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/module.h>
 251 #include <linux/types.h>
 252 #include <linux/fcntl.h>
 253 #include <linux/poll.h>
 254 #include <linux/init.h>
 255 #include <linux/smp_lock.h>
 256 #include <linux/fs.h>
 257 #include <linux/random.h>
 258 #include <linux/bootmem.h>
 259 #include <linux/cache.h>
 260 #include <linux/err.h>
 261 #include <linux/crypto.h>
 262
 263 #include <net/icmp.h>
 264 #include <net/tcp.h>
 265 #include <net/xfrm.h>
 266 #include <net/ip.h>
 267 #include <net/netdma.h>
 268
 269 #include <asm/uaccess.h>
 270 #include <asm/ioctls.h>
 271
 272 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
 273
 274 DEFINE_SNMP_STAT(struct tcp_mib, tcp_statistics) __read_mostly;
 275
 276 atomic_t tcp_orphan_count = ATOMIC_INIT(0);
 277
 278 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 279
 280 int sysctl_tcp_mem[3] __read_mostly;
 281 int sysctl_tcp_wmem[3] __read_mostly;
 282 int sysctl_tcp_rmem[3] __read_mostly;
 283
 284 EXPORT_SYMBOL(sysctl_tcp_mem);
 285 EXPORT_SYMBOL(sysctl_tcp_rmem);
 286 EXPORT_SYMBOL(sysctl_tcp_wmem);
 287
 288 atomic_t tcp_memory_allocated;  /* Current allocated memory. */
 289 atomic_t tcp_sockets_allocated; /* Current number of TCP sockets. */
 290
 291 EXPORT_SYMBOL(tcp_memory_allocated);
 292 EXPORT_SYMBOL(tcp_sockets_allocated);
 293
 294 /*
 295  * Pressure flag: try to collapse.
 296  * Technical note: it is used by multiple contexts non atomically.
 297  * All the sk_stream_mem_schedule() is of this nature: accounting
 298  * is strict, actions are advisory and have some latency.
 299  */
 300 int tcp_memory_pressure;
 301
 302 EXPORT_SYMBOL(tcp_memory_pressure);
 303
 304 void tcp_enter_memory_pressure(void)
 305 {
 306         if (!tcp_memory_pressure) {
 307                 NET_INC_STATS(LINUX_MIB_TCPMEMORYPRESSURES);
 308                 tcp_memory_pressure = 1;
 309         }
 310 }
 311
 312 EXPORT_SYMBOL(tcp_enter_memory_pressure);
 313
 314 /*
 315  *      Wait for a TCP event.
 316  *
 317  *      Note that we don't need to lock the socket, as the upper poll layers
 318  *      take care of normal races (between the test and the event) and we don't
 319  *      go look at any of the socket buffers directly.
 320  */
 321 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 322 {
 323         unsigned int mask;
 324         struct sock *sk = sock->sk;
 325         struct tcp_sock *tp = tcp_sk(sk);
 326
 327         poll_wait(file, sk->sk_sleep, wait);
 328         if (sk->sk_state == TCP_LISTEN)
 329                 return inet_csk_listen_poll(sk);
 330
 331         /* Socket is not locked. We are protected from async events
 332            by poll logic and correct handling of state changes
 333            made by another threads is impossible in any case.
 334          */
 335
 336         mask = 0;
 337         if (sk->sk_err)
 338                 mask = POLLERR;
 339
 340         /*
 341          * POLLHUP is certainly not done right. But poll() doesn't
 342          * have a notion of HUP in just one direction, and for a
 343          * socket the read side is more interesting.
 344          *
 345          * Some poll() documentation says that POLLHUP is incompatible
 346          * with the POLLOUT/POLLWR flags, so somebody should check this
 347          * all. But careful, it tends to be safer to return too many
 348          * bits than too few, and you can easily break real applications
 349          * if you don't tell them that something has hung up!
 350          *
 351          * Check-me.
 352          *
 353          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
 354          * our fs/select.c). It means that after we received EOF,
 355          * poll always returns immediately, making impossible poll() on write()
 356          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
 357          * if and only if shutdown has been made in both directions.
 358          * Actually, it is interesting to look how Solaris and DUX
 359          * solve this dilemma. I would prefer, if PULLHUP were maskable,
 360          * then we could set it on SND_SHUTDOWN. BTW examples given
 361          * in Stevens' books assume exactly this behaviour, it explains
 362          * why PULLHUP is incompatible with POLLOUT.    --ANK
 363          *
 364          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 365          * blocking on fresh not-connected or disconnected socket. --ANK
 366          */
 367         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
 368                 mask |= POLLHUP;
 369         if (sk->sk_shutdown & RCV_SHUTDOWN)
 370                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
 371
 372         /* Connected? */
 373         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
 374                 /* Potential race condition. If read of tp below will
 375                  * escape above sk->sk_state, we can be illegally awaken
 376                  * in SYN_* states. */
 377                 if ((tp->rcv_nxt != tp->copied_seq) &&
 378                     (tp->urg_seq != tp->copied_seq ||
 379                      tp->rcv_nxt != tp->copied_seq + 1 ||
 380                      sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data))
 381                         mask |= POLLIN | POLLRDNORM;
 382
 383                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 384                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
 385                                 mask |= POLLOUT | POLLWRNORM;
 386                         } else {  /* send SIGIO later */
 387                                 set_bit(SOCK_ASYNC_NOSPACE,
 388                                         &sk->sk_socket->flags);
 389                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 390
 391                                 /* Race breaker. If space is freed after
 392                                  * wspace test but before the flags are set,
 393                                  * IO signal will be lost.
 394                                  */
 395                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
 396                                         mask |= POLLOUT | POLLWRNORM;
 397                         }
 398                 }
 399
 400                 if (tp->urg_data & TCP_URG_VALID)
 401                         mask |= POLLPRI;
 402         }
 403         return mask;
 404 }
 405
 406 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 407 {
 408         struct tcp_sock *tp = tcp_sk(sk);
 409         int answ;
 410
 411         switch (cmd) {
 412         case SIOCINQ:
 413                 if (sk->sk_state == TCP_LISTEN)
 414                         return -EINVAL;
 415
 416                 lock_sock(sk);
 417                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 418                         answ = 0;
 419                 else if (sock_flag(sk, SOCK_URGINLINE) ||
 420                          !tp->urg_data ||
 421                          before(tp->urg_seq, tp->copied_seq) ||
 422                          !before(tp->urg_seq, tp->rcv_nxt)) {
 423                         answ = tp->rcv_nxt - tp->copied_seq;
 424
 425                         /* Subtract 1, if FIN is in queue. */
 426                         if (answ && !skb_queue_empty(&sk->sk_receive_queue))
 427                                 answ -=
 428                        ((struct sk_buff *)sk->sk_receive_queue.prev)->h.th->fin;
 429                 } else
 430                         answ = tp->urg_seq - tp->copied_seq;
 431                 release_sock(sk);
 432                 break;
 433         case SIOCATMARK:
 434                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 435                 break;
 436         case SIOCOUTQ:
 437                 if (sk->sk_state == TCP_LISTEN)
 438                         return -EINVAL;
 439
 440                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 441                         answ = 0;
 442                 else
 443                         answ = tp->write_seq - tp->snd_una;
 444                 break;
 445         default:
 446                 return -ENOIOCTLCMD;
 447         };
 448
 449         return put_user(answ, (int __user *)arg);
 450 }
 451
 452 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 453 {
 454         TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
 455         tp->pushed_seq = tp->write_seq;
 456 }
 457
 458 static inline int forced_push(struct tcp_sock *tp)
 459 {
 460         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 461 }
 462
 463 static inline void skb_entail(struct sock *sk, struct tcp_sock *tp,
 464                               struct sk_buff *skb)
 465 {
 466         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 467
 468         skb->csum    = 0;
 469         tcb->seq     = tcb->end_seq = tp->write_seq;
 470         tcb->flags   = TCPCB_FLAG_ACK;
 471         tcb->sacked  = 0;
 472         skb_header_release(skb);
 473         __skb_queue_tail(&sk->sk_write_queue, skb);
 474         sk_charge_skb(sk, skb);
 475         if (!sk->sk_send_head)
 476                 sk->sk_send_head = skb;
 477         if (tp->nonagle & TCP_NAGLE_PUSH)
 478                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 479 }
 480
 481 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
 482                                 struct sk_buff *skb)
 483 {
 484         if (flags & MSG_OOB) {
 485                 tp->urg_mode = 1;
 486                 tp->snd_up = tp->write_seq;
 487                 TCP_SKB_CB(skb)->sacked |= TCPCB_URG;
 488         }
 489 }
 490
 491 static inline void tcp_push(struct sock *sk, struct tcp_sock *tp, int flags,
 492                             int mss_now, int nonagle)
 493 {
 494         if (sk->sk_send_head) {
 495                 struct sk_buff *skb = sk->sk_write_queue.prev;
 496                 if (!(flags & MSG_MORE) || forced_push(tp))
 497                         tcp_mark_push(tp, skb);
 498                 tcp_mark_urg(tp, flags, skb);
 499                 __tcp_push_pending_frames(sk, tp, mss_now,
 500                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
 501         }
 502 }
 503
 504 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
 505                          size_t psize, int flags)
 506 {
 507         struct tcp_sock *tp = tcp_sk(sk);
 508         int mss_now, size_goal;
 509         int err;
 510         ssize_t copied;
 511         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 512
 513         /* Wait for a connection to finish. */
 514         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 515                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 516                         goto out_err;
 517
 518         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 519
 520         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 521         size_goal = tp->xmit_size_goal;
 522         copied = 0;
 523
 524         err = -EPIPE;
 525         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 526                 goto do_error;
 527
 528         while (psize > 0) {
 529                 struct sk_buff *skb = sk->sk_write_queue.prev;
 530                 struct page *page = pages[poffset / PAGE_SIZE];
