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
 261 #include <net/icmp.h>
 262 #include <net/tcp.h>
 263 #include <net/xfrm.h>
 264 #include <net/ip.h>
 265
 266
 267 #include <asm/uaccess.h>
 268 #include <asm/ioctls.h>
 269
 270 int sysctl_tcp_fin_timeout = TCP_FIN_TIMEOUT;
 271
 272 DEFINE_SNMP_STAT(struct tcp_mib, tcp_statistics) __read_mostly;
 273
 274 atomic_t tcp_orphan_count = ATOMIC_INIT(0);
 275
 276 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 277
 278 int sysctl_tcp_mem[3];
 279 int sysctl_tcp_wmem[3] = { 4 * 1024, 16 * 1024, 128 * 1024 };
 280 int sysctl_tcp_rmem[3] = { 4 * 1024, 87380, 87380 * 2 };
 281
 282 EXPORT_SYMBOL(sysctl_tcp_mem);
 283 EXPORT_SYMBOL(sysctl_tcp_rmem);
 284 EXPORT_SYMBOL(sysctl_tcp_wmem);
 285
 286 atomic_t tcp_memory_allocated;  /* Current allocated memory. */
 287 atomic_t tcp_sockets_allocated; /* Current number of TCP sockets. */
 288
 289 EXPORT_SYMBOL(tcp_memory_allocated);
 290 EXPORT_SYMBOL(tcp_sockets_allocated);
 291
 292 /*
 293  * Pressure flag: try to collapse.
 294  * Technical note: it is used by multiple contexts non atomically.
 295  * All the sk_stream_mem_schedule() is of this nature: accounting
 296  * is strict, actions are advisory and have some latency.
 297  */
 298 int tcp_memory_pressure;
 299
 300 EXPORT_SYMBOL(tcp_memory_pressure);
 301
 302 void tcp_enter_memory_pressure(void)
 303 {
 304         if (!tcp_memory_pressure) {
 305                 NET_INC_STATS(LINUX_MIB_TCPMEMORYPRESSURES);
 306                 tcp_memory_pressure = 1;
 307         }
 308 }
 309
 310 EXPORT_SYMBOL(tcp_enter_memory_pressure);
 311
 312 /*
 313  *      Wait for a TCP event.
 314  *
 315  *      Note that we don't need to lock the socket, as the upper poll layers
 316  *      take care of normal races (between the test and the event) and we don't
 317  *      go look at any of the socket buffers directly.
 318  */
 319 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 320 {
 321         unsigned int mask;
 322         struct sock *sk = sock->sk;
 323         struct tcp_sock *tp = tcp_sk(sk);
 324
 325         poll_wait(file, sk->sk_sleep, wait);
 326         if (sk->sk_state == TCP_LISTEN)
 327                 return inet_csk_listen_poll(sk);
 328
 329         /* Socket is not locked. We are protected from async events
 330            by poll logic and correct handling of state changes
 331            made by another threads is impossible in any case.
 332          */
 333
 334         mask = 0;
 335         if (sk->sk_err)
 336                 mask = POLLERR;
 337
 338         /*
 339          * POLLHUP is certainly not done right. But poll() doesn't
 340          * have a notion of HUP in just one direction, and for a
 341          * socket the read side is more interesting.
 342          *
 343          * Some poll() documentation says that POLLHUP is incompatible
 344          * with the POLLOUT/POLLWR flags, so somebody should check this
 345          * all. But careful, it tends to be safer to return too many
 346          * bits than too few, and you can easily break real applications
 347          * if you don't tell them that something has hung up!
 348          *
 349          * Check-me.
 350          *
 351          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
 352          * our fs/select.c). It means that after we received EOF,
 353          * poll always returns immediately, making impossible poll() on write()
 354          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
 355          * if and only if shutdown has been made in both directions.
 356          * Actually, it is interesting to look how Solaris and DUX
 357          * solve this dilemma. I would prefer, if PULLHUP were maskable,
 358          * then we could set it on SND_SHUTDOWN. BTW examples given
 359          * in Stevens' books assume exactly this behaviour, it explains
 360          * why PULLHUP is incompatible with POLLOUT.    --ANK
 361          *
 362          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 363          * blocking on fresh not-connected or disconnected socket. --ANK
 364          */
 365         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
 366                 mask |= POLLHUP;
 367         if (sk->sk_shutdown & RCV_SHUTDOWN)
 368                 mask |= POLLIN | POLLRDNORM;
 369
 370         /* Connected? */
 371         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
 372                 /* Potential race condition. If read of tp below will
 373                  * escape above sk->sk_state, we can be illegally awaken
 374                  * in SYN_* states. */
 375                 if ((tp->rcv_nxt != tp->copied_seq) &&
 376                     (tp->urg_seq != tp->copied_seq ||
 377                      tp->rcv_nxt != tp->copied_seq + 1 ||
 378                      sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data))
 379                         mask |= POLLIN | POLLRDNORM;
 380
 381                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 382                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
 383                                 mask |= POLLOUT | POLLWRNORM;
 384                         } else {  /* send SIGIO later */
 385                                 set_bit(SOCK_ASYNC_NOSPACE,
 386                                         &sk->sk_socket->flags);
 387                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 388
 389                                 /* Race breaker. If space is freed after
 390                                  * wspace test but before the flags are set,
 391                                  * IO signal will be lost.
 392                                  */
 393                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
 394                                         mask |= POLLOUT | POLLWRNORM;
 395                         }
 396                 }
 397
 398                 if (tp->urg_data & TCP_URG_VALID)
 399                         mask |= POLLPRI;
 400         }
 401         return mask;
 402 }
 403
 404 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 405 {
 406         struct tcp_sock *tp = tcp_sk(sk);
 407         int answ;
 408
 409         switch (cmd) {
 410         case SIOCINQ:
 411                 if (sk->sk_state == TCP_LISTEN)
 412                         return -EINVAL;
 413
 414                 lock_sock(sk);
 415                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 416                         answ = 0;
 417                 else if (sock_flag(sk, SOCK_URGINLINE) ||
 418                          !tp->urg_data ||
 419                          before(tp->urg_seq, tp->copied_seq) ||
 420                          !before(tp->urg_seq, tp->rcv_nxt)) {
 421                         answ = tp->rcv_nxt - tp->copied_seq;
 422
 423                         /* Subtract 1, if FIN is in queue. */
 424                         if (answ && !skb_queue_empty(&sk->sk_receive_queue))
 425                                 answ -=
 426                        ((struct sk_buff *)sk->sk_receive_queue.prev)->h.th->fin;
 427                 } else
 428                         answ = tp->urg_seq - tp->copied_seq;
 429                 release_sock(sk);
 430                 break;
 431         case SIOCATMARK:
 432                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 433                 break;
 434         case SIOCOUTQ:
 435                 if (sk->sk_state == TCP_LISTEN)
 436                         return -EINVAL;
 437
 438                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 439                         answ = 0;
 440                 else
 441                         answ = tp->write_seq - tp->snd_una;
 442                 break;
 443         default:
 444                 return -ENOIOCTLCMD;
 445         };
 446
 447         return put_user(answ, (int __user *)arg);
 448 }
 449
 450 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 451 {
 452         TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
 453         tp->pushed_seq = tp->write_seq;
 454 }
 455
 456 static inline int forced_push(struct tcp_sock *tp)
 457 {
 458         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 459 }
 460
 461 static inline void skb_entail(struct sock *sk, struct tcp_sock *tp,
 462                               struct sk_buff *skb)
 463 {
 464         skb->csum = 0;
 465         TCP_SKB_CB(skb)->seq = tp->write_seq;
 466         TCP_SKB_CB(skb)->end_seq = tp->write_seq;
 467         TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
 468         TCP_SKB_CB(skb)->sacked = 0;
 469         skb_header_release(skb);
 470         __skb_queue_tail(&sk->sk_write_queue, skb);
 471         sk_charge_skb(sk, skb);
 472         if (!sk->sk_send_head)
 473                 sk->sk_send_head = skb;
 474         if (tp->nonagle & TCP_NAGLE_PUSH)
 475                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 476 }
 477
 478 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
 479                                 struct sk_buff *skb)
 480 {
 481         if (flags & MSG_OOB) {
 482                 tp->urg_mode = 1;
 483                 tp->snd_up = tp->write_seq;
 484                 TCP_SKB_CB(skb)->sacked |= TCPCB_URG;
