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/config.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/smp_lock.h>
 257 #include <linux/fs.h>
 258 #include <linux/random.h>
 259 #include <linux/bootmem.h>
 260 #include <linux/cache.h>
 261 #include <linux/err.h>
 262
 263 #include <net/icmp.h>
 264 #include <net/tcp.h>
 265 #include <net/xfrm.h>
 266 #include <net/ip.h>
 267
 268
 269 #include <asm/uaccess.h>
 270 #include <asm/ioctls.h>
 271
 272 int sysctl_tcp_fin_timeout = TCP_FIN_TIMEOUT;
 273
 274 DEFINE_SNMP_STAT(struct tcp_mib, tcp_statistics) __read_mostly;
 275
 276 atomic_t tcp_orphan_count = ATOMIC_INIT(0);
 277
 278 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 279
 280 int sysctl_tcp_mem[3] __read_mostly;
 281 int sysctl_tcp_wmem[3] __read_mostly;
 282 int sysctl_tcp_rmem[3] __read_mostly;
 283
 284 EXPORT_SYMBOL(sysctl_tcp_mem);
 285 EXPORT_SYMBOL(sysctl_tcp_rmem);
 286 EXPORT_SYMBOL(sysctl_tcp_wmem);
 287
 288 atomic_t tcp_memory_allocated;  /* Current allocated memory. */
 289 atomic_t tcp_sockets_allocated; /* Current number of TCP sockets. */
 290
 291 EXPORT_SYMBOL(tcp_memory_allocated);
 292 EXPORT_SYMBOL(tcp_sockets_allocated);
 293
 294 /*
 295  * Pressure flag: try to collapse.
 296  * Technical note: it is used by multiple contexts non atomically.
 297  * All the sk_stream_mem_schedule() is of this nature: accounting
 298  * is strict, actions are advisory and have some latency.
 299  */
 300 int tcp_memory_pressure;
 301
 302 EXPORT_SYMBOL(tcp_memory_pressure);
 303
 304 void tcp_enter_memory_pressure(void)
 305 {
 306         if (!tcp_memory_pressure) {
 307                 NET_INC_STATS(LINUX_MIB_TCPMEMORYPRESSURES);
 308                 tcp_memory_pressure = 1;
 309         }
 310 }
 311
 312 EXPORT_SYMBOL(tcp_enter_memory_pressure);
 313
 314 /*
 315  *      Wait for a TCP event.
 316  *
 317  *      Note that we don't need to lock the socket, as the upper poll layers
 318  *      take care of normal races (between the test and the event) and we don't
 319  *      go look at any of the socket buffers directly.
 320  */
 321 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 322 {
 323         unsigned int mask;
 324         struct sock *sk = sock->sk;
 325         struct tcp_sock *tp = tcp_sk(sk);
 326
 327         poll_wait(file, sk->sk_sleep, wait);
 328         if (sk->sk_state == TCP_LISTEN)
 329                 return inet_csk_listen_poll(sk);
 330
 331         /* Socket is not locked. We are protected from async events
 332            by poll logic and correct handling of state changes
 333            made by another threads is impossible in any case.
 334          */
 335
 336         mask = 0;
 337         if (sk->sk_err)
 338                 mask = POLLERR;
 339
 340         /*
 341          * POLLHUP is certainly not done right. But poll() doesn't
 342          * have a notion of HUP in just one direction, and for a
 343          * socket the read side is more interesting.
 344          *
 345          * Some poll() documentation says that POLLHUP is incompatible
 346          * with the POLLOUT/POLLWR flags, so somebody should check this
 347          * all. But careful, it tends to be safer to return too many
 348          * bits than too few, and you can easily break real applications
 349          * if you don't tell them that something has hung up!
 350          *
 351          * Check-me.
 352          *
 353          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
 354          * our fs/select.c). It means that after we received EOF,
 355          * poll always returns immediately, making impossible poll() on write()
 356          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
 357          * if and only if shutdown has been made in both directions.
 358          * Actually, it is interesting to look how Solaris and DUX
 359          * solve this dilemma. I would prefer, if PULLHUP were maskable,
 360          * then we could set it on SND_SHUTDOWN. BTW examples given
 361          * in Stevens' books assume exactly this behaviour, it explains
 362          * why PULLHUP is incompatible with POLLOUT.    --ANK
 363          *
 364          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 365          * blocking on fresh not-connected or disconnected socket. --ANK
 366          */
 367         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
 368                 mask |= POLLHUP;
 369         if (sk->sk_shutdown & RCV_SHUTDOWN)
 370                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
 371
 372         /* Connected? */
 373         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
 374                 /* Potential race condition. If read of tp below will
 375                  * escape above sk->sk_state, we can be illegally awaken
 376                  * in SYN_* states. */
 377                 if ((tp->rcv_nxt != tp->copied_seq) &&
 378                     (tp->urg_seq != tp->copied_seq ||
 379                      tp->rcv_nxt != tp->copied_seq + 1 ||
 380                      sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data))
 381                         mask |= POLLIN | POLLRDNORM;
 382
 383                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 384                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
 385                                 mask |= POLLOUT | POLLWRNORM;
 386                         } else {  /* send SIGIO later */
 387                                 set_bit(SOCK_ASYNC_NOSPACE,
 388                                         &sk->sk_socket->flags);
 389                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 390
 391                                 /* Race breaker. If space is freed after
 392                                  * wspace test but before the flags are set,
 393                                  * IO signal will be lost.
 394                                  */
 395                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
 396                                         mask |= POLLOUT | POLLWRNORM;
 397                         }
 398                 }
 399
 400                 if (tp->urg_data & TCP_URG_VALID)
 401                         mask |= POLLPRI;
 402         }
 403         return mask;
 404 }
 405
 406 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 407 {
 408         struct tcp_sock *tp = tcp_sk(sk);
 409         int answ;
 410
 411         switch (cmd) {
 412         case SIOCINQ:
 413                 if (sk->sk_state == TCP_LISTEN)
 414                         return -EINVAL;
 415
 416                 lock_sock(sk);
 417                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 418                         answ = 0;
 419                 else if (sock_flag(sk, SOCK_URGINLINE) ||
 420                          !tp->urg_data ||
 421                          before(tp->urg_seq, tp->copied_seq) ||
 422                          !before(tp->urg_seq, tp->rcv_nxt)) {
 423                         answ = tp->rcv_nxt - tp->copied_seq;
 424
 425                         /* Subtract 1, if FIN is in queue. */
 426                         if (answ && !skb_queue_empty(&sk->sk_receive_queue))
 427                                 answ -=
 428                        ((struct sk_buff *)sk->sk_receive_queue.prev)->h.th->fin;
 429                 } else
 430                         answ = tp->urg_seq - tp->copied_seq;
 431                 release_sock(sk);
 432                 break;
 433         case SIOCATMARK:
 434                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 435                 break;
 436         case SIOCOUTQ:
 437                 if (sk->sk_state == TCP_LISTEN)
 438                         return -EINVAL;
 439
 440                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 441                         answ = 0;
 442                 else
 443                         answ = tp->write_seq - tp->snd_una;
 444                 break;
 445         default:
 446                 return -ENOIOCTLCMD;
 447         };
 448
 449         return put_user(answ, (int __user *)arg);
 450 }
 451
 452 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 453 {
 454         TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
 455         tp->pushed_seq = tp->write_seq;
 456 }
 457
 458 static inline int forced_push(struct tcp_sock *tp)
 459 {
 460         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 461 }
 462
 463 static inline void skb_entail(struct sock *sk, struct tcp_sock *tp,
 464                               struct sk_buff *skb)
 465 {
 466         skb->csum = 0;
 467         TCP_SKB_CB(skb)->seq = tp->write_seq;
 468         TCP_SKB_CB(skb)->end_seq = tp->write_seq;
 469         TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
 470         TCP_SKB_CB(skb)->sacked = 0;
 471         skb_header_release(skb);
 472         __skb_queue_tail(&sk->sk_write_queue, skb);
 473         sk_charge_skb(sk, skb);
 474         if (!sk->sk_send_head)
 475                 sk->sk_send_head = skb;
 476         if (tp->nonagle & TCP_NAGLE_PUSH)
 477                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 478 }
 479
 480 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
 481                                 struct sk_buff *skb)
 482 {
 483         if (flags & MSG_OOB) {
 484                 tp->urg_mode = 1;
 485                 tp->snd_up = tp->write_seq;
 486                 TCP_SKB_CB(skb)->sacked |= TCPCB_URG;
 487         }
 488 }
 489
 490 static inline void tcp_push(struct sock *sk, struct tcp_sock *tp, int flags,
 491                             int mss_now, int nonagle)
 492 {
 493         if (sk->sk_send_head) {
 494                 struct sk_buff *skb = sk->sk_write_queue.prev;
 495                 if (!(flags & MSG_MORE) || forced_push(tp))
 496                         tcp_mark_push(tp, skb);
 497                 tcp_mark_urg(tp, flags, skb);
 498                 __tcp_push_pending_frames(sk, tp, mss_now,
 499                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
 500         }
 501 }
 502
 503 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
 504                          size_t psize, int flags)
 505 {
 506         struct tcp_sock *tp = tcp_sk(sk);
 507         int mss_now, size_goal;
 508         int err;
 509         ssize_t copied;
 510         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 511
 512         /* Wait for a connection to finish. */
 513         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 514                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 515                         goto out_err;
 516
 517         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 518
 519         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 520         size_goal = tp->xmit_size_goal;
 521         copied = 0;
 522
 523         err = -EPIPE;
 524         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 525                 goto do_error;
 526
 527         while (psize > 0) {
 528                 struct sk_buff *skb = sk->sk_write_queue.prev;
 529                 struct page *page = pages[poffset / PAGE_SIZE];
 530                 int copy, i, can_coalesce;
