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  * Authors:     Ross Biro
   9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
  11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
  12  *              Florian La Roche, <flla@stud.uni-sb.de>
  13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
  15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
  16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
  17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  18  *              Jorge Cwik, <jorge@laser.satlink.net>
  19  *
  20  * Fixes:
  21  *              Alan Cox        :       Numerous verify_area() calls
  22  *              Alan Cox        :       Set the ACK bit on a reset
  23  *              Alan Cox        :       Stopped it crashing if it closed while
  24  *                                      sk->inuse=1 and was trying to connect
  25  *                                      (tcp_err()).
  26  *              Alan Cox        :       All icmp error handling was broken
  27  *                                      pointers passed where wrong and the
  28  *                                      socket was looked up backwards. Nobody
  29  *                                      tested any icmp error code obviously.
  30  *              Alan Cox        :       tcp_err() now handled properly. It
  31  *                                      wakes people on errors. poll
  32  *                                      behaves and the icmp error race
  33  *                                      has gone by moving it into sock.c
  34  *              Alan Cox        :       tcp_send_reset() fixed to work for
  35  *                                      everything not just packets for
  36  *                                      unknown sockets.
  37  *              Alan Cox        :       tcp option processing.
  38  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
  39  *                                      syn rule wrong]
  40  *              Herp Rosmanith  :       More reset fixes
  41  *              Alan Cox        :       No longer acks invalid rst frames.
  42  *                                      Acking any kind of RST is right out.
  43  *              Alan Cox        :       Sets an ignore me flag on an rst
  44  *                                      receive otherwise odd bits of prattle
  45  *                                      escape still
  46  *              Alan Cox        :       Fixed another acking RST frame bug.
  47  *                                      Should stop LAN workplace lockups.
  48  *              Alan Cox        :       Some tidyups using the new skb list
  49  *                                      facilities
  50  *              Alan Cox        :       sk->keepopen now seems to work
  51  *              Alan Cox        :       Pulls options out correctly on accepts
  52  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
  53  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
  54  *                                      bit to skb ops.
  55  *              Alan Cox        :       Tidied tcp_data to avoid a potential
  56  *                                      nasty.
  57  *              Alan Cox        :       Added some better commenting, as the
  58  *                                      tcp is hard to follow
  59  *              Alan Cox        :       Removed incorrect check for 20 * psh
  60  *      Michael O'Reilly        :       ack < copied bug fix.
  61  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
  62  *              Alan Cox        :       FIN with no memory -> CRASH
  63  *              Alan Cox        :       Added socket option proto entries.
  64  *                                      Also added awareness of them to accept.
  65  *              Alan Cox        :       Added TCP options (SOL_TCP)
  66  *              Alan Cox        :       Switched wakeup calls to callbacks,
  67  *                                      so the kernel can layer network
  68  *                                      sockets.
  69  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
  70  *              Alan Cox        :       Handle FIN (more) properly (we hope).
  71  *              Alan Cox        :       RST frames sent on unsynchronised
  72  *                                      state ack error.
  73  *              Alan Cox        :       Put in missing check for SYN bit.
  74  *              Alan Cox        :       Added tcp_select_window() aka NET2E
  75  *                                      window non shrink trick.
  76  *              Alan Cox        :       Added a couple of small NET2E timer
  77  *                                      fixes
  78  *              Charles Hedrick :       TCP fixes
  79  *              Toomas Tamm     :       TCP window fixes
  80  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
  81  *              Charles Hedrick :       Rewrote most of it to actually work
  82  *              Linus           :       Rewrote tcp_read() and URG handling
  83  *                                      completely
  84  *              Gerhard Koerting:       Fixed some missing timer handling
  85  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
  86  *              Gerhard Koerting:       PC/TCP workarounds
  87  *              Adam Caldwell   :       Assorted timer/timing errors
  88  *              Matthew Dillon  :       Fixed another RST bug
  89  *              Alan Cox        :       Move to kernel side addressing changes.
  90  *              Alan Cox        :       Beginning work on TCP fastpathing
  91  *                                      (not yet usable)
  92  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
  93  *              Alan Cox        :       TCP fast path debugging
  94  *              Alan Cox        :       Window clamping
  95  *              Michael Riepe   :       Bug in tcp_check()
  96  *              Matt Dillon     :       More TCP improvements and RST bug fixes
  97  *              Matt Dillon     :       Yet more small nasties remove from the
  98  *                                      TCP code (Be very nice to this man if
  99  *                                      tcp finally works 100%) 8)
 100  *              Alan Cox        :       BSD accept semantics.
 101  *              Alan Cox        :       Reset on closedown bug.
 102  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
 103  *              Michael Pall    :       Handle poll() after URG properly in
 104  *                                      all cases.
 105  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
 106  *                                      (multi URG PUSH broke rlogin).
 107  *              Michael Pall    :       Fix the multi URG PUSH problem in
 108  *                                      tcp_readable(), poll() after URG
 109  *                                      works now.
 110  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
 111  *                                      BSD api.
 112  *              Alan Cox        :       Changed the semantics of sk->socket to
 113  *                                      fix a race and a signal problem with
 114  *                                      accept() and async I/O.
 115  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
 116  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
 117  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
 118  *                                      clients/servers which listen in on
 119  *                                      fixed ports.
 120  *              Alan Cox        :       Cleaned the above up and shrank it to
 121  *                                      a sensible code size.
 122  *              Alan Cox        :       Self connect lockup fix.
 123  *              Alan Cox        :       No connect to multicast.
 124  *              Ross Biro       :       Close unaccepted children on master
 125  *                                      socket close.
 126  *              Alan Cox        :       Reset tracing code.
 127  *              Alan Cox        :       Spurious resets on shutdown.
 128  *              Alan Cox        :       Giant 15 minute/60 second timer error
 129  *              Alan Cox        :       Small whoops in polling before an
 130  *                                      accept.
 131  *              Alan Cox        :       Kept the state trace facility since
 132  *                                      it's handy for debugging.
 133  *              Alan Cox        :       More reset handler fixes.
 134  *              Alan Cox        :       Started rewriting the code based on
 135  *                                      the RFC's for other useful protocol
 136  *                                      references see: Comer, KA9Q NOS, and
 137  *                                      for a reference on the difference
 138  *                                      between specifications and how BSD
 139  *                                      works see the 4.4lite source.
 140  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
 141  *                                      close.
 142  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
 143  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
 144  *              Alan Cox        :       Reimplemented timers as per the RFC
 145  *                                      and using multiple timers for sanity.
 146  *              Alan Cox        :       Small bug fixes, and a lot of new
 147  *                                      comments.
 148  *              Alan Cox        :       Fixed dual reader crash by locking
 149  *                                      the buffers (much like datagram.c)
 150  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
 151  *                                      now gets fed up of retrying without
 152  *                                      (even a no space) answer.
 153  *              Alan Cox        :       Extracted closing code better
 154  *              Alan Cox        :       Fixed the closing state machine to
 155  *                                      resemble the RFC.
 156  *              Alan Cox        :       More 'per spec' fixes.
 157  *              Jorge Cwik      :       Even faster checksumming.
 158  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
 159  *                                      only frames. At least one pc tcp stack
 160  *                                      generates them.
 161  *              Alan Cox        :       Cache last socket.
 162  *              Alan Cox        :       Per route irtt.
 163  *              Matt Day        :       poll()->select() match BSD precisely on error
 164  *              Alan Cox        :       New buffers
 165  *              Marc Tamsky     :       Various sk->prot->retransmits and
 166  *                                      sk->retransmits misupdating fixed.
 167  *                                      Fixed tcp_write_timeout: stuck close,
 168  *                                      and TCP syn retries gets used now.
 169  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
 170  *                                      ack if state is TCP_CLOSED.
 171  *              Alan Cox        :       Look up device on a retransmit - routes may
 172  *                                      change. Doesn't yet cope with MSS shrink right
 173  *                                      but it's a start!
 174  *              Marc Tamsky     :       Closing in closing fixes.
 175  *              Mike Shaver     :       RFC1122 verifications.
 176  *              Alan Cox        :       rcv_saddr errors.
 177  *              Alan Cox        :       Block double connect().
 178  *              Alan Cox        :       Small hooks for enSKIP.
 179  *              Alexey Kuznetsov:       Path MTU discovery.
 180  *              Alan Cox        :       Support soft errors.
 181  *              Alan Cox        :       Fix MTU discovery pathological case
 182  *                                      when the remote claims no mtu!
 183  *              Marc Tamsky     :       TCP_CLOSE fix.
 184  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
 185  *                                      window but wrong (fixes NT lpd problems)
 186  *              Pedro Roque     :       Better TCP window handling, delayed ack.
 187  *              Joerg Reuter    :       No modification of locked buffers in
 188  *                                      tcp_do_retransmit()
 189  *              Eric Schenk     :       Changed receiver side silly window
 190  *                                      avoidance algorithm to BSD style
 191  *                                      algorithm. This doubles throughput
 192  *                                      against machines running Solaris,
 193  *                                      and seems to result in general
 194  *                                      improvement.
 195  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
 196  *      Willy Konynenberg       :       Transparent proxying support.
 197  *      Mike McLagan            :       Routing by source
 198  *              Keith Owens     :       Do proper merging with partial SKB's in
 199  *                                      tcp_do_sendmsg to avoid burstiness.
 200  *              Eric Schenk     :       Fix fast close down bug with
 201  *                                      shutdown() followed by close().
 202  *              Andi Kleen      :       Make poll agree with SIGIO
 203  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
 204  *                                      lingertime == 0 (RFC 793 ABORT Call)
 205  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
 206  *                                      csum_and_copy_from_user() if possible.
 207  *
 208  *              This program is free software; you can redistribute it and/or
 209  *              modify it under the terms of the GNU General Public License
 210  *              as published by the Free Software Foundation; either version
 211  *              2 of the License, or(at your option) any later version.
 212  *
 213  * Description of States:
 214  *
 215  *      TCP_SYN_SENT            sent a connection request, waiting for ack
 216  *
 217  *      TCP_SYN_RECV            received a connection request, sent ack,
 218  *                              waiting for final ack in three-way handshake.
 219  *
 220  *      TCP_ESTABLISHED         connection established
 221  *
 222  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
 223  *                              transmission of remaining buffered data
 224  *
 225  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
 226  *                              to shutdown
 227  *
 228  *      TCP_CLOSING             both sides have shutdown but we still have
 229  *                              data we have to finish sending
 230  *
 231  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
 232  *                              closed, can only be entered from FIN_WAIT2
 233  *                              or CLOSING.  Required because the other end
 234  *                              may not have gotten our last ACK causing it
 235  *                              to retransmit the data packet (which we ignore)
 236  *
 237  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
 238  *                              us to finish writing our data and to shutdown
 239  *                              (we have to close() to move on to LAST_ACK)
 240  *
 241  *      TCP_LAST_ACK            out side has shutdown after remote has
 242  *                              shutdown.  There may still be data in our
 243  *                              buffer that we have to finish sending
 244  *
 245  *      TCP_CLOSE               socket is finished
 246  */
 247
 248 #include <linux/kernel.h>
 249 #include <linux/module.h>
 250 #include <linux/types.h>
 251 #include <linux/fcntl.h>
 252 #include <linux/poll.h>
 253 #include <linux/init.h>
 254 #include <linux/fs.h>
 255 #include <linux/skbuff.h>
 256 #include <linux/scatterlist.h>
 257 #include <linux/splice.h>
 258 #include <linux/net.h>
 259 #include <linux/socket.h>
 260 #include <linux/random.h>
 261 #include <linux/bootmem.h>
 262 #include <linux/highmem.h>
 263 #include <linux/swap.h>
 264 #include <linux/cache.h>
 265 #include <linux/err.h>
 266 #include <linux/crypto.h>
 267
 268 #include <net/icmp.h>
 269 #include <net/tcp.h>
 270 #include <net/xfrm.h>
 271 #include <net/ip.h>
 272 #include <net/netdma.h>
 273 #include <net/sock.h>
 274
 275 #include <asm/uaccess.h>
 276 #include <asm/ioctls.h>
 277
 278 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
 279
 280 struct percpu_counter tcp_orphan_count;
 281 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 282
 283 int sysctl_tcp_mem[3] __read_mostly;
 284 int sysctl_tcp_wmem[3] __read_mostly;
 285 int sysctl_tcp_rmem[3] __read_mostly;
 286
 287 EXPORT_SYMBOL(sysctl_tcp_mem);
 288 EXPORT_SYMBOL(sysctl_tcp_rmem);
 289 EXPORT_SYMBOL(sysctl_tcp_wmem);
 290
 291 atomic_t tcp_memory_allocated;  /* Current allocated memory. */
 292 EXPORT_SYMBOL(tcp_memory_allocated);
 293
 294 /*
 295  * Current number of TCP sockets.