 531                 int copy, i, can_coalesce;
 532                 int offset = poffset % PAGE_SIZE;
 533                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
 534
 535                 if (!sk->sk_send_head || (copy = size_goal - skb->len) <= 0) {
 536 new_segment:
 537                         if (!sk_stream_memory_free(sk))
 538                                 goto wait_for_sndbuf;
 539
 540                         skb = sk_stream_alloc_pskb(sk, 0, 0,
 541                                                    sk->sk_allocation);
 542                         if (!skb)
 543                                 goto wait_for_memory;
 544
 545                         skb_entail(sk, tp, skb);
 546                         copy = size_goal;
 547                 }
 548
 549                 if (copy > size)
 550                         copy = size;
 551
 552                 i = skb_shinfo(skb)->nr_frags;
 553                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 554                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
 555                         tcp_mark_push(tp, skb);
 556                         goto new_segment;
 557                 }
 558                 if (!sk_stream_wmem_schedule(sk, copy))
 559                         goto wait_for_memory;
 560
 561                 if (can_coalesce) {
 562                         skb_shinfo(skb)->frags[i - 1].size += copy;
 563                 } else {
 564                         get_page(page);
 565                         skb_fill_page_desc(skb, i, page, offset, copy);
 566                 }
 567
 568                 skb->len += copy;
 569                 skb->data_len += copy;
 570                 skb->truesize += copy;
 571                 sk->sk_wmem_queued += copy;
 572                 sk->sk_forward_alloc -= copy;
 573                 skb->ip_summed = CHECKSUM_PARTIAL;
 574                 tp->write_seq += copy;
 575                 TCP_SKB_CB(skb)->end_seq += copy;
 576                 skb_shinfo(skb)->gso_segs = 0;
 577
 578                 if (!copied)
 579                         TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
 580
 581                 copied += copy;
 582                 poffset += copy;
 583                 if (!(psize -= copy))
 584                         goto out;
 585
 586                 if (skb->len < mss_now || (flags & MSG_OOB))
 587                         continue;
 588
 589                 if (forced_push(tp)) {
 590                         tcp_mark_push(tp, skb);
 591                         __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
 592                 } else if (skb == sk->sk_send_head)
 593                         tcp_push_one(sk, mss_now);
 594                 continue;
 595
 596 wait_for_sndbuf:
 597                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 598 wait_for_memory:
 599                 if (copied)
 600                         tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
 601
 602                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 603                         goto do_error;
 604
 605                 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 606                 size_goal = tp->xmit_size_goal;
 607         }
 608
 609 out:
 610         if (copied)
 611                 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
 612         return copied;
 613
 614 do_error:
 615         if (copied)
 616                 goto out;
 617 out_err:
 618         return sk_stream_error(sk, flags, err);
 619 }
 620
 621 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
 622                      size_t size, int flags)
 623 {
 624         ssize_t res;
 625         struct sock *sk = sock->sk;
 626
 627         if (!(sk->sk_route_caps & NETIF_F_SG) ||
 628             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
 629                 return sock_no_sendpage(sock, page, offset, size, flags);
 630
 631         lock_sock(sk);
 632         TCP_CHECK_TIMER(sk);
 633         res = do_tcp_sendpages(sk, &page, offset, size, flags);
 634         TCP_CHECK_TIMER(sk);
 635         release_sock(sk);
 636         return res;
 637 }
 638
 639 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
 640 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
 641
 642 static inline int select_size(struct sock *sk, struct tcp_sock *tp)
 643 {
 644         int tmp = tp->mss_cache;
 645
 646         if (sk->sk_route_caps & NETIF_F_SG) {
 647                 if (sk_can_gso(sk))
 648                         tmp = 0;
 649                 else {
 650                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
 651
 652                         if (tmp >= pgbreak &&
 653                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
 654                                 tmp = pgbreak;
 655                 }
 656         }
 657
 658         return tmp;
 659 }
 660
 661 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
 662                 size_t size)
 663 {
 664         struct iovec *iov;
 665         struct tcp_sock *tp = tcp_sk(sk);
 666         struct sk_buff *skb;
 667         int iovlen, flags;
 668         int mss_now, size_goal;
 669         int err, copied;
 670         long timeo;
 671
 672         lock_sock(sk);
 673         TCP_CHECK_TIMER(sk);
 674
 675         flags = msg->msg_flags;
 676         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 677
 678         /* Wait for a connection to finish. */
 679         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 680                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 681                         goto out_err;
 682
 683         /* This should be in poll */
 684         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 685
 686         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 687         size_goal = tp->xmit_size_goal;
 688
 689         /* Ok commence sending. */
 690         iovlen = msg->msg_iovlen;
 691         iov = msg->msg_iov;
 692         copied = 0;
 693
 694         err = -EPIPE;
 695         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 696                 goto do_error;
 697
 698         while (--iovlen >= 0) {
 699                 int seglen = iov->iov_len;
 700                 unsigned char __user *from = iov->iov_base;
 701
 702                 iov++;
 703
 704                 while (seglen > 0) {
 705                         int copy;
 706
 707                         skb = sk->sk_write_queue.prev;
 708
 709                         if (!sk->sk_send_head ||
 710                             (copy = size_goal - skb->len) <= 0) {
 711
 712 new_segment:
 713                                 /* Allocate new segment. If the interface is SG,
 714                                  * allocate skb fitting to single page.
 715                                  */
 716                                 if (!sk_stream_memory_free(sk))
 717                                         goto wait_for_sndbuf;
 718
 719                                 skb = sk_stream_alloc_pskb(sk, select_size(sk, tp),
 720                                                            0, sk->sk_allocation);
 721                                 if (!skb)
 722                                         goto wait_for_memory;
 723
 724                                 /*
 725                                  * Check whether we can use HW checksum.
 726                                  */
 727                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
 728                                         skb->ip_summed = CHECKSUM_PARTIAL;
 729
 730                                 skb_entail(sk, tp, skb);
 731                                 copy = size_goal;
 732                         }
 733
 734                         /* Try to append data to the end of skb. */
 735                         if (copy > seglen)
 736                                 copy = seglen;
 737
 738                         /* Where to copy to? */
 739                         if (skb_tailroom(skb) > 0) {
 740                                 /* We have some space in skb head. Superb! */
 741                                 if (copy > skb_tailroom(skb))
 742                                         copy = skb_tailroom(skb);
 743                                 if ((err = skb_add_data(skb, from, copy)) != 0)
 744                                         goto do_fault;
 745                         } else {
 746                                 int merge = 0;
 747                                 int i = skb_shinfo(skb)->nr_frags;
 748                                 struct page *page = TCP_PAGE(sk);
 749                                 int off = TCP_OFF(sk);
 750
 751                                 if (skb_can_coalesce(skb, i, page, off) &&
 752                                     off != PAGE_SIZE) {
 753                                         /* We can extend the last page
 754                                          * fragment. */
 755                                         merge = 1;
 756                                 } else if (i == MAX_SKB_FRAGS ||
 757                                            (!i &&
 758                                            !(sk->sk_route_caps & NETIF_F_SG))) {
 759                                         /* Need to add new fragment and cannot
 760                                          * do this because interface is non-SG,
 761                                          * or because all the page slots are
 762                                          * busy. */
 763                                         tcp_mark_push(tp, skb);
 764                                         goto new_segment;
 765                                 } else if (page) {
 766                                         if (off == PAGE_SIZE) {
 767                                                 put_page(page);
 768                                                 TCP_PAGE(sk) = page = NULL;
 769                                                 off = 0;
 770                                         }
 771                                 } else
 772                                         off = 0;
 773
 774                                 if (copy > PAGE_SIZE - off)
 775                                         copy = PAGE_SIZE - off;
 776
 777                                 if (!sk_stream_wmem_schedule(sk, copy))
 778                                         goto wait_for_memory;
 779
 780                                 if (!page) {
 781                                         /* Allocate new cache page. */
 782                                         if (!(page = sk_stream_alloc_page(sk)))
 783                                                 goto wait_for_memory;
 784                                 }
 785
 786                                 /* Time to copy data. We are close to
 787                                  * the end! */
 788                                 err = skb_copy_to_page(sk, from, skb, page,
 789                                                        off, copy);
 790                                 if (err) {
 791                                         /* If this page was new, give it to the
 792                                          * socket so it does not get leaked.
