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  * Version:     $Id: tcp.c,v 1.216 2002/02/01 22:01:04 davem Exp $
   9  *
  10  * Authors:     Ross Biro
  11  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  12  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
  13  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
  14  *              Florian La Roche, <flla@stud.uni-sb.de>
  15  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  16  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
  17  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
  18  *              Matthew Dillon, <dillon@apollo.west.oic.com>
  19  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  20  *              Jorge Cwik, <jorge@laser.satlink.net>
  21  *
  22  * Fixes:
  23  *              Alan Cox        :       Numerous verify_area() calls
  24  *              Alan Cox        :       Set the ACK bit on a reset
  25  *              Alan Cox        :       Stopped it crashing if it closed while
  26  *                                      sk->inuse=1 and was trying to connect
  27  *                                      (tcp_err()).
  28  *              Alan Cox        :       All icmp error handling was broken
  29  *                                      pointers passed where wrong and the
  30  *                                      socket was looked up backwards. Nobody
  31  *                                      tested any icmp error code obviously.
  32  *              Alan Cox        :       tcp_err() now handled properly. It
  33  *                                      wakes people on errors. poll
  34  *                                      behaves and the icmp error race
  35  *                                      has gone by moving it into sock.c
  36  *              Alan Cox        :       tcp_send_reset() fixed to work for
  37  *                                      everything not just packets for
  38  *                                      unknown sockets.
  39  *              Alan Cox        :       tcp option processing.
  40  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
  41  *                                      syn rule wrong]
  42  *              Herp Rosmanith  :       More reset fixes
  43  *              Alan Cox        :       No longer acks invalid rst frames.
  44  *                                      Acking any kind of RST is right out.
  45  *              Alan Cox        :       Sets an ignore me flag on an rst
  46  *                                      receive otherwise odd bits of prattle
  47  *                                      escape still
  48  *              Alan Cox        :       Fixed another acking RST frame bug.
  49  *                                      Should stop LAN workplace lockups.
  50  *              Alan Cox        :       Some tidyups using the new skb list
  51  *                                      facilities
  52  *              Alan Cox        :       sk->keepopen now seems to work
  53  *              Alan Cox        :       Pulls options out correctly on accepts
  54  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
  55  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
  56  *                                      bit to skb ops.
  57  *              Alan Cox        :       Tidied tcp_data to avoid a potential
  58  *                                      nasty.
  59  *              Alan Cox        :       Added some better commenting, as the
  60  *                                      tcp is hard to follow
  61  *              Alan Cox        :       Removed incorrect check for 20 * psh
  62  *      Michael O'Reilly        :       ack < copied bug fix.
  63  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
  64  *              Alan Cox        :       FIN with no memory -> CRASH
  65  *              Alan Cox        :       Added socket option proto entries.
  66  *                                      Also added awareness of them to accept.
  67  *              Alan Cox        :       Added TCP options (SOL_TCP)
  68  *              Alan Cox        :       Switched wakeup calls to callbacks,
  69  *                                      so the kernel can layer network
  70  *                                      sockets.
  71  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
  72  *              Alan Cox        :       Handle FIN (more) properly (we hope).
  73  *              Alan Cox        :       RST frames sent on unsynchronised
  74  *                                      state ack error.
  75  *              Alan Cox        :       Put in missing check for SYN bit.
  76  *              Alan Cox        :       Added tcp_select_window() aka NET2E
  77  *                                      window non shrink trick.
  78  *              Alan Cox        :       Added a couple of small NET2E timer
  79  *                                      fixes
  80  *              Charles Hedrick :       TCP fixes
  81  *              Toomas Tamm     :       TCP window fixes
  82  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
  83  *              Charles Hedrick :       Rewrote most of it to actually work
  84  *              Linus           :       Rewrote tcp_read() and URG handling
  85  *                                      completely
  86  *              Gerhard Koerting:       Fixed some missing timer handling
  87  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
  88  *              Gerhard Koerting:       PC/TCP workarounds
  89  *              Adam Caldwell   :       Assorted timer/timing errors
  90  *              Matthew Dillon  :       Fixed another RST bug
  91  *              Alan Cox        :       Move to kernel side addressing changes.
  92  *              Alan Cox        :       Beginning work on TCP fastpathing
  93  *                                      (not yet usable)
  94  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
  95  *              Alan Cox        :       TCP fast path debugging
  96  *              Alan Cox        :       Window clamping
  97  *              Michael Riepe   :       Bug in tcp_check()
  98  *              Matt Dillon     :       More TCP improvements and RST bug fixes
  99  *              Matt Dillon     :       Yet more small nasties remove from the
 100  *                                      TCP code (Be very nice to this man if
 101  *                                      tcp finally works 100%) 8)
 102  *              Alan Cox        :       BSD accept semantics.
 103  *              Alan Cox        :       Reset on closedown bug.
 104  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
 105  *              Michael Pall    :       Handle poll() after URG properly in
 106  *                                      all cases.
 107  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
 108  *                                      (multi URG PUSH broke rlogin).
 109  *              Michael Pall    :       Fix the multi URG PUSH problem in
 110  *                                      tcp_readable(), poll() after URG
 111  *                                      works now.
 112  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
 113  *                                      BSD api.
 114  *              Alan Cox        :       Changed the semantics of sk->socket to
 115  *                                      fix a race and a signal problem with
 116  *                                      accept() and async I/O.
 117  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
 118  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
 119  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
 120  *                                      clients/servers which listen in on
 121  *                                      fixed ports.
 122  *              Alan Cox        :       Cleaned the above up and shrank it to
 123  *                                      a sensible code size.
 124  *              Alan Cox        :       Self connect lockup fix.
 125  *              Alan Cox        :       No connect to multicast.
 126  *              Ross Biro       :       Close unaccepted children on master
 127  *                                      socket close.
 128  *              Alan Cox        :       Reset tracing code.
 129  *              Alan Cox        :       Spurious resets on shutdown.
 130  *              Alan Cox        :       Giant 15 minute/60 second timer error
 131  *              Alan Cox        :       Small whoops in polling before an
 132  *                                      accept.
 133  *              Alan Cox        :       Kept the state trace facility since
 134  *                                      it's handy for debugging.
 135  *              Alan Cox        :       More reset handler fixes.
 136  *              Alan Cox        :       Started rewriting the code based on
 137  *                                      the RFC's for other useful protocol
 138  *                                      references see: Comer, KA9Q NOS, and
 139  *                                      for a reference on the difference
 140  *                                      between specifications and how BSD
 141  *                                      works see the 4.4lite source.
 142  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
 143  *                                      close.
 144  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
 145  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
 146  *              Alan Cox        :       Reimplemented timers as per the RFC
 147  *                                      and using multiple timers for sanity.
 148  *              Alan Cox        :       Small bug fixes, and a lot of new
 149  *                                      comments.
 150  *              Alan Cox        :       Fixed dual reader crash by locking
 151  *                                      the buffers (much like datagram.c)
 152  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
 153  *                                      now gets fed up of retrying without
 154  *                                      (even a no space) answer.
 155  *              Alan Cox        :       Extracted closing code better
 156  *              Alan Cox        :       Fixed the closing state machine to
 157  *                                      resemble the RFC.
 158  *              Alan Cox        :       More 'per spec' fixes.
 159  *              Jorge Cwik      :       Even faster checksumming.
 160  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
 161  *                                      only frames. At least one pc tcp stack
 162  *                                      generates them.
 163  *              Alan Cox        :       Cache last socket.
 164  *              Alan Cox        :       Per route irtt.
 165  *              Matt Day        :       poll()->select() match BSD precisely on error
 166  *              Alan Cox        :       New buffers
 167  *              Marc Tamsky     :       Various sk->prot->retransmits and
 168  *                                      sk->retransmits misupdating fixed.
 169  *                                      Fixed tcp_write_timeout: stuck close,
 170  *                                      and TCP syn retries gets used now.
 171  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
 172  *                                      ack if state is TCP_CLOSED.
 173  *              Alan Cox        :       Look up device on a retransmit - routes may
 174  *                                      change. Doesn't yet cope with MSS shrink right
 175  *                                      but it's a start!
 176  *              Marc Tamsky     :       Closing in closing fixes.
 177  *              Mike Shaver     :       RFC1122 verifications.
 178  *              Alan Cox        :       rcv_saddr errors.
 179  *              Alan Cox        :       Block double connect().
 180  *              Alan Cox        :       Small hooks for enSKIP.
 181  *              Alexey Kuznetsov:       Path MTU discovery.
 182  *              Alan Cox        :       Support soft errors.
 183  *              Alan Cox        :       Fix MTU discovery pathological case
 184  *                                      when the remote claims no mtu!
 185  *              Marc Tamsky     :       TCP_CLOSE fix.
 186  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
 187  *                                      window but wrong (fixes NT lpd problems)
 188  *              Pedro Roque     :       Better TCP window handling, delayed ack.
 189  *              Joerg Reuter    :       No modification of locked buffers in
 190  *                                      tcp_do_retransmit()
 191  *              Eric Schenk     :       Changed receiver side silly window
 192  *                                      avoidance algorithm to BSD style
 193  *                                      algorithm. This doubles throughput
 194  *                                      against machines running Solaris,
 195  *                                      and seems to result in general
 196  *                                      improvement.
 197  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
 198  *      Willy Konynenberg       :       Transparent proxying support.
 199  *      Mike McLagan            :       Routing by source
 200  *              Keith Owens     :       Do proper merging with partial SKB's in
 201  *                                      tcp_do_sendmsg to avoid burstiness.
 202  *              Eric Schenk     :       Fix fast close down bug with
 203  *                                      shutdown() followed by close().
 204  *              Andi Kleen      :       Make poll agree with SIGIO
 205  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
 206  *                                      lingertime == 0 (RFC 793 ABORT Call)
 207  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
 208  *                                      csum_and_copy_from_user() if possible.
 209  *
 210  *              This program is free software; you can redistribute it and/or
 211  *              modify it under the terms of the GNU General Public License
 212  *              as published by the Free Software Foundation; either version
 213  *              2 of the License, or(at your option) any later version.
 214  *
 215  * Description of States:
 216  *
 217  *      TCP_SYN_SENT            sent a connection request, waiting for ack
 218  *
 219  *      TCP_SYN_RECV            received a connection request, sent ack,
 220  *                              waiting for final ack in three-way handshake.
