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 #include <linux/time.h>
 268 #include <linux/slab.h>
 269
 270 #include <net/icmp.h>
 271 #include <net/tcp.h>
 272 #include <net/xfrm.h>
 273 #include <net/ip.h>
 274 #include <net/netdma.h>
 275 #include <net/sock.h>
 276
 277 #include <asm/uaccess.h>
 278 #include <asm/ioctls.h>
 279
 280 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
 281
 282 struct percpu_counter tcp_orphan_count;
 283 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 284
 285 long sysctl_tcp_mem[3] __read_mostly;
 286 int sysctl_tcp_wmem[3] __read_mostly;
 287 int sysctl_tcp_rmem[3] __read_mostly;
 288
 289 EXPORT_SYMBOL(sysctl_tcp_mem);
 290 EXPORT_SYMBOL(sysctl_tcp_rmem);
 291 EXPORT_SYMBOL(sysctl_tcp_wmem);
 292
 293 atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
 294 EXPORT_SYMBOL(tcp_memory_allocated);
 295
 296 /*
 297  * Current number of TCP sockets.
 298  */
 299 struct percpu_counter tcp_sockets_allocated;
 300 EXPORT_SYMBOL(tcp_sockets_allocated);
 301
 302 /*
 303  * TCP splice context
 304  */
 305 struct tcp_splice_state {
 306         struct pipe_inode_info *pipe;
 307         size_t len;
 308         unsigned int flags;
 309 };
 310
 311 /*
 312  * Pressure flag: try to collapse.
 313  * Technical note: it is used by multiple contexts non atomically.
 314  * All the __sk_mem_schedule() is of this nature: accounting
 315  * is strict, actions are advisory and have some latency.
 316  */
 317 int tcp_memory_pressure __read_mostly;
 318 EXPORT_SYMBOL(tcp_memory_pressure);
 319
 320 void tcp_enter_memory_pressure(struct sock *sk)
 321 {
 322         if (!tcp_memory_pressure) {
 323                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
 324                 tcp_memory_pressure = 1;
 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         const struct tcp_sock *tp = tcp_sk(sk);
 378
 379         sock_poll_wait(file, sk_sleep(sk), 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 EXPORT_SYMBOL(tcp_poll);
 466
 467 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 468 {
 469         struct tcp_sock *tp = tcp_sk(sk);
 470         int answ;
 471
 472         switch (cmd) {
 473         case SIOCINQ:
 474                 if (sk->sk_state == TCP_LISTEN)
 475                         return -EINVAL;
 476
 477                 lock_sock(sk);
 478                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 479                         answ = 0;
 480                 else if (sock_flag(sk, SOCK_URGINLINE) ||
 481                          !tp->urg_data ||
 482                          before(tp->urg_seq, tp->copied_seq) ||
 483                          !before(tp->urg_seq, tp->rcv_nxt)) {
 484
 485                         answ = tp->rcv_nxt - tp->copied_seq;
 486
 487                         /* Subtract 1, if FIN was received */
 488                         if (answ && sock_flag(sk, SOCK_DONE))
 489                                 answ--;
 490                 } else
 491                         answ = tp->urg_seq - tp->copied_seq;
 492                 release_sock(sk);
 493                 break;
 494         case SIOCATMARK:
 495                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 496                 break;
 497         case SIOCOUTQ:
 498                 if (sk->sk_state == TCP_LISTEN)
 499                         return -EINVAL;
 500
 501                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 502                         answ = 0;
 503                 else
 504                         answ = tp->write_seq - tp->snd_una;
 505                 break;
 506         case SIOCOUTQNSD:
 507                 if (sk->sk_state == TCP_LISTEN)
 508                         return -EINVAL;
 509
 510                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 511                         answ = 0;
 512                 else
 513                         answ = tp->write_seq - tp->snd_nxt;
 514                 break;
 515         default:
 516                 return -ENOIOCTLCMD;
 517         }
 518
 519         return put_user(answ, (int __user *)arg);
 520 }
 521 EXPORT_SYMBOL(tcp_ioctl);
 522
 523 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 524 {
 525         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 526         tp->pushed_seq = tp->write_seq;
 527 }
 528
 529 static inline int forced_push(const struct tcp_sock *tp)
 530 {
 531         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 532 }
 533
 534 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
 535 {
 536         struct tcp_sock *tp = tcp_sk(sk);
 537         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 538
 539         skb->csum    = 0;
 540         tcb->seq     = tcb->end_seq = tp->write_seq;
 541         tcb->tcp_flags = TCPHDR_ACK;
 542         tcb->sacked  = 0;
 543         skb_header_release(skb);
 544         tcp_add_write_queue_tail(sk, skb);
 545         sk->sk_wmem_queued += skb->truesize;
 546         sk_mem_charge(sk, skb->truesize);
 547         if (tp->nonagle & TCP_NAGLE_PUSH)
 548                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 549 }
 550
 551 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
 552 {
 553         if (flags & MSG_OOB)
 554                 tp->snd_up = tp->write_seq;
 555 }
 556
 557 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
 558                             int nonagle)
 559 {
 560         if (tcp_send_head(sk)) {
 561                 struct tcp_sock *tp = tcp_sk(sk);
 562
 563                 if (!(flags & MSG_MORE) || forced_push(tp))
 564                         tcp_mark_push(tp, tcp_write_queue_tail(sk));
 565
 566                 tcp_mark_urg(tp, flags);
 567                 __tcp_push_pending_frames(sk, mss_now,
 568                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
 569         }
 570 }
 571
 572 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 573                                 unsigned int offset, size_t len)
 574 {
 575         struct tcp_splice_state *tss = rd_desc->arg.data;
 576         int ret;
 577
 578         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
 579                               tss->flags);
 580         if (ret > 0)
 581                 rd_desc->count -= ret;
 582         return ret;
 583 }
 584
 585 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 586 {
 587         /* Store TCP splice context information in read_descriptor_t. */
 588         read_descriptor_t rd_desc = {
 589                 .arg.data = tss,
 590                 .count    = tss->len,
 591         };
 592
 593         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 594 }
 595
 596 /**
 597  *  tcp_splice_read - splice data from TCP socket to a pipe
 598  * @sock:       socket to splice from
 599  * @ppos:       position (not valid)
 600  * @pipe:       pipe to splice to
 601  * @len:        number of bytes to splice
 602  * @flags:      splice modifier flags
 603  *
 604  * Description:
 605  *    Will read pages from given socket and fill them into a pipe.
 606  *
 607  **/
 608 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 609                         struct pipe_inode_info *pipe, size_t len,
 610                         unsigned int flags)
 611 {
 612         struct sock *sk = sock->sk;
 613         struct tcp_splice_state tss = {
 614                 .pipe = pipe,
 615                 .len = len,
 616                 .flags = flags,
 617         };
 618         long timeo;
 619         ssize_t spliced;
 620         int ret;
 621
 622         sock_rps_record_flow(sk);
 623         /*
 624          * We can't seek on a socket input
 625          */
 626         if (unlikely(*ppos))
 627                 return -ESPIPE;
 628
 629         ret = spliced = 0;
 630
 631         lock_sock(sk);
 632
 633         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 634         while (tss.len) {
 635                 ret = __tcp_splice_read(sk, &tss);
 636                 if (ret < 0)
 637                         break;
 638                 else if (!ret) {
 639                         if (spliced)
 640                                 break;
 641                         if (sock_flag(sk, SOCK_DONE))
 642                                 break;
 643                         if (sk->sk_err) {
 644                                 ret = sock_error(sk);
 645                                 break;
 646                         }
 647                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 648                                 break;
 649                         if (sk->sk_state == TCP_CLOSE) {
 650                                 /*
 651                                  * This occurs when user tries to read
 652                                  * from never connected socket.
 653                                  */
 654                                 if (!sock_flag(sk, SOCK_DONE))
 655                                         ret = -ENOTCONN;
 656                                 break;
 657                         }
 658                         if (!timeo) {
 659                                 ret = -EAGAIN;
 660                                 break;
 661                         }
 662                         /* if __tcp_splice_read() got nothing while we have
 663                          * an skb in receive queue, we do not want to loop.
 664                          * This might happen with URG data.
 665                          */
 666                         if (!skb_queue_empty(&sk->sk_receive_queue))
 667                                 break;
 668                         sk_wait_data(sk, &timeo);
 669                         if (signal_pending(current)) {
 670                                 ret = sock_intr_errno(timeo);
 671                                 break;
 672                         }
 673                         continue;
 674                 }
 675                 tss.len -= ret;
 676                 spliced += ret;
 677
 678                 if (!timeo)
 679                         break;
 680                 release_sock(sk);
 681                 lock_sock(sk);
 682
 683                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 684                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 685                     signal_pending(current))
 686                         break;
 687         }
 688
 689         release_sock(sk);
 690
 691         if (spliced)
 692                 return spliced;
 693
 694         return ret;
 695 }
 696 EXPORT_SYMBOL(tcp_splice_read);
 697
 698 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
 699 {
 700         struct sk_buff *skb;
 701
 702         /* The TCP header must be at least 32-bit aligned.  */
 703         size = ALIGN(size, 4);
 704
 705         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 706         if (skb) {
 707                 if (sk_wmem_schedule(sk, skb->truesize)) {
 708                         skb_reserve(skb, sk->sk_prot->max_header);
 709                         /*
 710                          * Make sure that we have exactly size bytes
 711                          * available to the caller, no more, no less.
