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 #include <linux/in.h>
 263
 264 #include <net/icmp.h>
 265 #include <net/tcp.h>
 266 #include <net/xfrm.h>
 267 #include <net/ip.h>
 268 #include <net/netdma.h>
 269
 270 #include <asm/uaccess.h>
 271 #include <asm/ioctls.h>
 272
 273 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
 274
 275 DEFINE_SNMP_STAT(struct tcp_mib, tcp_statistics) __read_mostly;
 276
 277 atomic_t tcp_orphan_count = ATOMIC_INIT(0);
 278
 279 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 280
 281 int sysctl_tcp_mem[3] __read_mostly;
 282 int sysctl_tcp_wmem[3] __read_mostly;
 283 int sysctl_tcp_rmem[3] __read_mostly;
 284
 285 EXPORT_SYMBOL(sysctl_tcp_mem);
 286 EXPORT_SYMBOL(sysctl_tcp_rmem);
 287 EXPORT_SYMBOL(sysctl_tcp_wmem);
 288
 289 atomic_t tcp_memory_allocated;  /* Current allocated memory. */
 290 atomic_t tcp_sockets_allocated; /* Current number of TCP sockets. */
 291
 292 EXPORT_SYMBOL(tcp_memory_allocated);
 293 EXPORT_SYMBOL(tcp_sockets_allocated);
 294
 295 /*
 296  * Pressure flag: try to collapse.
 297  * Technical note: it is used by multiple contexts non atomically.
 298  * All the sk_stream_mem_schedule() is of this nature: accounting
 299  * is strict, actions are advisory and have some latency.
 300  */
 301 int tcp_memory_pressure;
 302
 303 EXPORT_SYMBOL(tcp_memory_pressure);
 304
 305 void tcp_enter_memory_pressure(void)
 306 {
 307         if (!tcp_memory_pressure) {
 308                 NET_INC_STATS(LINUX_MIB_TCPMEMORYPRESSURES);
 309                 tcp_memory_pressure = 1;
 310         }
 311 }
 312
 313 EXPORT_SYMBOL(tcp_enter_memory_pressure);
 314
 315 /*
 316  *      Wait for a TCP event.
 317  *
 318  *      Note that we don't need to lock the socket, as the upper poll layers
 319  *      take care of normal races (between the test and the event) and we don't
 320  *      go look at any of the socket buffers directly.
 321  */
 322 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 323 {
 324         unsigned int mask;
 325         struct sock *sk = sock->sk;
 326         struct tcp_sock *tp = tcp_sk(sk);
 327
 328         poll_wait(file, sk->sk_sleep, wait);
 329         if (sk->sk_state == TCP_LISTEN)
 330                 return inet_csk_listen_poll(sk);
 331
 332         /* Socket is not locked. We are protected from async events
 333            by poll logic and correct handling of state changes
 334            made by another threads is impossible in any case.
 335          */
 336
 337         mask = 0;
 338         if (sk->sk_err)
 339                 mask = POLLERR;
 340
 341         /*
 342          * POLLHUP is certainly not done right. But poll() doesn't
 343          * have a notion of HUP in just one direction, and for a
 344          * socket the read side is more interesting.
 345          *
 346          * Some poll() documentation says that POLLHUP is incompatible
 347          * with the POLLOUT/POLLWR flags, so somebody should check this
 348          * all. But careful, it tends to be safer to return too many
 349          * bits than too few, and you can easily break real applications
 350          * if you don't tell them that something has hung up!
 351          *
 352          * Check-me.
 353          *
 354          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
 355          * our fs/select.c). It means that after we received EOF,
 356          * poll always returns immediately, making impossible poll() on write()
 357          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
 358          * if and only if shutdown has been made in both directions.
 359          * Actually, it is interesting to look how Solaris and DUX
 360          * solve this dilemma. I would prefer, if PULLHUP were maskable,
 361          * then we could set it on SND_SHUTDOWN. BTW examples given
 362          * in Stevens' books assume exactly this behaviour, it explains
 363          * why PULLHUP is incompatible with POLLOUT.    --ANK
 364          *
 365          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 366          * blocking on fresh not-connected or disconnected socket. --ANK
 367          */
 368         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
 369                 mask |= POLLHUP;
 370         if (sk->sk_shutdown & RCV_SHUTDOWN)
 371                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
 372
 373         /* Connected? */
 374         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
 375                 /* Potential race condition. If read of tp below will
 376                  * escape above sk->sk_state, we can be illegally awaken
 377                  * in SYN_* states. */
 378                 if ((tp->rcv_nxt != tp->copied_seq) &&
 379                     (tp->urg_seq != tp->copied_seq ||
 380                      tp->rcv_nxt != tp->copied_seq + 1 ||
 381                      sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data))
 382                         mask |= POLLIN | POLLRDNORM;
 383
 384                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 385                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
 386                                 mask |= POLLOUT | POLLWRNORM;
 387                         } else {  /* send SIGIO later */
 388                                 set_bit(SOCK_ASYNC_NOSPACE,
 389                                         &sk->sk_socket->flags);
 390                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 391
 392                                 /* Race breaker. If space is freed after
 393                                  * wspace test but before the flags are set,
 394                                  * IO signal will be lost.
 395                                  */
 396                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
 397                                         mask |= POLLOUT | POLLWRNORM;
 398                         }
 399                 }
 400
 401                 if (tp->urg_data & TCP_URG_VALID)
 402                         mask |= POLLPRI;
 403         }
 404         return mask;
 405 }
 406
 407 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 408 {
 409         struct tcp_sock *tp = tcp_sk(sk);
 410         int answ;
 411
 412         switch (cmd) {
 413         case SIOCINQ:
 414                 if (sk->sk_state == TCP_LISTEN)
 415                         return -EINVAL;
 416
 417                 lock_sock(sk);
 418                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 419                         answ = 0;
 420                 else if (sock_flag(sk, SOCK_URGINLINE) ||
 421                          !tp->urg_data ||
 422                          before(tp->urg_seq, tp->copied_seq) ||
 423                          !before(tp->urg_seq, tp->rcv_nxt)) {
 424                         answ = tp->rcv_nxt - tp->copied_seq;
 425
 426                         /* Subtract 1, if FIN is in queue. */
 427                         if (answ && !skb_queue_empty(&sk->sk_receive_queue))
 428                                 answ -=
 429                        ((struct sk_buff *)sk->sk_receive_queue.prev)->h.th->fin;
 430                 } else
 431                         answ = tp->urg_seq - tp->copied_seq;
 432                 release_sock(sk);
 433                 break;
 434         case SIOCATMARK:
 435                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 436                 break;
 437         case SIOCOUTQ:
 438                 if (sk->sk_state == TCP_LISTEN)
 439                         return -EINVAL;
 440
 441                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 442                         answ = 0;
 443                 else
 444                         answ = tp->write_seq - tp->snd_una;
 445                 break;
 446         default:
 447                 return -ENOIOCTLCMD;
 448         };
 449
 450         return put_user(answ, (int __user *)arg);
 451 }
 452
 453 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 454 {
 455         TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
 456         tp->pushed_seq = tp->write_seq;
 457 }
 458
 459 static inline int forced_push(struct tcp_sock *tp)
 460 {
 461         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 462 }
 463
 464 static inline void skb_entail(struct sock *sk, struct tcp_sock *tp,
 465                               struct sk_buff *skb)
 466 {
 467         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 468
 469         skb->csum    = 0;
 470         tcb->seq     = tcb->end_seq = tp->write_seq;
 471         tcb->flags   = TCPCB_FLAG_ACK;
 472         tcb->sacked  = 0;
 473         skb_header_release(skb);
 474         __skb_queue_tail(&sk->sk_write_queue, skb);
 475         sk_charge_skb(sk, skb);
 476         if (!sk->sk_send_head)
 477                 sk->sk_send_head = skb;
 478         if (tp->nonagle & TCP_NAGLE_PUSH)
 479                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 480 }
 481
 482 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
 483                                 struct sk_buff *skb)
 484 {
 485         if (flags & MSG_OOB) {
 486                 tp->urg_mode = 1;
 487                 tp->snd_up = tp->write_seq;
 488                 TCP_SKB_CB(skb)->sacked |= TCPCB_URG;
 489         }
 490 }
 491
 492 static inline void tcp_push(struct sock *sk, struct tcp_sock *tp, int flags,
 493                             int mss_now, int nonagle)
 494 {
 495         if (sk->sk_send_head) {
 496                 struct sk_buff *skb = sk->sk_write_queue.prev;
 497                 if (!(flags & MSG_MORE) || forced_push(tp))
 498                         tcp_mark_push(tp, skb);
 499                 tcp_mark_urg(tp, flags, skb);
 500                 __tcp_push_pending_frames(sk, tp, mss_now,
 501                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
 502         }
 503 }
 504
 505 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
 506                          size_t psize, int flags)
 507 {
 508         struct tcp_sock *tp = tcp_sk(sk);
 509         int mss_now, size_goal;
 510         int err;
 511         ssize_t copied;
 512         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 513
 514         /* Wait for a connection to finish. */
 515         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 516                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 517                         goto out_err;
 518
 519         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 520
 521         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 522         size_goal = tp->xmit_size_goal;
 523         copied = 0;
 524
 525         err = -EPIPE;
 526         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 527                 goto do_error;
 528
 529         while (psize > 0) {
 530                 struct sk_buff *skb = sk->sk_write_queue.prev;
 531                 struct page *page = pages[poffset / PAGE_SIZE];
 532                 int copy, i, can_coalesce;
 533                 int offset = poffset % PAGE_SIZE;
 534                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