 221  *
 222  *      TCP_ESTABLISHED         connection established
 223  *
 224  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
 225  *                              transmission of remaining buffered data
 226  *
 227  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
 228  *                              to shutdown
 229  *
 230  *      TCP_CLOSING             both sides have shutdown but we still have
 231  *                              data we have to finish sending
 232  *
 233  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
 234  *                              closed, can only be entered from FIN_WAIT2
 235  *                              or CLOSING.  Required because the other end
 236  *                              may not have gotten our last ACK causing it
 237  *                              to retransmit the data packet (which we ignore)
 238  *
 239  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
 240  *                              us to finish writing our data and to shutdown
 241  *                              (we have to close() to move on to LAST_ACK)
 242  *
 243  *      TCP_LAST_ACK            out side has shutdown after remote has
 244  *                              shutdown.  There may still be data in our
 245  *                              buffer that we have to finish sending
 246  *
 247  *      TCP_CLOSE               socket is finished
 248  */
 249
 250 #include <linux/kernel.h>
 251 #include <linux/module.h>
 252 #include <linux/types.h>
 253 #include <linux/fcntl.h>
 254 #include <linux/poll.h>
 255 #include <linux/init.h>
 256 #include <linux/fs.h>
 257 #include <linux/skbuff.h>
 258 #include <linux/splice.h>
 259 #include <linux/net.h>
 260 #include <linux/socket.h>
 261 #include <linux/random.h>
 262 #include <linux/bootmem.h>
 263 #include <linux/highmem.h>
 264 #include <linux/swap.h>
 265 #include <linux/cache.h>
 266 #include <linux/err.h>
 267 #include <linux/crypto.h>
 268
 269 #include <net/icmp.h>
 270 #include <net/tcp.h>
 271 #include <net/xfrm.h>
 272 #include <net/ip.h>
 273 #include <net/netdma.h>
 274 #include <net/sock.h>
 275
 276 #include <asm/uaccess.h>
 277 #include <asm/ioctls.h>
 278
 279 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
 280
 281 DEFINE_SNMP_STAT(struct tcp_mib, tcp_statistics) __read_mostly;
 282
 283 atomic_t tcp_orphan_count = ATOMIC_INIT(0);
 284
 285 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 286
 287 int sysctl_tcp_mem[3] __read_mostly;
 288 int sysctl_tcp_wmem[3] __read_mostly;
 289 int sysctl_tcp_rmem[3] __read_mostly;
 290
 291 EXPORT_SYMBOL(sysctl_tcp_mem);
 292 EXPORT_SYMBOL(sysctl_tcp_rmem);
 293 EXPORT_SYMBOL(sysctl_tcp_wmem);
 294
 295 atomic_t tcp_memory_allocated;  /* Current allocated memory. */
 296 atomic_t tcp_sockets_allocated; /* Current number of TCP sockets. */
 297
 298 EXPORT_SYMBOL(tcp_memory_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(void)
 321 {
 322         if (!tcp_memory_pressure) {
 323                 NET_INC_STATS(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 another 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 PULLHUP 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 PULLHUP 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                 /* Potential race condition. If read of tp below will
 391                  * escape above sk->sk_state, we can be illegally awaken
 392                  * in SYN_* states. */
 393                 if ((tp->rcv_nxt != tp->copied_seq) &&
 394                     (tp->urg_seq != tp->copied_seq ||
 395                      tp->rcv_nxt != tp->copied_seq + 1 ||
 396                      sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data))
 397                         mask |= POLLIN | POLLRDNORM;
 398
 399                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 400                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
 401                                 mask |= POLLOUT | POLLWRNORM;
 402                         } else {  /* send SIGIO later */
 403                                 set_bit(SOCK_ASYNC_NOSPACE,
 404                                         &sk->sk_socket->flags);
 405                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 406
 407                                 /* Race breaker. If space is freed after
 408                                  * wspace test but before the flags are set,
 409                                  * IO signal will be lost.
 410                                  */
 411                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
 412                                         mask |= POLLOUT | POLLWRNORM;
 413                         }
 414                 }
 415
 416                 if (tp->urg_data & TCP_URG_VALID)
 417                         mask |= POLLPRI;
 418         }
 419         return mask;
 420 }
 421
 422 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 423 {
 424         struct tcp_sock *tp = tcp_sk(sk);
 425         int answ;
 426
 427         switch (cmd) {
 428         case SIOCINQ:
 429                 if (sk->sk_state == TCP_LISTEN)
 430                         return -EINVAL;
 431
 432                 lock_sock(sk);
 433                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 434                         answ = 0;
 435                 else if (sock_flag(sk, SOCK_URGINLINE) ||
 436                          !tp->urg_data ||
 437                          before(tp->urg_seq, tp->copied_seq) ||
 438                          !before(tp->urg_seq, tp->rcv_nxt)) {
 439                         answ = tp->rcv_nxt - tp->copied_seq;
 440
 441                         /* Subtract 1, if FIN is in queue. */
 442                         if (answ && !skb_queue_empty(&sk->sk_receive_queue))
 443                                 answ -=
 444                        tcp_hdr((struct sk_buff *)sk->sk_receive_queue.prev)->fin;
 445                 } else
 446                         answ = tp->urg_seq - tp->copied_seq;
 447                 release_sock(sk);
 448                 break;
 449         case SIOCATMARK:
 450                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 451                 break;
 452         case SIOCOUTQ:
 453                 if (sk->sk_state == TCP_LISTEN)
 454                         return -EINVAL;
 455
 456                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 457                         answ = 0;
 458                 else
 459                         answ = tp->write_seq - tp->snd_una;
 460                 break;
 461         default:
 462                 return -ENOIOCTLCMD;
 463         }
 464
 465         return put_user(answ, (int __user *)arg);
 466 }
 467
 468 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 469 {
 470         TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
 471         tp->pushed_seq = tp->write_seq;
 472 }
 473
 474 static inline int forced_push(struct tcp_sock *tp)
 475 {
 476         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 477 }
 478
 479 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
 480 {
 481         struct tcp_sock *tp = tcp_sk(sk);
 482         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 483
 484         skb->csum    = 0;
 485         tcb->seq     = tcb->end_seq = tp->write_seq;
 486         tcb->flags   = TCPCB_FLAG_ACK;
 487         tcb->sacked  = 0;
 488         skb_header_release(skb);
 489         tcp_add_write_queue_tail(sk, skb);
 490         sk->sk_wmem_queued += skb->truesize;
 491         sk_mem_charge(sk, skb->truesize);
 492         if (tp->nonagle & TCP_NAGLE_PUSH)
 493                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 494 }
 495
 496 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
 497                                 struct sk_buff *skb)
 498 {
 499         if (flags & MSG_OOB) {
 500                 tp->urg_mode = 1;
 501                 tp->snd_up = tp->write_seq;
 502         }
 503 }
 504
 505 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
 506                             int nonagle)
 507 {
 508         struct tcp_sock *tp = tcp_sk(sk);
 509
 510         if (tcp_send_head(sk)) {
 511                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 512                 if (!(flags & MSG_MORE) || forced_push(tp))
 513                         tcp_mark_push(tp, skb);
 514                 tcp_mark_urg(tp, flags, skb);
 515                 __tcp_push_pending_frames(sk, mss_now,
 516                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
 517         }
 518 }
 519
 520 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 521                                 unsigned int offset, size_t len)
 522 {
 523         struct tcp_splice_state *tss = rd_desc->arg.data;
 524
 525         return skb_splice_bits(skb, offset, tss->pipe, tss->len, tss->flags);
 526 }
 527
 528 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 529 {
 530         /* Store TCP splice context information in read_descriptor_t. */
 531         read_descriptor_t rd_desc = {
 532                 .arg.data = tss,
 533         };
 534
 535         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 536 }
 537
 538 /**
 539  *  tcp_splice_read - splice data from TCP socket to a pipe
 540  * @sock:       socket to splice from
 541  * @ppos:       position (not valid)
 542  * @pipe:       pipe to splice to
 543  * @len:        number of bytes to splice
 544  * @flags:      splice modifier flags
 545  *
 546  * Description:
 547  *    Will read pages from given socket and fill them into a pipe.
 548  *
 549  **/
 550 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 551                         struct pipe_inode_info *pipe, size_t len,
 552                         unsigned int flags)
 553 {
 554         struct sock *sk = sock->sk;
 555         struct tcp_splice_state tss = {
 556                 .pipe = pipe,
 557                 .len = len,
 558                 .flags = flags,
 559         };
 560         long timeo;
 561         ssize_t spliced;
 562         int ret;
 563
 564         /*
 565          * We can't seek on a socket input
 566          */
 567         if (unlikely(*ppos))
 568                 return -ESPIPE;
 569
 570         ret = spliced = 0;
 571
 572         lock_sock(sk);
 573
 574         timeo = sock_rcvtimeo(sk, flags & SPLICE_F_NONBLOCK);
 575         while (tss.len) {
 576                 ret = __tcp_splice_read(sk, &tss);
 577                 if (ret < 0)
 578                         break;
 579                 else if (!ret) {
 580                         if (spliced)
 581                                 break;
 582                         if (flags & SPLICE_F_NONBLOCK) {
 583                                 ret = -EAGAIN;
 584                                 break;
 585                         }
 586                         if (sock_flag(sk, SOCK_DONE))
 587                                 break;
 588                         if (sk->sk_err) {
 589                                 ret = sock_error(sk);
 590                                 break;
 591                         }
 592                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 593                                 break;
 594                         if (sk->sk_state == TCP_CLOSE) {
 595                                 /*
 596                                  * This occurs when user tries to read
 597                                  * from never connected socket.
 598                                  */
 599                                 if (!sock_flag(sk, SOCK_DONE))
 600                                         ret = -ENOTCONN;
 601                                 break;
 602                         }
 603                         if (!timeo) {
 604                                 ret = -EAGAIN;
 605                                 break;
 606                         }
 607                         sk_wait_data(sk, &timeo);
 608                         if (signal_pending(current)) {
 609                                 ret = sock_intr_errno(timeo);
 610                                 break;
 611                         }
 612                         continue;
 613                 }
 614                 tss.len -= ret;
 615                 spliced += ret;
 616
 617                 release_sock(sk);
 618                 lock_sock(sk);
 619
 620                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 621                     (sk->sk_shutdown & RCV_SHUTDOWN) || !timeo ||
 622                     signal_pending(current))
 623                         break;
 624         }
 625
 626         release_sock(sk);
 627
 628         if (spliced)
 629                 return spliced;
 630
 631         return ret;
 632 }
 633
 634 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
 635 {
 636         struct sk_buff *skb;
 637
 638         /* The TCP header must be at least 32-bit aligned.  */
 639         size = ALIGN(size, 4);
 640
 641         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 642         if (skb) {
 643                 if (sk_wmem_schedule(sk, skb->truesize)) {
 644                         /*
 645                          * Make sure that we have exactly size bytes
 646                          * available to the caller, no more, no less.