 712                          */
 713                         skb->reserved_tailroom = skb->end - skb->tail - size;
 714                         return skb;
 715                 }
 716                 __kfree_skb(skb);
 717         } else {
 718                 sk->sk_prot->enter_memory_pressure(sk);
 719                 sk_stream_moderate_sndbuf(sk);
 720         }
 721         return NULL;
 722 }
 723
 724 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 725                                        int large_allowed)
 726 {
 727         struct tcp_sock *tp = tcp_sk(sk);
 728         u32 xmit_size_goal, old_size_goal;
 729
 730         xmit_size_goal = mss_now;
 731
 732         if (large_allowed && sk_can_gso(sk)) {
 733                 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
 734                                   inet_csk(sk)->icsk_af_ops->net_header_len -
 735                                   inet_csk(sk)->icsk_ext_hdr_len -
 736                                   tp->tcp_header_len);
 737
 738                 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
 739
 740                 /* We try hard to avoid divides here */
 741                 old_size_goal = tp->xmit_size_goal_segs * mss_now;
 742
 743                 if (likely(old_size_goal <= xmit_size_goal &&
 744                            old_size_goal + mss_now > xmit_size_goal)) {
 745                         xmit_size_goal = old_size_goal;
 746                 } else {
 747                         tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
 748                         xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
 749                 }
 750         }
 751
 752         return max(xmit_size_goal, mss_now);
 753 }
 754
 755 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 756 {
 757         int mss_now;
 758
 759         mss_now = tcp_current_mss(sk);
 760         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 761
 762         return mss_now;
 763 }
 764
 765 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
 766                          size_t psize, int flags)
 767 {
 768         struct tcp_sock *tp = tcp_sk(sk);
 769         int mss_now, size_goal;
 770         int err;
 771         ssize_t copied;
 772         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 773
 774         /* Wait for a connection to finish. */
 775         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 776                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 777                         goto out_err;
 778
 779         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 780
 781         mss_now = tcp_send_mss(sk, &size_goal, flags);
 782         copied = 0;
 783
 784         err = -EPIPE;
 785         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 786                 goto out_err;
 787
 788         while (psize > 0) {
 789                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 790                 struct page *page = pages[poffset / PAGE_SIZE];
 791                 int copy, i, can_coalesce;
 792                 int offset = poffset % PAGE_SIZE;
 793                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
 794
 795                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
 796 new_segment:
 797                         if (!sk_stream_memory_free(sk))
 798                                 goto wait_for_sndbuf;
 799
 800                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
 801                         if (!skb)
 802                                 goto wait_for_memory;
 803
 804                         skb_entail(sk, skb);
 805                         copy = size_goal;
 806                 }
 807
 808                 if (copy > size)
 809                         copy = size;
 810
 811                 i = skb_shinfo(skb)->nr_frags;
 812                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 813                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
 814                         tcp_mark_push(tp, skb);
 815                         goto new_segment;
 816                 }
 817                 if (!sk_wmem_schedule(sk, copy))
 818                         goto wait_for_memory;
 819
 820                 if (can_coalesce) {
 821                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
 822                 } else {
 823                         get_page(page);
 824                         skb_fill_page_desc(skb, i, page, offset, copy);
 825                 }
 826
 827                 skb->len += copy;
 828                 skb->data_len += copy;
 829                 skb->truesize += copy;
 830                 sk->sk_wmem_queued += copy;
 831                 sk_mem_charge(sk, copy);
 832                 skb->ip_summed = CHECKSUM_PARTIAL;
 833                 tp->write_seq += copy;
 834                 TCP_SKB_CB(skb)->end_seq += copy;
 835                 skb_shinfo(skb)->gso_segs = 0;
 836
 837                 if (!copied)
 838                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
 839
 840                 copied += copy;
 841                 poffset += copy;
 842                 if (!(psize -= copy))
 843                         goto out;
 844
 845                 if (skb->len < size_goal || (flags & MSG_OOB))
 846                         continue;
 847
 848                 if (forced_push(tp)) {
 849                         tcp_mark_push(tp, skb);
 850                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
 851                 } else if (skb == tcp_send_head(sk))
 852                         tcp_push_one(sk, mss_now);
 853                 continue;
 854
 855 wait_for_sndbuf:
 856                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 857 wait_for_memory:
 858                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
 859
 860                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 861                         goto do_error;
 862
 863                 mss_now = tcp_send_mss(sk, &size_goal, flags);
 864         }
 865
 866 out:
 867         if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
 868                 tcp_push(sk, flags, mss_now, tp->nonagle);
 869         return copied;
 870
 871 do_error:
 872         if (copied)
 873                 goto out;
 874 out_err:
 875         return sk_stream_error(sk, flags, err);
 876 }
 877
 878 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
 879                  size_t size, int flags)
 880 {
 881         ssize_t res;
 882
 883         if (!(sk->sk_route_caps & NETIF_F_SG) ||
 884             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
 885                 return sock_no_sendpage(sk->sk_socket, page, offset, size,
 886                                         flags);
 887
 888         lock_sock(sk);
 889         res = do_tcp_sendpages(sk, &page, offset, size, flags);
 890         release_sock(sk);
 891         return res;
 892 }
 893 EXPORT_SYMBOL(tcp_sendpage);
 894
 895 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
 896 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
 897
 898 static inline int select_size(const struct sock *sk, int sg)
 899 {
 900         const struct tcp_sock *tp = tcp_sk(sk);
 901         int tmp = tp->mss_cache;
 902
 903         if (sg) {
 904                 if (sk_can_gso(sk))
 905                         tmp = 0;
 906                 else {
 907                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
 908
 909                         if (tmp >= pgbreak &&
 910                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
 911                                 tmp = pgbreak;
 912                 }
 913         }
 914
 915         return tmp;
 916 }
 917
 918 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
 919                 size_t size)
 920 {
 921         struct iovec *iov;
 922         struct tcp_sock *tp = tcp_sk(sk);
 923         struct sk_buff *skb;
 924         int iovlen, flags;
 925         int mss_now, size_goal;
 926         int sg, err, copied;
 927         long timeo;
 928
 929         lock_sock(sk);
 930
 931         flags = msg->msg_flags;
 932         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 933
 934         /* Wait for a connection to finish. */
 935         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 936                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 937                         goto out_err;
 938
 939         /* This should be in poll */
 940         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 941
 942         mss_now = tcp_send_mss(sk, &size_goal, flags);
 943
 944         /* Ok commence sending. */
 945         iovlen = msg->msg_iovlen;
 946         iov = msg->msg_iov;
 947         copied = 0;
 948
 949         err = -EPIPE;
 950         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 951                 goto out_err;
 952
 953         sg = sk->sk_route_caps & NETIF_F_SG;
 954
 955         while (--iovlen >= 0) {
 956                 size_t seglen = iov->iov_len;
 957                 unsigned char __user *from = iov->iov_base;
 958
 959                 iov++;
 960
 961                 while (seglen > 0) {
 962                         int copy = 0;
 963                         int max = size_goal;
 964
 965                         skb = tcp_write_queue_tail(sk);
 966                         if (tcp_send_head(sk)) {
 967                                 if (skb->ip_summed == CHECKSUM_NONE)
 968                                         max = mss_now;
 969                                 copy = max - skb->len;
 970                         }
 971
 972                         if (copy <= 0) {
 973 new_segment:
 974                                 /* Allocate new segment. If the interface is SG,
 975                                  * allocate skb fitting to single page.
 976                                  */
 977                                 if (!sk_stream_memory_free(sk))
 978                                         goto wait_for_sndbuf;
 979
 980                                 skb = sk_stream_alloc_skb(sk,
 981                                                           select_size(sk, sg),
 982                                                           sk->sk_allocation);
 983                                 if (!skb)
 984                                         goto wait_for_memory;
 985
 986                                 /*
 987                                  * Check whether we can use HW checksum.
 988                                  */
 989                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
 990                                         skb->ip_summed = CHECKSUM_PARTIAL;
 991
 992                                 skb_entail(sk, skb);
 993                                 copy = size_goal;
 994                                 max = size_goal;
 995                         }
 996
 997                         /* Try to append data to the end of skb. */
 998                         if (copy > seglen)
 999                                 copy = seglen;
1000
1001                         /* Where to copy to? */
1002                         if (skb_availroom(skb) > 0) {
1003                                 /* We have some space in skb head. Superb! */
1004                                 copy = min_t(int, copy, skb_availroom(skb));
1005                                 err = skb_add_data_nocache(sk, skb, from, copy);
1006                                 if (err)
1007                                         goto do_fault;
1008                         } else {
1009                                 int merge = 0;
1010                                 int i = skb_shinfo(skb)->nr_frags;
1011                                 struct page *page = TCP_PAGE(sk);
1012                                 int off = TCP_OFF(sk);
1013
1014                                 if (skb_can_coalesce(skb, i, page, off) &&
1015                                     off != PAGE_SIZE) {
1016                                         /* We can extend the last page
1017                                          * fragment. */
1018                                         merge = 1;
1019                                 } else if (i == MAX_SKB_FRAGS || !sg) {
1020                                         /* Need to add new fragment and cannot
1021                                          * do this because interface is non-SG,
1022                                          * or because all the page slots are
1023                                          * busy. */
1024                                         tcp_mark_push(tp, skb);
1025                                         goto new_segment;
1026                                 } else if (page) {
1027                                         if (off == PAGE_SIZE) {
1028                                                 put_page(page);
1029                                                 TCP_PAGE(sk) = page = NULL;
1030                                                 off = 0;
1031                                         }
1032                                 } else
1033                                         off = 0;
1034
1035                                 if (copy > PAGE_SIZE - off)
1036                                         copy = PAGE_SIZE - off;
1037
1038                                 if (!sk_wmem_schedule(sk, copy))
1039                                         goto wait_for_memory;
1040
1041                                 if (!page) {
1042                                         /* Allocate new cache page. */
1043                                         if (!(page = sk_stream_alloc_page(sk)))
1044                                                 goto wait_for_memory;
1045                                 }
1046
1047                                 /* Time to copy data. We are close to
1048                                  * the end! */
1049                                 err = skb_copy_to_page_nocache(sk, from, skb,
1050                                                                page, off, copy);
1051                                 if (err) {
1052                                         /* If this page was new, give it to the
1053                                          * socket so it does not get leaked.
1054                                          */
1055                                         if (!TCP_PAGE(sk)) {
1056                                                 TCP_PAGE(sk) = page;
1057                                                 TCP_OFF(sk) = 0;
1058                                         }
1059                                         goto do_error;
1060                                 }
1061
1062                                 /* Update the skb. */
1063                                 if (merge) {
1064                                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1065                                 } else {
1066                                         skb_fill_page_desc(skb, i, page, off, copy);
1067                                         if (TCP_PAGE(sk)) {
1068                                                 get_page(page);
1069                                         } else if (off + copy < PAGE_SIZE) {
1070                                                 get_page(page);
1071                                                 TCP_PAGE(sk) = page;
1072                                         }
1073                                 }
1074
1075                                 TCP_OFF(sk) = off + copy;
1076                         }
1077
1078                         if (!copied)
1079                                 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1080
1081                         tp->write_seq += copy;
1082                         TCP_SKB_CB(skb)->end_seq += copy;
1083                         skb_shinfo(skb)->gso_segs = 0;
1084
1085                         from += copy;
1086                         copied += copy;
1087                         if ((seglen -= copy) == 0 && iovlen == 0)
1088                                 goto out;
1089
1090                         if (skb->len < max || (flags & MSG_OOB))
1091                                 continue;
1092
1093                         if (forced_push(tp)) {
1094                                 tcp_mark_push(tp, skb);
1095                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1096                         } else if (skb == tcp_send_head(sk))
1097                                 tcp_push_one(sk, mss_now);
1098                         continue;
1099
1100 wait_for_sndbuf:
1101                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1102 wait_for_memory:
1103                         if (copied)
1104                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1105
1106                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1107                                 goto do_error;
1108
1109                         mss_now = tcp_send_mss(sk, &size_goal, flags);
1110                 }
1111         }
1112
1113 out:
1114         if (copied)
1115                 tcp_push(sk, flags, mss_now, tp->nonagle);
1116         release_sock(sk);
1117         return copied;
1118
1119 do_fault:
1120         if (!skb->len) {
1121                 tcp_unlink_write_queue(skb, sk);
1122                 /* It is the one place in all of TCP, except connection
1123                  * reset, where we can be unlinking the send_head.