 531                 int offset = poffset % PAGE_SIZE;
 532                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
 533
 534                 if (!sk->sk_send_head || (copy = size_goal - skb->len) <= 0) {
 535 new_segment:
 536                         if (!sk_stream_memory_free(sk))
 537                                 goto wait_for_sndbuf;
 538
 539                         skb = sk_stream_alloc_pskb(sk, 0, 0,
 540                                                    sk->sk_allocation);
 541                         if (!skb)
 542                                 goto wait_for_memory;
 543
 544                         skb_entail(sk, tp, skb);
 545                         copy = size_goal;
 546                 }
 547
 548                 if (copy > size)
 549                         copy = size;
 550
 551                 i = skb_shinfo(skb)->nr_frags;
 552                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 553                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
 554                         tcp_mark_push(tp, skb);
 555                         goto new_segment;
 556                 }
 557                 if (!sk_stream_wmem_schedule(sk, copy))
 558                         goto wait_for_memory;
 559
 560                 if (can_coalesce) {
 561                         skb_shinfo(skb)->frags[i - 1].size += copy;
 562                 } else {
 563                         get_page(page);
 564                         skb_fill_page_desc(skb, i, page, offset, copy);
 565                 }
 566
 567                 skb->len += copy;
 568                 skb->data_len += copy;
 569                 skb->truesize += copy;
 570                 sk->sk_wmem_queued += copy;
 571                 sk->sk_forward_alloc -= copy;
 572                 skb->ip_summed = CHECKSUM_HW;
 573                 tp->write_seq += copy;
 574                 TCP_SKB_CB(skb)->end_seq += copy;
 575                 skb_shinfo(skb)->gso_segs = 0;
 576
 577                 if (!copied)
 578                         TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
 579
 580                 copied += copy;
 581                 poffset += copy;
 582                 if (!(psize -= copy))
 583                         goto out;
 584
 585                 if (skb->len < mss_now || (flags & MSG_OOB))
 586                         continue;
 587
 588                 if (forced_push(tp)) {
 589                         tcp_mark_push(tp, skb);
 590                         __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
 591                 } else if (skb == sk->sk_send_head)
 592                         tcp_push_one(sk, mss_now);
 593                 continue;
 594
 595 wait_for_sndbuf:
 596                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 597 wait_for_memory:
 598                 if (copied)
 599                         tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
 600
 601                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 602                         goto do_error;
 603
 604                 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 605                 size_goal = tp->xmit_size_goal;
 606         }
 607
 608 out:
 609         if (copied)
 610                 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
 611         return copied;
 612
 613 do_error:
 614         if (copied)
 615                 goto out;
 616 out_err:
 617         return sk_stream_error(sk, flags, err);
 618 }
 619
 620 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
 621                      size_t size, int flags)
 622 {
 623         ssize_t res;
 624         struct sock *sk = sock->sk;
 625
 626         if (!(sk->sk_route_caps & NETIF_F_SG) ||
 627             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
 628                 return sock_no_sendpage(sock, page, offset, size, flags);
 629
 630         lock_sock(sk);
 631         TCP_CHECK_TIMER(sk);
 632         res = do_tcp_sendpages(sk, &page, offset, size, flags);
 633         TCP_CHECK_TIMER(sk);
 634         release_sock(sk);
 635         return res;
 636 }
 637
 638 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
 639 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
 640
 641 static inline int select_size(struct sock *sk, struct tcp_sock *tp)
 642 {
 643         int tmp = tp->mss_cache;
 644
 645         if (sk->sk_route_caps & NETIF_F_SG) {
 646                 if (sk->sk_route_caps & NETIF_F_TSO)
 647                         tmp = 0;
 648                 else {
 649                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
 650
 651                         if (tmp >= pgbreak &&
 652                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
 653                                 tmp = pgbreak;
 654                 }
 655         }
 656
 657         return tmp;
 658 }
 659
 660 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
 661                 size_t size)
 662 {
 663         struct iovec *iov;
 664         struct tcp_sock *tp = tcp_sk(sk);
 665         struct sk_buff *skb;
 666         int iovlen, flags;
 667         int mss_now, size_goal;
 668         int err, copied;
 669         long timeo;
 670
 671         lock_sock(sk);
 672         TCP_CHECK_TIMER(sk);
 673
 674         flags = msg->msg_flags;
 675         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 676
 677         /* Wait for a connection to finish. */
 678         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 679                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 680                         goto out_err;
 681
 682         /* This should be in poll */
 683         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 684
 685         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 686         size_goal = tp->xmit_size_goal;
 687
 688         /* Ok commence sending. */
 689         iovlen = msg->msg_iovlen;
 690         iov = msg->msg_iov;
 691         copied = 0;
 692
 693         err = -EPIPE;
 694         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 695                 goto do_error;
 696
 697         while (--iovlen >= 0) {
 698                 int seglen = iov->iov_len;
 699                 unsigned char __user *from = iov->iov_base;
 700
 701                 iov++;
 702
 703                 while (seglen > 0) {
 704                         int copy;
 705
 706                         skb = sk->sk_write_queue.prev;
 707
 708                         if (!sk->sk_send_head ||
 709                             (copy = size_goal - skb->len) <= 0) {
 710
 711 new_segment:
 712                                 /* Allocate new segment. If the interface is SG,
 713                                  * allocate skb fitting to single page.
 714                                  */
 715                                 if (!sk_stream_memory_free(sk))
 716                                         goto wait_for_sndbuf;
 717
 718                                 skb = sk_stream_alloc_pskb(sk, select_size(sk, tp),
 719                                                            0, sk->sk_allocation);
 720                                 if (!skb)
 721                                         goto wait_for_memory;
 722
 723                                 /*
 724                                  * Check whether we can use HW checksum.
 725                                  */
 726                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
 727                                         skb->ip_summed = CHECKSUM_HW;
 728
 729                                 skb_entail(sk, tp, skb);
 730                                 copy = size_goal;
 731                         }
 732
 733                         /* Try to append data to the end of skb. */
 734                         if (copy > seglen)
 735                                 copy = seglen;
 736
 737                         /* Where to copy to? */
 738                         if (skb_tailroom(skb) > 0) {
 739                                 /* We have some space in skb head. Superb! */
 740                                 if (copy > skb_tailroom(skb))
 741                                         copy = skb_tailroom(skb);
 742                                 if ((err = skb_add_data(skb, from, copy)) != 0)
 743                                         goto do_fault;
 744                         } else {
 745                                 int merge = 0;
 746                                 int i = skb_shinfo(skb)->nr_frags;
 747                                 struct page *page = TCP_PAGE(sk);
 748                                 int off = TCP_OFF(sk);
 749
 750                                 if (skb_can_coalesce(skb, i, page, off) &&
 751                                     off != PAGE_SIZE) {
 752                                         /* We can extend the last page
 753                                          * fragment. */
 754                                         merge = 1;
 755                                 } else if (i == MAX_SKB_FRAGS ||
 756                                            (!i &&
 757                                            !(sk->sk_route_caps & NETIF_F_SG))) {
 758                                         /* Need to add new fragment and cannot
 759                                          * do this because interface is non-SG,
 760                                          * or because all the page slots are
 761                                          * busy. */
 762                                         tcp_mark_push(tp, skb);
 763                                         goto new_segment;
 764                                 } else if (page) {
 765                                         if (off == PAGE_SIZE) {
 766                                                 put_page(page);
 767                                                 TCP_PAGE(sk) = page = NULL;
 768                                                 off = 0;
 769                                         }
 770                                 } else
 771                                         off = 0;
 772
 773                                 if (copy > PAGE_SIZE - off)
 774                                         copy = PAGE_SIZE - off;
 775
 776                                 if (!sk_stream_wmem_schedule(sk, copy))
 777                                         goto wait_for_memory;
 778
 779                                 if (!page) {
 780                                         /* Allocate new cache page. */
 781                                         if (!(page = sk_stream_alloc_page(sk)))
 782                                                 goto wait_for_memory;
 783                                 }
 784
 785                                 /* Time to copy data. We are close to
 786                                  * the end! */
 787                                 err = skb_copy_to_page(sk, from, skb, page,
 788                                                        off, copy);
 789                                 if (err) {
 790                                         /* If this page was new, give it to the
 791                                          * socket so it does not get leaked.