 647                          */
 648                         skb_reserve(skb, skb_tailroom(skb) - size);
 649                         return skb;
 650                 }
 651                 __kfree_skb(skb);
 652         } else {
 653                 sk->sk_prot->enter_memory_pressure();
 654                 sk_stream_moderate_sndbuf(sk);
 655         }
 656         return NULL;
 657 }
 658
 659 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
 660                          size_t psize, int flags)
 661 {
 662         struct tcp_sock *tp = tcp_sk(sk);
 663         int mss_now, size_goal;
 664         int err;
 665         ssize_t copied;
 666         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 667
 668         /* Wait for a connection to finish. */
 669         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 670                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 671                         goto out_err;
 672
 673         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 674
 675         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 676         size_goal = tp->xmit_size_goal;
 677         copied = 0;
 678
 679         err = -EPIPE;
 680         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 681                 goto do_error;
 682
 683         while (psize > 0) {
 684                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 685                 struct page *page = pages[poffset / PAGE_SIZE];
 686                 int copy, i, can_coalesce;
 687                 int offset = poffset % PAGE_SIZE;
 688                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
 689
 690                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
 691 new_segment:
 692                         if (!sk_stream_memory_free(sk))
 693                                 goto wait_for_sndbuf;
 694
 695                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
 696                         if (!skb)
 697                                 goto wait_for_memory;
 698
 699                         skb_entail(sk, skb);
 700                         copy = size_goal;
 701                 }
 702
 703                 if (copy > size)
 704                         copy = size;
 705
 706                 i = skb_shinfo(skb)->nr_frags;
 707                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 708                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
 709                         tcp_mark_push(tp, skb);
 710                         goto new_segment;
 711                 }
 712                 if (!sk_wmem_schedule(sk, copy))
 713                         goto wait_for_memory;
 714
 715                 if (can_coalesce) {
 716                         skb_shinfo(skb)->frags[i - 1].size += copy;
 717                 } else {
 718                         get_page(page);
 719                         skb_fill_page_desc(skb, i, page, offset, copy);
 720                 }
 721
 722                 skb->len += copy;
 723                 skb->data_len += copy;
 724                 skb->truesize += copy;
 725                 sk->sk_wmem_queued += copy;
 726                 sk_mem_charge(sk, copy);
 727                 skb->ip_summed = CHECKSUM_PARTIAL;
 728                 tp->write_seq += copy;
 729                 TCP_SKB_CB(skb)->end_seq += copy;
 730                 skb_shinfo(skb)->gso_segs = 0;
 731
 732                 if (!copied)
 733                         TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
 734
 735                 copied += copy;
 736                 poffset += copy;
 737                 if (!(psize -= copy))
 738                         goto out;
 739
 740                 if (skb->len < size_goal || (flags & MSG_OOB))
 741                         continue;
 742
 743                 if (forced_push(tp)) {
 744                         tcp_mark_push(tp, skb);
 745                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
 746                 } else if (skb == tcp_send_head(sk))
 747                         tcp_push_one(sk, mss_now);
 748                 continue;
 749
 750 wait_for_sndbuf:
 751                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 752 wait_for_memory:
 753                 if (copied)
 754                         tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
 755
 756                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 757                         goto do_error;
 758
 759                 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 760                 size_goal = tp->xmit_size_goal;
 761         }
 762
 763 out:
 764         if (copied)
 765                 tcp_push(sk, flags, mss_now, tp->nonagle);
 766         return copied;
 767
 768 do_error:
 769         if (copied)
 770                 goto out;
 771 out_err:
 772         return sk_stream_error(sk, flags, err);
 773 }
 774
 775 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
 776                      size_t size, int flags)
 777 {
 778         ssize_t res;
 779         struct sock *sk = sock->sk;
 780
 781         if (!(sk->sk_route_caps & NETIF_F_SG) ||
 782             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
 783                 return sock_no_sendpage(sock, page, offset, size, flags);
 784
 785         lock_sock(sk);
 786         TCP_CHECK_TIMER(sk);
 787         res = do_tcp_sendpages(sk, &page, offset, size, flags);
 788         TCP_CHECK_TIMER(sk);
 789         release_sock(sk);
 790         return res;
 791 }
 792
 793 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
 794 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
 795
 796 static inline int select_size(struct sock *sk)
 797 {
 798         struct tcp_sock *tp = tcp_sk(sk);
 799         int tmp = tp->mss_cache;
 800
 801         if (sk->sk_route_caps & NETIF_F_SG) {
 802                 if (sk_can_gso(sk))
 803                         tmp = 0;
 804                 else {
 805                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
 806
 807                         if (tmp >= pgbreak &&
 808                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
 809                                 tmp = pgbreak;
 810                 }
 811         }
 812
 813         return tmp;
 814 }
 815
 816 int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
 817                 size_t size)
 818 {
 819         struct sock *sk = sock->sk;
 820         struct iovec *iov;
 821         struct tcp_sock *tp = tcp_sk(sk);
 822         struct sk_buff *skb;
 823         int iovlen, flags;
 824         int mss_now, size_goal;
 825         int err, copied;
 826         long timeo;
 827
 828         lock_sock(sk);
 829         TCP_CHECK_TIMER(sk);
 830
 831         flags = msg->msg_flags;
 832         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 833
 834         /* Wait for a connection to finish. */
 835         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 836                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 837                         goto out_err;
 838
 839         /* This should be in poll */
 840         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 841
 842         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 843         size_goal = tp->xmit_size_goal;
 844
 845         /* Ok commence sending. */
 846         iovlen = msg->msg_iovlen;
 847         iov = msg->msg_iov;
 848         copied = 0;
 849
 850         err = -EPIPE;
 851         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 852                 goto do_error;
 853
 854         while (--iovlen >= 0) {
 855                 int seglen = iov->iov_len;
 856                 unsigned char __user *from = iov->iov_base;
 857
 858                 iov++;
 859
 860                 while (seglen > 0) {
 861                         int copy;
 862
 863                         skb = tcp_write_queue_tail(sk);
 864
 865                         if (!tcp_send_head(sk) ||
 866                             (copy = size_goal - skb->len) <= 0) {
 867
 868 new_segment:
 869                                 /* Allocate new segment. If the interface is SG,
 870                                  * allocate skb fitting to single page.
 871                                  */
 872                                 if (!sk_stream_memory_free(sk))
 873                                         goto wait_for_sndbuf;
 874
 875                                 skb = sk_stream_alloc_skb(sk, select_size(sk),
 876                                                 sk->sk_allocation);
 877                                 if (!skb)
 878                                         goto wait_for_memory;
 879
 880                                 /*
 881                                  * Check whether we can use HW checksum.
 882                                  */
 883                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
 884                                         skb->ip_summed = CHECKSUM_PARTIAL;
 885
 886                                 skb_entail(sk, skb);
 887                                 copy = size_goal;
 888                         }
 889
 890                         /* Try to append data to the end of skb. */
 891                         if (copy > seglen)
 892                                 copy = seglen;
 893
 894                         /* Where to copy to? */
 895                         if (skb_tailroom(skb) > 0) {
 896                                 /* We have some space in skb head. Superb! */
 897                                 if (copy > skb_tailroom(skb))
 898                                         copy = skb_tailroom(skb);
 899                                 if ((err = skb_add_data(skb, from, copy)) != 0)
 900                                         goto do_fault;
 901                         } else {
 902                                 int merge = 0;
 903                                 int i = skb_shinfo(skb)->nr_frags;
 904                                 struct page *page = TCP_PAGE(sk);
 905                                 int off = TCP_OFF(sk);
 906
 907                                 if (skb_can_coalesce(skb, i, page, off) &&
 908                                     off != PAGE_SIZE) {
 909                                         /* We can extend the last page
 910                                          * fragment. */
 911                                         merge = 1;
 912                                 } else if (i == MAX_SKB_FRAGS ||
 913                                            (!i &&
 914                                            !(sk->sk_route_caps & NETIF_F_SG))) {
 915                                         /* Need to add new fragment and cannot
 916                                          * do this because interface is non-SG,
 917                                          * or because all the page slots are
 918                                          * busy. */
 919                                         tcp_mark_push(tp, skb);
 920                                         goto new_segment;
 921                                 } else if (page) {
 922                                         if (off == PAGE_SIZE) {
 923                                                 put_page(page);
 924                                                 TCP_PAGE(sk) = page = NULL;
 925                                                 off = 0;
 926                                         }
 927                                 } else
 928                                         off = 0;
 929
 930                                 if (copy > PAGE_SIZE - off)
 931                                         copy = PAGE_SIZE - off;
 932
 933                                 if (!sk_wmem_schedule(sk, copy))
 934                                         goto wait_for_memory;
 935
 936                                 if (!page) {
 937                                         /* Allocate new cache page. */
 938                                         if (!(page = sk_stream_alloc_page(sk)))
 939                                                 goto wait_for_memory;
 940                                 }
 941
 942                                 /* Time to copy data. We are close to
 943                                  * the end! */
 944                                 err = skb_copy_to_page(sk, from, skb, page,
 945                                                        off, copy);
 946                                 if (err) {
 947                                         /* If this page was new, give it to the
 948                                          * socket so it does not get leaked.
 949                                          */
 950                                         if (!TCP_PAGE(sk)) {
 951                                                 TCP_PAGE(sk) = page;
 952                                                 TCP_OFF(sk) = 0;
 953                                         }
 954                                         goto do_error;
 955                                 }
 956
 957                                 /* Update the skb. */
 958                                 if (merge) {
 959                                         skb_shinfo(skb)->frags[i - 1].size +=
 960                                                                         copy;
 961                                 } else {
 962                                         skb_fill_page_desc(skb, i, page, off, copy);
 963                                         if (TCP_PAGE(sk)) {
 964                                                 get_page(page);
 965                                         } else if (off + copy < PAGE_SIZE) {
 966                                                 get_page(page);
 967                                                 TCP_PAGE(sk) = page;
 968                                         }
 969                                 }
 970
 971                                 TCP_OFF(sk) = off + copy;
 972                         }
 973
 974                         if (!copied)
 975                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
 976
 977                         tp->write_seq += copy;
 978                         TCP_SKB_CB(skb)->end_seq += copy;
 979                         skb_shinfo(skb)->gso_segs = 0;
 980
 981                         from += copy;
 982                         copied += copy;
 983                         if ((seglen -= copy) == 0 && iovlen == 0)
 984                                 goto out;
 985
 986                         if (skb->len < size_goal || (flags & MSG_OOB))
 987                                 continue;
 988
 989                         if (forced_push(tp)) {
 990                                 tcp_mark_push(tp, skb);
 991                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
 992                         } else if (skb == tcp_send_head(sk))
 993                                 tcp_push_one(sk, mss_now);
 994                         continue;
 995
 996 wait_for_sndbuf:
 997                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 998 wait_for_memory:
 999                         if (copied)
1000                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1001
1002                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1003                                 goto do_error;
1004
1005                         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
1006                         size_goal = tp->xmit_size_goal;
1007                 }
1008         }
1009
1010 out:
1011         if (copied)
1012                 tcp_push(sk, flags, mss_now, tp->nonagle);
1013         TCP_CHECK_TIMER(sk);
1014         release_sock(sk);
1015         return copied;
1016
1017 do_fault:
1018         if (!skb->len) {
1019                 tcp_unlink_write_queue(skb, sk);
1020                 /* It is the one place in all of TCP, except connection
1021                  * reset, where we can be unlinking the send_head.