1124                  */
1125                 tcp_check_send_head(sk, skb);
1126                 sk_wmem_free_skb(sk, skb);
1127         }
1128
1129 do_error:
1130         if (copied)
1131                 goto out;
1132 out_err:
1133         err = sk_stream_error(sk, flags, err);
1134         release_sock(sk);
1135         return err;
1136 }
1137 EXPORT_SYMBOL(tcp_sendmsg);
1138
1139 /*
1140  *      Handle reading urgent data. BSD has very simple semantics for
1141  *      this, no blocking and very strange errors 8)
1142  */
1143
1144 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1145 {
1146         struct tcp_sock *tp = tcp_sk(sk);
1147
1148         /* No URG data to read. */
1149         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1150             tp->urg_data == TCP_URG_READ)
1151                 return -EINVAL; /* Yes this is right ! */
1152
1153         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1154                 return -ENOTCONN;
1155
1156         if (tp->urg_data & TCP_URG_VALID) {
1157                 int err = 0;
1158                 char c = tp->urg_data;
1159
1160                 if (!(flags & MSG_PEEK))
1161                         tp->urg_data = TCP_URG_READ;
1162
1163                 /* Read urgent data. */
1164                 msg->msg_flags |= MSG_OOB;
1165
1166                 if (len > 0) {
1167                         if (!(flags & MSG_TRUNC))
1168                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1169                         len = 1;
1170                 } else
1171                         msg->msg_flags |= MSG_TRUNC;
1172
1173                 return err ? -EFAULT : len;
1174         }
1175
1176         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1177                 return 0;
1178
1179         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1180          * the available implementations agree in this case:
1181          * this call should never block, independent of the
1182          * blocking state of the socket.
1183          * Mike <pall@rz.uni-karlsruhe.de>
1184          */
1185         return -EAGAIN;
1186 }
1187
1188 /* Clean up the receive buffer for full frames taken by the user,
1189  * then send an ACK if necessary.  COPIED is the number of bytes
1190  * tcp_recvmsg has given to the user so far, it speeds up the
1191  * calculation of whether or not we must ACK for the sake of
1192  * a window update.
1193  */
1194 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1195 {
1196         struct tcp_sock *tp = tcp_sk(sk);
1197         int time_to_ack = 0;
1198
1199         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1200
1201         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1202              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1203              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1204
1205         if (inet_csk_ack_scheduled(sk)) {
1206                 const struct inet_connection_sock *icsk = inet_csk(sk);
1207                    /* Delayed ACKs frequently hit locked sockets during bulk
1208                     * receive. */
1209                 if (icsk->icsk_ack.blocked ||
1210                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1211                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1212                     /*
1213                      * If this read emptied read buffer, we send ACK, if
1214                      * connection is not bidirectional, user drained
1215                      * receive buffer and there was a small segment
1216                      * in queue.
1217                      */
1218                     (copied > 0 &&
1219                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1220                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1221                        !icsk->icsk_ack.pingpong)) &&
1222                       !atomic_read(&sk->sk_rmem_alloc)))
1223                         time_to_ack = 1;
1224         }
1225
1226         /* We send an ACK if we can now advertise a non-zero window
1227          * which has been raised "significantly".
1228          *
1229          * Even if window raised up to infinity, do not send window open ACK
1230          * in states, where we will not receive more. It is useless.
1231          */
1232         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1233                 __u32 rcv_window_now = tcp_receive_window(tp);
1234
1235                 /* Optimize, __tcp_select_window() is not cheap. */
1236                 if (2*rcv_window_now <= tp->window_clamp) {
1237                         __u32 new_window = __tcp_select_window(sk);
1238
1239                         /* Send ACK now, if this read freed lots of space
1240                          * in our buffer. Certainly, new_window is new window.
1241                          * We can advertise it now, if it is not less than current one.
1242                          * "Lots" means "at least twice" here.
1243                          */
1244                         if (new_window && new_window >= 2 * rcv_window_now)
1245                                 time_to_ack = 1;
1246                 }
1247         }
1248         if (time_to_ack)
1249                 tcp_send_ack(sk);
1250 }
1251
1252 static void tcp_prequeue_process(struct sock *sk)
1253 {
1254         struct sk_buff *skb;
1255         struct tcp_sock *tp = tcp_sk(sk);
1256
1257         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1258
1259         /* RX process wants to run with disabled BHs, though it is not
1260          * necessary */
1261         local_bh_disable();
1262         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1263                 sk_backlog_rcv(sk, skb);
1264         local_bh_enable();
1265
1266         /* Clear memory counter. */
1267         tp->ucopy.memory = 0;
1268 }
1269
1270 #ifdef CONFIG_NET_DMA
1271 static void tcp_service_net_dma(struct sock *sk, bool wait)
1272 {
1273         dma_cookie_t done, used;
1274         dma_cookie_t last_issued;
1275         struct tcp_sock *tp = tcp_sk(sk);
1276
1277         if (!tp->ucopy.dma_chan)
1278                 return;
1279
1280         last_issued = tp->ucopy.dma_cookie;
1281         dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1282
1283         do {
1284                 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1285                                               last_issued, &done,
1286                                               &used) == DMA_SUCCESS) {
1287                         /* Safe to free early-copied skbs now */
1288                         __skb_queue_purge(&sk->sk_async_wait_queue);
1289                         break;
1290                 } else {
1291                         struct sk_buff *skb;
1292                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1293                                (dma_async_is_complete(skb->dma_cookie, done,
1294                                                       used) == DMA_SUCCESS)) {
1295                                 __skb_dequeue(&sk->sk_async_wait_queue);
1296                                 kfree_skb(skb);
1297                         }
1298                 }
1299         } while (wait);
1300 }
1301 #endif
1302
1303 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1304 {
1305         struct sk_buff *skb;
1306         u32 offset;
1307
1308         skb_queue_walk(&sk->sk_receive_queue, skb) {
1309                 offset = seq - TCP_SKB_CB(skb)->seq;
1310                 if (tcp_hdr(skb)->syn)
1311                         offset--;
1312                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1313                         *off = offset;
1314                         return skb;
1315                 }
1316         }
1317         return NULL;
1318 }
1319
1320 /*
1321  * This routine provides an alternative to tcp_recvmsg() for routines
1322  * that would like to handle copying from skbuffs directly in 'sendfile'
1323  * fashion.
1324  * Note:
1325  *      - It is assumed that the socket was locked by the caller.
1326  *      - The routine does not block.
1327  *      - At present, there is no support for reading OOB data
1328  *        or for 'peeking' the socket using this routine
1329  *        (although both would be easy to implement).
1330  */
1331 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1332                   sk_read_actor_t recv_actor)
1333 {
1334         struct sk_buff *skb;
1335         struct tcp_sock *tp = tcp_sk(sk);
1336         u32 seq = tp->copied_seq;
1337         u32 offset;
1338         int copied = 0;
1339
1340         if (sk->sk_state == TCP_LISTEN)
1341                 return -ENOTCONN;
1342         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1343                 if (offset < skb->len) {
1344                         int used;
1345                         size_t len;
1346
1347                         len = skb->len - offset;
1348                         /* Stop reading if we hit a patch of urgent data */
1349                         if (tp->urg_data) {
1350                                 u32 urg_offset = tp->urg_seq - seq;
1351                                 if (urg_offset < len)
1352                                         len = urg_offset;
1353                                 if (!len)
1354                                         break;
1355                         }
1356                         used = recv_actor(desc, skb, offset, len);
1357                         if (used < 0) {
1358                                 if (!copied)
1359                                         copied = used;
1360                                 break;
1361                         } else if (used <= len) {
1362                                 seq += used;
1363                                 copied += used;
1364                                 offset += used;
1365                         }
1366                         /*
1367                          * If recv_actor drops the lock (e.g. TCP splice
1368                          * receive) the skb pointer might be invalid when
1369                          * getting here: tcp_collapse might have deleted it
1370                          * while aggregating skbs from the socket queue.
1371                          */
1372                         skb = tcp_recv_skb(sk, seq-1, &offset);
1373                         if (!skb || (offset+1 != skb->len))
1374                                 break;
1375                 }
1376                 if (tcp_hdr(skb)->fin) {
1377                         sk_eat_skb(sk, skb, 0);
1378                         ++seq;
1379                         break;
1380                 }
1381                 sk_eat_skb(sk, skb, 0);
1382                 if (!desc->count)
1383                         break;
1384                 tp->copied_seq = seq;
1385         }
1386         tp->copied_seq = seq;
1387
1388         tcp_rcv_space_adjust(sk);
1389
1390         /* Clean up data we have read: This will do ACK frames. */
1391         if (copied > 0)
1392                 tcp_cleanup_rbuf(sk, copied);
1393         return copied;
1394 }
1395 EXPORT_SYMBOL(tcp_read_sock);
1396
1397 /*
1398  *      This routine copies from a sock struct into the user buffer.
1399  *
1400  *      Technical note: in 2.3 we work on _locked_ socket, so that
1401  *      tricks with *seq access order and skb->users are not required.
1402  *      Probably, code can be easily improved even more.
1403  */
1404
1405 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1406                 size_t len, int nonblock, int flags, int *addr_len)
1407 {
1408         struct tcp_sock *tp = tcp_sk(sk);
1409         int copied = 0;
1410         u32 peek_seq;
1411         u32 *seq;
1412         unsigned long used;
1413         int err;
1414         int target;             /* Read at least this many bytes */
1415         long timeo;
1416         struct task_struct *user_recv = NULL;
1417         int copied_early = 0;
1418         struct sk_buff *skb;
1419         u32 urg_hole = 0;
1420
1421         lock_sock(sk);
1422
1423         err = -ENOTCONN;
1424         if (sk->sk_state == TCP_LISTEN)
1425                 goto out;
1426
1427         timeo = sock_rcvtimeo(sk, nonblock);
1428
1429         /* Urgent data needs to be handled specially. */
1430         if (flags & MSG_OOB)
1431                 goto recv_urg;
1432
1433         seq = &tp->copied_seq;
1434         if (flags & MSG_PEEK) {
1435                 peek_seq = tp->copied_seq;
1436                 seq = &peek_seq;
1437         }
1438
1439         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1440
1441 #ifdef CONFIG_NET_DMA
1442         tp->ucopy.dma_chan = NULL;
1443         preempt_disable();
1444         skb = skb_peek_tail(&sk->sk_receive_queue);
1445         {
1446                 int available = 0;
1447
1448                 if (skb)
1449                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1450                 if ((available < target) &&
1451                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1452                     !sysctl_tcp_low_latency &&
1453                     dma_find_channel(DMA_MEMCPY)) {
1454                         preempt_enable_no_resched();
1455                         tp->ucopy.pinned_list =
1456                                         dma_pin_iovec_pages(msg->msg_iov, len);
1457                 } else {
1458                         preempt_enable_no_resched();
1459                 }
1460         }
1461 #endif
1462
1463         do {
1464                 u32 offset;
1465
1466                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1467                 if (tp->urg_data && tp->urg_seq == *seq) {
1468                         if (copied)
1469                                 break;
1470                         if (signal_pending(current)) {
1471                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1472                                 break;
1473                         }
1474                 }
1475
1476                 /* Next get a buffer. */
1477
1478                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1479                         /* Now that we have two receive queues this
1480                          * shouldn't happen.