 792                                          */
 793                                         if (!TCP_PAGE(sk)) {
 794                                                 TCP_PAGE(sk) = page;
 795                                                 TCP_OFF(sk) = 0;
 796                                         }
 797                                         goto do_error;
 798                                 }
 799
 800                                 /* Update the skb. */
 801                                 if (merge) {
 802                                         skb_shinfo(skb)->frags[i - 1].size +=
 803                                                                         copy;
 804                                 } else {
 805                                         skb_fill_page_desc(skb, i, page, off, copy);
 806                                         if (TCP_PAGE(sk)) {
 807                                                 get_page(page);
 808                                         } else if (off + copy < PAGE_SIZE) {
 809                                                 get_page(page);
 810                                                 TCP_PAGE(sk) = page;
 811                                         }
 812                                 }
 813
 814                                 TCP_OFF(sk) = off + copy;
 815                         }
 816
 817                         if (!copied)
 818                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
 819
 820                         tp->write_seq += copy;
 821                         TCP_SKB_CB(skb)->end_seq += copy;
 822                         skb_shinfo(skb)->gso_segs = 0;
 823
 824                         from += copy;
 825                         copied += copy;
 826                         if ((seglen -= copy) == 0 && iovlen == 0)
 827                                 goto out;
 828
 829                         if (skb->len < mss_now || (flags & MSG_OOB))
 830                                 continue;
 831
 832                         if (forced_push(tp)) {
 833                                 tcp_mark_push(tp, skb);
 834                                 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
 835                         } else if (skb == sk->sk_send_head)
 836                                 tcp_push_one(sk, mss_now);
 837                         continue;
 838
 839 wait_for_sndbuf:
 840                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 841 wait_for_memory:
 842                         if (copied)
 843                                 tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
 844
 845                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 846                                 goto do_error;
 847
 848                         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 849                         size_goal = tp->xmit_size_goal;
 850                 }
 851         }
 852
 853 out:
 854         if (copied)
 855                 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
 856         TCP_CHECK_TIMER(sk);
 857         release_sock(sk);
 858         return copied;
 859
 860 do_fault:
 861         if (!skb->len) {
 862                 if (sk->sk_send_head == skb)
 863                         sk->sk_send_head = NULL;
 864                 __skb_unlink(skb, &sk->sk_write_queue);
 865                 sk_stream_free_skb(sk, skb);
 866         }
 867
 868 do_error:
 869         if (copied)
 870                 goto out;
 871 out_err:
 872         err = sk_stream_error(sk, flags, err);
 873         TCP_CHECK_TIMER(sk);
 874         release_sock(sk);
 875         return err;
 876 }
 877
 878 /*
 879  *      Handle reading urgent data. BSD has very simple semantics for
 880  *      this, no blocking and very strange errors 8)
 881  */
 882
 883 static int tcp_recv_urg(struct sock *sk, long timeo,
 884                         struct msghdr *msg, int len, int flags,
 885                         int *addr_len)
 886 {
 887         struct tcp_sock *tp = tcp_sk(sk);
 888
 889         /* No URG data to read. */
 890         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
 891             tp->urg_data == TCP_URG_READ)
 892                 return -EINVAL; /* Yes this is right ! */
 893
 894         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
 895                 return -ENOTCONN;
 896
 897         if (tp->urg_data & TCP_URG_VALID) {
 898                 int err = 0;
 899                 char c = tp->urg_data;
 900
 901                 if (!(flags & MSG_PEEK))
 902                         tp->urg_data = TCP_URG_READ;
 903
 904                 /* Read urgent data. */
 905                 msg->msg_flags |= MSG_OOB;
 906
 907                 if (len > 0) {
 908                         if (!(flags & MSG_TRUNC))
 909                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
 910                         len = 1;
 911                 } else
 912                         msg->msg_flags |= MSG_TRUNC;
 913
 914                 return err ? -EFAULT : len;
 915         }
 916
 917         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
 918                 return 0;
 919
 920         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
 921          * the available implementations agree in this case:
 922          * this call should never block, independent of the
 923          * blocking state of the socket.
 924          * Mike <pall@rz.uni-karlsruhe.de>
 925          */
 926         return -EAGAIN;
 927 }
 928
 929 /* Clean up the receive buffer for full frames taken by the user,
 930  * then send an ACK if necessary.  COPIED is the number of bytes
 931  * tcp_recvmsg has given to the user so far, it speeds up the
 932  * calculation of whether or not we must ACK for the sake of
 933  * a window update.
 934  */
 935 void cleanup_rbuf(struct sock *sk, int copied)
 936 {
 937         struct tcp_sock *tp = tcp_sk(sk);
 938         int time_to_ack = 0;
 939
 940 #if TCP_DEBUG
 941         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
 942
 943         BUG_TRAP(!skb || before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
 944 #endif
 945
 946         if (inet_csk_ack_scheduled(sk)) {
 947                 const struct inet_connection_sock *icsk = inet_csk(sk);
 948                    /* Delayed ACKs frequently hit locked sockets during bulk
 949                     * receive. */
 950                 if (icsk->icsk_ack.blocked ||
 951                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
 952                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
 953                     /*
 954                      * If this read emptied read buffer, we send ACK, if
 955                      * connection is not bidirectional, user drained
 956                      * receive buffer and there was a small segment
 957                      * in queue.
 958                      */
 959                     (copied > 0 && (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
 960                      !icsk->icsk_ack.pingpong && !atomic_read(&sk->sk_rmem_alloc)))
 961                         time_to_ack = 1;
 962         }
 963
 964         /* We send an ACK if we can now advertise a non-zero window
 965          * which has been raised "significantly".
 966          *
 967          * Even if window raised up to infinity, do not send window open ACK
 968          * in states, where we will not receive more. It is useless.
 969          */
 970         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
 971                 __u32 rcv_window_now = tcp_receive_window(tp);
 972
 973                 /* Optimize, __tcp_select_window() is not cheap. */
 974                 if (2*rcv_window_now <= tp->window_clamp) {
 975                         __u32 new_window = __tcp_select_window(sk);
 976
 977                         /* Send ACK now, if this read freed lots of space
 978                          * in our buffer. Certainly, new_window is new window.
 979                          * We can advertise it now, if it is not less than current one.
 980                          * "Lots" means "at least twice" here.
 981                          */
 982                         if (new_window && new_window >= 2 * rcv_window_now)
 983                                 time_to_ack = 1;
 984                 }
 985         }
 986         if (time_to_ack)
 987                 tcp_send_ack(sk);
 988 }
 989
 990 static void tcp_prequeue_process(struct sock *sk)
 991 {
 992         struct sk_buff *skb;
 993         struct tcp_sock *tp = tcp_sk(sk);
 994
 995         NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED);
 996
 997         /* RX process wants to run with disabled BHs, though it is not
 998          * necessary */
 999         local_bh_disable();
1000         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1001                 sk->sk_backlog_rcv(sk, skb);
1002         local_bh_enable();
1003
1004         /* Clear memory counter. */
1005         tp->ucopy.memory = 0;
1006 }
1007
1008 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1009 {
1010         struct sk_buff *skb;
1011         u32 offset;
1012
1013         skb_queue_walk(&sk->sk_receive_queue, skb) {
1014                 offset = seq - TCP_SKB_CB(skb)->seq;
1015                 if (skb->h.th->syn)
1016                         offset--;
1017                 if (offset < skb->len || skb->h.th->fin) {
1018                         *off = offset;
1019                         return skb;
1020                 }
1021         }
1022         return NULL;
1023 }
1024
1025 /*
1026  * This routine provides an alternative to tcp_recvmsg() for routines
1027  * that would like to handle copying from skbuffs directly in 'sendfile'
1028  * fashion.