1022                  */
1023                 tcp_check_send_head(sk, skb);
1024                 sk_wmem_free_skb(sk, skb);
1025         }
1026
1027 do_error:
1028         if (copied)
1029                 goto out;
1030 out_err:
1031         err = sk_stream_error(sk, flags, err);
1032         TCP_CHECK_TIMER(sk);
1033         release_sock(sk);
1034         return err;
1035 }
1036
1037 /*
1038  *      Handle reading urgent data. BSD has very simple semantics for
1039  *      this, no blocking and very strange errors 8)
1040  */
1041
1042 static int tcp_recv_urg(struct sock *sk, long timeo,
1043                         struct msghdr *msg, int len, int flags,
1044                         int *addr_len)
1045 {
1046         struct tcp_sock *tp = tcp_sk(sk);
1047
1048         /* No URG data to read. */
1049         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1050             tp->urg_data == TCP_URG_READ)
1051                 return -EINVAL; /* Yes this is right ! */
1052
1053         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1054                 return -ENOTCONN;
1055
1056         if (tp->urg_data & TCP_URG_VALID) {
1057                 int err = 0;
1058                 char c = tp->urg_data;
1059
1060                 if (!(flags & MSG_PEEK))
1061                         tp->urg_data = TCP_URG_READ;
1062
1063                 /* Read urgent data. */
1064                 msg->msg_flags |= MSG_OOB;
1065
1066                 if (len > 0) {
1067                         if (!(flags & MSG_TRUNC))
1068                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1069                         len = 1;
1070                 } else
1071                         msg->msg_flags |= MSG_TRUNC;
1072
1073                 return err ? -EFAULT : len;
1074         }
1075
1076         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1077                 return 0;
1078
1079         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1080          * the available implementations agree in this case:
1081          * this call should never block, independent of the
1082          * blocking state of the socket.
1083          * Mike <pall@rz.uni-karlsruhe.de>
1084          */
1085         return -EAGAIN;
1086 }
1087
1088 /* Clean up the receive buffer for full frames taken by the user,
1089  * then send an ACK if necessary.  COPIED is the number of bytes
1090  * tcp_recvmsg has given to the user so far, it speeds up the
1091  * calculation of whether or not we must ACK for the sake of
1092  * a window update.
1093  */
1094 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1095 {
1096         struct tcp_sock *tp = tcp_sk(sk);
1097         int time_to_ack = 0;
1098
1099 #if TCP_DEBUG
1100         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1101
1102         BUG_TRAP(!skb || before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
1103 #endif
1104
1105         if (inet_csk_ack_scheduled(sk)) {
1106                 const struct inet_connection_sock *icsk = inet_csk(sk);
1107                    /* Delayed ACKs frequently hit locked sockets during bulk
1108                     * receive. */
1109                 if (icsk->icsk_ack.blocked ||
1110                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1111                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1112                     /*
1113                      * If this read emptied read buffer, we send ACK, if
1114                      * connection is not bidirectional, user drained
1115                      * receive buffer and there was a small segment
1116                      * in queue.
1117                      */
1118                     (copied > 0 &&
1119                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1120                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1121                        !icsk->icsk_ack.pingpong)) &&
1122                       !atomic_read(&sk->sk_rmem_alloc)))
1123                         time_to_ack = 1;
1124         }
1125
1126         /* We send an ACK if we can now advertise a non-zero window
1127          * which has been raised "significantly".
1128          *
1129          * Even if window raised up to infinity, do not send window open ACK
1130          * in states, where we will not receive more. It is useless.
1131          */
1132         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1133                 __u32 rcv_window_now = tcp_receive_window(tp);
1134
1135                 /* Optimize, __tcp_select_window() is not cheap. */
1136                 if (2*rcv_window_now <= tp->window_clamp) {
1137                         __u32 new_window = __tcp_select_window(sk);
1138
1139                         /* Send ACK now, if this read freed lots of space
1140                          * in our buffer. Certainly, new_window is new window.
1141                          * We can advertise it now, if it is not less than current one.
1142                          * "Lots" means "at least twice" here.
1143                          */
1144                         if (new_window && new_window >= 2 * rcv_window_now)
1145                                 time_to_ack = 1;
1146                 }
1147         }
1148         if (time_to_ack)
1149                 tcp_send_ack(sk);
1150 }
1151
1152 static void tcp_prequeue_process(struct sock *sk)
1153 {
1154         struct sk_buff *skb;
1155         struct tcp_sock *tp = tcp_sk(sk);
1156
1157         NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED);
1158
1159         /* RX process wants to run with disabled BHs, though it is not
1160          * necessary */
1161         local_bh_disable();
1162         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1163                 sk->sk_backlog_rcv(sk, skb);
1164         local_bh_enable();
1165
1166         /* Clear memory counter. */
1167         tp->ucopy.memory = 0;
1168 }
1169
1170 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1171 {
1172         struct sk_buff *skb;
1173         u32 offset;
1174
1175         skb_queue_walk(&sk->sk_receive_queue, skb) {
1176                 offset = seq - TCP_SKB_CB(skb)->seq;
1177                 if (tcp_hdr(skb)->syn)
1178                         offset--;
1179                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1180                         *off = offset;
1181                         return skb;
1182                 }
1183         }
1184         return NULL;
1185 }
1186
1187 /*
1188  * This routine provides an alternative to tcp_recvmsg() for routines
1189  * that would like to handle copying from skbuffs directly in 'sendfile'
1190  * fashion.
1191  * Note:
1192  *      - It is assumed that the socket was locked by the caller.
1193  *      - The routine does not block.
1194  *      - At present, there is no support for reading OOB data
1195  *        or for 'peeking' the socket using this routine
1196  *        (although both would be easy to implement).
1197  */
1198 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1199                   sk_read_actor_t recv_actor)
1200 {
1201         struct sk_buff *skb;
1202         struct tcp_sock *tp = tcp_sk(sk);
1203         u32 seq = tp->copied_seq;
1204         u32 offset;
1205         int copied = 0;
1206
1207         if (sk->sk_state == TCP_LISTEN)
1208                 return -ENOTCONN;
1209         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1210                 if (offset < skb->len) {
1211                         size_t used, len;
1212
1213                         len = skb->len - offset;
1214                         /* Stop reading if we hit a patch of urgent data */
1215                         if (tp->urg_data) {
1216                                 u32 urg_offset = tp->urg_seq - seq;
1217                                 if (urg_offset < len)
1218                                         len = urg_offset;
1219                                 if (!len)
1220                                         break;
1221                         }
1222                         used = recv_actor(desc, skb, offset, len);
1223                         if (used < 0) {
1224                                 if (!copied)
1225                                         copied = used;
1226                                 break;
1227                         } else if (used <= len) {
1228                                 seq += used;
1229                                 copied += used;
1230                                 offset += used;
1231                         }
1232                         /*
1233                          * If recv_actor drops the lock (e.g. TCP splice
1234                          * receive) the skb pointer might be invalid when
1235                          * getting here: tcp_collapse might have deleted it
1236                          * while aggregating skbs from the socket queue.
1237                          */
1238                         skb = tcp_recv_skb(sk, seq-1, &offset);
1239                         if (!skb || (offset+1 != skb->len))
1240                                 break;
1241                 }
1242                 if (tcp_hdr(skb)->fin) {
1243                         sk_eat_skb(sk, skb, 0);
1244                         ++seq;
1245                         break;
1246                 }
1247                 sk_eat_skb(sk, skb, 0);
1248                 if (!desc->count)
1249                         break;
1250         }
1251         tp->copied_seq = seq;
1252
1253         tcp_rcv_space_adjust(sk);
1254
1255         /* Clean up data we have read: This will do ACK frames. */
1256         if (copied > 0)
1257                 tcp_cleanup_rbuf(sk, copied);
1258         return copied;
1259 }
1260
1261 /*
1262  *      This routine copies from a sock struct into the user buffer.
1263  *
1264  *      Technical note: in 2.3 we work on _locked_ socket, so that
1265  *      tricks with *seq access order and skb->users are not required.
1266  *      Probably, code can be easily improved even more.
1267  */
1268
1269 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1270                 size_t len, int nonblock, int flags, int *addr_len)
1271 {
1272         struct tcp_sock *tp = tcp_sk(sk);
1273         int copied = 0;
1274         u32 peek_seq;
1275         u32 *seq;
1276         unsigned long used;
1277         int err;
1278         int target;             /* Read at least this many bytes */
1279         long timeo;
1280         struct task_struct *user_recv = NULL;
1281         int copied_early = 0;
1282         struct sk_buff *skb;
1283
1284         lock_sock(sk);
1285
1286         TCP_CHECK_TIMER(sk);
1287
1288         err = -ENOTCONN;
1289         if (sk->sk_state == TCP_LISTEN)
1290                 goto out;
1291
1292         timeo = sock_rcvtimeo(sk, nonblock);
1293
1294         /* Urgent data needs to be handled specially. */
1295         if (flags & MSG_OOB)
1296                 goto recv_urg;
1297
1298         seq = &tp->copied_seq;
1299         if (flags & MSG_PEEK) {
1300                 peek_seq = tp->copied_seq;
1301                 seq = &peek_seq;
1302         }
1303
1304         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1305
1306 #ifdef CONFIG_NET_DMA
1307         tp->ucopy.dma_chan = NULL;
1308         preempt_disable();
1309         skb = skb_peek_tail(&sk->sk_receive_queue);
1310         {
1311                 int available = 0;
1312
1313                 if (skb)
1314                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1315                 if ((available < target) &&
1316                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1317                     !sysctl_tcp_low_latency &&
1318                     __get_cpu_var(softnet_data).net_dma) {
1319                         preempt_enable_no_resched();
1320                         tp->ucopy.pinned_list =
1321                                         dma_pin_iovec_pages(msg->msg_iov, len);
1322                 } else {
1323                         preempt_enable_no_resched();
1324                 }
1325         }
1326 #endif
1327
1328         do {
1329                 u32 offset;
1330
1331                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1332                 if (tp->urg_data && tp->urg_seq == *seq) {
1333                         if (copied)
1334                                 break;
1335                         if (signal_pending(current)) {
1336                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1337                                 break;
1338                         }
1339                 }
1340
1341                 /* Next get a buffer. */
1342
1343                 skb = skb_peek(&sk->sk_receive_queue);
1344                 do {
1345                         if (!skb)
1346                                 break;
1347
1348                         /* Now that we have two receive queues this
1349                          * shouldn't happen.
1350                          */
1351                         if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1352                                 printk(KERN_INFO "recvmsg bug: copied %X "
1353                                        "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1354                                 break;
1355                         }
1356                         offset = *seq - TCP_SKB_CB(skb)->seq;
1357                         if (tcp_hdr(skb)->syn)
1358                                 offset--;
1359                         if (offset < skb->len)
1360                                 goto found_ok_skb;
1361                         if (tcp_hdr(skb)->fin)
1362                                 goto found_fin_ok;
1363                         BUG_TRAP(flags & MSG_PEEK);
1364                         skb = skb->next;
1365                 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1366
1367                 /* Well, if we have backlog, try to process it now yet. */
1368
1369                 if (copied >= target && !sk->sk_backlog.tail)
1370                         break;
1371
1372                 if (copied) {
1373                         if (sk->sk_err ||
1374                             sk->sk_state == TCP_CLOSE ||
1375                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1376                             !timeo ||
1377                             signal_pending(current) ||
1378                             (flags & MSG_PEEK))
1379                                 break;
1380                 } else {
1381                         if (sock_flag(sk, SOCK_DONE))
1382                                 break;
1383
1384                         if (sk->sk_err) {
1385                                 copied = sock_error(sk);
1386                                 break;
1387                         }
1388
1389                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1390                                 break;
1391
1392                         if (sk->sk_state == TCP_CLOSE) {
1393                                 if (!sock_flag(sk, SOCK_DONE)) {
1394                                         /* This occurs when user tries to read
1395                                          * from never connected socket.