 793                                          */
 794                                         if (!TCP_PAGE(sk)) {
 795                                                 TCP_PAGE(sk) = page;
 796                                                 TCP_OFF(sk) = 0;
 797                                         }
 798                                         goto do_error;
 799                                 }
 800
 801                                 /* Update the skb. */
 802                                 if (merge) {
 803                                         skb_shinfo(skb)->frags[i - 1].size +=
 804                                                                         copy;
 805                                 } else {
 806                                         skb_fill_page_desc(skb, i, page, off, copy);
 807                                         if (TCP_PAGE(sk)) {
 808                                                 get_page(page);
 809                                         } else if (off + copy < PAGE_SIZE) {
 810                                                 get_page(page);
 811                                                 TCP_PAGE(sk) = page;
 812                                         }
 813                                 }
 814
 815                                 TCP_OFF(sk) = off + copy;
 816                         }
 817
 818                         if (!copied)
 819                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
 820
 821                         tp->write_seq += copy;
 822                         TCP_SKB_CB(skb)->end_seq += copy;
 823                         skb_shinfo(skb)->gso_segs = 0;
 824
 825                         from += copy;
 826                         copied += copy;
 827                         if ((seglen -= copy) == 0 && iovlen == 0)
 828                                 goto out;
 829
 830                         if (skb->len < mss_now || (flags & MSG_OOB))
 831                                 continue;
 832
 833                         if (forced_push(tp)) {
 834                                 tcp_mark_push(tp, skb);
 835                                 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
 836                         } else if (skb == sk->sk_send_head)
 837                                 tcp_push_one(sk, mss_now);
 838                         continue;
 839
 840 wait_for_sndbuf:
 841                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 842 wait_for_memory:
 843                         if (copied)
 844                                 tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
 845
 846                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 847                                 goto do_error;
 848
 849                         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 850                         size_goal = tp->xmit_size_goal;
 851                 }
 852         }
 853
 854 out:
 855         if (copied)
 856                 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
 857         TCP_CHECK_TIMER(sk);
 858         release_sock(sk);
 859         return copied;
 860
 861 do_fault:
 862         if (!skb->len) {
 863                 if (sk->sk_send_head == skb)
 864                         sk->sk_send_head = NULL;
 865                 __skb_unlink(skb, &sk->sk_write_queue);
 866                 sk_stream_free_skb(sk, skb);
 867         }
 868
 869 do_error:
 870         if (copied)
 871                 goto out;
 872 out_err:
 873         err = sk_stream_error(sk, flags, err);
 874         TCP_CHECK_TIMER(sk);
 875         release_sock(sk);
 876         return err;
 877 }
 878
 879 /*
 880  *      Handle reading urgent data. BSD has very simple semantics for
 881  *      this, no blocking and very strange errors 8)
 882  */
 883
 884 static int tcp_recv_urg(struct sock *sk, long timeo,
 885                         struct msghdr *msg, int len, int flags,
 886                         int *addr_len)
 887 {
 888         struct tcp_sock *tp = tcp_sk(sk);
 889
 890         /* No URG data to read. */
 891         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
 892             tp->urg_data == TCP_URG_READ)
 893                 return -EINVAL; /* Yes this is right ! */
 894
 895         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
 896                 return -ENOTCONN;
 897
 898         if (tp->urg_data & TCP_URG_VALID) {
 899                 int err = 0;
 900                 char c = tp->urg_data;
 901
 902                 if (!(flags & MSG_PEEK))
 903                         tp->urg_data = TCP_URG_READ;
 904
 905                 /* Read urgent data. */
 906                 msg->msg_flags |= MSG_OOB;
 907
 908                 if (len > 0) {
 909                         if (!(flags & MSG_TRUNC))
 910                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
 911                         len = 1;
 912                 } else
 913                         msg->msg_flags |= MSG_TRUNC;
 914
 915                 return err ? -EFAULT : len;
 916         }
 917
 918         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
 919                 return 0;
 920
 921         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
 922          * the available implementations agree in this case:
 923          * this call should never block, independent of the
 924          * blocking state of the socket.
 925          * Mike <pall@rz.uni-karlsruhe.de>
 926          */
 927         return -EAGAIN;
 928 }
 929
 930 /* Clean up the receive buffer for full frames taken by the user,
 931  * then send an ACK if necessary.  COPIED is the number of bytes
 932  * tcp_recvmsg has given to the user so far, it speeds up the
 933  * calculation of whether or not we must ACK for the sake of
 934  * a window update.
 935  */
 936 void tcp_cleanup_rbuf(struct sock *sk, int copied)
 937 {
 938         struct tcp_sock *tp = tcp_sk(sk);
 939         int time_to_ack = 0;
 940
 941 #if TCP_DEBUG
 942         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
 943
 944         BUG_TRAP(!skb || before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
 945 #endif
 946
 947         if (inet_csk_ack_scheduled(sk)) {
 948                 const struct inet_connection_sock *icsk = inet_csk(sk);
 949                    /* Delayed ACKs frequently hit locked sockets during bulk
 950                     * receive. */
 951                 if (icsk->icsk_ack.blocked ||
 952                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
 953                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
 954                     /*
 955                      * If this read emptied read buffer, we send ACK, if
 956                      * connection is not bidirectional, user drained
 957                      * receive buffer and there was a small segment
 958                      * in queue.
 959                      */
 960                     (copied > 0 &&
 961                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
 962                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
 963                        !icsk->icsk_ack.pingpong)) &&
 964                       !atomic_read(&sk->sk_rmem_alloc)))
 965                         time_to_ack = 1;
 966         }
 967
 968         /* We send an ACK if we can now advertise a non-zero window
 969          * which has been raised "significantly".
 970          *
 971          * Even if window raised up to infinity, do not send window open ACK
 972          * in states, where we will not receive more. It is useless.
 973          */
 974         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
 975                 __u32 rcv_window_now = tcp_receive_window(tp);
 976
 977                 /* Optimize, __tcp_select_window() is not cheap. */
 978                 if (2*rcv_window_now <= tp->window_clamp) {
 979                         __u32 new_window = __tcp_select_window(sk);
 980
 981                         /* Send ACK now, if this read freed lots of space
 982                          * in our buffer. Certainly, new_window is new window.
 983                          * We can advertise it now, if it is not less than current one.
 984                          * "Lots" means "at least twice" here.
 985                          */
 986                         if (new_window && new_window >= 2 * rcv_window_now)
 987                                 time_to_ack = 1;
 988                 }
 989         }
 990         if (time_to_ack)
 991                 tcp_send_ack(sk);
 992 }
 993
 994 static void tcp_prequeue_process(struct sock *sk)
 995 {
 996         struct sk_buff *skb;
 997         struct tcp_sock *tp = tcp_sk(sk);
 998
 999         NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED);
1000
1001         /* RX process wants to run with disabled BHs, though it is not
1002          * necessary */
1003         local_bh_disable();
1004         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1005                 sk->sk_backlog_rcv(sk, skb);
1006         local_bh_enable();
1007
1008         /* Clear memory counter. */
1009         tp->ucopy.memory = 0;
1010 }
1011
1012 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1013 {
1014         struct sk_buff *skb;
1015         u32 offset;
1016
1017         skb_queue_walk(&sk->sk_receive_queue, skb) {
1018                 offset = seq - TCP_SKB_CB(skb)->seq;
1019                 if (skb->h.th->syn)
1020                         offset--;
1021                 if (offset < skb->len || skb->h.th->fin) {
1022                         *off = offset;
1023                         return skb;
1024                 }
1025         }
1026         return NULL;
1027 }
1028
1029 /*
1030  * This routine provides an alternative to tcp_recvmsg() for routines
1031  * that would like to handle copying from skbuffs directly in 'sendfile'
1032  * fashion.
1033  * Note:
1034  *      - It is assumed that the socket was locked by the caller.
1035  *      - The routine does not block.
1036  *      - At present, there is no support for reading OOB data
1037  *        or for 'peeking' the socket using this routine
1038  *        (although both would be easy to implement).
1039  */
1040 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1041                   sk_read_actor_t recv_actor)
1042 {
1043         struct sk_buff *skb;
1044         struct tcp_sock *tp = tcp_sk(sk);
1045         u32 seq = tp->copied_seq;
1046         u32 offset;
1047         int copied = 0;
1048
1049         if (sk->sk_state == TCP_LISTEN)
1050                 return -ENOTCONN;
1051         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1052                 if (offset < skb->len) {
1053                         size_t used, len;
1054
1055                         len = skb->len - offset;
1056                         /* Stop reading if we hit a patch of urgent data */
1057                         if (tp->urg_data) {
1058                                 u32 urg_offset = tp->urg_seq - seq;
1059                                 if (urg_offset < len)
1060                                         len = urg_offset;
1061                                 if (!len)
1062                                         break;
1063                         }
1064                         used = recv_actor(desc, skb, offset, len);
1065                         if (used <= len) {
1066                                 seq += used;
1067                                 copied += used;
1068                                 offset += used;
1069                         }
1070                         if (offset != skb->len)
1071                                 break;
1072                 }
1073                 if (skb->h.th->fin) {
1074                         sk_eat_skb(sk, skb, 0);
1075                         ++seq;
1076                         break;
1077                 }
1078                 sk_eat_skb(sk, skb, 0);
1079                 if (!desc->count)
1080                         break;
1081         }
1082         tp->copied_seq = seq;
1083
1084         tcp_rcv_space_adjust(sk);
1085
1086         /* Clean up data we have read: This will do ACK frames. */
1087         if (copied)
1088                 tcp_cleanup_rbuf(sk, copied);
1089         return copied;
1090 }
1091
1092 /*
1093  *      This routine copies from a sock struct into the user buffer.
1094  *
1095  *      Technical note: in 2.3 we work on _locked_ socket, so that
1096  *      tricks with *seq access order and skb->users are not required.
1097  *      Probably, code can be easily improved even more.
