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 /* Convert seconds to retransmits based on initial and max timeout */
 330 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
 331 {
 332         u8 res = 0;
 333
 334         if (seconds > 0) {
 335                 int period = timeout;
 336
 337                 res = 1;
 338                 while (seconds > period && res < 255) {
 339                         res++;
 340                         timeout <<= 1;
 341                         if (timeout > rto_max)
 342                                 timeout = rto_max;
 343                         period += timeout;
 344                 }
 345         }
 346         return res;
 347 }
 348
 349 /* Convert retransmits to seconds based on initial and max timeout */
 350 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
 351 {
 352         int period = 0;
 353
 354         if (retrans > 0) {
 355                 period = timeout;
 356                 while (--retrans) {
 357                         timeout <<= 1;
 358                         if (timeout > rto_max)
 359                                 timeout = rto_max;
 360                         period += timeout;
 361                 }
 362         }
 363         return period;
 364 }
 365
 366 /*
 367  *      Wait for a TCP event.
 368  *
 369  *      Note that we don't need to lock the socket, as the upper poll layers
 370  *      take care of normal races (between the test and the event) and we don't
 371  *      go look at any of the socket buffers directly.
 372  */
 373 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 374 {
 375         unsigned int mask;
 376         struct sock *sk = sock->sk;
 377         struct tcp_sock *tp = tcp_sk(sk);
 378
 379         sock_poll_wait(file, sk->sk_sleep, wait);
 380         if (sk->sk_state == TCP_LISTEN)
 381                 return inet_csk_listen_poll(sk);
 382
 383         /* Socket is not locked. We are protected from async events
 384          * by poll logic and correct handling of state changes
 385          * made by other threads is impossible in any case.
 386          */
 387
 388         mask = 0;
 389
 390         /*
 391          * POLLHUP is certainly not done right. But poll() doesn't
 392          * have a notion of HUP in just one direction, and for a
 393          * socket the read side is more interesting.
 394          *
 395          * Some poll() documentation says that POLLHUP is incompatible
 396          * with the POLLOUT/POLLWR flags, so somebody should check this
 397          * all. But careful, it tends to be safer to return too many
 398          * bits than too few, and you can easily break real applications
 399          * if you don't tell them that something has hung up!
 400          *
 401          * Check-me.
 402          *
 403          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
 404          * our fs/select.c). It means that after we received EOF,
 405          * poll always returns immediately, making impossible poll() on write()
 406          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
 407          * if and only if shutdown has been made in both directions.
 408          * Actually, it is interesting to look how Solaris and DUX
 409          * solve this dilemma. I would prefer, if POLLHUP were maskable,
 410          * then we could set it on SND_SHUTDOWN. BTW examples given
 411          * in Stevens' books assume exactly this behaviour, it explains
 412          * why POLLHUP is incompatible with POLLOUT.    --ANK
 413          *
 414          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 415          * blocking on fresh not-connected or disconnected socket. --ANK
 416          */
 417         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
 418                 mask |= POLLHUP;
 419         if (sk->sk_shutdown & RCV_SHUTDOWN)
 420                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
 421
 422         /* Connected? */
 423         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
 424                 int target = sock_rcvlowat(sk, 0, INT_MAX);
 425
 426                 if (tp->urg_seq == tp->copied_seq &&
 427                     !sock_flag(sk, SOCK_URGINLINE) &&
 428                     tp->urg_data)
 429                         target--;
 430
 431                 /* Potential race condition. If read of tp below will
 432                  * escape above sk->sk_state, we can be illegally awaken
 433                  * in SYN_* states. */
 434                 if (tp->rcv_nxt - tp->copied_seq >= target)
 435                         mask |= POLLIN | POLLRDNORM;
 436
 437                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 438                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
 439                                 mask |= POLLOUT | POLLWRNORM;
 440                         } else {  /* send SIGIO later */
 441                                 set_bit(SOCK_ASYNC_NOSPACE,
 442                                         &sk->sk_socket->flags);
 443                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 444
 445                                 /* Race breaker. If space is freed after
 446                                  * wspace test but before the flags are set,
 447                                  * IO signal will be lost.
 448                                  */
 449                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
 450                                         mask |= POLLOUT | POLLWRNORM;
 451                         }
 452                 } else
 453                         mask |= POLLOUT | POLLWRNORM;
 454
 455                 if (tp->urg_data & TCP_URG_VALID)
 456                         mask |= POLLPRI;
 457         }
 458         /* This barrier is coupled with smp_wmb() in tcp_reset() */
 459         smp_rmb();
 460         if (sk->sk_err)
 461                 mask |= POLLERR;
 462
 463         return mask;
 464 }
 465
 466 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 467 {
 468         struct tcp_sock *tp = tcp_sk(sk);
 469         int answ;
 470
 471         switch (cmd) {
 472         case SIOCINQ:
 473                 if (sk->sk_state == TCP_LISTEN)
 474                         return -EINVAL;
 475
 476                 lock_sock(sk);
 477                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 478                         answ = 0;
 479                 else if (sock_flag(sk, SOCK_URGINLINE) ||
 480                          !tp->urg_data ||
 481                          before(tp->urg_seq, tp->copied_seq) ||
 482                          !before(tp->urg_seq, tp->rcv_nxt)) {
 483                         struct sk_buff *skb;
 484
 485                         answ = tp->rcv_nxt - tp->copied_seq;
 486
 487                         /* Subtract 1, if FIN is in queue. */
 488                         skb = skb_peek_tail(&sk->sk_receive_queue);
 489                         if (answ && skb)
 490                                 answ -= tcp_hdr(skb)->fin;
 491                 } else
 492                         answ = tp->urg_seq - tp->copied_seq;
 493                 release_sock(sk);
 494                 break;
 495         case SIOCATMARK:
 496                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 497                 break;
 498         case SIOCOUTQ:
 499                 if (sk->sk_state == TCP_LISTEN)
 500                         return -EINVAL;
 501
 502                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 503                         answ = 0;
 504                 else
 505                         answ = tp->write_seq - tp->snd_una;
 506                 break;
 507         default:
 508                 return -ENOIOCTLCMD;
 509         }
 510
 511         return put_user(answ, (int __user *)arg);
 512 }
 513
 514 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 515 {
 516         TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
 517         tp->pushed_seq = tp->write_seq;
 518 }
 519
 520 static inline int forced_push(struct tcp_sock *tp)
 521 {
 522         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 523 }
 524
 525 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
 526 {
 527         struct tcp_sock *tp = tcp_sk(sk);
 528         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 529
 530         skb->csum    = 0;
 531         tcb->seq     = tcb->end_seq = tp->write_seq;
 532         tcb->flags   = TCPCB_FLAG_ACK;
 533         tcb->sacked  = 0;
 534         skb_header_release(skb);
 535         tcp_add_write_queue_tail(sk, skb);
 536         sk->sk_wmem_queued += skb->truesize;
 537         sk_mem_charge(sk, skb->truesize);
 538         if (tp->nonagle & TCP_NAGLE_PUSH)
 539                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 540 }
 541
 542 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
 543                                 struct sk_buff *skb)
 544 {
 545         if (flags & MSG_OOB)
 546                 tp->snd_up = tp->write_seq;
 547 }
 548
 549 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
 550                             int nonagle)
 551 {
 552         struct tcp_sock *tp = tcp_sk(sk);
 553
 554         if (tcp_send_head(sk)) {
 555                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 556                 if (!(flags & MSG_MORE) || forced_push(tp))
 557                         tcp_mark_push(tp, skb);
 558                 tcp_mark_urg(tp, flags, skb);
 559                 __tcp_push_pending_frames(sk, mss_now,
 560                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
 561         }
 562 }
 563
 564 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 565                                 unsigned int offset, size_t len)
 566 {
 567         struct tcp_splice_state *tss = rd_desc->arg.data;
 568         int ret;
 569
 570         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
 571                               tss->flags);
 572         if (ret > 0)
 573                 rd_desc->count -= ret;
 574         return ret;
 575 }
 576
 577 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 578 {
 579         /* Store TCP splice context information in read_descriptor_t. */
 580         read_descriptor_t rd_desc = {
 581                 .arg.data = tss,
 582                 .count    = tss->len,
 583         };
 584
 585         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 586 }
 587
 588 /**
 589  *  tcp_splice_read - splice data from TCP socket to a pipe
 590  * @sock:       socket to splice from
 591  * @ppos:       position (not valid)
 592  * @pipe:       pipe to splice to
 593  * @len:        number of bytes to splice
 594  * @flags:      splice modifier flags
 595  *
 596  * Description:
 597  *    Will read pages from given socket and fill them into a pipe.
 598  *
 599  **/
 600 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 601                         struct pipe_inode_info *pipe, size_t len,
 602                         unsigned int flags)
 603 {
 604         struct sock *sk = sock->sk;
 605         struct tcp_splice_state tss = {
 606                 .pipe = pipe,
 607                 .len = len,
 608                 .flags = flags,
 609         };
 610         long timeo;
 611         ssize_t spliced;
 612         int ret;
 613
 614         /*
 615          * We can't seek on a socket input
 616          */
 617         if (unlikely(*ppos))
 618                 return -ESPIPE;
 619
 620         ret = spliced = 0;
 621
 622         lock_sock(sk);
 623
 624         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 625         while (tss.len) {
 626                 ret = __tcp_splice_read(sk, &tss);
 627                 if (ret < 0)
 628                         break;
 629                 else if (!ret) {
 630                         if (spliced)
 631                                 break;
 632                         if (sock_flag(sk, SOCK_DONE))
 633                                 break;
 634                         if (sk->sk_err) {
 635                                 ret = sock_error(sk);
 636                                 break;
 637                         }
 638                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 639                                 break;
 640                         if (sk->sk_state == TCP_CLOSE) {
 641                                 /*
 642                                  * This occurs when user tries to read
 643                                  * from never connected socket.
 644                                  */
 645                                 if (!sock_flag(sk, SOCK_DONE))
 646                                         ret = -ENOTCONN;
 647                                 break;
 648                         }
 649                         if (!timeo) {
 650                                 ret = -EAGAIN;
 651                                 break;
 652                         }
 653                         sk_wait_data(sk, &timeo);
 654                         if (signal_pending(current)) {
 655                                 ret = sock_intr_errno(timeo);
 656                                 break;
 657                         }
 658                         continue;
 659                 }
 660                 tss.len -= ret;
 661                 spliced += ret;
 662
 663                 if (!timeo)
 664                         break;
 665                 release_sock(sk);
 666                 lock_sock(sk);
 667
 668                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 669                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 670                     signal_pending(current))
 671                         break;
 672         }
 673
 674         release_sock(sk);
 675
 676         if (spliced)
 677                 return spliced;
 678
 679         return ret;
 680 }
 681
 682 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
 683 {
 684         struct sk_buff *skb;
 685
 686         /* The TCP header must be at least 32-bit aligned.  */
 687         size = ALIGN(size, 4);
 688
 689         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 690         if (skb) {
 691                 if (sk_wmem_schedule(sk, skb->truesize)) {
 692                         /*
 693                          * Make sure that we have exactly size bytes
 694                          * available to the caller, no more, no less.