 535
 536                 if (!sk->sk_send_head || (copy = size_goal - skb->len) <= 0) {
 537 new_segment:
 538                         if (!sk_stream_memory_free(sk))
 539                                 goto wait_for_sndbuf;
 540
 541                         skb = sk_stream_alloc_pskb(sk, 0, 0,
 542                                                    sk->sk_allocation);
 543                         if (!skb)
 544                                 goto wait_for_memory;
 545
 546                         skb_entail(sk, tp, skb);
 547                         copy = size_goal;
 548                 }
 549
 550                 if (copy > size)
 551                         copy = size;
 552
 553                 i = skb_shinfo(skb)->nr_frags;
 554                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 555                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
 556                         tcp_mark_push(tp, skb);
 557                         goto new_segment;
 558                 }
 559                 if (!sk_stream_wmem_schedule(sk, copy))
 560                         goto wait_for_memory;
 561
 562                 if (can_coalesce) {
 563                         skb_shinfo(skb)->frags[i - 1].size += copy;
 564                 } else {
 565                         get_page(page);
 566                         skb_fill_page_desc(skb, i, page, offset, copy);
 567                 }
 568
 569                 skb->len += copy;
 570                 skb->data_len += copy;
 571                 skb->truesize += copy;
 572                 sk->sk_wmem_queued += copy;
 573                 sk->sk_forward_alloc -= copy;
 574                 skb->ip_summed = CHECKSUM_PARTIAL;
 575                 tp->write_seq += copy;
 576                 TCP_SKB_CB(skb)->end_seq += copy;
 577                 skb_shinfo(skb)->gso_segs = 0;
 578
 579                 if (!copied)
 580                         TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
 581
 582                 copied += copy;
 583                 poffset += copy;
 584                 if (!(psize -= copy))
 585                         goto out;
 586
 587                 if (skb->len < mss_now || (flags & MSG_OOB))
 588                         continue;
 589
 590                 if (forced_push(tp)) {
 591                         tcp_mark_push(tp, skb);
 592                         __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
 593                 } else if (skb == sk->sk_send_head)
 594                         tcp_push_one(sk, mss_now);
 595                 continue;
 596
 597 wait_for_sndbuf:
 598                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 599 wait_for_memory:
 600                 if (copied)
 601                         tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
 602
 603                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 604                         goto do_error;
 605
 606                 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 607                 size_goal = tp->xmit_size_goal;
 608         }
 609
 610 out:
 611         if (copied)
 612                 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
 613         return copied;
 614
 615 do_error:
 616         if (copied)
 617                 goto out;
 618 out_err:
 619         return sk_stream_error(sk, flags, err);
 620 }
 621
 622 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
 623                      size_t size, int flags)
 624 {
 625         ssize_t res;
 626         struct sock *sk = sock->sk;
 627
 628         if (!(sk->sk_route_caps & NETIF_F_SG) ||
 629             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
 630                 return sock_no_sendpage(sock, page, offset, size, flags);
 631
 632         lock_sock(sk);
 633         TCP_CHECK_TIMER(sk);
 634         res = do_tcp_sendpages(sk, &page, offset, size, flags);
 635         TCP_CHECK_TIMER(sk);
 636         release_sock(sk);
 637         return res;
 638 }
 639
 640 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
 641 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
 642
 643 static inline int select_size(struct sock *sk, struct tcp_sock *tp)
 644 {
 645         int tmp = tp->mss_cache;
 646
 647         if (sk->sk_route_caps & NETIF_F_SG) {
 648                 if (sk_can_gso(sk))
 649                         tmp = 0;
 650                 else {
 651                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
 652
 653                         if (tmp >= pgbreak &&
 654                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
 655                                 tmp = pgbreak;
 656                 }
 657         }
 658
 659         return tmp;
 660 }
 661
 662 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
 663                 size_t size)
 664 {
 665         struct iovec *iov;
 666         struct tcp_sock *tp = tcp_sk(sk);
 667         struct sk_buff *skb;
 668         int iovlen, flags;
 669         int mss_now, size_goal;
 670         int err, copied;
 671         long timeo;
 672
 673         lock_sock(sk);
 674         TCP_CHECK_TIMER(sk);
 675
 676         flags = msg->msg_flags;
 677         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 678
 679         /* Wait for a connection to finish. */
 680         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 681                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 682                         goto out_err;
 683
 684         /* This should be in poll */
 685         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 686
 687         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 688         size_goal = tp->xmit_size_goal;
 689
 690         /* Ok commence sending. */
 691         iovlen = msg->msg_iovlen;
 692         iov = msg->msg_iov;
 693         copied = 0;
 694
 695         err = -EPIPE;
 696         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 697                 goto do_error;
 698
 699         while (--iovlen >= 0) {
 700                 int seglen = iov->iov_len;
 701                 unsigned char __user *from = iov->iov_base;
 702
 703                 iov++;
 704
 705                 while (seglen > 0) {
 706                         int copy;
 707
 708                         skb = sk->sk_write_queue.prev;
 709
 710                         if (!sk->sk_send_head ||
 711                             (copy = size_goal - skb->len) <= 0) {
 712
 713 new_segment:
 714                                 /* Allocate new segment. If the interface is SG,
 715                                  * allocate skb fitting to single page.
 716                                  */
 717                                 if (!sk_stream_memory_free(sk))
 718                                         goto wait_for_sndbuf;
 719
 720                                 skb = sk_stream_alloc_pskb(sk, select_size(sk, tp),
 721                                                            0, sk->sk_allocation);
 722                                 if (!skb)
 723                                         goto wait_for_memory;
 724
 725                                 /*
 726                                  * Check whether we can use HW checksum.
 727                                  */
 728                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
 729                                         skb->ip_summed = CHECKSUM_PARTIAL;
 730
 731                                 skb_entail(sk, tp, skb);
 732                                 copy = size_goal;
 733                         }
 734
 735                         /* Try to append data to the end of skb. */
 736                         if (copy > seglen)
 737                                 copy = seglen;
 738
 739                         /* Where to copy to? */
 740                         if (skb_tailroom(skb) > 0) {
 741                                 /* We have some space in skb head. Superb! */
 742                                 if (copy > skb_tailroom(skb))
 743                                         copy = skb_tailroom(skb);
 744                                 if ((err = skb_add_data(skb, from, copy)) != 0)
 745                                         goto do_fault;
 746                         } else {
 747                                 int merge = 0;
 748                                 int i = skb_shinfo(skb)->nr_frags;
 749                                 struct page *page = TCP_PAGE(sk);
 750                                 int off = TCP_OFF(sk);
 751
 752                                 if (skb_can_coalesce(skb, i, page, off) &&
 753                                     off != PAGE_SIZE) {
 754                                         /* We can extend the last page
 755                                          * fragment. */
 756                                         merge = 1;
 757                                 } else if (i == MAX_SKB_FRAGS ||
 758                                            (!i &&
 759                                            !(sk->sk_route_caps & NETIF_F_SG))) {
 760                                         /* Need to add new fragment and cannot
 761                                          * do this because interface is non-SG,
 762                                          * or because all the page slots are
 763                                          * busy. */
 764                                         tcp_mark_push(tp, skb);
 765                                         goto new_segment;
 766                                 } else if (page) {
 767                                         if (off == PAGE_SIZE) {
 768                                                 put_page(page);
 769                                                 TCP_PAGE(sk) = page = NULL;
 770                                                 off = 0;
 771                                         }
 772                                 } else
 773                                         off = 0;
 774
 775                                 if (copy > PAGE_SIZE - off)
 776                                         copy = PAGE_SIZE - off;
 777
 778                                 if (!sk_stream_wmem_schedule(sk, copy))
 779                                         goto wait_for_memory;
 780
 781                                 if (!page) {
 782                                         /* Allocate new cache page. */
 783                                         if (!(page = sk_stream_alloc_page(sk)))
 784                                                 goto wait_for_memory;
 785                                 }
 786
 787                                 /* Time to copy data. We are close to
 788                                  * the end! */
 789                                 err = skb_copy_to_page(sk, from, skb, page,
 790                                                        off, copy);
 791                                 if (err) {
 792                                         /* If this page was new, give it to the
 793                                          * socket so it does not get leaked.