1481                          */
1482                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1483                                  "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1484                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1485                                  flags))
1486                                 break;
1487
1488                         offset = *seq - TCP_SKB_CB(skb)->seq;
1489                         if (tcp_hdr(skb)->syn)
1490                                 offset--;
1491                         if (offset < skb->len)
1492                                 goto found_ok_skb;
1493                         if (tcp_hdr(skb)->fin)
1494                                 goto found_fin_ok;
1495                         WARN(!(flags & MSG_PEEK),
1496                              "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1497                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1498                 }
1499
1500                 /* Well, if we have backlog, try to process it now yet. */
1501
1502                 if (copied >= target && !sk->sk_backlog.tail)
1503                         break;
1504
1505                 if (copied) {
1506                         if (sk->sk_err ||
1507                             sk->sk_state == TCP_CLOSE ||
1508                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1509                             !timeo ||
1510                             signal_pending(current))
1511                                 break;
1512                 } else {
1513                         if (sock_flag(sk, SOCK_DONE))
1514                                 break;
1515
1516                         if (sk->sk_err) {
1517                                 copied = sock_error(sk);
1518                                 break;
1519                         }
1520
1521                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1522                                 break;
1523
1524                         if (sk->sk_state == TCP_CLOSE) {
1525                                 if (!sock_flag(sk, SOCK_DONE)) {
1526                                         /* This occurs when user tries to read
1527                                          * from never connected socket.
1528                                          */
1529                                         copied = -ENOTCONN;
1530                                         break;
1531                                 }
1532                                 break;
1533                         }
1534
1535                         if (!timeo) {
1536                                 copied = -EAGAIN;
1537                                 break;
1538                         }
1539
1540                         if (signal_pending(current)) {
1541                                 copied = sock_intr_errno(timeo);
1542                                 break;
1543                         }
1544                 }
1545
1546                 tcp_cleanup_rbuf(sk, copied);
1547
1548                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1549                         /* Install new reader */
1550                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1551                                 user_recv = current;
1552                                 tp->ucopy.task = user_recv;
1553                                 tp->ucopy.iov = msg->msg_iov;
1554                         }
1555
1556                         tp->ucopy.len = len;
1557
1558                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1559                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1560
1561                         /* Ugly... If prequeue is not empty, we have to
1562                          * process it before releasing socket, otherwise
1563                          * order will be broken at second iteration.
1564                          * More elegant solution is required!!!
1565                          *
1566                          * Look: we have the following (pseudo)queues:
1567                          *
1568                          * 1. packets in flight
1569                          * 2. backlog
1570                          * 3. prequeue
1571                          * 4. receive_queue
1572                          *
1573                          * Each queue can be processed only if the next ones
1574                          * are empty. At this point we have empty receive_queue.
1575                          * But prequeue _can_ be not empty after 2nd iteration,
1576                          * when we jumped to start of loop because backlog
1577                          * processing added something to receive_queue.
1578                          * We cannot release_sock(), because backlog contains
1579                          * packets arrived _after_ prequeued ones.
1580                          *
1581                          * Shortly, algorithm is clear --- to process all
1582                          * the queues in order. We could make it more directly,
1583                          * requeueing packets from backlog to prequeue, if
1584                          * is not empty. It is more elegant, but eats cycles,
1585                          * unfortunately.
1586                          */
1587                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1588                                 goto do_prequeue;
1589
1590                         /* __ Set realtime policy in scheduler __ */
1591                 }
1592
1593 #ifdef CONFIG_NET_DMA
1594                 if (tp->ucopy.dma_chan) {
1595                         if (tp->rcv_wnd == 0 &&
1596                             !skb_queue_empty(&sk->sk_async_wait_queue)) {
1597                                 tcp_service_net_dma(sk, true);
1598                                 tcp_cleanup_rbuf(sk, copied);
1599                         } else
1600                                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1601                 }
1602 #endif
1603                 if (copied >= target) {
1604                         /* Do not sleep, just process backlog. */
1605                         release_sock(sk);
1606                         lock_sock(sk);
1607                 } else
1608                         sk_wait_data(sk, &timeo);
1609
1610 #ifdef CONFIG_NET_DMA
1611                 tcp_service_net_dma(sk, false);  /* Don't block */
1612                 tp->ucopy.wakeup = 0;
1613 #endif
1614
1615                 if (user_recv) {
1616                         int chunk;
1617
1618                         /* __ Restore normal policy in scheduler __ */
1619
1620                         if ((chunk = len - tp->ucopy.len) != 0) {
1621                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1622                                 len -= chunk;
1623                                 copied += chunk;
1624                         }
1625
1626                         if (tp->rcv_nxt == tp->copied_seq &&
1627                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1628 do_prequeue:
1629                                 tcp_prequeue_process(sk);
1630
1631                                 if ((chunk = len - tp->ucopy.len) != 0) {
1632                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1633                                         len -= chunk;
1634                                         copied += chunk;
1635                                 }
1636                         }
1637                 }
1638                 if ((flags & MSG_PEEK) &&
1639                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1640                         if (net_ratelimit())
1641                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1642                                        current->comm, task_pid_nr(current));
1643                         peek_seq = tp->copied_seq;
1644                 }
1645                 continue;
1646
1647         found_ok_skb:
1648                 /* Ok so how much can we use? */
1649                 used = skb->len - offset;
1650                 if (len < used)
1651                         used = len;
1652
1653                 /* Do we have urgent data here? */
1654                 if (tp->urg_data) {
1655                         u32 urg_offset = tp->urg_seq - *seq;
1656                         if (urg_offset < used) {
1657                                 if (!urg_offset) {
1658                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1659                                                 ++*seq;
1660                                                 urg_hole++;
1661                                                 offset++;
1662                                                 used--;
1663                                                 if (!used)
1664                                                         goto skip_copy;
1665                                         }
1666                                 } else
1667                                         used = urg_offset;
1668                         }
1669                 }
1670
1671                 if (!(flags & MSG_TRUNC)) {
1672 #ifdef CONFIG_NET_DMA
1673                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1674                                 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1675
1676                         if (tp->ucopy.dma_chan) {
1677                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1678                                         tp->ucopy.dma_chan, skb, offset,
1679                                         msg->msg_iov, used,
1680                                         tp->ucopy.pinned_list);
1681
1682                                 if (tp->ucopy.dma_cookie < 0) {
1683
1684                                         printk(KERN_ALERT "dma_cookie < 0\n");
1685
1686                                         /* Exception. Bailout! */
1687                                         if (!copied)
1688                                                 copied = -EFAULT;
1689                                         break;
1690                                 }
1691
1692                                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1693
1694                                 if ((offset + used) == skb->len)
1695                                         copied_early = 1;
1696
1697                         } else
1698 #endif
1699                         {
1700                                 err = skb_copy_datagram_iovec(skb, offset,
1701                                                 msg->msg_iov, used);
1702                                 if (err) {
1703                                         /* Exception. Bailout! */
1704                                         if (!copied)
1705                                                 copied = -EFAULT;
1706                                         break;
1707                                 }
1708                         }
1709                 }
1710
1711                 *seq += used;
1712                 copied += used;
1713                 len -= used;
1714
1715                 tcp_rcv_space_adjust(sk);
1716
1717 skip_copy:
1718                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1719                         tp->urg_data = 0;
1720                         tcp_fast_path_check(sk);
1721                 }
1722                 if (used + offset < skb->len)
1723                         continue;
1724
1725                 if (tcp_hdr(skb)->fin)
1726                         goto found_fin_ok;
1727                 if (!(flags & MSG_PEEK)) {
1728                         sk_eat_skb(sk, skb, copied_early);
1729                         copied_early = 0;
1730                 }
1731                 continue;
1732
1733         found_fin_ok:
1734                 /* Process the FIN. */
1735                 ++*seq;
1736                 if (!(flags & MSG_PEEK)) {
1737                         sk_eat_skb(sk, skb, copied_early);
1738                         copied_early = 0;
1739                 }
1740                 break;
1741         } while (len > 0);
1742
1743         if (user_recv) {
1744                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1745                         int chunk;
1746
1747                         tp->ucopy.len = copied > 0 ? len : 0;
1748
1749                         tcp_prequeue_process(sk);
1750
1751                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1752                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1753                                 len -= chunk;
1754                                 copied += chunk;
1755                         }
1756                 }
1757
1758                 tp->ucopy.task = NULL;
1759                 tp->ucopy.len = 0;
1760         }
1761
1762 #ifdef CONFIG_NET_DMA
1763         tcp_service_net_dma(sk, true);  /* Wait for queue to drain */
1764         tp->ucopy.dma_chan = NULL;
1765
1766         if (tp->ucopy.pinned_list) {
1767                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1768                 tp->ucopy.pinned_list = NULL;
1769         }
1770 #endif
1771
1772         /* According to UNIX98, msg_name/msg_namelen are ignored
1773          * on connected socket. I was just happy when found this 8) --ANK
1774          */
1775
1776         /* Clean up data we have read: This will do ACK frames. */
1777         tcp_cleanup_rbuf(sk, copied);
1778
1779         release_sock(sk);
1780         return copied;
1781
1782 out:
1783         release_sock(sk);
1784         return err;
1785
1786 recv_urg:
1787         err = tcp_recv_urg(sk, msg, len, flags);
1788         goto out;
1789 }
1790 EXPORT_SYMBOL(tcp_recvmsg);
1791
1792 void tcp_set_state(struct sock *sk, int state)
1793 {
1794         int oldstate = sk->sk_state;
1795
1796         switch (state) {
1797         case TCP_ESTABLISHED:
1798                 if (oldstate != TCP_ESTABLISHED)
1799                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1800                 break;
1801
1802         case TCP_CLOSE:
1803                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1804                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1805
1806                 sk->sk_prot->unhash(sk);
1807                 if (inet_csk(sk)->icsk_bind_hash &&
1808                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1809                         inet_put_port(sk);
1810                 /* fall through */
1811         default:
1812                 if (oldstate == TCP_ESTABLISHED)
1813                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1814         }
1815
1816         /* Change state AFTER socket is unhashed to avoid closed
1817          * socket sitting in hash tables.
1818          */
1819         sk->sk_state = state;
1820
1821 #ifdef STATE_TRACE
1822         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1823 #endif
1824 }
1825 EXPORT_SYMBOL_GPL(tcp_set_state);
1826
1827 /*
1828  *      State processing on a close. This implements the state shift for
1829  *      sending our FIN frame. Note that we only send a FIN for some
1830  *      states. A shutdown() may have already sent the FIN, or we may be
1831  *      closed.