 695                          */
 696                         skb_reserve(skb, skb_tailroom(skb) - size);
 697                         return skb;
 698                 }
 699                 __kfree_skb(skb);
 700         } else {
 701                 sk->sk_prot->enter_memory_pressure(sk);
 702                 sk_stream_moderate_sndbuf(sk);
 703         }
 704         return NULL;
 705 }
 706
 707 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 708                                        int large_allowed)
 709 {
 710         struct tcp_sock *tp = tcp_sk(sk);
 711         u32 xmit_size_goal, old_size_goal;
 712
 713         xmit_size_goal = mss_now;
 714
 715         if (large_allowed && sk_can_gso(sk)) {
 716                 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
 717                                   inet_csk(sk)->icsk_af_ops->net_header_len -
 718                                   inet_csk(sk)->icsk_ext_hdr_len -
 719                                   tp->tcp_header_len);
 720
 721                 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
 722
 723                 /* We try hard to avoid divides here */
 724                 old_size_goal = tp->xmit_size_goal_segs * mss_now;
 725
 726                 if (likely(old_size_goal <= xmit_size_goal &&
 727                            old_size_goal + mss_now > xmit_size_goal)) {
 728                         xmit_size_goal = old_size_goal;
 729                 } else {
 730                         tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
 731                         xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
 732                 }
 733         }
 734
 735         return max(xmit_size_goal, mss_now);
 736 }
 737
 738 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 739 {
 740         int mss_now;
 741
 742         mss_now = tcp_current_mss(sk);
 743         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 744
 745         return mss_now;
 746 }
 747
 748 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
 749                          size_t psize, int flags)
 750 {
 751         struct tcp_sock *tp = tcp_sk(sk);
 752         int mss_now, size_goal;
 753         int err;
 754         ssize_t copied;
 755         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 756
 757         /* Wait for a connection to finish. */
 758         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 759                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 760                         goto out_err;
 761
 762         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 763
 764         mss_now = tcp_send_mss(sk, &size_goal, flags);
 765         copied = 0;
 766
 767         err = -EPIPE;
 768         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 769                 goto out_err;
 770
 771         while (psize > 0) {
 772                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 773                 struct page *page = pages[poffset / PAGE_SIZE];
 774                 int copy, i, can_coalesce;
 775                 int offset = poffset % PAGE_SIZE;
 776                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
 777
 778                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
 779 new_segment:
 780                         if (!sk_stream_memory_free(sk))
 781                                 goto wait_for_sndbuf;
 782
 783                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
 784                         if (!skb)
 785                                 goto wait_for_memory;
 786
 787                         skb_entail(sk, skb);
 788                         copy = size_goal;
 789                 }
 790
 791                 if (copy > size)
 792                         copy = size;
 793
 794                 i = skb_shinfo(skb)->nr_frags;
 795                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 796                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
 797                         tcp_mark_push(tp, skb);
 798                         goto new_segment;
 799                 }
 800                 if (!sk_wmem_schedule(sk, copy))
 801                         goto wait_for_memory;
 802
 803                 if (can_coalesce) {
 804                         skb_shinfo(skb)->frags[i - 1].size += copy;
 805                 } else {
 806                         get_page(page);
 807                         skb_fill_page_desc(skb, i, page, offset, copy);
 808                 }
 809
 810                 skb->len += copy;
 811                 skb->data_len += copy;
 812                 skb->truesize += copy;
 813                 sk->sk_wmem_queued += copy;
 814                 sk_mem_charge(sk, copy);
 815                 skb->ip_summed = CHECKSUM_PARTIAL;
 816                 tp->write_seq += copy;
 817                 TCP_SKB_CB(skb)->end_seq += copy;
 818                 skb_shinfo(skb)->gso_segs = 0;
 819
 820                 if (!copied)
 821                         TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
 822
 823                 copied += copy;
 824                 poffset += copy;
 825                 if (!(psize -= copy))
 826                         goto out;
 827
 828                 if (skb->len < size_goal || (flags & MSG_OOB))
 829                         continue;
 830
 831                 if (forced_push(tp)) {
 832                         tcp_mark_push(tp, skb);
 833                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
 834                 } else if (skb == tcp_send_head(sk))
 835                         tcp_push_one(sk, mss_now);
 836                 continue;
 837
 838 wait_for_sndbuf:
 839                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 840 wait_for_memory:
 841                 if (copied)
 842                         tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
 843
 844                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 845                         goto do_error;
 846
 847                 mss_now = tcp_send_mss(sk, &size_goal, flags);
 848         }
 849
 850 out:
 851         if (copied)
 852                 tcp_push(sk, flags, mss_now, tp->nonagle);
 853         return copied;
 854
 855 do_error:
 856         if (copied)
 857                 goto out;
 858 out_err:
 859         return sk_stream_error(sk, flags, err);
 860 }
 861
 862 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
 863                      size_t size, int flags)
 864 {
 865         ssize_t res;
 866         struct sock *sk = sock->sk;
 867
 868         if (!(sk->sk_route_caps & NETIF_F_SG) ||
 869             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
 870                 return sock_no_sendpage(sock, page, offset, size, flags);
 871
 872         lock_sock(sk);
 873         TCP_CHECK_TIMER(sk);
 874         res = do_tcp_sendpages(sk, &page, offset, size, flags);
 875         TCP_CHECK_TIMER(sk);
 876         release_sock(sk);
 877         return res;
 878 }
 879
 880 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
 881 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
 882
 883 static inline int select_size(struct sock *sk)
 884 {
 885         struct tcp_sock *tp = tcp_sk(sk);
 886         int tmp = tp->mss_cache;
 887
 888         if (sk->sk_route_caps & NETIF_F_SG) {
 889                 if (sk_can_gso(sk))
 890                         tmp = 0;
 891                 else {
 892                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
 893
 894                         if (tmp >= pgbreak &&
 895                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
 896                                 tmp = pgbreak;
 897                 }
 898         }
 899
 900         return tmp;
 901 }
 902
 903 int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
 904                 size_t size)
 905 {
 906         struct sock *sk = sock->sk;
 907         struct iovec *iov;
 908         struct tcp_sock *tp = tcp_sk(sk);
 909         struct sk_buff *skb;
 910         int iovlen, flags;
 911         int mss_now, size_goal;
 912         int err, copied;
 913         long timeo;
 914
 915         lock_sock(sk);
 916         TCP_CHECK_TIMER(sk);
 917
 918         flags = msg->msg_flags;
 919         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 920
 921         /* Wait for a connection to finish. */
 922         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 923                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 924                         goto out_err;
 925
 926         /* This should be in poll */
 927         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 928
 929         mss_now = tcp_send_mss(sk, &size_goal, flags);
 930
 931         /* Ok commence sending. */
 932         iovlen = msg->msg_iovlen;
 933         iov = msg->msg_iov;
 934         copied = 0;
 935
 936         err = -EPIPE;
 937         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 938                 goto out_err;
 939
 940         while (--iovlen >= 0) {
 941                 size_t seglen = iov->iov_len;
 942                 unsigned char __user *from = iov->iov_base;
 943
 944                 iov++;
 945
 946                 while (seglen > 0) {
 947                         int copy = 0;
 948                         int max = size_goal;
 949
 950                         skb = tcp_write_queue_tail(sk);
 951                         if (tcp_send_head(sk)) {
 952                                 if (skb->ip_summed == CHECKSUM_NONE)
 953                                         max = mss_now;
 954                                 copy = max - skb->len;
 955                         }
 956
 957                         if (copy <= 0) {
 958 new_segment:
 959                                 /* Allocate new segment. If the interface is SG,
 960                                  * allocate skb fitting to single page.
 961                                  */
 962                                 if (!sk_stream_memory_free(sk))
 963                                         goto wait_for_sndbuf;
 964
 965                                 skb = sk_stream_alloc_skb(sk, select_size(sk),
 966                                                 sk->sk_allocation);
 967                                 if (!skb)
 968                                         goto wait_for_memory;
 969
 970                                 /*
 971                                  * Check whether we can use HW checksum.
 972                                  */
 973                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
 974                                         skb->ip_summed = CHECKSUM_PARTIAL;
 975
 976                                 skb_entail(sk, skb);
 977                                 copy = size_goal;
 978                                 max = size_goal;
 979                         }
 980
 981                         /* Try to append data to the end of skb. */
 982                         if (copy > seglen)
 983                                 copy = seglen;
 984
 985                         /* Where to copy to? */
 986                         if (skb_tailroom(skb) > 0) {
 987                                 /* We have some space in skb head. Superb! */
 988                                 if (copy > skb_tailroom(skb))
 989                                         copy = skb_tailroom(skb);
 990                                 if ((err = skb_add_data(skb, from, copy)) != 0)
 991                                         goto do_fault;
 992                         } else {
 993                                 int merge = 0;
 994                                 int i = skb_shinfo(skb)->nr_frags;
 995                                 struct page *page = TCP_PAGE(sk);
 996                                 int off = TCP_OFF(sk);
 997
 998                                 if (skb_can_coalesce(skb, i, page, off) &&
 999                                     off != PAGE_SIZE) {
1000                                         /* We can extend the last page
1001                                          * fragment. */
1002                                         merge = 1;
1003                                 } else if (i == MAX_SKB_FRAGS ||
1004                                            (!i &&
1005                                            !(sk->sk_route_caps & NETIF_F_SG))) {
1006                                         /* Need to add new fragment and cannot
1007                                          * do this because interface is non-SG,
1008                                          * or because all the page slots are
1009                                          * busy. */
1010                                         tcp_mark_push(tp, skb);
1011                                         goto new_segment;
1012                                 } else if (page) {
1013                                         if (off == PAGE_SIZE) {
1014                                                 put_page(page);
1015                                                 TCP_PAGE(sk) = page = NULL;
1016                                                 off = 0;
1017                                         }
1018                                 } else
1019                                         off = 0;
1020
1021                                 if (copy > PAGE_SIZE - off)
1022                                         copy = PAGE_SIZE - off;
1023
1024                                 if (!sk_wmem_schedule(sk, copy))
1025                                         goto wait_for_memory;
1026
1027                                 if (!page) {
1028                                         /* Allocate new cache page. */
1029                                         if (!(page = sk_stream_alloc_page(sk)))
1030                                                 goto wait_for_memory;
1031                                 }
1032
1033                                 /* Time to copy data. We are close to
1034                                  * the end! */
1035                                 err = skb_copy_to_page(sk, from, skb, page,
1036                                                        off, copy);
1037                                 if (err) {
1038                                         /* If this page was new, give it to the
1039                                          * socket so it does not get leaked.
