         skb_fill_page_desc(skb, i, page, off, copy);
 968                                         if (TCP_PAGE(sk)) {
 969                                                 get_page(page);
 970                                         } else if (off + copy < PAGE_SIZE) {
 971                                                 get_page(page);
 972                                                 TCP_PAGE(sk) = page;
 973                                         }
 974                                 }
 975
 976                                 TCP_OFF(sk) = off + copy;
 977                         }
 978
 979                         if (!copied)
 980                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
 981
 982                         tp->write_seq += copy;
 983                         TCP_SKB_CB(skb)->end_seq += copy;
 984                         skb_shinfo(skb)->gso_segs = 0;
 985
 986                         from += copy;
 987                         copied += copy;
 988                         if ((seglen -= copy) == 0 && iovlen == 0)
 989                                 goto out;
 990
 991                         if (skb->len < size_goal || (flags & MSG_OOB))
 992                                 continue;
 993
 994                         if (forced_push(tp)) {
 995                                 tcp_mark_push(tp, skb);
 996                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
 997                         } else if (skb == tcp_send_head(sk))
 998                                 tcp_push_one(sk, mss_now);
 999                         continue;
1000
1001 wait_for_sndbuf:
1002                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1003 wait_for_memory:
1004                         if (copied)
1005                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1006
1007                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1008                                 goto do_error;
1009
1010                         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
1011                         size_goal = tp->xmit_size_goal;
1012                 }
1013         }
1014
1015 out:
1016         if (copied)
1017                 tcp_push(sk, flags, mss_now, tp->nonagle);
1018         TCP_CHECK_TIMER(sk);
1019         release_sock(sk);
1020         return copied;
1021
1022 do_fault:
1023         if (!skb->len) {
1024                 tcp_unlink_write_queue(skb, sk);
1025                 /* It is the one place in all of TCP, except connection
1026                  * reset, where we can be unlinking the send_head.
1027                  */
1028                 tcp_check_send_head(sk, skb);
1029                 sk_wmem_free_skb(sk, skb);
1030         }
1031
1032 do_error:
1033         if (copied)
1034                 goto out;
1035 out_err:
1036         err = sk_stream_error(sk, flags, err);
1037         TCP_CHECK_TIMER(sk);
1038         release_sock(sk);
1039         return err;
1040 }
1041
1042 /*
1043  *      Handle reading urgent data. BSD has very simple semantics for
1044  *      this, no blocking and very strange errors 8)
1045  */
1046
1047 static int tcp_recv_urg(struct sock *sk, long timeo,
1048                         struct msghdr *msg, int len, int flags,
1049                         int *addr_len)
1050 {
1051         struct tcp_sock *tp = tcp_sk(sk);
1052
1053         /* No URG data to read. */
1054         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1055             tp->urg_data == TCP_URG_READ)
1056                 return -EINVAL; /* Yes this is right ! */
1057
1058         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1059                 return -ENOTCONN;
1060
1061         if (tp->urg_data & TCP_URG_VALID) {
1062                 int err = 0;
1063                 char c = tp->urg_data;
1064
1065                 if (!(flags & MSG_PEEK))
1066                         tp->urg_data = TCP_URG_READ;
1067
1068                 /* Read urgent data. */
1069                 msg->msg_flags |= MSG_OOB;
1070
1071                 if (len > 0) {
1072                         if (!(flags & MSG_TRUNC))
1073                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1074                         len = 1;
1075                 } else
1076                         msg->msg_flags |= MSG_TRUNC;
1077
1078                 return err ? -EFAULT : len;
1079         }
1080
1081         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1082                 return 0;
1083
1084         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1085          * the available implementations agree in this case:
1086          * this call should never block, independent of the
1087          * blocking state of the socket.
1088          * Mike <pall@rz.uni-karlsruhe.de>
1089          */
1090         return -EAGAIN;
1091 }
1092
1093 /* Clean up the receive buffer for full frames taken by the user,
1094  * then send an ACK if necessary.  COPIED is the number of bytes
1095  * tcp_recvmsg has given to the user so far, it speeds up the
1096  * calculation of whether or not we must ACK for the sake of
1097  * a window update.
1098  */
1099 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1100 {
1101         struct tcp_sock *tp = tcp_sk(sk);
1102         int time_to_ack = 0;
1103
1104 #if TCP_DEBUG
1105         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1106
1107         WARN_ON(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
1108 #endif
1109
1110         if (inet_csk_ack_scheduled(sk)) {
1111                 const struct inet_connection_sock *icsk = inet_csk(sk);
1112                    /* Delayed ACKs frequently hit locked sockets during bulk
1113                     * receive. */
1114                 if (icsk->icsk_ack.blocked ||
1115                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1116                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1117                     /*
1118                      * If this read emptied read buffer, we send ACK, if
1119                      * connection is not bidirectional, user drained
1120                      * receive buffer and there was a small segment
1121                      * in queue.
1122                      */
1123                     (copied > 0 &&
1124                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1125                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1126                        !icsk->icsk_ack.pingpong)) &&
1127                       !atomic_read(&sk->sk_rmem_alloc)))
1128                         time_to_ack = 1;
1129         }
1130
1131         /* We send an ACK if we can now advertise a non-zero window
1132          * which has been raised "significantly".
1133          *
1134          * Even if window raised up to infinity, do not send window open ACK
1135          * in states, where we will not receive more. It is useless.
1136          */
1137         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1138                 __u32 rcv_window_now = tcp_receive_window(tp);
1139
1140                 /* Optimize, __tcp_select_window() is not cheap. */
1141                 if (2*rcv_window_now <= tp->window_clamp) {
1142                         __u32 new_window = __tcp_select_window(sk);
1143
1144                         /* Send ACK now, if this read freed lots of space
1145                          * in our buffer. Certainly, new_window is new window.
1146                          * We can advertise it now, if it is not less than current one.
1147                          * "Lots" means "at least twice" here.
1148                          */
1149                         if (new_window && new_window >= 2 * rcv_window_now)
1150                                 time_to_ack = 1;
1151                 }
1152         }
1153         if (time_to_ack)
1154                 tcp_send_ack(sk);
1155 }
1156
1157 static void tcp_prequeue_process(struct sock *sk)
1158 {
1159         struct sk_buff *skb;
1160         struct tcp_sock *tp = tcp_sk(sk);
1161
1162         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1163
1164         /* RX process wants to run with disabled BHs, though it is not
1165          * necessary */
1166         local_bh_disable();
1167         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1168                 sk->sk_backlog_rcv(sk, skb);
1169         local_bh_enable();
1170
1171         /* Clear memory counter. */
1172         tp->ucopy.memory = 0;
1173 }
1174
1175 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1176 {
1177         struct sk_buff *skb;
1178         u32 offset;
1179
1180         skb_queue_walk(&sk->sk_receive_queue, skb) {
1181                 offset = seq - TCP_SKB_CB(skb)->seq;
1182                 if (tcp_hdr(skb)->syn)
1183                         offset--;
1184                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1185                         *off = offset;
1186                         return skb;
1187                 }
1188         }
1189         return NULL;
1190 }
1191
1192 /*
1193  * This routine provides an alternative to tcp_recvmsg() for routines
1194  * that would like to handle copying from skbuffs directly in 'sendfile'
1195  * fashion.
1196  * Note:
1197  *      - It is assumed that the socket was locked by the caller.
1198  *      - The routine does not block.