 485         }
 486 }
 487
 488 static inline void tcp_push(struct sock *sk, struct tcp_sock *tp, int flags,
 489                             int mss_now, int nonagle)
 490 {
 491         if (sk->sk_send_head) {
 492                 struct sk_buff *skb = sk->sk_write_queue.prev;
 493                 if (!(flags & MSG_MORE) || forced_push(tp))
 494                         tcp_mark_push(tp, skb);
 495                 tcp_mark_urg(tp, flags, skb);
 496                 __tcp_push_pending_frames(sk, tp, mss_now,
 497                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
 498         }
 499 }
 500
 501 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
 502                          size_t psize, int flags)
 503 {
 504         struct tcp_sock *tp = tcp_sk(sk);
 505         int mss_now, size_goal;
 506         int err;
 507         ssize_t copied;
 508         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 509
 510         /* Wait for a connection to finish. */
 511         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 512                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 513                         goto out_err;
 514
 515         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 516
 517         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 518         size_goal = tp->xmit_size_goal;
 519         copied = 0;
 520
 521         err = -EPIPE;
 522         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 523                 goto do_error;
 524
 525         while (psize > 0) {
 526                 struct sk_buff *skb = sk->sk_write_queue.prev;
 527                 struct page *page = pages[poffset / PAGE_SIZE];
 528                 int copy, i, can_coalesce;
 529                 int offset = poffset % PAGE_SIZE;
 530                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
 531
 532                 if (!sk->sk_send_head || (copy = size_goal - skb->len) <= 0) {
 533 new_segment:
 534                         if (!sk_stream_memory_free(sk))
 535                                 goto wait_for_sndbuf;
 536
 537                         skb = sk_stream_alloc_pskb(sk, 0, 0,
 538                                                    sk->sk_allocation);
 539                         if (!skb)
 540                                 goto wait_for_memory;
 541
 542                         skb_entail(sk, tp, skb);
 543                         copy = size_goal;
 544                 }
 545
 546                 if (copy > size)
 547                         copy = size;
 548
 549                 i = skb_shinfo(skb)->nr_frags;
 550                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 551                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
 552                         tcp_mark_push(tp, skb);
 553                         goto new_segment;
 554                 }
 555                 if (!sk_stream_wmem_schedule(sk, copy))
 556                         goto wait_for_memory;
 557
 558                 if (can_coalesce) {
 559                         skb_shinfo(skb)->frags[i - 1].size += copy;
 560                 } else {
 561                         get_page(page);
 562                         skb_fill_page_desc(skb, i, page, offset, copy);
 563                 }
 564
 565                 skb->len += copy;
 566                 skb->data_len += copy;
 567                 skb->truesize += copy;
 568                 sk->sk_wmem_queued += copy;
 569                 sk->sk_forward_alloc -= copy;
 570                 skb->ip_summed = CHECKSUM_HW;
 571                 tp->write_seq += copy;
 572                 TCP_SKB_CB(skb)->end_seq += copy;
 573                 skb_shinfo(skb)->tso_segs = 0;
 574
 575                 if (!copied)
 576                         TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
 577
 578                 copied += copy;
 579                 poffset += copy;
 580                 if (!(psize -= copy))
 581                         goto out;
 582
 583                 if (skb->len < mss_now || (flags & MSG_OOB))
 584                         continue;
 585
 586                 if (forced_push(tp)) {
 587                         tcp_mark_push(tp, skb);
 588                         __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
 589                 } else if (skb == sk->sk_send_head)
 590                         tcp_push_one(sk, mss_now);
 591                 continue;
 592
 593 wait_for_sndbuf:
 594                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 595 wait_for_memory:
 596                 if (copied)
 597                         tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
 598
 599                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 600                         goto do_error;
 601
 602                 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 603                 size_goal = tp->xmit_size_goal;
 604         }
 605
 606 out:
 607         if (copied)
 608                 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
 609         return copied;
 610
 611 do_error:
 612         if (copied)
 613                 goto out;
 614 out_err:
 615         return sk_stream_error(sk, flags, err);
 616 }
 617
 618 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
 619                      size_t size, int flags)
 620 {
 621         ssize_t res;
 622         struct sock *sk = sock->sk;
 623
 624 #define TCP_ZC_CSUM_FLAGS (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM | NETIF_F_HW_CSUM)
 625
 626         if (!(sk->sk_route_caps & NETIF_F_SG) ||
 627             !(sk->sk_route_caps & TCP_ZC_CSUM_FLAGS))
 628                 return sock_no_sendpage(sock, page, offset, size, flags);
 629
 630 #undef TCP_ZC_CSUM_FLAGS
 631
 632         lock_sock(sk);
 633         TCP_CHECK_TIMER(sk);
 634         res = do_tcp_sendpages(sk, &page, offset, size, flags);
 635         TCP_CHECK_TIMER(sk);
 636         release_sock(sk);
 637         return res;
 638 }
 639
 640 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
 641 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
 642
 643 static inline int select_size(struct sock *sk, struct tcp_sock *tp)
 644 {
 645         int tmp = tp->mss_cache;
 646
 647         if (sk->sk_route_caps & NETIF_F_SG) {
 648                 if (sk->sk_route_caps & NETIF_F_TSO)
 649                         tmp = 0;
 650                 else {
 651                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
 652
 653                         if (tmp >= pgbreak &&
 654                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
 655                                 tmp = pgbreak;
 656                 }
 657         }
 658
 659         return tmp;
 660 }
 661
 662 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
 663                 size_t size)
 664 {
 665         struct iovec *iov;
 666         struct tcp_sock *tp = tcp_sk(sk);
 667         struct sk_buff *skb;
 668         int iovlen, flags;
 669         int mss_now, size_goal;
 670         int err, copied;
 671         long timeo;
 672
 673         lock_sock(sk);
 674         TCP_CHECK_TIMER(sk);
 675
 676         flags = msg->msg_flags;
 677         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 678
 679         /* Wait for a connection to finish. */
 680         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 681                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 682                         goto out_err;
 683
 684         /* This should be in poll */
 685         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 686
 687         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 688         size_goal = tp->xmit_size_goal;
 689
 690         /* Ok commence sending. */
 691         iovlen = msg->msg_iovlen;
 692         iov = msg->msg_iov;
 693         copied = 0;
 694
 695         err = -EPIPE;
 696         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 697                 goto do_error;
 698
 699         while (--iovlen >= 0) {
 700                 int seglen = iov->iov_len;
 701                 unsigned char __user *from = iov->iov_base;
 702
 703                 iov++;
 704
 705                 while (seglen > 0) {
 706                         int copy;
 707
 708                         skb = sk->sk_write_queue.prev;
 709
 710                         if (!sk->sk_send_head ||
 711                             (copy = size_goal - skb->len) <= 0) {
 712
 713 new_segment:
 714                                 /* Allocate new segment. If the interface is SG,
 715                                  * allocate skb fitting to single page.
 716                                  */
 717                                 if (!sk_stream_memory_free(sk))
 718                                         goto wait_for_sndbuf;
 719
 720                                 skb = sk_stream_alloc_pskb(sk, select_size(sk, tp),
 721                                                            0, sk->sk_allocation);
 722                                 if (!skb)
 723                                         goto wait_for_memory;
 724
 725                                 /*
 726                                  * Check whether we can use HW checksum.