1098  */
1099
1100 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1101                 size_t len, int nonblock, int flags, int *addr_len)
1102 {
1103         struct tcp_sock *tp = tcp_sk(sk);
1104         int copied = 0;
1105         u32 peek_seq;
1106         u32 *seq;
1107         unsigned long used;
1108         int err;
1109         int target;             /* Read at least this many bytes */
1110         long timeo;
1111         struct task_struct *user_recv = NULL;
1112         int copied_early = 0;
1113
1114         lock_sock(sk);
1115
1116         TCP_CHECK_TIMER(sk);
1117
1118         err = -ENOTCONN;
1119         if (sk->sk_state == TCP_LISTEN)
1120                 goto out;
1121
1122         timeo = sock_rcvtimeo(sk, nonblock);
1123
1124         /* Urgent data needs to be handled specially. */
1125         if (flags & MSG_OOB)
1126                 goto recv_urg;
1127
1128         seq = &tp->copied_seq;
1129         if (flags & MSG_PEEK) {
1130                 peek_seq = tp->copied_seq;
1131                 seq = &peek_seq;
1132         }
1133
1134         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1135
1136 #ifdef CONFIG_NET_DMA
1137         tp->ucopy.dma_chan = NULL;
1138         preempt_disable();
1139         if ((len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1140             !sysctl_tcp_low_latency && __get_cpu_var(softnet_data).net_dma) {
1141                 preempt_enable_no_resched();
1142                 tp->ucopy.pinned_list = dma_pin_iovec_pages(msg->msg_iov, len);
1143         } else
1144                 preempt_enable_no_resched();
1145 #endif
1146
1147         do {
1148                 struct sk_buff *skb;
1149                 u32 offset;
1150
1151                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1152                 if (tp->urg_data && tp->urg_seq == *seq) {
1153                         if (copied)
1154                                 break;
1155                         if (signal_pending(current)) {
1156                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1157                                 break;
1158                         }
1159                 }
1160
1161                 /* Next get a buffer. */
1162
1163                 skb = skb_peek(&sk->sk_receive_queue);
1164                 do {
1165                         if (!skb)
1166                                 break;
1167
1168                         /* Now that we have two receive queues this
1169                          * shouldn't happen.
1170                          */
1171                         if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1172                                 printk(KERN_INFO "recvmsg bug: copied %X "
1173                                        "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1174                                 break;
1175                         }
1176                         offset = *seq - TCP_SKB_CB(skb)->seq;
1177                         if (skb->h.th->syn)
1178                                 offset--;
1179                         if (offset < skb->len)
1180                                 goto found_ok_skb;
1181                         if (skb->h.th->fin)
1182                                 goto found_fin_ok;
1183                         BUG_TRAP(flags & MSG_PEEK);
1184                         skb = skb->next;
1185                 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1186
1187                 /* Well, if we have backlog, try to process it now yet. */
1188
1189                 if (copied >= target && !sk->sk_backlog.tail)
1190                         break;
1191
1192                 if (copied) {
1193                         if (sk->sk_err ||
1194                             sk->sk_state == TCP_CLOSE ||
1195                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1196                             !timeo ||
1197                             signal_pending(current) ||
1198                             (flags & MSG_PEEK))
1199                                 break;
1200                 } else {
1201                         if (sock_flag(sk, SOCK_DONE))
1202                                 break;
1203
1204                         if (sk->sk_err) {
1205                                 copied = sock_error(sk);
1206                                 break;
1207                         }
1208
1209                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1210                                 break;
1211
1212                         if (sk->sk_state == TCP_CLOSE) {
1213                                 if (!sock_flag(sk, SOCK_DONE)) {
1214                                         /* This occurs when user tries to read
1215                                          * from never connected socket.
1216                                          */
1217                                         copied = -ENOTCONN;
1218                                         break;
1219                                 }
1220                                 break;
1221                         }
1222
1223                         if (!timeo) {
1224                                 copied = -EAGAIN;
1225                                 break;
1226                         }
1227
1228                         if (signal_pending(current)) {
1229                                 copied = sock_intr_errno(timeo);
1230                                 break;
1231                         }
1232                 }
1233
1234                 tcp_cleanup_rbuf(sk, copied);
1235
1236                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1237                         /* Install new reader */
1238                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1239                                 user_recv = current;
1240                                 tp->ucopy.task = user_recv;
1241                                 tp->ucopy.iov = msg->msg_iov;
1242                         }
1243
1244                         tp->ucopy.len = len;
1245
1246                         BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1247                                  (flags & (MSG_PEEK | MSG_TRUNC)));
1248
1249                         /* Ugly... If prequeue is not empty, we have to
1250                          * process it before releasing socket, otherwise
1251                          * order will be broken at second iteration.
1252                          * More elegant solution is required!!!
1253                          *
1254                          * Look: we have the following (pseudo)queues:
1255                          *
1256                          * 1. packets in flight
1257                          * 2. backlog
1258                          * 3. prequeue
1259                          * 4. receive_queue
1260                          *
1261                          * Each queue can be processed only if the next ones
1262                          * are empty. At this point we have empty receive_queue.
1263                          * But prequeue _can_ be not empty after 2nd iteration,
1264                          * when we jumped to start of loop because backlog
1265                          * processing added something to receive_queue.
1266                          * We cannot release_sock(), because backlog contains
1267                          * packets arrived _after_ prequeued ones.
1268                          *
1269                          * Shortly, algorithm is clear --- to process all
1270                          * the queues in order. We could make it more directly,
1271                          * requeueing packets from backlog to prequeue, if
1272                          * is not empty. It is more elegant, but eats cycles,
1273                          * unfortunately.
1274                          */
1275                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1276                                 goto do_prequeue;
1277
1278                         /* __ Set realtime policy in scheduler __ */
1279                 }
1280
1281                 if (copied >= target) {
1282                         /* Do not sleep, just process backlog. */
1283                         release_sock(sk);
1284                         lock_sock(sk);
1285                 } else
1286                         sk_wait_data(sk, &timeo);
1287
1288 #ifdef CONFIG_NET_DMA
1289                 tp->ucopy.wakeup = 0;
1290 #endif
1291
1292                 if (user_recv) {
1293                         int chunk;
1294
1295                         /* __ Restore normal policy in scheduler __ */
1296
1297                         if ((chunk = len - tp->ucopy.len) != 0) {
1298                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1299                                 len -= chunk;
1300                                 copied += chunk;
1301                         }
1302
1303                         if (tp->rcv_nxt == tp->copied_seq &&
1304                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1305 do_prequeue:
1306                                 tcp_prequeue_process(sk);
1307
1308                                 if ((chunk = len - tp->ucopy.len) != 0) {
1309                                         NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1310                                         len -= chunk;
1311                                         copied += chunk;
1312                                 }
1313                         }
1314                 }
1315                 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1316                         if (net_ratelimit())
1317                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1318                                        current->comm, current->pid);
1319                         peek_seq = tp->copied_seq;
1320                 }
1321                 continue;
1322
1323         found_ok_skb:
1324                 /* Ok so how much can we use? */
1325                 used = skb->len - offset;
1326                 if (len < used)
1327                         used = len;
1328
1329                 /* Do we have urgent data here? */
1330                 if (tp->urg_data) {
1331                         u32 urg_offset = tp->urg_seq - *seq;
1332                         if (urg_offset < used) {
1333                                 if (!urg_offset) {
1334                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1335                                                 ++*seq;
1336                                                 offset++;
1337                                                 used--;
1338                                                 if (!used)
1339                                                         goto skip_copy;
1340                                         }
1341                                 } else
1342                                         used = urg_offset;
1343                         }
1344                 }
1345
1346                 if (!(flags & MSG_TRUNC)) {
1347 #ifdef CONFIG_NET_DMA
1348                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1349                                 tp->ucopy.dma_chan = get_softnet_dma();
1350
1351                         if (tp->ucopy.dma_chan) {
1352                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1353                                         tp->ucopy.dma_chan, skb, offset,
1354                                         msg->msg_iov, used,
1355                                         tp->ucopy.pinned_list);
1356
1357                                 if (tp->ucopy.dma_cookie < 0) {
1358
1359                                         printk(KERN_ALERT "dma_cookie < 0\n");
1360
1361                                         /* Exception. Bailout! */
1362                                         if (!copied)
1363                                                 copied = -EFAULT;
1364                                         break;
1365                                 }
1366                                 if ((offset + used) == skb->len)
1367                                         copied_early = 1;
1368
1369                         } else
1370 #endif
1371                         {
1372                                 err = skb_copy_datagram_iovec(skb, offset,
1373                                                 msg->msg_iov, used);
1374                                 if (err) {
1375                                         /* Exception. Bailout! */
1376                                         if (!copied)
1377                                                 copied = -EFAULT;
1378                                         break;
1379                                 }
1380                         }
1381                 }
1382
1383                 *seq += used;
1384                 copied += used;
1385                 len -= used;
1386
1387                 tcp_rcv_space_adjust(sk);
1388
1389 skip_copy:
1390                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1391                         tp->urg_data = 0;
1392                         tcp_fast_path_check(sk, tp);
1393                 }
1394                 if (used + offset < skb->len)
1395                         continue;
1396
1397                 if (skb->h.th->fin)
1398                         goto found_fin_ok;
1399                 if (!(flags & MSG_PEEK)) {
1400                         sk_eat_skb(sk, skb, copied_early);
1401                         copied_early = 0;
1402                 }
1403                 continue;
1404
1405         found_fin_ok:
1406                 /* Process the FIN. */
1407                 ++*seq;
1408                 if (!(flags & MSG_PEEK)) {
1409                         sk_eat_skb(sk, skb, copied_early);
1410                         copied_early = 0;
1411                 }
1412                 break;
1413         } while (len > 0);
1414
1415         if (user_recv) {
1416                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1417                         int chunk;
1418
1419                         tp->ucopy.len = copied > 0 ? len : 0;
1420
1421                         tcp_prequeue_process(sk);
1422
1423                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1424                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1425                                 len -= chunk;
1426                                 copied += chunk;
1427                         }
1428                 }
1429
1430                 tp->ucopy.task = NULL;
1431                 tp->ucopy.len = 0;
1432         }
1433
1434 #ifdef CONFIG_NET_DMA
1435         if (tp->ucopy.dma_chan) {
1436                 struct sk_buff *skb;
1437                 dma_cookie_t done, used;
1438
1439                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1440
1441                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1442                                                  tp->ucopy.dma_cookie, &done,
1443                                                  &used) == DMA_IN_PROGRESS) {
1444                         /* do partial cleanup of sk_async_wait_queue */
1445                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1446                                (dma_async_is_complete(skb->dma_cookie, done,
1447                                                       used) == DMA_SUCCESS)) {
1448                                 __skb_dequeue(&sk->sk_async_wait_queue);
1449                                 kfree_skb(skb);
1450                         }
1451                 }
1452
1453                 /* Safe to free early-copied skbs now */
1454                 __skb_queue_purge(&sk->sk_async_wait_queue);
1455                 dma_chan_put(tp->ucopy.dma_chan);
1456                 tp->ucopy.dma_chan = NULL;
1457         }
1458         if (tp->ucopy.pinned_list) {
1459                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1460                 tp->ucopy.pinned_list = NULL;
1461         }
1462 #endif
1463
1464         /* According to UNIX98, msg_name/msg_namelen are ignored
1465          * on connected socket. I was just happy when found this 8) --ANK
1466          */
1467
1468         /* Clean up data we have read: This will do ACK frames. */
1469         tcp_cleanup_rbuf(sk, copied);
1470
1471         TCP_CHECK_TIMER(sk);
1472         release_sock(sk);
1473         return copied;
1474
1475 out:
1476         TCP_CHECK_TIMER(sk);
1477         release_sock(sk);
1478         return err;
1479
1480 recv_urg:
1481         err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1482         goto out;
1483 }
1484
1485 /*
1486  *      State processing on a close. This implements the state shift for
1487  *      sending our FIN frame. Note that we only send a FIN for some
1488  *      states. A shutdown() may have already sent the FIN, or we may be
1489  *      closed.