1199  *      - At present, there is no support for reading OOB data
1200  *        or for 'peeking' the socket using this routine
1201  *        (although both would be easy to implement).
1202  */
1203 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1204                   sk_read_actor_t recv_actor)
1205 {
1206         struct sk_buff *skb;
1207         struct tcp_sock *tp = tcp_sk(sk);
1208         u32 seq = tp->copied_seq;
1209         u32 offset;
1210         int copied = 0;
1211
1212         if (sk->sk_state == TCP_LISTEN)
1213                 return -ENOTCONN;
1214         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1215                 if (offset < skb->len) {
1216                         int used;
1217                         size_t len;
1218
1219                         len = skb->len - offset;
1220                         /* Stop reading if we hit a patch of urgent data */
1221                         if (tp->urg_data) {
1222                                 u32 urg_offset = tp->urg_seq - seq;
1223                                 if (urg_offset < len)
1224                                         len = urg_offset;
1225                                 if (!len)
1226                                         break;
1227                         }
1228                         used = recv_actor(desc, skb, offset, len);
1229                         if (used < 0) {
1230                                 if (!copied)
1231                                         copied = used;
1232                                 break;
1233                         } else if (used <= len) {
1234                                 seq += used;
1235                                 copied += used;
1236                                 offset += used;
1237                         }
1238                         /*
1239                          * If recv_actor drops the lock (e.g. TCP splice
1240                          * receive) the skb pointer might be invalid when
1241                          * getting here: tcp_collapse might have deleted it
1242                          * while aggregating skbs from the socket queue.
1243                          */
1244                         skb = tcp_recv_skb(sk, seq-1, &offset);
1245                         if (!skb || (offset+1 != skb->len))
1246                                 break;
1247                 }
1248                 if (tcp_hdr(skb)->fin) {
1249                         sk_eat_skb(sk, skb, 0);
1250                         ++seq;
1251                         break;
1252                 }
1253                 sk_eat_skb(sk, skb, 0);
1254                 if (!desc->count)
1255                         break;
1256         }
1257         tp->copied_seq = seq;
1258
1259         tcp_rcv_space_adjust(sk);
1260
1261         /* Clean up data we have read: This will do ACK frames. */
1262         if (copied > 0)
1263                 tcp_cleanup_rbuf(sk, copied);
1264         return copied;
1265 }
1266
1267 /*
1268  *      This routine copies from a sock struct into the user buffer.
1269  *
1270  *      Technical note: in 2.3 we work on _locked_ socket, so that
1271  *      tricks with *seq access order and skb->users are not required.
1272  *      Probably, code can be easily improved even more.
1273  */
1274
1275 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1276                 size_t len, int nonblock, int flags, int *addr_len)
1277 {
1278         struct tcp_sock *tp = tcp_sk(sk);
1279         int copied = 0;
1280         u32 peek_seq;
1281         u32 *seq;
1282         unsigned long used;
1283         int err;
1284         int target;             /* Read at least this many bytes */
1285         long timeo;
1286         struct task_struct *user_recv = NULL;
1287         int copied_early = 0;
1288         struct sk_buff *skb;
1289
1290         lock_sock(sk);
1291
1292         TCP_CHECK_TIMER(sk);
1293
1294         err = -ENOTCONN;
1295         if (sk->sk_state == TCP_LISTEN)
1296                 goto out;
1297
1298         timeo = sock_rcvtimeo(sk, nonblock);
1299
1300         /* Urgent data needs to be handled specially. */
1301         if (flags & MSG_OOB)
1302                 goto recv_urg;
1303
1304         seq = &tp->copied_seq;
1305         if (flags & MSG_PEEK) {
1306                 peek_seq = tp->copied_seq;
1307                 seq = &peek_seq;
1308         }
1309
1310         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1311
1312 #ifdef CONFIG_NET_DMA
1313         tp->ucopy.dma_chan = NULL;
1314         preempt_disable();
1315         skb = skb_peek_tail(&sk->sk_receive_queue);
1316         {
1317                 int available = 0;
1318
1319                 if (skb)
1320                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1321                 if ((available < target) &&
1322                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1323                     !sysctl_tcp_low_latency &&
1324                     __get_cpu_var(softnet_data).net_dma) {
1325                         preempt_enable_no_resched();
1326                         tp->ucopy.pinned_list =
1327                                         dma_pin_iovec_pages(msg->msg_iov, len);
1328                 } else {
1329                         preempt_enable_no_resched();
1330                 }
1331         }
1332 #endif
1333
1334         do {
1335                 u32 offset;
1336
1337                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1338                 if (tp->urg_data && tp->urg_seq == *seq) {
1339                         if (copied)
1340                                 break;
1341                         if (signal_pending(current)) {
1342                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1343                                 break;
1344                         }
1345                 }
1346
1347                 /* Next get a buffer. */
1348
1349                 skb = skb_peek(&sk->sk_receive_queue);
1350                 do {
1351                         if (!skb)
1352                                 break;
1353
1354                         /* Now that we have two receive queues this
1355                          * shouldn't happen.
1356                          */
1357                         if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1358                                 printk(KERN_INFO "recvmsg bug: copied %X "
1359                                        "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1360                                 break;
1361                         }
1362                         offset = *seq - TCP_SKB_CB(skb)->seq;
1363                         if (tcp_hdr(skb)->syn)
1364                                 offset--;
1365                         if (offset < skb->len)
1366                                 goto found_ok_skb;
1367                         if (tcp_hdr(skb)->fin)
1368                                 goto found_fin_ok;
1369                         WARN_ON(!(flags & MSG_PEEK));
1370                         skb = skb->next;
1371                 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1372
1373                 /* Well, if we have backlog, try to process it now yet. */
1374
1375                 if (copied >= target && !sk->sk_backlog.tail)
1376                         break;
1377
1378                 if (copied) {
1379                         if (sk->sk_err ||
1380                             sk->sk_state == TCP_CLOSE ||
1381                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1382                             !timeo ||
1383                             signal_pending(current) ||
1384                             (flags & MSG_PEEK))
1385                                 break;
1386                 } else {
1387                         if (sock_flag(sk, SOCK_DONE))
1388                                 break;
1389
1390                         if (sk->sk_err) {
1391                                 copied = sock_error(sk);
1392                                 break;
1393                         }
1394
1395                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1396                                 break;
1397
1398                         if (sk->sk_state == TCP_CLOSE) {
1399                                 if (!sock_flag(sk, SOCK_DONE)) {
1400                                         /* This occurs when user tries to read
1401                                          * from never connected socket.
1402                                          */
1403                                         copied = -ENOTCONN;
1404                                         break;
1405                                 }
1406                                 break;
1407                         }
1408
1409                         if (!timeo) {
1410                                 copied = -EAGAIN;
1411                                 break;
1412                         }
1413
1414                         if (signal_pending(current)) {
1415                                 copied = sock_intr_errno(timeo);
1416                                 break;
1417                         }
1418                 }
1419
1420                 tcp_cleanup_rbuf(sk, copied);
1421
1422                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1423                         /* Install new reader */
1424                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1425                                 user_recv = current;
1426                                 tp->ucopy.task = user_recv;
1427                                 tp->ucopy.iov = msg->msg_iov;
1428                         }
1429
1430                         tp->ucopy.len = len;
1431
1432                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1433                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1434
1435                         /* Ugly... If prequeue is not empty, we have to
1436                          * process it before releasing socket, otherwise
1437                          * order will be broken at second iteration.