 727                                  */
 728                                 if (sk->sk_route_caps &
 729                                     (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM |
 730                                      NETIF_F_HW_CSUM))
 731                                         skb->ip_summed = CHECKSUM_HW;
 732
 733                                 skb_entail(sk, tp, skb);
 734                                 copy = size_goal;
 735                         }
 736
 737                         /* Try to append data to the end of skb. */
 738                         if (copy > seglen)
 739                                 copy = seglen;
 740
 741                         /* Where to copy to? */
 742                         if (skb_tailroom(skb) > 0) {
 743                                 /* We have some space in skb head. Superb! */
 744                                 if (copy > skb_tailroom(skb))
 745                                         copy = skb_tailroom(skb);
 746                                 if ((err = skb_add_data(skb, from, copy)) != 0)
 747                                         goto do_fault;
 748                         } else {
 749                                 int merge = 0;
 750                                 int i = skb_shinfo(skb)->nr_frags;
 751                                 struct page *page = TCP_PAGE(sk);
 752                                 int off = TCP_OFF(sk);
 753
 754                                 if (skb_can_coalesce(skb, i, page, off) &&
 755                                     off != PAGE_SIZE) {
 756                                         /* We can extend the last page
 757                                          * fragment. */
 758                                         merge = 1;
 759                                 } else if (i == MAX_SKB_FRAGS ||
 760                                            (!i &&
 761                                            !(sk->sk_route_caps & NETIF_F_SG))) {
 762                                         /* Need to add new fragment and cannot
 763                                          * do this because interface is non-SG,
 764                                          * or because all the page slots are
 765                                          * busy. */
 766                                         tcp_mark_push(tp, skb);
 767                                         goto new_segment;
 768                                 } else if (page) {
 769                                         if (off == PAGE_SIZE) {
 770                                                 put_page(page);
 771                                                 TCP_PAGE(sk) = page = NULL;
 772                                         }
 773                                 }
 774
 775                                 if (!page) {
 776                                         /* Allocate new cache page. */
 777                                         if (!(page = sk_stream_alloc_page(sk)))
 778                                                 goto wait_for_memory;
 779                                         off = 0;
 780                                 }
 781
 782                                 if (copy > PAGE_SIZE - off)
 783                                         copy = PAGE_SIZE - off;
 784
 785                                 /* Time to copy data. We are close to
 786                                  * the end! */
 787                                 err = skb_copy_to_page(sk, from, skb, page,
 788                                                        off, copy);
 789                                 if (err) {
 790                                         /* If this page was new, give it to the
 791                                          * socket so it does not get leaked.
 792                                          */
 793                                         if (!TCP_PAGE(sk)) {
 794                                                 TCP_PAGE(sk) = page;
 795                                                 TCP_OFF(sk) = 0;
 796                                         }
 797                                         goto do_error;
 798                                 }
 799
 800                                 /* Update the skb. */
 801                                 if (merge) {
 802                                         skb_shinfo(skb)->frags[i - 1].size +=
 803                                                                         copy;
 804                                 } else {
 805                                         skb_fill_page_desc(skb, i, page, off, copy);
 806                                         if (TCP_PAGE(sk)) {
 807                                                 get_page(page);
 808                                         } else if (off + copy < PAGE_SIZE) {
 809                                                 get_page(page);
 810                                                 TCP_PAGE(sk) = page;
 811                                         }
 812                                 }
 813
 814                                 TCP_OFF(sk) = off + copy;
 815                         }
 816
 817                         if (!copied)
 818                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
 819
 820                         tp->write_seq += copy;
 821                         TCP_SKB_CB(skb)->end_seq += copy;
 822                         skb_shinfo(skb)->tso_segs = 0;
 823
 824                         from += copy;
 825                         copied += copy;
 826                         if ((seglen -= copy) == 0 && iovlen == 0)
 827                                 goto out;
 828
 829                         if (skb->len < mss_now || (flags & MSG_OOB))
 830                                 continue;
 831
 832                         if (forced_push(tp)) {
 833                                 tcp_mark_push(tp, skb);
 834                                 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
 835                         } else if (skb == sk->sk_send_head)
 836                                 tcp_push_one(sk, mss_now);
 837                         continue;
 838
 839 wait_for_sndbuf:
 840                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 841 wait_for_memory:
 842                         if (copied)
 843                                 tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
 844
 845                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 846                                 goto do_error;
 847
 848                         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 849                         size_goal = tp->xmit_size_goal;
 850                 }
 851         }
 852
 853 out:
 854         if (copied)
 855                 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
 856         TCP_CHECK_TIMER(sk);
 857         release_sock(sk);
 858         return copied;
 859
 860 do_fault:
 861         if (!skb->len) {
 862                 if (sk->sk_send_head == skb)
 863                         sk->sk_send_head = NULL;
 864                 __skb_unlink(skb, &sk->sk_write_queue);
 865                 sk_stream_free_skb(sk, skb);
 866         }
 867
 868 do_error:
 869         if (copied)
 870                 goto out;
 871 out_err:
 872         err = sk_stream_error(sk, flags, err);
 873         TCP_CHECK_TIMER(sk);
 874         release_sock(sk);
 875         return err;
 876 }
 877
 878 /*
 879  *      Handle reading urgent data. BSD has very simple semantics for
 880  *      this, no blocking and very strange errors 8)
 881  */
 882
 883 static int tcp_recv_urg(struct sock *sk, long timeo,
 884                         struct msghdr *msg, int len, int flags,
 885                         int *addr_len)
 886 {
 887         struct tcp_sock *tp = tcp_sk(sk);
 888
 889         /* No URG data to read. */
 890         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
 891             tp->urg_data == TCP_URG_READ)
 892                 return -EINVAL; /* Yes this is right ! */
 893
 894         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
 895                 return -ENOTCONN;
 896
 897         if (tp->urg_data & TCP_URG_VALID) {
 898                 int err = 0;
 899                 char c = tp->urg_data;
 900
 901                 if (!(flags & MSG_PEEK))
 902                         tp->urg_data = TCP_URG_READ;
 903
 904                 /* Read urgent data. */
 905                 msg->msg_flags |= MSG_OOB;
 906
 907                 if (len > 0) {
 908                         if (!(flags & MSG_TRUNC))
 909                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
 910                         len = 1;
 911                 } else
 912                         msg->msg_flags |= MSG_TRUNC;
 913
 914                 return err ? -EFAULT : len;
 915         }
 916
 917         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
 918                 return 0;
 919
 920         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
 921          * the available implementations agree in this case:
 922          * this call should never block, independent of the
 923          * blocking state of the socket.
 924          * Mike <pall@rz.uni-karlsruhe.de>
 925          */
 926         return -EAGAIN;
 927 }
 928
 929 /* Clean up the receive buffer for full frames taken by the user,
 930  * then send an ACK if necessary.  COPIED is the number of bytes
 931  * tcp_recvmsg has given to the user so far, it speeds up the
 932  * calculation of whether or not we must ACK for the sake of
 933  * a window update.
 934  */
 935 static void cleanup_rbuf(struct sock *sk, int copied)
 936 {
 937         struct tcp_sock *tp = tcp_sk(sk);
 938         int time_to_ack = 0;
 939
 940 #if TCP_DEBUG
 941         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
 942
 943         BUG_TRAP(!skb || before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
 944 #endif
 945
 946         if (inet_csk_ack_scheduled(sk)) {
 947                 const struct inet_connection_sock *icsk = inet_csk(sk);
 948                    /* Delayed ACKs frequently hit locked sockets during bulk
 949                     * receive. */
 950                 if (icsk->icsk_ack.blocked ||
 951                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
 952                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
 953                     /*
 954                      * If this read emptied read buffer, we send ACK, if
 955                      * connection is not bidirectional, user drained
 956                      * receive buffer and there was a small segment
 957                      * in queue.
 958                      */
 959                     (copied > 0 && (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
 960                      !icsk->icsk_ack.pingpong && !atomic_read(&sk->sk_rmem_alloc)))
 961                         time_to_ack = 1;
 962         }
 963
 964         /* We send an ACK if we can now advertise a non-zero window
 965          * which has been raised "significantly".
 966          *
 967          * Even if window raised up to infinity, do not send window open ACK
 968          * in states, where we will not receive more. It is useless.