 296  */
 297 struct percpu_counter tcp_sockets_allocated;
 298 EXPORT_SYMBOL(tcp_sockets_allocated);
 299
 300 /*
 301  * TCP splice context
 302  */
 303 struct tcp_splice_state {
 304         struct pipe_inode_info *pipe;
 305         size_t len;
 306         unsigned int flags;
 307 };
 308
 309 /*
 310  * Pressure flag: try to collapse.
 311  * Technical note: it is used by multiple contexts non atomically.
 312  * All the __sk_mem_schedule() is of this nature: accounting
 313  * is strict, actions are advisory and have some latency.
 314  */
 315 int tcp_memory_pressure __read_mostly;
 316
 317 EXPORT_SYMBOL(tcp_memory_pressure);
 318
 319 void tcp_enter_memory_pressure(struct sock *sk)
 320 {
 321         if (!tcp_memory_pressure) {
 322                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
 323                 tcp_memory_pressure = 1;
 324         }
 325 }
 326
 327 EXPORT_SYMBOL(tcp_enter_memory_pressure);
 328
 329 /*
 330  *      Wait for a TCP event.
 331  *
 332  *      Note that we don't need to lock the socket, as the upper poll layers
 333  *      take care of normal races (between the test and the event) and we don't
 334  *      go look at any of the socket buffers directly.
 335  */
 336 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 337 {
 338         unsigned int mask;
 339         struct sock *sk = sock->sk;
 340         struct tcp_sock *tp = tcp_sk(sk);
 341
 342         poll_wait(file, sk->sk_sleep, wait);
 343         if (sk->sk_state == TCP_LISTEN)
 344                 return inet_csk_listen_poll(sk);
 345
 346         /* Socket is not locked. We are protected from async events
 347          * by poll logic and correct handling of state changes
 348          * made by other threads is impossible in any case.
 349          */
 350
 351         mask = 0;
 352         if (sk->sk_err)
 353                 mask = POLLERR;
 354
 355         /*
 356          * POLLHUP is certainly not done right. But poll() doesn't
 357          * have a notion of HUP in just one direction, and for a
 358          * socket the read side is more interesting.
 359          *
 360          * Some poll() documentation says that POLLHUP is incompatible
 361          * with the POLLOUT/POLLWR flags, so somebody should check this
 362          * all. But careful, it tends to be safer to return too many
 363          * bits than too few, and you can easily break real applications
 364          * if you don't tell them that something has hung up!
 365          *
 366          * Check-me.
 367          *
 368          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
 369          * our fs/select.c). It means that after we received EOF,
 370          * poll always returns immediately, making impossible poll() on write()
 371          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
 372          * if and only if shutdown has been made in both directions.
 373          * Actually, it is interesting to look how Solaris and DUX
 374          * solve this dilemma. I would prefer, if POLLHUP were maskable,
 375          * then we could set it on SND_SHUTDOWN. BTW examples given
 376          * in Stevens' books assume exactly this behaviour, it explains
 377          * why POLLHUP is incompatible with POLLOUT.    --ANK
 378          *
 379          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 380          * blocking on fresh not-connected or disconnected socket. --ANK
 381          */
 382         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
 383                 mask |= POLLHUP;
 384         if (sk->sk_shutdown & RCV_SHUTDOWN)
 385                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
 386
 387         /* Connected? */
 388         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
 389                 int target = sock_rcvlowat(sk, 0, INT_MAX);
 390
 391                 if (tp->urg_seq == tp->copied_seq &&
 392                     !sock_flag(sk, SOCK_URGINLINE) &&
 393                     tp->urg_data)
 394                         target--;
 395
 396                 /* Potential race condition. If read of tp below will
 397                  * escape above sk->sk_state, we can be illegally awaken
 398                  * in SYN_* states. */
 399                 if (tp->rcv_nxt - tp->copied_seq >= target)
 400                         mask |= POLLIN | POLLRDNORM;
 401
 402                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 403                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
 404                                 mask |= POLLOUT | POLLWRNORM;
 405                         } else {  /* send SIGIO later */
 406                                 set_bit(SOCK_ASYNC_NOSPACE,
 407                                         &sk->sk_socket->flags);
 408                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 409
 410                                 /* Race breaker. If space is freed after
 411                                  * wspace test but before the flags are set,
 412                                  * IO signal will be lost.
 413                                  */
 414                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
 415                                         mask |= POLLOUT | POLLWRNORM;
 416                         }
 417                 }
 418
 419                 if (tp->urg_data & TCP_URG_VALID)
 420                         mask |= POLLPRI;
 421         }
 422         return mask;
 423 }
 424
 425 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 426 {
 427         struct tcp_sock *tp = tcp_sk(sk);
 428         int answ;
 429
 430         switch (cmd) {
 431         case SIOCINQ:
 432                 if (sk->sk_state == TCP_LISTEN)
 433                         return -EINVAL;
 434
 435                 lock_sock(sk);
 436                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 437                         answ = 0;
 438                 else if (sock_flag(sk, SOCK_URGINLINE) ||
 439                          !tp->urg_data ||
 440                          before(tp->urg_seq, tp->copied_seq) ||
 441                          !before(tp->urg_seq, tp->rcv_nxt)) {
 442                         answ = tp->rcv_nxt - tp->copied_seq;
 443
 444                         /* Subtract 1, if FIN is in queue. */
 445                         if (answ && !skb_queue_empty(&sk->sk_receive_queue))
 446                                 answ -=
 447                        tcp_hdr((struct sk_buff *)sk->sk_receive_queue.prev)->fin;
 448                 } else
 449                         answ = tp->urg_seq - tp->copied_seq;
 450                 release_sock(sk);
 451                 break;
 452         case SIOCATMARK:
 453                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 454                 break;
 455         case SIOCOUTQ:
 456                 if (sk->sk_state == TCP_LISTEN)
 457                         return -EINVAL;
 458
 459                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 460                         answ = 0;
 461                 else
 462                         answ = tp->write_seq - tp->snd_una;
 463                 break;
 464         default:
 465                 return -ENOIOCTLCMD;
 466         }
 467
 468         return put_user(answ, (int __user *)arg);
 469 }
 470
 471 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 472 {
 473         TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
 474         tp->pushed_seq = tp->write_seq;
 475 }
 476
 477 static inline int forced_push(struct tcp_sock *tp)
 478 {
 479         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 480 }
 481
 482 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
 483 {
 484         struct tcp_sock *tp = tcp_sk(sk);
 485         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 486
 487         skb->csum    = 0;
 488         tcb->seq     = tcb->end_seq = tp->write_seq;
 489         tcb->flags   = TCPCB_FLAG_ACK;
 490         tcb->sacked  = 0;
 491         skb_header_release(skb);
 492         tcp_add_write_queue_tail(sk, skb);
 493         sk->sk_wmem_queued += skb->truesize;
 494         sk_mem_charge(sk, skb->truesize);
 495         if (tp->nonagle & TCP_NAGLE_PUSH)
 496                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 497 }
 498
 499 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
 500                                 struct sk_buff *skb)
 501 {
 502         if (flags & MSG_OOB)
 503                 tp->snd_up = tp->write_seq;
 504 }
 505
 506 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
 507                             int nonagle)
 508 {
 509         struct tcp_sock *tp = tcp_sk(sk);
 510
 511         if (tcp_send_head(sk)) {
 512                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 513                 if (!(flags & MSG_MORE) || forced_push(tp))
 514                         tcp_mark_push(tp, skb);
 515                 tcp_mark_urg(tp, flags, skb);
 516                 __tcp_push_pending_frames(sk, mss_now,
 517                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
 518         }
 519 }
 520
 521 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 522                                 unsigned int offset, size_t len)
 523 {
 524         struct tcp_splice_state *tss = rd_desc->arg.data;
 525         int ret;
 526
 527         ret = skb_splice_bits(skb, offset, tss->pipe, rd_desc->count, tss->flags);
 528         if (ret > 0)
 529                 rd_desc->count -= ret;
 530         return ret;
 531 }
 532
 533 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 534 {
 535         /* Store TCP splice context information in read_descriptor_t. */
 536         read_descriptor_t rd_desc = {
 537                 .arg.data = tss,
 538                 .count    = tss->len,
 539         };
 540
 541         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 542 }
 543
 544 /**
 545  *  tcp_splice_read - splice data from TCP socket to a pipe
 546  * @sock:       socket to splice from
 547  * @ppos:       position (not valid)
 548  * @pipe:       pipe to splice to
 549  * @len:        number of bytes to splice
 550  * @flags:      splice modifier flags
 551  *
 552  * Description:
 553  *    Will read pages from given socket and fill them into a pipe.
 554  *
 555  **/
 556 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 557                         struct pipe_inode_info *pipe, size_t len,
 558                         unsigned int flags)
 559 {
 560         struct sock *sk = sock->sk;
 561         struct tcp_splice_state tss = {
 562                 .pipe = pipe,
 563                 .len = len,
 564                 .flags = flags,
 565         };
 566         long timeo;
 567         ssize_t spliced;
 568         int ret;
 569
 570         /*
 571          * We can't seek on a socket input
 572          */
 573         if (unlikely(*ppos))
 574                 return -ESPIPE;
 575
 576         ret = spliced = 0;
 577
 578         lock_sock(sk);
 579
 580         timeo = sock_rcvtimeo(sk, flags & SPLICE_F_NONBLOCK);
 581         while (tss.len) {
 582                 ret = __tcp_splice_read(sk, &tss);
 583                 if (ret < 0)
 584                         break;
 585                 else if (!ret) {
 586                         if (spliced)
 587                                 break;
 588                         if (sock_flag(sk, SOCK_DONE))
 589                                 break;
 590                         if (sk->sk_err) {
 591                                 ret = sock_error(sk);
 592                                 break;
 593                         }
 594                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 595                                 break;
 596                         if (sk->sk_state == TCP_CLOSE) {
 597                                 /*
 598                                  * This occurs when user tries to read
 599                                  * from never connected socket.
 600                                  */
 601                                 if (!sock_flag(sk, SOCK_DONE))
 602                                         ret = -ENOTCONN;
 603                                 break;
 604                         }
 605                         if (!timeo) {
 606                                 ret = -EAGAIN;
 607                                 break;
 608                         }
 609                         sk_wait_data(sk, &timeo);
 610                         if (signal_pending(current)) {
 611                                 ret = sock_intr_errno(timeo);
 612                                 break;
 613                         }
 614                         continue;
 615                 }
 616                 tss.len -= ret;
 617                 spliced += ret;
 618
 619                 if (!timeo)
 620                         break;
 621                 release_sock(sk);
 622                 lock_sock(sk);
 623
 624                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 625                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 626                     signal_pending(current))
 627                         break;
 628         }
 629
 630         release_sock(sk);
 631
 632         if (spliced)
 633                 return spliced;
 634
 635         return ret;
 636 }
 637
 638 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
 639 {
 640         struct sk_buff *skb;
 641
 642         /* The TCP header must be at least 32-bit aligned.  */
 643         size = ALIGN(size, 4);
 644
 645         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 646         if (skb) {
 647                 if (sk_wmem_schedule(sk, skb->truesize)) {
 648                         /*
 649                          * Make sure that we have exactly size bytes
 650                          * available to the caller, no more, no less.