1438                          * More elegant solution is required!!!
1439                          *
1440                          * Look: we have the following (pseudo)queues:
1441                          *
1442                          * 1. packets in flight
1443                          * 2. backlog
1444                          * 3. prequeue
1445                          * 4. receive_queue
1446                          *
1447                          * Each queue can be processed only if the next ones
1448                          * are empty. At this point we have empty receive_queue.
1449                          * But prequeue _can_ be not empty after 2nd iteration,
1450                          * when we jumped to start of loop because backlog
1451                          * processing added something to receive_queue.
1452                          * We cannot release_sock(), because backlog contains
1453                          * packets arrived _after_ prequeued ones.
1454                          *
1455                          * Shortly, algorithm is clear --- to process all
1456                          * the queues in order. We could make it more directly,
1457                          * requeueing packets from backlog to prequeue, if
1458                          * is not empty. It is more elegant, but eats cycles,
1459                          * unfortunately.
1460                          */
1461                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1462                                 goto do_prequeue;
1463
1464                         /* __ Set realtime policy in scheduler __ */
1465                 }
1466
1467                 if (copied >= target) {
1468                         /* Do not sleep, just process backlog. */
1469                         release_sock(sk);
1470                         lock_sock(sk);
1471                 } else
1472                         sk_wait_data(sk, &timeo);
1473
1474 #ifdef CONFIG_NET_DMA
1475                 tp->ucopy.wakeup = 0;
1476 #endif
1477
1478                 if (user_recv) {
1479                         int chunk;
1480
1481                         /* __ Restore normal policy in scheduler __ */
1482
1483                         if ((chunk = len - tp->ucopy.len) != 0) {
1484                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1485                                 len -= chunk;
1486                                 copied += chunk;
1487                         }
1488
1489                         if (tp->rcv_nxt == tp->copied_seq &&
1490                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1491 do_prequeue:
1492                                 tcp_prequeue_process(sk);
1493
1494                                 if ((chunk = len - tp->ucopy.len) != 0) {
1495                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1496                                         len -= chunk;
1497                                         copied += chunk;
1498                                 }
1499                         }
1500                 }
1501                 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1502                         if (net_ratelimit())
1503                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1504                                        current->comm, task_pid_nr(current));
1505                         peek_seq = tp->copied_seq;
1506                 }
1507                 continue;
1508
1509         found_ok_skb:
1510                 /* Ok so how much can we use? */
1511                 used = skb->len - offset;
1512                 if (len < used)
1513                         used = len;
1514
1515                 /* Do we have urgent data here? */
1516                 if (tp->urg_data) {
1517                         u32 urg_offset = tp->urg_seq - *seq;
1518                         if (urg_offset < used) {
1519                                 if (!urg_offset) {
1520                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1521                                                 ++*seq;
1522                                                 offset++;
1523                                                 used--;
1524                                                 if (!used)
1525                                                         goto skip_copy;
1526                                         }
1527                                 } else
1528                                         used = urg_offset;
1529                         }
1530                 }
1531
1532                 if (!(flags & MSG_TRUNC)) {
1533 #ifdef CONFIG_NET_DMA
1534                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1535                                 tp->ucopy.dma_chan = get_softnet_dma();
1536
1537                         if (tp->ucopy.dma_chan) {
1538                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1539                                         tp->ucopy.dma_chan, skb, offset,
1540                                         msg->msg_iov, used,
1541                                         tp->ucopy.pinned_list);
1542
1543                                 if (tp->ucopy.dma_cookie < 0) {
1544
1545                                         printk(KERN_ALERT "dma_cookie < 0\n");
1546
1547                                         /* Exception. Bailout! */
1548                                         if (!copied)
1549                                                 copied = -EFAULT;
1550                                         break;
1551                                 }
1552                                 if ((offset + used) == skb->len)
1553                                         copied_early = 1;
1554
1555                         } else
1556 #endif
1557                         {
1558                                 err = skb_copy_datagram_iovec(skb, offset,
1559                                                 msg->msg_iov, used);
1560                                 if (err) {
1561                                         /* Exception. Bailout! */
1562                                         if (!copied)
1563                                                 copied = -EFAULT;
1564                                         break;
1565                                 }
1566                         }
1567                 }
1568
1569                 *seq += used;
1570                 copied += used;
1571                 len -= used;
1572
1573                 tcp_rcv_space_adjust(sk);
1574
1575 skip_copy:
1576                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1577                         tp->urg_data = 0;
1578                         tcp_fast_path_check(sk);
1579                 }
1580                 if (used + offset < skb->len)
1581                         continue;
1582
1583                 if (tcp_hdr(skb)->fin)
1584                         goto found_fin_ok;
1585                 if (!(flags & MSG_PEEK)) {
1586                         sk_eat_skb(sk, skb, copied_early);
1587                         copied_early = 0;
1588                 }
1589                 continue;
1590
1591         found_fin_ok:
1592                 /* Process the FIN. */
1593                 ++*seq;
1594                 if (!(flags & MSG_PEEK)) {
1595                         sk_eat_skb(sk, skb, copied_early);
1596                         copied_early = 0;
1597                 }
1598                 break;
1599         } while (len > 0);
1600
1601         if (user_recv) {
1602                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1603                         int chunk;
1604
1605                         tp->ucopy.len = copied > 0 ? len : 0;
1606
1607                         tcp_prequeue_process(sk);
1608
1609                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1610                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1611                                 len -= chunk;
1612                                 copied += chunk;
1613                         }
1614                 }
1615
1616                 tp->ucopy.task = NULL;
1617                 tp->ucopy.len = 0;
1618         }
1619
1620 #ifdef CONFIG_NET_DMA
1621         if (tp->ucopy.dma_chan) {
1622                 dma_cookie_t done, used;
1623
1624                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1625
1626                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1627                                                  tp->ucopy.dma_cookie, &done,
1628                                                  &used) == DMA_IN_PROGRESS) {
1629                         /* do partial cleanup of sk_async_wait_queue */
1630                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1631                                (dma_async_is_complete(skb->dma_cookie, done,
1632                                                       used) == DMA_SUCCESS)) {
1633                                 __skb_dequeue(&sk->sk_async_wait_queue);
1634                                 kfree_skb(skb);
1635                         }
1636                 }
1637
1638                 /* Safe to free early-copied skbs now */
1639                 __skb_queue_purge(&sk->sk_async_wait_queue);
1640                 dma_chan_put(tp->ucopy.dma_chan);
1641                 tp->ucopy.dma_chan = NULL;
1642         }
1643         if (tp->ucopy.pinned_list) {
1644                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1645                 tp->ucopy.pinned_list = NULL;
1646         }
1647 #endif
1648
1649         /* According to UNIX98, msg_name/msg_namelen are ignored
1650          * on connected socket. I was just happy when found this 8) --ANK
1651          */
1652
1653         /* Clean up data we have read: This will do ACK frames. */
1654         tcp_cleanup_rbuf(sk, copied);
1655
1656         TCP_CHECK_TIMER(sk);
1657         release_sock(sk);
1658         return copied;
1659
1660 out:
1661         TCP_CHECK_TIMER(sk);
1662         release_sock(sk);
1663         return err;
1664
1665 recv_urg:
1666         err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1667         goto out;
1668 }
1669
1670 void tcp_set_state(struct sock *sk, int state)
1671 {
1672         int oldstate = sk->sk_state;
1673
1674         switch (state) {
1675         case TCP_ESTABLISHED:
1676                 if (oldstate != TCP_ESTABLISHED)
1677                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1678                 break;
1679
1680         case TCP_CLOSE:
1681                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1682                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1683
1684                 sk->sk_prot->unhash(sk);
1685                 if (inet_csk(sk)->icsk_bind_hash &&
1686                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1687                         inet_put_port(sk);
1688                 /* fall through */
1689         default:
1690                 if (oldstate==TCP_ESTABLISHED)
1691                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1692         }
1693
1694         /* Change state AFTER socket is unhashed to avoid closed
1695          * socket sitting in hash tables.