 794                                          */
 795                                         if (!TCP_PAGE(sk)) {
 796                                                 TCP_PAGE(sk) = page;
 797                                                 TCP_OFF(sk) = 0;
 798                                         }
 799                                         goto do_error;
 800                                 }
 801
 802                                 /* Update the skb. */
 803                                 if (merge) {
 804                                         skb_shinfo(skb)->frags[i - 1].size +=
 805                                                                         copy;
 806                                 } else {
 807                                         skb_fill_page_desc(skb, i, page, off, copy);
 808                                         if (TCP_PAGE(sk)) {
 809                                                 get_page(page);
 810                                         } else if (off + copy < PAGE_SIZE) {
 811                                                 get_page(page);
 812                                                 TCP_PAGE(sk) = page;
 813                                         }
 814                                 }
 815
 816                                 TCP_OFF(sk) = off + copy;
 817                         }
 818
 819                         if (!copied)
 820                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
 821
 822                         tp->write_seq += copy;
 823                         TCP_SKB_CB(skb)->end_seq += copy;
 824                         skb_shinfo(skb)->gso_segs = 0;
 825
 826                         from += copy;
 827                         copied += copy;
 828                         if ((seglen -= copy) == 0 && iovlen == 0)
 829                                 goto out;
 830
 831                         if (skb->len < mss_now || (flags & MSG_OOB))
 832                                 continue;
 833
 834                         if (forced_push(tp)) {
 835                                 tcp_mark_push(tp, skb);
 836                                 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
 837                         } else if (skb == sk->sk_send_head)
 838                                 tcp_push_one(sk, mss_now);
 839                         continue;
 840
 841 wait_for_sndbuf:
 842                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 843 wait_for_memory:
 844                         if (copied)
 845                                 tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
 846
 847                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 848                                 goto do_error;
 849
 850                         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 851                         size_goal = tp->xmit_size_goal;
 852                 }
 853         }
 854
 855 out:
 856         if (copied)
 857                 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
 858         TCP_CHECK_TIMER(sk);
 859         release_sock(sk);
 860         return copied;
 861
 862 do_fault:
 863         if (!skb->len) {
 864                 if (sk->sk_send_head == skb)
 865                         sk->sk_send_head = NULL;
 866                 __skb_unlink(skb, &sk->sk_write_queue);
 867                 sk_stream_free_skb(sk, skb);
 868         }
 869
 870 do_error:
 871         if (copied)
 872                 goto out;
 873 out_err:
 874         err = sk_stream_error(sk, flags, err);
 875         TCP_CHECK_TIMER(sk);
 876         release_sock(sk);
 877         return err;
 878 }
 879
 880 /*
 881  *      Handle reading urgent data. BSD has very simple semantics for
 882  *      this, no blocking and very strange errors 8)
 883  */
 884
 885 static int tcp_recv_urg(struct sock *sk, long timeo,
 886                         struct msghdr *msg, int len, int flags,
 887                         int *addr_len)
 888 {
 889         struct tcp_sock *tp = tcp_sk(sk);
 890
 891         /* No URG data to read. */
 892         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
 893             tp->urg_data == TCP_URG_READ)
 894                 return -EINVAL; /* Yes this is right ! */
 895
 896         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
 897                 return -ENOTCONN;
 898
 899         if (tp->urg_data & TCP_URG_VALID) {
 900                 int err = 0;
 901                 char c = tp->urg_data;
 902
 903                 if (!(flags & MSG_PEEK))
 904                         tp->urg_data = TCP_URG_READ;
 905
 906                 /* Read urgent data. */
 907                 msg->msg_flags |= MSG_OOB;
 908
 909                 if (len > 0) {
 910                         if (!(flags & MSG_TRUNC))
 911                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
 912                         len = 1;
 913                 } else
 914                         msg->msg_flags |= MSG_TRUNC;
 915
 916                 return err ? -EFAULT : len;
 917         }
 918
 919         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
 920                 return 0;
 921
 922         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
 923          * the available implementations agree in this case:
 924          * this call should never block, independent of the
 925          * blocking state of the socket.
 926          * Mike <pall@rz.uni-karlsruhe.de>
 927          */
 928         return -EAGAIN;
 929 }
 930
 931 /* Clean up the receive buffer for full frames taken by the user,
 932  * then send an ACK if necessary.  COPIED is the number of bytes
 933  * tcp_recvmsg has given to the user so far, it speeds up the
 934  * calculation of whether or not we must ACK for the sake of
 935  * a window update.
 936  */
 937 void tcp_cleanup_rbuf(struct sock *sk, int copied)
 938 {
 939         struct tcp_sock *tp = tcp_sk(sk);
 940         int time_to_ack = 0;
 941
 942 #if TCP_DEBUG
 943         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
 944
 945         BUG_TRAP(!skb || before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
 946 #endif
 947
 948         if (inet_csk_ack_scheduled(sk)) {
 949                 const struct inet_connection_sock *icsk = inet_csk(sk);
 950                    /* Delayed ACKs frequently hit locked sockets during bulk
 951                     * receive. */
 952                 if (icsk->icsk_ack.blocked ||
 953                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
 954                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
 955                     /*
 956                      * If this read emptied read buffer, we send ACK, if
 957                      * connection is not bidirectional, user drained
 958                      * receive buffer and there was a small segment
 959                      * in queue.
 960                      */
 961                     (copied > 0 &&
 962                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
 963                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
 964                        !icsk->icsk_ack.pingpong)) &&
 965                       !atomic_read(&sk->sk_rmem_alloc)))
 966                         time_to_ack = 1;
 967         }
 968
 969         /* We send an ACK if we can now advertise a non-zero window
 970          * which has been raised "significantly".
 971          *
 972          * Even if window raised up to infinity, do not send window open ACK
 973          * in states, where we will not receive more. It is useless.
 974          */
 975         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
 976                 __u32 rcv_window_now = tcp_receive_window(tp);
 977
 978                 /* Optimize, __tcp_select_window() is not cheap. */
 979                 if (2*rcv_window_now <= tp->window_clamp) {
 980                         __u32 new_window = __tcp_select_window(sk);
 981
 982                         /* Send ACK now, if this read freed lots of space
 983                          * in our buffer. Certainly, new_window is new window.
 984                          * We can advertise it now, if it is not less than current one.
 985                          * "Lots" means "at least twice" here.
 986                          */
 987                         if (new_window && new_window >= 2 * rcv_window_now)
 988                                 time_to_ack = 1;
 989                 }
 990         }
 991         if (time_to_ack)
 992                 tcp_send_ack(sk);
 993 }
 994
 995 static void tcp_prequeue_process(struct sock *sk)
 996 {
 997         struct sk_buff *skb;
 998         struct tcp_sock *tp = tcp_sk(sk);
 999
1000         NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED);
1001
1002         /* RX process wants to run with disabled BHs, though it is not
1003          * necessary */
1004         local_bh_disable();
1005         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1006                 sk->sk_backlog_rcv(sk, skb);
1007         local_bh_enable();
1008
1009         /* Clear memory counter. */
1010         tp->ucopy.memory = 0;
1011 }
1012
1013 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1014 {
1015         struct sk_buff *skb;
1016         u32 offset;
1017
1018         skb_queue_walk(&sk->sk_receive_queue, skb) {
1019                 offset = seq - TCP_SKB_CB(skb)->seq;
1020                 if (skb->h.th->syn)
1021                         offset--;
1022                 if (offset < skb->len || skb->h.th->fin) {
1023                         *off = offset;
1024                         return skb;
1025                 }
1026         }
1027         return NULL;
1028 }
1029
1030 /*
1031  * This routine provides an alternative to tcp_recvmsg() for routines
1032  * that would like to handle copying from skbuffs directly in 'sendfile'
1033  * fashion.
1034  * Note:
1035  *      - It is assumed that the socket was locked by the caller.
1036  *      - The routine does not block.
1037  *      - At present, there is no support for reading OOB data
1038  *        or for 'peeking' the socket using this routine
1039  *        (although both would be easy to implement).
1040  */
1041 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1042                   sk_read_actor_t recv_actor)
1043 {
1044         struct sk_buff *skb;
1045         struct tcp_sock *tp = tcp_sk(sk);
1046         u32 seq = tp->copied_seq;
1047         u32 offset;
1048         int copied = 0;
1049
1050         if (sk->sk_state == TCP_LISTEN)
1051                 return -ENOTCONN;
1052         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1053                 if (offset < skb->len) {
1054                         size_t used, len;
1055
1056                         len = skb->len - offset;
1057                         /* Stop reading if we hit a patch of urgent data */
1058                         if (tp->urg_data) {
1059                                 u32 urg_offset = tp->urg_seq - seq;
1060                                 if (urg_offset < len)
1061                                         len = urg_offset;
1062                                 if (!len)
1063                                         break;
1064                         }
1065                         used = recv_actor(desc, skb, offset, len);
1066                         if (used <= len) {
1067                                 seq += used;
1068                                 copied += used;
1069                                 offset += used;
1070                         }
1071                         if (offset != skb->len)
1072                                 break;
1073                 }
1074                 if (skb->h.th->fin) {
1075                         sk_eat_skb(sk, skb, 0);
1076                         ++seq;
1077                         break;
1078                 }
1079                 sk_eat_skb(sk, skb, 0);
1080                 if (!desc->count)
1081                         break;
1082         }
1083         tp->copied_seq = seq;
1084
1085         tcp_rcv_space_adjust(sk);
1086
1087         /* Clean up data we have read: This will do ACK frames. */
1088         if (copied)
1089                 tcp_cleanup_rbuf(sk, copied);
1090         return copied;
1091 }
1092
1093 /*
1094  *      This routine copies from a sock struct into the user buffer.