1040                                          */
1041                                         if (!TCP_PAGE(sk)) {
1042                                                 TCP_PAGE(sk) = page;
1043                                                 TCP_OFF(sk) = 0;
1044                                         }
1045                                         goto do_error;
1046                                 }
1047
1048                                 /* Update the skb. */
1049                                 if (merge) {
1050                                         skb_shinfo(skb)->frags[i - 1].size +=
1051                                                                         copy;
1052                                 } else {
1053                                         skb_fill_page_desc(skb, i, page, off, copy);
1054                                         if (TCP_PAGE(sk)) {
1055                                                 get_page(page);
1056                                         } else if (off + copy < PAGE_SIZE) {
1057                                                 get_page(page);
1058                                                 TCP_PAGE(sk) = page;
1059                                         }
1060                                 }
1061
1062                                 TCP_OFF(sk) = off + copy;
1063                         }
1064
1065                         if (!copied)
1066                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
1067
1068                         tp->write_seq += copy;
1069                         TCP_SKB_CB(skb)->end_seq += copy;
1070                         skb_shinfo(skb)->gso_segs = 0;
1071
1072                         from += copy;
1073                         copied += copy;
1074                         if ((seglen -= copy) == 0 && iovlen == 0)
1075                                 goto out;
1076
1077                         if (skb->len < max || (flags & MSG_OOB))
1078                                 continue;
1079
1080                         if (forced_push(tp)) {
1081                                 tcp_mark_push(tp, skb);
1082                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1083                         } else if (skb == tcp_send_head(sk))
1084                                 tcp_push_one(sk, mss_now);
1085                         continue;
1086
1087 wait_for_sndbuf:
1088                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1089 wait_for_memory:
1090                         if (copied)
1091                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1092
1093                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1094                                 goto do_error;
1095
1096                         mss_now = tcp_send_mss(sk, &size_goal, flags);
1097                 }
1098         }
1099
1100 out:
1101         if (copied)
1102                 tcp_push(sk, flags, mss_now, tp->nonagle);
1103         TCP_CHECK_TIMER(sk);
1104         release_sock(sk);
1105         return copied;
1106
1107 do_fault:
1108         if (!skb->len) {
1109                 tcp_unlink_write_queue(skb, sk);
1110                 /* It is the one place in all of TCP, except connection
1111                  * reset, where we can be unlinking the send_head.
1112                  */
1113                 tcp_check_send_head(sk, skb);
1114                 sk_wmem_free_skb(sk, skb);
1115         }
1116
1117 do_error:
1118         if (copied)
1119                 goto out;
1120 out_err:
1121         err = sk_stream_error(sk, flags, err);
1122         TCP_CHECK_TIMER(sk);
1123         release_sock(sk);
1124         return err;
1125 }
1126
1127 /*
1128  *      Handle reading urgent data. BSD has very simple semantics for
1129  *      this, no blocking and very strange errors 8)
1130  */
1131
1132 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1133 {
1134         struct tcp_sock *tp = tcp_sk(sk);
1135
1136         /* No URG data to read. */
1137         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1138             tp->urg_data == TCP_URG_READ)
1139                 return -EINVAL; /* Yes this is right ! */
1140
1141         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1142                 return -ENOTCONN;
1143
1144         if (tp->urg_data & TCP_URG_VALID) {
1145                 int err = 0;
1146                 char c = tp->urg_data;
1147
1148                 if (!(flags & MSG_PEEK))
1149                         tp->urg_data = TCP_URG_READ;
1150
1151                 /* Read urgent data. */
1152                 msg->msg_flags |= MSG_OOB;
1153
1154                 if (len > 0) {
1155                         if (!(flags & MSG_TRUNC))
1156                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1157                         len = 1;
1158                 } else
1159                         msg->msg_flags |= MSG_TRUNC;
1160
1161                 return err ? -EFAULT : len;
1162         }
1163
1164         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1165                 return 0;
1166
1167         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1168          * the available implementations agree in this case:
1169          * this call should never block, independent of the
1170          * blocking state of the socket.
1171          * Mike <pall@rz.uni-karlsruhe.de>
1172          */
1173         return -EAGAIN;
1174 }
1175
1176 /* Clean up the receive buffer for full frames taken by the user,
1177  * then send an ACK if necessary.  COPIED is the number of bytes
1178  * tcp_recvmsg has given to the user so far, it speeds up the
1179  * calculation of whether or not we must ACK for the sake of
1180  * a window update.
1181  */
1182 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1183 {
1184         struct tcp_sock *tp = tcp_sk(sk);
1185         int time_to_ack = 0;
1186
1187 #if TCP_DEBUG
1188         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1189
1190         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1191              KERN_INFO "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1192              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1193 #endif
1194
1195         if (inet_csk_ack_scheduled(sk)) {
1196                 const struct inet_connection_sock *icsk = inet_csk(sk);
1197                    /* Delayed ACKs frequently hit locked sockets during bulk
1198                     * receive. */
1199                 if (icsk->icsk_ack.blocked ||
1200                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1201                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1202                     /*
1203                      * If this read emptied read buffer, we send ACK, if
1204                      * connection is not bidirectional, user drained
1205                      * receive buffer and there was a small segment
1206                      * in queue.
1207                      */
1208                     (copied > 0 &&
1209                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1210                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1211                        !icsk->icsk_ack.pingpong)) &&
1212                       !atomic_read(&sk->sk_rmem_alloc)))
1213                         time_to_ack = 1;
1214         }
1215
1216         /* We send an ACK if we can now advertise a non-zero window
1217          * which has been raised "significantly".
1218          *
1219          * Even if window raised up to infinity, do not send window open ACK
1220          * in states, where we will not receive more. It is useless.
1221          */
1222         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1223                 __u32 rcv_window_now = tcp_receive_window(tp);
1224
1225                 /* Optimize, __tcp_select_window() is not cheap. */
1226                 if (2*rcv_window_now <= tp->window_clamp) {
1227                         __u32 new_window = __tcp_select_window(sk);
1228
1229                         /* Send ACK now, if this read freed lots of space
1230                          * in our buffer. Certainly, new_window is new window.
1231                          * We can advertise it now, if it is not less than current one.
1232                          * "Lots" means "at least twice" here.
1233                          */
1234                         if (new_window && new_window >= 2 * rcv_window_now)
1235                                 time_to_ack = 1;
1236                 }
1237         }
1238         if (time_to_ack)
1239                 tcp_send_ack(sk);
1240 }
1241
1242 static void tcp_prequeue_process(struct sock *sk)
1243 {
1244         struct sk_buff *skb;
1245         struct tcp_sock *tp = tcp_sk(sk);
1246
1247         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1248
1249         /* RX process wants to run with disabled BHs, though it is not
1250          * necessary */
1251         local_bh_disable();
1252         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1253                 sk_backlog_rcv(sk, skb);
1254         local_bh_enable();
1255
1256         /* Clear memory counter. */
1257         tp->ucopy.memory = 0;
1258 }
1259
1260 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1261 {
1262         struct sk_buff *skb;
1263         u32 offset;
1264
1265         skb_queue_walk(&sk->sk_receive_queue, skb) {
1266                 offset = seq - TCP_SKB_CB(skb)->seq;
1267                 if (tcp_hdr(skb)->syn)
1268                         offset--;
1269                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1270                         *off = offset;
1271                         return skb;
1272                 }
1273         }
1274         return NULL;
1275 }
1276
1277 /*
1278  * This routine provides an alternative to tcp_recvmsg() for routines
1279  * that would like to handle copying from skbuffs directly in 'sendfile'
1280  * fashion.
1281  * Note:
1282  *      - It is assumed that the socket was locked by the caller.
1283  *      - The routine does not block.
1284  *      - At present, there is no support for reading OOB data
1285  *        or for 'peeking' the socket using this routine
1286  *        (although both would be easy to implement).
1287  */
1288 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1289                   sk_read_actor_t recv_actor)
1290 {
1291         struct sk_buff *skb;
1292         struct tcp_sock *tp = tcp_sk(sk);
1293         u32 seq = tp->copied_seq;
1294         u32 offset;
1295         int copied = 0;
1296
1297         if (sk->sk_state == TCP_LISTEN)
1298                 return -ENOTCONN;
1299         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1300                 if (offset < skb->len) {
1301                         int used;
1302                         size_t len;
1303
1304                         len = skb->len - offset;
1305                         /* Stop reading if we hit a patch of urgent data */
1306                         if (tp->urg_data) {
1307                                 u32 urg_offset = tp->urg_seq - seq;
1308                                 if (urg_offset < len)
1309                                         len = urg_offset;
1310                                 if (!len)
1311                                         break;
1312                         }
1313                         used = recv_actor(desc, skb, offset, len);
1314                         if (used < 0) {
1315                                 if (!copied)
1316                                         copied = used;
1317                                 break;
1318                         } else if (used <= len) {
1319                                 seq += used;
1320                                 copied += used;
1321                                 offset += used;
1322                         }
1323                         /*
1324                          * If recv_actor drops the lock (e.g. TCP splice
1325                          * receive) the skb pointer might be invalid when
1326                          * getting here: tcp_collapse might have deleted it
1327                          * while aggregating skbs from the socket queue.
1328                          */
1329                         skb = tcp_recv_skb(sk, seq-1, &offset);
1330                         if (!skb || (offset+1 != skb->len))
1331                                 break;
1332                 }
1333                 if (tcp_hdr(skb)->fin) {
1334                         sk_eat_skb(sk, skb, 0);
1335                         ++seq;
1336                         break;
1337                 }
1338                 sk_eat_skb(sk, skb, 0);
1339                 if (!desc->count)
1340                         break;
1341                 tp->copied_seq = seq;
1342         }
1343         tp->copied_seq = seq;
1344
1345         tcp_rcv_space_adjust(sk);
1346
1347         /* Clean up data we have read: This will do ACK frames. */
1348         if (copied > 0)
1349                 tcp_cleanup_rbuf(sk, copied);
1350         return copied;
1351 }
1352
1353 /*
1354  *      This routine copies from a sock struct into the user buffer.