1696          */
1697         sk->sk_state = state;
1698
1699 #ifdef STATE_TRACE
1700         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n",sk, statename[oldstate],statename[state]);
1701 #endif
1702 }
1703 EXPORT_SYMBOL_GPL(tcp_set_state);
1704
1705 /*
1706  *      State processing on a close. This implements the state shift for
1707  *      sending our FIN frame. Note that we only send a FIN for some
1708  *      states. A shutdown() may have already sent the FIN, or we may be
1709  *      closed.
1710  */
1711
1712 static const unsigned char new_state[16] = {
1713   /* current state:        new state:      action:      */
1714   /* (Invalid)          */ TCP_CLOSE,
1715   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1716   /* TCP_SYN_SENT       */ TCP_CLOSE,
1717   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1718   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1719   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1720   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1721   /* TCP_CLOSE          */ TCP_CLOSE,
1722   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1723   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1724   /* TCP_LISTEN         */ TCP_CLOSE,
1725   /* TCP_CLOSING        */ TCP_CLOSING,
1726 };
1727
1728 static int tcp_close_state(struct sock *sk)
1729 {
1730         int next = (int)new_state[sk->sk_state];
1731         int ns = next & TCP_STATE_MASK;
1732
1733         tcp_set_state(sk, ns);
1734
1735         return next & TCP_ACTION_FIN;
1736 }
1737
1738 /*
1739  *      Shutdown the sending side of a connection. Much like close except
1740  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1741  */
1742
1743 void tcp_shutdown(struct sock *sk, int how)
1744 {
1745         /*      We need to grab some memory, and put together a FIN,
1746          *      and then put it into the queue to be sent.
1747          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1748          */
1749         if (!(how & SEND_SHUTDOWN))
1750                 return;
1751
1752         /* If we've already sent a FIN, or it's a closed state, skip this. */
1753         if ((1 << sk->sk_state) &
1754             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1755              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1756                 /* Clear out any half completed packets.  FIN if needed. */
1757                 if (tcp_close_state(sk))
1758                         tcp_send_fin(sk);
1759         }
1760 }
1761
1762 void tcp_close(struct sock *sk, long timeout)
1763 {
1764         struct sk_buff *skb;
1765         int data_was_unread = 0;
1766         int state;
1767
1768         lock_sock(sk);
1769         sk->sk_shutdown = SHUTDOWN_MASK;
1770
1771         if (sk->sk_state == TCP_LISTEN) {
1772                 tcp_set_state(sk, TCP_CLOSE);
1773
1774                 /* Special case. */
1775                 inet_csk_listen_stop(sk);
1776
1777                 goto adjudge_to_death;
1778         }
1779
1780         /*  We need to flush the recv. buffs.  We do this only on the
1781          *  descriptor close, not protocol-sourced closes, because the
1782          *  reader process may not have drained the data yet!
1783          */
1784         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1785                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1786                           tcp_hdr(skb)->fin;
1787                 data_was_unread += len;
1788                 __kfree_skb(skb);
1789         }
1790
1791         sk_mem_reclaim(sk);
1792
1793         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1794          * data was lost. To witness the awful effects of the old behavior of
1795          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1796          * GET in an FTP client, suspend the process, wait for the client to
1797          * advertise a zero window, then kill -9 the FTP client, wheee...
1798          * Note: timeout is always zero in such a case.
1799          */
1800         if (data_was_unread) {
1801                 /* Unread data was tossed, zap the connection. */
1802                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1803                 tcp_set_state(sk, TCP_CLOSE);
1804                 tcp_send_active_reset(sk, GFP_KERNEL);
1805         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1806                 /* Check zero linger _after_ checking for unread data. */
1807                 sk->sk_prot->disconnect(sk, 0);
1808                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1809         } else if (tcp_close_state(sk)) {
1810                 /* We FIN if the application ate all the data before
1811                  * zapping the connection.
1812                  */
1813
1814                 /* RED-PEN. Formally speaking, we have broken TCP state
1815                  * machine. State transitions:
1816                  *
1817                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1818                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1819                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1820                  *
1821                  * are legal only when FIN has been sent (i.e. in window),
1822                  * rather than queued out of window. Purists blame.
1823                  *
1824                  * F.e. "RFC state" is ESTABLISHED,
1825                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1826                  *
1827                  * The visible declinations are that sometimes
1828                  * we enter time-wait state, when it is not required really
1829                  * (harmless), do not send active resets, when they are
1830                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1831                  * they look as CLOSING or LAST_ACK for Linux)
1832                  * Probably, I missed some more holelets.
1833                  *                                              --ANK
1834                  */
1835                 tcp_send_fin(sk);
1836         }
1837
1838         sk_stream_wait_close(sk, timeout);
1839
1840 adjudge_to_death:
1841         state = sk->sk_state;
1842         sock_hold(sk);
1843         sock_orphan(sk);
1844         atomic_inc(sk->sk_prot->orphan_count);
1845
1846         /* It is the last release_sock in its life. It will remove backlog. */
1847         release_sock(sk);
1848
1849
1850         /* Now socket is owned by kernel and we acquire BH lock
1851            to finish close. No need to check for user refs.
1852          */
1853         local_bh_disable();
1854         bh_lock_sock(sk);
1855         WARN_ON(sock_owned_by_user(sk));
1856
1857         /* Have we already been destroyed by a softirq or backlog? */
1858         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1859                 goto out;
1860
1861         /*      This is a (useful) BSD violating of the RFC. There is a
1862          *      problem with TCP as specified in that the other end could
1863          *      keep a socket open forever with no application left this end.
1864          *      We use a 3 minute timeout (about the same as BSD) then kill
1865          *      our end. If they send after that then tough - BUT: long enough
1866          *      that we won't make the old 4*rto = almost no time - whoops
1867          *      reset mistake.