1095  *
1096  *      Technical note: in 2.3 we work on _locked_ socket, so that
1097  *      tricks with *seq access order and skb->users are not required.
1098  *      Probably, code can be easily improved even more.
1099  */
1100
1101 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1102                 size_t len, int nonblock, int flags, int *addr_len)
1103 {
1104         struct tcp_sock *tp = tcp_sk(sk);
1105         int copied = 0;
1106         u32 peek_seq;
1107         u32 *seq;
1108         unsigned long used;
1109         int err;
1110         int target;             /* Read at least this many bytes */
1111         long timeo;
1112         struct task_struct *user_recv = NULL;
1113         int copied_early = 0;
1114
1115         lock_sock(sk);
1116
1117         TCP_CHECK_TIMER(sk);
1118
1119         err = -ENOTCONN;
1120         if (sk->sk_state == TCP_LISTEN)
1121                 goto out;
1122
1123         timeo = sock_rcvtimeo(sk, nonblock);
1124
1125         /* Urgent data needs to be handled specially. */
1126         if (flags & MSG_OOB)
1127                 goto recv_urg;
1128
1129         seq = &tp->copied_seq;
1130         if (flags & MSG_PEEK) {
1131                 peek_seq = tp->copied_seq;
1132                 seq = &peek_seq;
1133         }
1134
1135         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1136
1137 #ifdef CONFIG_NET_DMA
1138         tp->ucopy.dma_chan = NULL;
1139         preempt_disable();
1140         if ((len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1141             !sysctl_tcp_low_latency && __get_cpu_var(softnet_data).net_dma) {
1142                 preempt_enable_no_resched();
1143                 tp->ucopy.pinned_list = dma_pin_iovec_pages(msg->msg_iov, len);
1144         } else
1145                 preempt_enable_no_resched();
1146 #endif
1147
1148         do {
1149                 struct sk_buff *skb;
1150                 u32 offset;
1151
1152                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1153                 if (tp->urg_data && tp->urg_seq == *seq) {
1154                         if (copied)
1155                                 break;
1156                         if (signal_pending(current)) {
1157                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1158                                 break;
1159                         }
1160                 }
1161
1162                 /* Next get a buffer. */
1163
1164                 skb = skb_peek(&sk->sk_receive_queue);
1165                 do {
1166                         if (!skb)
1167                                 break;
1168
1169                         /* Now that we have two receive queues this
1170                          * shouldn't happen.
1171                          */
1172                         if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1173                                 printk(KERN_INFO "recvmsg bug: copied %X "
1174                                        "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1175                                 break;
1176                         }
1177                         offset = *seq - TCP_SKB_CB(skb)->seq;
1178                         if (skb->h.th->syn)
1179                                 offset--;
1180                         if (offset < skb->len)
1181                                 goto found_ok_skb;
1182                         if (skb->h.th->fin)
1183                                 goto found_fin_ok;
1184                         BUG_TRAP(flags & MSG_PEEK);
1185                         skb = skb->next;
1186                 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1187
1188                 /* Well, if we have backlog, try to process it now yet. */
1189
1190                 if (copied >= target && !sk->sk_backlog.tail)
1191                         break;
1192
1193                 if (copied) {
1194                         if (sk->sk_err ||
1195                             sk->sk_state == TCP_CLOSE ||
1196                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1197                             !timeo ||
1198                             signal_pending(current) ||
1199                             (flags & MSG_PEEK))
1200                                 break;
1201                 } else {
1202                         if (sock_flag(sk, SOCK_DONE))
1203                                 break;
1204
1205                         if (sk->sk_err) {
1206                                 copied = sock_error(sk);
1207                                 break;
1208                         }
1209
1210                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1211                                 break;
1212
1213                         if (sk->sk_state == TCP_CLOSE) {
1214                                 if (!sock_flag(sk, SOCK_DONE)) {
1215                                         /* This occurs when user tries to read
1216                                          * from never connected socket.
1217                                          */
1218                                         copied = -ENOTCONN;
1219                                         break;
1220                                 }
1221                                 break;
1222                         }
1223
1224                         if (!timeo) {
1225                                 copied = -EAGAIN;
1226                                 break;
1227                         }
1228
1229                         if (signal_pending(current)) {
1230                                 copied = sock_intr_errno(timeo);
1231                                 break;
1232                         }
1233                 }
1234
1235                 tcp_cleanup_rbuf(sk, copied);
1236
1237                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1238                         /* Install new reader */
1239                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1240                                 user_recv = current;
1241                                 tp->ucopy.task = user_recv;
1242                                 tp->ucopy.iov = msg->msg_iov;
1243                         }
1244
1245                         tp->ucopy.len = len;
1246
1247                         BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1248                                  (flags & (MSG_PEEK | MSG_TRUNC)));
1249
1250                         /* Ugly... If prequeue is not empty, we have to
1251                          * process it before releasing socket, otherwise
1252                          * order will be broken at second iteration.
1253                          * More elegant solution is required!!!
1254                          *
1255                          * Look: we have the following (pseudo)queues:
1256                          *
1257                          * 1. packets in flight
1258                          * 2. backlog
1259                          * 3. prequeue
1260                          * 4. receive_queue
1261                          *
1262                          * Each queue can be processed only if the next ones
1263                          * are empty. At this point we have empty receive_queue.
1264                          * But prequeue _can_ be not empty after 2nd iteration,
1265                          * when we jumped to start of loop because backlog
1266                          * processing added something to receive_queue.
1267                          * We cannot release_sock(), because backlog contains
1268                          * packets arrived _after_ prequeued ones.
1269                          *
1270                          * Shortly, algorithm is clear --- to process all
1271                          * the queues in order. We could make it more directly,
1272                          * requeueing packets from backlog to prequeue, if
1273                          * is not empty. It is more elegant, but eats cycles,
1274                          * unfortunately.
1275                          */
1276                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1277                                 goto do_prequeue;
1278
1279                         /* __ Set realtime policy in scheduler __ */
1280                 }
1281
1282                 if (copied >= target) {
1283                         /* Do not sleep, just process backlog. */
1284                         release_sock(sk);
1285                         lock_sock(sk);
1286                 } else
1287                         sk_wait_data(sk, &timeo);
1288
1289 #ifdef CONFIG_NET_DMA
1290                 tp->ucopy.wakeup = 0;
1291 #endif
1292
1293                 if (user_recv) {
1294                         int chunk;
1295
1296                         /* __ Restore normal policy in scheduler __ */
1297
1298                         if ((chunk = len - tp->ucopy.len) != 0) {
1299                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1300                                 len -= chunk;
1301                                 copied += chunk;
1302                         }
1303
1304                         if (tp->rcv_nxt == tp->copied_seq &&
1305                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1306 do_prequeue:
1307                                 tcp_prequeue_process(sk);
1308
1309                                 if ((chunk = len - tp->ucopy.len) != 0) {
1310                                         NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1311                                         len -= chunk;
1312                                         copied += chunk;
1313                                 }
1314                         }
1315                 }
1316                 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1317                         if (net_ratelimit())
1318                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1319                                        current->comm, current->pid);
1320                         peek_seq = tp->copied_seq;
1321                 }
1322                 continue;
1323
1324         found_ok_skb:
1325                 /* Ok so how much can we use? */
1326                 used = skb->len - offset;
1327                 if (len < used)
1328                         used = len;
1329
1330                 /* Do we have urgent data here? */
1331                 if (tp->urg_data) {
1332                         u32 urg_offset = tp->urg_seq - *seq;
1333                         if (urg_offset < used) {
1334                                 if (!urg_offset) {
1335                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1336                                                 ++*seq;
1337                                                 offset++;
1338                                                 used--;
1339                                                 if (!used)
1340                                                         goto skip_copy;
1341                                         }
1342                                 } else
1343                                         used = urg_offset;
1344                         }
1345                 }
1346
1347                 if (!(flags & MSG_TRUNC)) {
1348 #ifdef CONFIG_NET_DMA
1349                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1350                                 tp->ucopy.dma_chan = get_softnet_dma();
1351
1352                         if (tp->ucopy.dma_chan) {
1353                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1354                                         tp->ucopy.dma_chan, skb, offset,
1355                                         msg->msg_iov, used,
1356                                         tp->ucopy.pinned_list);
1357
1358                                 if (tp->ucopy.dma_cookie < 0) {
1359
1360                                         printk(KERN_ALERT "dma_cookie < 0\n");
1361
1362                                         /* Exception. Bailout! */
1363                                         if (!copied)