1490  */
1491
1492 static const unsigned char new_state[16] = {
1493   /* current state:        new state:      action:      */
1494   /* (Invalid)          */ TCP_CLOSE,
1495   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1496   /* TCP_SYN_SENT       */ TCP_CLOSE,
1497   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1498   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1499   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1500   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1501   /* TCP_CLOSE          */ TCP_CLOSE,
1502   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1503   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1504   /* TCP_LISTEN         */ TCP_CLOSE,
1505   /* TCP_CLOSING        */ TCP_CLOSING,
1506 };
1507
1508 static int tcp_close_state(struct sock *sk)
1509 {
1510         int next = (int)new_state[sk->sk_state];
1511         int ns = next & TCP_STATE_MASK;
1512
1513         tcp_set_state(sk, ns);
1514
1515         return next & TCP_ACTION_FIN;
1516 }
1517
1518 /*
1519  *      Shutdown the sending side of a connection. Much like close except
1520  *      that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1521  */
1522
1523 void tcp_shutdown(struct sock *sk, int how)
1524 {
1525         /*      We need to grab some memory, and put together a FIN,
1526          *      and then put it into the queue to be sent.
1527          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1528          */
1529         if (!(how & SEND_SHUTDOWN))
1530                 return;
1531
1532         /* If we've already sent a FIN, or it's a closed state, skip this. */
1533         if ((1 << sk->sk_state) &
1534             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1535              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1536                 /* Clear out any half completed packets.  FIN if needed. */
1537                 if (tcp_close_state(sk))
1538                         tcp_send_fin(sk);
1539         }
1540 }
1541
1542 void tcp_close(struct sock *sk, long timeout)
1543 {
1544         struct sk_buff *skb;
1545         int data_was_unread = 0;
1546         int state;
1547
1548         lock_sock(sk);
1549         sk->sk_shutdown = SHUTDOWN_MASK;
1550
1551         if (sk->sk_state == TCP_LISTEN) {
1552                 tcp_set_state(sk, TCP_CLOSE);
1553
1554                 /* Special case. */
1555                 inet_csk_listen_stop(sk);
1556
1557                 goto adjudge_to_death;
1558         }
1559
1560         /*  We need to flush the recv. buffs.  We do this only on the
1561          *  descriptor close, not protocol-sourced closes, because the
1562          *  reader process may not have drained the data yet!
1563          */
1564         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1565                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1566                           skb->h.th->fin;
1567                 data_was_unread += len;
1568                 __kfree_skb(skb);
1569         }
1570
1571         sk_stream_mem_reclaim(sk);
1572
1573         /* As outlined in draft-ietf-tcpimpl-prob-03.txt, section
1574          * 3.10, we send a RST here because data was lost.  To
1575          * witness the awful effects of the old behavior of always
1576          * doing a FIN, run an older 2.1.x kernel or 2.0.x, start
1577          * a bulk GET in an FTP client, suspend the process, wait
1578          * for the client to advertise a zero window, then kill -9
1579          * the FTP client, wheee...  Note: timeout is always zero
1580          * in such a case.
1581          */
1582         if (data_was_unread) {
1583                 /* Unread data was tossed, zap the connection. */
1584                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1585                 tcp_set_state(sk, TCP_CLOSE);
1586                 tcp_send_active_reset(sk, GFP_KERNEL);
1587         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1588                 /* Check zero linger _after_ checking for unread data. */
1589                 sk->sk_prot->disconnect(sk, 0);
1590                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1591         } else if (tcp_close_state(sk)) {
1592                 /* We FIN if the application ate all the data before
1593                  * zapping the connection.
1594                  */
1595
1596                 /* RED-PEN. Formally speaking, we have broken TCP state
1597                  * machine. State transitions:
1598                  *
1599                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1600                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1601                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1602                  *
1603                  * are legal only when FIN has been sent (i.e. in window),
1604                  * rather than queued out of window. Purists blame.
1605                  *
1606                  * F.e. "RFC state" is ESTABLISHED,
1607                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1608                  *
1609                  * The visible declinations are that sometimes
1610                  * we enter time-wait state, when it is not required really
1611                  * (harmless), do not send active resets, when they are
1612                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1613                  * they look as CLOSING or LAST_ACK for Linux)
1614                  * Probably, I missed some more holelets.
1615                  *                                              --ANK
1616                  */
1617                 tcp_send_fin(sk);
1618         }
1619
1620         sk_stream_wait_close(sk, timeout);
1621
1622 adjudge_to_death:
1623         state = sk->sk_state;
1624         sock_hold(sk);
1625         sock_orphan(sk);
1626         atomic_inc(sk->sk_prot->orphan_count);
1627
1628         /* It is the last release_sock in its life. It will remove backlog. */
1629         release_sock(sk);
1630
1631
1632         /* Now socket is owned by kernel and we acquire BH lock
1633            to finish close. No need to check for user refs.
1634          */
1635         local_bh_disable();
1636         bh_lock_sock(sk);
1637         BUG_TRAP(!sock_owned_by_user(sk));
1638
1639         /* Have we already been destroyed by a softirq or backlog? */
1640         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1641                 goto out;
1642
1643         /*      This is a (useful) BSD violating of the RFC. There is a
1644          *      problem with TCP as specified in that the other end could
1645          *      keep a socket open forever with no application left this end.
1646          *      We use a 3 minute timeout (about the same as BSD) then kill
1647          *      our end. If they send after that then tough - BUT: long enough
1648          *      that we won't make the old 4*rto = almost no time - whoops
1649          *      reset mistake.