 969          */
 970         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
 971                 __u32 rcv_window_now = tcp_receive_window(tp);
 972
 973                 /* Optimize, __tcp_select_window() is not cheap. */
 974                 if (2*rcv_window_now <= tp->window_clamp) {
 975                         __u32 new_window = __tcp_select_window(sk);
 976
 977                         /* Send ACK now, if this read freed lots of space
 978                          * in our buffer. Certainly, new_window is new window.
 979                          * We can advertise it now, if it is not less than current one.
 980                          * "Lots" means "at least twice" here.
 981                          */
 982                         if (new_window && new_window >= 2 * rcv_window_now)
 983                                 time_to_ack = 1;
 984                 }
 985         }
 986         if (time_to_ack)
 987                 tcp_send_ack(sk);
 988 }
 989
 990 static void tcp_prequeue_process(struct sock *sk)
 991 {
 992         struct sk_buff *skb;
 993         struct tcp_sock *tp = tcp_sk(sk);
 994
 995         NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED);
 996
 997         /* RX process wants to run with disabled BHs, though it is not
 998          * necessary */
 999         local_bh_disable();
1000         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1001                 sk->sk_backlog_rcv(sk, skb);
1002         local_bh_enable();
1003
1004         /* Clear memory counter. */
1005         tp->ucopy.memory = 0;
1006 }
1007
1008 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1009 {
1010         struct sk_buff *skb;
1011         u32 offset;
1012
1013         skb_queue_walk(&sk->sk_receive_queue, skb) {
1014                 offset = seq - TCP_SKB_CB(skb)->seq;
1015                 if (skb->h.th->syn)
1016                         offset--;
1017                 if (offset < skb->len || skb->h.th->fin) {
1018                         *off = offset;
1019                         return skb;
1020                 }
1021         }
1022         return NULL;
1023 }
1024
1025 /*
1026  * This routine provides an alternative to tcp_recvmsg() for routines
1027  * that would like to handle copying from skbuffs directly in 'sendfile'
1028  * fashion.
1029  * Note:
1030  *      - It is assumed that the socket was locked by the caller.
1031  *      - The routine does not block.
1032  *      - At present, there is no support for reading OOB data
1033  *        or for 'peeking' the socket using this routine
1034  *        (although both would be easy to implement).
1035  */
1036 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1037                   sk_read_actor_t recv_actor)
1038 {
1039         struct sk_buff *skb;
1040         struct tcp_sock *tp = tcp_sk(sk);
1041         u32 seq = tp->copied_seq;
1042         u32 offset;
1043         int copied = 0;
1044
1045         if (sk->sk_state == TCP_LISTEN)
1046                 return -ENOTCONN;
1047         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1048                 if (offset < skb->len) {
1049                         size_t used, len;
1050
1051                         len = skb->len - offset;
1052                         /* Stop reading if we hit a patch of urgent data */
1053                         if (tp->urg_data) {
1054                                 u32 urg_offset = tp->urg_seq - seq;
1055                                 if (urg_offset < len)
1056                                         len = urg_offset;
1057                                 if (!len)
1058                                         break;
1059                         }
1060                         used = recv_actor(desc, skb, offset, len);
1061                         if (used <= len) {
1062                                 seq += used;
1063                                 copied += used;
1064                                 offset += used;
1065                         }
1066                         if (offset != skb->len)
1067                                 break;
1068                 }
1069                 if (skb->h.th->fin) {
1070                         sk_eat_skb(sk, skb);
1071                         ++seq;
1072                         break;
1073                 }
1074                 sk_eat_skb(sk, skb);
1075                 if (!desc->count)
1076                         break;
1077         }
1078         tp->copied_seq = seq;
1079
1080         tcp_rcv_space_adjust(sk);
1081
1082         /* Clean up data we have read: This will do ACK frames. */
1083         if (copied)
1084                 cleanup_rbuf(sk, copied);
1085         return copied;
1086 }
1087
1088 /*
1089  *      This routine copies from a sock struct into the user buffer.
1090  *
1091  *      Technical note: in 2.3 we work on _locked_ socket, so that
1092  *      tricks with *seq access order and skb->users are not required.
1093  *      Probably, code can be easily improved even more.
1094  */
1095
1096 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1097                 size_t len, int nonblock, int flags, int *addr_len)
1098 {
1099         struct tcp_sock *tp = tcp_sk(sk);
1100         int copied = 0;
1101         u32 peek_seq;
1102         u32 *seq;
1103         unsigned long used;
1104         int err;
1105         int target;             /* Read at least this many bytes */
1106         long timeo;
1107         struct task_struct *user_recv = NULL;
1108
1109         lock_sock(sk);
1110
1111         TCP_CHECK_TIMER(sk);
1112
1113         err = -ENOTCONN;
1114         if (sk->sk_state == TCP_LISTEN)
1115                 goto out;
1116
1117         timeo = sock_rcvtimeo(sk, nonblock);
1118
1119         /* Urgent data needs to be handled specially. */
1120         if (flags & MSG_OOB)
1121                 goto recv_urg;
1122
1123         seq = &tp->copied_seq;
1124         if (flags & MSG_PEEK) {
1125                 peek_seq = tp->copied_seq;
1126                 seq = &peek_seq;
1127         }
1128
1129         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1130
1131         do {
1132                 struct sk_buff *skb;
1133                 u32 offset;
1134
1135                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1136                 if (tp->urg_data && tp->urg_seq == *seq) {
1137                         if (copied)
1138                                 break;
1139                         if (signal_pending(current)) {
1140                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1141                                 break;
1142                         }
1143                 }
1144
1145                 /* Next get a buffer. */
1146
1147                 skb = skb_peek(&sk->sk_receive_queue);
1148                 do {
1149                         if (!skb)
1150                                 break;
1151
1152                         /* Now that we have two receive queues this
1153                          * shouldn't happen.
1154                          */
1155                         if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1156                                 printk(KERN_INFO "recvmsg bug: copied %X "
1157                                        "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1158                                 break;
1159                         }
1160                         offset = *seq - TCP_SKB_CB(skb)->seq;
1161                         if (skb->h.th->syn)
1162                                 offset--;
1163                         if (offset < skb->len)
1164                                 goto found_ok_skb;
1165                         if (skb->h.th->fin)
1166                                 goto found_fin_ok;
1167                         BUG_TRAP(flags & MSG_PEEK);
1168                         skb = skb->next;
1169                 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1170
1171                 /* Well, if we have backlog, try to process it now yet. */
1172
1173                 if (copied >= target && !sk->sk_backlog.tail)
1174                         break;
1175
1176                 if (copied) {
1177                         if (sk->sk_err ||
1178                             sk->sk_state == TCP_CLOSE ||
1179                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1180                             !timeo ||
1181                             signal_pending(current) ||
1182                             (flags & MSG_PEEK))
1183                                 break;
1184                 } else {
1185                         if (sock_flag(sk, SOCK_DONE))
1186                                 break;
1187
1188                         if (sk->sk_err) {
1189                                 copied = sock_error(sk);
1190                                 break;
1191                         }
1192
1193                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1194                                 break;
1195
1196                         if (sk->sk_state == TCP_CLOSE) {
1197                                 if (!sock_flag(sk, SOCK_DONE)) {
1198                                         /* This occurs when user tries to read
1199                                          * from never connected socket.
1200                                          */
1201                                         copied = -ENOTCONN;
1202                                         break;
1203                                 }
1204                                 break;
1205                         }
1206
1207                         if (!timeo) {
1208                                 copied = -EAGAIN;
1209                                 break;
1210                         }
1211
1212                         if (signal_pending(current)) {
1213                                 copied = sock_intr_errno(timeo);
1214                                 break;
1215                         }
1216                 }
1217
1218                 cleanup_rbuf(sk, copied);
1219
1220                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1221                         /* Install new reader */
1222                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1223                                 user_recv = current;
1224                                 tp->ucopy.task = user_recv;
1225                                 tp->ucopy.iov = msg->msg_iov;
1226                         }
1227
1228                         tp->ucopy.len = len;
1229
1230                         BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1231                                  (flags & (MSG_PEEK | MSG_TRUNC)));
1232
1233                         /* Ugly... If prequeue is not empty, we have to
1234                          * process it before releasing socket, otherwise
1235                          * order will be broken at second iteration.