 651                          */
 652                         skb_reserve(skb, skb_tailroom(skb) - size);
 653                         return skb;
 654                 }
 655                 __kfree_skb(skb);
 656         } else {
 657                 sk->sk_prot->enter_memory_pressure(sk);
 658                 sk_stream_moderate_sndbuf(sk);
 659         }
 660         return NULL;
 661 }
 662
 663 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
 664                          size_t psize, int flags)
 665 {
 666         struct tcp_sock *tp = tcp_sk(sk);
 667         int mss_now, size_goal;
 668         int err;
 669         ssize_t copied;
 670         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 671
 672         /* Wait for a connection to finish. */
 673         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 674                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 675                         goto out_err;
 676
 677         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 678
 679         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 680         size_goal = tp->xmit_size_goal;
 681         copied = 0;
 682
 683         err = -EPIPE;
 684         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 685                 goto do_error;
 686
 687         while (psize > 0) {
 688                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 689                 struct page *page = pages[poffset / PAGE_SIZE];
 690                 int copy, i, can_coalesce;
 691                 int offset = poffset % PAGE_SIZE;
 692                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
 693
 694                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
 695 new_segment:
 696                         if (!sk_stream_memory_free(sk))
 697                                 goto wait_for_sndbuf;
 698
 699                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
 700                         if (!skb)
 701                                 goto wait_for_memory;
 702
 703                         skb_entail(sk, skb);
 704                         copy = size_goal;
 705                 }
 706
 707                 if (copy > size)
 708                         copy = size;
 709
 710                 i = skb_shinfo(skb)->nr_frags;
 711                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 712                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
 713                         tcp_mark_push(tp, skb);
 714                         goto new_segment;
 715                 }
 716                 if (!sk_wmem_schedule(sk, copy))
 717                         goto wait_for_memory;
 718
 719                 if (can_coalesce) {
 720                         skb_shinfo(skb)->frags[i - 1].size += copy;
 721                 } else {
 722                         get_page(page);
 723                         skb_fill_page_desc(skb, i, page, offset, copy);
 724                 }
 725
 726                 skb->len += copy;
 727                 skb->data_len += copy;
 728                 skb->truesize += copy;
 729                 sk->sk_wmem_queued += copy;
 730                 sk_mem_charge(sk, copy);
 731                 skb->ip_summed = CHECKSUM_PARTIAL;
 732                 tp->write_seq += copy;
 733                 TCP_SKB_CB(skb)->end_seq += copy;
 734                 skb_shinfo(skb)->gso_segs = 0;
 735
 736                 if (!copied)
 737                         TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
 738
 739                 copied += copy;
 740                 poffset += copy;
 741                 if (!(psize -= copy))
 742                         goto out;
 743
 744                 if (skb->len < size_goal || (flags & MSG_OOB))
 745                         continue;
 746
 747                 if (forced_push(tp)) {
 748                         tcp_mark_push(tp, skb);
 749                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
 750                 } else if (skb == tcp_send_head(sk))
 751                         tcp_push_one(sk, mss_now);
 752                 continue;
 753
 754 wait_for_sndbuf:
 755                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 756 wait_for_memory:
 757                 if (copied)
 758                         tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
 759
 760                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 761                         goto do_error;
 762
 763                 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 764                 size_goal = tp->xmit_size_goal;
 765         }
 766
 767 out:
 768         if (copied)
 769                 tcp_push(sk, flags, mss_now, tp->nonagle);
 770         return copied;
 771
 772 do_error:
 773         if (copied)
 774                 goto out;
 775 out_err:
 776         return sk_stream_error(sk, flags, err);
 777 }
 778
 779 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
 780                      size_t size, int flags)
 781 {
 782         ssize_t res;
 783         struct sock *sk = sock->sk;
 784
 785         if (!(sk->sk_route_caps & NETIF_F_SG) ||
 786             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
 787                 return sock_no_sendpage(sock, page, offset, size, flags);
 788
 789         lock_sock(sk);
 790         TCP_CHECK_TIMER(sk);
 791         res = do_tcp_sendpages(sk, &page, offset, size, flags);
 792         TCP_CHECK_TIMER(sk);
 793         release_sock(sk);
 794         return res;
 795 }
 796
 797 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
 798 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
 799
 800 static inline int select_size(struct sock *sk)
 801 {
 802         struct tcp_sock *tp = tcp_sk(sk);
 803         int tmp = tp->mss_cache;
 804
 805         if (sk->sk_route_caps & NETIF_F_SG) {
 806                 if (sk_can_gso(sk))
 807                         tmp = 0;
 808                 else {
 809                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
 810
 811                         if (tmp >= pgbreak &&
 812                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
 813                                 tmp = pgbreak;
 814                 }
 815         }
 816
 817         return tmp;
 818 }
 819
 820 int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
 821                 size_t size)
 822 {
 823         struct sock *sk = sock->sk;
 824         struct iovec *iov;
 825         struct tcp_sock *tp = tcp_sk(sk);
 826         struct sk_buff *skb;
 827         int iovlen, flags;
 828         int mss_now, size_goal;
 829         int err, copied;
 830         long timeo;
 831
 832         lock_sock(sk);
 833         TCP_CHECK_TIMER(sk);
 834
 835         flags = msg->msg_flags;
 836         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 837
 838         /* Wait for a connection to finish. */
 839         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 840                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 841                         goto out_err;
 842
 843         /* This should be in poll */
 844         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 845
 846         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
 847         size_goal = tp->xmit_size_goal;
 848
 849         /* Ok commence sending. */
 850         iovlen = msg->msg_iovlen;
 851         iov = msg->msg_iov;
 852         copied = 0;
 853
 854         err = -EPIPE;
 855         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 856                 goto do_error;
 857
 858         while (--iovlen >= 0) {
 859                 int seglen = iov->iov_len;
 860                 unsigned char __user *from = iov->iov_base;
 861
 862                 iov++;
 863
 864                 while (seglen > 0) {
 865                         int copy;
 866
 867                         skb = tcp_write_queue_tail(sk);
 868
 869                         if (!tcp_send_head(sk) ||
 870                             (copy = size_goal - skb->len) <= 0) {
 871
 872 new_segment:
 873                                 /* Allocate new segment. If the interface is SG,
 874                                  * allocate skb fitting to single page.
 875                                  */
 876                                 if (!sk_stream_memory_free(sk))
 877                                         goto wait_for_sndbuf;
 878
 879                                 skb = sk_stream_alloc_skb(sk, select_size(sk),
 880                                                 sk->sk_allocation);
 881                                 if (!skb)
 882                                         goto wait_for_memory;
 883
 884                                 /*
 885                                  * Check whether we can use HW checksum.
 886                                  */
 887                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
 888                                         skb->ip_summed = CHECKSUM_PARTIAL;
 889
 890                                 skb_entail(sk, skb);
 891                                 copy = size_goal;
 892                         }
 893
 894                         /* Try to append data to the end of skb. */
 895                         if (copy > seglen)
 896                                 copy = seglen;
 897
 898                         /* Where to copy to? */
 899                         if (skb_tailroom(skb) > 0) {
 900                                 /* We have some space in skb head. Superb! */
 901                                 if (copy > skb_tailroom(skb))
 902                                         copy = skb_tailroom(skb);
 903                                 if ((err = skb_add_data(skb, from, copy)) != 0)
 904                                         goto do_fault;
 905                         } else {
 906                                 int merge = 0;
 907                                 int i = skb_shinfo(skb)->nr_frags;
 908                                 struct page *page = TCP_PAGE(sk);
 909                                 int off = TCP_OFF(sk);
 910
 911                                 if (skb_can_coalesce(skb, i, page, off) &&
 912                                     off != PAGE_SIZE) {
 913                                         /* We can extend the last page
 914                                          * fragment. */
 915                                         merge = 1;
 916                                 } else if (i == MAX_SKB_FRAGS ||
 917                                            (!i &&
 918                                            !(sk->sk_route_caps & NETIF_F_SG))) {
 919                                         /* Need to add new fragment and cannot
 920                                          * do this because interface is non-SG,
 921                                          * or because all the page slots are
 922                                          * busy. */
 923                                         tcp_mark_push(tp, skb);
 924                                         goto new_segment;
 925                                 } else if (page) {
 926                                         if (off == PAGE_SIZE) {
 927                                                 put_page(page);
 928                                                 TCP_PAGE(sk) = page = NULL;
 929                                                 off = 0;
 930                                         }
 931                                 } else
 932                                         off = 0;
 933
 934                                 if (copy > PAGE_SIZE - off)
 935                                         copy = PAGE_SIZE - off;
 936
 937                                 if (!sk_wmem_schedule(sk, copy))
 938                                         goto wait_for_memory;
 939
 940                                 if (!page) {
 941                                         /* Allocate new cache page. */
 942                                         if (!(page = sk_stream_alloc_page(sk)))
 943                                                 goto wait_for_memory;
 944                                 }
 945
 946                                 /* Time to copy data. We are close to
 947                                  * the end! */
 948                                 err = skb_copy_to_page(sk, from, skb, page,
 949                                                        off, copy);
 950                                 if (err) {
 951                                         /* If this page was new, give it to the
 952                                          * socket so it does not get leaked.
 953                                          */
 954                                         if (!TCP_PAGE(sk)) {
 955                                                 TCP_PAGE(sk) = page;
 956                                                 TCP_OFF(sk) = 0;
 957                                         }
 958                                         goto do_error;
 959                                 }
 960
 961                                 /* Update the skb. */
 962                                 if (merge) {
 963                                         skb_shinfo(skb)->frags[i - 1].size +=
 964                                                                         copy;
 965                                 } else {
 966                                         skb_fill_page_desc(skb, i, page, off, copy);
 967                                         if (TCP_PAGE(sk)) {
 968                                                 get_page(page);
 969                                         } else if (off + copy < PAGE_SIZE) {
 970                                                 get_page(page);
 971                                                 TCP_PAGE(sk) = page;
 972                                         }
 973                                 }
 974
 975                                 TCP_OFF(sk) = off + copy;
 976                         }
 977
 978                         if (!copied)
 979                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
 980
 981                         tp->write_seq += copy;
 982                         TCP_SKB_CB(skb)->end_seq += copy;
 983                         skb_shinfo(skb)->gso_segs = 0;
 984
 985                         from += copy;
 986                         copied += copy;
 987                         if ((seglen -= copy) == 0 && iovlen == 0)
 988                                 goto out;
 989
 990                         if (skb->len < size_goal || (flags & MSG_OOB))
 991                                 continue;
 992
 993                         if (forced_push(tp)) {
 994                                 tcp_mark_push(tp, skb);
 995                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
 996                         } else if (skb == tcp_send_head(sk))
 997                                 tcp_push_one(sk, mss_now);
 998                         continue;
 999
1000 wait_for_sndbuf:
1001                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1002 wait_for_memory:
1003                         if (copied)
1004                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1005
1006                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1007                                 goto do_error;
1008
1009                         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
1010                         size_goal = tp->xmit_size_goal;
1011                 }
1012         }
1013
1014 out:
1015         if (copied)
1016                 tcp_push(sk, flags, mss_now, tp->nonagle);
1017         TCP_CHECK_TIMER(sk);
1018         release_sock(sk);
1019         return copied;
1020
1021 do_fault:
1022         if (!skb->len) {
1023                 tcp_unlink_write_queue(skb, sk);
1024                 /* It is the one place in all of TCP, except connection
1025                  * reset, where we can be unlinking the send_head.