1832  */
1833
1834 static const unsigned char new_state[16] = {
1835   /* current state:        new state:      action:      */
1836   /* (Invalid)          */ TCP_CLOSE,
1837   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1838   /* TCP_SYN_SENT       */ TCP_CLOSE,
1839   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1840   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1841   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1842   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1843   /* TCP_CLOSE          */ TCP_CLOSE,
1844   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1845   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1846   /* TCP_LISTEN         */ TCP_CLOSE,
1847   /* TCP_CLOSING        */ TCP_CLOSING,
1848 };
1849
1850 static int tcp_close_state(struct sock *sk)
1851 {
1852         int next = (int)new_state[sk->sk_state];
1853         int ns = next & TCP_STATE_MASK;
1854
1855         tcp_set_state(sk, ns);
1856
1857         return next & TCP_ACTION_FIN;
1858 }
1859
1860 /*
1861  *      Shutdown the sending side of a connection. Much like close except
1862  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1863  */
1864
1865 void tcp_shutdown(struct sock *sk, int how)
1866 {
1867         /*      We need to grab some memory, and put together a FIN,
1868          *      and then put it into the queue to be sent.
1869          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1870          */
1871         if (!(how & SEND_SHUTDOWN))
1872                 return;
1873
1874         /* If we've already sent a FIN, or it's a closed state, skip this. */
1875         if ((1 << sk->sk_state) &
1876             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1877              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1878                 /* Clear out any half completed packets.  FIN if needed. */
1879                 if (tcp_close_state(sk))
1880                         tcp_send_fin(sk);
1881         }
1882 }
1883 EXPORT_SYMBOL(tcp_shutdown);
1884
1885 void tcp_close(struct sock *sk, long timeout)
1886 {
1887         struct sk_buff *skb;
1888         int data_was_unread = 0;
1889         int state;
1890
1891         lock_sock(sk);
1892         sk->sk_shutdown = SHUTDOWN_MASK;
1893
1894         if (sk->sk_state == TCP_LISTEN) {
1895                 tcp_set_state(sk, TCP_CLOSE);
1896
1897                 /* Special case. */
1898                 inet_csk_listen_stop(sk);
1899
1900                 goto adjudge_to_death;
1901         }
1902
1903         /*  We need to flush the recv. buffs.  We do this only on the
1904          *  descriptor close, not protocol-sourced closes, because the
1905          *  reader process may not have drained the data yet!
1906          */
1907         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1908                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1909                           tcp_hdr(skb)->fin;
1910                 data_was_unread += len;
1911                 __kfree_skb(skb);
1912         }
1913
1914         sk_mem_reclaim(sk);
1915
1916         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
1917         if (sk->sk_state == TCP_CLOSE)
1918                 goto adjudge_to_death;
1919
1920         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1921          * data was lost. To witness the awful effects of the old behavior of
1922          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1923          * GET in an FTP client, suspend the process, wait for the client to
1924          * advertise a zero window, then kill -9 the FTP client, wheee...
1925          * Note: timeout is always zero in such a case.
1926          */
1927         if (data_was_unread) {
1928                 /* Unread data was tossed, zap the connection. */
1929                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1930                 tcp_set_state(sk, TCP_CLOSE);
1931                 tcp_send_active_reset(sk, sk->sk_allocation);
1932         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1933                 /* Check zero linger _after_ checking for unread data. */
1934                 sk->sk_prot->disconnect(sk, 0);
1935                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1936         } else if (tcp_close_state(sk)) {
1937                 /* We FIN if the application ate all the data before
1938                  * zapping the connection.
1939                  */
1940
1941                 /* RED-PEN. Formally speaking, we have broken TCP state
1942                  * machine. State transitions:
1943                  *
1944                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1945                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1946                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1947                  *
1948                  * are legal only when FIN has been sent (i.e. in window),
1949                  * rather than queued out of window. Purists blame.
1950                  *
1951                  * F.e. "RFC state" is ESTABLISHED,
1952                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1953                  *
1954                  * The visible declinations are that sometimes
1955                  * we enter time-wait state, when it is not required really
1956                  * (harmless), do not send active resets, when they are
1957                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1958                  * they look as CLOSING or LAST_ACK for Linux)
1959                  * Probably, I missed some more holelets.
1960                  *                                              --ANK
1961                  */
1962                 tcp_send_fin(sk);
1963         }
1964
1965         sk_stream_wait_close(sk, timeout);
1966
1967 adjudge_to_death:
1968         state = sk->sk_state;
1969         sock_hold(sk);
1970         sock_orphan(sk);
1971
1972         /* It is the last release_sock in its life. It will remove backlog. */
1973         release_sock(sk);
1974
1975
1976         /* Now socket is owned by kernel and we acquire BH lock
1977            to finish close. No need to check for user refs.
1978          */
1979         local_bh_disable();
1980         bh_lock_sock(sk);
1981         WARN_ON(sock_owned_by_user(sk));
1982
1983         percpu_counter_inc(sk->sk_prot->orphan_count);
1984
1985         /* Have we already been destroyed by a softirq or backlog? */
1986         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1987                 goto out;
1988
1989         /*      This is a (useful) BSD violating of the RFC. There is a
1990          *      problem with TCP as specified in that the other end could
1991          *      keep a socket open forever with no application left this end.
1992          *      We use a 3 minute timeout (about the same as BSD) then kill
1993          *      our end. If they send after that then tough - BUT: long enough
1994          *      that we won't make the old 4*rto = almost no time - whoops
1995          *      reset mistake.
1996          *
1997          *      Nope, it was not mistake. It is really desired behaviour
1998          *      f.e. on http servers, when such sockets are useless, but
1999          *      consume significant resources. Let's do it with special
2000          *      linger2 option.                                 --ANK
2001          */
2002
2003         if (sk->sk_state == TCP_FIN_WAIT2) {
2004                 struct tcp_sock *tp = tcp_sk(sk);
2005                 if (tp->linger2 < 0) {
2006                         tcp_set_state(sk, TCP_CLOSE);
2007                         tcp_send_active_reset(sk, GFP_ATOMIC);
2008                         NET_INC_STATS_BH(sock_net(sk),
2009                                         LINUX_MIB_TCPABORTONLINGER);
2010                 } else {
2011                         const int tmo = tcp_fin_time(sk);
2012
2013                         if (tmo > TCP_TIMEWAIT_LEN) {
2014                                 inet_csk_reset_keepalive_timer(sk,
2015                                                 tmo - TCP_TIMEWAIT_LEN);
2016                         } else {
2017                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2018                                 goto out;
2019                         }
2020                 }
2021         }
2022         if (sk->sk_state != TCP_CLOSE) {
2023                 sk_mem_reclaim(sk);
2024                 if (tcp_too_many_orphans(sk, 0)) {
2025                         if (net_ratelimit())
2026                                 printk(KERN_INFO "TCP: too many of orphaned "
2027                                        "sockets\n");
2028                         tcp_set_state(sk, TCP_CLOSE);
2029                         tcp_send_active_reset(sk, GFP_ATOMIC);
2030                         NET_INC_STATS_BH(sock_net(sk),
2031                                         LINUX_MIB_TCPABORTONMEMORY);
2032                 }
2033         }
2034
2035         if (sk->sk_state == TCP_CLOSE)
2036                 inet_csk_destroy_sock(sk);
2037         /* Otherwise, socket is reprieved until protocol close. */
2038
2039 out:
2040         bh_unlock_sock(sk);
2041         local_bh_enable();
2042         sock_put(sk);
2043 }
2044 EXPORT_SYMBOL(tcp_close);
2045
2046 /* These states need RST on ABORT according to RFC793 */
2047
2048 static inline int tcp_need_reset(int state)
2049 {
2050         return (1 << state) &
2051                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2052                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2053 }
2054
2055 int tcp_disconnect(struct sock *sk, int flags)
2056 {
2057         struct inet_sock *inet = inet_sk(sk);
2058         struct inet_connection_sock *icsk = inet_csk(sk);
2059         struct tcp_sock *tp = tcp_sk(sk);
2060         int err = 0;
2061         int old_state = sk->sk_state;
2062
2063         if (old_state != TCP_CLOSE)
2064                 tcp_set_state(sk, TCP_CLOSE);
2065
2066         /* ABORT function of RFC793 */
2067         if (old_state == TCP_LISTEN) {
2068                 inet_csk_listen_stop(sk);
2069         } else if (tcp_need_reset(old_state) ||
2070                    (tp->snd_nxt != tp->write_seq &&
2071                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2072                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2073                  * states
2074                  */
2075                 tcp_send_active_reset(sk, gfp_any());
2076                 sk->sk_err = ECONNRESET;
2077         } else if (old_state == TCP_SYN_SENT)
2078                 sk->sk_err = ECONNRESET;
2079
2080         tcp_clear_xmit_timers(sk);
2081         __skb_queue_purge(&sk->sk_receive_queue);
2082         tcp_write_queue_purge(sk);
2083         __skb_queue_purge(&tp->out_of_order_queue);
2084 #ifdef CONFIG_NET_DMA
2085         __skb_queue_purge(&sk->sk_async_wait_queue);
2086 #endif
2087
2088         inet->inet_dport = 0;
2089
2090         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2091                 inet_reset_saddr(sk);
2092
2093         sk->sk_shutdown = 0;
2094         sock_reset_flag(sk, SOCK_DONE);
2095         tp->srtt = 0;
2096         if ((tp->write_seq += tp->max_window + 2) == 0)
2097                 tp->write_seq = 1;
2098         icsk->icsk_backoff = 0;
2099         tp->snd_cwnd = 2;
2100         icsk->icsk_probes_out = 0;
2101         tp->packets_out = 0;
2102         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2103         tp->snd_cwnd_cnt = 0;
2104         tp->bytes_acked = 0;
2105         tp->window_clamp = 0;
2106         tcp_set_ca_state(sk, TCP_CA_Open);
2107         tcp_clear_retrans(tp);
2108         inet_csk_delack_init(sk);
2109         /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2110          * issue in __tcp_select_window()
2111          */
2112         icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2113         tcp_init_send_head(sk);
2114         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2115         __sk_dst_reset(sk);
2116
2117         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2118
2119         sk->sk_error_report(sk);
2120         return err;
2121 }
2122 EXPORT_SYMBOL(tcp_disconnect);
2123
2124 /*
2125  *      Socket option code for TCP.