1029  * Note:
1030  *      - It is assumed that the socket was locked by the caller.
1031  *      - The routine does not block.
1032  *      - At present, there is no support for reading OOB data
1033  *        or for 'peeking' the socket using this routine
1034  *        (although both would be easy to implement).
1035  */
1036 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1037                   sk_read_actor_t recv_actor)
1038 {
1039         struct sk_buff *skb;
1040         struct tcp_sock *tp = tcp_sk(sk);
1041         u32 seq = tp->copied_seq;
1042         u32 offset;
1043         int copied = 0;
1044
1045         if (sk->sk_state == TCP_LISTEN)
1046                 return -ENOTCONN;
1047         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1048                 if (offset < skb->len) {
1049                         size_t used, len;
1050
1051                         len = skb->len - offset;
1052                         /* Stop reading if we hit a patch of urgent data */
1053                         if (tp->urg_data) {
1054                                 u32 urg_offset = tp->urg_seq - seq;
1055                                 if (urg_offset < len)
1056                                         len = urg_offset;
1057                                 if (!len)
1058                                         break;
1059                         }
1060                         used = recv_actor(desc, skb, offset, len);
1061                         if (used <= len) {
1062                                 seq += used;
1063                                 copied += used;
1064                                 offset += used;
1065                         }
1066                         if (offset != skb->len)
1067                                 break;
1068                 }
1069                 if (skb->h.th->fin) {
1070                         sk_eat_skb(sk, skb);
1071                         ++seq;
1072                         break;
1073                 }
1074                 sk_eat_skb(sk, skb);
1075                 if (!desc->count)
1076                         break;
1077         }
1078         tp->copied_seq = seq;
1079
1080         tcp_rcv_space_adjust(sk);
1081
1082         /* Clean up data we have read: This will do ACK frames. */
1083         if (copied)
1084                 cleanup_rbuf(sk, copied);
1085         return copied;
1086 }
1087
1088 /*
1089  *      This routine copies from a sock struct into the user buffer.
1090  *
1091  *      Technical note: in 2.3 we work on _locked_ socket, so that
1092  *      tricks with *seq access order and skb->users are not required.
1093  *      Probably, code can be easily improved even more.
1094  */
1095
1096 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1097                 size_t len, int nonblock, int flags, int *addr_len)
1098 {
1099         struct tcp_sock *tp = tcp_sk(sk);
1100         int copied = 0;
1101         u32 peek_seq;
1102         u32 *seq;
1103         unsigned long used;
1104         int err;
1105         int target;             /* Read at least this many bytes */
1106         long timeo;
1107         struct task_struct *user_recv = NULL;
1108
1109         lock_sock(sk);
1110
1111         TCP_CHECK_TIMER(sk);
1112
1113         err = -ENOTCONN;
1114         if (sk->sk_state == TCP_LISTEN)
1115                 goto out;
1116
1117         timeo = sock_rcvtimeo(sk, nonblock);
1118
1119         /* Urgent data needs to be handled specially. */
1120         if (flags & MSG_OOB)
1121                 goto recv_urg;
1122
1123         seq = &tp->copied_seq;
1124         if (flags & MSG_PEEK) {
1125                 peek_seq = tp->copied_seq;
1126                 seq = &peek_seq;
1127         }
1128
1129         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1130
1131         do {
1132                 struct sk_buff *skb;
1133                 u32 offset;
1134
1135                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1136                 if (tp->urg_data && tp->urg_seq == *seq) {
1137                         if (copied)
1138                                 break;
1139                         if (signal_pending(current)) {
1140                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1141                                 break;
1142                         }
1143                 }
1144
1145                 /* Next get a buffer. */
1146
1147                 skb = skb_peek(&sk->sk_receive_queue);
1148                 do {
1149                         if (!skb)
1150                                 break;
1151
1152                         /* Now that we have two receive queues this
1153                          * shouldn't happen.
1154                          */
1155                         if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1156                                 printk(KERN_INFO "recvmsg bug: copied %X "
1157                                        "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1158                                 break;
1159                         }
1160                         offset = *seq - TCP_SKB_CB(skb)->seq;
1161                         if (skb->h.th->syn)
1162                                 offset--;
1163                         if (offset < skb->len)
1164                                 goto found_ok_skb;
1165                         if (skb->h.th->fin)
1166                                 goto found_fin_ok;
1167                         BUG_TRAP(flags & MSG_PEEK);
1168                         skb = skb->next;
1169                 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1170
1171                 /* Well, if we have backlog, try to process it now yet. */
1172
1173                 if (copied >= target && !sk->sk_backlog.tail)
1174                         break;
1175
1176                 if (copied) {
1177                         if (sk->sk_err ||
1178                             sk->sk_state == TCP_CLOSE ||
1179                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1180                             !timeo ||
1181                             signal_pending(current) ||
1182                             (flags & MSG_PEEK))
1183                                 break;
1184                 } else {
1185                         if (sock_flag(sk, SOCK_DONE))
1186                                 break;
1187
1188                         if (sk->sk_err) {
1189                                 copied = sock_error(sk);
1190                                 break;
1191                         }
1192
1193                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1194                                 break;
1195
1196                         if (sk->sk_state == TCP_CLOSE) {
1197                                 if (!sock_flag(sk, SOCK_DONE)) {
1198                                         /* This occurs when user tries to read
1199                                          * from never connected socket.
1200                                          */
1201                                         copied = -ENOTCONN;
1202                                         break;
1203                                 }
1204                                 break;
1205                         }
1206
1207                         if (!timeo) {
1208                                 copied = -EAGAIN;
1209                                 break;
1210                         }
1211
1212                         if (signal_pending(current)) {
1213                                 copied = sock_intr_errno(timeo);
1214                                 break;
1215                         }
1216                 }
1217
1218                 cleanup_rbuf(sk, copied);
1219
1220                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1221                         /* Install new reader */
1222                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1223                                 user_recv = current;
1224                                 tp->ucopy.task = user_recv;
1225                                 tp->ucopy.iov = msg->msg_iov;
1226                         }
1227
1228                         tp->ucopy.len = len;
1229
1230                         BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1231                                  (flags & (MSG_PEEK | MSG_TRUNC)));
1232
1233                         /* Ugly... If prequeue is not empty, we have to
1234                          * process it before releasing socket, otherwise
1235                          * order will be broken at second iteration.
1236                          * More elegant solution is required!!!
1237                          *
1238                          * Look: we have the following (pseudo)queues:
1239                          *
1240                          * 1. packets in flight
1241                          * 2. backlog
1242                          * 3. prequeue
1243                          * 4. receive_queue
1244                          *
1245                          * Each queue can be processed only if the next ones
1246                          * are empty. At this point we have empty receive_queue.
1247                          * But prequeue _can_ be not empty after 2nd iteration,
1248                          * when we jumped to start of loop because backlog
1249                          * processing added something to receive_queue.
1250                          * We cannot release_sock(), because backlog contains
1251                          * packets arrived _after_ prequeued ones.
1252                          *
1253                          * Shortly, algorithm is clear --- to process all
1254                          * the queues in order. We could make it more directly,
1255                          * requeueing packets from backlog to prequeue, if
1256                          * is not empty. It is more elegant, but eats cycles,
1257                          * unfortunately.