1236                          * More elegant solution is required!!!
1237                          *
1238                          * Look: we have the following (pseudo)queues:
1239                          *
1240                          * 1. packets in flight
1241                          * 2. backlog
1242                          * 3. prequeue
1243                          * 4. receive_queue
1244                          *
1245                          * Each queue can be processed only if the next ones
1246                          * are empty. At this point we have empty receive_queue.
1247                          * But prequeue _can_ be not empty after 2nd iteration,
1248                          * when we jumped to start of loop because backlog
1249                          * processing added something to receive_queue.
1250                          * We cannot release_sock(), because backlog contains
1251                          * packets arrived _after_ prequeued ones.
1252                          *
1253                          * Shortly, algorithm is clear --- to process all
1254                          * the queues in order. We could make it more directly,
1255                          * requeueing packets from backlog to prequeue, if
1256                          * is not empty. It is more elegant, but eats cycles,
1257                          * unfortunately.
1258                          */
1259                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1260                                 goto do_prequeue;
1261
1262                         /* __ Set realtime policy in scheduler __ */
1263                 }
1264
1265                 if (copied >= target) {
1266                         /* Do not sleep, just process backlog. */
1267                         release_sock(sk);
1268                         lock_sock(sk);
1269                 } else
1270                         sk_wait_data(sk, &timeo);
1271
1272                 if (user_recv) {
1273                         int chunk;
1274
1275                         /* __ Restore normal policy in scheduler __ */
1276
1277                         if ((chunk = len - tp->ucopy.len) != 0) {
1278                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1279                                 len -= chunk;
1280                                 copied += chunk;
1281                         }
1282
1283                         if (tp->rcv_nxt == tp->copied_seq &&
1284                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1285 do_prequeue:
1286                                 tcp_prequeue_process(sk);
1287
1288                                 if ((chunk = len - tp->ucopy.len) != 0) {
1289                                         NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1290                                         len -= chunk;
1291                                         copied += chunk;
1292                                 }
1293                         }
1294                 }
1295                 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1296                         if (net_ratelimit())
1297                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1298                                        current->comm, current->pid);
1299                         peek_seq = tp->copied_seq;
1300                 }
1301                 continue;
1302
1303         found_ok_skb:
1304                 /* Ok so how much can we use? */
1305                 used = skb->len - offset;
1306                 if (len < used)
1307                         used = len;
1308
1309                 /* Do we have urgent data here? */
1310                 if (tp->urg_data) {
1311                         u32 urg_offset = tp->urg_seq - *seq;
1312                         if (urg_offset < used) {
1313                                 if (!urg_offset) {
1314                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1315                                                 ++*seq;
1316                                                 offset++;
1317                                                 used--;
1318                                                 if (!used)
1319                                                         goto skip_copy;
1320                                         }
1321                                 } else
1322                                         used = urg_offset;
1323                         }
1324                 }
1325
1326                 if (!(flags & MSG_TRUNC)) {
1327                         err = skb_copy_datagram_iovec(skb, offset,
1328                                                       msg->msg_iov, used);
1329                         if (err) {
1330                                 /* Exception. Bailout! */
1331                                 if (!copied)
1332                                         copied = -EFAULT;
1333                                 break;
1334                         }
1335                 }
1336
1337                 *seq += used;
1338                 copied += used;
1339                 len -= used;
1340
1341                 tcp_rcv_space_adjust(sk);
1342
1343 skip_copy:
1344                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1345                         tp->urg_data = 0;
1346                         tcp_fast_path_check(sk, tp);
1347                 }
1348                 if (used + offset < skb->len)
1349                         continue;
1350
1351                 if (skb->h.th->fin)
1352                         goto found_fin_ok;
1353                 if (!(flags & MSG_PEEK))
1354                         sk_eat_skb(sk, skb);
1355                 continue;
1356
1357         found_fin_ok:
1358                 /* Process the FIN. */
1359                 ++*seq;
1360                 if (!(flags & MSG_PEEK))
1361                         sk_eat_skb(sk, skb);
1362                 break;
1363         } while (len > 0);
1364
1365         if (user_recv) {
1366                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1367                         int chunk;
1368
1369                         tp->ucopy.len = copied > 0 ? len : 0;
1370
1371                         tcp_prequeue_process(sk);
1372
1373                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1374                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1375                                 len -= chunk;
1376                                 copied += chunk;
1377                         }
1378                 }
1379
1380                 tp->ucopy.task = NULL;
1381                 tp->ucopy.len = 0;
1382         }
1383
1384         /* According to UNIX98, msg_name/msg_namelen are ignored
1385          * on connected socket. I was just happy when found this 8) --ANK
1386          */
1387
1388         /* Clean up data we have read: This will do ACK frames. */
1389         cleanup_rbuf(sk, copied);
1390
1391         TCP_CHECK_TIMER(sk);
1392         release_sock(sk);
1393         return copied;
1394
1395 out:
1396         TCP_CHECK_TIMER(sk);
1397         release_sock(sk);
1398         return err;
1399
1400 recv_urg:
1401         err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1402         goto out;
1403 }
1404
1405 /*
1406  *      State processing on a close. This implements the state shift for
1407  *      sending our FIN frame. Note that we only send a FIN for some
1408  *      states. A shutdown() may have already sent the FIN, or we may be
1409  *      closed.
1410  */
1411
1412 static unsigned char new_state[16] = {
1413   /* current state:        new state:      action:      */
1414   /* (Invalid)          */ TCP_CLOSE,
1415   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1416   /* TCP_SYN_SENT       */ TCP_CLOSE,
1417   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1418   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1419   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1420   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1421   /* TCP_CLOSE          */ TCP_CLOSE,
1422   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1423   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1424   /* TCP_LISTEN         */ TCP_CLOSE,
1425   /* TCP_CLOSING        */ TCP_CLOSING,
1426 };
1427
1428 static int tcp_close_state(struct sock *sk)
1429 {
1430         int next = (int)new_state[sk->sk_state];
1431         int ns = next & TCP_STATE_MASK;
1432
1433         tcp_set_state(sk, ns);
1434
1435         return next & TCP_ACTION_FIN;
1436 }
1437
1438 /*
1439  *      Shutdown the sending side of a connection. Much like close except
1440  *      that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1441  */
1442
1443 void tcp_shutdown(struct sock *sk, int how)
1444 {
1445         /*      We need to grab some memory, and put together a FIN,
1446          *      and then put it into the queue to be sent.
1447          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1448          */
1449         if (!(how & SEND_SHUTDOWN))
1450                 return;
1451
1452         /* If we've already sent a FIN, or it's a closed state, skip this. */
1453         if ((1 << sk->sk_state) &
1454             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1455              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1456                 /* Clear out any half completed packets.  FIN if needed. */
1457                 if (tcp_close_state(sk))
1458                         tcp_send_fin(sk);
1459         }
1460 }
1461
1462 void tcp_close(struct sock *sk, long timeout)
1463 {
1464         struct sk_buff *skb;
1465         int data_was_unread = 0;
1466
1467         lock_sock(sk);
1468         sk->sk_shutdown = SHUTDOWN_MASK;
1469
1470         if (sk->sk_state == TCP_LISTEN) {
1471                 tcp_set_state(sk, TCP_CLOSE);
1472
1473                 /* Special case. */
1474                 inet_csk_listen_stop(sk);
1475
1476                 goto adjudge_to_death;
1477         }
1478
1479         /*  We need to flush the recv. buffs.  We do this only on the
1480          *  descriptor close, not protocol-sourced closes, because the
1481          *  reader process may not have drained the data yet!
1482          */
1483         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1484                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1485                           skb->h.th->fin;
1486                 data_was_unread += len;
1487                 __kfree_skb(skb);
1488         }
1489
1490         sk_stream_mem_reclaim(sk);
1491
1492         /* As outlined in draft-ietf-tcpimpl-prob-03.txt, section
1493          * 3.10, we send a RST here because data was lost.  To
1494          * witness the awful effects of the old behavior of always
1495          * doing a FIN, run an older 2.1.x kernel or 2.0.x, start
1496          * a bulk GET in an FTP client, suspend the process, wait
1497          * for the client to advertise a zero window, then kill -9
1498          * the FTP client, wheee...  Note: timeout is always zero
1499          * in such a case.