2126  */
2127 static int do_tcp_setsockopt(struct sock *sk, int level,
2128                 int optname, char __user *optval, unsigned int optlen)
2129 {
2130         struct tcp_sock *tp = tcp_sk(sk);
2131         struct inet_connection_sock *icsk = inet_csk(sk);
2132         int val;
2133         int err = 0;
2134
2135         /* These are data/string values, all the others are ints */
2136         switch (optname) {
2137         case TCP_CONGESTION: {
2138                 char name[TCP_CA_NAME_MAX];
2139
2140                 if (optlen < 1)
2141                         return -EINVAL;
2142
2143                 val = strncpy_from_user(name, optval,
2144                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2145                 if (val < 0)
2146                         return -EFAULT;
2147                 name[val] = 0;
2148
2149                 lock_sock(sk);
2150                 err = tcp_set_congestion_control(sk, name);
2151                 release_sock(sk);
2152                 return err;
2153         }
2154         case TCP_COOKIE_TRANSACTIONS: {
2155                 struct tcp_cookie_transactions ctd;
2156                 struct tcp_cookie_values *cvp = NULL;
2157
2158                 if (sizeof(ctd) > optlen)
2159                         return -EINVAL;
2160                 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2161                         return -EFAULT;
2162
2163                 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2164                     ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2165                         return -EINVAL;
2166
2167                 if (ctd.tcpct_cookie_desired == 0) {
2168                         /* default to global value */
2169                 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2170                            ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2171                            ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2172                         return -EINVAL;
2173                 }
2174
2175                 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2176                         /* Supercedes all other values */
2177                         lock_sock(sk);
2178                         if (tp->cookie_values != NULL) {
2179                                 kref_put(&tp->cookie_values->kref,
2180                                          tcp_cookie_values_release);
2181                                 tp->cookie_values = NULL;
2182                         }
2183                         tp->rx_opt.cookie_in_always = 0; /* false */
2184                         tp->rx_opt.cookie_out_never = 1; /* true */
2185                         release_sock(sk);
2186                         return err;
2187                 }
2188
2189                 /* Allocate ancillary memory before locking.
2190                  */
2191                 if (ctd.tcpct_used > 0 ||
2192                     (tp->cookie_values == NULL &&
2193                      (sysctl_tcp_cookie_size > 0 ||
2194                       ctd.tcpct_cookie_desired > 0 ||
2195                       ctd.tcpct_s_data_desired > 0))) {
2196                         cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2197                                       GFP_KERNEL);
2198                         if (cvp == NULL)
2199                                 return -ENOMEM;
2200
2201                         kref_init(&cvp->kref);
2202                 }
2203                 lock_sock(sk);
2204                 tp->rx_opt.cookie_in_always =
2205                         (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2206                 tp->rx_opt.cookie_out_never = 0; /* false */
2207
2208                 if (tp->cookie_values != NULL) {
2209                         if (cvp != NULL) {
2210                                 /* Changed values are recorded by a changed
2211                                  * pointer, ensuring the cookie will differ,
2212                                  * without separately hashing each value later.
2213                                  */
2214                                 kref_put(&tp->cookie_values->kref,
2215                                          tcp_cookie_values_release);
2216                         } else {
2217                                 cvp = tp->cookie_values;
2218                         }
2219                 }
2220
2221                 if (cvp != NULL) {
2222                         cvp->cookie_desired = ctd.tcpct_cookie_desired;
2223
2224                         if (ctd.tcpct_used > 0) {
2225                                 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2226                                        ctd.tcpct_used);
2227                                 cvp->s_data_desired = ctd.tcpct_used;
2228                                 cvp->s_data_constant = 1; /* true */
2229                         } else {
2230                                 /* No constant payload data. */
2231                                 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2232                                 cvp->s_data_constant = 0; /* false */
2233                         }
2234
2235                         tp->cookie_values = cvp;
2236                 }
2237                 release_sock(sk);
2238                 return err;
2239         }
2240         default:
2241                 /* fallthru */
2242                 break;
2243         }
2244
2245         if (optlen < sizeof(int))
2246                 return -EINVAL;
2247
2248         if (get_user(val, (int __user *)optval))
2249                 return -EFAULT;
2250
2251         lock_sock(sk);
2252
2253         switch (optname) {
2254         case TCP_MAXSEG:
2255                 /* Values greater than interface MTU won't take effect. However
2256                  * at the point when this call is done we typically don't yet
2257                  * know which interface is going to be used */
2258                 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2259                         err = -EINVAL;
2260                         break;
2261                 }
2262                 tp->rx_opt.user_mss = val;
2263                 break;
2264
2265         case TCP_NODELAY:
2266                 if (val) {
2267                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2268                          * this option on corked socket is remembered, but
2269                          * it is not activated until cork is cleared.
2270                          *
2271                          * However, when TCP_NODELAY is set we make
2272                          * an explicit push, which overrides even TCP_CORK
2273                          * for currently queued segments.
2274                          */
2275                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2276                         tcp_push_pending_frames(sk);
2277                 } else {
2278                         tp->nonagle &= ~TCP_NAGLE_OFF;
2279                 }
2280                 break;
2281
2282         case TCP_THIN_LINEAR_TIMEOUTS:
2283                 if (val < 0 || val > 1)
2284                         err = -EINVAL;
2285                 else
2286                         tp->thin_lto = val;
2287                 break;
2288
2289         case TCP_THIN_DUPACK:
2290                 if (val < 0 || val > 1)
2291                         err = -EINVAL;
2292                 else
2293                         tp->thin_dupack = val;
2294                 break;
2295
2296         case TCP_CORK:
2297                 /* When set indicates to always queue non-full frames.
2298                  * Later the user clears this option and we transmit
2299                  * any pending partial frames in the queue.  This is
2300                  * meant to be used alongside sendfile() to get properly
2301                  * filled frames when the user (for example) must write
2302                  * out headers with a write() call first and then use
2303                  * sendfile to send out the data parts.
2304                  *
2305                  * TCP_CORK can be set together with TCP_NODELAY and it is
2306                  * stronger than TCP_NODELAY.
2307                  */
2308                 if (val) {
2309                         tp->nonagle |= TCP_NAGLE_CORK;
2310                 } else {
2311                         tp->nonagle &= ~TCP_NAGLE_CORK;
2312                         if (tp->nonagle&TCP_NAGLE_OFF)
2313                                 tp->nonagle |= TCP_NAGLE_PUSH;
2314                         tcp_push_pending_frames(sk);
2315                 }
2316                 break;
2317
2318         case TCP_KEEPIDLE:
2319                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2320                         err = -EINVAL;
2321                 else {
2322                         tp->keepalive_time = val * HZ;
2323                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2324                             !((1 << sk->sk_state) &
2325                               (TCPF_CLOSE | TCPF_LISTEN))) {
2326                                 u32 elapsed = keepalive_time_elapsed(tp);
2327                                 if (tp->keepalive_time > elapsed)
2328                                         elapsed = tp->keepalive_time - elapsed;
2329                                 else
2330                                         elapsed = 0;
2331                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2332                         }
2333                 }
2334                 break;
2335         case TCP_KEEPINTVL:
2336                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2337                         err = -EINVAL;
2338                 else
2339                         tp->keepalive_intvl = val * HZ;
2340                 break;
2341         case TCP_KEEPCNT:
2342                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2343                         err = -EINVAL;
2344                 else
2345                         tp->keepalive_probes = val;
2346                 break;
2347         case TCP_SYNCNT:
2348                 if (val < 1 || val > MAX_TCP_SYNCNT)
2349                         err = -EINVAL;
2350                 else
2351                         icsk->icsk_syn_retries = val;
2352                 break;
2353
2354         case TCP_LINGER2:
2355                 if (val < 0)
2356                         tp->linger2 = -1;
2357                 else if (val > sysctl_tcp_fin_timeout / HZ)
2358                         tp->linger2 = 0;
2359                 else
2360                         tp->linger2 = val * HZ;
2361                 break;
2362
2363         case TCP_DEFER_ACCEPT:
2364                 /* Translate value in seconds to number of retransmits */
2365                 icsk->icsk_accept_queue.rskq_defer_accept =
2366                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2367                                         TCP_RTO_MAX / HZ);
2368                 break;
2369
2370         case TCP_WINDOW_CLAMP:
2371                 if (!val) {
2372                         if (sk->sk_state != TCP_CLOSE) {
2373                                 err = -EINVAL;
2374                                 break;
2375                         }
2376                         tp->window_clamp = 0;
2377                 } else
2378                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2379                                                 SOCK_MIN_RCVBUF / 2 : val;
2380                 break;
2381
2382         case TCP_QUICKACK:
2383                 if (!val) {
2384                         icsk->icsk_ack.pingpong = 1;
2385                 } else {
2386                         icsk->icsk_ack.pingpong = 0;
2387                         if ((1 << sk->sk_state) &
2388                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2389                             inet_csk_ack_scheduled(sk)) {
2390                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2391                                 tcp_cleanup_rbuf(sk, 1);
2392                                 if (!(val & 1))
2393                                         icsk->icsk_ack.pingpong = 1;
2394                         }
2395                 }
2396                 break;
2397
2398 #ifdef CONFIG_TCP_MD5SIG
2399         case TCP_MD5SIG:
2400                 /* Read the IP->Key mappings from userspace */
2401                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2402                 break;
2403 #endif
2404         case TCP_USER_TIMEOUT:
2405                 /* Cap the max timeout in ms TCP will retry/retrans
2406                  * before giving up and aborting (ETIMEDOUT) a connection.