1026                  */
1027                 tcp_check_send_head(sk, skb);
1028                 sk_wmem_free_skb(sk, skb);
1029         }
1030
1031 do_error:
1032         if (copied)
1033                 goto out;
1034 out_err:
1035         err = sk_stream_error(sk, flags, err);
1036         TCP_CHECK_TIMER(sk);
1037         release_sock(sk);
1038         return err;
1039 }
1040
1041 /*
1042  *      Handle reading urgent data. BSD has very simple semantics for
1043  *      this, no blocking and very strange errors 8)
1044  */
1045
1046 static int tcp_recv_urg(struct sock *sk, long timeo,
1047                         struct msghdr *msg, int len, int flags,
1048                         int *addr_len)
1049 {
1050         struct tcp_sock *tp = tcp_sk(sk);
1051
1052         /* No URG data to read. */
1053         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1054             tp->urg_data == TCP_URG_READ)
1055                 return -EINVAL; /* Yes this is right ! */
1056
1057         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1058                 return -ENOTCONN;
1059
1060         if (tp->urg_data & TCP_URG_VALID) {
1061                 int err = 0;
1062                 char c = tp->urg_data;
1063
1064                 if (!(flags & MSG_PEEK))
1065                         tp->urg_data = TCP_URG_READ;
1066
1067                 /* Read urgent data. */
1068                 msg->msg_flags |= MSG_OOB;
1069
1070                 if (len > 0) {
1071                         if (!(flags & MSG_TRUNC))
1072                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1073                         len = 1;
1074                 } else
1075                         msg->msg_flags |= MSG_TRUNC;
1076
1077                 return err ? -EFAULT : len;
1078         }
1079
1080         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1081                 return 0;
1082
1083         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1084          * the available implementations agree in this case:
1085          * this call should never block, independent of the
1086          * blocking state of the socket.
1087          * Mike <pall@rz.uni-karlsruhe.de>
1088          */
1089         return -EAGAIN;
1090 }
1091
1092 /* Clean up the receive buffer for full frames taken by the user,
1093  * then send an ACK if necessary.  COPIED is the number of bytes
1094  * tcp_recvmsg has given to the user so far, it speeds up the
1095  * calculation of whether or not we must ACK for the sake of
1096  * a window update.
1097  */
1098 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1099 {
1100         struct tcp_sock *tp = tcp_sk(sk);
1101         int time_to_ack = 0;
1102
1103 #if TCP_DEBUG
1104         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1105
1106         WARN_ON(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
1107 #endif
1108
1109         if (inet_csk_ack_scheduled(sk)) {
1110                 const struct inet_connection_sock *icsk = inet_csk(sk);
1111                    /* Delayed ACKs frequently hit locked sockets during bulk
1112                     * receive. */
1113                 if (icsk->icsk_ack.blocked ||
1114                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1115                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1116                     /*
1117                      * If this read emptied read buffer, we send ACK, if
1118                      * connection is not bidirectional, user drained
1119                      * receive buffer and there was a small segment
1120                      * in queue.
1121                      */
1122                     (copied > 0 &&
1123                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1124                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1125                        !icsk->icsk_ack.pingpong)) &&
1126                       !atomic_read(&sk->sk_rmem_alloc)))
1127                         time_to_ack = 1;
1128         }
1129
1130         /* We send an ACK if we can now advertise a non-zero window
1131          * which has been raised "significantly".
1132          *
1133          * Even if window raised up to infinity, do not send window open ACK
1134          * in states, where we will not receive more. It is useless.
1135          */
1136         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1137                 __u32 rcv_window_now = tcp_receive_window(tp);
1138
1139                 /* Optimize, __tcp_select_window() is not cheap. */
1140                 if (2*rcv_window_now <= tp->window_clamp) {
1141                         __u32 new_window = __tcp_select_window(sk);
1142
1143                         /* Send ACK now, if this read freed lots of space
1144                          * in our buffer. Certainly, new_window is new window.
1145                          * We can advertise it now, if it is not less than current one.
1146                          * "Lots" means "at least twice" here.
1147                          */
1148                         if (new_window && new_window >= 2 * rcv_window_now)
1149                                 time_to_ack = 1;
1150                 }
1151         }
1152         if (time_to_ack)
1153                 tcp_send_ack(sk);
1154 }
1155
1156 static void tcp_prequeue_process(struct sock *sk)
1157 {
1158         struct sk_buff *skb;
1159         struct tcp_sock *tp = tcp_sk(sk);
1160
1161         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1162
1163         /* RX process wants to run with disabled BHs, though it is not
1164          * necessary */
1165         local_bh_disable();
1166         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1167                 sk_backlog_rcv(sk, skb);
1168         local_bh_enable();
1169
1170         /* Clear memory counter. */
1171         tp->ucopy.memory = 0;
1172 }
1173
1174 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1175 {
1176         struct sk_buff *skb;
1177         u32 offset;
1178
1179         skb_queue_walk(&sk->sk_receive_queue, skb) {
1180                 offset = seq - TCP_SKB_CB(skb)->seq;
1181                 if (tcp_hdr(skb)->syn)
1182                         offset--;
1183                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1184                         *off = offset;
1185                         return skb;
1186                 }
1187         }
1188         return NULL;
1189 }
1190
1191 /*
1192  * This routine provides an alternative to tcp_recvmsg() for routines
1193  * that would like to handle copying from skbuffs directly in 'sendfile'
1194  * fashion.
1195  * Note:
1196  *      - It is assumed that the socket was locked by the caller.
1197  *      - The routine does not block.
1198  *      - At present, there is no support for reading OOB data
1199  *        or for 'peeking' the socket using this routine
1200  *        (although both would be easy to implement).
1201  */
1202 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1203                   sk_read_actor_t recv_actor)
1204 {
1205         struct sk_buff *skb;
1206         struct tcp_sock *tp = tcp_sk(sk);
1207         u32 seq = tp->copied_seq;
1208         u32 offset;
1209         int copied = 0;
1210
1211         if (sk->sk_state == TCP_LISTEN)
1212                 return -ENOTCONN;
1213         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1214                 if (offset < skb->len) {
1215                         int used;
1216                         size_t len;
1217
1218                         len = skb->len - offset;
1219                         /* Stop reading if we hit a patch of urgent data */
1220                         if (tp->urg_data) {
1221                                 u32 urg_offset = tp->urg_seq - seq;
1222                                 if (urg_offset < len)
1223                                         len = urg_offset;
1224                                 if (!len)
1225                                         break;
1226                         }
1227                         used = recv_actor(desc, skb, offset, len);
1228                         if (used < 0) {
1229                                 if (!copied)
1230                                         copied = used;
1231                                 break;
1232                         } else if (used <= len) {
1233                                 seq += used;
1234                                 copied += used;
1235                                 offset += used;
1236                         }
1237                         /*
1238                          * If recv_actor drops the lock (e.g. TCP splice
1239                          * receive) the skb pointer might be invalid when
1240                          * getting here: tcp_collapse might have deleted it
1241                          * while aggregating skbs from the socket queue.
1242                          */
1243                         skb = tcp_recv_skb(sk, seq-1, &offset);
1244                         if (!skb || (offset+1 != skb->len))
1245                                 break;
1246                 }
1247                 if (tcp_hdr(skb)->fin) {
1248                         sk_eat_skb(sk, skb, 0);
1249                         ++seq;
1250                         break;
1251                 }
1252                 sk_eat_skb(sk, skb, 0);
1253                 if (!desc->count)
1254                         break;
1255         }
1256         tp->copied_seq = seq;
1257
1258         tcp_rcv_space_adjust(sk);
1259
1260         /* Clean up data we have read: This will do ACK frames. */
1261         if (copied > 0)
1262                 tcp_cleanup_rbuf(sk, copied);
1263         return copied;
1264 }
1265
1266 /*
1267  *      This routine copies from a sock struct into the user buffer.
1268  *
1269  *      Technical note: in 2.3 we work on _locked_ socket, so that
1270  *      tricks with *seq access order and skb->users are not required.
1271  *      Probably, code can be easily improved even more.
1272  */
1273
1274 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1275                 size_t len, int nonblock, int flags, int *addr_len)
1276 {
1277         struct tcp_sock *tp = tcp_sk(sk);
1278         int copied = 0;
1279         u32 peek_seq;
1280         u32 *seq;
1281         unsigned long used;
1282         int err;
1283         int target;             /* Read at least this many bytes */
1284         long timeo;
1285         struct task_struct *user_recv = NULL;
1286         int copied_early = 0;
1287         struct sk_buff *skb;
1288
1289         lock_sock(sk);
1290
1291         TCP_CHECK_TIMER(sk);
1292
1293         err = -ENOTCONN;
1294         if (sk->sk_state == TCP_LISTEN)
1295                 goto out;
1296
1297         timeo = sock_rcvtimeo(sk, nonblock);
1298
1299         /* Urgent data needs to be handled specially. */
1300         if (flags & MSG_OOB)
1301                 goto recv_urg;
1302
1303         seq = &tp->copied_seq;
1304         if (flags & MSG_PEEK) {
1305                 peek_seq = tp->copied_seq;
1306                 seq = &peek_seq;
1307         }
1308
1309         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1310
1311 #ifdef CONFIG_NET_DMA
1312         tp->ucopy.dma_chan = NULL;
1313         preempt_disable();
1314         skb = skb_peek_tail(&sk->sk_receive_queue);
1315         {
1316                 int available = 0;
1317
1318                 if (skb)
1319                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1320                 if ((available < target) &&
1321                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1322                     !sysctl_tcp_low_latency &&
1323                     dma_find_channel(DMA_MEMCPY)) {
1324                         preempt_enable_no_resched();
1325                         tp->ucopy.pinned_list =
1326                                         dma_pin_iovec_pages(msg->msg_iov, len);
1327                 } else {
1328                         preempt_enable_no_resched();
1329                 }
1330         }
1331 #endif
1332
1333         do {
1334                 u32 offset;
1335
1336                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1337                 if (tp->urg_data && tp->urg_seq == *seq) {
1338                         if (copied)
1339                                 break;
1340                         if (signal_pending(current)) {
1341                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1342                                 break;
1343                         }
1344                 }
1345
1346                 /* Next get a buffer. */
1347
1348                 skb = skb_peek(&sk->sk_receive_queue);
1349                 do {
1350                         if (!skb)
1351                                 break;
1352
1353                         /* Now that we have two receive queues this
1354                          * shouldn't happen.
1355                          */
1356                         if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1357                                 printk(KERN_INFO "recvmsg bug: copied %X "
1358                                        "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1359                                 break;
1360                         }
1361                         offset = *seq - TCP_SKB_CB(skb)->seq;
1362                         if (tcp_hdr(skb)->syn)
1363                                 offset--;
1364                         if (offset < skb->len)
1365                                 goto found_ok_skb;
1366                         if (tcp_hdr(skb)->fin)
1367                                 goto found_fin_ok;
1368                         WARN_ON(!(flags & MSG_PEEK));
1369                         skb = skb->next;
1370                 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1371
1372                 /* Well, if we have backlog, try to process it now yet. */
1373
1374                 if (copied >= target && !sk->sk_backlog.tail)
1375                         break;
1376
1377                 if (copied) {
1378                         if (sk->sk_err ||
1379                             sk->sk_state == TCP_CLOSE ||
1380                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1381                             !timeo ||
1382                             signal_pending(current))
1383                                 break;
1384                 } else {
1385                         if (sock_flag(sk, SOCK_DONE))
1386                                 break;
1387
1388                         if (sk->sk_err) {
1389                                 copied = sock_error(sk);
1390                                 break;
1391                         }
1392
1393                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1394                                 break;
1395
1396                         if (sk->sk_state == TCP_CLOSE) {
1397                                 if (!sock_flag(sk, SOCK_DONE)) {
1398                                         /* This occurs when user tries to read
1399                                          * from never connected socket.