1355  *
1356  *      Technical note: in 2.3 we work on _locked_ socket, so that
1357  *      tricks with *seq access order and skb->users are not required.
1358  *      Probably, code can be easily improved even more.
1359  */
1360
1361 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1362                 size_t len, int nonblock, int flags, int *addr_len)
1363 {
1364         struct tcp_sock *tp = tcp_sk(sk);
1365         int copied = 0;
1366         u32 peek_seq;
1367         u32 *seq;
1368         unsigned long used;
1369         int err;
1370         int target;             /* Read at least this many bytes */
1371         long timeo;
1372         struct task_struct *user_recv = NULL;
1373         int copied_early = 0;
1374         struct sk_buff *skb;
1375         u32 urg_hole = 0;
1376
1377         lock_sock(sk);
1378
1379         TCP_CHECK_TIMER(sk);
1380
1381         err = -ENOTCONN;
1382         if (sk->sk_state == TCP_LISTEN)
1383                 goto out;
1384
1385         timeo = sock_rcvtimeo(sk, nonblock);
1386
1387         /* Urgent data needs to be handled specially. */
1388         if (flags & MSG_OOB)
1389                 goto recv_urg;
1390
1391         seq = &tp->copied_seq;
1392         if (flags & MSG_PEEK) {
1393                 peek_seq = tp->copied_seq;
1394                 seq = &peek_seq;
1395         }
1396
1397         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1398
1399 #ifdef CONFIG_NET_DMA
1400         tp->ucopy.dma_chan = NULL;
1401         preempt_disable();
1402         skb = skb_peek_tail(&sk->sk_receive_queue);
1403         {
1404                 int available = 0;
1405
1406                 if (skb)
1407                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1408                 if ((available < target) &&
1409                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1410                     !sysctl_tcp_low_latency &&
1411                     dma_find_channel(DMA_MEMCPY)) {
1412                         preempt_enable_no_resched();
1413                         tp->ucopy.pinned_list =
1414                                         dma_pin_iovec_pages(msg->msg_iov, len);
1415                 } else {
1416                         preempt_enable_no_resched();
1417                 }
1418         }
1419 #endif
1420
1421         do {
1422                 u32 offset;
1423
1424                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1425                 if (tp->urg_data && tp->urg_seq == *seq) {
1426                         if (copied)
1427                                 break;
1428                         if (signal_pending(current)) {
1429                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1430                                 break;
1431                         }
1432                 }
1433
1434                 /* Next get a buffer. */
1435
1436                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1437                         /* Now that we have two receive queues this
1438                          * shouldn't happen.
1439                          */
1440                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1441                              KERN_INFO "recvmsg bug: copied %X "
1442                                        "seq %X rcvnxt %X fl %X\n", *seq,
1443                                        TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1444                                        flags))
1445                                 break;
1446
1447                         offset = *seq - TCP_SKB_CB(skb)->seq;
1448                         if (tcp_hdr(skb)->syn)
1449                                 offset--;
1450                         if (offset < skb->len)
1451                                 goto found_ok_skb;
1452                         if (tcp_hdr(skb)->fin)
1453                                 goto found_fin_ok;
1454                         WARN(!(flags & MSG_PEEK), KERN_INFO "recvmsg bug 2: "
1455                                         "copied %X seq %X rcvnxt %X fl %X\n",
1456                                         *seq, TCP_SKB_CB(skb)->seq,
1457                                         tp->rcv_nxt, flags);
1458                 }
1459
1460                 /* Well, if we have backlog, try to process it now yet. */
1461
1462                 if (copied >= target && !sk->sk_backlog.tail)
1463                         break;
1464
1465                 if (copied) {
1466                         if (sk->sk_err ||
1467                             sk->sk_state == TCP_CLOSE ||
1468                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1469                             !timeo ||
1470                             signal_pending(current))
1471                                 break;
1472                 } else {
1473                         if (sock_flag(sk, SOCK_DONE))
1474                                 break;
1475
1476                         if (sk->sk_err) {
1477                                 copied = sock_error(sk);
1478                                 break;
1479                         }
1480
1481                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1482                                 break;
1483
1484                         if (sk->sk_state == TCP_CLOSE) {
1485                                 if (!sock_flag(sk, SOCK_DONE)) {
1486                                         /* This occurs when user tries to read
1487                                          * from never connected socket.
1488                                          */
1489                                         copied = -ENOTCONN;
1490                                         break;
1491                                 }
1492                                 break;
1493                         }
1494
1495                         if (!timeo) {
1496                                 copied = -EAGAIN;
1497                                 break;
1498                         }
1499
1500                         if (signal_pending(current)) {
1501                                 copied = sock_intr_errno(timeo);
1502                                 break;
1503                         }
1504                 }
1505
1506                 tcp_cleanup_rbuf(sk, copied);
1507
1508                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1509                         /* Install new reader */
1510                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1511                                 user_recv = current;
1512                                 tp->ucopy.task = user_recv;
1513                                 tp->ucopy.iov = msg->msg_iov;
1514                         }
1515
1516                         tp->ucopy.len = len;
1517
1518                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1519                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1520
1521                         /* Ugly... If prequeue is not empty, we have to
1522                          * process it before releasing socket, otherwise
1523                          * order will be broken at second iteration.
1524                          * More elegant solution is required!!!
1525                          *
1526                          * Look: we have the following (pseudo)queues:
1527                          *
1528                          * 1. packets in flight
1529                          * 2. backlog
1530                          * 3. prequeue
1531                          * 4. receive_queue
1532                          *
1533                          * Each queue can be processed only if the next ones
1534                          * are empty. At this point we have empty receive_queue.
1535                          * But prequeue _can_ be not empty after 2nd iteration,
1536                          * when we jumped to start of loop because backlog
1537                          * processing added something to receive_queue.
1538                          * We cannot release_sock(), because backlog contains
1539                          * packets arrived _after_ prequeued ones.
1540                          *
1541                          * Shortly, algorithm is clear --- to process all
1542                          * the queues in order. We could make it more directly,
1543                          * requeueing packets from backlog to prequeue, if
1544                          * is not empty. It is more elegant, but eats cycles,
1545                          * unfortunately.
1546                          */
1547                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1548                                 goto do_prequeue;
1549
1550                         /* __ Set realtime policy in scheduler __ */
1551                 }
1552
1553                 if (copied >= target) {
1554                         /* Do not sleep, just process backlog. */
1555                         release_sock(sk);
1556                         lock_sock(sk);
1557                 } else
1558                         sk_wait_data(sk, &timeo);
1559
1560 #ifdef CONFIG_NET_DMA
1561                 tp->ucopy.wakeup = 0;
1562 #endif
1563
1564                 if (user_recv) {
1565                         int chunk;
1566
1567                         /* __ Restore normal policy in scheduler __ */
1568
1569                         if ((chunk = len - tp->ucopy.len) != 0) {
1570                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1571                                 len -= chunk;
1572                                 copied += chunk;
1573                         }
1574
1575                         if (tp->rcv_nxt == tp->copied_seq &&
1576                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1577 do_prequeue:
1578                                 tcp_prequeue_process(sk);
1579
1580                                 if ((chunk = len - tp->ucopy.len) != 0) {
1581                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1582                                         len -= chunk;
1583                                         copied += chunk;
1584                                 }
1585                         }
1586                 }
1587                 if ((flags & MSG_PEEK) &&
1588                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1589                         if (net_ratelimit())
1590                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1591                                        current->comm, task_pid_nr(current));
1592                         peek_seq = tp->copied_seq;
1593                 }
1594                 continue;
1595
1596         found_ok_skb:
1597                 /* Ok so how much can we use? */
1598                 used = skb->len - offset;
1599                 if (len < used)
1600                         used = len;
1601
1602                 /* Do we have urgent data here? */
1603                 if (tp->urg_data) {
1604                         u32 urg_offset = tp->urg_seq - *seq;
1605                         if (urg_offset < used) {
1606                                 if (!urg_offset) {
1607                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1608                                                 ++*seq;
1609                                                 urg_hole++;
1610                                                 offset++;
1611                                                 used--;
1612                                                 if (!used)
1613                                                         goto skip_copy;
1614                                         }
1615                                 } else
1616                                         used = urg_offset;
1617                         }
1618                 }
1619
1620                 if (!(flags & MSG_TRUNC)) {
1621 #ifdef CONFIG_NET_DMA
1622                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1623                                 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1624
1625                         if (tp->ucopy.dma_chan) {
1626                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1627                                         tp->ucopy.dma_chan, skb, offset,
1628                                         msg->msg_iov, used,
1629                                         tp->ucopy.pinned_list);
1630
1631                                 if (tp->ucopy.dma_cookie < 0) {
1632
1633                                         printk(KERN_ALERT "dma_cookie < 0\n");
1634
1635                                         /* Exception. Bailout! */
1636                                         if (!copied)
1637                                                 copied = -EFAULT;
1638                                         break;
1639                                 }
1640                                 if ((offset + used) == skb->len)
1641                                         copied_early = 1;
1642
1643                         } else
1644 #endif
1645                         {
1646                                 err = skb_copy_datagram_iovec(skb, offset,
1647                                                 msg->msg_iov, used);
1648                                 if (err) {
1649                                         /* Exception. Bailout! */
1650                                         if (!copied)
1651                                                 copied = -EFAULT;
1652                                         break;
1653                                 }
1654                         }
1655                 }
1656
1657                 *seq += used;
1658                 copied += used;
1659                 len -= used;
1660
1661                 tcp_rcv_space_adjust(sk);
1662
1663 skip_copy:
1664                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1665                         tp->urg_data = 0;
1666                         tcp_fast_path_check(sk);
1667                 }
1668                 if (used + offset < skb->len)
1669                         continue;
1670
1671                 if (tcp_hdr(skb)->fin)
1672                         goto found_fin_ok;
1673                 if (!(flags & MSG_PEEK)) {
1674                         sk_eat_skb(sk, skb, copied_early);
1675                         copied_early = 0;
1676                 }
1677                 continue;
1678
1679         found_fin_ok:
1680                 /* Process the FIN. */
1681                 ++*seq;
1682                 if (!(flags & MSG_PEEK)) {
1683                         sk_eat_skb(sk, skb, copied_early);
1684                         copied_early = 0;
1685                 }
1686                 break;
1687         } while (len > 0);
1688
1689         if (user_recv) {
1690                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1691                         int chunk;
1692
1693                         tp->ucopy.len = copied > 0 ? len : 0;
1694
1695                         tcp_prequeue_process(sk);
1696
1697                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1698                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1699                                 len -= chunk;
1700                                 copied += chunk;
1701                         }
1702                 }
1703
1704                 tp->ucopy.task = NULL;
1705                 tp->ucopy.len = 0;
1706         }
1707
1708 #ifdef CONFIG_NET_DMA
1709         if (tp->ucopy.dma_chan) {
1710                 dma_cookie_t done, used;
1711
1712                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1713
1714                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1715                                                  tp->ucopy.dma_cookie, &done,
1716                                                  &used) == DMA_IN_PROGRESS) {
1717                         /* do partial cleanup of sk_async_wait_queue */
1718                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1719                                (dma_async_is_complete(skb->dma_cookie, done,
1720                                                       used) == DMA_SUCCESS)) {
1721                                 __skb_dequeue(&sk->sk_async_wait_queue);
1722                                 kfree_skb(skb);
1723                         }
1724                 }
1725
1726                 /* Safe to free early-copied skbs now */
1727                 __skb_queue_purge(&sk->sk_async_wait_queue);
1728                 tp->ucopy.dma_chan = NULL;
1729         }
1730         if (tp->ucopy.pinned_list) {
1731                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1732                 tp->ucopy.pinned_list = NULL;
1733         }
1734 #endif
1735
1736         /* According to UNIX98, msg_name/msg_namelen are ignored
1737          * on connected socket. I was just happy when found this 8) --ANK
1738          */
1739
1740         /* Clean up data we have read: This will do ACK frames. */
1741         tcp_cleanup_rbuf(sk, copied);
1742
1743         TCP_CHECK_TIMER(sk);
1744         release_sock(sk);
1745         return copied;
1746
1747 out:
1748         TCP_CHECK_TIMER(sk);
1749         release_sock(sk);
1750         return err;
1751
1752 recv_urg:
1753         err = tcp_recv_urg(sk, msg, len, flags);
1754         goto out;
1755 }
1756
1757 void tcp_set_state(struct sock *sk, int state)
1758 {
1759         int oldstate = sk->sk_state;
1760
1761         switch (state) {
1762         case TCP_ESTABLISHED:
1763                 if (oldstate != TCP_ESTABLISHED)
1764                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1765                 break;
1766
1767         case TCP_CLOSE:
1768                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1769                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1770
1771                 sk->sk_prot->unhash(sk);
1772                 if (inet_csk(sk)->icsk_bind_hash &&
1773                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1774                         inet_put_port(sk);
1775                 /* fall through */
1776         default:
1777                 if (oldstate == TCP_ESTABLISHED)
1778                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1779         }
1780
1781         /* Change state AFTER socket is unhashed to avoid closed
1782          * socket sitting in hash tables.