1650          *
1651          *      Nope, it was not mistake. It is really desired behaviour
1652          *      f.e. on http servers, when such sockets are useless, but
1653          *      consume significant resources. Let's do it with special
1654          *      linger2 option.                                 --ANK
1655          */
1656
1657         if (sk->sk_state == TCP_FIN_WAIT2) {
1658                 struct tcp_sock *tp = tcp_sk(sk);
1659                 if (tp->linger2 < 0) {
1660                         tcp_set_state(sk, TCP_CLOSE);
1661                         tcp_send_active_reset(sk, GFP_ATOMIC);
1662                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1663                 } else {
1664                         const int tmo = tcp_fin_time(sk);
1665
1666                         if (tmo > TCP_TIMEWAIT_LEN) {
1667                                 inet_csk_reset_keepalive_timer(sk,
1668                                                 tmo - TCP_TIMEWAIT_LEN);
1669                         } else {
1670                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1671                                 goto out;
1672                         }
1673                 }
1674         }
1675         if (sk->sk_state != TCP_CLOSE) {
1676                 sk_stream_mem_reclaim(sk);
1677                 if (atomic_read(sk->sk_prot->orphan_count) > sysctl_tcp_max_orphans ||
1678                     (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
1679                      atomic_read(&tcp_memory_allocated) > sysctl_tcp_mem[2])) {
1680                         if (net_ratelimit())
1681                                 printk(KERN_INFO "TCP: too many of orphaned "
1682                                        "sockets\n");
1683                         tcp_set_state(sk, TCP_CLOSE);
1684                         tcp_send_active_reset(sk, GFP_ATOMIC);
1685                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1686                 }
1687         }
1688
1689         if (sk->sk_state == TCP_CLOSE)
1690                 inet_csk_destroy_sock(sk);
1691         /* Otherwise, socket is reprieved until protocol close. */
1692
1693 out:
1694         bh_unlock_sock(sk);
1695         local_bh_enable();
1696         sock_put(sk);
1697 }
1698
1699 /* These states need RST on ABORT according to RFC793 */
1700
1701 static inline int tcp_need_reset(int state)
1702 {
1703         return (1 << state) &
1704                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1705                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1706 }
1707
1708 int tcp_disconnect(struct sock *sk, int flags)
1709 {
1710         struct inet_sock *inet = inet_sk(sk);
1711         struct inet_connection_sock *icsk = inet_csk(sk);
1712         struct tcp_sock *tp = tcp_sk(sk);
1713         int err = 0;
1714         int old_state = sk->sk_state;
1715
1716         if (old_state != TCP_CLOSE)
1717                 tcp_set_state(sk, TCP_CLOSE);
1718
1719         /* ABORT function of RFC793 */
1720         if (old_state == TCP_LISTEN) {
1721                 inet_csk_listen_stop(sk);
1722         } else if (tcp_need_reset(old_state) ||
1723                    (tp->snd_nxt != tp->write_seq &&
1724                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1725                 /* The last check adjusts for discrepancy of Linux wrt. RFC
1726                  * states
1727                  */
1728                 tcp_send_active_reset(sk, gfp_any());
1729                 sk->sk_err = ECONNRESET;
1730         } else if (old_state == TCP_SYN_SENT)
1731                 sk->sk_err = ECONNRESET;
1732
1733         tcp_clear_xmit_timers(sk);
1734         __skb_queue_purge(&sk->sk_receive_queue);
1735         sk_stream_writequeue_purge(sk);
1736         __skb_queue_purge(&tp->out_of_order_queue);
1737 #ifdef CONFIG_NET_DMA
1738         __skb_queue_purge(&sk->sk_async_wait_queue);
1739 #endif
1740
1741         inet->dport = 0;
1742
1743         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1744                 inet_reset_saddr(sk);
1745
1746         sk->sk_shutdown = 0;
1747         sock_reset_flag(sk, SOCK_DONE);
1748         tp->srtt = 0;
1749         if ((tp->write_seq += tp->max_window + 2) == 0)
1750                 tp->write_seq = 1;
1751         icsk->icsk_backoff = 0;
1752         tp->snd_cwnd = 2;
1753         icsk->icsk_probes_out = 0;
1754         tp->packets_out = 0;
1755         tp->snd_ssthresh = 0x7fffffff;
1756         tp->snd_cwnd_cnt = 0;
1757         tp->bytes_acked = 0;
1758         tcp_set_ca_state(sk, TCP_CA_Open);
1759         tcp_clear_retrans(tp);
1760         inet_csk_delack_init(sk);
1761         sk->sk_send_head = NULL;
1762         tp->rx_opt.saw_tstamp = 0;
1763         tcp_sack_reset(&tp->rx_opt);
1764         __sk_dst_reset(sk);
1765
1766         BUG_TRAP(!inet->num || icsk->icsk_bind_hash);
1767
1768         sk->sk_error_report(sk);
1769         return err;
1770 }
1771
1772 /*
1773  *      Socket option code for TCP.
1774  */
1775 static int do_tcp_setsockopt(struct sock *sk, int level,
1776                 int optname, char __user *optval, int optlen)
1777 {
1778         struct tcp_sock *tp = tcp_sk(sk);
1779         struct inet_connection_sock *icsk = inet_csk(sk);
1780         int val;
1781         int err = 0;
1782
1783         /* This is a string value all the others are int's */
1784         if (optname == TCP_CONGESTION) {
1785                 char name[TCP_CA_NAME_MAX];
1786
1787                 if (optlen < 1)
1788                         return -EINVAL;
1789
1790                 val = strncpy_from_user(name, optval,
1791                                         min(TCP_CA_NAME_MAX-1, optlen));
1792                 if (val < 0)
1793                         return -EFAULT;
1794                 name[val] = 0;
1795
1796                 lock_sock(sk);
1797                 err = tcp_set_congestion_control(sk, name);
1798                 release_sock(sk);
1799                 return err;
1800         }
1801
1802         if (optlen < sizeof(int))
1803                 return -EINVAL;
1804
1805         if (get_user(val, (int __user *)optval))
1806                 return -EFAULT;
1807
1808         lock_sock(sk);
1809
1810         switch (optname) {
1811         case TCP_MAXSEG:
1812                 /* Values greater than interface MTU won't take effect. However
1813                  * at the point when this call is done we typically don't yet
1814                  * know which interface is going to be used */
1815                 if (val < 8 || val > MAX_TCP_WINDOW) {
1816                         err = -EINVAL;
1817                         break;
1818                 }
1819                 tp->rx_opt.user_mss = val;
1820                 break;
1821
1822         case TCP_NODELAY:
1823                 if (val) {
1824                         /* TCP_NODELAY is weaker than TCP_CORK, so that
1825                          * this option on corked socket is remembered, but
1826                          * it is not activated until cork is cleared.
1827                          *
1828                          * However, when TCP_NODELAY is set we make
1829                          * an explicit push, which overrides even TCP_CORK
1830                          * for currently queued segments.
1831                          */
1832                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
1833                         tcp_push_pending_frames(sk, tp);
1834                 } else {
1835                         tp->nonagle &= ~TCP_NAGLE_OFF;
1836                 }
1837                 break;
1838
1839         case TCP_CORK:
1840                 /* When set indicates to always queue non-full frames.
1841                  * Later the user clears this option and we transmit
1842                  * any pending partial frames in the queue.  This is
1843                  * meant to be used alongside sendfile() to get properly
1844                  * filled frames when the user (for example) must write
1845                  * out headers with a write() call first and then use
1846                  * sendfile to send out the data parts.
1847                  *
1848                  * TCP_CORK can be set together with TCP_NODELAY and it is
1849                  * stronger than TCP_NODELAY.
1850                  */
1851                 if (val) {
1852                         tp->nonagle |= TCP_NAGLE_CORK;
1853                 } else {
1854                         tp->nonagle &= ~TCP_NAGLE_CORK;
1855                         if (tp->nonagle&TCP_NAGLE_OFF)
1856                                 tp->nonagle |= TCP_NAGLE_PUSH;
1857                         tcp_push_pending_frames(sk, tp);
1858                 }
1859                 break;
1860
1861         case TCP_KEEPIDLE:
1862                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
1863                         err = -EINVAL;
1864                 else {
1865                         tp->keepalive_time = val * HZ;
1866                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
1867                             !((1 << sk->sk_state) &
1868                               (TCPF_CLOSE | TCPF_LISTEN))) {
1869                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
1870                                 if (tp->keepalive_time > elapsed)
1871                                         elapsed = tp->keepalive_time - elapsed;
1872                                 else
1873                                         elapsed = 0;
1874                                 inet_csk_reset_keepalive_timer(sk, elapsed);
1875                         }
1876                 }
1877                 break;
1878         case TCP_KEEPINTVL:
1879                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
1880                         err = -EINVAL;
1881                 else
1882                         tp->keepalive_intvl = val * HZ;
1883                 break;
1884         case TCP_KEEPCNT:
1885                 if (val < 1 || val > MAX_TCP_KEEPCNT)
1886                         err = -EINVAL;
1887                 else
1888                         tp->keepalive_probes = val;
1889                 break;
1890         case TCP_SYNCNT:
1891                 if (val < 1 || val > MAX_TCP_SYNCNT)
1892                         err = -EINVAL;
1893                 else
1894                         icsk->icsk_syn_retries = val;
1895                 break;
1896
1897         case TCP_LINGER2:
1898                 if (val < 0)
1899                         tp->linger2 = -1;
1900                 else if (val > sysctl_tcp_fin_timeout / HZ)
1901                         tp->linger2 = 0;
1902                 else
1903                         tp->linger2 = val * HZ;
1904                 break;
1905
1906         case TCP_DEFER_ACCEPT:
1907                 icsk->icsk_accept_queue.rskq_defer_accept = 0;
1908                 if (val > 0) {
1909                         /* Translate value in seconds to number of
1910                          * retransmits */
1911                         while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
1912                                val > ((TCP_TIMEOUT_INIT / HZ) <<
1913                                        icsk->icsk_accept_queue.rskq_defer_accept))