1400                                          */
1401                                         copied = -ENOTCONN;
1402                                         break;
1403                                 }
1404                                 break;
1405                         }
1406
1407                         if (!timeo) {
1408                                 copied = -EAGAIN;
1409                                 break;
1410                         }
1411
1412                         if (signal_pending(current)) {
1413                                 copied = sock_intr_errno(timeo);
1414                                 break;
1415                         }
1416                 }
1417
1418                 tcp_cleanup_rbuf(sk, copied);
1419
1420                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1421                         /* Install new reader */
1422                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1423                                 user_recv = current;
1424                                 tp->ucopy.task = user_recv;
1425                                 tp->ucopy.iov = msg->msg_iov;
1426                         }
1427
1428                         tp->ucopy.len = len;
1429
1430                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1431                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1432
1433                         /* Ugly... If prequeue is not empty, we have to
1434                          * process it before releasing socket, otherwise
1435                          * order will be broken at second iteration.
1436                          * More elegant solution is required!!!
1437                          *
1438                          * Look: we have the following (pseudo)queues:
1439                          *
1440                          * 1. packets in flight
1441                          * 2. backlog
1442                          * 3. prequeue
1443                          * 4. receive_queue
1444                          *
1445                          * Each queue can be processed only if the next ones
1446                          * are empty. At this point we have empty receive_queue.
1447                          * But prequeue _can_ be not empty after 2nd iteration,
1448                          * when we jumped to start of loop because backlog
1449                          * processing added something to receive_queue.
1450                          * We cannot release_sock(), because backlog contains
1451                          * packets arrived _after_ prequeued ones.
1452                          *
1453                          * Shortly, algorithm is clear --- to process all
1454                          * the queues in order. We could make it more directly,
1455                          * requeueing packets from backlog to prequeue, if
1456                          * is not empty. It is more elegant, but eats cycles,
1457                          * unfortunately.
1458                          */
1459                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1460                                 goto do_prequeue;
1461
1462                         /* __ Set realtime policy in scheduler __ */
1463                 }
1464
1465                 if (copied >= target) {
1466                         /* Do not sleep, just process backlog. */
1467                         release_sock(sk);
1468                         lock_sock(sk);
1469                 } else
1470                         sk_wait_data(sk, &timeo);
1471
1472 #ifdef CONFIG_NET_DMA
1473                 tp->ucopy.wakeup = 0;
1474 #endif
1475
1476                 if (user_recv) {
1477                         int chunk;
1478
1479                         /* __ Restore normal policy in scheduler __ */
1480
1481                         if ((chunk = len - tp->ucopy.len) != 0) {
1482                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1483                                 len -= chunk;
1484                                 copied += chunk;
1485                         }
1486
1487                         if (tp->rcv_nxt == tp->copied_seq &&
1488                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1489 do_prequeue:
1490                                 tcp_prequeue_process(sk);
1491
1492                                 if ((chunk = len - tp->ucopy.len) != 0) {
1493                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1494                                         len -= chunk;
1495                                         copied += chunk;
1496                                 }
1497                         }
1498                 }
1499                 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1500                         if (net_ratelimit())
1501                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1502                                        current->comm, task_pid_nr(current));
1503                         peek_seq = tp->copied_seq;
1504                 }
1505                 continue;
1506
1507         found_ok_skb:
1508                 /* Ok so how much can we use? */
1509                 used = skb->len - offset;
1510                 if (len < used)
1511                         used = len;
1512
1513                 /* Do we have urgent data here? */
1514                 if (tp->urg_data) {
1515                         u32 urg_offset = tp->urg_seq - *seq;
1516                         if (urg_offset < used) {
1517                                 if (!urg_offset) {
1518                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1519                                                 ++*seq;
1520                                                 offset++;
1521                                                 used--;
1522                                                 if (!used)
1523                                                         goto skip_copy;
1524                                         }
1525                                 } else
1526                                         used = urg_offset;
1527                         }
1528                 }
1529
1530                 if (!(flags & MSG_TRUNC)) {
1531 #ifdef CONFIG_NET_DMA
1532                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1533                                 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1534
1535                         if (tp->ucopy.dma_chan) {
1536                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1537                                         tp->ucopy.dma_chan, skb, offset,
1538                                         msg->msg_iov, used,
1539                                         tp->ucopy.pinned_list);
1540
1541                                 if (tp->ucopy.dma_cookie < 0) {
1542
1543                                         printk(KERN_ALERT "dma_cookie < 0\n");
1544
1545                                         /* Exception. Bailout! */
1546                                         if (!copied)
1547                                                 copied = -EFAULT;
1548                                         break;
1549                                 }
1550                                 if ((offset + used) == skb->len)
1551                                         copied_early = 1;
1552
1553                         } else
1554 #endif
1555                         {
1556                                 err = skb_copy_datagram_iovec(skb, offset,
1557                                                 msg->msg_iov, used);
1558                                 if (err) {
1559                                         /* Exception. Bailout! */
1560                                         if (!copied)
1561                                                 copied = -EFAULT;
1562                                         break;
1563                                 }
1564                         }
1565                 }
1566
1567                 *seq += used;
1568                 copied += used;
1569                 len -= used;
1570
1571                 tcp_rcv_space_adjust(sk);
1572
1573 skip_copy:
1574                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1575                         tp->urg_data = 0;
1576                         tcp_fast_path_check(sk);
1577                 }
1578                 if (used + offset < skb->len)
1579                         continue;
1580
1581                 if (tcp_hdr(skb)->fin)
1582                         goto found_fin_ok;
1583                 if (!(flags & MSG_PEEK)) {
1584                         sk_eat_skb(sk, skb, copied_early);
1585                         copied_early = 0;
1586                 }
1587                 continue;
1588
1589         found_fin_ok:
1590                 /* Process the FIN. */
1591                 ++*seq;
1592                 if (!(flags & MSG_PEEK)) {
1593                         sk_eat_skb(sk, skb, copied_early);
1594                         copied_early = 0;
1595                 }
1596                 break;
1597         } while (len > 0);
1598
1599         if (user_recv) {
1600                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1601                         int chunk;
1602
1603                         tp->ucopy.len = copied > 0 ? len : 0;
1604
1605                         tcp_prequeue_process(sk);
1606
1607                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1608                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1609                                 len -= chunk;
1610                                 copied += chunk;
1611                         }
1612                 }
1613
1614                 tp->ucopy.task = NULL;
1615                 tp->ucopy.len = 0;
1616         }
1617
1618 #ifdef CONFIG_NET_DMA
1619         if (tp->ucopy.dma_chan) {
1620                 dma_cookie_t done, used;
1621
1622                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1623
1624                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1625                                                  tp->ucopy.dma_cookie, &done,
1626                                                  &used) == DMA_IN_PROGRESS) {
1627                         /* do partial cleanup of sk_async_wait_queue */
1628                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1629                                (dma_async_is_complete(skb->dma_cookie, done,
1630                                                       used) == DMA_SUCCESS)) {
1631                                 __skb_dequeue(&sk->sk_async_wait_queue);
1632                                 kfree_skb(skb);
1633                         }
1634                 }
1635
1636                 /* Safe to free early-copied skbs now */
1637                 __skb_queue_purge(&sk->sk_async_wait_queue);
1638                 tp->ucopy.dma_chan = NULL;
1639         }
1640         if (tp->ucopy.pinned_list) {
1641                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1642                 tp->ucopy.pinned_list = NULL;
1643         }
1644 #endif
1645
1646         /* According to UNIX98, msg_name/msg_namelen are ignored
1647          * on connected socket. I was just happy when found this 8) --ANK
1648          */
1649
1650         /* Clean up data we have read: This will do ACK frames. */
1651         tcp_cleanup_rbuf(sk, copied);
1652
1653         TCP_CHECK_TIMER(sk);
1654         release_sock(sk);
1655         return copied;
1656
1657 out:
1658         TCP_CHECK_TIMER(sk);
1659         release_sock(sk);
1660         return err;
1661
1662 recv_urg:
1663         err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1664         goto out;
1665 }
1666
1667 void tcp_set_state(struct sock *sk, int state)
1668 {
1669         int oldstate = sk->sk_state;
1670
1671         switch (state) {
1672         case TCP_ESTABLISHED:
1673                 if (oldstate != TCP_ESTABLISHED)
1674                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1675                 break;
1676
1677         case TCP_CLOSE:
1678                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1679                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1680
1681                 sk->sk_prot->unhash(sk);
1682                 if (inet_csk(sk)->icsk_bind_hash &&
1683                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1684                         inet_put_port(sk);
1685                 /* fall through */
1686         default:
1687                 if (oldstate == TCP_ESTABLISHED)
1688                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1689         }
1690
1691         /* Change state AFTER socket is unhashed to avoid closed
1692          * socket sitting in hash tables.
1693          */
1694         sk->sk_state = state;
1695
1696 #ifdef STATE_TRACE
1697         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1698 #endif
1699 }
1700 EXPORT_SYMBOL_GPL(tcp_set_state);
1701
1702 /*
1703  *      State processing on a close. This implements the state shift for
1704  *      sending our FIN frame. Note that we only send a FIN for some
1705  *      states. A shutdown() may have already sent the FIN, or we may be
1706  *      closed.
1707  */
1708
1709 static const unsigned char new_state[16] = {
1710   /* current state:        new state:      action:      */
1711   /* (Invalid)          */ TCP_CLOSE,
1712   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1713   /* TCP_SYN_SENT       */ TCP_CLOSE,
1714   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1715   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1716   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1717   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1718   /* TCP_CLOSE          */ TCP_CLOSE,
1719   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1720   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1721   /* TCP_LISTEN         */ TCP_CLOSE,
1722   /* TCP_CLOSING        */ TCP_CLOSING,
1723 };
1724
1725 static int tcp_close_state(struct sock *sk)
1726 {
1727         int next = (int)new_state[sk->sk_state];
1728         int ns = next & TCP_STATE_MASK;
1729
1730         tcp_set_state(sk, ns);
1731
1732         return next & TCP_ACTION_FIN;
1733 }
1734
1735 /*
1736  *      Shutdown the sending side of a connection. Much like close except
1737  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1738  */
1739
1740 void tcp_shutdown(struct sock *sk, int how)
1741 {
1742         /*      We need to grab some memory, and put together a FIN,
1743          *      and then put it into the queue to be sent.
1744          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1745          */
1746         if (!(how & SEND_SHUTDOWN))
1747                 return;
1748
1749         /* If we've already sent a FIN, or it's a closed state, skip this. */
1750         if ((1 << sk->sk_state) &
1751             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1752              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1753                 /* Clear out any half completed packets.  FIN if needed. */
1754                 if (tcp_close_state(sk))
1755                         tcp_send_fin(sk);
1756         }
1757 }
1758
1759 void tcp_close(struct sock *sk, long timeout)
1760 {
1761         struct sk_buff *skb;
1762         int data_was_unread = 0;
1763         int state;
1764
1765         lock_sock(sk);
1766         sk->sk_shutdown = SHUTDOWN_MASK;
1767
1768         if (sk->sk_state == TCP_LISTEN) {
1769                 tcp_set_state(sk, TCP_CLOSE);
1770
1771                 /* Special case. */
1772                 inet_csk_listen_stop(sk);
1773
1774                 goto adjudge_to_death;
1775         }
1776
1777         /*  We need to flush the recv. buffs.  We do this only on the
1778          *  descriptor close, not protocol-sourced closes, because the
1779          *  reader process may not have drained the data yet!
1780          */
1781         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1782                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1783                           tcp_hdr(skb)->fin;
1784                 data_was_unread += len;
1785                 __kfree_skb(skb);
1786         }
1787
1788         sk_mem_reclaim(sk);
1789
1790         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1791          * data was lost. To witness the awful effects of the old behavior of
1792          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1793          * GET in an FTP client, suspend the process, wait for the client to
1794          * advertise a zero window, then kill -9 the FTP client, wheee...
1795          * Note: timeout is always zero in such a case.
1796          */
1797         if (data_was_unread) {
1798                 /* Unread data was tossed, zap the connection. */
1799                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1800                 tcp_set_state(sk, TCP_CLOSE);
1801                 tcp_send_active_reset(sk, GFP_KERNEL);
1802         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1803                 /* Check zero linger _after_ checking for unread data. */
1804                 sk->sk_prot->disconnect(sk, 0);
1805                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1806         } else if (tcp_close_state(sk)) {
1807                 /* We FIN if the application ate all the data before
1808                  * zapping the connection.