1783          */
1784         sk->sk_state = state;
1785
1786 #ifdef STATE_TRACE
1787         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1788 #endif
1789 }
1790 EXPORT_SYMBOL_GPL(tcp_set_state);
1791
1792 /*
1793  *      State processing on a close. This implements the state shift for
1794  *      sending our FIN frame. Note that we only send a FIN for some
1795  *      states. A shutdown() may have already sent the FIN, or we may be
1796  *      closed.
1797  */
1798
1799 static const unsigned char new_state[16] = {
1800   /* current state:        new state:      action:      */
1801   /* (Invalid)          */ TCP_CLOSE,
1802   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1803   /* TCP_SYN_SENT       */ TCP_CLOSE,
1804   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1805   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1806   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1807   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1808   /* TCP_CLOSE          */ TCP_CLOSE,
1809   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1810   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1811   /* TCP_LISTEN         */ TCP_CLOSE,
1812   /* TCP_CLOSING        */ TCP_CLOSING,
1813 };
1814
1815 static int tcp_close_state(struct sock *sk)
1816 {
1817         int next = (int)new_state[sk->sk_state];
1818         int ns = next & TCP_STATE_MASK;
1819
1820         tcp_set_state(sk, ns);
1821
1822         return next & TCP_ACTION_FIN;
1823 }
1824
1825 /*
1826  *      Shutdown the sending side of a connection. Much like close except
1827  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1828  */
1829
1830 void tcp_shutdown(struct sock *sk, int how)
1831 {
1832         /*      We need to grab some memory, and put together a FIN,
1833          *      and then put it into the queue to be sent.
1834          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1835          */
1836         if (!(how & SEND_SHUTDOWN))
1837                 return;
1838
1839         /* If we've already sent a FIN, or it's a closed state, skip this. */
1840         if ((1 << sk->sk_state) &
1841             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1842              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1843                 /* Clear out any half completed packets.  FIN if needed. */
1844                 if (tcp_close_state(sk))
1845                         tcp_send_fin(sk);
1846         }
1847 }
1848
1849 void tcp_close(struct sock *sk, long timeout)
1850 {
1851         struct sk_buff *skb;
1852         int data_was_unread = 0;
1853         int state;
1854
1855         lock_sock(sk);
1856         sk->sk_shutdown = SHUTDOWN_MASK;
1857
1858         if (sk->sk_state == TCP_LISTEN) {
1859                 tcp_set_state(sk, TCP_CLOSE);
1860
1861                 /* Special case. */
1862                 inet_csk_listen_stop(sk);
1863
1864                 goto adjudge_to_death;
1865         }
1866
1867         /*  We need to flush the recv. buffs.  We do this only on the
1868          *  descriptor close, not protocol-sourced closes, because the
1869          *  reader process may not have drained the data yet!
1870          */
1871         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1872                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1873                           tcp_hdr(skb)->fin;
1874                 data_was_unread += len;
1875                 __kfree_skb(skb);
1876         }
1877
1878         sk_mem_reclaim(sk);
1879
1880         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1881          * data was lost. To witness the awful effects of the old behavior of
1882          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1883          * GET in an FTP client, suspend the process, wait for the client to
1884          * advertise a zero window, then kill -9 the FTP client, wheee...
1885          * Note: timeout is always zero in such a case.
1886          */
1887         if (data_was_unread) {
1888                 /* Unread data was tossed, zap the connection. */
1889                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1890                 tcp_set_state(sk, TCP_CLOSE);
1891                 tcp_send_active_reset(sk, sk->sk_allocation);
1892         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1893                 /* Check zero linger _after_ checking for unread data. */
1894                 sk->sk_prot->disconnect(sk, 0);
1895                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1896         } else if (tcp_close_state(sk)) {
1897                 /* We FIN if the application ate all the data before
1898                  * zapping the connection.
1899                  */
1900
1901                 /* RED-PEN. Formally speaking, we have broken TCP state
1902                  * machine. State transitions:
1903                  *
1904                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1905                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1906                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1907                  *
1908                  * are legal only when FIN has been sent (i.e. in window),
1909                  * rather than queued out of window. Purists blame.
1910                  *
1911                  * F.e. "RFC state" is ESTABLISHED,
1912                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1913                  *
1914                  * The visible declinations are that sometimes
1915                  * we enter time-wait state, when it is not required really
1916                  * (harmless), do not send active resets, when they are
1917                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1918                  * they look as CLOSING or LAST_ACK for Linux)
1919                  * Probably, I missed some more holelets.
1920                  *                                              --ANK
1921                  */
1922                 tcp_send_fin(sk);
1923         }
1924
1925         sk_stream_wait_close(sk, timeout);
1926
1927 adjudge_to_death:
1928         state = sk->sk_state;
1929         sock_hold(sk);
1930         sock_orphan(sk);
1931
1932         /* It is the last release_sock in its life. It will remove backlog. */
1933         release_sock(sk);
1934
1935
1936         /* Now socket is owned by kernel and we acquire BH lock
1937            to finish close. No need to check for user refs.
1938          */
1939         local_bh_disable();
1940         bh_lock_sock(sk);
1941         WARN_ON(sock_owned_by_user(sk));
1942
1943         percpu_counter_inc(sk->sk_prot->orphan_count);
1944
1945         /* Have we already been destroyed by a softirq or backlog? */
1946         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1947                 goto out;
1948
1949         /*      This is a (useful) BSD violating of the RFC. There is a
1950          *      problem with TCP as specified in that the other end could
1951          *      keep a socket open forever with no application left this end.
1952          *      We use a 3 minute timeout (about the same as BSD) then kill
1953          *      our end. If they send after that then tough - BUT: long enough
1954          *      that we won't make the old 4*rto = almost no time - whoops
1955          *      reset mistake.