1868          *
1869          *      Nope, it was not mistake. It is really desired behaviour
1870          *      f.e. on http servers, when such sockets are useless, but
1871          *      consume significant resources. Let's do it with special
1872          *      linger2 option.                                 --ANK
1873          */
1874
1875         if (sk->sk_state == TCP_FIN_WAIT2) {
1876                 struct tcp_sock *tp = tcp_sk(sk);
1877                 if (tp->linger2 < 0) {
1878                         tcp_set_state(sk, TCP_CLOSE);
1879                         tcp_send_active_reset(sk, GFP_ATOMIC);
1880                         NET_INC_STATS_BH(sock_net(sk),
1881                                         LINUX_MIB_TCPABORTONLINGER);
1882                 } else {
1883                         const int tmo = tcp_fin_time(sk);
1884
1885                         if (tmo > TCP_TIMEWAIT_LEN) {
1886                                 inet_csk_reset_keepalive_timer(sk,
1887                                                 tmo - TCP_TIMEWAIT_LEN);
1888                         } else {
1889                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1890                                 goto out;
1891                         }
1892                 }
1893         }
1894         if (sk->sk_state != TCP_CLOSE) {
1895                 sk_mem_reclaim(sk);
1896                 if (tcp_too_many_orphans(sk,
1897                                 atomic_read(sk->sk_prot->orphan_count))) {
1898                         if (net_ratelimit())
1899                                 printk(KERN_INFO "TCP: too many of orphaned "
1900                                        "sockets\n");
1901                         tcp_set_state(sk, TCP_CLOSE);
1902                         tcp_send_active_reset(sk, GFP_ATOMIC);
1903                         NET_INC_STATS_BH(sock_net(sk),
1904                                         LINUX_MIB_TCPABORTONMEMORY);
1905                 }
1906         }
1907
1908         if (sk->sk_state == TCP_CLOSE)
1909                 inet_csk_destroy_sock(sk);
1910         /* Otherwise, socket is reprieved until protocol close. */
1911
1912 out:
1913         bh_unlock_sock(sk);
1914         local_bh_enable();
1915         sock_put(sk);
1916 }
1917
1918 /* These states need RST on ABORT according to RFC793 */
1919
1920 static inline int tcp_need_reset(int state)
1921 {
1922         return (1 << state) &
1923                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1924                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1925 }
1926
1927 int tcp_disconnect(struct sock *sk, int flags)
1928 {
1929         struct inet_sock *inet = inet_sk(sk);
1930         struct inet_connection_sock *icsk = inet_csk(sk);
1931         struct tcp_sock *tp = tcp_sk(sk);
1932         int err = 0;
1933         int old_state = sk->sk_state;
1934
1935         if (old_state != TCP_CLOSE)
1936                 tcp_set_state(sk, TCP_CLOSE);
1937
1938         /* ABORT function of RFC793 */
1939         if (old_state == TCP_LISTEN) {
1940                 inet_csk_listen_stop(sk);
1941         } else if (tcp_need_reset(old_state) ||
1942                    (tp->snd_nxt != tp->write_seq &&
1943                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1944                 /* The last check adjusts for discrepancy of Linux wrt. RFC
1945                  * states
1946                  */
1947                 tcp_send_active_reset(sk, gfp_any());
1948                 sk->sk_err = ECONNRESET;
1949         } else if (old_state == TCP_SYN_SENT)
1950                 sk->sk_err = ECONNRESET;
1951
1952         tcp_clear_xmit_timers(sk);
1953         __skb_queue_purge(&sk->sk_receive_queue);
1954         tcp_write_queue_purge(sk);
1955         __skb_queue_purge(&tp->out_of_order_queue);
1956 #ifdef CONFIG_NET_DMA
1957         __skb_queue_purge(&sk->sk_async_wait_queue);
1958 #endif
1959
1960         inet->dport = 0;
1961
1962         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1963                 inet_reset_saddr(sk);
1964
1965         sk->sk_shutdown = 0;
1966         sock_reset_flag(sk, SOCK_DONE);
1967         tp->srtt = 0;
1968         if ((tp->write_seq += tp->max_window + 2) == 0)
1969                 tp->write_seq = 1;
1970         icsk->icsk_backoff = 0;
1971         tp->snd_cwnd = 2;
1972         icsk->icsk_probes_out = 0;
1973         tp->packets_out = 0;
1974         tp->snd_ssthresh = 0x7fffffff;
1975         tp->snd_cwnd_cnt = 0;
1976         tp->bytes_acked = 0;
1977         tcp_set_ca_state(sk, TCP_CA_Open);
1978         tcp_clear_retrans(tp);
1979         inet_csk_delack_init(sk);
1980         tcp_init_send_head(sk);
1981         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
1982         __sk_dst_reset(sk);
1983
1984         WARN_ON(inet->num && !icsk->icsk_bind_hash);
1985
1986         sk->sk_error_report(sk);
1987         return err;
1988 }
1989
1990 /*
1991  *      Socket option code for TCP.
1992  */
1993 static int do_tcp_setsockopt(struct sock *sk, int level,
1994                 int optname, char __user *optval, int optlen)
1995 {
1996         struct tcp_sock *tp = tcp_sk(sk);
1997         struct inet_connection_sock *icsk = inet_csk(sk);
1998         int val;
1999         int err = 0;
2000
2001         /* This is a string value all the others are int's */
2002         if (optname == TCP_CONGESTION) {
2003                 char name[TCP_CA_NAME_MAX];
2004
2005                 if (optlen < 1)
2006                         return -EINVAL;
2007
2008                 val = strncpy_from_user(name, optval,
2009                                         min(TCP_CA_NAME_MAX-1, optlen));
2010                 if (val < 0)
2011                         return -EFAULT;
2012                 name[val] = 0;
2013
2014                 lock_sock(sk);
2015                 err = tcp_set_congestion_control(sk, name);
2016                 release_sock(sk);
2017                 return err;
2018         }
2019
2020         if (optlen < sizeof(int))
2021                 return -EINVAL;
2022
2023         if (get_user(val, (int __user *)optval))
2024                 return -EFAULT;
2025
2026         lock_sock(sk);
2027
2028         switch (optname) {
2029         case TCP_MAXSEG:
2030                 /* Values greater than interface MTU won't take effect. However
2031                  * at the point when this call is done we typically don't yet
2032                  * know which interface is going to be used */
2033                 if (val < 8 || val > MAX_TCP_WINDOW) {
2034                         err = -EINVAL;
2035                         break;
2036                 }
2037                 tp->rx_opt.user_mss = val;
2038                 break;
2039
2040         case TCP_NODELAY:
2041                 if (val) {
2042                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2043                          * this option on corked socket is remembered, but
2044                          * it is not activated until cork is cleared.
2045                          *
2046                          * However, when TCP_NODELAY is set we make
2047                          * an explicit push, which overrides even TCP_CORK
2048                          * for currently queued segments.
2049                          */
2050                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2051                         tcp_push_pending_frames(sk);
2052                 } else {
2053                         tp->nonagle &= ~TCP_NAGLE_OFF;
2054                 }
2055                 break;
2056
2057         case TCP_CORK:
2058                 /* When set indicates to always queue non-full frames.
2059                  * Later the user clears this option and we transmit
2060                  * any pending partial frames in the queue.  This is
2061                  * meant to be used alongside sendfile() to get properly
2062                  * filled frames when the user (for example) must write
2063                  * out headers with a write() call first and then use
2064                  * sendfile to send out the data parts.
2065                  *
2066                  * TCP_CORK can be set together with TCP_NODELAY and it is
2067                  * stronger than TCP_NODELAY.