1809                  */
1810
1811                 /* RED-PEN. Formally speaking, we have broken TCP state
1812                  * machine. State transitions:
1813                  *
1814                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1815                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1816                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1817                  *
1818                  * are legal only when FIN has been sent (i.e. in window),
1819                  * rather than queued out of window. Purists blame.
1820                  *
1821                  * F.e. "RFC state" is ESTABLISHED,
1822                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1823                  *
1824                  * The visible declinations are that sometimes
1825                  * we enter time-wait state, when it is not required really
1826                  * (harmless), do not send active resets, when they are
1827                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1828                  * they look as CLOSING or LAST_ACK for Linux)
1829                  * Probably, I missed some more holelets.
1830                  *                                              --ANK
1831                  */
1832                 tcp_send_fin(sk);
1833         }
1834
1835         sk_stream_wait_close(sk, timeout);
1836
1837 adjudge_to_death:
1838         state = sk->sk_state;
1839         sock_hold(sk);
1840         sock_orphan(sk);
1841
1842         /* It is the last release_sock in its life. It will remove backlog. */
1843         release_sock(sk);
1844
1845
1846         /* Now socket is owned by kernel and we acquire BH lock
1847            to finish close. No need to check for user refs.
1848          */
1849         local_bh_disable();
1850         bh_lock_sock(sk);
1851         WARN_ON(sock_owned_by_user(sk));
1852
1853         percpu_counter_inc(sk->sk_prot->orphan_count);
1854
1855         /* Have we already been destroyed by a softirq or backlog? */
1856         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1857                 goto out;
1858
1859         /*      This is a (useful) BSD violating of the RFC. There is a
1860          *      problem with TCP as specified in that the other end could
1861          *      keep a socket open forever with no application left this end.
1862          *      We use a 3 minute timeout (about the same as BSD) then kill
1863          *      our end. If they send after that then tough - BUT: long enough
1864          *      that we won't make the old 4*rto = almost no time - whoops
1865          *      reset mistake.
1866          *
1867          *      Nope, it was not mistake. It is really desired behaviour
1868          *      f.e. on http servers, when such sockets are useless, but
1869          *      consume significant resources. Let's do it with special
1870          *      linger2 option.                                 --ANK
1871          */
1872
1873         if (sk->sk_state == TCP_FIN_WAIT2) {
1874                 struct tcp_sock *tp = tcp_sk(sk);
1875                 if (tp->linger2 < 0) {
1876                         tcp_set_state(sk, TCP_CLOSE);
1877                         tcp_send_active_reset(sk, GFP_ATOMIC);
1878                         NET_INC_STATS_BH(sock_net(sk),
1879                                         LINUX_MIB_TCPABORTONLINGER);
1880                 } else {
1881                         const int tmo = tcp_fin_time(sk);
1882
1883                         if (tmo > TCP_TIMEWAIT_LEN) {
1884                                 inet_csk_reset_keepalive_timer(sk,
1885                                                 tmo - TCP_TIMEWAIT_LEN);
1886                         } else {
1887                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1888                                 goto out;
1889                         }
1890                 }
1891         }
1892         if (sk->sk_state != TCP_CLOSE) {
1893                 int orphan_count = percpu_counter_read_positive(
1894                                                 sk->sk_prot->orphan_count);
1895
1896                 sk_mem_reclaim(sk);
1897                 if (tcp_too_many_orphans(sk, orphan_count)) {
1898                         if (net_ratelimit())
1899                                 printk(KERN_INFO "TCP: too many of orphaned "
1900                                        "sockets\n");
1901                         tcp_set_state(sk, TCP_CLOSE);
1902                         tcp_send_active_reset(sk, GFP_ATOMIC);
1903                         NET_INC_STATS_BH(sock_net(sk),
1904                                         LINUX_MIB_TCPABORTONMEMORY);
1905                 }
1906         }
1907
1908         if (sk->sk_state == TCP_CLOSE)
1909                 inet_csk_destroy_sock(sk);
1910         /* Otherwise, socket is reprieved until protocol close. */
1911
1912 out:
1913         bh_unlock_sock(sk);
1914         local_bh_enable();
1915         sock_put(sk);
1916 }
1917
1918 /* These states need RST on ABORT according to RFC793 */
1919
1920 static inline int tcp_need_reset(int state)
1921 {
1922         return (1 << state) &
1923                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1924                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1925 }
1926
1927 int tcp_disconnect(struct sock *sk, int flags)
1928 {
1929         struct inet_sock *inet = inet_sk(sk);
1930         struct inet_connection_sock *icsk = inet_csk(sk);
1931         struct tcp_sock *tp = tcp_sk(sk);
1932         int err = 0;
1933         int old_state = sk->sk_state;
1934
1935         if (old_state != TCP_CLOSE)
1936                 tcp_set_state(sk, TCP_CLOSE);
1937
1938         /* ABORT function of RFC793 */
1939         if (old_state == TCP_LISTEN) {
1940                 inet_csk_listen_stop(sk);
1941         } else if (tcp_need_reset(old_state) ||
1942                    (tp->snd_nxt != tp->write_seq &&
1943                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1944                 /* The last check adjusts for discrepancy of Linux wrt. RFC
1945                  * states
1946                  */
1947                 tcp_send_active_reset(sk, gfp_any());
1948                 sk->sk_err = ECONNRESET;
1949         } else if (old_state == TCP_SYN_SENT)
1950                 sk->sk_err = ECONNRESET;
1951
1952         tcp_clear_xmit_timers(sk);
1953         __skb_queue_purge(&sk->sk_receive_queue);
1954         tcp_write_queue_purge(sk);
1955         __skb_queue_purge(&tp->out_of_order_queue);
1956 #ifdef CONFIG_NET_DMA
1957         __skb_queue_purge(&sk->sk_async_wait_queue);
1958 #endif
1959
1960         inet->dport = 0;
1961
1962         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1963                 inet_reset_saddr(sk);
1964
1965         sk->sk_shutdown = 0;
1966         sock_reset_flag(sk, SOCK_DONE);
1967         tp->srtt = 0;
1968         if ((tp->write_seq += tp->max_window + 2) == 0)
1969                 tp->write_seq = 1;
1970         icsk->icsk_backoff = 0;
1971         tp->snd_cwnd = 2;
1972         icsk->icsk_probes_out = 0;
1973         tp->packets_out = 0;
1974         tp->snd_ssthresh = 0x7fffffff;
1975         tp->snd_cwnd_cnt = 0;
1976         tp->bytes_acked = 0;
1977         tcp_set_ca_state(sk, TCP_CA_Open);
1978         tcp_clear_retrans(tp);
1979         inet_csk_delack_init(sk);
1980         tcp_init_send_head(sk);
1981         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
1982         __sk_dst_reset(sk);
1983
1984         WARN_ON(inet->num && !icsk->icsk_bind_hash);
1985
1986         sk->sk_error_report(sk);
1987         return err;
1988 }
1989
1990 /*
1991  *      Socket option code for TCP.
1992  */
1993 static int do_tcp_setsockopt(struct sock *sk, int level,
1994                 int optname, char __user *optval, int optlen)
1995 {
1996         struct tcp_sock *tp = tcp_sk(sk);
1997         struct inet_connection_sock *icsk = inet_csk(sk);
1998         int val;
1999         int err = 0;
2000
2001         /* This is a string value all the others are int's */
2002         if (optname == TCP_CONGESTION) {
2003                 char name[TCP_CA_NAME_MAX];
2004
2005                 if (optlen < 1)
2006                         return -EINVAL;
2007
2008                 val = strncpy_from_user(name, optval,
2009                                         min(TCP_CA_NAME_MAX-1, optlen));
2010                 if (val < 0)
2011                         return -EFAULT;
2012                 name[val] = 0;
2013
2014                 lock_sock(sk);
2015                 err = tcp_set_congestion_control(sk, name);
2016                 release_sock(sk);
2017                 return err;
2018         }
2019
2020         if (optlen < sizeof(int))
2021                 return -EINVAL;
2022
2023         if (get_user(val, (int __user *)optval))
2024                 return -EFAULT;
2025
2026         lock_sock(sk);
2027
2028         switch (optname) {
2029         case TCP_MAXSEG:
2030                 /* Values greater than interface MTU won't take effect. However
2031                  * at the point when this call is done we typically don't yet
2032                  * know which interface is going to be used */
2033                 if (val < 8 || val > MAX_TCP_WINDOW) {
2034                         err = -EINVAL;
2035                         break;
2036                 }
2037                 tp->rx_opt.user_mss = val;
2038                 break;
2039
2040         case TCP_NODELAY:
2041                 if (val) {
2042                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2043                          * this option on corked socket is remembered, but
2044                          * it is not activated until cork is cleared.
2045                          *
2046                          * However, when TCP_NODELAY is set we make
2047                          * an explicit push, which overrides even TCP_CORK
2048                          * for currently queued segments.
2049                          */
2050                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2051                         tcp_push_pending_frames(sk);
2052                 } else {
2053                         tp->nonagle &= ~TCP_NAGLE_OFF;
2054                 }
2055                 break;
2056
2057         case TCP_CORK:
2058                 /* When set indicates to always queue non-full frames.
2059                  * Later the user clears this option and we transmit
2060                  * any pending partial frames in the queue.  This is
2061                  * meant to be used alongside sendfile() to get properly
2062                  * filled frames when the user (for example) must write
2063                  * out headers with a write() call first and then use
2064                  * sendfile to send out the data parts.
2065                  *
2066                  * TCP_CORK can be set together with TCP_NODELAY and it is
2067                  * stronger than TCP_NODELAY.