1396                                          */
1397                                         copied = -ENOTCONN;
1398                                         break;
1399                                 }
1400                                 break;
1401                         }
1402
1403                         if (!timeo) {
1404                                 copied = -EAGAIN;
1405                                 break;
1406                         }
1407
1408                         if (signal_pending(current)) {
1409                                 copied = sock_intr_errno(timeo);
1410                                 break;
1411                         }
1412                 }
1413
1414                 tcp_cleanup_rbuf(sk, copied);
1415
1416                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1417                         /* Install new reader */
1418                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1419                                 user_recv = current;
1420                                 tp->ucopy.task = user_recv;
1421                                 tp->ucopy.iov = msg->msg_iov;
1422                         }
1423
1424                         tp->ucopy.len = len;
1425
1426                         BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1427                                  (flags & (MSG_PEEK | MSG_TRUNC)));
1428
1429                         /* Ugly... If prequeue is not empty, we have to
1430                          * process it before releasing socket, otherwise
1431                          * order will be broken at second iteration.
1432                          * More elegant solution is required!!!
1433                          *
1434                          * Look: we have the following (pseudo)queues:
1435                          *
1436                          * 1. packets in flight
1437                          * 2. backlog
1438                          * 3. prequeue
1439                          * 4. receive_queue
1440                          *
1441                          * Each queue can be processed only if the next ones
1442                          * are empty. At this point we have empty receive_queue.
1443                          * But prequeue _can_ be not empty after 2nd iteration,
1444                          * when we jumped to start of loop because backlog
1445                          * processing added something to receive_queue.
1446                          * We cannot release_sock(), because backlog contains
1447                          * packets arrived _after_ prequeued ones.
1448                          *
1449                          * Shortly, algorithm is clear --- to process all
1450                          * the queues in order. We could make it more directly,
1451                          * requeueing packets from backlog to prequeue, if
1452                          * is not empty. It is more elegant, but eats cycles,
1453                          * unfortunately.
1454                          */
1455                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1456                                 goto do_prequeue;
1457
1458                         /* __ Set realtime policy in scheduler __ */
1459                 }
1460
1461                 if (copied >= target) {
1462                         /* Do not sleep, just process backlog. */
1463                         release_sock(sk);
1464                         lock_sock(sk);
1465                 } else
1466                         sk_wait_data(sk, &timeo);
1467
1468 #ifdef CONFIG_NET_DMA
1469                 tp->ucopy.wakeup = 0;
1470 #endif
1471
1472                 if (user_recv) {
1473                         int chunk;
1474
1475                         /* __ Restore normal policy in scheduler __ */
1476
1477                         if ((chunk = len - tp->ucopy.len) != 0) {
1478                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1479                                 len -= chunk;
1480                                 copied += chunk;
1481                         }
1482
1483                         if (tp->rcv_nxt == tp->copied_seq &&
1484                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1485 do_prequeue:
1486                                 tcp_prequeue_process(sk);
1487
1488                                 if ((chunk = len - tp->ucopy.len) != 0) {
1489                                         NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1490                                         len -= chunk;
1491                                         copied += chunk;
1492                                 }
1493                         }
1494                 }
1495                 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1496                         if (net_ratelimit())
1497                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1498                                        current->comm, task_pid_nr(current));
1499                         peek_seq = tp->copied_seq;
1500                 }
1501                 continue;
1502
1503         found_ok_skb:
1504                 /* Ok so how much can we use? */
1505                 used = skb->len - offset;
1506                 if (len < used)
1507                         used = len;
1508
1509                 /* Do we have urgent data here? */
1510                 if (tp->urg_data) {
1511                         u32 urg_offset = tp->urg_seq - *seq;
1512                         if (urg_offset < used) {
1513                                 if (!urg_offset) {
1514                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1515                                                 ++*seq;
1516                                                 offset++;
1517                                                 used--;
1518                                                 if (!used)
1519                                                         goto skip_copy;
1520                                         }
1521                                 } else
1522                                         used = urg_offset;
1523                         }
1524                 }
1525
1526                 if (!(flags & MSG_TRUNC)) {
1527 #ifdef CONFIG_NET_DMA
1528                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1529                                 tp->ucopy.dma_chan = get_softnet_dma();
1530
1531                         if (tp->ucopy.dma_chan) {
1532                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1533                                         tp->ucopy.dma_chan, skb, offset,
1534                                         msg->msg_iov, used,
1535                                         tp->ucopy.pinned_list);
1536
1537                                 if (tp->ucopy.dma_cookie < 0) {
1538
1539                                         printk(KERN_ALERT "dma_cookie < 0\n");
1540
1541                                         /* Exception. Bailout! */
1542                                         if (!copied)
1543                                                 copied = -EFAULT;
1544                                         break;
1545                                 }
1546                                 if ((offset + used) == skb->len)
1547                                         copied_early = 1;
1548
1549                         } else
1550 #endif
1551                         {
1552                                 err = skb_copy_datagram_iovec(skb, offset,
1553                                                 msg->msg_iov, used);
1554                                 if (err) {
1555                                         /* Exception. Bailout! */
1556                                         if (!copied)
1557                                                 copied = -EFAULT;
1558                                         break;
1559                                 }
1560                         }
1561                 }
1562
1563                 *seq += used;
1564                 copied += used;
1565                 len -= used;
1566
1567                 tcp_rcv_space_adjust(sk);
1568
1569 skip_copy:
1570                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1571                         tp->urg_data = 0;
1572                         tcp_fast_path_check(sk);
1573                 }
1574                 if (used + offset < skb->len)
1575                         continue;
1576
1577                 if (tcp_hdr(skb)->fin)
1578                         goto found_fin_ok;
1579                 if (!(flags & MSG_PEEK)) {
1580                         sk_eat_skb(sk, skb, copied_early);
1581                         copied_early = 0;
1582                 }
1583                 continue;
1584
1585         found_fin_ok:
1586                 /* Process the FIN. */
1587                 ++*seq;
1588                 if (!(flags & MSG_PEEK)) {
1589                         sk_eat_skb(sk, skb, copied_early);
1590                         copied_early = 0;
1591                 }
1592                 break;
1593         } while (len > 0);
1594
1595         if (user_recv) {
1596                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1597                         int chunk;
1598
1599                         tp->ucopy.len = copied > 0 ? len : 0;
1600
1601                         tcp_prequeue_process(sk);
1602
1603                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1604                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1605                                 len -= chunk;
1606                                 copied += chunk;
1607                         }
1608                 }
1609
1610                 tp->ucopy.task = NULL;
1611                 tp->ucopy.len = 0;
1612         }
1613
1614 #ifdef CONFIG_NET_DMA
1615         if (tp->ucopy.dma_chan) {
1616                 dma_cookie_t done, used;
1617
1618                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1619
1620                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1621                                                  tp->ucopy.dma_cookie, &done,
1622                                                  &used) == DMA_IN_PROGRESS) {
1623                         /* do partial cleanup of sk_async_wait_queue */
1624                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1625                                (dma_async_is_complete(skb->dma_cookie, done,
1626                                                       used) == DMA_SUCCESS)) {
1627                                 __skb_dequeue(&sk->sk_async_wait_queue);
1628                                 kfree_skb(skb);
1629                         }
1630                 }
1631
1632                 /* Safe to free early-copied skbs now */
1633                 __skb_queue_purge(&sk->sk_async_wait_queue);
1634                 dma_chan_put(tp->ucopy.dma_chan);
1635                 tp->ucopy.dma_chan = NULL;
1636         }
1637         if (tp->ucopy.pinned_list) {
1638                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1639                 tp->ucopy.pinned_list = NULL;
1640         }
1641 #endif
1642
1643         /* According to UNIX98, msg_name/msg_namelen are ignored
1644          * on connected socket. I was just happy when found this 8) --ANK
1645          */
1646
1647         /* Clean up data we have read: This will do ACK frames. */
1648         tcp_cleanup_rbuf(sk, copied);
1649
1650         TCP_CHECK_TIMER(sk);
1651         release_sock(sk);
1652         return copied;
1653
1654 out:
1655         TCP_CHECK_TIMER(sk);
1656         release_sock(sk);
1657         return err;
1658
1659 recv_urg:
1660         err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1661         goto out;
1662 }
1663
1664 void tcp_set_state(struct sock *sk, int state)
1665 {
1666         int oldstate = sk->sk_state;
1667
1668         switch (state) {
1669         case TCP_ESTABLISHED:
1670                 if (oldstate != TCP_ESTABLISHED)
1671                         TCP_INC_STATS(TCP_MIB_CURRESTAB);
1672                 break;
1673
1674         case TCP_CLOSE:
1675                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1676                         TCP_INC_STATS(TCP_MIB_ESTABRESETS);
1677
1678                 sk->sk_prot->unhash(sk);
1679                 if (inet_csk(sk)->icsk_bind_hash &&
1680                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1681                         inet_put_port(sk);
1682                 /* fall through */
1683         default:
1684                 if (oldstate==TCP_ESTABLISHED)
1685                         TCP_DEC_STATS(TCP_MIB_CURRESTAB);
1686         }
1687
1688         /* Change state AFTER socket is unhashed to avoid closed
1689          * socket sitting in hash tables.
1690          */
1691         sk->sk_state = state;
1692
1693 #ifdef STATE_TRACE
1694         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n",sk, statename[oldstate],statename[state]);
1695 #endif
1696 }
1697 EXPORT_SYMBOL_GPL(tcp_set_state);
1698
1699 /*
1700  *      State processing on a close. This implements the state shift for
1701  *      sending our FIN frame. Note that we only send a FIN for some
1702  *      states. A shutdown() may have already sent the FIN, or we may be
1703  *      closed.
1704  */
1705
1706 static const unsigned char new_state[16] = {
1707   /* current state:        new state:      action:      */
1708   /* (Invalid)          */ TCP_CLOSE,
1709   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1710   /* TCP_SYN_SENT       */ TCP_CLOSE,
1711   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1712   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1713   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1714   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1715   /* TCP_CLOSE          */ TCP_CLOSE,
1716   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1717   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1718   /* TCP_LISTEN         */ TCP_CLOSE,
1719   /* TCP_CLOSING        */ TCP_CLOSING,
1720 };
1721
1722 static int tcp_close_state(struct sock *sk)
1723 {
1724         int next = (int)new_state[sk->sk_state];
1725         int ns = next & TCP_STATE_MASK;
1726
1727         tcp_set_state(sk, ns);
1728
1729         return next & TCP_ACTION_FIN;
1730 }
1731
1732 /*
1733  *      Shutdown the sending side of a connection. Much like close except
1734  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1735  */
1736
1737 void tcp_shutdown(struct sock *sk, int how)
1738 {
1739         /*      We need to grab some memory, and put together a FIN,
1740          *      and then put it into the queue to be sent.
