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