1364                                                 copied = -EFAULT;
1365                                         break;
1366                                 }
1367                                 if ((offset + used) == skb->len)
1368                                         copied_early = 1;
1369
1370                         } else
1371 #endif
1372                         {
1373                                 err = skb_copy_datagram_iovec(skb, offset,
1374                                                 msg->msg_iov, used);
1375                                 if (err) {
1376                                         /* Exception. Bailout! */
1377                                         if (!copied)
1378                                                 copied = -EFAULT;
1379                                         break;
1380                                 }
1381                         }
1382                 }
1383
1384                 *seq += used;
1385                 copied += used;
1386                 len -= used;
1387
1388                 tcp_rcv_space_adjust(sk);
1389
1390 skip_copy:
1391                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1392                         tp->urg_data = 0;
1393                         tcp_fast_path_check(sk, tp);
1394                 }
1395                 if (used + offset < skb->len)
1396                         continue;
1397
1398                 if (skb->h.th->fin)
1399                         goto found_fin_ok;
1400                 if (!(flags & MSG_PEEK)) {
1401                         sk_eat_skb(sk, skb, copied_early);
1402                         copied_early = 0;
1403                 }
1404                 continue;
1405
1406         found_fin_ok:
1407                 /* Process the FIN. */
1408                 ++*seq;
1409                 if (!(flags & MSG_PEEK)) {
1410                         sk_eat_skb(sk, skb, copied_early);
1411                         copied_early = 0;
1412                 }
1413                 break;
1414         } while (len > 0);
1415
1416         if (user_recv) {
1417                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1418                         int chunk;
1419
1420                         tp->ucopy.len = copied > 0 ? len : 0;
1421
1422                         tcp_prequeue_process(sk);
1423
1424                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1425                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1426                                 len -= chunk;
1427                                 copied += chunk;
1428                         }
1429                 }
1430
1431                 tp->ucopy.task = NULL;
1432                 tp->ucopy.len = 0;
1433         }
1434
1435 #ifdef CONFIG_NET_DMA
1436         if (tp->ucopy.dma_chan) {
1437                 struct sk_buff *skb;
1438                 dma_cookie_t done, used;
1439
1440                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1441
1442                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1443                                                  tp->ucopy.dma_cookie, &done,
1444                                                  &used) == DMA_IN_PROGRESS) {
1445                         /* do partial cleanup of sk_async_wait_queue */
1446                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1447                                (dma_async_is_complete(skb->dma_cookie, done,
1448                                                       used) == DMA_SUCCESS)) {
1449                                 __skb_dequeue(&sk->sk_async_wait_queue);
1450                                 kfree_skb(skb);
1451                         }
1452                 }
1453
1454                 /* Safe to free early-copied skbs now */
1455                 __skb_queue_purge(&sk->sk_async_wait_queue);
1456                 dma_chan_put(tp->ucopy.dma_chan);
1457                 tp->ucopy.dma_chan = NULL;
1458         }
1459         if (tp->ucopy.pinned_list) {
1460                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1461                 tp->ucopy.pinned_list = NULL;
1462         }
1463 #endif
1464
1465         /* According to UNIX98, msg_name/msg_namelen are ignored
1466          * on connected socket. I was just happy when found this 8) --ANK
1467          */
1468
1469         /* Clean up data we have read: This will do ACK frames. */
1470         tcp_cleanup_rbuf(sk, copied);
1471
1472         TCP_CHECK_TIMER(sk);
1473         release_sock(sk);
1474         return copied;
1475
1476 out:
1477         TCP_CHECK_TIMER(sk);
1478         release_sock(sk);
1479         return err;
1480
1481 recv_urg:
1482         err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1483         goto out;
1484 }
1485
1486 /*
1487  *      State processing on a close. This implements the state shift for
1488  *      sending our FIN frame. Note that we only send a FIN for some
1489  *      states. A shutdown() may have already sent the FIN, or we may be
1490  *      closed.
1491  */
1492
1493 static const unsigned char new_state[16] = {
1494   /* current state:        new state:      action:      */
1495   /* (Invalid)          */ TCP_CLOSE,
1496   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1497   /* TCP_SYN_SENT       */ TCP_CLOSE,
1498   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1499   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1500   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1501   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1502   /* TCP_CLOSE          */ TCP_CLOSE,
1503   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1504   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1505   /* TCP_LISTEN         */ TCP_CLOSE,
1506   /* TCP_CLOSING        */ TCP_CLOSING,
1507 };
1508
1509 static int tcp_close_state(struct sock *sk)
1510 {
1511         int next = (int)new_state[sk->sk_state];
1512         int ns = next & TCP_STATE_MASK;
1513
1514         tcp_set_state(sk, ns);
1515
1516         return next & TCP_ACTION_FIN;
1517 }
1518
1519 /*
1520  *      Shutdown the sending side of a connection. Much like close except
1521  *      that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1522  */
1523
1524 void tcp_shutdown(struct sock *sk, int how)
1525 {
1526         /*      We need to grab some memory, and put together a FIN,
1527          *      and then put it into the queue to be sent.
1528          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1529          */
1530         if (!(how & SEND_SHUTDOWN))
1531                 return;
1532
1533         /* If we've already sent a FIN, or it's a closed state, skip this. */
1534         if ((1 << sk->sk_state) &
1535             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1536              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1537                 /* Clear out any half completed packets.  FIN if needed. */
1538                 if (tcp_close_state(sk))
1539                         tcp_send_fin(sk);
1540         }
1541 }
1542
1543 void tcp_close(struct sock *sk, long timeout)
1544 {
1545         struct sk_buff *skb;
1546         int data_was_unread = 0;
1547         int state;
1548
1549         lock_sock(sk);
1550         sk->sk_shutdown = SHUTDOWN_MASK;
1551
1552         if (sk->sk_state == TCP_LISTEN) {
1553                 tcp_set_state(sk, TCP_CLOSE);
1554
1555                 /* Special case. */
1556                 inet_csk_listen_stop(sk);
1557
1558                 goto adjudge_to_death;
1559         }
1560
1561         /*  We need to flush the recv. buffs.  We do this only on the
1562          *  descriptor close, not protocol-sourced closes, because the
1563          *  reader process may not have drained the data yet!
1564          */
1565         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1566                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1567                           skb->h.th->fin;
1568                 data_was_unread += len;
1569                 __kfree_skb(skb);
1570         }
1571
1572         sk_stream_mem_reclaim(sk);
1573
1574         /* As outlined in draft-ietf-tcpimpl-prob-03.txt, section
1575          * 3.10, we send a RST here because data was lost.  To
1576          * witness the awful effects of the old behavior of always
1577          * doing a FIN, run an older 2.1.x kernel or 2.0.x, start
1578          * a bulk GET in an FTP client, suspend the process, wait
1579          * for the client to advertise a zero window, then kill -9
1580          * the FTP client, wheee...  Note: timeout is always zero
1581          * in such a case.
1582          */
1583         if (data_was_unread) {
1584                 /* Unread data was tossed, zap the connection. */
1585                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1586                 tcp_set_state(sk, TCP_CLOSE);
1587                 tcp_send_active_reset(sk, GFP_KERNEL);
1588         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1589                 /* Check zero linger _after_ checking for unread data. */
1590                 sk->sk_prot->disconnect(sk, 0);
1591                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1592         } else if (tcp_close_state(sk)) {
1593                 /* We FIN if the application ate all the data before
1594                  * zapping the connection.
1595                  */
1596
1597                 /* RED-PEN. Formally speaking, we have broken TCP state
1598                  * machine. State transitions:
1599                  *
1600                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1601                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1602                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1603                  *
1604                  * are legal only when FIN has been sent (i.e. in window),
1605                  * rather than queued out of window. Purists blame.
1606                  *
1607                  * F.e. "RFC state" is ESTABLISHED,
1608                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1609                  *
1610                  * The visible declinations are that sometimes
1611                  * we enter time-wait state, when it is not required really
1612                  * (harmless), do not send active resets, when they are
1613                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1614                  * they look as CLOSING or LAST_ACK for Linux)
1615                  * Probably, I missed some more holelets.
1616                  *                                              --ANK
1617                  */
1618                 tcp_send_fin(sk);
1619         }
1620
1621         sk_stream_wait_close(sk, timeout);
1622
1623 adjudge_to_death:
1624         state = sk->sk_state;
1625         sock_hold(sk);
1626         sock_orphan(sk);
1627         atomic_inc(sk->sk_prot->orphan_count);
1628
1629         /* It is the last release_sock in its life. It will remove backlog. */
1630         release_sock(sk);
1631
1632
1633         /* Now socket is owned by kernel and we acquire BH lock
1634            to finish close. No need to check for user refs.
1635          */
1636         local_bh_disable();
1637         bh_lock_sock(sk);
1638         BUG_TRAP(!sock_owned_by_user(sk));
1639
1640         /* Have we already been destroyed by a softirq or backlog? */
1641         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1642                 goto out;
1643
1644         /*      This is a (useful) BSD violating of the RFC. There is a
1645          *      problem with TCP as specified in that the other end could
1646          *      keep a socket open forever with no application left this end.
1647          *      We use a 3 minute timeout (about the same as BSD) then kill
1648          *      our end. If they send after that then tough - BUT: long enough
1649          *      that we won't make the old 4*rto = almost no time - whoops
1650          *      reset mistake.