1956          *
1957          *      Nope, it was not mistake. It is really desired behaviour
1958          *      f.e. on http servers, when such sockets are useless, but
1959          *      consume significant resources. Let's do it with special
1960          *      linger2 option.                                 --ANK
1961          */
1962
1963         if (sk->sk_state == TCP_FIN_WAIT2) {
1964                 struct tcp_sock *tp = tcp_sk(sk);
1965                 if (tp->linger2 < 0) {
1966                         tcp_set_state(sk, TCP_CLOSE);
1967                         tcp_send_active_reset(sk, GFP_ATOMIC);
1968                         NET_INC_STATS_BH(sock_net(sk),
1969                                         LINUX_MIB_TCPABORTONLINGER);
1970                 } else {
1971                         const int tmo = tcp_fin_time(sk);
1972
1973                         if (tmo > TCP_TIMEWAIT_LEN) {
1974                                 inet_csk_reset_keepalive_timer(sk,
1975                                                 tmo - TCP_TIMEWAIT_LEN);
1976                         } else {
1977                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1978                                 goto out;
1979                         }
1980                 }
1981         }
1982         if (sk->sk_state != TCP_CLOSE) {
1983                 sk_mem_reclaim(sk);
1984                 if (tcp_too_many_orphans(sk, 0)) {
1985                         if (net_ratelimit())
1986                                 printk(KERN_INFO "TCP: too many of orphaned "
1987                                        "sockets\n");
1988                         tcp_set_state(sk, TCP_CLOSE);
1989                         tcp_send_active_reset(sk, GFP_ATOMIC);
1990                         NET_INC_STATS_BH(sock_net(sk),
1991                                         LINUX_MIB_TCPABORTONMEMORY);
1992                 }
1993         }
1994
1995         if (sk->sk_state == TCP_CLOSE)
1996                 inet_csk_destroy_sock(sk);
1997         /* Otherwise, socket is reprieved until protocol close. */
1998
1999 out:
2000         bh_unlock_sock(sk);
2001         local_bh_enable();
2002         sock_put(sk);
2003 }
2004
2005 /* These states need RST on ABORT according to RFC793 */
2006
2007 static inline int tcp_need_reset(int state)
2008 {
2009         return (1 << state) &
2010                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2011                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2012 }
2013
2014 int tcp_disconnect(struct sock *sk, int flags)
2015 {
2016         struct inet_sock *inet = inet_sk(sk);
2017         struct inet_connection_sock *icsk = inet_csk(sk);
2018         struct tcp_sock *tp = tcp_sk(sk);
2019         int err = 0;
2020         int old_state = sk->sk_state;
2021
2022         if (old_state != TCP_CLOSE)
2023                 tcp_set_state(sk, TCP_CLOSE);
2024
2025         /* ABORT function of RFC793 */
2026         if (old_state == TCP_LISTEN) {
2027                 inet_csk_listen_stop(sk);
2028         } else if (tcp_need_reset(old_state) ||
2029                    (tp->snd_nxt != tp->write_seq &&
2030                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2031                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2032                  * states
2033                  */
2034                 tcp_send_active_reset(sk, gfp_any());
2035                 sk->sk_err = ECONNRESET;
2036         } else if (old_state == TCP_SYN_SENT)
2037                 sk->sk_err = ECONNRESET;
2038
2039         tcp_clear_xmit_timers(sk);
2040         __skb_queue_purge(&sk->sk_receive_queue);
2041         tcp_write_queue_purge(sk);
2042         __skb_queue_purge(&tp->out_of_order_queue);
2043 #ifdef CONFIG_NET_DMA
2044         __skb_queue_purge(&sk->sk_async_wait_queue);
2045 #endif
2046
2047         inet->dport = 0;
2048
2049         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2050                 inet_reset_saddr(sk);
2051
2052         sk->sk_shutdown = 0;
2053         sock_reset_flag(sk, SOCK_DONE);
2054         tp->srtt = 0;
2055         if ((tp->write_seq += tp->max_window + 2) == 0)
2056                 tp->write_seq = 1;
2057         icsk->icsk_backoff = 0;
2058         tp->snd_cwnd = 2;
2059         icsk->icsk_probes_out = 0;
2060         tp->packets_out = 0;
2061         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2062         tp->snd_cwnd_cnt = 0;
2063         tp->bytes_acked = 0;
2064         tcp_set_ca_state(sk, TCP_CA_Open);
2065         tcp_clear_retrans(tp);
2066         inet_csk_delack_init(sk);
2067         tcp_init_send_head(sk);
2068         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2069         __sk_dst_reset(sk);
2070
2071         WARN_ON(inet->num && !icsk->icsk_bind_hash);
2072
2073         sk->sk_error_report(sk);
2074         return err;
2075 }
2076
2077 /*
2078  *      Socket option code for TCP.
2079  */
2080 static int do_tcp_setsockopt(struct sock *sk, int level,
2081                 int optname, char __user *optval, unsigned int optlen)
2082 {
2083         struct tcp_sock *tp = tcp_sk(sk);
2084         struct inet_connection_sock *icsk = inet_csk(sk);
2085         int val;
2086         int err = 0;
2087
2088         /* This is a string value all the others are int's */
2089         if (optname == TCP_CONGESTION) {
2090                 char name[TCP_CA_NAME_MAX];
2091
2092                 if (optlen < 1)
2093                         return -EINVAL;
2094
2095                 val = strncpy_from_user(name, optval,
2096                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2097                 if (val < 0)
2098                         return -EFAULT;
2099                 name[val] = 0;
2100
2101                 lock_sock(sk);
2102                 err = tcp_set_congestion_control(sk, name);
2103                 release_sock(sk);
2104                 return err;
2105         }
2106
2107         if (optlen < sizeof(int))
2108                 return -EINVAL;
2109
2110         if (get_user(val, (int __user *)optval))
2111                 return -EFAULT;
2112
2113         lock_sock(sk);
2114
2115         switch (optname) {
2116         case TCP_MAXSEG:
2117                 /* Values greater than interface MTU won't take effect. However
2118                  * at the point when this call is done we typically don't yet
2119                  * know which interface is going to be used */
2120                 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2121                         err = -EINVAL;
2122                         break;
2123                 }
2124                 tp->rx_opt.user_mss = val;
2125                 break;
2126
2127         case TCP_NODELAY:
2128                 if (val) {
2129                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2130                          * this option on corked socket is remembered, but
2131                          * it is not activated until cork is cleared.
2132                          *
2133                          * However, when TCP_NODELAY is set we make
2134                          * an explicit push, which overrides even TCP_CORK
2135                          * for currently queued segments.
2136                          */
2137                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2138                         tcp_push_pending_frames(sk);
2139                 } else {
2140                         tp->nonagle &= ~TCP_NAGLE_OFF;
2141                 }
2142                 break;
2143
2144         case TCP_CORK:
2145                 /* When set indicates to always queue non-full frames.
2146                  * Later the user clears this option and we transmit
2147                  * any pending partial frames in the queue.  This is
2148                  * meant to be used alongside sendfile() to get properly
2149                  * filled frames when the user (for example) must write
2150                  * out headers with a write() call first and then use
2151                  * sendfile to send out the data parts.
2152                  *
2153                  * TCP_CORK can be set together with TCP_NODELAY and it is
2154                  * stronger than TCP_NODELAY.
2155                  */
2156                 if (val) {
2157                         tp->nonagle |= TCP_NAGLE_CORK;
2158                 } else {
2159                         tp->nonagle &= ~TCP_NAGLE_CORK;
2160                         if (tp->nonagle&TCP_NAGLE_OFF)
2161                                 tp->nonagle |= TCP_NAGLE_PUSH;
2162                         tcp_push_pending_frames(sk);
2163                 }
2164                 break;
2165
2166         case TCP_KEEPIDLE:
2167                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2168                         err = -EINVAL;
2169                 else {
2170                         tp->keepalive_time = val * HZ;
2171                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2172                             !((1 << sk->sk_state) &
2173                               (TCPF_CLOSE | TCPF_LISTEN))) {
2174                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2175                                 if (tp->keepalive_time > elapsed)
2176                                         elapsed = tp->keepalive_time - elapsed;
2177                                 else
2178                                         elapsed = 0;
2179                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2180                         }
2181                 }
2182                 break;
2183         case TCP_KEEPINTVL:
2184                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2185                         err = -EINVAL;
2186                 else
2187                         tp->keepalive_intvl = val * HZ;
2188                 break;
2189         case TCP_KEEPCNT:
2190                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2191                         err = -EINVAL;
2192                 else
2193                         tp->keepalive_probes = val;
2194                 break;
2195         case TCP_SYNCNT:
2196                 if (val < 1 || val > MAX_TCP_SYNCNT)
2197                         err = -EINVAL;
2198                 else
2199                         icsk->icsk_syn_retries = val;
2200                 break;
2201
2202         case TCP_LINGER2:
2203                 if (val < 0)
2204                         tp->linger2 = -1;
2205                 else if (val > sysctl_tcp_fin_timeout / HZ)
2206                         tp->linger2 = 0;
2207                 else
2208                         tp->linger2 = val * HZ;
2209                 break;
2210
2211         case TCP_DEFER_ACCEPT:
2212                 /* Translate value in seconds to number of retransmits */
2213                 icsk->icsk_accept_queue.rskq_defer_accept =
2214                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2215                                         TCP_RTO_MAX / HZ);
2216                 break;
2217
2218         case TCP_WINDOW_CLAMP:
2219                 if (!val) {
2220                         if (sk->sk_state != TCP_CLOSE) {
2221                                 err = -EINVAL;
2222                                 break;
2223                         }
2224                         tp->window_clamp = 0;
2225                 } else
2226                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2227                                                 SOCK_MIN_RCVBUF / 2 : val;
2228                 break;
2229
2230         case TCP_QUICKACK:
2231                 if (!val) {
2232                         icsk->icsk_ack.pingpong = 1;
2233                 } else {
2234                         icsk->icsk_ack.pingpong = 0;
2235                         if ((1 << sk->sk_state) &
2236                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2237                             inet_csk_ack_scheduled(sk)) {
2238                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2239                                 tcp_cleanup_rbuf(sk, 1);
2240                                 if (!(val & 1))
2241                                         icsk->icsk_ack.pingpong = 1;
2242                         }
2243                 }
2244                 break;
2245
2246 #ifdef CONFIG_TCP_MD5SIG
2247         case TCP_MD5SIG:
2248                 /* Read the IP->Key mappings from userspace */
2249                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2250                 break;
2251 #endif
2252
2253         default:
2254                 err = -ENOPROTOOPT;
2255                 break;
2256         }
2257
2258         release_sock(sk);
2259         return err;
2260 }
2261
2262 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2263                    unsigned int optlen)
2264 {
2265         struct inet_connection_sock *icsk = inet_csk(sk);
2266
2267         if (level != SOL_TCP)
2268                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2269                                                      optval, optlen);
2270         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2271 }
2272
2273 #ifdef CONFIG_COMPAT
2274 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2275                           char __user *optval, unsigned int optlen)