1741          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1742          */
1743         if (!(how & SEND_SHUTDOWN))
1744                 return;
1745
1746         /* If we've already sent a FIN, or it's a closed state, skip this. */
1747         if ((1 << sk->sk_state) &
1748             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1749              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1750                 /* Clear out any half completed packets.  FIN if needed. */
1751                 if (tcp_close_state(sk))
1752                         tcp_send_fin(sk);
1753         }
1754 }
1755
1756 void tcp_close(struct sock *sk, long timeout)
1757 {
1758         struct sk_buff *skb;
1759         int data_was_unread = 0;
1760         int state;
1761
1762         lock_sock(sk);
1763         sk->sk_shutdown = SHUTDOWN_MASK;
1764
1765         if (sk->sk_state == TCP_LISTEN) {
1766                 tcp_set_state(sk, TCP_CLOSE);
1767
1768                 /* Special case. */
1769                 inet_csk_listen_stop(sk);
1770
1771                 goto adjudge_to_death;
1772         }
1773
1774         /*  We need to flush the recv. buffs.  We do this only on the
1775          *  descriptor close, not protocol-sourced closes, because the
1776          *  reader process may not have drained the data yet!
1777          */
1778         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1779                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1780                           tcp_hdr(skb)->fin;
1781                 data_was_unread += len;
1782                 __kfree_skb(skb);
1783         }
1784
1785         sk_mem_reclaim(sk);
1786
1787         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1788          * data was lost. To witness the awful effects of the old behavior of
1789          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1790          * GET in an FTP client, suspend the process, wait for the client to
1791          * advertise a zero window, then kill -9 the FTP client, wheee...
1792          * Note: timeout is always zero in such a case.
1793          */
1794         if (data_was_unread) {
1795                 /* Unread data was tossed, zap the connection. */
1796                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1797                 tcp_set_state(sk, TCP_CLOSE);
1798                 tcp_send_active_reset(sk, GFP_KERNEL);
1799         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1800                 /* Check zero linger _after_ checking for unread data. */
1801                 sk->sk_prot->disconnect(sk, 0);
1802                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1803         } else if (tcp_close_state(sk)) {
1804                 /* We FIN if the application ate all the data before
1805                  * zapping the connection.
1806                  */
1807
1808                 /* RED-PEN. Formally speaking, we have broken TCP state
1809                  * machine. State transitions:
1810                  *
1811                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1812                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1813                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1814                  *
1815                  * are legal only when FIN has been sent (i.e. in window),
1816                  * rather than queued out of window. Purists blame.
1817                  *
1818                  * F.e. "RFC state" is ESTABLISHED,
1819                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1820                  *
1821                  * The visible declinations are that sometimes
1822                  * we enter time-wait state, when it is not required really
1823                  * (harmless), do not send active resets, when they are
1824                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1825                  * they look as CLOSING or LAST_ACK for Linux)
1826                  * Probably, I missed some more holelets.
1827                  *                                              --ANK
1828                  */
1829                 tcp_send_fin(sk);
1830         }
1831
1832         sk_stream_wait_close(sk, timeout);
1833
1834 adjudge_to_death:
1835         state = sk->sk_state;
1836         sock_hold(sk);
1837         sock_orphan(sk);
1838         atomic_inc(sk->sk_prot->orphan_count);
1839
1840         /* It is the last release_sock in its life. It will remove backlog. */
1841         release_sock(sk);
1842
1843
1844         /* Now socket is owned by kernel and we acquire BH lock
1845            to finish close. No need to check for user refs.
1846          */
1847         local_bh_disable();
1848         bh_lock_sock(sk);
1849         BUG_TRAP(!sock_owned_by_user(sk));
1850
1851         /* Have we already been destroyed by a softirq or backlog? */
1852         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1853                 goto out;
1854
1855         /*      This is a (useful) BSD violating of the RFC. There is a
1856          *      problem with TCP as specified in that the other end could
1857          *      keep a socket open forever with no application left this end.
1858          *      We use a 3 minute timeout (about the same as BSD) then kill
1859          *      our end. If they send after that then tough - BUT: long enough
1860          *      that we won't make the old 4*rto = almost no time - whoops
1861          *      reset mistake.
1862          *
1863          *      Nope, it was not mistake. It is really desired behaviour
1864          *      f.e. on http servers, when such sockets are useless, but
1865          *      consume significant resources. Let's do it with special
1866          *      linger2 option.                                 --ANK
1867          */
1868
1869         if (sk->sk_state == TCP_FIN_WAIT2) {
1870                 struct tcp_sock *tp = tcp_sk(sk);
1871                 if (tp->linger2 < 0) {
1872                         tcp_set_state(sk, TCP_CLOSE);
1873                         tcp_send_active_reset(sk, GFP_ATOMIC);
1874                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1875                 } else {
1876                         const int tmo = tcp_fin_time(sk);
1877
1878                         if (tmo > TCP_TIMEWAIT_LEN) {
1879                                 inet_csk_reset_keepalive_timer(sk,
1880                                                 tmo - TCP_TIMEWAIT_LEN);
1881                         } else {
1882                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1883                                 goto out;
1884                         }
1885                 }
1886         }
1887         if (sk->sk_state != TCP_CLOSE) {
1888                 sk_mem_reclaim(sk);
1889                 if (tcp_too_many_orphans(sk,
1890                                 atomic_read(sk->sk_prot->orphan_count))) {
1891                         if (net_ratelimit())
1892                                 printk(KERN_INFO "TCP: too many of orphaned "
1893                                        "sockets\n");
1894                         tcp_set_state(sk, TCP_CLOSE);
1895                         tcp_send_active_reset(sk, GFP_ATOMIC);
1896                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1897                 }
1898         }
1899
1900         if (sk->sk_state == TCP_CLOSE)
1901                 inet_csk_destroy_sock(sk);
1902         /* Otherwise, socket is reprieved until protocol close. */
1903
1904 out:
1905         bh_unlock_sock(sk);
1906         local_bh_enable();
1907         sock_put(sk);
1908 }
1909
1910 /* These states need RST on ABORT according to RFC793 */
1911
1912 static inline int tcp_need_reset(int state)
1913 {
1914         return (1 << state) &
1915                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1916                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1917 }
1918
1919 int tcp_disconnect(struct sock *sk, int flags)
1920 {
1921         struct inet_sock *inet = inet_sk(sk);
1922         struct inet_connection_sock *icsk = inet_csk(sk);
1923         struct tcp_sock *tp = tcp_sk(sk);
1924         int err = 0;
1925         int old_state = sk->sk_state;
1926
1927         if (old_state != TCP_CLOSE)
1928                 tcp_set_state(sk, TCP_CLOSE);
1929
1930         /* ABORT function of RFC793 */
1931         if (old_state == TCP_LISTEN) {
1932                 inet_csk_listen_stop(sk);
1933         } else if (tcp_need_reset(old_state) ||
1934                    (tp->snd_nxt != tp->write_seq &&
1935                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1936                 /* The last check adjusts for discrepancy of Linux wrt. RFC
1937                  * states
1938                  */
1939                 tcp_send_active_reset(sk, gfp_any());
1940                 sk->sk_err = ECONNRESET;
1941         } else if (old_state == TCP_SYN_SENT)
1942                 sk->sk_err = ECONNRESET;
1943
1944         tcp_clear_xmit_timers(sk);
1945         __skb_queue_purge(&sk->sk_receive_queue);
1946         tcp_write_queue_purge(sk);
1947         __skb_queue_purge(&tp->out_of_order_queue);
1948 #ifdef CONFIG_NET_DMA
1949         __skb_queue_purge(&sk->sk_async_wait_queue);
1950 #endif
1951
1952         inet->dport = 0;
1953
1954         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1955                 inet_reset_saddr(sk);
1956
1957         sk->sk_shutdown = 0;
1958         sock_reset_flag(sk, SOCK_DONE);
1959         tp->srtt = 0;
1960         if ((tp->write_seq += tp->max_window + 2) == 0)
1961                 tp->write_seq = 1;
1962         icsk->icsk_backoff = 0;
1963         tp->snd_cwnd = 2;
1964         icsk->icsk_probes_out = 0;
1965         tp->packets_out = 0;
1966         tp->snd_ssthresh = 0x7fffffff;
1967         tp->snd_cwnd_cnt = 0;
1968         tp->bytes_acked = 0;
1969         tcp_set_ca_state(sk, TCP_CA_Open);
1970         tcp_clear_retrans(tp);
1971         inet_csk_delack_init(sk);
1972         tcp_init_send_head(sk);
1973         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
1974         __sk_dst_reset(sk);
1975
1976         BUG_TRAP(!inet->num || icsk->icsk_bind_hash);
1977
1978         sk->sk_error_report(sk);
1979         return err;
1980 }
1981
1982 /*
1983  *      Socket option code for TCP.
1984  */
1985 static int do_tcp_setsockopt(struct sock *sk, int level,
1986                 int optname, char __user *optval, int optlen)
1987 {
1988         struct tcp_sock *tp = tcp_sk(sk);
1989         struct inet_connection_sock *icsk = inet_csk(sk);
1990         int val;
1991         int err = 0;
1992
1993         /* This is a string value all the others are int's */
1994         if (optname == TCP_CONGESTION) {
1995                 char name[TCP_CA_NAME_MAX];
1996
1997                 if (optlen < 1)
1998                         return -EINVAL;
1999
2000                 val = strncpy_from_user(name, optval,
2001                                         min(TCP_CA_NAME_MAX-1, optlen));
2002                 if (val < 0)
2003                         return -EFAULT;
2004                 name[val] = 0;
2005
2006                 lock_sock(sk);
2007                 err = tcp_set_congestion_control(sk, name);
2008                 release_sock(sk);
2009                 return err;
2010         }
2011
2012         if (optlen < sizeof(int))
2013                 return -EINVAL;
2014
2015         if (get_user(val, (int __user *)optval))
2016                 return -EFAULT;
2017
2018         lock_sock(sk);
2019
2020         switch (optname) {
2021         case TCP_MAXSEG:
2022                 /* Values greater than interface MTU won't take effect. However
2023                  * at the point when this call is done we typically don't yet
2024                  * know which interface is going to be used */
2025                 if (val < 8 || val > MAX_TCP_WINDOW) {
2026                         err = -EINVAL;
2027                         break;
2028                 }
2029                 tp->rx_opt.user_mss = val;
2030                 break;
2031
2032         case TCP_NODELAY:
2033                 if (val) {
2034                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2035                          * this option on corked socket is remembered, but
2036                          * it is not activated until cork is cleared.
2037                          *
2038                          * However, when TCP_NODELAY is set we make
2039                          * an explicit push, which overrides even TCP_CORK
2040                          * for currently queued segments.