2068                  */
2069                 if (val) {
2070                         tp->nonagle |= TCP_NAGLE_CORK;
2071                 } else {
2072                         tp->nonagle &= ~TCP_NAGLE_CORK;
2073                         if (tp->nonagle&TCP_NAGLE_OFF)
2074                                 tp->nonagle |= TCP_NAGLE_PUSH;
2075                         tcp_push_pending_frames(sk);
2076                 }
2077                 break;
2078
2079         case TCP_KEEPIDLE:
2080                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2081                         err = -EINVAL;
2082                 else {
2083                         tp->keepalive_time = val * HZ;
2084                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2085                             !((1 << sk->sk_state) &
2086                               (TCPF_CLOSE | TCPF_LISTEN))) {
2087                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2088                                 if (tp->keepalive_time > elapsed)
2089                                         elapsed = tp->keepalive_time - elapsed;
2090                                 else
2091                                         elapsed = 0;
2092                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2093                         }
2094                 }
2095                 break;
2096         case TCP_KEEPINTVL:
2097                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2098                         err = -EINVAL;
2099                 else
2100                         tp->keepalive_intvl = val * HZ;
2101                 break;
2102         case TCP_KEEPCNT:
2103                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2104                         err = -EINVAL;
2105                 else
2106                         tp->keepalive_probes = val;
2107                 break;
2108         case TCP_SYNCNT:
2109                 if (val < 1 || val > MAX_TCP_SYNCNT)
2110                         err = -EINVAL;
2111                 else
2112                         icsk->icsk_syn_retries = val;
2113                 break;
2114
2115         case TCP_LINGER2:
2116                 if (val < 0)
2117                         tp->linger2 = -1;
2118                 else if (val > sysctl_tcp_fin_timeout / HZ)
2119                         tp->linger2 = 0;
2120                 else
2121                         tp->linger2 = val * HZ;
2122                 break;
2123
2124         case TCP_DEFER_ACCEPT:
2125                 icsk->icsk_accept_queue.rskq_defer_accept = 0;
2126                 if (val > 0) {
2127                         /* Translate value in seconds to number of
2128                          * retransmits */
2129                         while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
2130                                val > ((TCP_TIMEOUT_INIT / HZ) <<
2131                                        icsk->icsk_accept_queue.rskq_defer_accept))
2132                                 icsk->icsk_accept_queue.rskq_defer_accept++;
2133                         icsk->icsk_accept_queue.rskq_defer_accept++;
2134                 }
2135                 break;
2136
2137         case TCP_WINDOW_CLAMP:
2138                 if (!val) {
2139                         if (sk->sk_state != TCP_CLOSE) {
2140                                 err = -EINVAL;
2141                                 break;
2142                         }
2143                         tp->window_clamp = 0;
2144                 } else
2145                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2146                                                 SOCK_MIN_RCVBUF / 2 : val;
2147                 break;
2148
2149         case TCP_QUICKACK:
2150                 if (!val) {
2151                         icsk->icsk_ack.pingpong = 1;
2152                 } else {
2153                         icsk->icsk_ack.pingpong = 0;
2154                         if ((1 << sk->sk_state) &
2155                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2156                             inet_csk_ack_scheduled(sk)) {
2157                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2158                                 tcp_cleanup_rbuf(sk, 1);
2159                                 if (!(val & 1))
2160                                         icsk->icsk_ack.pingpong = 1;
2161                         }
2162                 }
2163                 break;
2164
2165 #ifdef CONFIG_TCP_MD5SIG
2166         case TCP_MD5SIG:
2167                 /* Read the IP->Key mappings from userspace */
2168                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2169                 break;
2170 #endif
2171
2172         default:
2173                 err = -ENOPROTOOPT;
2174                 break;
2175         }
2176
2177         release_sock(sk);
2178         return err;
2179 }
2180
2181 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2182                    int optlen)
2183 {
2184         struct inet_connection_sock *icsk = inet_csk(sk);
2185
2186         if (level != SOL_TCP)
2187                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2188                                                      optval, optlen);
2189         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2190 }
2191
2192 #ifdef CONFIG_COMPAT
2193 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2194                           char __user *optval, int optlen)
2195 {
2196         if (level != SOL_TCP)
2197                 return inet_csk_compat_setsockopt(sk, level, optname,
2198                                                   optval, optlen);
2199         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2200 }
2201
2202 EXPORT_SYMBOL(compat_tcp_setsockopt);
2203 #endif
2204
2205 /* Return information about state of tcp endpoint in API format. */
2206 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2207 {
2208         struct tcp_sock *tp = tcp_sk(sk);
2209         const struct inet_connection_sock *icsk = inet_csk(sk);
2210         u32 now = tcp_time_stamp;
2211
2212         memset(info, 0, sizeof(*info));
2213
2214         info->tcpi_state = sk->sk_state;
2215         info->tcpi_ca_state = icsk->icsk_ca_state;
2216         info->tcpi_retransmits = icsk->icsk_retransmits;
2217         info->tcpi_probes = icsk->icsk_probes_out;
2218         info->tcpi_backoff = icsk->icsk_backoff;
2219
2220         if (tp->rx_opt.tstamp_ok)
2221                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2222         if (tcp_is_sack(tp))
2223                 info->tcpi_options |= TCPI_OPT_SACK;
2224         if (tp->rx_opt.wscale_ok) {
2225                 info->tcpi_options |= TCPI_OPT_WSCALE;
2226                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2227                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2228         }
2229
2230         if (tp->ecn_flags&TCP_ECN_OK)
2231                 info->tcpi_options |= TCPI_OPT_ECN;
2232
2233         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2234         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2235         info->tcpi_snd_mss = tp->mss_cache;
2236         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2237
2238         if (sk->sk_state == TCP_LISTEN) {
2239                 info->tcpi_unacked = sk->sk_ack_backlog;
2240                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2241         } else {
2242                 info->tcpi_unacked = tp->packets_out;
2243                 info->tcpi_sacked = tp->sacked_out;
2244         }
2245         info->tcpi_lost = tp->lost_out;
2246         info->tcpi_retrans = tp->retrans_out;
2247         info->tcpi_fackets = tp->fackets_out;
2248
2249         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2250         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2251         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2252
2253         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2254         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2255         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2256         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2257         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2258         info->tcpi_snd_cwnd = tp->snd_cwnd;
2259         info->tcpi_advmss = tp->advmss;
2260         info->tcpi_reordering = tp->reordering;
2261
2262         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2263         info->tcpi_rcv_space = tp->rcvq_space.space;
2264
2265         info->tcpi_total_retrans = tp->total_retrans;
2266 }
2267
2268 EXPORT_SYMBOL_GPL(tcp_get_info);
2269
2270 static int do_tcp_getsockopt(struct sock *sk, int level,
2271                 int optname, char __user *optval, int __user *optlen)
2272 {
2273         struct inet_connection_sock *icsk = inet_csk(sk);
2274         struct tcp_sock *tp = tcp_sk(sk);
2275         int val, len;
2276
2277         if (get_user(len, optlen))
2278                 return -EFAULT;
2279
2280         len = min_t(unsigned int, len, sizeof(int));
2281
2282         if (len < 0)
2283                 return -EINVAL;
2284
2285         switch (optname) {
2286         case TCP_MAXSEG:
2287                 val = tp->mss_cache;
2288                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2289                         val = tp->rx_opt.user_mss;
2290                 break;
2291         case TCP_NODELAY:
2292                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2293                 break;
2294         case TCP_CORK:
2295                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2296                 break;
2297         case TCP_KEEPIDLE:
2298                 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2299                 break;
2300         case TCP_KEEPINTVL:
2301                 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2302                 break;
2303         case TCP_KEEPCNT:
2304                 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2305                 break;
2306         case TCP_SYNCNT:
2307                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2308                 break;
2309         case TCP_LINGER2:
2310                 val = tp->linger2;
2311                 if (val >= 0)
2312                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2313                 break;
2314         case TCP_DEFER_ACCEPT:
2315                 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2316                         ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2317                 break;
2318         case TCP_WINDOW_CLAMP:
2319                 val = tp->window_clamp;
2320                 break;
2321         case TCP_INFO: {
2322                 struct tcp_info info;
2323
2324                 if (get_user(len, optlen))
2325                         return -EFAULT;
2326
2327                 tcp_get_info(sk, &info);
2328
2329                 len = min_t(unsigned int, len, sizeof(info));
2330                 if (put_user(len, optlen))
2331                         return -EFAULT;
2332                 if (copy_to_user(optval, &info, len))
2333                         return -EFAULT;
2334                 return 0;
2335         }
2336         case TCP_QUICKACK:
2337                 val = !icsk->icsk_ack.pingpong;
2338                 break;
2339
2340         case TCP_CONGESTION:
2341                 if (get_user(len, optlen))
2342                         return -EFAULT;
2343                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2344                 if (put_user(len, optlen))
2345                         return -EFAULT;
2346                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2347                         return -EFAULT;
2348                 return 0;
2349         default:
2350                 return -ENOPROTOOPT;
2351         }
2352
2353         if (put_user(len, optlen))
2354                 return -EFAULT;
2355         if (copy_to_user(optval, &val, len))
2356                 return -EFAULT;
2357         return 0;
2358 }
2359
2360 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2361                    int __user *optlen)
2362 {
2363         struct inet_connection_sock *icsk = inet_csk(sk);
2364
2365         if (level != SOL_TCP)
2366                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2367                                                      optval, optlen);
2368         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2369 }
2370
2371 #ifdef CONFIG_COMPAT
2372 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2373                           char __user *optval, int __user *optlen)
2374 {
2375         if (level != SOL_TCP)
2376                 return inet_csk_compat_getsockopt(sk, level, optname,
2377                                                   optval, optlen);
2378         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2379 }
2380
2381 EXPORT_SYMBOL(compat_tcp_getsockopt);
2382 #endif
2383
2384 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2385 {
2386         struct sk_buff *segs = ERR_PTR(-EINVAL);
2387         struct tcphdr *th;
2388         unsigned thlen;
2389         unsigned int seq;
2390         __be32 delta;
2391         unsigned int oldlen;
2392         unsigned int mss;
2393
2394         if (!pskb_may_pull(skb, sizeof(*th)))
2395                 goto out;
2396
2397         th = tcp_hdr(skb);
2398         thlen = th->doff * 4;
2399         if (thlen < sizeof(*th))
2400                 goto out;
2401
2402         if (!pskb_may_pull(skb, thlen))
2403                 goto out;
2404
2405         oldlen = (u16)~skb->len;
2406         __skb_pull(skb, thlen);
2407
2408         mss = skb_shinfo(skb)->gso_size;
2409         if (unlikely(skb->len <= mss))
2410                 goto out;
2411
2412         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2413                 /* Packet is from an untrusted source, reset gso_segs. */
2414                 int type = skb_shinfo(skb)->gso_type;
2415
2416                 if (unlikely(type &
2417                              ~(SKB_GSO_TCPV4 |
2418                                SKB_GSO_DODGY |
2419                                SKB_GSO_TCP_ECN |
2420                                SKB_GSO_TCPV6 |
2421                                0) ||
2422                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2423                         goto out;
2424
2425                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2426
2427                 segs = NULL;
2428                 goto out;
2429         }
2430
2431         segs = skb_segment(skb, features);
2432         if (IS_ERR(segs))
2433                 goto out;
2434
2435         delta = htonl(oldlen + (thlen + mss));
2436
2437         skb = segs;
2438         th = tcp_hdr(skb);
2439         seq = ntohl(th->seq);
2440
2441         do {
2442                 th->fin = th->psh = 0;
2443
2444                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2445                                        (__force u32)delta));
2446                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2447                         th->check =
2448                              csum_fold(csum_partial(skb_transport_header(skb),
2449                                                     thlen, skb->csum));
2450
2451                 seq += mss;
2452                 skb = skb->next;
2453                 th = tcp_hdr(skb);
2454
2455                 th->seq = htonl(seq);
2456                 th->cwr = 0;
2457         } while (skb->next);
2458
2459         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2460                       skb->data_len);
2461         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2462                                 (__force u32)delta));
2463         if (skb->ip_summed != CHECKSUM_PARTIAL)
2464                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2465                                                    thlen, skb->csum));