2276 {
2277         if (level != SOL_TCP)
2278                 return inet_csk_compat_setsockopt(sk, level, optname,
2279                                                   optval, optlen);
2280         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2281 }
2282
2283 EXPORT_SYMBOL(compat_tcp_setsockopt);
2284 #endif
2285
2286 /* Return information about state of tcp endpoint in API format. */
2287 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2288 {
2289         struct tcp_sock *tp = tcp_sk(sk);
2290         const struct inet_connection_sock *icsk = inet_csk(sk);
2291         u32 now = tcp_time_stamp;
2292
2293         memset(info, 0, sizeof(*info));
2294
2295         info->tcpi_state = sk->sk_state;
2296         info->tcpi_ca_state = icsk->icsk_ca_state;
2297         info->tcpi_retransmits = icsk->icsk_retransmits;
2298         info->tcpi_probes = icsk->icsk_probes_out;
2299         info->tcpi_backoff = icsk->icsk_backoff;
2300
2301         if (tp->rx_opt.tstamp_ok)
2302                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2303         if (tcp_is_sack(tp))
2304                 info->tcpi_options |= TCPI_OPT_SACK;
2305         if (tp->rx_opt.wscale_ok) {
2306                 info->tcpi_options |= TCPI_OPT_WSCALE;
2307                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2308                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2309         }
2310
2311         if (tp->ecn_flags&TCP_ECN_OK)
2312                 info->tcpi_options |= TCPI_OPT_ECN;
2313
2314         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2315         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2316         info->tcpi_snd_mss = tp->mss_cache;
2317         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2318
2319         if (sk->sk_state == TCP_LISTEN) {
2320                 info->tcpi_unacked = sk->sk_ack_backlog;
2321                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2322         } else {
2323                 info->tcpi_unacked = tp->packets_out;
2324                 info->tcpi_sacked = tp->sacked_out;
2325         }
2326         info->tcpi_lost = tp->lost_out;
2327         info->tcpi_retrans = tp->retrans_out;
2328         info->tcpi_fackets = tp->fackets_out;
2329
2330         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2331         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2332         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2333
2334         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2335         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2336         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2337         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2338         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2339         info->tcpi_snd_cwnd = tp->snd_cwnd;
2340         info->tcpi_advmss = tp->advmss;
2341         info->tcpi_reordering = tp->reordering;
2342
2343         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2344         info->tcpi_rcv_space = tp->rcvq_space.space;
2345
2346         info->tcpi_total_retrans = tp->total_retrans;
2347 }
2348
2349 EXPORT_SYMBOL_GPL(tcp_get_info);
2350
2351 static int do_tcp_getsockopt(struct sock *sk, int level,
2352                 int optname, char __user *optval, int __user *optlen)
2353 {
2354         struct inet_connection_sock *icsk = inet_csk(sk);
2355         struct tcp_sock *tp = tcp_sk(sk);
2356         int val, len;
2357
2358         if (get_user(len, optlen))
2359                 return -EFAULT;
2360
2361         len = min_t(unsigned int, len, sizeof(int));
2362
2363         if (len < 0)
2364                 return -EINVAL;
2365
2366         switch (optname) {
2367         case TCP_MAXSEG:
2368                 val = tp->mss_cache;
2369                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2370                         val = tp->rx_opt.user_mss;
2371                 break;
2372         case TCP_NODELAY:
2373                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2374                 break;
2375         case TCP_CORK:
2376                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2377                 break;
2378         case TCP_KEEPIDLE:
2379                 val = keepalive_time_when(tp) / HZ;
2380                 break;
2381         case TCP_KEEPINTVL:
2382                 val = keepalive_intvl_when(tp) / HZ;
2383                 break;
2384         case TCP_KEEPCNT:
2385                 val = keepalive_probes(tp);
2386                 break;
2387         case TCP_SYNCNT:
2388                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2389                 break;
2390         case TCP_LINGER2:
2391                 val = tp->linger2;
2392                 if (val >= 0)
2393                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2394                 break;
2395         case TCP_DEFER_ACCEPT:
2396                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2397                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2398                 break;
2399         case TCP_WINDOW_CLAMP:
2400                 val = tp->window_clamp;
2401                 break;
2402         case TCP_INFO: {
2403                 struct tcp_info info;
2404
2405                 if (get_user(len, optlen))
2406                         return -EFAULT;
2407
2408                 tcp_get_info(sk, &info);
2409
2410                 len = min_t(unsigned int, len, sizeof(info));
2411                 if (put_user(len, optlen))
2412                         return -EFAULT;
2413                 if (copy_to_user(optval, &info, len))
2414                         return -EFAULT;
2415                 return 0;
2416         }
2417         case TCP_QUICKACK:
2418                 val = !icsk->icsk_ack.pingpong;
2419                 break;
2420
2421         case TCP_CONGESTION:
2422                 if (get_user(len, optlen))
2423                         return -EFAULT;
2424                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2425                 if (put_user(len, optlen))
2426                         return -EFAULT;
2427                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2428                         return -EFAULT;
2429                 return 0;
2430         default:
2431                 return -ENOPROTOOPT;
2432         }
2433
2434         if (put_user(len, optlen))
2435                 return -EFAULT;
2436         if (copy_to_user(optval, &val, len))
2437                 return -EFAULT;
2438         return 0;
2439 }
2440
2441 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2442                    int __user *optlen)
2443 {
2444         struct inet_connection_sock *icsk = inet_csk(sk);
2445
2446         if (level != SOL_TCP)
2447                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2448                                                      optval, optlen);
2449         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2450 }
2451
2452 #ifdef CONFIG_COMPAT
2453 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2454                           char __user *optval, int __user *optlen)
2455 {
2456         if (level != SOL_TCP)
2457                 return inet_csk_compat_getsockopt(sk, level, optname,
2458                                                   optval, optlen);
2459         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2460 }
2461
2462 EXPORT_SYMBOL(compat_tcp_getsockopt);
2463 #endif
2464
2465 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2466 {
2467         struct sk_buff *segs = ERR_PTR(-EINVAL);
2468         struct tcphdr *th;
2469         unsigned thlen;
2470         unsigned int seq;
2471         __be32 delta;
2472         unsigned int oldlen;
2473         unsigned int mss;
2474
2475         if (!pskb_may_pull(skb, sizeof(*th)))
2476                 goto out;
2477
2478         th = tcp_hdr(skb);
2479         thlen = th->doff * 4;
2480         if (thlen < sizeof(*th))
2481                 goto out;
2482
2483         if (!pskb_may_pull(skb, thlen))
2484                 goto out;
2485
2486         oldlen = (u16)~skb->len;
2487         __skb_pull(skb, thlen);
2488
2489         mss = skb_shinfo(skb)->gso_size;
2490         if (unlikely(skb->len <= mss))
2491                 goto out;
2492
2493         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2494                 /* Packet is from an untrusted source, reset gso_segs. */
2495                 int type = skb_shinfo(skb)->gso_type;
2496
2497                 if (unlikely(type &
2498                              ~(SKB_GSO_TCPV4 |
2499                                SKB_GSO_DODGY |
2500                                SKB_GSO_TCP_ECN |
2501                                SKB_GSO_TCPV6 |
2502                                0) ||
2503                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2504                         goto out;
2505
2506                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2507
2508                 segs = NULL;
2509                 goto out;
2510         }
2511
2512         segs = skb_segment(skb, features);
2513         if (IS_ERR(segs))
2514                 goto out;
2515
2516         delta = htonl(oldlen + (thlen + mss));
2517
2518         skb = segs;
2519         th = tcp_hdr(skb);
2520         seq = ntohl(th->seq);
2521
2522         do {
2523                 th->fin = th->psh = 0;
2524
2525                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2526                                        (__force u32)delta));
2527                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2528                         th->check =
2529                              csum_fold(csum_partial(skb_transport_header(skb),
2530                                                     thlen, skb->csum));
2531
2532                 seq += mss;
2533                 skb = skb->next;
2534                 th = tcp_hdr(skb);
2535
2536                 th->seq = htonl(seq);
2537                 th->cwr = 0;
2538         } while (skb->next);
2539
2540         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2541                       skb->data_len);
2542         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2543                                 (__force u32)delta));
2544         if (skb->ip_summed != CHECKSUM_PARTIAL)
2545                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2546                                                    thlen, skb->csum));
2547
2548 out:
2549         return segs;
2550 }
2551 EXPORT_SYMBOL(tcp_tso_segment);
2552
2553 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2554 {
2555         struct sk_buff **pp = NULL;
2556         struct sk_buff *p;
2557         struct tcphdr *th;
2558         struct tcphdr *th2;
2559         unsigned int len;
2560         unsigned int thlen;
2561         unsigned int flags;
2562         unsigned int mss = 1;
2563         unsigned int hlen;
2564         unsigned int off;
2565         int flush = 1;
2566         int i;
2567
2568         off = skb_gro_offset(skb);
2569         hlen = off + sizeof(*th);
2570         th = skb_gro_header_fast(skb, off);
2571         if (skb_gro_header_hard(skb, hlen)) {
2572                 th = skb_gro_header_slow(skb, hlen, off);
2573                 if (unlikely(!th))
2574                         goto out;
2575         }
2576
2577         thlen = th->doff * 4;
2578         if (thlen < sizeof(*th))
2579                 goto out;
2580
2581         hlen = off + thlen;
2582         if (skb_gro_header_hard(skb, hlen)) {
2583                 th = skb_gro_header_slow(skb, hlen, off);
2584                 if (unlikely(!th))
2585                         goto out;
2586         }
2587
2588         skb_gro_pull(skb, thlen);
2589
2590         len = skb_gro_len(skb);
2591         flags = tcp_flag_word(th);
2592
2593         for (; (p = *head); head = &p->next) {
2594                 if (!NAPI_GRO_CB(p)->same_flow)
2595                         continue;
2596
2597                 th2 = tcp_hdr(p);
2598
2599                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2600                         NAPI_GRO_CB(p)->same_flow = 0;
2601                         continue;
2602                 }
2603
2604                 goto found;
2605         }
2606
2607         goto out_check_final;
2608
2609 found:
2610         flush = NAPI_GRO_CB(p)->flush;
2611         flush |= flags & TCP_FLAG_CWR;
2612         flush |= (flags ^ tcp_flag_word(th2)) &
2613                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH);
2614         flush |= th->ack_seq ^ th2->ack_seq;
2615         for (i = sizeof(*th); i < thlen; i += 4)
2616                 flush |= *(u32 *)((u8 *)th + i) ^
2617                          *(u32 *)((u8 *)th2 + i);
2618
2619         mss = skb_shinfo(p)->gso_size;
2620
2621         flush |= (len - 1) >= mss;
2622         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2623
2624         if (flush || skb_gro_receive(head, skb)) {
2625                 mss = 1;
2626                 goto out_check_final;
2627         }
2628
2629         p = *head;
2630         th2 = tcp_hdr(p);
2631         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2632