1500          */
1501         if (data_was_unread) {
1502                 /* Unread data was tossed, zap the connection. */
1503                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1504                 tcp_set_state(sk, TCP_CLOSE);
1505                 tcp_send_active_reset(sk, GFP_KERNEL);
1506         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1507                 /* Check zero linger _after_ checking for unread data. */
1508                 sk->sk_prot->disconnect(sk, 0);
1509                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1510         } else if (tcp_close_state(sk)) {
1511                 /* We FIN if the application ate all the data before
1512                  * zapping the connection.
1513                  */
1514
1515                 /* RED-PEN. Formally speaking, we have broken TCP state
1516                  * machine. State transitions:
1517                  *
1518                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1519                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1520                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1521                  *
1522                  * are legal only when FIN has been sent (i.e. in window),
1523                  * rather than queued out of window. Purists blame.
1524                  *
1525                  * F.e. "RFC state" is ESTABLISHED,
1526                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1527                  *
1528                  * The visible declinations are that sometimes
1529                  * we enter time-wait state, when it is not required really
1530                  * (harmless), do not send active resets, when they are
1531                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1532                  * they look as CLOSING or LAST_ACK for Linux)
1533                  * Probably, I missed some more holelets.
1534                  *                                              --ANK
1535                  */
1536                 tcp_send_fin(sk);
1537         }
1538
1539         sk_stream_wait_close(sk, timeout);
1540
1541 adjudge_to_death:
1542         /* It is the last release_sock in its life. It will remove backlog. */
1543         release_sock(sk);
1544
1545
1546         /* Now socket is owned by kernel and we acquire BH lock
1547            to finish close. No need to check for user refs.
1548          */
1549         local_bh_disable();
1550         bh_lock_sock(sk);
1551         BUG_TRAP(!sock_owned_by_user(sk));
1552
1553         sock_hold(sk);
1554         sock_orphan(sk);
1555
1556         /*      This is a (useful) BSD violating of the RFC. There is a
1557          *      problem with TCP as specified in that the other end could
1558          *      keep a socket open forever with no application left this end.
1559          *      We use a 3 minute timeout (about the same as BSD) then kill
1560          *      our end. If they send after that then tough - BUT: long enough
1561          *      that we won't make the old 4*rto = almost no time - whoops
1562          *      reset mistake.
1563          *
1564          *      Nope, it was not mistake. It is really desired behaviour
1565          *      f.e. on http servers, when such sockets are useless, but
1566          *      consume significant resources. Let's do it with special
1567          *      linger2 option.                                 --ANK
1568          */
1569
1570         if (sk->sk_state == TCP_FIN_WAIT2) {
1571                 struct tcp_sock *tp = tcp_sk(sk);
1572                 if (tp->linger2 < 0) {
1573                         tcp_set_state(sk, TCP_CLOSE);
1574                         tcp_send_active_reset(sk, GFP_ATOMIC);
1575                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1576                 } else {
1577                         const int tmo = tcp_fin_time(sk);
1578
1579                         if (tmo > TCP_TIMEWAIT_LEN) {
1580                                 inet_csk_reset_keepalive_timer(sk, tcp_fin_time(sk));
1581                         } else {
1582                                 atomic_inc(sk->sk_prot->orphan_count);
1583                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1584                                 goto out;
1585                         }
1586                 }
1587         }
1588         if (sk->sk_state != TCP_CLOSE) {
1589                 sk_stream_mem_reclaim(sk);
1590                 if (atomic_read(sk->sk_prot->orphan_count) > sysctl_tcp_max_orphans ||
1591                     (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
1592                      atomic_read(&tcp_memory_allocated) > sysctl_tcp_mem[2])) {
1593                         if (net_ratelimit())
1594                                 printk(KERN_INFO "TCP: too many of orphaned "
1595                                        "sockets\n");
1596                         tcp_set_state(sk, TCP_CLOSE);
1597                         tcp_send_active_reset(sk, GFP_ATOMIC);
1598                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1599                 }
1600         }
1601         atomic_inc(sk->sk_prot->orphan_count);
1602
1603         if (sk->sk_state == TCP_CLOSE)
1604                 inet_csk_destroy_sock(sk);
1605         /* Otherwise, socket is reprieved until protocol close. */
1606
1607 out:
1608         bh_unlock_sock(sk);
1609         local_bh_enable();
1610         sock_put(sk);
1611 }
1612
1613 /* These states need RST on ABORT according to RFC793 */
1614
1615 static inline int tcp_need_reset(int state)
1616 {
1617         return (1 << state) &
1618                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1619                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1620 }
1621
1622 int tcp_disconnect(struct sock *sk, int flags)
1623 {
1624         struct inet_sock *inet = inet_sk(sk);
1625         struct inet_connection_sock *icsk = inet_csk(sk);
1626         struct tcp_sock *tp = tcp_sk(sk);
1627         int err = 0;
1628         int old_state = sk->sk_state;
1629
1630         if (old_state != TCP_CLOSE)
1631                 tcp_set_state(sk, TCP_CLOSE);
1632
1633         /* ABORT function of RFC793 */
1634         if (old_state == TCP_LISTEN) {
1635                 inet_csk_listen_stop(sk);
1636         } else if (tcp_need_reset(old_state) ||
1637                    (tp->snd_nxt != tp->write_seq &&
1638                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1639                 /* The last check adjusts for discrepance of Linux wrt. RFC
1640                  * states
1641                  */
1642                 tcp_send_active_reset(sk, gfp_any());
1643                 sk->sk_err = ECONNRESET;
1644         } else if (old_state == TCP_SYN_SENT)
1645                 sk->sk_err = ECONNRESET;
1646
1647         tcp_clear_xmit_timers(sk);
1648         __skb_queue_purge(&sk->sk_receive_queue);
1649         sk_stream_writequeue_purge(sk);
1650         __skb_queue_purge(&tp->out_of_order_queue);
1651
1652         inet->dport = 0;
1653
1654         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1655                 inet_reset_saddr(sk);
1656
1657         sk->sk_shutdown = 0;
1658         sock_reset_flag(sk, SOCK_DONE);
1659         tp->srtt = 0;
1660         if ((tp->write_seq += tp->max_window + 2) == 0)
1661                 tp->write_seq = 1;
1662         icsk->icsk_backoff = 0;
1663         tp->snd_cwnd = 2;
1664         icsk->icsk_probes_out = 0;
1665         tp->packets_out = 0;
1666         tp->snd_ssthresh = 0x7fffffff;
1667         tp->snd_cwnd_cnt = 0;
1668         tcp_set_ca_state(sk, TCP_CA_Open);
1669         tcp_clear_retrans(tp);
1670         inet_csk_delack_init(sk);
1671         sk->sk_send_head = NULL;
1672         tp->rx_opt.saw_tstamp = 0;
1673         tcp_sack_reset(&tp->rx_opt);
1674         __sk_dst_reset(sk);
1675
1676         BUG_TRAP(!inet->num || icsk->icsk_bind_hash);
1677
1678         sk->sk_error_report(sk);
1679         return err;
1680 }
1681
1682 /*
1683  *      Socket option code for TCP.