1258                          */
1259                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1260                                 goto do_prequeue;
1261
1262                         /* __ Set realtime policy in scheduler __ */
1263                 }
1264
1265                 if (copied >= target) {
1266                         /* Do not sleep, just process backlog. */
1267                         release_sock(sk);
1268                         lock_sock(sk);
1269                 } else
1270                         sk_wait_data(sk, &timeo);
1271
1272                 if (user_recv) {
1273                         int chunk;
1274
1275                         /* __ Restore normal policy in scheduler __ */
1276
1277                         if ((chunk = len - tp->ucopy.len) != 0) {
1278                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1279                                 len -= chunk;
1280                                 copied += chunk;
1281                         }
1282
1283                         if (tp->rcv_nxt == tp->copied_seq &&
1284                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1285 do_prequeue:
1286                                 tcp_prequeue_process(sk);
1287
1288                                 if ((chunk = len - tp->ucopy.len) != 0) {
1289                                         NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1290                                         len -= chunk;
1291                                         copied += chunk;
1292                                 }
1293                         }
1294                 }
1295                 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1296                         if (net_ratelimit())
1297                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1298                                        current->comm, current->pid);
1299                         peek_seq = tp->copied_seq;
1300                 }
1301                 continue;
1302
1303         found_ok_skb:
1304                 /* Ok so how much can we use? */
1305                 used = skb->len - offset;
1306                 if (len < used)
1307                         used = len;
1308
1309                 /* Do we have urgent data here? */
1310                 if (tp->urg_data) {
1311                         u32 urg_offset = tp->urg_seq - *seq;
1312                         if (urg_offset < used) {
1313                                 if (!urg_offset) {
1314                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1315                                                 ++*seq;
1316                                                 offset++;
1317                                                 used--;
1318                                                 if (!used)
1319                                                         goto skip_copy;
1320                                         }
1321                                 } else
1322                                         used = urg_offset;
1323                         }
1324                 }
1325
1326                 if (!(flags & MSG_TRUNC)) {
1327                         err = skb_copy_datagram_iovec(skb, offset,
1328                                                       msg->msg_iov, used);
1329                         if (err) {
1330                                 /* Exception. Bailout! */
1331                                 if (!copied)
1332                                         copied = -EFAULT;
1333                                 break;
1334                         }
1335                 }
1336
1337                 *seq += used;
1338                 copied += used;
1339                 len -= used;
1340
1341                 tcp_rcv_space_adjust(sk);
1342
1343 skip_copy:
1344                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1345                         tp->urg_data = 0;
1346                         tcp_fast_path_check(sk, tp);
1347                 }
1348                 if (used + offset < skb->len)
1349                         continue;
1350
1351                 if (skb->h.th->fin)
1352                         goto found_fin_ok;
1353                 if (!(flags & MSG_PEEK))
1354                         sk_eat_skb(sk, skb);
1355                 continue;
1356
1357         found_fin_ok:
1358                 /* Process the FIN. */
1359                 ++*seq;
1360                 if (!(flags & MSG_PEEK))
1361                         sk_eat_skb(sk, skb);
1362                 break;
1363         } while (len > 0);
1364
1365         if (user_recv) {
1366                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1367                         int chunk;
1368
1369                         tp->ucopy.len = copied > 0 ? len : 0;
1370
1371                         tcp_prequeue_process(sk);
1372
1373                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1374                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1375                                 len -= chunk;
1376                                 copied += chunk;
1377                         }
1378                 }
1379
1380                 tp->ucopy.task = NULL;
1381                 tp->ucopy.len = 0;
1382         }
1383
1384         /* According to UNIX98, msg_name/msg_namelen are ignored
1385          * on connected socket. I was just happy when found this 8) --ANK
1386          */
1387
1388         /* Clean up data we have read: This will do ACK frames. */
1389         cleanup_rbuf(sk, copied);
1390
1391         TCP_CHECK_TIMER(sk);
1392         release_sock(sk);
1393         return copied;
1394
1395 out:
1396         TCP_CHECK_TIMER(sk);
1397         release_sock(sk);
1398         return err;
1399
1400 recv_urg:
1401         err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1402         goto out;
1403 }
1404
1405 /*
1406  *      State processing on a close. This implements the state shift for
1407  *      sending our FIN frame. Note that we only send a FIN for some
1408  *      states. A shutdown() may have already sent the FIN, or we may be
1409  *      closed.
1410  */
1411
1412 static const unsigned char new_state[16] = {
1413   /* current state:        new state:      action:      */
1414   /* (Invalid)          */ TCP_CLOSE,
1415   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1416   /* TCP_SYN_SENT       */ TCP_CLOSE,
1417   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1418   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1419   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1420   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1421   /* TCP_CLOSE          */ TCP_CLOSE,
1422   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1423   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1424   /* TCP_LISTEN         */ TCP_CLOSE,
1425   /* TCP_CLOSING        */ TCP_CLOSING,
1426 };
1427
1428 static int tcp_close_state(struct sock *sk)
1429 {
1430         int next = (int)new_state[sk->sk_state];
1431         int ns = next & TCP_STATE_MASK;
1432
1433         tcp_set_state(sk, ns);
1434
1435         return next & TCP_ACTION_FIN;
1436 }
1437
1438 /*
1439  *      Shutdown the sending side of a connection. Much like close except
1440  *      that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1441  */
1442
1443 void tcp_shutdown(struct sock *sk, int how)
1444 {
1445         /*      We need to grab some memory, and put together a FIN,
1446          *      and then put it into the queue to be sent.
1447          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1448          */
1449         if (!(how & SEND_SHUTDOWN))
1450                 return;
1451
1452         /* If we've already sent a FIN, or it's a closed state, skip this. */
1453         if ((1 << sk->sk_state) &
1454             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1455              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1456                 /* Clear out any half completed packets.  FIN if needed. */
1457                 if (tcp_close_state(sk))
1458                         tcp_send_fin(sk);
1459         }
1460 }
1461
1462 void tcp_close(struct sock *sk, long timeout)
1463 {
1464         struct sk_buff *skb;
1465         int data_was_unread = 0;
1466         int state;
1467
1468         lock_sock(sk);
1469         sk->sk_shutdown = SHUTDOWN_MASK;
1470
1471         if (sk->sk_state == TCP_LISTEN) {
1472                 tcp_set_state(sk, TCP_CLOSE);
1473
1474                 /* Special case. */
1475                 inet_csk_listen_stop(sk);
1476
1477                 goto adjudge_to_death;
1478         }
1479
1480         /*  We need to flush the recv. buffs.  We do this only on the
1481          *  descriptor close, not protocol-sourced closes, because the
1482          *  reader process may not have drained the data yet!
1483          */
1484         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1485                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1486                           skb->h.th->fin;
1487                 data_was_unread += len;
1488                 __kfree_skb(skb);
1489         }
1490
1491         sk_stream_mem_reclaim(sk);
1492
1493         /* As outlined in draft-ietf-tcpimpl-prob-03.txt, section
1494          * 3.10, we send a RST here because data was lost.  To
1495          * witness the awful effects of the old behavior of always
1496          * doing a FIN, run an older 2.1.x kernel or 2.0.x, start
1497          * a bulk GET in an FTP client, suspend the process, wait
1498          * for the client to advertise a zero window, then kill -9
1499          * the FTP client, wheee...  Note: timeout is always zero
1500          * in such a case.
1501          */
1502         if (data_was_unread) {
1503                 /* Unread data was tossed, zap the connection. */
1504                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1505                 tcp_set_state(sk, TCP_CLOSE);
1506                 tcp_send_active_reset(sk, GFP_KERNEL);
1507         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1508                 /* Check zero linger _after_ checking for unread data. */
1509                 sk->sk_prot->disconnect(sk, 0);
1510                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1511         } else if (tcp_close_state(sk)) {
1512                 /* We FIN if the application ate all the data before
1513                  * zapping the connection.
1514                  */
1515
1516                 /* RED-PEN. Formally speaking, we have broken TCP state
1517                  * machine. State transitions:
1518                  *
1519                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1520                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1521                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1522                  *
1523                  * are legal only when FIN has been sent (i.e. in window),
1524                  * rather than queued out of window. Purists blame.
1525                  *
1526                  * F.e. "RFC state" is ESTABLISHED,
1527                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1528                  *
1529                  * The visible declinations are that sometimes
1530                  * we enter time-wait state, when it is not required really
1531                  * (harmless), do not send active resets, when they are
1532                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1533                  * they look as CLOSING or LAST_ACK for Linux)
1534                  * Probably, I missed some more holelets.
1535                  *                                              --ANK
1536                  */
1537                 tcp_send_fin(sk);
1538         }
1539
1540         sk_stream_wait_close(sk, timeout);
1541
1542 adjudge_to_death:
1543         state = sk->sk_state;
1544         sock_hold(sk);
1545         sock_orphan(sk);
1546         atomic_inc(sk->sk_prot->orphan_count);
1547
1548         /* It is the last release_sock in its life. It will remove backlog. */
1549         release_sock(sk);
1550
1551
1552         /* Now socket is owned by kernel and we acquire BH lock
1553            to finish close. No need to check for user refs.