1914                                 icsk->icsk_accept_queue.rskq_defer_accept++;
1915                         icsk->icsk_accept_queue.rskq_defer_accept++;
1916                 }
1917                 break;
1918
1919         case TCP_WINDOW_CLAMP:
1920                 if (!val) {
1921                         if (sk->sk_state != TCP_CLOSE) {
1922                                 err = -EINVAL;
1923                                 break;
1924                         }
1925                         tp->window_clamp = 0;
1926                 } else
1927                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
1928                                                 SOCK_MIN_RCVBUF / 2 : val;
1929                 break;
1930
1931         case TCP_QUICKACK:
1932                 if (!val) {
1933                         icsk->icsk_ack.pingpong = 1;
1934                 } else {
1935                         icsk->icsk_ack.pingpong = 0;
1936                         if ((1 << sk->sk_state) &
1937                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
1938                             inet_csk_ack_scheduled(sk)) {
1939                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
1940                                 tcp_cleanup_rbuf(sk, 1);
1941                                 if (!(val & 1))
1942                                         icsk->icsk_ack.pingpong = 1;
1943                         }
1944                 }
1945                 break;
1946
1947 #ifdef CONFIG_TCP_MD5SIG
1948         case TCP_MD5SIG:
1949                 /* Read the IP->Key mappings from userspace */
1950                 err = tp->af_specific->md5_parse(sk, optval, optlen);
1951                 break;
1952 #endif
1953
1954         default:
1955                 err = -ENOPROTOOPT;
1956                 break;
1957         };
1958         release_sock(sk);
1959         return err;
1960 }
1961
1962 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
1963                    int optlen)
1964 {
1965         struct inet_connection_sock *icsk = inet_csk(sk);
1966
1967         if (level != SOL_TCP)
1968                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
1969                                                      optval, optlen);
1970         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1971 }
1972
1973 #ifdef CONFIG_COMPAT
1974 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
1975                           char __user *optval, int optlen)
1976 {
1977         if (level != SOL_TCP)
1978                 return inet_csk_compat_setsockopt(sk, level, optname,
1979                                                   optval, optlen);
1980         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1981 }
1982
1983 EXPORT_SYMBOL(compat_tcp_setsockopt);
1984 #endif
1985
1986 /* Return information about state of tcp endpoint in API format. */
1987 void tcp_get_info(struct sock *sk, struct tcp_info *info)
1988 {
1989         struct tcp_sock *tp = tcp_sk(sk);
1990         const struct inet_connection_sock *icsk = inet_csk(sk);
1991         u32 now = tcp_time_stamp;
1992
1993         memset(info, 0, sizeof(*info));
1994
1995         info->tcpi_state = sk->sk_state;
1996         info->tcpi_ca_state = icsk->icsk_ca_state;
1997         info->tcpi_retransmits = icsk->icsk_retransmits;
1998         info->tcpi_probes = icsk->icsk_probes_out;
1999         info->tcpi_backoff = icsk->icsk_backoff;
2000
2001         if (tp->rx_opt.tstamp_ok)
2002                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2003         if (tp->rx_opt.sack_ok)
2004                 info->tcpi_options |= TCPI_OPT_SACK;
2005         if (tp->rx_opt.wscale_ok) {
2006                 info->tcpi_options |= TCPI_OPT_WSCALE;
2007                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2008                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2009         }
2010
2011         if (tp->ecn_flags&TCP_ECN_OK)
2012                 info->tcpi_options |= TCPI_OPT_ECN;
2013
2014         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2015         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2016         info->tcpi_snd_mss = tp->mss_cache;
2017         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2018
2019         info->tcpi_unacked = tp->packets_out;
2020         info->tcpi_sacked = tp->sacked_out;
2021         info->tcpi_lost = tp->lost_out;
2022         info->tcpi_retrans = tp->retrans_out;
2023         info->tcpi_fackets = tp->fackets_out;
2024
2025         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2026         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2027         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2028
2029         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2030         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2031         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2032         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2033         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2034         info->tcpi_snd_cwnd = tp->snd_cwnd;
2035         info->tcpi_advmss = tp->advmss;
2036         info->tcpi_reordering = tp->reordering;
2037
2038         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2039         info->tcpi_rcv_space = tp->rcvq_space.space;
2040
2041         info->tcpi_total_retrans = tp->total_retrans;
2042 }
2043
2044 EXPORT_SYMBOL_GPL(tcp_get_info);
2045
2046 static int do_tcp_getsockopt(struct sock *sk, int level,
2047                 int optname, char __user *optval, int __user *optlen)
2048 {
2049         struct inet_connection_sock *icsk = inet_csk(sk);
2050         struct tcp_sock *tp = tcp_sk(sk);
2051         int val, len;
2052
2053         if (get_user(len, optlen))
2054                 return -EFAULT;
2055
2056         len = min_t(unsigned int, len, sizeof(int));
2057
2058         if (len < 0)
2059                 return -EINVAL;
2060
2061         switch (optname) {
2062         case TCP_MAXSEG:
2063                 val = tp->mss_cache;
2064                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2065                         val = tp->rx_opt.user_mss;
2066                 break;
2067         case TCP_NODELAY:
2068                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2069                 break;
2070         case TCP_CORK:
2071                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2072                 break;
2073         case TCP_KEEPIDLE:
2074                 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2075                 break;
2076         case TCP_KEEPINTVL:
2077                 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2078                 break;
2079         case TCP_KEEPCNT:
2080                 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2081                 break;
2082         case TCP_SYNCNT:
2083                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2084                 break;
2085         case TCP_LINGER2:
2086                 val = tp->linger2;
2087                 if (val >= 0)
2088                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2089                 break;
2090         case TCP_DEFER_ACCEPT:
2091                 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2092                         ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2093                 break;
2094         case TCP_WINDOW_CLAMP:
2095                 val = tp->window_clamp;
2096                 break;
2097         case TCP_INFO: {
2098                 struct tcp_info info;
2099
2100                 if (get_user(len, optlen))
2101                         return -EFAULT;
2102
2103                 tcp_get_info(sk, &info);
2104
2105                 len = min_t(unsigned int, len, sizeof(info));
2106                 if (put_user(len, optlen))
2107                         return -EFAULT;
2108                 if (copy_to_user(optval, &info, len))
2109                         return -EFAULT;
2110                 return 0;
2111         }
2112         case TCP_QUICKACK:
2113                 val = !icsk->icsk_ack.pingpong;
2114                 break;
2115
2116         case TCP_CONGESTION:
2117                 if (get_user(len, optlen))
2118                         return -EFAULT;
2119                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2120                 if (put_user(len, optlen))
2121                         return -EFAULT;
2122                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2123                         return -EFAULT;
2124                 return 0;
2125         default:
2126                 return -ENOPROTOOPT;
2127         };
2128
2129         if (put_user(len, optlen))
2130                 return -EFAULT;
2131         if (copy_to_user(optval, &val, len))
2132                 return -EFAULT;
2133         return 0;
2134 }
2135
2136 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2137                    int __user *optlen)
2138 {
2139         struct inet_connection_sock *icsk = inet_csk(sk);
2140
2141         if (level != SOL_TCP)
2142                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2143                                                      optval, optlen);
2144         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2145 }
2146
2147 #ifdef CONFIG_COMPAT
2148 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2149                           char __user *optval, int __user *optlen)
2150 {
2151         if (level != SOL_TCP)
2152                 return inet_csk_compat_getsockopt(sk, level, optname,
2153                                                   optval, optlen);
2154         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2155 }
2156
2157 EXPORT_SYMBOL(compat_tcp_getsockopt);
2158 #endif
2159
2160 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2161 {
2162         struct sk_buff *segs = ERR_PTR(-EINVAL);
2163         struct tcphdr *th;
2164         unsigned thlen;
2165         unsigned int seq;
2166         __be32 delta;
2167         unsigned int oldlen;
2168         unsigned int len;
2169
2170         if (!pskb_may_pull(skb, sizeof(*th)))
2171                 goto out;
2172
2173         th = skb->h.th;
2174         thlen = th->doff * 4;
2175         if (thlen < sizeof(*th))
2176                 goto out;
2177
2178         if (!pskb_may_pull(skb, thlen))
2179                 goto out;
2180
2181         oldlen = (u16)~skb->len;
2182         __skb_pull(skb, thlen);
2183
2184         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2185                 /* Packet is from an untrusted source, reset gso_segs. */
2186                 int type = skb_shinfo(skb)->gso_type;
2187                 int mss;
2188
2189                 if (unlikely(type &
2190                              ~(SKB_GSO_TCPV4 |
2191                                SKB_GSO_DODGY |
2192                                SKB_GSO_TCP_ECN |
2193                                SKB_GSO_TCPV6 |
2194                                0) ||
2195                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2196                         goto out;
2197
2198                 mss = skb_shinfo(skb)->gso_size;
2199                 skb_shinfo(skb)->gso_segs = (skb->len + mss - 1) / mss;
2200
2201                 segs = NULL;
2202                 goto out;
2203         }
2204