1684  */
1685 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
1686                    int optlen)
1687 {
1688         struct tcp_sock *tp = tcp_sk(sk);
1689         struct inet_connection_sock *icsk = inet_csk(sk);
1690         int val;
1691         int err = 0;
1692
1693         if (level != SOL_TCP)
1694                 return tp->af_specific->setsockopt(sk, level, optname,
1695                                                    optval, optlen);
1696
1697         /* This is a string value all the others are int's */
1698         if (optname == TCP_CONGESTION) {
1699                 char name[TCP_CA_NAME_MAX];
1700
1701                 if (optlen < 1)
1702                         return -EINVAL;
1703
1704                 val = strncpy_from_user(name, optval,
1705                                         min(TCP_CA_NAME_MAX-1, optlen));
1706                 if (val < 0)
1707                         return -EFAULT;
1708                 name[val] = 0;
1709
1710                 lock_sock(sk);
1711                 err = tcp_set_congestion_control(sk, name);
1712                 release_sock(sk);
1713                 return err;
1714         }
1715
1716         if (optlen < sizeof(int))
1717                 return -EINVAL;
1718
1719         if (get_user(val, (int __user *)optval))
1720                 return -EFAULT;
1721
1722         lock_sock(sk);
1723
1724         switch (optname) {
1725         case TCP_MAXSEG:
1726                 /* Values greater than interface MTU won't take effect. However
1727                  * at the point when this call is done we typically don't yet
1728                  * know which interface is going to be used */
1729                 if (val < 8 || val > MAX_TCP_WINDOW) {
1730                         err = -EINVAL;
1731                         break;
1732                 }
1733                 tp->rx_opt.user_mss = val;
1734                 break;
1735
1736         case TCP_NODELAY:
1737                 if (val) {
1738                         /* TCP_NODELAY is weaker than TCP_CORK, so that
1739                          * this option on corked socket is remembered, but
1740                          * it is not activated until cork is cleared.
1741                          *
1742                          * However, when TCP_NODELAY is set we make
1743                          * an explicit push, which overrides even TCP_CORK
1744                          * for currently queued segments.
1745                          */
1746                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
1747                         tcp_push_pending_frames(sk, tp);
1748                 } else {
1749                         tp->nonagle &= ~TCP_NAGLE_OFF;
1750                 }
1751                 break;
1752
1753         case TCP_CORK:
1754                 /* When set indicates to always queue non-full frames.
1755                  * Later the user clears this option and we transmit
1756                  * any pending partial frames in the queue.  This is
1757                  * meant to be used alongside sendfile() to get properly
1758                  * filled frames when the user (for example) must write
1759                  * out headers with a write() call first and then use
1760                  * sendfile to send out the data parts.
1761                  *
1762                  * TCP_CORK can be set together with TCP_NODELAY and it is
1763                  * stronger than TCP_NODELAY.
1764                  */
1765                 if (val) {
1766                         tp->nonagle |= TCP_NAGLE_CORK;
1767                 } else {
1768                         tp->nonagle &= ~TCP_NAGLE_CORK;
1769                         if (tp->nonagle&TCP_NAGLE_OFF)
1770                                 tp->nonagle |= TCP_NAGLE_PUSH;
1771                         tcp_push_pending_frames(sk, tp);
1772                 }
1773                 break;
1774
1775         case TCP_KEEPIDLE:
1776                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
1777                         err = -EINVAL;
1778                 else {
1779                         tp->keepalive_time = val * HZ;
1780                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
1781                             !((1 << sk->sk_state) &
1782                               (TCPF_CLOSE | TCPF_LISTEN))) {
1783                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
1784                                 if (tp->keepalive_time > elapsed)
1785                                         elapsed = tp->keepalive_time - elapsed;
1786                                 else
1787                                         elapsed = 0;
1788                                 inet_csk_reset_keepalive_timer(sk, elapsed);
1789                         }
1790                 }
1791                 break;
1792         case TCP_KEEPINTVL:
1793                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
1794                         err = -EINVAL;
1795                 else
1796                         tp->keepalive_intvl = val * HZ;
1797                 break;
1798         case TCP_KEEPCNT:
1799                 if (val < 1 || val > MAX_TCP_KEEPCNT)
1800                         err = -EINVAL;
1801                 else
1802                         tp->keepalive_probes = val;
1803                 break;
1804         case TCP_SYNCNT:
1805                 if (val < 1 || val > MAX_TCP_SYNCNT)
1806                         err = -EINVAL;
1807                 else
1808                         icsk->icsk_syn_retries = val;
1809                 break;
1810
1811         case TCP_LINGER2:
1812                 if (val < 0)
1813                         tp->linger2 = -1;
1814                 else if (val > sysctl_tcp_fin_timeout / HZ)
1815                         tp->linger2 = 0;
1816                 else
1817                         tp->linger2 = val * HZ;
1818                 break;
1819
1820         case TCP_DEFER_ACCEPT:
1821                 icsk->icsk_accept_queue.rskq_defer_accept = 0;
1822                 if (val > 0) {
1823                         /* Translate value in seconds to number of
1824                          * retransmits */
1825                         while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
1826                                val > ((TCP_TIMEOUT_INIT / HZ) <<
1827                                        icsk->icsk_accept_queue.rskq_defer_accept))
1828                                 icsk->icsk_accept_queue.rskq_defer_accept++;
1829                         icsk->icsk_accept_queue.rskq_defer_accept++;
1830                 }
1831                 break;
1832
1833         case TCP_WINDOW_CLAMP:
1834                 if (!val) {
1835                         if (sk->sk_state != TCP_CLOSE) {
1836                                 err = -EINVAL;
1837                                 break;
1838                         }
1839                         tp->window_clamp = 0;
1840                 } else
1841                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
1842                                                 SOCK_MIN_RCVBUF / 2 : val;
1843                 break;
1844
1845         case TCP_QUICKACK:
1846                 if (!val) {
1847                         icsk->icsk_ack.pingpong = 1;
1848                 } else {
1849                         icsk->icsk_ack.pingpong = 0;
1850                         if ((1 << sk->sk_state) &
1851                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
1852                             inet_csk_ack_scheduled(sk)) {
1853                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
1854                                 cleanup_rbuf(sk, 1);
1855                                 if (!(val & 1))
1856                                         icsk->icsk_ack.pingpong = 1;
1857                         }
1858                 }
1859                 break;
1860
1861         default:
1862                 err = -ENOPROTOOPT;
1863                 break;
1864         };
1865         release_sock(sk);
1866         return err;
1867 }
1868
1869 /* Return information about state of tcp endpoint in API format. */
1870 void tcp_get_info(struct sock *sk, struct tcp_info *info)
1871 {
1872         struct tcp_sock *tp = tcp_sk(sk);
1873         const struct inet_connection_sock *icsk = inet_csk(sk);
1874         u32 now = tcp_time_stamp;
1875
1876         memset(info, 0, sizeof(*info));
1877
1878         info->tcpi_state = sk->sk_state;
1879         info->tcpi_ca_state = icsk->icsk_ca_state;
1880         info->tcpi_retransmits = icsk->icsk_retransmits;
1881         info->tcpi_probes = icsk->icsk_probes_out;
1882         info->tcpi_backoff = icsk->icsk_backoff;
1883
1884         if (tp->rx_opt.tstamp_ok)
1885                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
1886         if (tp->rx_opt.sack_ok)
1887                 info->tcpi_options |= TCPI_OPT_SACK;
1888         if (tp->rx_opt.wscale_ok) {
1889                 info->tcpi_options |= TCPI_OPT_WSCALE;
1890                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
1891                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
1892         }
1893
1894         if (tp->ecn_flags&TCP_ECN_OK)
1895                 info->tcpi_options |= TCPI_OPT_ECN;
1896
1897         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
1898         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
1899         info->tcpi_snd_mss = tp->mss_cache;
1900         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
1901
1902         info->tcpi_unacked = tp->packets_out;
1903         info->tcpi_sacked = tp->sacked_out;
1904         info->tcpi_lost = tp->lost_out;
1905         info->tcpi_retrans = tp->retrans_out;
1906         info->tcpi_fackets = tp->fackets_out;
1907
1908         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