2068                  */
2069                 if (val) {
2070                         tp->nonagle |= TCP_NAGLE_CORK;
2071                 } else {
2072                         tp->nonagle &= ~TCP_NAGLE_CORK;
2073                         if (tp->nonagle&TCP_NAGLE_OFF)
2074                                 tp->nonagle |= TCP_NAGLE_PUSH;
2075                         tcp_push_pending_frames(sk);
2076                 }
2077                 break;
2078
2079         case TCP_KEEPIDLE:
2080                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2081                         err = -EINVAL;
2082                 else {
2083                         tp->keepalive_time = val * HZ;
2084                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2085                             !((1 << sk->sk_state) &
2086                               (TCPF_CLOSE | TCPF_LISTEN))) {
2087                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2088                                 if (tp->keepalive_time > elapsed)
2089                                         elapsed = tp->keepalive_time - elapsed;
2090                                 else
2091                                         elapsed = 0;
2092                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2093                         }
2094                 }
2095                 break;
2096         case TCP_KEEPINTVL:
2097                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2098                         err = -EINVAL;
2099                 else
2100                         tp->keepalive_intvl = val * HZ;
2101                 break;
2102         case TCP_KEEPCNT:
2103                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2104                         err = -EINVAL;
2105                 else
2106                         tp->keepalive_probes = val;
2107                 break;
2108         case TCP_SYNCNT:
2109                 if (val < 1 || val > MAX_TCP_SYNCNT)
2110                         err = -EINVAL;
2111                 else
2112                         icsk->icsk_syn_retries = val;
2113                 break;
2114
2115         case TCP_LINGER2:
2116                 if (val < 0)
2117                         tp->linger2 = -1;
2118                 else if (val > sysctl_tcp_fin_timeout / HZ)
2119                         tp->linger2 = 0;
2120                 else
2121                         tp->linger2 = val * HZ;
2122                 break;
2123
2124         case TCP_DEFER_ACCEPT:
2125                 icsk->icsk_accept_queue.rskq_defer_accept = 0;
2126                 if (val > 0) {
2127                         /* Translate value in seconds to number of
2128                          * retransmits */
2129                         while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
2130                                val > ((TCP_TIMEOUT_INIT / HZ) <<
2131                                        icsk->icsk_accept_queue.rskq_defer_accept))
2132                                 icsk->icsk_accept_queue.rskq_defer_accept++;
2133                         icsk->icsk_accept_queue.rskq_defer_accept++;
2134                 }
2135                 break;
2136
2137         case TCP_WINDOW_CLAMP:
2138                 if (!val) {
2139                         if (sk->sk_state != TCP_CLOSE) {
2140                                 err = -EINVAL;
2141                                 break;
2142                         }
2143                         tp->window_clamp = 0;
2144                 } else
2145                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2146                                                 SOCK_MIN_RCVBUF / 2 : val;
2147                 break;
2148
2149         case TCP_QUICKACK:
2150                 if (!val) {
2151                         icsk->icsk_ack.pingpong = 1;
2152                 } else {
2153                         icsk->icsk_ack.pingpong = 0;
2154                         if ((1 << sk->sk_state) &
2155                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2156                             inet_csk_ack_scheduled(sk)) {
2157                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2158                                 tcp_cleanup_rbuf(sk, 1);
2159                                 if (!(val & 1))
2160                                         icsk->icsk_ack.pingpong = 1;
2161                         }
2162                 }
2163                 break;
2164
2165 #ifdef CONFIG_TCP_MD5SIG
2166         case TCP_MD5SIG:
2167                 /* Read the IP->Key mappings from userspace */
2168                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2169                 break;
2170 #endif
2171
2172         default:
2173                 err = -ENOPROTOOPT;
2174                 break;
2175         }
2176
2177         release_sock(sk);
2178         return err;
2179 }
2180
2181 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2182                    int optlen)
2183 {
2184         struct inet_connection_sock *icsk = inet_csk(sk);
2185
2186         if (level != SOL_TCP)
2187                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2188                                                      optval, optlen);
2189         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2190 }
2191
2192 #ifdef CONFIG_COMPAT
2193 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2194                           char __user *optval, int optlen)
2195 {
2196         if (level != SOL_TCP)
2197                 return inet_csk_compat_setsockopt(sk, level, optname,
2198                                                   optval, optlen);
2199         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2200 }
2201
2202 EXPORT_SYMBOL(compat_tcp_setsockopt);
2203 #endif
2204
2205 /* Return information about state of tcp endpoint in API format. */
2206 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2207 {
2208         struct tcp_sock *tp = tcp_sk(sk);
2209         const struct inet_connection_sock *icsk = inet_csk(sk);
2210         u32 now = tcp_time_stamp;
2211
2212         memset(info, 0, sizeof(*info));
2213
2214         info->tcpi_state = sk->sk_state;
2215         info->tcpi_ca_state = icsk->icsk_ca_state;
2216         info->tcpi_retransmits = icsk->icsk_retransmits;
2217         info->tcpi_probes = icsk->icsk_probes_out;
2218         info->tcpi_backoff = icsk->icsk_backoff;
2219
2220         if (tp->rx_opt.tstamp_ok)
2221                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2222         if (tcp_is_sack(tp))
2223                 info->tcpi_options |= TCPI_OPT_SACK;
2224         if (tp->rx_opt.wscale_ok) {
2225                 info->tcpi_options |= TCPI_OPT_WSCALE;
2226                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2227                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2228         }
2229
2230         if (tp->ecn_flags&TCP_ECN_OK)
2231                 info->tcpi_options |= TCPI_OPT_ECN;
2232
2233         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2234         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2235         info->tcpi_snd_mss = tp->mss_cache;
2236         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2237
2238         if (sk->sk_state == TCP_LISTEN) {
2239                 info->tcpi_unacked = sk->sk_ack_backlog;
2240                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2241         } else {
2242                 info->tcpi_unacked = tp->packets_out;
2243                 info->tcpi_sacked = tp->sacked_out;
2244         }
2245         info->tcpi_lost = tp->lost_out;
2246         info->tcpi_retrans = tp->retrans_out;
2247         info->tcpi_fackets = tp->fackets_out;
2248
2249         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2250         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2251         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2252
2253         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2254         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2255         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2256         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2257         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2258         info->tcpi_snd_cwnd = tp->snd_cwnd;
2259         info->tcpi_advmss = tp->advmss;
2260         info->tcpi_reordering = tp->reordering;
2261
2262         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2263         info->tcpi_rcv_space = tp->rcvq_space.space;
2264
2265         info->tcpi_total_retrans = tp->total_retrans;
2266 }
2267
2268 EXPORT_SYMBOL_GPL(tcp_get_info);
2269
2270 static int do_tcp_getsockopt(struct sock *sk, int level,
2271                 int optname, char __user *optval, int __user *optlen)
2272 {
2273         struct inet_connection_sock *icsk = inet_csk(sk);
2274         struct tcp_sock *tp = tcp_sk(sk);
2275         int val, len;
2276
2277         if (get_user(len, optlen))
2278                 return -EFAULT;
2279
2280         len = min_t(unsigned int, len, sizeof(int));
2281
2282         if (len < 0)
2283                 return -EINVAL;
2284
2285         switch (optname) {
2286         case TCP_MAXSEG:
2287                 val = tp->mss_cache;
2288                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2289                         val = tp->rx_opt.user_mss;
2290                 break;
2291         case TCP_NODELAY:
2292                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2293                 break;
2294         case TCP_CORK:
2295                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2296                 break;
2297         case TCP_KEEPIDLE:
2298                 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2299                 break;
2300         case TCP_KEEPINTVL:
2301                 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2302                 break;
2303         case TCP_KEEPCNT:
2304                 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2305                 break;
2306         case TCP_SYNCNT:
2307                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2308                 break;
2309         case TCP_LINGER2:
2310                 val = tp->linger2;
2311                 if (val >= 0)
2312                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2313                 break;
2314         case TCP_DEFER_ACCEPT:
2315                 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2316                         ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2317                 break;
2318         case TCP_WINDOW_CLAMP:
2319                 val = tp->window_clamp;
2320                 break;
2321         case TCP_INFO: {