2633 out_check_final:
2634         flush = len < mss;
2635         flush |= flags & (TCP_FLAG_URG | TCP_FLAG_PSH | TCP_FLAG_RST |
2636                           TCP_FLAG_SYN | TCP_FLAG_FIN);
2637
2638         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2639                 pp = head;
2640
2641 out:
2642         NAPI_GRO_CB(skb)->flush |= flush;
2643
2644         return pp;
2645 }
2646 EXPORT_SYMBOL(tcp_gro_receive);
2647
2648 int tcp_gro_complete(struct sk_buff *skb)
2649 {
2650         struct tcphdr *th = tcp_hdr(skb);
2651
2652         skb->csum_start = skb_transport_header(skb) - skb->head;
2653         skb->csum_offset = offsetof(struct tcphdr, check);
2654         skb->ip_summed = CHECKSUM_PARTIAL;
2655
2656         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2657
2658         if (th->cwr)
2659                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2660
2661         return 0;
2662 }
2663 EXPORT_SYMBOL(tcp_gro_complete);
2664
2665 #ifdef CONFIG_TCP_MD5SIG
2666 static unsigned long tcp_md5sig_users;
2667 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2668 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2669
2670 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2671 {
2672         int cpu;
2673         for_each_possible_cpu(cpu) {
2674                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2675                 if (p) {
2676                         if (p->md5_desc.tfm)
2677                                 crypto_free_hash(p->md5_desc.tfm);
2678                         kfree(p);
2679                         p = NULL;
2680                 }
2681         }
2682         free_percpu(pool);
2683 }
2684
2685 void tcp_free_md5sig_pool(void)
2686 {
2687         struct tcp_md5sig_pool **pool = NULL;
2688
2689         spin_lock_bh(&tcp_md5sig_pool_lock);
2690         if (--tcp_md5sig_users == 0) {
2691                 pool = tcp_md5sig_pool;
2692                 tcp_md5sig_pool = NULL;
2693         }
2694         spin_unlock_bh(&tcp_md5sig_pool_lock);
2695         if (pool)
2696                 __tcp_free_md5sig_pool(pool);
2697 }
2698
2699 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2700
2701 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(struct sock *sk)
2702 {
2703         int cpu;
2704         struct tcp_md5sig_pool **pool;
2705
2706         pool = alloc_percpu(struct tcp_md5sig_pool *);
2707         if (!pool)
2708                 return NULL;
2709
2710         for_each_possible_cpu(cpu) {
2711                 struct tcp_md5sig_pool *p;
2712                 struct crypto_hash *hash;
2713
2714                 p = kzalloc(sizeof(*p), sk->sk_allocation);
2715                 if (!p)
2716                         goto out_free;
2717                 *per_cpu_ptr(pool, cpu) = p;
2718
2719                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2720                 if (!hash || IS_ERR(hash))
2721                         goto out_free;
2722
2723                 p->md5_desc.tfm = hash;
2724         }
2725         return pool;
2726 out_free:
2727         __tcp_free_md5sig_pool(pool);
2728         return NULL;
2729 }
2730
2731 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(struct sock *sk)
2732 {
2733         struct tcp_md5sig_pool **pool;
2734         int alloc = 0;
2735
2736 retry:
2737         spin_lock_bh(&tcp_md5sig_pool_lock);
2738         pool = tcp_md5sig_pool;
2739         if (tcp_md5sig_users++ == 0) {
2740                 alloc = 1;
2741                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2742         } else if (!pool) {
2743                 tcp_md5sig_users--;
2744                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2745                 cpu_relax();
2746                 goto retry;
2747         } else
2748                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2749
2750         if (alloc) {
2751                 /* we cannot hold spinlock here because this may sleep. */
2752                 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool(sk);
2753                 spin_lock_bh(&tcp_md5sig_pool_lock);
2754                 if (!p) {
2755                         tcp_md5sig_users--;
2756                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2757                         return NULL;
2758                 }
2759                 pool = tcp_md5sig_pool;
2760                 if (pool) {
2761                         /* oops, it has already been assigned. */
2762                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2763                         __tcp_free_md5sig_pool(p);
2764                 } else {
2765                         tcp_md5sig_pool = pool = p;
2766                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2767                 }
2768         }
2769         return pool;
2770 }
2771
2772 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2773
2774 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2775 {
2776         struct tcp_md5sig_pool **p;
2777         spin_lock_bh(&tcp_md5sig_pool_lock);
2778         p = tcp_md5sig_pool;
2779         if (p)
2780                 tcp_md5sig_users++;
2781         spin_unlock_bh(&tcp_md5sig_pool_lock);
2782         return (p ? *per_cpu_ptr(p, cpu) : NULL);
2783 }
2784
2785 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2786
2787 void __tcp_put_md5sig_pool(void)
2788 {
2789         tcp_free_md5sig_pool();
2790 }
2791
2792 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2793
2794 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2795                         struct tcphdr *th)
2796 {
2797         struct scatterlist sg;
2798         int err;
2799
2800         __sum16 old_checksum = th->check;
2801         th->check = 0;
2802         /* options aren't included in the hash */
2803         sg_init_one(&sg, th, sizeof(struct tcphdr));
2804         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
2805         th->check = old_checksum;
2806         return err;
2807 }
2808
2809 EXPORT_SYMBOL(tcp_md5_hash_header);
2810
2811 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2812                           struct sk_buff *skb, unsigned header_len)
2813 {
2814         struct scatterlist sg;
2815         const struct tcphdr *tp = tcp_hdr(skb);
2816         struct hash_desc *desc = &hp->md5_desc;
2817         unsigned i;
2818         const unsigned head_data_len = skb_headlen(skb) > header_len ?
2819                                        skb_headlen(skb) - header_len : 0;
2820         const struct skb_shared_info *shi = skb_shinfo(skb);
2821
2822         sg_init_table(&sg, 1);
2823
2824         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
2825         if (crypto_hash_update(desc, &sg, head_data_len))
2826                 return 1;
2827
2828         for (i = 0; i < shi->nr_frags; ++i) {
2829                 const struct skb_frag_struct *f = &shi->frags[i];
2830                 sg_set_page(&sg, f->page, f->size, f->page_offset);
2831                 if (crypto_hash_update(desc, &sg, f->size))
2832                         return 1;
2833         }
2834
2835         return 0;
2836 }
2837
2838 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
2839
2840 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
2841 {
2842         struct scatterlist sg;
2843
2844         sg_init_one(&sg, key->key, key->keylen);
2845         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
2846 }
2847
2848 EXPORT_SYMBOL(tcp_md5_hash_key);
2849
2850 #endif
2851
2852 void tcp_done(struct sock *sk)
2853 {
2854         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2855                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
2856
2857         tcp_set_state(sk, TCP_CLOSE);
2858         tcp_clear_xmit_timers(sk);
2859
2860         sk->sk_shutdown = SHUTDOWN_MASK;
2861
2862         if (!sock_flag(sk, SOCK_DEAD))
2863                 sk->sk_state_change(sk);
2864         else
2865                 inet_csk_destroy_sock(sk);
2866 }
2867 EXPORT_SYMBOL_GPL(tcp_done);
2868
2869 extern struct tcp_congestion_ops tcp_reno;
2870
2871 static __initdata unsigned long thash_entries;
2872 static int __init set_thash_entries(char *str)
2873 {
2874         if (!str)
2875                 return 0;
2876         thash_entries = simple_strtoul(str, &str, 0);
2877         return 1;
2878 }
2879 __setup("thash_entries=", set_thash_entries);
2880
2881 void __init tcp_init(void)
2882 {
2883         struct sk_buff *skb = NULL;
2884         unsigned long nr_pages, limit;
2885         int i, max_share, cnt;
2886
2887         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
2888
2889         percpu_counter_init(&tcp_sockets_allocated, 0);
2890         percpu_counter_init(&tcp_orphan_count, 0);
2891         tcp_hashinfo.bind_bucket_cachep =
2892                 kmem_cache_create("tcp_bind_bucket",
2893                                   sizeof(struct inet_bind_bucket), 0,
2894                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
2895
2896         /* Size and allocate the main established and bind bucket
2897          * hash tables.
2898          *
2899          * The methodology is similar to that of the buffer cache.
2900          */
2901         tcp_hashinfo.ehash =
2902                 alloc_large_system_hash("TCP established",
2903                                         sizeof(struct inet_ehash_bucket),
2904                                         thash_entries,
2905                                         (totalram_pages >= 128 * 1024) ?
2906                                         13 : 15,
2907                                         0,
2908                                         &tcp_hashinfo.ehash_size,
2909                                         NULL,
2910                                         thash_entries ? 0 : 512 * 1024);
2911         tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
2912         for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
2913                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
2914                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
2915         }
2916         if (inet_ehash_locks_alloc(&tcp_hashinfo))
2917                 panic("TCP: failed to alloc ehash_locks");
2918         tcp_hashinfo.bhash =
2919                 alloc_large_system_hash("TCP bind",
2920                                         sizeof(struct inet_bind_hashbucket),
2921                                         tcp_hashinfo.ehash_size,
2922                                         (totalram_pages >= 128 * 1024) ?
2923                                         13 : 15,
2924                                         0,
2925                                         &tcp_hashinfo.bhash_size,
2926                                         NULL,
2927                                         64 * 1024);
2928         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2929         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2930                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2931                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2932         }
2933
2934
2935         cnt = tcp_hashinfo.ehash_size;
2936
2937         tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
2938         sysctl_tcp_max_orphans = cnt / 2;
2939         sysctl_max_syn_backlog = max(128, cnt / 256);
2940
2941         /* Set the pressure threshold to be a fraction of global memory that
2942          * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2943          * memory, with a floor of 128 pages, and a ceiling that prevents an
2944          * integer overflow.
2945          */
2946         nr_pages = totalram_pages - totalhigh_pages;
2947         limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
2948         limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
2949         limit = max(limit, 128UL);
2950         limit = min(limit, INT_MAX * 4UL / 3 / 2);
2951         sysctl_tcp_mem[0] = limit / 4 * 3;
2952         sysctl_tcp_mem[1] = limit;
2953         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2954
2955         /* Set per-socket limits to no more than 1/128 the pressure threshold */
2956         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2957         max_share = min(4UL*1024*1024, limit);
2958
2959         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
2960         sysctl_tcp_wmem[1] = 16*1024;
2961         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2962
2963         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
2964         sysctl_tcp_rmem[1] = 87380;
2965         sysctl_tcp_rmem[2] = max(87380, max_share);
2966
2967         printk(KERN_INFO "TCP: Hash tables configured "
2968                "(established %d bind %d)\n",
2969                tcp_hashinfo.ehash_size, tcp_hashinfo.bhash_size);
2970
2971         tcp_register_congestion_control(&tcp_reno);
2972 }
2973
2974 EXPORT_SYMBOL(tcp_close);
2975 EXPORT_SYMBOL(tcp_disconnect);
2976 EXPORT_SYMBOL(tcp_getsockopt);
2977 EXPORT_SYMBOL(tcp_ioctl);
2978 EXPORT_SYMBOL(tcp_poll);
2979 EXPORT_SYMBOL(tcp_read_sock);
2980 EXPORT_SYMBOL(tcp_recvmsg);
2981 EXPORT_SYMBOL(tcp_sendmsg);
2982 EXPORT_SYMBOL(tcp_splice_read);
2983 EXPORT_SYMBOL(tcp_sendpage);
2984 EXPORT_SYMBOL(tcp_setsockopt);
2985 EXPORT_SYMBOL(tcp_shutdown);