1909         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
1910         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
1911
1912         info->tcpi_pmtu = tp->pmtu_cookie;
1913         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
1914         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
1915         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
1916         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
1917         info->tcpi_snd_cwnd = tp->snd_cwnd;
1918         info->tcpi_advmss = tp->advmss;
1919         info->tcpi_reordering = tp->reordering;
1920
1921         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
1922         info->tcpi_rcv_space = tp->rcvq_space.space;
1923
1924         info->tcpi_total_retrans = tp->total_retrans;
1925 }
1926
1927 EXPORT_SYMBOL_GPL(tcp_get_info);
1928
1929 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
1930                    int __user *optlen)
1931 {
1932         struct inet_connection_sock *icsk = inet_csk(sk);
1933         struct tcp_sock *tp = tcp_sk(sk);
1934         int val, len;
1935
1936         if (level != SOL_TCP)
1937                 return tp->af_specific->getsockopt(sk, level, optname,
1938                                                    optval, optlen);
1939
1940         if (get_user(len, optlen))
1941                 return -EFAULT;
1942
1943         len = min_t(unsigned int, len, sizeof(int));
1944
1945         if (len < 0)
1946                 return -EINVAL;
1947
1948         switch (optname) {
1949         case TCP_MAXSEG:
1950                 val = tp->mss_cache;
1951                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
1952                         val = tp->rx_opt.user_mss;
1953                 break;
1954         case TCP_NODELAY:
1955                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
1956                 break;
1957         case TCP_CORK:
1958                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
1959                 break;
1960         case TCP_KEEPIDLE:
1961                 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
1962                 break;
1963         case TCP_KEEPINTVL:
1964                 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
1965                 break;
1966         case TCP_KEEPCNT:
1967                 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
1968                 break;
1969         case TCP_SYNCNT:
1970                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
1971                 break;
1972         case TCP_LINGER2:
1973                 val = tp->linger2;
1974                 if (val >= 0)
1975                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
1976                 break;
1977         case TCP_DEFER_ACCEPT:
1978                 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
1979                         ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
1980                 break;
1981         case TCP_WINDOW_CLAMP:
1982                 val = tp->window_clamp;
1983                 break;
1984         case TCP_INFO: {
1985                 struct tcp_info info;
1986
1987                 if (get_user(len, optlen))
1988                         return -EFAULT;
1989
1990                 tcp_get_info(sk, &info);
1991
1992                 len = min_t(unsigned int, len, sizeof(info));
1993                 if (put_user(len, optlen))
1994                         return -EFAULT;
1995                 if (copy_to_user(optval, &info, len))
1996                         return -EFAULT;
1997                 return 0;
1998         }
1999         case TCP_QUICKACK:
2000                 val = !icsk->icsk_ack.pingpong;
2001                 break;
2002
2003         case TCP_CONGESTION:
2004                 if (get_user(len, optlen))
2005                         return -EFAULT;
2006                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2007                 if (put_user(len, optlen))
2008                         return -EFAULT;
2009                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2010                         return -EFAULT;
2011                 return 0;
2012         default:
2013                 return -ENOPROTOOPT;
2014         };
2015
2016         if (put_user(len, optlen))
2017                 return -EFAULT;
2018         if (copy_to_user(optval, &val, len))
2019                 return -EFAULT;
2020         return 0;
2021 }
2022
2023
2024 extern void __skb_cb_too_small_for_tcp(int, int);
2025 extern struct tcp_congestion_ops tcp_reno;
2026
2027 static __initdata unsigned long thash_entries;
2028 static int __init set_thash_entries(char *str)
2029 {
2030         if (!str)
2031                 return 0;
2032         thash_entries = simple_strtoul(str, &str, 0);
2033         return 1;
2034 }
2035 __setup("thash_entries=", set_thash_entries);
2036
2037 void __init tcp_init(void)
2038 {
2039         struct sk_buff *skb = NULL;
2040         int order, i;
2041
2042         if (sizeof(struct tcp_skb_cb) > sizeof(skb->cb))
2043                 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb),
2044                                            sizeof(skb->cb));
2045
2046         tcp_hashinfo.bind_bucket_cachep =
2047                 kmem_cache_create("tcp_bind_bucket",
2048                                   sizeof(struct inet_bind_bucket), 0,
2049                                   SLAB_HWCACHE_ALIGN, NULL, NULL);
2050         if (!tcp_hashinfo.bind_bucket_cachep)
2051                 panic("tcp_init: Cannot alloc tcp_bind_bucket cache.");
2052
2053         /* Size and allocate the main established and bind bucket
2054          * hash tables.
2055          *
2056          * The methodology is similar to that of the buffer cache.
2057          */
2058         tcp_hashinfo.ehash =
2059                 alloc_large_system_hash("TCP established",
2060                                         sizeof(struct inet_ehash_bucket),
2061                                         thash_entries,
2062                                         (num_physpages >= 128 * 1024) ?
2063                                                 (25 - PAGE_SHIFT) :
2064                                                 (27 - PAGE_SHIFT),
2065                                         HASH_HIGHMEM,
2066                                         &tcp_hashinfo.ehash_size,
2067                                         NULL,
2068                                         0);
2069         tcp_hashinfo.ehash_size = (1 << tcp_hashinfo.ehash_size) >> 1;
2070         for (i = 0; i < (tcp_hashinfo.ehash_size << 1); i++) {
2071                 rwlock_init(&tcp_hashinfo.ehash[i].lock);
2072                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].chain);
2073         }
2074
2075         tcp_hashinfo.bhash =
2076                 alloc_large_system_hash("TCP bind",
2077                                         sizeof(struct inet_bind_hashbucket),
2078                                         tcp_hashinfo.ehash_size,
2079                                         (num_physpages >= 128 * 1024) ?
2080                                                 (25 - PAGE_SHIFT) :
2081                                                 (27 - PAGE_SHIFT),
2082                                         HASH_HIGHMEM,
2083                                         &tcp_hashinfo.bhash_size,
2084                                         NULL,
2085                                         64 * 1024);
2086         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2087         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2088                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2089                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2090         }
2091
2092         /* Try to be a bit smarter and adjust defaults depending
2093          * on available memory.
2094          */
2095         for (order = 0; ((1 << order) << PAGE_SHIFT) <
2096                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2097                         order++)
2098                 ;
2099         if (order >= 4) {
2100                 sysctl_local_port_range[0] = 32768;
2101                 sysctl_local_port_range[1] = 61000;
2102                 tcp_death_row.sysctl_max_tw_buckets = 180000;
2103                 sysctl_tcp_max_orphans = 4096 << (order - 4);
2104                 sysctl_max_syn_backlog = 1024;
2105         } else if (order < 3) {
2106                 sysctl_local_port_range[0] = 1024 * (3 - order);
2107                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2108                 sysctl_tcp_max_orphans >>= (3 - order);
2109                 sysctl_max_syn_backlog = 128;
2110         }
2111         tcp_hashinfo.port_rover = sysctl_local_port_range[0] - 1;
2112
2113         sysctl_tcp_mem[0] =  768 << order;
2114         sysctl_tcp_mem[1] = 1024 << order;
2115         sysctl_tcp_mem[2] = 1536 << order;
2116
2117         if (order < 3) {
2118                 sysctl_tcp_wmem[2] = 64 * 1024;
2119                 sysctl_tcp_rmem[0] = PAGE_SIZE;
2120                 sysctl_tcp_rmem[1] = 43689;
2121                 sysctl_tcp_rmem[2] = 2 * 43689;
2122         }
2123
2124         printk(KERN_INFO "TCP: Hash tables configured "
2125                "(established %d bind %d)\n",
2126                tcp_hashinfo.ehash_size << 1, tcp_hashinfo.bhash_size);
2127
2128         tcp_register_congestion_control(&tcp_reno);
2129 }
2130
2131 EXPORT_SYMBOL(tcp_close);
2132 EXPORT_SYMBOL(tcp_disconnect);
2133 EXPORT_SYMBOL(tcp_getsockopt);
2134 EXPORT_SYMBOL(tcp_ioctl);
2135 EXPORT_SYMBOL(tcp_poll);
2136 EXPORT_SYMBOL(tcp_read_sock);
2137 EXPORT_SYMBOL(tcp_recvmsg);
2138 EXPORT_SYMBOL(tcp_sendmsg);
2139 EXPORT_SYMBOL(tcp_sendpage);
2140 EXPORT_SYMBOL(tcp_setsockopt);
2141 EXPORT_SYMBOL(tcp_shutdown);
2142 EXPORT_SYMBOL(tcp_statistics);