2322                 struct tcp_info info;
2323
2324                 if (get_user(len, optlen))
2325                         return -EFAULT;
2326
2327                 tcp_get_info(sk, &info);
2328
2329                 len = min_t(unsigned int, len, sizeof(info));
2330                 if (put_user(len, optlen))
2331                         return -EFAULT;
2332                 if (copy_to_user(optval, &info, len))
2333                         return -EFAULT;
2334                 return 0;
2335         }
2336         case TCP_QUICKACK:
2337                 val = !icsk->icsk_ack.pingpong;
2338                 break;
2339
2340         case TCP_CONGESTION:
2341                 if (get_user(len, optlen))
2342                         return -EFAULT;
2343                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2344                 if (put_user(len, optlen))
2345                         return -EFAULT;
2346                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2347                         return -EFAULT;
2348                 return 0;
2349         default:
2350                 return -ENOPROTOOPT;
2351         }
2352
2353         if (put_user(len, optlen))
2354                 return -EFAULT;
2355         if (copy_to_user(optval, &val, len))
2356                 return -EFAULT;
2357         return 0;
2358 }
2359
2360 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2361                    int __user *optlen)
2362 {
2363         struct inet_connection_sock *icsk = inet_csk(sk);
2364
2365         if (level != SOL_TCP)
2366                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2367                                                      optval, optlen);
2368         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2369 }
2370
2371 #ifdef CONFIG_COMPAT
2372 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2373                           char __user *optval, int __user *optlen)
2374 {
2375         if (level != SOL_TCP)
2376                 return inet_csk_compat_getsockopt(sk, level, optname,
2377                                                   optval, optlen);
2378         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2379 }
2380
2381 EXPORT_SYMBOL(compat_tcp_getsockopt);
2382 #endif
2383
2384 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2385 {
2386         struct sk_buff *segs = ERR_PTR(-EINVAL);
2387         struct tcphdr *th;
2388         unsigned thlen;
2389         unsigned int seq;
2390         __be32 delta;
2391         unsigned int oldlen;
2392         unsigned int len;
2393
2394         if (!pskb_may_pull(skb, sizeof(*th)))
2395                 goto out;
2396
2397         th = tcp_hdr(skb);
2398         thlen = th->doff * 4;
2399         if (thlen < sizeof(*th))
2400                 goto out;
2401
2402         if (!pskb_may_pull(skb, thlen))
2403                 goto out;
2404
2405         oldlen = (u16)~skb->len;
2406         __skb_pull(skb, thlen);
2407
2408         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2409                 /* Packet is from an untrusted source, reset gso_segs. */
2410                 int type = skb_shinfo(skb)->gso_type;
2411                 int mss;
2412
2413                 if (unlikely(type &
2414                              ~(SKB_GSO_TCPV4 |
2415                                SKB_GSO_DODGY |
2416                                SKB_GSO_TCP_ECN |
2417                                SKB_GSO_TCPV6 |
2418                                0) ||
2419                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2420                         goto out;
2421
2422                 mss = skb_shinfo(skb)->gso_size;
2423                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2424
2425                 segs = NULL;
2426                 goto out;
2427         }
2428
2429         segs = skb_segment(skb, features);
2430         if (IS_ERR(segs))
2431                 goto out;
2432
2433         len = skb_shinfo(skb)->gso_size;
2434         delta = htonl(oldlen + (thlen + len));
2435
2436         skb = segs;
2437         th = tcp_hdr(skb);
2438         seq = ntohl(th->seq);
2439
2440         do {
2441                 th->fin = th->psh = 0;
2442
2443                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2444                                        (__force u32)delta));
2445                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2446                         th->check =
2447                              csum_fold(csum_partial(skb_transport_header(skb),
2448                                                     thlen, skb->csum));
2449
2450                 seq += len;
2451                 skb = skb->next;
2452                 th = tcp_hdr(skb);
2453
2454                 th->seq = htonl(seq);
2455                 th->cwr = 0;
2456         } while (skb->next);
2457
2458         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2459                       skb->data_len);
2460         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2461                                 (__force u32)delta));
2462         if (skb->ip_summed != CHECKSUM_PARTIAL)
2463                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2464                                                    thlen, skb->csum));
2465
2466 out:
2467         return segs;
2468 }
2469 EXPORT_SYMBOL(tcp_tso_segment);
2470
2471 #ifdef CONFIG_TCP_MD5SIG
2472 static unsigned long tcp_md5sig_users;
2473 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2474 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2475
2476 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2477 {
2478         int cpu;
2479         for_each_possible_cpu(cpu) {
2480                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2481                 if (p) {
2482                         if (p->md5_desc.tfm)
2483                                 crypto_free_hash(p->md5_desc.tfm);
2484                         kfree(p);
2485                         p = NULL;
2486                 }
2487         }
2488         free_percpu(pool);
2489 }
2490
2491 void tcp_free_md5sig_pool(void)
2492 {
2493         struct tcp_md5sig_pool **pool = NULL;
2494
2495         spin_lock_bh(&tcp_md5sig_pool_lock);
2496         if (--tcp_md5sig_users == 0) {
2497                 pool = tcp_md5sig_pool;
2498                 tcp_md5sig_pool = NULL;
2499         }
2500         spin_unlock_bh(&tcp_md5sig_pool_lock);
2501         if (pool)
2502                 __tcp_free_md5sig_pool(pool);
2503 }
2504
2505 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2506
2507 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(void)
2508 {
2509         int cpu;
2510         struct tcp_md5sig_pool **pool;
2511
2512         pool = alloc_percpu(struct tcp_md5sig_pool *);
2513         if (!pool)
2514                 return NULL;
2515
2516         for_each_possible_cpu(cpu) {
2517                 struct tcp_md5sig_pool *p;
2518                 struct crypto_hash *hash;
2519
2520                 p = kzalloc(sizeof(*p), GFP_KERNEL);
2521                 if (!p)
2522                         goto out_free;
2523                 *per_cpu_ptr(pool, cpu) = p;
2524
2525                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2526                 if (!hash || IS_ERR(hash))
2527                         goto out_free;
2528
2529                 p->md5_desc.tfm = hash;
2530         }
2531         return pool;
2532 out_free:
2533         __tcp_free_md5sig_pool(pool);
2534         return NULL;
2535 }
2536
2537 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(void)
2538 {
2539         struct tcp_md5sig_pool **pool;
2540         int alloc = 0;
2541
2542 retry:
2543         spin_lock_bh(&tcp_md5sig_pool_lock);
2544         pool = tcp_md5sig_pool;
2545         if (tcp_md5sig_users++ == 0) {
2546                 alloc = 1;
2547                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2548         } else if (!pool) {
2549                 tcp_md5sig_users--;
2550                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2551                 cpu_relax();
2552                 goto retry;
2553         } else
2554                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2555
2556         if (alloc) {
2557                 /* we cannot hold spinlock here because this may sleep. */
2558                 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool();
2559                 spin_lock_bh(&tcp_md5sig_pool_lock);
2560                 if (!p) {
2561                         tcp_md5sig_users--;
2562                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2563                         return NULL;
2564                 }
2565                 pool = tcp_md5sig_pool;
2566                 if (pool) {
2567                         /* oops, it has already been assigned. */
2568                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2569                         __tcp_free_md5sig_pool(p);
2570                 } else {
2571                         tcp_md5sig_pool = pool = p;
2572                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2573                 }
2574         }
2575         return pool;
2576 }
2577
2578 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2579
2580 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2581 {
2582         struct tcp_md5sig_pool **p;
2583         spin_lock_bh(&tcp_md5sig_pool_lock);
2584         p = tcp_md5sig_pool;
2585         if (p)
2586                 tcp_md5sig_users++;
2587         spin_unlock_bh(&tcp_md5sig_pool_lock);
2588         return (p ? *per_cpu_ptr(p, cpu) : NULL);
2589 }
2590
2591 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2592
2593 void __tcp_put_md5sig_pool(void)
2594 {
2595         tcp_free_md5sig_pool();
2596 }
2597
2598 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2599
2600 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2601                         struct tcphdr *th)
2602 {
2603         struct scatterlist sg;
2604         int err;
2605
2606         __sum16 old_checksum = th->check;
2607         th->check = 0;
2608         /* options aren't included in the hash */
2609         sg_init_one(&sg, th, sizeof(struct tcphdr));
2610         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
2611         th->check = old_checksum;
2612         return err;
2613 }
2614
2615 EXPORT_SYMBOL(tcp_md5_hash_header);
2616
2617 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2618                           struct sk_buff *skb, unsigned header_len)
2619 {
2620         struct scatterlist sg;
2621         const struct tcphdr *tp = tcp_hdr(skb);
2622         struct hash_desc *desc = &hp->md5_desc;
2623         unsigned i;
2624         const unsigned head_data_len = skb_headlen(skb) > header_len ?