2407                  */
2408                 if (val < 0)
2409                         err = -EINVAL;
2410                 else
2411                         icsk->icsk_user_timeout = msecs_to_jiffies(val);
2412                 break;
2413         default:
2414                 err = -ENOPROTOOPT;
2415                 break;
2416         }
2417
2418         release_sock(sk);
2419         return err;
2420 }
2421
2422 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2423                    unsigned int optlen)
2424 {
2425         const struct inet_connection_sock *icsk = inet_csk(sk);
2426
2427         if (level != SOL_TCP)
2428                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2429                                                      optval, optlen);
2430         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2431 }
2432 EXPORT_SYMBOL(tcp_setsockopt);
2433
2434 #ifdef CONFIG_COMPAT
2435 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2436                           char __user *optval, unsigned int optlen)
2437 {
2438         if (level != SOL_TCP)
2439                 return inet_csk_compat_setsockopt(sk, level, optname,
2440                                                   optval, optlen);
2441         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2442 }
2443 EXPORT_SYMBOL(compat_tcp_setsockopt);
2444 #endif
2445
2446 /* Return information about state of tcp endpoint in API format. */
2447 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2448 {
2449         const struct tcp_sock *tp = tcp_sk(sk);
2450         const struct inet_connection_sock *icsk = inet_csk(sk);
2451         u32 now = tcp_time_stamp;
2452
2453         memset(info, 0, sizeof(*info));
2454
2455         info->tcpi_state = sk->sk_state;
2456         info->tcpi_ca_state = icsk->icsk_ca_state;
2457         info->tcpi_retransmits = icsk->icsk_retransmits;
2458         info->tcpi_probes = icsk->icsk_probes_out;
2459         info->tcpi_backoff = icsk->icsk_backoff;
2460
2461         if (tp->rx_opt.tstamp_ok)
2462                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2463         if (tcp_is_sack(tp))
2464                 info->tcpi_options |= TCPI_OPT_SACK;
2465         if (tp->rx_opt.wscale_ok) {
2466                 info->tcpi_options |= TCPI_OPT_WSCALE;
2467                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2468                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2469         }
2470
2471         if (tp->ecn_flags & TCP_ECN_OK)
2472                 info->tcpi_options |= TCPI_OPT_ECN;
2473         if (tp->ecn_flags & TCP_ECN_SEEN)
2474                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2475
2476         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2477         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2478         info->tcpi_snd_mss = tp->mss_cache;
2479         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2480
2481         if (sk->sk_state == TCP_LISTEN) {
2482                 info->tcpi_unacked = sk->sk_ack_backlog;
2483                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2484         } else {
2485                 info->tcpi_unacked = tp->packets_out;
2486                 info->tcpi_sacked = tp->sacked_out;
2487         }
2488         info->tcpi_lost = tp->lost_out;
2489         info->tcpi_retrans = tp->retrans_out;
2490         info->tcpi_fackets = tp->fackets_out;
2491
2492         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2493         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2494         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2495
2496         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2497         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2498         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2499         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2500         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2501         info->tcpi_snd_cwnd = tp->snd_cwnd;
2502         info->tcpi_advmss = tp->advmss;
2503         info->tcpi_reordering = tp->reordering;
2504
2505         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2506         info->tcpi_rcv_space = tp->rcvq_space.space;
2507
2508         info->tcpi_total_retrans = tp->total_retrans;
2509 }
2510 EXPORT_SYMBOL_GPL(tcp_get_info);
2511
2512 static int do_tcp_getsockopt(struct sock *sk, int level,
2513                 int optname, char __user *optval, int __user *optlen)
2514 {
2515         struct inet_connection_sock *icsk = inet_csk(sk);
2516         struct tcp_sock *tp = tcp_sk(sk);
2517         int val, len;
2518
2519         if (get_user(len, optlen))
2520                 return -EFAULT;
2521
2522         len = min_t(unsigned int, len, sizeof(int));
2523
2524         if (len < 0)
2525                 return -EINVAL;
2526
2527         switch (optname) {
2528         case TCP_MAXSEG:
2529                 val = tp->mss_cache;
2530                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2531                         val = tp->rx_opt.user_mss;
2532                 break;
2533         case TCP_NODELAY:
2534                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2535                 break;
2536         case TCP_CORK:
2537                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2538                 break;
2539         case TCP_KEEPIDLE:
2540                 val = keepalive_time_when(tp) / HZ;
2541                 break;
2542         case TCP_KEEPINTVL:
2543                 val = keepalive_intvl_when(tp) / HZ;
2544                 break;
2545         case TCP_KEEPCNT:
2546                 val = keepalive_probes(tp);
2547                 break;
2548         case TCP_SYNCNT:
2549                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2550                 break;
2551         case TCP_LINGER2:
2552                 val = tp->linger2;
2553                 if (val >= 0)
2554                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2555                 break;
2556         case TCP_DEFER_ACCEPT:
2557                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2558                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2559                 break;
2560         case TCP_WINDOW_CLAMP:
2561                 val = tp->window_clamp;
2562                 break;
2563         case TCP_INFO: {
2564                 struct tcp_info info;
2565
2566                 if (get_user(len, optlen))
2567                         return -EFAULT;
2568
2569                 tcp_get_info(sk, &info);
2570
2571                 len = min_t(unsigned int, len, sizeof(info));
2572                 if (put_user(len, optlen))
2573                         return -EFAULT;
2574                 if (copy_to_user(optval, &info, len))
2575                         return -EFAULT;
2576                 return 0;
2577         }
2578         case TCP_QUICKACK:
2579                 val = !icsk->icsk_ack.pingpong;
2580                 break;
2581
2582         case TCP_CONGESTION:
2583                 if (get_user(len, optlen))
2584                         return -EFAULT;
2585                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2586                 if (put_user(len, optlen))
2587                         return -EFAULT;
2588                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2589                         return -EFAULT;
2590                 return 0;
2591
2592         case TCP_COOKIE_TRANSACTIONS: {
2593                 struct tcp_cookie_transactions ctd;
2594                 struct tcp_cookie_values *cvp = tp->cookie_values;
2595
2596                 if (get_user(len, optlen))
2597                         return -EFAULT;
2598                 if (len < sizeof(ctd))
2599                         return -EINVAL;
2600
2601                 memset(&ctd, 0, sizeof(ctd));
2602                 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2603                                    TCP_COOKIE_IN_ALWAYS : 0)
2604                                 | (tp->rx_opt.cookie_out_never ?
2605                                    TCP_COOKIE_OUT_NEVER : 0);
2606
2607                 if (cvp != NULL) {
2608                         ctd.tcpct_flags |= (cvp->s_data_in ?
2609                                             TCP_S_DATA_IN : 0)
2610                                          | (cvp->s_data_out ?
2611                                             TCP_S_DATA_OUT : 0);
2612
2613                         ctd.tcpct_cookie_desired = cvp->cookie_desired;
2614                         ctd.tcpct_s_data_desired = cvp->s_data_desired;
2615
2616                         memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2617                                cvp->cookie_pair_size);
2618                         ctd.tcpct_used = cvp->cookie_pair_size;
2619                 }
2620
2621                 if (put_user(sizeof(ctd), optlen))
2622                         return -EFAULT;
2623                 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2624                         return -EFAULT;
2625                 return 0;
2626         }
2627         case TCP_THIN_LINEAR_TIMEOUTS:
2628                 val = tp->thin_lto;
2629                 break;
2630         case TCP_THIN_DUPACK:
2631                 val = tp->thin_dupack;
2632                 break;
2633
2634         case TCP_USER_TIMEOUT:
2635                 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2636                 break;
2637         default:
2638                 return -ENOPROTOOPT;
2639         }
2640
2641         if (put_user(len, optlen))
2642                 return -EFAULT;
2643         if (copy_to_user(optval, &val, len))
2644                 return -EFAULT;
2645         return 0;
2646 }
2647
2648 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2649                    int __user *optlen)
2650 {
2651         struct inet_connection_sock *icsk = inet_csk(sk);
2652
2653         if (level != SOL_TCP)
2654                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2655                                                      optval, optlen);
2656         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2657 }
2658 EXPORT_SYMBOL(tcp_getsockopt);
2659
2660 #ifdef CONFIG_COMPAT
2661 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2662                           char __user *optval, int __user *optlen)
2663 {
2664         if (level != SOL_TCP)
2665                 return inet_csk_compat_getsockopt(sk, level, optname,
2666                                                   optval, optlen);
2667         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2668 }
2669 EXPORT_SYMBOL(compat_tcp_getsockopt);
2670 #endif
2671
2672 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, u32 features)
2673 {
2674         struct sk_buff *segs = ERR_PTR(-EINVAL);
2675         struct tcphdr *th;
2676         unsigned thlen;
2677         unsigned int seq;
2678         __be32 delta;
2679         unsigned int oldlen;
2680         unsigned int mss;
2681
2682         if (!pskb_may_pull(skb, sizeof(*th)))
2683                 goto out;
2684
2685         th = tcp_hdr(skb);
2686         thlen = th->doff * 4;
2687         if (thlen < sizeof(*th))
2688                 goto out;
2689
2690         if (!pskb_may_pull(skb, thlen))
2691                 goto out;
2692
2693         oldlen = (u16)~skb->len;
2694         __skb_pull(skb, thlen);
2695
2696         mss = skb_shinfo(skb)->gso_size;
2697         if (unlikely(skb->len <= mss))
2698                 goto out;
2699
2700         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2701                 /* Packet is from an untrusted source, reset gso_segs. */
2702                 int type = skb_shinfo(skb)->gso_type;
2703
2704                 if (unlikely(type &
2705                              ~(SKB_GSO_TCPV4 |
2706                                SKB_GSO_DODGY |
2707                                SKB_GSO_TCP_ECN |
2708                                SKB_GSO_TCPV6 |
2709                                0) ||
2710                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2711                         goto out;
2712
2713                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2714
2715                 segs = NULL;
2716                 goto out;
2717         }
2718
2719         segs = skb_segment(skb, features);
2720         if (IS_ERR(segs))
2721                 goto out;
2722
2723         delta = htonl(oldlen + (thlen + mss));
2724
2725         skb = segs;
2726         th = tcp_hdr(skb);
2727         seq = ntohl(th->seq);
2728
2729         do {
2730                 th->fin = th->psh = 0;
2731
2732                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2733                                        (__force u32)delta));
2734                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2735                         th->check =
2736                              csum_fold(csum_partial(skb_transport_header(skb),
2737                                                     thlen, skb->csum));
2738
2739                 seq += mss;
2740                 skb = skb->next;
2741                 th = tcp_hdr(skb);
2742
2743                 th->seq = htonl(seq);
2744                 th->cwr = 0;
2745         } while (skb->next);
2746
2747         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2748                       skb->data_len);
2749         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2750                                 (__force u32)delta));
2751         if (skb->ip_summed != CHECKSUM_PARTIAL)
2752                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2753                                                    thlen, skb->csum));
2754
2755 out:
2756         return segs;
2757 }
2758 EXPORT_SYMBOL(tcp_tso_segment);
2759
2760 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2761 {
2762         struct sk_buff **pp = NULL;
2763         struct sk_buff *p;
2764         struct tcphdr *th;
2765         struct tcphdr *th2;
2766         unsigned int len;
2767         unsigned int thlen;
2768         __be32 flags;
2769         unsigned int mss = 1;
2770         unsigned int hlen;
2771         unsigned int off;