1554          */
1555         local_bh_disable();
1556         bh_lock_sock(sk);
1557         BUG_TRAP(!sock_owned_by_user(sk));
1558
1559         /* Have we already been destroyed by a softirq or backlog? */
1560         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1561                 goto out;
1562
1563         /*      This is a (useful) BSD violating of the RFC. There is a
1564          *      problem with TCP as specified in that the other end could
1565          *      keep a socket open forever with no application left this end.
1566          *      We use a 3 minute timeout (about the same as BSD) then kill
1567          *      our end. If they send after that then tough - BUT: long enough
1568          *      that we won't make the old 4*rto = almost no time - whoops
1569          *      reset mistake.
1570          *
1571          *      Nope, it was not mistake. It is really desired behaviour
1572          *      f.e. on http servers, when such sockets are useless, but
1573          *      consume significant resources. Let's do it with special
1574          *      linger2 option.                                 --ANK
1575          */
1576
1577         if (sk->sk_state == TCP_FIN_WAIT2) {
1578                 struct tcp_sock *tp = tcp_sk(sk);
1579                 if (tp->linger2 < 0) {
1580                         tcp_set_state(sk, TCP_CLOSE);
1581                         tcp_send_active_reset(sk, GFP_ATOMIC);
1582                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1583                 } else {
1584                         const int tmo = tcp_fin_time(sk);
1585
1586                         if (tmo > TCP_TIMEWAIT_LEN) {
1587                                 inet_csk_reset_keepalive_timer(sk, tcp_fin_time(sk));
1588                         } else {
1589                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1590                                 goto out;
1591                         }
1592                 }
1593         }
1594         if (sk->sk_state != TCP_CLOSE) {
1595                 sk_stream_mem_reclaim(sk);
1596                 if (atomic_read(sk->sk_prot->orphan_count) > sysctl_tcp_max_orphans ||
1597                     (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
1598                      atomic_read(&tcp_memory_allocated) > sysctl_tcp_mem[2])) {
1599                         if (net_ratelimit())
1600                                 printk(KERN_INFO "TCP: too many of orphaned "
1601                                        "sockets\n");
1602                         tcp_set_state(sk, TCP_CLOSE);
1603                         tcp_send_active_reset(sk, GFP_ATOMIC);
1604                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1605                 }
1606         }
1607
1608         if (sk->sk_state == TCP_CLOSE)
1609                 inet_csk_destroy_sock(sk);
1610         /* Otherwise, socket is reprieved until protocol close. */
1611
1612 out:
1613         bh_unlock_sock(sk);
1614         local_bh_enable();
1615         sock_put(sk);
1616 }
1617
1618 /* These states need RST on ABORT according to RFC793 */
1619
1620 static inline int tcp_need_reset(int state)
1621 {
1622         return (1 << state) &
1623                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1624                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1625 }
1626
1627 int tcp_disconnect(struct sock *sk, int flags)
1628 {
1629         struct inet_sock *inet = inet_sk(sk);
1630         struct inet_connection_sock *icsk = inet_csk(sk);
1631         struct tcp_sock *tp = tcp_sk(sk);
1632         int err = 0;
1633         int old_state = sk->sk_state;
1634
1635         if (old_state != TCP_CLOSE)
1636                 tcp_set_state(sk, TCP_CLOSE);
1637
1638         /* ABORT function of RFC793 */
1639         if (old_state == TCP_LISTEN) {
1640                 inet_csk_listen_stop(sk);
1641         } else if (tcp_need_reset(old_state) ||
1642                    (tp->snd_nxt != tp->write_seq &&
1643                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1644                 /* The last check adjusts for discrepancy of Linux wrt. RFC
1645                  * states
1646                  */
1647                 tcp_send_active_reset(sk, gfp_any());
1648                 sk->sk_err = ECONNRESET;
1649         } else if (old_state == TCP_SYN_SENT)
1650                 sk->sk_err = ECONNRESET;
1651
1652         tcp_clear_xmit_timers(sk);
1653         __skb_queue_purge(&sk->sk_receive_queue);
1654         sk_stream_writequeue_purge(sk);
1655         __skb_queue_purge(&tp->out_of_order_queue);
1656
1657         inet->dport = 0;
1658
1659         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1660                 inet_reset_saddr(sk);
1661
1662         sk->sk_shutdown = 0;
1663         sock_reset_flag(sk, SOCK_DONE);
1664         tp->srtt = 0;
1665         if ((tp->write_seq += tp->max_window + 2) == 0)
1666                 tp->write_seq = 1;
1667         icsk->icsk_backoff = 0;
1668         tp->snd_cwnd = 2;
1669         icsk->icsk_probes_out = 0;
1670         tp->packets_out = 0;
1671         tp->snd_ssthresh = 0x7fffffff;
1672         tp->snd_cwnd_cnt = 0;
1673         tp->bytes_acked = 0;
1674         tcp_set_ca_state(sk, TCP_CA_Open);
1675         tcp_clear_retrans(tp);
1676         inet_csk_delack_init(sk);
1677         sk->sk_send_head = NULL;
1678         tp->rx_opt.saw_tstamp = 0;
1679         tcp_sack_reset(&tp->rx_opt);
1680         __sk_dst_reset(sk);
1681
1682         BUG_TRAP(!inet->num || icsk->icsk_bind_hash);
1683
1684         sk->sk_error_report(sk);
1685         return err;
1686 }
1687
1688 /*
1689  *      Socket option code for TCP.
1690  */
1691 static int do_tcp_setsockopt(struct sock *sk, int level,
1692                 int optname, char __user *optval, int optlen)
1693 {
1694         struct tcp_sock *tp = tcp_sk(sk);
1695         struct inet_connection_sock *icsk = inet_csk(sk);
1696         int val;
1697         int err = 0;
1698
1699         /* This is a string value all the others are int's */
1700         if (optname == TCP_CONGESTION) {
1701                 char name[TCP_CA_NAME_MAX];
1702
1703                 if (optlen < 1)
1704                         return -EINVAL;
1705
1706                 val = strncpy_from_user(name, optval,
1707                                         min(TCP_CA_NAME_MAX-1, optlen));
1708                 if (val < 0)
1709                         return -EFAULT;
1710                 name[val] = 0;
1711
1712                 lock_sock(sk);
1713                 err = tcp_set_congestion_control(sk, name);
1714                 release_sock(sk);
1715                 return err;
1716         }
1717
1718         if (optlen < sizeof(int))
1719                 return -EINVAL;
1720
1721         if (get_user(val, (int __user *)optval))
1722                 return -EFAULT;
1723
1724         lock_sock(sk);
1725
1726         switch (optname) {
1727         case TCP_MAXSEG:
1728                 /* Values greater than interface MTU won't take effect. However
1729                  * at the point when this call is done we typically don't yet
1730                  * know which interface is going to be used */
1731                 if (val < 8 || val > MAX_TCP_WINDOW) {
1732                         err = -EINVAL;
1733                         break;
1734                 }
1735                 tp->rx_opt.user_mss = val;
1736                 break;
1737
1738         case TCP_NODELAY:
1739                 if (val) {
1740                         /* TCP_NODELAY is weaker than TCP_CORK, so that
1741                          * this option on corked socket is remembered, but
1742                          * it is not activated until cork is cleared.
1743                          *
1744                          * However, when TCP_NODELAY is set we make
1745                          * an explicit push, which overrides even TCP_CORK
1746                          * for currently queued segments.
1747                          */
1748                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
1749                         tcp_push_pending_frames(sk, tp);
1750                 } else {
1751                         tp->nonagle &= ~TCP_NAGLE_OFF;
1752                 }
1753                 break;
1754
1755         case TCP_CORK:
1756                 /* When set indicates to always queue non-full frames.
1757                  * Later the user clears this option and we transmit
1758                  * any pending partial frames in the queue.  This is
1759                  * meant to be used alongside sendfile() to get properly
1760                  * filled frames when the user (for example) must write
1761                  * out headers with a write() call first and then use
1762                  * sendfile to send out the data parts.
1763                  *
1764                  * TCP_CORK can be set together with TCP_NODELAY and it is
1765                  * stronger than TCP_NODELAY.