2041                          */
2042                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2043                         tcp_push_pending_frames(sk);
2044                 } else {
2045                         tp->nonagle &= ~TCP_NAGLE_OFF;
2046                 }
2047                 break;
2048
2049         case TCP_CORK:
2050                 /* When set indicates to always queue non-full frames.
2051                  * Later the user clears this option and we transmit
2052                  * any pending partial frames in the queue.  This is
2053                  * meant to be used alongside sendfile() to get properly
2054                  * filled frames when the user (for example) must write
2055                  * out headers with a write() call first and then use
2056                  * sendfile to send out the data parts.
2057                  *
2058                  * TCP_CORK can be set together with TCP_NODELAY and it is
2059                  * stronger than TCP_NODELAY.
2060                  */
2061                 if (val) {
2062                         tp->nonagle |= TCP_NAGLE_CORK;
2063                 } else {
2064                         tp->nonagle &= ~TCP_NAGLE_CORK;
2065                         if (tp->nonagle&TCP_NAGLE_OFF)
2066                                 tp->nonagle |= TCP_NAGLE_PUSH;
2067                         tcp_push_pending_frames(sk);
2068                 }
2069                 break;
2070
2071         case TCP_KEEPIDLE:
2072                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2073                         err = -EINVAL;
2074                 else {
2075                         tp->keepalive_time = val * HZ;
2076                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2077                             !((1 << sk->sk_state) &
2078                               (TCPF_CLOSE | TCPF_LISTEN))) {
2079                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2080                                 if (tp->keepalive_time > elapsed)
2081                                         elapsed = tp->keepalive_time - elapsed;
2082                                 else
2083                                         elapsed = 0;
2084                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2085                         }
2086                 }
2087                 break;
2088         case TCP_KEEPINTVL:
2089                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2090                         err = -EINVAL;
2091                 else
2092                         tp->keepalive_intvl = val * HZ;
2093                 break;
2094         case TCP_KEEPCNT:
2095                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2096                         err = -EINVAL;
2097                 else
2098                         tp->keepalive_probes = val;
2099                 break;
2100         case TCP_SYNCNT:
2101                 if (val < 1 || val > MAX_TCP_SYNCNT)
2102                         err = -EINVAL;
2103                 else
2104                         icsk->icsk_syn_retries = val;
2105                 break;
2106
2107         case TCP_LINGER2:
2108                 if (val < 0)
2109                         tp->linger2 = -1;
2110                 else if (val > sysctl_tcp_fin_timeout / HZ)
2111                         tp->linger2 = 0;
2112                 else
2113                         tp->linger2 = val * HZ;
2114                 break;
2115
2116         case TCP_DEFER_ACCEPT:
2117                 icsk->icsk_accept_queue.rskq_defer_accept = 0;
2118                 if (val > 0) {
2119                         /* Translate value in seconds to number of
2120                          * retransmits */
2121                         while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
2122                                val > ((TCP_TIMEOUT_INIT / HZ) <<
2123                                        icsk->icsk_accept_queue.rskq_defer_accept))
2124                                 icsk->icsk_accept_queue.rskq_defer_accept++;
2125                         icsk->icsk_accept_queue.rskq_defer_accept++;
2126                 }
2127                 break;
2128
2129         case TCP_WINDOW_CLAMP:
2130                 if (!val) {
2131                         if (sk->sk_state != TCP_CLOSE) {
2132                                 err = -EINVAL;
2133                                 break;
2134                         }
2135                         tp->window_clamp = 0;
2136                 } else
2137                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2138                                                 SOCK_MIN_RCVBUF / 2 : val;
2139                 break;
2140
2141         case TCP_QUICKACK:
2142                 if (!val) {
2143                         icsk->icsk_ack.pingpong = 1;
2144                 } else {
2145                         icsk->icsk_ack.pingpong = 0;
2146                         if ((1 << sk->sk_state) &
2147                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2148                             inet_csk_ack_scheduled(sk)) {
2149                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2150                                 tcp_cleanup_rbuf(sk, 1);
2151                                 if (!(val & 1))
2152                                         icsk->icsk_ack.pingpong = 1;
2153                         }
2154                 }
2155                 break;
2156
2157 #ifdef CONFIG_TCP_MD5SIG
2158         case TCP_MD5SIG:
2159                 /* Read the IP->Key mappings from userspace */
2160                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2161                 break;
2162 #endif
2163
2164         default:
2165                 err = -ENOPROTOOPT;
2166                 break;
2167         }
2168
2169         release_sock(sk);
2170         return err;
2171 }
2172
2173 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2174                    int optlen)
2175 {
2176         struct inet_connection_sock *icsk = inet_csk(sk);
2177
2178         if (level != SOL_TCP)
2179                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2180                                                      optval, optlen);
2181         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2182 }
2183
2184 #ifdef CONFIG_COMPAT
2185 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2186                           char __user *optval, int optlen)
2187 {
2188         if (level != SOL_TCP)
2189                 return inet_csk_compat_setsockopt(sk, level, optname,
2190                                                   optval, optlen);
2191         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2192 }
2193
2194 EXPORT_SYMBOL(compat_tcp_setsockopt);
2195 #endif
2196
2197 /* Return information about state of tcp endpoint in API format. */
2198 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2199 {
2200         struct tcp_sock *tp = tcp_sk(sk);
2201         const struct inet_connection_sock *icsk = inet_csk(sk);
2202         u32 now = tcp_time_stamp;
2203
2204         memset(info, 0, sizeof(*info));
2205
2206         info->tcpi_state = sk->sk_state;
2207         info->tcpi_ca_state = icsk->icsk_ca_state;
2208         info->tcpi_retransmits = icsk->icsk_retransmits;
2209         info->tcpi_probes = icsk->icsk_probes_out;
2210         info->tcpi_backoff = icsk->icsk_backoff;
2211
2212         if (tp->rx_opt.tstamp_ok)
2213                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2214         if (tcp_is_sack(tp))
2215                 info->tcpi_options |= TCPI_OPT_SACK;
2216         if (tp->rx_opt.wscale_ok) {
2217                 info->tcpi_options |= TCPI_OPT_WSCALE;
2218                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2219                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2220         }
2221
2222         if (tp->ecn_flags&TCP_ECN_OK)
2223                 info->tcpi_options |= TCPI_OPT_ECN;
2224
2225         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2226         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2227         info->tcpi_snd_mss = tp->mss_cache;
2228         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2229
2230         if (sk->sk_state == TCP_LISTEN) {
2231                 info->tcpi_unacked = sk->sk_ack_backlog;
2232                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2233         } else {
2234                 info->tcpi_unacked = tp->packets_out;
2235                 info->tcpi_sacked = tp->sacked_out;
2236         }
2237         info->tcpi_lost = tp->lost_out;
2238         info->tcpi_retrans = tp->retrans_out;
2239         info->tcpi_fackets = tp->fackets_out;
2240
2241         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2242         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2243         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2244
2245         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2246         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2247         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2248         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2249         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2250         info->tcpi_snd_cwnd = tp->snd_cwnd;
2251         info->tcpi_advmss = tp->advmss;
2252         info->tcpi_reordering = tp->reordering;
2253
2254         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2255         info->tcpi_rcv_space = tp->rcvq_space.space;
2256
2257         info->tcpi_total_retrans = tp->total_retrans;
2258 }
2259
2260 EXPORT_SYMBOL_GPL(tcp_get_info);
2261
2262 static int do_tcp_getsockopt(struct sock *sk, int level,
2263                 int optname, char __user *optval, int __user *optlen)
2264 {
2265         struct inet_connection_sock *icsk = inet_csk(sk);
2266         struct tcp_sock *tp = tcp_sk(sk);
2267         int val, len;
2268
2269         if (get_user(len, optlen))
2270                 return -EFAULT;
2271
2272         len = min_t(unsigned int, len, sizeof(int));
2273
2274         if (len < 0)
2275                 return -EINVAL;
2276
2277         switch (optname) {
2278         case TCP_MAXSEG:
2279                 val = tp->mss_cache;
2280                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2281                         val = tp->rx_opt.user_mss;
2282                 break;
2283         case TCP_NODELAY:
2284                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2285                 break;
2286         case TCP_CORK:
2287                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2288                 break;
2289         case TCP_KEEPIDLE:
2290                 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2291                 break;
2292         case TCP_KEEPINTVL:
2293                 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2294                 break;
2295         case TCP_KEEPCNT:
2296                 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2297                 break;
2298         case TCP_SYNCNT:
2299                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2300                 break;
2301         case TCP_LINGER2:
2302                 val = tp->linger2;
2303                 if (val >= 0)
2304                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2305                 break;
2306         case TCP_DEFER_ACCEPT:
2307                 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2308                         ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2309                 break;
2310         case TCP_WINDOW_CLAMP:
2311                 val = tp->window_clamp;
2312                 break;
2313         case TCP_INFO: {
2314                 struct tcp_info info;
2315
2316                 if (get_user(len, optlen))
2317                         return -EFAULT;
2318
2319                 tcp_get_info(sk, &info);
2320
2321                 len = min_t(unsigned int, len, sizeof(info));
2322                 if (put_user(len, optlen))
2323                         return -EFAULT;
2324                 if (copy_to_user(optval, &info, len))
2325                         return -EFAULT;
2326                 return 0;
2327         }
2328         case TCP_QUICKACK:
2329                 val = !icsk->icsk_ack.pingpong;
2330                 break;
2331
2332         case TCP_CONGESTION:
2333                 if (get_user(len, optlen))
2334                         return -EFAULT;
2335                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2336                 if (put_user(len, optlen))
2337                         return -EFAULT;
2338                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2339                         return -EFAULT;
2340                 return 0;
2341         default:
2342                 return -ENOPROTOOPT;
2343         }
2344
2345         if (put_user(len, optlen))
2346                 return -EFAULT;
2347         if (copy_to_user(optval, &val, len))
2348                 return -EFAULT;
2349         return 0;
2350 }
2351
2352 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2353                    int __user *optlen)
2354 {
2355         struct inet_connection_sock *icsk = inet_csk(sk);
2356
2357         if (level != SOL_TCP)
2358                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2359                                                      optval, optlen);
2360         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2361 }
2362
2363 #ifdef CONFIG_COMPAT
2364 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2365                           char __user *optval, int __user *optlen)
2366 {
2367         if (level != SOL_TCP)
2368                 return inet_csk_compat_getsockopt(sk, level, optname,
2369                                                   optval, optlen);