2625                                        skb_headlen(skb) - header_len : 0;
2626         const struct skb_shared_info *shi = skb_shinfo(skb);
2627
2628         sg_init_table(&sg, 1);
2629
2630         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
2631         if (crypto_hash_update(desc, &sg, head_data_len))
2632                 return 1;
2633
2634         for (i = 0; i < shi->nr_frags; ++i) {
2635                 const struct skb_frag_struct *f = &shi->frags[i];
2636                 sg_set_page(&sg, f->page, f->size, f->page_offset);
2637                 if (crypto_hash_update(desc, &sg, f->size))
2638                         return 1;
2639         }
2640
2641         return 0;
2642 }
2643
2644 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
2645
2646 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
2647 {
2648         struct scatterlist sg;
2649
2650         sg_init_one(&sg, key->key, key->keylen);
2651         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
2652 }
2653
2654 EXPORT_SYMBOL(tcp_md5_hash_key);
2655
2656 #endif
2657
2658 void tcp_done(struct sock *sk)
2659 {
2660         if(sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2661                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
2662
2663         tcp_set_state(sk, TCP_CLOSE);
2664         tcp_clear_xmit_timers(sk);
2665
2666         sk->sk_shutdown = SHUTDOWN_MASK;
2667
2668         if (!sock_flag(sk, SOCK_DEAD))
2669                 sk->sk_state_change(sk);
2670         else
2671                 inet_csk_destroy_sock(sk);
2672 }
2673 EXPORT_SYMBOL_GPL(tcp_done);
2674
2675 extern struct tcp_congestion_ops tcp_reno;
2676
2677 static __initdata unsigned long thash_entries;
2678 static int __init set_thash_entries(char *str)
2679 {
2680         if (!str)
2681                 return 0;
2682         thash_entries = simple_strtoul(str, &str, 0);
2683         return 1;
2684 }
2685 __setup("thash_entries=", set_thash_entries);
2686
2687 void __init tcp_init(void)
2688 {
2689         struct sk_buff *skb = NULL;
2690         unsigned long nr_pages, limit;
2691         int order, i, max_share;
2692
2693         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
2694
2695         tcp_hashinfo.bind_bucket_cachep =
2696                 kmem_cache_create("tcp_bind_bucket",
2697                                   sizeof(struct inet_bind_bucket), 0,
2698                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
2699
2700         /* Size and allocate the main established and bind bucket
2701          * hash tables.
2702          *
2703          * The methodology is similar to that of the buffer cache.
2704          */
2705         tcp_hashinfo.ehash =
2706                 alloc_large_system_hash("TCP established",
2707                                         sizeof(struct inet_ehash_bucket),
2708                                         thash_entries,
2709                                         (num_physpages >= 128 * 1024) ?
2710                                         13 : 15,
2711                                         0,
2712                                         &tcp_hashinfo.ehash_size,
2713                                         NULL,
2714                                         thash_entries ? 0 : 512 * 1024);
2715         tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
2716         for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
2717                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].chain);
2718                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].twchain);
2719         }
2720         if (inet_ehash_locks_alloc(&tcp_hashinfo))
2721                 panic("TCP: failed to alloc ehash_locks");
2722         tcp_hashinfo.bhash =
2723                 alloc_large_system_hash("TCP bind",
2724                                         sizeof(struct inet_bind_hashbucket),
2725                                         tcp_hashinfo.ehash_size,
2726                                         (num_physpages >= 128 * 1024) ?
2727                                         13 : 15,
2728                                         0,
2729                                         &tcp_hashinfo.bhash_size,
2730                                         NULL,
2731                                         64 * 1024);
2732         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2733         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2734                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2735                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2736         }
2737
2738         /* Try to be a bit smarter and adjust defaults depending
2739          * on available memory.
2740          */
2741         for (order = 0; ((1 << order) << PAGE_SHIFT) <
2742                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2743                         order++)
2744                 ;
2745         if (order >= 4) {
2746                 tcp_death_row.sysctl_max_tw_buckets = 180000;
2747                 sysctl_tcp_max_orphans = 4096 << (order - 4);
2748                 sysctl_max_syn_backlog = 1024;
2749         } else if (order < 3) {
2750                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2751                 sysctl_tcp_max_orphans >>= (3 - order);
2752                 sysctl_max_syn_backlog = 128;
2753         }
2754
2755         /* Set the pressure threshold to be a fraction of global memory that
2756          * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2757          * memory, with a floor of 128 pages.
2758          */
2759         nr_pages = totalram_pages - totalhigh_pages;
2760         limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
2761         limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
2762         limit = max(limit, 128UL);
2763         sysctl_tcp_mem[0] = limit / 4 * 3;
2764         sysctl_tcp_mem[1] = limit;
2765         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2766
2767         /* Set per-socket limits to no more than 1/128 the pressure threshold */
2768         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2769         max_share = min(4UL*1024*1024, limit);
2770
2771         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
2772         sysctl_tcp_wmem[1] = 16*1024;
2773         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2774
2775         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
2776         sysctl_tcp_rmem[1] = 87380;
2777         sysctl_tcp_rmem[2] = max(87380, max_share);
2778
2779         printk(KERN_INFO "TCP: Hash tables configured "
2780                "(established %d bind %d)\n",
2781                tcp_hashinfo.ehash_size, tcp_hashinfo.bhash_size);
2782
2783         tcp_register_congestion_control(&tcp_reno);
2784 }
2785
2786 EXPORT_SYMBOL(tcp_close);
2787 EXPORT_SYMBOL(tcp_disconnect);
2788 EXPORT_SYMBOL(tcp_getsockopt);
2789 EXPORT_SYMBOL(tcp_ioctl);
2790 EXPORT_SYMBOL(tcp_poll);
2791 EXPORT_SYMBOL(tcp_read_sock);
2792 EXPORT_SYMBOL(tcp_recvmsg);
2793 EXPORT_SYMBOL(tcp_sendmsg);
2794 EXPORT_SYMBOL(tcp_splice_read);
2795 EXPORT_SYMBOL(tcp_sendpage);
2796 EXPORT_SYMBOL(tcp_setsockopt);
2797 EXPORT_SYMBOL(tcp_shutdown);