1651          *
1652          *      Nope, it was not mistake. It is really desired behaviour
1653          *      f.e. on http servers, when such sockets are useless, but
1654          *      consume significant resources. Let's do it with special
1655          *      linger2 option.                                 --ANK
1656          */
1657
1658         if (sk->sk_state == TCP_FIN_WAIT2) {
1659                 struct tcp_sock *tp = tcp_sk(sk);
1660                 if (tp->linger2 < 0) {
1661                         tcp_set_state(sk, TCP_CLOSE);
1662                         tcp_send_active_reset(sk, GFP_ATOMIC);
1663                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1664                 } else {
1665                         const int tmo = tcp_fin_time(sk);
1666
1667                         if (tmo > TCP_TIMEWAIT_LEN) {
1668                                 inet_csk_reset_keepalive_timer(sk,
1669                                                 tmo - TCP_TIMEWAIT_LEN);
1670                         } else {
1671                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1672                                 goto out;
1673                         }
1674                 }
1675         }
1676         if (sk->sk_state != TCP_CLOSE) {
1677                 sk_stream_mem_reclaim(sk);
1678                 if (atomic_read(sk->sk_prot->orphan_count) > sysctl_tcp_max_orphans ||
1679                     (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
1680                      atomic_read(&tcp_memory_allocated) > sysctl_tcp_mem[2])) {
1681                         if (net_ratelimit())
1682                                 printk(KERN_INFO "TCP: too many of orphaned "
1683                                        "sockets\n");
1684                         tcp_set_state(sk, TCP_CLOSE);
1685                         tcp_send_active_reset(sk, GFP_ATOMIC);
1686                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1687                 }
1688         }
1689
1690         if (sk->sk_state == TCP_CLOSE)
1691                 inet_csk_destroy_sock(sk);
1692         /* Otherwise, socket is reprieved until protocol close. */
1693
1694 out:
1695         bh_unlock_sock(sk);
1696         local_bh_enable();
1697         sock_put(sk);
1698 }
1699
1700 /* These states need RST on ABORT according to RFC793 */
1701
1702 static inline int tcp_need_reset(int state)
1703 {
1704         return (1 << state) &
1705                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1706                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1707 }
1708
1709 int tcp_disconnect(struct sock *sk, int flags)
1710 {
1711         struct inet_sock *inet = inet_sk(sk);
1712         struct inet_connection_sock *icsk = inet_csk(sk);
1713         struct tcp_sock *tp = tcp_sk(sk);
1714         int err = 0;
1715         int old_state = sk->sk_state;
1716
1717         if (old_state != TCP_CLOSE)
1718                 tcp_set_state(sk, TCP_CLOSE);
1719
1720         /* ABORT function of RFC793 */
1721         if (old_state == TCP_LISTEN) {
1722                 inet_csk_listen_stop(sk);
1723         } else if (tcp_need_reset(old_state) ||
1724                    (tp->snd_nxt != tp->write_seq &&
1725                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1726                 /* The last check adjusts for discrepancy of Linux wrt. RFC
1727                  * states
1728                  */
1729                 tcp_send_active_reset(sk, gfp_any());
1730                 sk->sk_err = ECONNRESET;
1731         } else if (old_state == TCP_SYN_SENT)
1732                 sk->sk_err = ECONNRESET;
1733
1734         tcp_clear_xmit_timers(sk);
1735         __skb_queue_purge(&sk->sk_receive_queue);
1736         sk_stream_writequeue_purge(sk);
1737         __skb_queue_purge(&tp->out_of_order_queue);
1738 #ifdef CONFIG_NET_DMA
1739         __skb_queue_purge(&sk->sk_async_wait_queue);
1740 #endif
1741
1742         inet->dport = 0;
1743
1744         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1745                 inet_reset_saddr(sk);
1746
1747         sk->sk_shutdown = 0;
1748         sock_reset_flag(sk, SOCK_DONE);
1749         tp->srtt = 0;
1750         if ((tp->write_seq += tp->max_window + 2) == 0)
1751                 tp->write_seq = 1;
1752         icsk->icsk_backoff = 0;
1753         tp->snd_cwnd = 2;
1754         icsk->icsk_probes_out = 0;
1755         tp->packets_out = 0;
1756         tp->snd_ssthresh = 0x7fffffff;
1757         tp->snd_cwnd_cnt = 0;
1758         tp->bytes_acked = 0;
1759         tcp_set_ca_state(sk, TCP_CA_Open);
1760         tcp_clear_retrans(tp);
1761         inet_csk_delack_init(sk);
1762         sk->sk_send_head = NULL;
1763         tp->rx_opt.saw_tstamp = 0;
1764         tcp_sack_reset(&tp->rx_opt);
1765         __sk_dst_reset(sk);
1766
1767         BUG_TRAP(!inet->num || icsk->icsk_bind_hash);
1768
1769         sk->sk_error_report(sk);
1770         return err;
1771 }
1772
1773 /*
1774  *      Socket option code for TCP.
1775  */
1776 static int do_tcp_setsockopt(struct sock *sk, int level,
1777                 int optname, char __user *optval, int optlen)
1778 {
1779         struct tcp_sock *tp = tcp_sk(sk);
1780         struct inet_connection_sock *icsk = inet_csk(sk);
1781         int val;
1782         int err = 0;
1783
1784         /* This is a string value all the others are int's */
1785         if (optname == TCP_CONGESTION) {
1786                 char name[TCP_CA_NAME_MAX];
1787
1788                 if (optlen < 1)
1789                         return -EINVAL;
1790
1791                 val = strncpy_from_user(name, optval,
1792                                         min(TCP_CA_NAME_MAX-1, optlen));
1793                 if (val < 0)
1794                         return -EFAULT;
1795                 name[val] = 0;
1796
1797                 lock_sock(sk);
1798                 err = tcp_set_congestion_control(sk, name);
1799                 release_sock(sk);
1800                 return err;
1801         }
1802
1803         if (optlen < sizeof(int))
1804                 return -EINVAL;
1805
1806         if (get_user(val, (int __user *)optval))
1807                 return -EFAULT;
1808
1809         lock_sock(sk);
1810
1811         switch (optname) {
1812         case TCP_MAXSEG:
1813                 /* Values greater than interface MTU won't take effect. However
1814                  * at the point when this call is done we typically don't yet
1815                  * know which interface is going to be used */
1816                 if (val < 8 || val > MAX_TCP_WINDOW) {
1817                         err = -EINVAL;
1818                         break;
1819                 }
1820                 tp->rx_opt.user_mss = val;
1821                 break;
1822
1823         case TCP_NODELAY:
1824                 if (val) {
1825                         /* TCP_NODELAY is weaker than TCP_CORK, so that
1826                          * this option on corked socket is remembered, but
1827                          * it is not activated until cork is cleared.
1828                          *
1829                          * However, when TCP_NODELAY is set we make
1830                          * an explicit push, which overrides even TCP_CORK
1831                          * for currently queued segments.
1832                          */
1833                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
1834                         tcp_push_pending_frames(sk, tp);
1835                 } else {
1836                         tp->nonagle &= ~TCP_NAGLE_OFF;
1837                 }
1838                 break;
1839
1840         case TCP_CORK:
1841                 /* When set indicates to always queue non-full frames.
1842                  * Later the user clears this option and we transmit
1843                  * any pending partial frames in the queue.  This is
1844                  * meant to be used alongside sendfile() to get properly
1845                  * filled frames when the user (for example) must write
1846                  * out headers with a write() call first and then use
1847                  * sendfile to send out the data parts.
1848                  *
1849                  * TCP_CORK can be set together with TCP_NODELAY and it is
1850                  * stronger than TCP_NODELAY.