1766                  */
1767                 if (val) {
1768                         tp->nonagle |= TCP_NAGLE_CORK;
1769                 } else {
1770                         tp->nonagle &= ~TCP_NAGLE_CORK;
1771                         if (tp->nonagle&TCP_NAGLE_OFF)
1772                                 tp->nonagle |= TCP_NAGLE_PUSH;
1773                         tcp_push_pending_frames(sk, tp);
1774                 }
1775                 break;
1776
1777         case TCP_KEEPIDLE:
1778                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
1779                         err = -EINVAL;
1780                 else {
1781                         tp->keepalive_time = val * HZ;
1782                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
1783                             !((1 << sk->sk_state) &
1784                               (TCPF_CLOSE | TCPF_LISTEN))) {
1785                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
1786                                 if (tp->keepalive_time > elapsed)
1787                                         elapsed = tp->keepalive_time - elapsed;
1788                                 else
1789                                         elapsed = 0;
1790                                 inet_csk_reset_keepalive_timer(sk, elapsed);
1791                         }
1792                 }
1793                 break;
1794         case TCP_KEEPINTVL:
1795                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
1796                         err = -EINVAL;
1797                 else
1798                         tp->keepalive_intvl = val * HZ;
1799                 break;
1800         case TCP_KEEPCNT:
1801                 if (val < 1 || val > MAX_TCP_KEEPCNT)
1802                         err = -EINVAL;
1803                 else
1804                         tp->keepalive_probes = val;
1805                 break;
1806         case TCP_SYNCNT:
1807                 if (val < 1 || val > MAX_TCP_SYNCNT)
1808                         err = -EINVAL;
1809                 else
1810                         icsk->icsk_syn_retries = val;
1811                 break;
1812
1813         case TCP_LINGER2:
1814                 if (val < 0)
1815                         tp->linger2 = -1;
1816                 else if (val > sysctl_tcp_fin_timeout / HZ)
1817                         tp->linger2 = 0;
1818                 else
1819                         tp->linger2 = val * HZ;
1820                 break;
1821
1822         case TCP_DEFER_ACCEPT:
1823                 icsk->icsk_accept_queue.rskq_defer_accept = 0;
1824                 if (val > 0) {
1825                         /* Translate value in seconds to number of
1826                          * retransmits */
1827                         while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
1828                                val > ((TCP_TIMEOUT_INIT / HZ) <<
1829                                        icsk->icsk_accept_queue.rskq_defer_accept))
1830                                 icsk->icsk_accept_queue.rskq_defer_accept++;
1831                         icsk->icsk_accept_queue.rskq_defer_accept++;
1832                 }
1833                 break;
1834
1835         case TCP_WINDOW_CLAMP:
1836                 if (!val) {
1837                         if (sk->sk_state != TCP_CLOSE) {
1838                                 err = -EINVAL;
1839                                 break;
1840                         }
1841                         tp->window_clamp = 0;
1842                 } else
1843                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
1844                                                 SOCK_MIN_RCVBUF / 2 : val;
1845                 break;
1846
1847         case TCP_QUICKACK:
1848                 if (!val) {
1849                         icsk->icsk_ack.pingpong = 1;
1850                 } else {
1851                         icsk->icsk_ack.pingpong = 0;
1852                         if ((1 << sk->sk_state) &
1853                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
1854                             inet_csk_ack_scheduled(sk)) {
1855                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
1856                                 cleanup_rbuf(sk, 1);
1857                                 if (!(val & 1))
1858                                         icsk->icsk_ack.pingpong = 1;
1859                         }
1860                 }
1861                 break;
1862
1863         default:
1864                 err = -ENOPROTOOPT;
1865                 break;
1866         };
1867         release_sock(sk);
1868         return err;
1869 }
1870
1871 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
1872                    int optlen)
1873 {
1874         struct inet_connection_sock *icsk = inet_csk(sk);
1875
1876         if (level != SOL_TCP)
1877                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
1878                                                      optval, optlen);
1879         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1880 }
1881
1882 #ifdef CONFIG_COMPAT
1883 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
1884                           char __user *optval, int optlen)
1885 {
1886         if (level != SOL_TCP)
1887                 return inet_csk_compat_setsockopt(sk, level, optname,
1888                                                   optval, optlen);
1889         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1890 }
1891
1892 EXPORT_SYMBOL(compat_tcp_setsockopt);
1893 #endif
1894
1895 /* Return information about state of tcp endpoint in API format. */
1896 void tcp_get_info(struct sock *sk, struct tcp_info *info)
1897 {
1898         struct tcp_sock *tp = tcp_sk(sk);
1899         const struct inet_connection_sock *icsk = inet_csk(sk);
1900         u32 now = tcp_time_stamp;
1901
1902         memset(info, 0, sizeof(*info));
1903
1904         info->tcpi_state = sk->sk_state;
1905         info->tcpi_ca_state = icsk->icsk_ca_state;
1906         info->tcpi_retransmits = icsk->icsk_retransmits;
1907         info->tcpi_probes = icsk->icsk_probes_out;
1908         info->tcpi_backoff = icsk->icsk_backoff;
1909
1910         if (tp->rx_opt.tstamp_ok)
1911                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
1912         if (tp->rx_opt.sack_ok)
1913                 info->tcpi_options |= TCPI_OPT_SACK;
1914         if (tp->rx_opt.wscale_ok) {
1915                 info->tcpi_options |= TCPI_OPT_WSCALE;
1916                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
1917                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
1918         }
1919
1920         if (tp->ecn_flags&TCP_ECN_OK)
1921                 info->tcpi_options |= TCPI_OPT_ECN;
1922
1923         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
1924         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
1925         info->tcpi_snd_mss = tp->mss_cache;
1926         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
1927
1928         info->tcpi_unacked = tp->packets_out;
1929         info->tcpi_sacked = tp->sacked_out;
1930         info->tcpi_lost = tp->lost_out;
1931         info->tcpi_retrans = tp->retrans_out;
1932         info->tcpi_fackets = tp->fackets_out;
1933
1934         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
1935         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
1936         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
1937
1938         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
1939         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
1940         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
1941         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
1942         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
1943         info->tcpi_snd_cwnd = tp->snd_cwnd;
1944         info->tcpi_advmss = tp->advmss;
1945         info->tcpi_reordering = tp->reordering;
1946
1947         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
1948         info->tcpi_rcv_space = tp->rcvq_space.space;
1949
1950         info->tcpi_total_retrans = tp->total_retrans;
1951 }
1952
1953 EXPORT_SYMBOL_GPL(tcp_get_info);
1954
1955 static int do_tcp_getsockopt(struct sock *sk, int level,
1956                 int optname, char __user *optval, int __user *optlen)
1957 {
1958         struct inet_connection_sock *icsk = inet_csk(sk);
1959         struct tcp_sock *tp = tcp_sk(sk);
1960         int val, len;
1961
1962         if (get_user(len, optlen))
1963                 return -EFAULT;
1964
1965         len = min_t(unsigned int, len, sizeof(int));
1966
1967         if (len < 0)
1968                 return -EINVAL;
1969
1970         switch (optname) {
1971         case TCP_MAXSEG:
1972                 val = tp->mss_cache;
1973                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
1974                         val = tp->rx_opt.user_mss;
1975                 break;
1976         case TCP_NODELAY:
1977                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
1978                 break;
1979         case TCP_CORK:
1980                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
1981                 break;
1982         case TCP_KEEPIDLE:
1983                 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
1984                 break;
1985         case TCP_KEEPINTVL:
1986                 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
1987                 break;
1988         case TCP_KEEPCNT:
1989                 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
1990                 break;
1991         case TCP_SYNCNT:
1992                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
1993                 break;
1994         case TCP_LINGER2:
1995                 val = tp->linger2;
1996                 if (val >= 0)
1997                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
1998                 break;
1999         case TCP_DEFER_ACCEPT:
2000                 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2001                         ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2002                 break;