2205         segs = skb_segment(skb, features);
2206         if (IS_ERR(segs))
2207                 goto out;
2208
2209         len = skb_shinfo(skb)->gso_size;
2210         delta = htonl(oldlen + (thlen + len));
2211
2212         skb = segs;
2213         th = skb->h.th;
2214         seq = ntohl(th->seq);
2215
2216         do {
2217                 th->fin = th->psh = 0;
2218
2219                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2220                                        (__force u32)delta));
2221                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2222                         th->check = csum_fold(csum_partial(skb->h.raw, thlen,
2223                                                            skb->csum));
2224
2225                 seq += len;
2226                 skb = skb->next;
2227                 th = skb->h.th;
2228
2229                 th->seq = htonl(seq);
2230                 th->cwr = 0;
2231         } while (skb->next);
2232
2233         delta = htonl(oldlen + (skb->tail - skb->h.raw) + skb->data_len);
2234         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2235                                 (__force u32)delta));
2236         if (skb->ip_summed != CHECKSUM_PARTIAL)
2237                 th->check = csum_fold(csum_partial(skb->h.raw, thlen,
2238                                                    skb->csum));
2239
2240 out:
2241         return segs;
2242 }
2243 EXPORT_SYMBOL(tcp_tso_segment);
2244
2245 #ifdef CONFIG_TCP_MD5SIG
2246 static unsigned long tcp_md5sig_users;
2247 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2248 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2249
2250 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2251 {
2252         int cpu;
2253         for_each_possible_cpu(cpu) {
2254                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2255                 if (p) {
2256                         if (p->md5_desc.tfm)
2257                                 crypto_free_hash(p->md5_desc.tfm);
2258                         kfree(p);
2259                         p = NULL;
2260                 }
2261         }
2262         free_percpu(pool);
2263 }
2264
2265 void tcp_free_md5sig_pool(void)
2266 {
2267         struct tcp_md5sig_pool **pool = NULL;
2268
2269         spin_lock(&tcp_md5sig_pool_lock);
2270         if (--tcp_md5sig_users == 0) {
2271                 pool = tcp_md5sig_pool;
2272                 tcp_md5sig_pool = NULL;
2273         }
2274         spin_unlock(&tcp_md5sig_pool_lock);
2275         if (pool)
2276                 __tcp_free_md5sig_pool(pool);
2277 }
2278
2279 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2280
2281 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(void)
2282 {
2283         int cpu;
2284         struct tcp_md5sig_pool **pool;
2285
2286         pool = alloc_percpu(struct tcp_md5sig_pool *);
2287         if (!pool)
2288                 return NULL;
2289
2290         for_each_possible_cpu(cpu) {
2291                 struct tcp_md5sig_pool *p;
2292                 struct crypto_hash *hash;
2293
2294                 p = kzalloc(sizeof(*p), GFP_KERNEL);
2295                 if (!p)
2296                         goto out_free;
2297                 *per_cpu_ptr(pool, cpu) = p;
2298
2299                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2300                 if (!hash || IS_ERR(hash))
2301                         goto out_free;
2302
2303                 p->md5_desc.tfm = hash;
2304         }
2305         return pool;
2306 out_free:
2307         __tcp_free_md5sig_pool(pool);
2308         return NULL;
2309 }
2310
2311 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(void)
2312 {
2313         struct tcp_md5sig_pool **pool;
2314         int alloc = 0;
2315
2316 retry:
2317         spin_lock(&tcp_md5sig_pool_lock);
2318         pool = tcp_md5sig_pool;
2319         if (tcp_md5sig_users++ == 0) {
2320                 alloc = 1;
2321                 spin_unlock(&tcp_md5sig_pool_lock);
2322         } else if (!pool) {
2323                 tcp_md5sig_users--;
2324                 spin_unlock(&tcp_md5sig_pool_lock);
2325                 cpu_relax();
2326                 goto retry;
2327         } else
2328                 spin_unlock(&tcp_md5sig_pool_lock);
2329
2330         if (alloc) {
2331                 /* we cannot hold spinlock here because this may sleep. */
2332                 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool();
2333                 spin_lock(&tcp_md5sig_pool_lock);
2334                 if (!p) {
2335                         tcp_md5sig_users--;
2336                         spin_unlock(&tcp_md5sig_pool_lock);
2337                         return NULL;
2338                 }
2339                 pool = tcp_md5sig_pool;
2340                 if (pool) {
2341                         /* oops, it has already been assigned. */
2342                         spin_unlock(&tcp_md5sig_pool_lock);
2343                         __tcp_free_md5sig_pool(p);
2344                 } else {
2345                         tcp_md5sig_pool = pool = p;
2346                         spin_unlock(&tcp_md5sig_pool_lock);
2347                 }
2348         }
2349         return pool;
2350 }
2351
2352 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2353
2354 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2355 {
2356         struct tcp_md5sig_pool **p;
2357         spin_lock(&tcp_md5sig_pool_lock);
2358         p = tcp_md5sig_pool;
2359         if (p)
2360                 tcp_md5sig_users++;
2361         spin_unlock(&tcp_md5sig_pool_lock);
2362         return (p ? *per_cpu_ptr(p, cpu) : NULL);
2363 }
2364
2365 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2366
2367 void __tcp_put_md5sig_pool(void)
2368 {
2369         tcp_free_md5sig_pool();
2370 }
2371
2372 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2373 #endif
2374
2375 extern void __skb_cb_too_small_for_tcp(int, int);
2376 extern struct tcp_congestion_ops tcp_reno;
2377
2378 static __initdata unsigned long thash_entries;
2379 static int __init set_thash_entries(char *str)
2380 {
2381         if (!str)
2382                 return 0;
2383         thash_entries = simple_strtoul(str, &str, 0);
2384         return 1;
2385 }
2386 __setup("thash_entries=", set_thash_entries);
2387
2388 void __init tcp_init(void)
2389 {
2390         struct sk_buff *skb = NULL;
2391         unsigned long limit;
2392         int order, i, max_share;
2393
2394         if (sizeof(struct tcp_skb_cb) > sizeof(skb->cb))
2395                 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb),
2396                                            sizeof(skb->cb));
2397
2398         tcp_hashinfo.bind_bucket_cachep =
2399                 kmem_cache_create("tcp_bind_bucket",
2400                                   sizeof(struct inet_bind_bucket), 0,
2401                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
2402
2403         /* Size and allocate the main established and bind bucket
2404          * hash tables.
2405          *
2406          * The methodology is similar to that of the buffer cache.
2407          */
2408         tcp_hashinfo.ehash =
2409                 alloc_large_system_hash("TCP established",
2410                                         sizeof(struct inet_ehash_bucket),
2411                                         thash_entries,
2412                                         (num_physpages >= 128 * 1024) ?
2413                                         13 : 15,
2414                                         0,
2415                                         &tcp_hashinfo.ehash_size,
2416                                         NULL,
2417                                         0);
2418         tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
2419         for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
2420                 rwlock_init(&tcp_hashinfo.ehash[i].lock);
2421                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].chain);
2422                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].twchain);
2423         }
2424
2425         tcp_hashinfo.bhash =
2426                 alloc_large_system_hash("TCP bind",
2427                                         sizeof(struct inet_bind_hashbucket),
2428                                         tcp_hashinfo.ehash_size,
2429                                         (num_physpages >= 128 * 1024) ?
2430                                         13 : 15,
2431                                         0,
2432                                         &tcp_hashinfo.bhash_size,
2433                                         NULL,
2434                                         64 * 1024);
2435         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2436         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2437                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2438                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2439         }
2440
2441         /* Try to be a bit smarter and adjust defaults depending
2442          * on available memory.
2443          */
2444         for (order = 0; ((1 << order) << PAGE_SHIFT) <
2445                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2446                         order++)
2447                 ;
2448         if (order >= 4) {
2449                 sysctl_local_port_range[0] = 32768;
2450                 sysctl_local_port_range[1] = 61000;
2451                 tcp_death_row.sysctl_max_tw_buckets = 180000;
2452                 sysctl_tcp_max_orphans = 4096 << (order - 4);
2453                 sysctl_max_syn_backlog = 1024;
2454         } else if (order < 3) {
2455                 sysctl_local_port_range[0] = 1024 * (3 - order);
2456                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2457                 sysctl_tcp_max_orphans >>= (3 - order);
2458                 sysctl_max_syn_backlog = 128;
2459         }
2460
2461         /* Allow no more than 3/4 kernel memory (usually less) allocated to TCP */
2462         sysctl_tcp_mem[0] = (1536 / sizeof (struct inet_bind_hashbucket)) << order;
2463         sysctl_tcp_mem[1] = sysctl_tcp_mem[0] * 4 / 3;
2464         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2465
2466         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2467         max_share = min(4UL*1024*1024, limit);
2468
2469         sysctl_tcp_wmem[0] = SK_STREAM_MEM_QUANTUM;
2470         sysctl_tcp_wmem[1] = 16*1024;
2471         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2472
2473         sysctl_tcp_rmem[0] = SK_STREAM_MEM_QUANTUM;
2474         sysctl_tcp_rmem[1] = 87380;
2475         sysctl_tcp_rmem[2] = max(87380, max_share);
2476
2477         printk(KERN_INFO "TCP: Hash tables configured "
2478                "(established %d bind %d)\n",
2479                tcp_hashinfo.ehash_size, tcp_hashinfo.bhash_size);
2480
2481         tcp_register_congestion_control(&tcp_reno);
2482 }
2483
2484 EXPORT_SYMBOL(tcp_close);
2485 EXPORT_SYMBOL(tcp_disconnect);
2486 EXPORT_SYMBOL(tcp_getsockopt);
2487 EXPORT_SYMBOL(tcp_ioctl);
2488 EXPORT_SYMBOL(tcp_poll);
2489 EXPORT_SYMBOL(tcp_read_sock);
2490 EXPORT_SYMBOL(tcp_recvmsg);
2491 EXPORT_SYMBOL(tcp_sendmsg);
2492 EXPORT_SYMBOL(tcp_sendpage);
2493 EXPORT_SYMBOL(tcp_setsockopt);
2494 EXPORT_SYMBOL(tcp_shutdown);
2495 EXPORT_SYMBOL(tcp_statistics);