1851                  */
1852                 if (val) {
1853                         tp->nonagle |= TCP_NAGLE_CORK;
1854                 } else {
1855                         tp->nonagle &= ~TCP_NAGLE_CORK;
1856                         if (tp->nonagle&TCP_NAGLE_OFF)
1857                                 tp->nonagle |= TCP_NAGLE_PUSH;
1858                         tcp_push_pending_frames(sk, tp);
1859                 }
1860                 break;
1861
1862         case TCP_KEEPIDLE:
1863                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
1864                         err = -EINVAL;
1865                 else {
1866                         tp->keepalive_time = val * HZ;
1867                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
1868                             !((1 << sk->sk_state) &
1869                               (TCPF_CLOSE | TCPF_LISTEN))) {
1870                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
1871                                 if (tp->keepalive_time > elapsed)
1872                                         elapsed = tp->keepalive_time - elapsed;
1873                                 else
1874                                         elapsed = 0;
1875                                 inet_csk_reset_keepalive_timer(sk, elapsed);
1876                         }
1877                 }
1878                 break;
1879         case TCP_KEEPINTVL:
1880                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
1881                         err = -EINVAL;
1882                 else
1883                         tp->keepalive_intvl = val * HZ;
1884                 break;
1885         case TCP_KEEPCNT:
1886                 if (val < 1 || val > MAX_TCP_KEEPCNT)
1887                         err = -EINVAL;
1888                 else
1889                         tp->keepalive_probes = val;
1890                 break;
1891         case TCP_SYNCNT:
1892                 if (val < 1 || val > MAX_TCP_SYNCNT)
1893                         err = -EINVAL;
1894                 else
1895                         icsk->icsk_syn_retries = val;
1896                 break;
1897
1898         case TCP_LINGER2:
1899                 if (val < 0)
1900                         tp->linger2 = -1;
1901                 else if (val > sysctl_tcp_fin_timeout / HZ)
1902                         tp->linger2 = 0;
1903                 else
1904                         tp->linger2 = val * HZ;
1905                 break;
1906
1907         case TCP_DEFER_ACCEPT:
1908                 icsk->icsk_accept_queue.rskq_defer_accept = 0;
1909                 if (val > 0) {
1910                         /* Translate value in seconds to number of
1911                          * retransmits */
1912                         while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
1913                                val > ((TCP_TIMEOUT_INIT / HZ) <<
1914                                        icsk->icsk_accept_queue.rskq_defer_accept))
1915                                 icsk->icsk_accept_queue.rskq_defer_accept++;
1916                         icsk->icsk_accept_queue.rskq_defer_accept++;
1917                 }
1918                 break;
1919
1920         case TCP_WINDOW_CLAMP:
1921                 if (!val) {
1922                         if (sk->sk_state != TCP_CLOSE) {
1923                                 err = -EINVAL;
1924                                 break;
1925                         }
1926                         tp->window_clamp = 0;
1927                 } else
1928                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
1929                                                 SOCK_MIN_RCVBUF / 2 : val;
1930                 break;
1931
1932         case TCP_QUICKACK:
1933                 if (!val) {
1934                         icsk->icsk_ack.pingpong = 1;
1935                 } else {
1936                         icsk->icsk_ack.pingpong = 0;
1937                         if ((1 << sk->sk_state) &
1938                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
1939                             inet_csk_ack_scheduled(sk)) {
1940                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
1941                                 tcp_cleanup_rbuf(sk, 1);
1942                                 if (!(val & 1))
1943                                         icsk->icsk_ack.pingpong = 1;
1944                         }
1945                 }
1946                 break;
1947
1948 #ifdef CONFIG_TCP_MD5SIG
1949         case TCP_MD5SIG:
1950                 /* Read the IP->Key mappings from userspace */
1951                 err = tp->af_specific->md5_parse(sk, optval, optlen);
1952                 break;
1953 #endif
1954
1955         default:
1956                 err = -ENOPROTOOPT;
1957                 break;
1958         };
1959         release_sock(sk);
1960         return err;
1961 }
1962
1963 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
1964                    int optlen)
1965 {
1966         struct inet_connection_sock *icsk = inet_csk(sk);
1967
1968         if (level != SOL_TCP)
1969                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
1970                                                      optval, optlen);
1971         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1972 }
1973
1974 #ifdef CONFIG_COMPAT
1975 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
1976                           char __user *optval, int optlen)
1977 {
1978         if (level != SOL_TCP)
1979                 return inet_csk_compat_setsockopt(sk, level, optname,
1980                                                   optval, optlen);
1981         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1982 }
1983
1984 EXPORT_SYMBOL(compat_tcp_setsockopt);
1985 #endif
1986
1987 /* Return information about state of tcp endpoint in API format. */
1988 void tcp_get_info(struct sock *sk, struct tcp_info *info)
1989 {
1990         struct tcp_sock *tp = tcp_sk(sk);
1991         const struct inet_connection_sock *icsk = inet_csk(sk);
1992         u32 now = tcp_time_stamp;
1993
1994         memset(info, 0, sizeof(*info));
1995
1996         info->tcpi_state = sk->sk_state;
1997         info->tcpi_ca_state = icsk->icsk_ca_state;
1998         info->tcpi_retransmits = icsk->icsk_retransmits;
1999         info->tcpi_probes = icsk->icsk_probes_out;
2000         info->tcpi_backoff = icsk->icsk_backoff;
2001
2002         if (tp->rx_opt.tstamp_ok)
2003                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2004         if (tp->rx_opt.sack_ok)
2005                 info->tcpi_options |= TCPI_OPT_SACK;
2006         if (tp->rx_opt.wscale_ok) {
2007                 info->tcpi_options |= TCPI_OPT_WSCALE;
2008                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2009                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2010         }
2011
2012         if (tp->ecn_flags&TCP_ECN_OK)
2013                 info->tcpi_options |= TCPI_OPT_ECN;
2014
2015         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2016         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2017         info->tcpi_snd_mss = tp->mss_cache;
2018         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2019
2020         info->tcpi_unacked = tp->packets_out;
2021         info->tcpi_sacked = tp->sacked_out;
2022         info->tcpi_lost = tp->lost_out;
2023         info->tcpi_retrans = tp->retrans_out;
2024         info->tcpi_fackets = tp->fackets_out;
2025
2026         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2027         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2028         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2029
2030         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2031         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2032         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2033         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2034         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2035         info->tcpi_snd_cwnd = tp->snd_cwnd;
2036         info->tcpi_advmss = tp->advmss;
2037         info->tcpi_reordering = tp->reordering;
2038
2039         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2040         info->tcpi_rcv_space = tp->rcvq_space.space;
2041
2042         info->tcpi_total_retrans = tp->total_retrans;
2043 }
2044
2045 EXPORT_SYMBOL_GPL(tcp_get_info);
2046
2047 static int do_tcp_getsockopt(struct sock *sk, int level,
2048                 int optname, char __user *optval, int __user *optlen)
2049 {
2050         struct inet_connection_sock *icsk = inet_csk(sk);
2051         struct tcp_sock *tp = tcp_sk(sk);
2052         int val, len;
2053
2054         if (get_user(len, optlen))
2055                 return -EFAULT;
2056
2057         len = min_t(unsigned int, len, sizeof(int));
2058
2059         if (len < 0)
2060                 return -EINVAL;
2061
2062         switch (optname) {
2063         case TCP_MAXSEG:
2064                 val = tp->mss_cache;
2065                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2066                         val = tp->rx_opt.user_mss;
2067                 break;
2068         case TCP_NODELAY:
2069                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2070                 break;
2071         case TCP_CORK:
2072                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2073                 break;
2074         case TCP_KEEPIDLE:
2075                 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2076                 break;
2077         case TCP_KEEPINTVL:
2078                 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2079                 break;
2080         case TCP_KEEPCNT:
2081                 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2082                 break;
2083         case TCP_SYNCNT:
2084                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2085                 break;
2086         case TCP_LINGER2:
2087                 val = tp->linger2;
2088                 if (val >= 0)
2089                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2090                 break;
2091         case TCP_DEFER_ACCEPT:
2092                 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2093                         ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2094                 break;
2095         case TCP_WINDOW_CLAMP:
2096                 val = tp->window_clamp;
2097                 break;
2098         case TCP_INFO: {
2099                 struct tcp_info info;
2100
2101                 if (get_user(len, optlen))
2102                         return -EFAULT;
2103
2104                 tcp_get_info(sk, &info);
2105
2106                 len = min_t(unsigned int, len, sizeof(info));
2107                 if (put_user(len, optlen))
2108                         return -EFAULT;
2109                 if (copy_to_user(optval, &info, len))
2110                         return -EFAULT;
2111                 return 0;
2112         }
2113         case TCP_QUICKACK:
2114                 val = !icsk->icsk_ack.pingpong;
2115                 break;
2116
2117         case TCP_CONGESTION:
2118                 if (get_user(len, optlen))
2119                         return -EFAULT;
2120                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2121                 if (put_user(len, optlen))
2122                         return -EFAULT;
2123                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2124                         return -EFAULT;
2125                 return 0;
2126         default:
2127                 return -ENOPROTOOPT;
2128         };
2129
2130         if (put_user(len, optlen))
2131                 return -EFAULT;
2132         if (copy_to_user(optval, &val, len))
2133                 return -EFAULT;
2134         return 0;
2135 }
2136
2137 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2138                    int __user *optlen)
2139 {
2140         struct inet_connection_sock *icsk = inet_csk(sk);
2141
2142         if (level != SOL_TCP)
