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