2370         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2371 }
2372
2373 EXPORT_SYMBOL(compat_tcp_getsockopt);
2374 #endif
2375
2376 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2377 {
2378         struct sk_buff *segs = ERR_PTR(-EINVAL);
2379         struct tcphdr *th;
2380         unsigned thlen;
2381         unsigned int seq;
2382         __be32 delta;
2383         unsigned int oldlen;
2384         unsigned int len;
2385
2386         if (!pskb_may_pull(skb, sizeof(*th)))
2387                 goto out;
2388
2389         th = tcp_hdr(skb);
2390         thlen = th->doff * 4;
2391         if (thlen < sizeof(*th))
2392                 goto out;
2393
2394         if (!pskb_may_pull(skb, thlen))
2395                 goto out;
2396
2397         oldlen = (u16)~skb->len;
2398         __skb_pull(skb, thlen);
2399
2400         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2401                 /* Packet is from an untrusted source, reset gso_segs. */
2402                 int type = skb_shinfo(skb)->gso_type;
2403                 int mss;
2404
2405                 if (unlikely(type &
2406                              ~(SKB_GSO_TCPV4 |
2407                                SKB_GSO_DODGY |
2408                                SKB_GSO_TCP_ECN |
2409                                SKB_GSO_TCPV6 |
2410                                0) ||
2411                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2412                         goto out;
2413
2414                 mss = skb_shinfo(skb)->gso_size;
2415                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2416
2417                 segs = NULL;
2418                 goto out;
2419         }
2420
2421         segs = skb_segment(skb, features);
2422         if (IS_ERR(segs))
2423                 goto out;
2424
2425         len = skb_shinfo(skb)->gso_size;
2426         delta = htonl(oldlen + (thlen + len));
2427
2428         skb = segs;
2429         th = tcp_hdr(skb);
2430         seq = ntohl(th->seq);
2431
2432         do {
2433                 th->fin = th->psh = 0;
2434
2435                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2436                                        (__force u32)delta));
2437                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2438                         th->check =
2439                              csum_fold(csum_partial(skb_transport_header(skb),
2440                                                     thlen, skb->csum));
2441
2442                 seq += len;
2443                 skb = skb->next;
2444                 th = tcp_hdr(skb);
2445
2446                 th->seq = htonl(seq);
2447                 th->cwr = 0;
2448         } while (skb->next);
2449
2450         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2451                       skb->data_len);
2452         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2453                                 (__force u32)delta));
2454         if (skb->ip_summed != CHECKSUM_PARTIAL)
2455                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2456                                                    thlen, skb->csum));
2457
2458 out:
2459         return segs;
2460 }
2461 EXPORT_SYMBOL(tcp_tso_segment);
2462
2463 #ifdef CONFIG_TCP_MD5SIG
2464 static unsigned long tcp_md5sig_users;
2465 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2466 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2467
2468 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2469 {
2470         int cpu;
2471         for_each_possible_cpu(cpu) {
2472                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2473                 if (p) {
2474                         if (p->md5_desc.tfm)
2475                                 crypto_free_hash(p->md5_desc.tfm);
2476                         kfree(p);
2477                         p = NULL;
2478                 }
2479         }
2480         free_percpu(pool);
2481 }
2482
2483 void tcp_free_md5sig_pool(void)
2484 {
2485         struct tcp_md5sig_pool **pool = NULL;
2486
2487         spin_lock_bh(&tcp_md5sig_pool_lock);
2488         if (--tcp_md5sig_users == 0) {
2489                 pool = tcp_md5sig_pool;
2490                 tcp_md5sig_pool = NULL;
2491         }
2492         spin_unlock_bh(&tcp_md5sig_pool_lock);
2493         if (pool)
2494                 __tcp_free_md5sig_pool(pool);
2495 }
2496
2497 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2498
2499 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(void)
2500 {
2501         int cpu;
2502         struct tcp_md5sig_pool **pool;
2503
2504         pool = alloc_percpu(struct tcp_md5sig_pool *);
2505         if (!pool)
2506                 return NULL;
2507
2508         for_each_possible_cpu(cpu) {
2509                 struct tcp_md5sig_pool *p;
2510                 struct crypto_hash *hash;
2511
2512                 p = kzalloc(sizeof(*p), GFP_KERNEL);
2513                 if (!p)
2514                         goto out_free;
2515                 *per_cpu_ptr(pool, cpu) = p;
2516
2517                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2518                 if (!hash || IS_ERR(hash))
2519                         goto out_free;
2520
2521                 p->md5_desc.tfm = hash;
2522         }
2523         return pool;
2524 out_free:
2525         __tcp_free_md5sig_pool(pool);
2526         return NULL;
2527 }
2528
2529 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(void)
2530 {
2531         struct tcp_md5sig_pool **pool;
2532         int alloc = 0;
2533
2534 retry:
2535         spin_lock_bh(&tcp_md5sig_pool_lock);
2536         pool = tcp_md5sig_pool;
2537         if (tcp_md5sig_users++ == 0) {
2538                 alloc = 1;
2539                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2540         } else if (!pool) {
2541                 tcp_md5sig_users--;
2542                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2543                 cpu_relax();
2544                 goto retry;
2545         } else
2546                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2547
2548         if (alloc) {
2549                 /* we cannot hold spinlock here because this may sleep. */
2550                 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool();
2551                 spin_lock_bh(&tcp_md5sig_pool_lock);
2552                 if (!p) {
2553                         tcp_md5sig_users--;
2554                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2555                         return NULL;
2556                 }
2557                 pool = tcp_md5sig_pool;
2558                 if (pool) {
2559                         /* oops, it has already been assigned. */
2560                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2561                         __tcp_free_md5sig_pool(p);
2562                 } else {
2563                         tcp_md5sig_pool = pool = p;
2564                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2565                 }
2566         }
2567         return pool;
2568 }
2569
2570 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2571
2572 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2573 {
2574         struct tcp_md5sig_pool **p;
2575         spin_lock_bh(&tcp_md5sig_pool_lock);
2576         p = tcp_md5sig_pool;
2577         if (p)
2578                 tcp_md5sig_users++;
2579         spin_unlock_bh(&tcp_md5sig_pool_lock);
2580         return (p ? *per_cpu_ptr(p, cpu) : NULL);
2581 }
2582
2583 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2584
2585 void __tcp_put_md5sig_pool(void)
2586 {
2587         tcp_free_md5sig_pool();
2588 }
2589
2590 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2591 #endif
2592
2593 void tcp_done(struct sock *sk)
2594 {
2595         if(sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2596                 TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS);
2597
2598         tcp_set_state(sk, TCP_CLOSE);
2599         tcp_clear_xmit_timers(sk);
2600
2601         sk->sk_shutdown = SHUTDOWN_MASK;
2602
2603         if (!sock_flag(sk, SOCK_DEAD))
2604                 sk->sk_state_change(sk);
2605         else
2606                 inet_csk_destroy_sock(sk);
2607 }
2608 EXPORT_SYMBOL_GPL(tcp_done);
2609
2610 extern struct tcp_congestion_ops tcp_reno;
2611
2612 static __initdata unsigned long thash_entries;
2613 static int __init set_thash_entries(char *str)
2614 {
2615         if (!str)
2616                 return 0;
2617         thash_entries = simple_strtoul(str, &str, 0);
2618         return 1;
2619 }
2620 __setup("thash_entries=", set_thash_entries);
2621
2622 void __init tcp_init(void)
2623 {
2624         struct sk_buff *skb = NULL;
2625         unsigned long nr_pages, limit;
2626         int order, i, max_share;
2627
2628         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
2629
2630         tcp_hashinfo.bind_bucket_cachep =
2631                 kmem_cache_create("tcp_bind_bucket",
2632                                   sizeof(struct inet_bind_bucket), 0,
2633                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
2634
2635         /* Size and allocate the main established and bind bucket
2636          * hash tables.
2637          *
2638          * The methodology is similar to that of the buffer cache.
2639          */
2640         tcp_hashinfo.ehash =
2641                 alloc_large_system_hash("TCP established",
2642                                         sizeof(struct inet_ehash_bucket),
2643                                         thash_entries,
2644                                         (num_physpages >= 128 * 1024) ?
2645                                         13 : 15,
2646                                         0,
2647                                         &tcp_hashinfo.ehash_size,
2648                                         NULL,
2649                                         thash_entries ? 0 : 512 * 1024);
2650         tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
2651         for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
2652                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].chain);
2653                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].twchain);
2654         }
2655         if (inet_ehash_locks_alloc(&tcp_hashinfo))
2656                 panic("TCP: failed to alloc ehash_locks");
2657         tcp_hashinfo.bhash =
2658                 alloc_large_system_hash("TCP bind",
2659                                         sizeof(struct inet_bind_hashbucket),
2660                                         tcp_hashinfo.ehash_size,
2661                                         (num_physpages >= 128 * 1024) ?
2662                                         13 : 15,
2663                                         0,
2664                                         &tcp_hashinfo.bhash_size,
2665                                         NULL,
2666                                         64 * 1024);
2667         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2668         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2669                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2670                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2671         }
2672
2673         /* Try to be a bit smarter and adjust defaults depending
2674          * on available memory.
2675          */
2676         for (order = 0; ((1 << order) << PAGE_SHIFT) <
2677                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2678                         order++)
2679                 ;
2680         if (order >= 4) {
2681                 tcp_death_row.sysctl_max_tw_buckets = 180000;
2682                 sysctl_tcp_max_orphans = 4096 << (order - 4);
2683                 sysctl_max_syn_backlog = 1024;
2684         } else if (order < 3) {
2685                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2686                 sysctl_tcp_max_orphans >>= (3 - order);
2687                 sysctl_max_syn_backlog = 128;
2688         }
2689
2690         /* Set the pressure threshold to be a fraction of global memory that
2691          * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2692          * memory, with a floor of 128 pages.
2693          */
2694         nr_pages = totalram_pages - totalhigh_pages;
2695         limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
2696         limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
2697         limit = max(limit, 128UL);
2698         sysctl_tcp_mem[0] = limit / 4 * 3;
2699         sysctl_tcp_mem[1] = limit;
2700         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2701
2702         /* Set per-socket limits to no more than 1/128 the pressure threshold */
2703         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2704         max_share = min(4UL*1024*1024, limit);
2705
2706         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
2707         sysctl_tcp_wmem[1] = 16*1024;
2708         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2709
2710         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
2711         sysctl_tcp_rmem[1] = 87380;
2712         sysctl_tcp_rmem[2] = max(87380, max_share);
2713
2714         printk(KERN_INFO "TCP: Hash tables configured "
2715                "(established %d bind %d)\n",
2716                tcp_hashinfo.ehash_size, tcp_hashinfo.bhash_size);
2717
2718         tcp_register_congestion_control(&tcp_reno);
2719 }
2720
2721 EXPORT_SYMBOL(tcp_close);
2722 EXPORT_SYMBOL(tcp_disconnect);
2723 EXPORT_SYMBOL(tcp_getsockopt);
2724 EXPORT_SYMBOL(tcp_ioctl);
2725 EXPORT_SYMBOL(tcp_poll);
2726 EXPORT_SYMBOL(tcp_read_sock);
2727 EXPORT_SYMBOL(tcp_recvmsg);
2728 EXPORT_SYMBOL(tcp_sendmsg);
2729 EXPORT_SYMBOL(tcp_splice_read);
2730 EXPORT_SYMBOL(tcp_sendpage);
2731 EXPORT_SYMBOL(tcp_setsockopt);
2732 EXPORT_SYMBOL(tcp_shutdown);
2733 EXPORT_SYMBOL(tcp_statistics);