2143                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2144                                                      optval, optlen);
2145         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2146 }
2147
2148 #ifdef CONFIG_COMPAT
2149 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2150                           char __user *optval, int __user *optlen)
2151 {
2152         if (level != SOL_TCP)
2153                 return inet_csk_compat_getsockopt(sk, level, optname,
2154                                                   optval, optlen);
2155         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2156 }
2157
2158 EXPORT_SYMBOL(compat_tcp_getsockopt);
2159 #endif
2160
2161 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2162 {
2163         struct sk_buff *segs = ERR_PTR(-EINVAL);
2164         struct tcphdr *th;
2165         unsigned thlen;
2166         unsigned int seq;
2167         __be32 delta;
2168         unsigned int oldlen;
2169         unsigned int len;
2170
2171         if (!pskb_may_pull(skb, sizeof(*th)))
2172                 goto out;
2173
2174         th = skb->h.th;
2175         thlen = th->doff * 4;
2176         if (thlen < sizeof(*th))
2177                 goto out;
2178
2179         if (!pskb_may_pull(skb, thlen))
2180                 goto out;
2181
2182         oldlen = (u16)~skb->len;
2183         __skb_pull(skb, thlen);
2184
2185         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2186                 /* Packet is from an untrusted source, reset gso_segs. */
2187                 int type = skb_shinfo(skb)->gso_type;
2188                 int mss;
2189
2190                 if (unlikely(type &
2191                              ~(SKB_GSO_TCPV4 |
2192                                SKB_GSO_DODGY |
2193                                SKB_GSO_TCP_ECN |
2194                                SKB_GSO_TCPV6 |
2195                                0) ||
2196                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2197                         goto out;
2198
2199                 mss = skb_shinfo(skb)->gso_size;
2200                 skb_shinfo(skb)->gso_segs = (skb->len + mss - 1) / mss;
2201
2202                 segs = NULL;
2203                 goto out;
2204         }
2205
2206         segs = skb_segment(skb, features);
2207         if (IS_ERR(segs))
2208                 goto out;
2209
2210         len = skb_shinfo(skb)->gso_size;
2211         delta = htonl(oldlen + (thlen + len));
2212
2213         skb = segs;
2214         th = skb->h.th;
2215         seq = ntohl(th->seq);
2216
2217         do {
2218                 th->fin = th->psh = 0;
2219
2220                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2221                                        (__force u32)delta));
2222                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2223                         th->check = csum_fold(csum_partial(skb->h.raw, thlen,
2224                                                            skb->csum));
2225
2226                 seq += len;
2227                 skb = skb->next;
2228                 th = skb->h.th;
2229
2230                 th->seq = htonl(seq);
2231                 th->cwr = 0;
2232         } while (skb->next);
2233
2234         delta = htonl(oldlen + (skb->tail - skb->h.raw) + skb->data_len);
2235         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2236                                 (__force u32)delta));
2237         if (skb->ip_summed != CHECKSUM_PARTIAL)
2238                 th->check = csum_fold(csum_partial(skb->h.raw, thlen,
2239                                                    skb->csum));
2240
2241 out:
2242         return segs;
2243 }
2244 EXPORT_SYMBOL(tcp_tso_segment);
2245
2246 #ifdef CONFIG_TCP_MD5SIG
2247 static unsigned long tcp_md5sig_users;
2248 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2249 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2250
2251 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2252 {
2253         int cpu;
2254         for_each_possible_cpu(cpu) {
2255                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2256                 if (p) {
2257                         if (p->md5_desc.tfm)
2258                                 crypto_free_hash(p->md5_desc.tfm);
2259                         kfree(p);
2260                         p = NULL;
2261                 }
2262         }
2263         free_percpu(pool);
2264 }
2265
2266 void tcp_free_md5sig_pool(void)
2267 {
2268         struct tcp_md5sig_pool **pool = NULL;
2269
2270         spin_lock_bh(&tcp_md5sig_pool_lock);
2271         if (--tcp_md5sig_users == 0) {
2272                 pool = tcp_md5sig_pool;
2273                 tcp_md5sig_pool = NULL;
2274         }
2275         spin_unlock_bh(&tcp_md5sig_pool_lock);
2276         if (pool)
2277                 __tcp_free_md5sig_pool(pool);
2278 }
2279
2280 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2281
2282 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(void)
2283 {
2284         int cpu;
2285         struct tcp_md5sig_pool **pool;
2286
2287         pool = alloc_percpu(struct tcp_md5sig_pool *);
2288         if (!pool)
2289                 return NULL;
2290
2291         for_each_possible_cpu(cpu) {
2292                 struct tcp_md5sig_pool *p;
2293                 struct crypto_hash *hash;
2294
2295                 p = kzalloc(sizeof(*p), GFP_KERNEL);
2296                 if (!p)
2297                         goto out_free;
2298                 *per_cpu_ptr(pool, cpu) = p;
2299
2300                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2301                 if (!hash || IS_ERR(hash))
2302                         goto out_free;
2303
2304                 p->md5_desc.tfm = hash;
2305         }
2306         return pool;
2307 out_free:
2308         __tcp_free_md5sig_pool(pool);
2309         return NULL;
2310 }
2311
2312 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(void)
2313 {
2314         struct tcp_md5sig_pool **pool;
2315         int alloc = 0;
2316
2317 retry:
2318         spin_lock_bh(&tcp_md5sig_pool_lock);
2319         pool = tcp_md5sig_pool;
2320         if (tcp_md5sig_users++ == 0) {
2321                 alloc = 1;
2322                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2323         } else if (!pool) {
2324                 tcp_md5sig_users--;
2325                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2326                 cpu_relax();
2327                 goto retry;
2328         } else
2329                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2330
2331         if (alloc) {
2332                 /* we cannot hold spinlock here because this may sleep. */
2333                 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool();
2334                 spin_lock_bh(&tcp_md5sig_pool_lock);
2335                 if (!p) {
2336                         tcp_md5sig_users--;
2337                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2338                         return NULL;
2339                 }
2340                 pool = tcp_md5sig_pool;
2341                 if (pool) {
2342                         /* oops, it has already been assigned. */
2343                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2344                         __tcp_free_md5sig_pool(p);
2345                 } else {
2346                         tcp_md5sig_pool = pool = p;
2347                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2348                 }
2349         }
2350         return pool;
2351 }
2352
2353 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2354
2355 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2356 {
2357         struct tcp_md5sig_pool **p;
2358         spin_lock_bh(&tcp_md5sig_pool_lock);
2359         p = tcp_md5sig_pool;
2360         if (p)
2361                 tcp_md5sig_users++;
2362         spin_unlock_bh(&tcp_md5sig_pool_lock);
2363         return (p ? *per_cpu_ptr(p, cpu) : NULL);
2364 }
2365
2366 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2367
2368 void __tcp_put_md5sig_pool(void)
2369 {
2370         tcp_free_md5sig_pool();
2371 }
2372
2373 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2374 #endif
2375
2376 extern void __skb_cb_too_small_for_tcp(int, int);
2377 extern struct tcp_congestion_ops tcp_reno;
2378
2379 static __initdata unsigned long thash_entries;
2380 static int __init set_thash_entries(char *str)
2381 {
2382         if (!str)
2383                 return 0;
2384         thash_entries = simple_strtoul(str, &str, 0);
2385         return 1;
2386 }
2387 __setup("thash_entries=", set_thash_entries);
2388
2389 void __init tcp_init(void)
2390 {
2391         struct sk_buff *skb = NULL;
2392         unsigned long limit;
2393         int order, i, max_share;
2394
2395         if (sizeof(struct tcp_skb_cb) > sizeof(skb->cb))
2396                 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb),
2397                                            sizeof(skb->cb));
2398
2399         tcp_hashinfo.bind_bucket_cachep =
2400                 kmem_cache_create("tcp_bind_bucket",
2401                                   sizeof(struct inet_bind_bucket), 0,
2402                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
2403
2404         /* Size and allocate the main established and bind bucket
2405          * hash tables.
2406          *
2407          * The methodology is similar to that of the buffer cache.
2408          */
2409         tcp_hashinfo.ehash =
2410                 alloc_large_system_hash("TCP established",
2411                                         sizeof(struct inet_ehash_bucket),
2412                                         thash_entries,
2413                                         (num_physpages >= 128 * 1024) ?
2414                                         13 : 15,
2415                                         0,
2416                                         &tcp_hashinfo.ehash_size,
2417                                         NULL,
2418                                         0);
2419         tcp_hashinfo.ehash_size = (1 << tcp_hashinfo.ehash_size) >> 1;
2420         for (i = 0; i < (tcp_hashinfo.ehash_size << 1); i++) {
2421                 rwlock_init(&tcp_hashinfo.ehash[i].lock);
2422                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].chain);
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         /* Set the pressure threshold to be a fraction of global memory that
2462          * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2463          * memory, with a floor of 128 pages.
2464          */
2465         limit = min(nr_all_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
2466         limit = (limit * (nr_all_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
2467         limit = max(limit, 128UL);
2468         sysctl_tcp_mem[0] = limit / 4 * 3;
2469         sysctl_tcp_mem[1] = limit;
2470         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2471
2472         /* Set per-socket limits to no more than 1/128 the pressure threshold */
2473         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2474         max_share = min(4UL*1024*1024, limit);
2475
2476         sysctl_tcp_wmem[0] = SK_STREAM_MEM_QUANTUM;
2477         sysctl_tcp_wmem[1] = 16*1024;
2478         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2479
2480         sysctl_tcp_rmem[0] = SK_STREAM_MEM_QUANTUM;
2481         sysctl_tcp_rmem[1] = 87380;
2482         sysctl_tcp_rmem[2] = max(87380, max_share);
2483
2484         printk(KERN_INFO "TCP: Hash tables configured "
2485                "(established %d bind %d)\n",
2486                tcp_hashinfo.ehash_size << 1, tcp_hashinfo.bhash_size);
2487
2488         tcp_register_congestion_control(&tcp_reno);
2489 }
2490
2491 EXPORT_SYMBOL(tcp_close);
2492 EXPORT_SYMBOL(tcp_disconnect);
2493 EXPORT_SYMBOL(tcp_getsockopt);
2494 EXPORT_SYMBOL(tcp_ioctl);
2495 EXPORT_SYMBOL(tcp_poll);
2496 EXPORT_SYMBOL(tcp_read_sock);
2497 EXPORT_SYMBOL(tcp_recvmsg);
2498 EXPORT_SYMBOL(tcp_sendmsg);
2499 EXPORT_SYMBOL(tcp_sendpage);
2500 EXPORT_SYMBOL(tcp_setsockopt);
2501 EXPORT_SYMBOL(tcp_shutdown);
2502 EXPORT_SYMBOL(tcp_statistics);