2003         case TCP_WINDOW_CLAMP:
2004                 val = tp->window_clamp;
2005                 break;
2006         case TCP_INFO: {
2007                 struct tcp_info info;
2008
2009                 if (get_user(len, optlen))
2010                         return -EFAULT;
2011
2012                 tcp_get_info(sk, &info);
2013
2014                 len = min_t(unsigned int, len, sizeof(info));
2015                 if (put_user(len, optlen))
2016                         return -EFAULT;
2017                 if (copy_to_user(optval, &info, len))
2018                         return -EFAULT;
2019                 return 0;
2020         }
2021         case TCP_QUICKACK:
2022                 val = !icsk->icsk_ack.pingpong;
2023                 break;
2024
2025         case TCP_CONGESTION:
2026                 if (get_user(len, optlen))
2027                         return -EFAULT;
2028                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2029                 if (put_user(len, optlen))
2030                         return -EFAULT;
2031                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2032                         return -EFAULT;
2033                 return 0;
2034         default:
2035                 return -ENOPROTOOPT;
2036         };
2037
2038         if (put_user(len, optlen))
2039                 return -EFAULT;
2040         if (copy_to_user(optval, &val, len))
2041                 return -EFAULT;
2042         return 0;
2043 }
2044
2045 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2046                    int __user *optlen)
2047 {
2048         struct inet_connection_sock *icsk = inet_csk(sk);
2049
2050         if (level != SOL_TCP)
2051                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2052                                                      optval, optlen);
2053         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2054 }
2055
2056 #ifdef CONFIG_COMPAT
2057 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2058                           char __user *optval, int __user *optlen)
2059 {
2060         if (level != SOL_TCP)
2061                 return inet_csk_compat_getsockopt(sk, level, optname,
2062                                                   optval, optlen);
2063         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2064 }
2065
2066 EXPORT_SYMBOL(compat_tcp_getsockopt);
2067 #endif
2068
2069 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2070 {
2071         struct sk_buff *segs = ERR_PTR(-EINVAL);
2072         struct tcphdr *th;
2073         unsigned thlen;
2074         unsigned int seq;
2075         unsigned int delta;
2076         unsigned int oldlen;
2077         unsigned int len;
2078
2079         if (!pskb_may_pull(skb, sizeof(*th)))
2080                 goto out;
2081
2082         th = skb->h.th;
2083         thlen = th->doff * 4;
2084         if (thlen < sizeof(*th))
2085                 goto out;
2086
2087         if (!pskb_may_pull(skb, thlen))
2088                 goto out;
2089
2090         oldlen = (u16)~skb->len;
2091         __skb_pull(skb, thlen);
2092
2093         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2094                 /* Packet is from an untrusted source, reset gso_segs. */
2095                 int mss = skb_shinfo(skb)->gso_size;
2096
2097                 skb_shinfo(skb)->gso_segs = (skb->len + mss - 1) / mss;
2098
2099                 segs = NULL;
2100                 goto out;
2101         }
2102
2103         segs = skb_segment(skb, features);
2104         if (IS_ERR(segs))
2105                 goto out;
2106
2107         len = skb_shinfo(skb)->gso_size;
2108         delta = htonl(oldlen + (thlen + len));
2109
2110         skb = segs;
2111         th = skb->h.th;
2112         seq = ntohl(th->seq);
2113
2114         do {
2115                 th->fin = th->psh = 0;
2116
2117                 th->check = ~csum_fold(th->check + delta);
2118                 if (skb->ip_summed != CHECKSUM_HW)
2119                         th->check = csum_fold(csum_partial(skb->h.raw, thlen,
2120                                                            skb->csum));
2121
2122                 seq += len;
2123                 skb = skb->next;
2124                 th = skb->h.th;
2125
2126                 th->seq = htonl(seq);
2127                 th->cwr = 0;
2128         } while (skb->next);
2129
2130         delta = htonl(oldlen + (skb->tail - skb->h.raw) + skb->data_len);
2131         th->check = ~csum_fold(th->check + delta);
2132         if (skb->ip_summed != CHECKSUM_HW)
2133                 th->check = csum_fold(csum_partial(skb->h.raw, thlen,
2134                                                    skb->csum));
2135
2136 out:
2137         return segs;
2138 }
2139
2140 extern void __skb_cb_too_small_for_tcp(int, int);
2141 extern struct tcp_congestion_ops tcp_reno;
2142
2143 static __initdata unsigned long thash_entries;
2144 static int __init set_thash_entries(char *str)
2145 {
2146         if (!str)
2147                 return 0;
2148         thash_entries = simple_strtoul(str, &str, 0);
2149         return 1;
2150 }
2151 __setup("thash_entries=", set_thash_entries);
2152
2153 void __init tcp_init(void)
2154 {
2155         struct sk_buff *skb = NULL;
2156         unsigned long limit;
2157         int order, i, max_share;
2158
2159         if (sizeof(struct tcp_skb_cb) > sizeof(skb->cb))
2160                 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb),
2161                                            sizeof(skb->cb));
2162
2163         tcp_hashinfo.bind_bucket_cachep =
2164                 kmem_cache_create("tcp_bind_bucket",
2165                                   sizeof(struct inet_bind_bucket), 0,
2166                                   SLAB_HWCACHE_ALIGN, NULL, NULL);
2167         if (!tcp_hashinfo.bind_bucket_cachep)
2168                 panic("tcp_init: Cannot alloc tcp_bind_bucket cache.");
2169
2170         /* Size and allocate the main established and bind bucket
2171          * hash tables.
2172          *
2173          * The methodology is similar to that of the buffer cache.
2174          */
2175         tcp_hashinfo.ehash =
2176                 alloc_large_system_hash("TCP established",
2177                                         sizeof(struct inet_ehash_bucket),
2178                                         thash_entries,
2179                                         (num_physpages >= 128 * 1024) ?
2180                                         13 : 15,
2181                                         HASH_HIGHMEM,
2182                                         &tcp_hashinfo.ehash_size,
2183                                         NULL,
2184                                         0);
2185         tcp_hashinfo.ehash_size = (1 << tcp_hashinfo.ehash_size) >> 1;
2186         for (i = 0; i < (tcp_hashinfo.ehash_size << 1); i++) {
2187                 rwlock_init(&tcp_hashinfo.ehash[i].lock);
2188                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].chain);
2189         }
2190
2191         tcp_hashinfo.bhash =
2192                 alloc_large_system_hash("TCP bind",
2193                                         sizeof(struct inet_bind_hashbucket),
2194                                         tcp_hashinfo.ehash_size,
2195                                         (num_physpages >= 128 * 1024) ?
2196                                         13 : 15,
2197                                         HASH_HIGHMEM,
2198                                         &tcp_hashinfo.bhash_size,
2199                                         NULL,
2200                                         64 * 1024);
2201         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2202         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2203                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2204                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2205         }
2206
2207         /* Try to be a bit smarter and adjust defaults depending
2208          * on available memory.
2209          */
2210         for (order = 0; ((1 << order) << PAGE_SHIFT) <
2211                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2212                         order++)
2213                 ;
2214         if (order >= 4) {
2215                 sysctl_local_port_range[0] = 32768;
2216                 sysctl_local_port_range[1] = 61000;
2217                 tcp_death_row.sysctl_max_tw_buckets = 180000;
2218                 sysctl_tcp_max_orphans = 4096 << (order - 4);
2219                 sysctl_max_syn_backlog = 1024;
2220         } else if (order < 3) {
2221                 sysctl_local_port_range[0] = 1024 * (3 - order);
2222                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2223                 sysctl_tcp_max_orphans >>= (3 - order);
2224                 sysctl_max_syn_backlog = 128;
2225         }
2226
2227         sysctl_tcp_mem[0] =  768 << order;
2228         sysctl_tcp_mem[1] = 1024 << order;
2229         sysctl_tcp_mem[2] = 1536 << order;
2230
2231         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2232         max_share = min(4UL*1024*1024, limit);
2233
2234         sysctl_tcp_wmem[0] = SK_STREAM_MEM_QUANTUM;
2235         sysctl_tcp_wmem[1] = 16*1024;
2236         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2237
2238         sysctl_tcp_rmem[0] = SK_STREAM_MEM_QUANTUM;
2239         sysctl_tcp_rmem[1] = 87380;
2240         sysctl_tcp_rmem[2] = max(87380, max_share);
2241
2242         printk(KERN_INFO "TCP: Hash tables configured "
2243                "(established %d bind %d)\n",
2244                tcp_hashinfo.ehash_size << 1, tcp_hashinfo.bhash_size);
2245
2246         tcp_register_congestion_control(&tcp_reno);
2247 }
2248
2249 EXPORT_SYMBOL(tcp_close);
2250 EXPORT_SYMBOL(tcp_disconnect);
2251 EXPORT_SYMBOL(tcp_getsockopt);
2252 EXPORT_SYMBOL(tcp_ioctl);
2253 EXPORT_SYMBOL(tcp_poll);
2254 EXPORT_SYMBOL(tcp_read_sock);
2255 EXPORT_SYMBOL(tcp_recvmsg);
2256 EXPORT_SYMBOL(tcp_sendmsg);
2257 EXPORT_SYMBOL(tcp_sendpage);
2258 EXPORT_SYMBOL(tcp_setsockopt);
2259 EXPORT_SYMBOL(tcp_shutdown);
2260 EXPORT_SYMBOL(tcp_statistics);
2261 EXPORT_SYMBOL_GPL(cleanup_rbuf);