2466
2467 out:
2468         return segs;
2469 }
2470 EXPORT_SYMBOL(tcp_tso_segment);
2471
2472 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2473 {
2474         struct sk_buff **pp = NULL;
2475         struct sk_buff *p;
2476         struct tcphdr *th;
2477         struct tcphdr *th2;
2478         unsigned int thlen;
2479         unsigned int flags;
2480         unsigned int total;
2481         unsigned int mss = 1;
2482         int flush = 1;
2483
2484         if (!pskb_may_pull(skb, sizeof(*th)))
2485                 goto out;
2486
2487         th = tcp_hdr(skb);
2488         thlen = th->doff * 4;
2489         if (thlen < sizeof(*th))
2490                 goto out;
2491
2492         if (!pskb_may_pull(skb, thlen))
2493                 goto out;
2494
2495         th = tcp_hdr(skb);
2496         __skb_pull(skb, thlen);
2497
2498         flags = tcp_flag_word(th);
2499
2500         for (; (p = *head); head = &p->next) {
2501                 if (!NAPI_GRO_CB(p)->same_flow)
2502                         continue;
2503
2504                 th2 = tcp_hdr(p);
2505
2506                 if (th->source != th2->source || th->dest != th2->dest) {
2507                         NAPI_GRO_CB(p)->same_flow = 0;
2508                         continue;
2509                 }
2510
2511                 goto found;
2512         }
2513
2514         goto out_check_final;
2515
2516 found:
2517         flush = NAPI_GRO_CB(p)->flush;
2518         flush |= flags & TCP_FLAG_CWR;
2519         flush |= (flags ^ tcp_flag_word(th2)) &
2520                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH);
2521         flush |= th->ack_seq != th2->ack_seq || th->window != th2->window;
2522         flush |= memcmp(th + 1, th2 + 1, thlen - sizeof(*th));
2523
2524         total = p->len;
2525         mss = skb_shinfo(p)->gso_size;
2526
2527         flush |= skb->len > mss || skb->len <= 0;
2528         flush |= ntohl(th2->seq) + total != ntohl(th->seq);
2529
2530         if (flush || skb_gro_receive(head, skb)) {
2531                 mss = 1;
2532                 goto out_check_final;
2533         }
2534
2535         p = *head;
2536         th2 = tcp_hdr(p);
2537         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2538
2539 out_check_final:
2540         flush = skb->len < mss;
2541         flush |= flags & (TCP_FLAG_URG | TCP_FLAG_PSH | TCP_FLAG_RST |
2542                           TCP_FLAG_SYN | TCP_FLAG_FIN);
2543
2544         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2545                 pp = head;
2546
2547 out:
2548         NAPI_GRO_CB(skb)->flush |= flush;
2549
2550         return pp;
2551 }
2552 EXPORT_SYMBOL(tcp_gro_receive);
2553
2554 int tcp_gro_complete(struct sk_buff *skb)
2555 {
2556         struct tcphdr *th = tcp_hdr(skb);
2557
2558         skb->csum_start = skb_transport_header(skb) - skb->head;
2559         skb->csum_offset = offsetof(struct tcphdr, check);
2560         skb->ip_summed = CHECKSUM_PARTIAL;
2561
2562         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2563
2564         if (th->cwr)
2565                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2566
2567         return 0;
2568 }
2569 EXPORT_SYMBOL(tcp_gro_complete);
2570
2571 #ifdef CONFIG_TCP_MD5SIG
2572 static unsigned long tcp_md5sig_users;
2573 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2574 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2575
2576 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2577 {
2578         int cpu;
2579         for_each_possible_cpu(cpu) {
2580                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2581                 if (p) {
2582                         if (p->md5_desc.tfm)
2583                                 crypto_free_hash(p->md5_desc.tfm);
2584                         kfree(p);
2585                         p = NULL;
2586                 }
2587         }
2588         free_percpu(pool);
2589 }
2590
2591 void tcp_free_md5sig_pool(void)
2592 {
2593         struct tcp_md5sig_pool **pool = NULL;
2594
2595         spin_lock_bh(&tcp_md5sig_pool_lock);
2596         if (--tcp_md5sig_users == 0) {
2597                 pool = tcp_md5sig_pool;
2598                 tcp_md5sig_pool = NULL;
2599         }
2600         spin_unlock_bh(&tcp_md5sig_pool_lock);
2601         if (pool)
2602                 __tcp_free_md5sig_pool(pool);
2603 }
2604
2605 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2606
2607 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(void)
2608 {
2609         int cpu;
2610         struct tcp_md5sig_pool **pool;
2611
2612         pool = alloc_percpu(struct tcp_md5sig_pool *);
2613         if (!pool)
2614                 return NULL;
2615
2616         for_each_possible_cpu(cpu) {
2617                 struct tcp_md5sig_pool *p;
2618                 struct crypto_hash *hash;
2619
2620                 p = kzalloc(sizeof(*p), GFP_KERNEL);
2621                 if (!p)
2622                         goto out_free;
2623                 *per_cpu_ptr(pool, cpu) = p;
2624
2625                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2626                 if (!hash || IS_ERR(hash))
2627                         goto out_free;
2628
2629                 p->md5_desc.tfm = hash;
2630         }
2631         return pool;
2632 out_free:
2633         __tcp_free_md5sig_pool(pool);
2634         return NULL;
2635 }
2636
2637 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(void)
2638 {
2639         struct tcp_md5sig_pool **pool;
2640         int alloc = 0;
2641
2642 retry:
2643         spin_lock_bh(&tcp_md5sig_pool_lock);
2644         pool = tcp_md5sig_pool;
2645         if (tcp_md5sig_users++ == 0) {
2646                 alloc = 1;
2647                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2648         } else if (!pool) {
2649                 tcp_md5sig_users--;
2650                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2651                 cpu_relax();
2652                 goto retry;
2653         } else
2654                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2655
2656         if (alloc) {
2657                 /* we cannot hold spinlock here because this may sleep. */
2658                 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool();
2659                 spin_lock_bh(&tcp_md5sig_pool_lock);
2660                 if (!p) {
2661                         tcp_md5sig_users--;
2662                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2663                         return NULL;
2664                 }
2665                 pool = tcp_md5sig_pool;
2666                 if (pool) {
2667                         /* oops, it has already been assigned. */
2668                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2669                         __tcp_free_md5sig_pool(p);
2670                 } else {
2671                         tcp_md5sig_pool = pool = p;
2672                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2673                 }
2674         }
2675         return pool;
2676 }
2677
2678 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2679
2680 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2681 {
2682         struct tcp_md5sig_pool **p;
2683         spin_lock_bh(&tcp_md5sig_pool_lock);
2684         p = tcp_md5sig_pool;
2685         if (p)
2686                 tcp_md5sig_users++;
2687         spin_unlock_bh(&tcp_md5sig_pool_lock);
2688         return (p ? *per_cpu_ptr(p, cpu) : NULL);
2689 }
2690
2691 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2692
2693 void __tcp_put_md5sig_pool(void)
2694 {
2695         tcp_free_md5sig_pool();
2696 }
2697
2698 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2699
2700 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2701                         struct tcphdr *th)
2702 {
2703         struct scatterlist sg;
2704         int err;
2705
2706         __sum16 old_checksum = th->check;
2707         th->check = 0;
2708         /* options aren't included in the hash */
2709         sg_init_one(&sg, th, sizeof(struct tcphdr));
2710         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
2711         th->check = old_checksum;
2712         return err;
2713 }
2714
2715 EXPORT_SYMBOL(tcp_md5_hash_header);
2716
2717 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2718                           struct sk_buff *skb, unsigned header_len)
2719 {
2720         struct scatterlist sg;
2721         const struct tcphdr *tp = tcp_hdr(skb);
2722         struct hash_desc *desc = &hp->md5_desc;
2723         unsigned i;
2724         const unsigned head_data_len = skb_headlen(skb) > header_len ?
2725                                        skb_headlen(skb) - header_len : 0;
2726         const struct skb_shared_info *shi = skb_shinfo(skb);
2727
2728         sg_init_table(&sg, 1);
2729
2730         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
2731         if (crypto_hash_update(desc, &sg, head_data_len))
2732                 return 1;
2733
2734         for (i = 0; i < shi->nr_frags; ++i) {
2735                 const struct skb_frag_struct *f = &shi->frags[i];
2736                 sg_set_page(&sg, f->page, f->size, f->page_offset);
2737                 if (crypto_hash_update(desc, &sg, f->size))
2738                         return 1;
2739         }
2740
2741         return 0;
2742 }
2743
2744 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
2745
2746 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
2747 {
2748         struct scatterlist sg;
2749
2750         sg_init_one(&sg, key->key, key->keylen);
2751         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
2752 }
2753
2754 EXPORT_SYMBOL(tcp_md5_hash_key);
2755
2756 #endif
2757
2758 void tcp_done(struct sock *sk)
2759 {
2760         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2761                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
2762
2763         tcp_set_state(sk, TCP_CLOSE);
2764         tcp_clear_xmit_timers(sk);
2765
2766         sk->sk_shutdown = SHUTDOWN_MASK;
2767
2768         if (!sock_flag(sk, SOCK_DEAD))
2769                 sk->sk_state_change(sk);
2770         else
2771                 inet_csk_destroy_sock(sk);
2772 }
2773 EXPORT_SYMBOL_GPL(tcp_done);
2774
2775 extern struct tcp_congestion_ops tcp_reno;
2776
2777 static __initdata unsigned long thash_entries;
2778 static int __init set_thash_entries(char *str)
2779 {
2780         if (!str)
2781                 return 0;
2782         thash_entries = simple_strtoul(str, &str, 0);
2783         return 1;
2784 }
2785 __setup("thash_entries=", set_thash_entries);
2786
2787 void __init tcp_init(void)
2788 {
2789         struct sk_buff *skb = NULL;
2790         unsigned long nr_pages, limit;
2791         int order, i, max_share;
2792
2793         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
2794
2795         percpu_counter_init(&tcp_sockets_allocated, 0);
2796         percpu_counter_init(&tcp_orphan_count, 0);
2797         tcp_hashinfo.bind_bucket_cachep =
2798                 kmem_cache_create("tcp_bind_bucket",
2799                                   sizeof(struct inet_bind_bucket), 0,
2800                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
2801
2802         /* Size and allocate the main established and bind bucket
2803          * hash tables.
2804          *
2805          * The methodology is similar to that of the buffer cache.
2806          */
2807         tcp_hashinfo.ehash =
2808                 alloc_large_system_hash("TCP established",
2809                                         sizeof(struct inet_ehash_bucket),
2810                                         thash_entries,
2811                                         (num_physpages >= 128 * 1024) ?
2812                                         13 : 15,
2813                                         0,
2814                                         &tcp_hashinfo.ehash_size,
2815                                         NULL,
2816                                         thash_entries ? 0 : 512 * 1024);
2817         tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
2818         for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
2819                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
2820                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
2821         }
2822         if (inet_ehash_locks_alloc(&tcp_hashinfo))
2823                 panic("TCP: failed to alloc ehash_locks");
2824         tcp_hashinfo.bhash =
2825                 alloc_large_system_hash("TCP bind",
2826                                         sizeof(struct inet_bind_hashbucket),
2827                                         tcp_hashinfo.ehash_size,
2828                                         (num_physpages >= 128 * 1024) ?
2829                                         13 : 15,
2830                                         0,
2831                                         &tcp_hashinfo.bhash_size,
2832                                         NULL,
2833                                         64 * 1024);
2834         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2835         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2836                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2837                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2838         }
2839
2840         /* Try to be a bit smarter and adjust defaults depending
2841          * on available memory.
2842          */
2843         for (order = 0; ((1 << order) << PAGE_SHIFT) <
2844                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2845                         order++)
2846                 ;
2847         if (order >= 4) {
2848                 tcp_death_row.sysctl_max_tw_buckets = 180000;
2849                 sysctl_tcp_max_orphans = 4096 << (order - 4);
2850                 sysctl_max_syn_backlog = 1024;
2851         } else if (order < 3) {
2852                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2853                 sysctl_tcp_max_orphans >>= (3 - order);
2854                 sysctl_max_syn_backlog = 128;
2855         }
2856
2857         /* Set the pressure threshold to be a fraction of global memory that
2858          * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2859          * memory, with a floor of 128 pages.
2860          */
2861         nr_pages = totalram_pages - totalhigh_pages;
2862         limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
2863         limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
2864         limit = max(limit, 128UL);
2865         sysctl_tcp_mem[0] = limit / 4 * 3;
2866         sysctl_tcp_mem[1] = limit;
2867         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2868
2869         /* Set per-socket limits to no more than 1/128 the pressure threshold */
2870         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2871         max_share = min(4UL*1024*1024, limit);
2872
2873         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
2874         sysctl_tcp_wmem[1] = 16*1024;
2875         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2876
2877         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
2878         sysctl_tcp_rmem[1] = 87380;
2879         sysctl_tcp_rmem[2] = max(87380, max_share);
2880
2881         printk(KERN_INFO "TCP: Hash tables configured "
2882                "(established %d bind %d)\n",
2883                tcp_hashinfo.ehash_size, tcp_hashinfo.bhash_size);
2884
2885         tcp_register_congestion_control(&tcp_reno);
2886 }
2887
2888 EXPORT_SYMBOL(tcp_close);
2889 EXPORT_SYMBOL(tcp_disconnect);
2890 EXPORT_SYMBOL(tcp_getsockopt);
2891 EXPORT_SYMBOL(tcp_ioctl);
2892 EXPORT_SYMBOL(tcp_poll);
2893 EXPORT_SYMBOL(tcp_read_sock);
2894 EXPORT_SYMBOL(tcp_recvmsg);
2895 EXPORT_SYMBOL(tcp_sendmsg);
2896 EXPORT_SYMBOL(tcp_splice_read);
2897 EXPORT_SYMBOL(tcp_sendpage);
2898 EXPORT_SYMBOL(tcp_setsockopt);
2899 EXPORT_SYMBOL(tcp_shutdown);