2772         int flush = 1;
2773         int i;
2774
2775         off = skb_gro_offset(skb);
2776         hlen = off + sizeof(*th);
2777         th = skb_gro_header_fast(skb, off);
2778         if (skb_gro_header_hard(skb, hlen)) {
2779                 th = skb_gro_header_slow(skb, hlen, off);
2780                 if (unlikely(!th))
2781                         goto out;
2782         }
2783
2784         thlen = th->doff * 4;
2785         if (thlen < sizeof(*th))
2786                 goto out;
2787
2788         hlen = off + thlen;
2789         if (skb_gro_header_hard(skb, hlen)) {
2790                 th = skb_gro_header_slow(skb, hlen, off);
2791                 if (unlikely(!th))
2792                         goto out;
2793         }
2794
2795         skb_gro_pull(skb, thlen);
2796
2797         len = skb_gro_len(skb);
2798         flags = tcp_flag_word(th);
2799
2800         for (; (p = *head); head = &p->next) {
2801                 if (!NAPI_GRO_CB(p)->same_flow)
2802                         continue;
2803
2804                 th2 = tcp_hdr(p);
2805
2806                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2807                         NAPI_GRO_CB(p)->same_flow = 0;
2808                         continue;
2809                 }
2810
2811                 goto found;
2812         }
2813
2814         goto out_check_final;
2815
2816 found:
2817         flush = NAPI_GRO_CB(p)->flush;
2818         flush |= (__force int)(flags & TCP_FLAG_CWR);
2819         flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
2820                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
2821         flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
2822         for (i = sizeof(*th); i < thlen; i += 4)
2823                 flush |= *(u32 *)((u8 *)th + i) ^
2824                          *(u32 *)((u8 *)th2 + i);
2825
2826         mss = skb_shinfo(p)->gso_size;
2827
2828         flush |= (len - 1) >= mss;
2829         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2830
2831         if (flush || skb_gro_receive(head, skb)) {
2832                 mss = 1;
2833                 goto out_check_final;
2834         }
2835
2836         p = *head;
2837         th2 = tcp_hdr(p);
2838         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2839
2840 out_check_final:
2841         flush = len < mss;
2842         flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
2843                                         TCP_FLAG_RST | TCP_FLAG_SYN |
2844                                         TCP_FLAG_FIN));
2845
2846         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2847                 pp = head;
2848
2849 out:
2850         NAPI_GRO_CB(skb)->flush |= flush;
2851
2852         return pp;
2853 }
2854 EXPORT_SYMBOL(tcp_gro_receive);
2855
2856 int tcp_gro_complete(struct sk_buff *skb)
2857 {
2858         struct tcphdr *th = tcp_hdr(skb);
2859
2860         skb->csum_start = skb_transport_header(skb) - skb->head;
2861         skb->csum_offset = offsetof(struct tcphdr, check);
2862         skb->ip_summed = CHECKSUM_PARTIAL;
2863
2864         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2865
2866         if (th->cwr)
2867                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2868
2869         return 0;
2870 }
2871 EXPORT_SYMBOL(tcp_gro_complete);
2872
2873 #ifdef CONFIG_TCP_MD5SIG
2874 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
2875 static DEFINE_MUTEX(tcp_md5sig_mutex);
2876 static bool tcp_md5sig_pool_populated = false;
2877
2878 static void __tcp_alloc_md5sig_pool(void)
2879 {
2880         int cpu;
2881
2882         for_each_possible_cpu(cpu) {
2883                 if (!per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm) {
2884                         struct crypto_hash *hash;
2885
2886                         hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2887                         if (IS_ERR_OR_NULL(hash))
2888                                 return;
2889                         per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm = hash;
2890                 }
2891         }
2892         /* before setting tcp_md5sig_pool_populated, we must commit all writes
2893          * to memory. See smp_rmb() in tcp_get_md5sig_pool()
2894          */
2895         smp_wmb();
2896         tcp_md5sig_pool_populated = true;
2897 }
2898
2899 bool tcp_alloc_md5sig_pool(void)
2900 {
2901         if (unlikely(!tcp_md5sig_pool_populated)) {
2902                 mutex_lock(&tcp_md5sig_mutex);
2903
2904                 if (!tcp_md5sig_pool_populated)
2905                         __tcp_alloc_md5sig_pool();
2906
2907                 mutex_unlock(&tcp_md5sig_mutex);
2908         }
2909         return tcp_md5sig_pool_populated;
2910 }
2911 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2912
2913
2914 /**
2915  *      tcp_get_md5sig_pool - get md5sig_pool for this user
2916  *
2917  *      We use percpu structure, so if we succeed, we exit with preemption
2918  *      and BH disabled, to make sure another thread or softirq handling
2919  *      wont try to get same context.
2920  */
2921 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2922 {
2923         local_bh_disable();
2924
2925         if (tcp_md5sig_pool_populated) {
2926                 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
2927                 smp_rmb();
2928                 return this_cpu_ptr(&tcp_md5sig_pool);
2929         }
2930         local_bh_enable();
2931         return NULL;
2932 }
2933 EXPORT_SYMBOL(tcp_get_md5sig_pool);
2934
2935 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2936                         const struct tcphdr *th)
2937 {
2938         struct scatterlist sg;
2939         struct tcphdr hdr;
2940         int err;
2941
2942         /* We are not allowed to change tcphdr, make a local copy */
2943         memcpy(&hdr, th, sizeof(hdr));
2944         hdr.check = 0;
2945
2946         /* options aren't included in the hash */
2947         sg_init_one(&sg, &hdr, sizeof(hdr));
2948         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
2949         return err;
2950 }
2951 EXPORT_SYMBOL(tcp_md5_hash_header);
2952
2953 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2954                           const struct sk_buff *skb, unsigned int header_len)
2955 {
2956         struct scatterlist sg;
2957         const struct tcphdr *tp = tcp_hdr(skb);
2958         struct hash_desc *desc = &hp->md5_desc;
2959         unsigned i;
2960         const unsigned head_data_len = skb_headlen(skb) > header_len ?
2961                                        skb_headlen(skb) - header_len : 0;
2962         const struct skb_shared_info *shi = skb_shinfo(skb);
2963         struct sk_buff *frag_iter;
2964
2965         sg_init_table(&sg, 1);
2966
2967         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
2968         if (crypto_hash_update(desc, &sg, head_data_len))
2969                 return 1;
2970
2971         for (i = 0; i < shi->nr_frags; ++i) {
2972                 const struct skb_frag_struct *f = &shi->frags[i];
2973                 unsigned int offset = f->page_offset;
2974                 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
2975
2976                 sg_set_page(&sg, page, skb_frag_size(f),
2977                             offset_in_page(offset));
2978                 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
2979                         return 1;
2980         }
2981
2982         skb_walk_frags(skb, frag_iter)
2983                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
2984                         return 1;
2985
2986         return 0;
2987 }
2988 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
2989
2990 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
2991 {
2992         struct scatterlist sg;
2993
2994         sg_init_one(&sg, key->key, key->keylen);
2995         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
2996 }
2997 EXPORT_SYMBOL(tcp_md5_hash_key);
2998
2999 #endif
3000
3001 /**
3002  * Each Responder maintains up to two secret values concurrently for
3003  * efficient secret rollover.  Each secret value has 4 states:
3004  *
3005  * Generating.  (tcp_secret_generating != tcp_secret_primary)
3006  *    Generates new Responder-Cookies, but not yet used for primary
3007  *    verification.  This is a short-term state, typically lasting only
3008  *    one round trip time (RTT).
3009  *
3010  * Primary.  (tcp_secret_generating == tcp_secret_primary)
3011  *    Used both for generation and primary verification.
3012  *
3013  * Retiring.  (tcp_secret_retiring != tcp_secret_secondary)
3014  *    Used for verification, until the first failure that can be
3015  *    verified by the newer Generating secret.  At that time, this
3016  *    cookie's state is changed to Secondary, and the Generating
3017  *    cookie's state is changed to Primary.  This is a short-term state,
3018  *    typically lasting only one round trip time (RTT).
3019  *
3020  * Secondary.  (tcp_secret_retiring == tcp_secret_secondary)
3021  *    Used for secondary verification, after primary verification
3022  *    failures.  This state lasts no more than twice the Maximum Segment
3023  *    Lifetime (2MSL).  Then, the secret is discarded.
3024  */
3025 struct tcp_cookie_secret {
3026         /* The secret is divided into two parts.  The digest part is the
3027          * equivalent of previously hashing a secret and saving the state,
3028          * and serves as an initialization vector (IV).  The message part
3029          * serves as the trailing secret.
3030          */
3031         u32                             secrets[COOKIE_WORKSPACE_WORDS];
3032         unsigned long                   expires;
3033 };
3034
3035 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3036 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3037 #define TCP_SECRET_LIFE (HZ * 600)
3038
3039 static struct tcp_cookie_secret tcp_secret_one;
3040 static struct tcp_cookie_secret tcp_secret_two;
3041
3042 /* Essentially a circular list, without dynamic allocation. */
3043 static struct tcp_cookie_secret *tcp_secret_generating;
3044 static struct tcp_cookie_secret *tcp_secret_primary;
3045 static struct tcp_cookie_secret *tcp_secret_retiring;
3046 static struct tcp_cookie_secret *tcp_secret_secondary;
3047
3048 static DEFINE_SPINLOCK(tcp_secret_locker);
3049
3050 /* Select a pseudo-random word in the cookie workspace.
3051  */
3052 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3053 {
3054         return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3055 }
3056
3057 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3058  * Called in softirq context.
3059  * Returns: 0 for success.
3060  */
3061 int tcp_cookie_generator(u32 *bakery)
3062 {
3063         unsigned long jiffy = jiffies;
3064
3065         if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3066                 spin_lock_bh(&tcp_secret_locker);
3067                 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3068                         /* refreshed by another */
3069                         memcpy(bakery,
3070                                &tcp_secret_generating->secrets[0],
3071                                COOKIE_WORKSPACE_WORDS);
3072                 } else {
3073                         /* still needs refreshing */
3074                         get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3075
3076                         /* The first time, paranoia assumes that the
3077                          * randomization function isn't as strong.  But,
3078                          * this secret initialization is delayed until
3079                          * the last possible moment (packet arrival).
3080                          * Although that time is observable, it is
3081                          * unpredictably variable.  Mash in the most
3082                          * volatile clock bits available, and expire the
3083                          * secret extra quickly.
3084                          */
3085                         if (unlikely(tcp_secret_primary->expires ==
3086                                      tcp_secret_secondary->expires)) {
3087                                 struct timespec tv;
3088
3089                                 getnstimeofday(&tv);
3090                                 bakery[COOKIE_DIGEST_WORDS+0] ^=
3091                                         (u32)tv.tv_nsec;
3092
3093                                 tcp_secret_secondary->expires = jiffy
3094                                         + TCP_SECRET_1MSL
3095                                         + (0x0f & tcp_cookie_work(bakery, 0));
3096                         } else {
3097                                 tcp_secret_secondary->expires = jiffy
3098                                         + TCP_SECRET_LIFE
3099                                         + (0xff & tcp_cookie_work(bakery, 1));
3100                                 tcp_secret_primary->expires = jiffy
3101                                         + TCP_SECRET_2MSL
3102                                         + (0x1f & tcp_cookie_work(bakery, 2));
3103                         }
3104                         memcpy(&tcp_secret_secondary->secrets[0],
3105                                bakery, COOKIE_WORKSPACE_WORDS);
3106
3107                         rcu_assign_pointer(tcp_secret_generating,
3108                                            tcp_secret_secondary);
3109                         rcu_assign_pointer(tcp_secret_retiring,
3110                                            tcp_secret_primary);
3111                         /*
3112                          * Neither call_rcu() nor synchronize_rcu() needed.
3113                          * Retiring data is not freed.  It is replaced after
3114                          * further (locked) pointer updates, and a quiet time
3115                          * (minimum 1MSL, maximum LIFE - 2MSL).
3116                          */
3117                 }
3118                 spin_unlock_bh(&tcp_secret_locker);
3119         } else {
3120                 rcu_read_lock_bh();
3121                 memcpy(bakery,
3122                        &rcu_dereference(tcp_secret_generating)->secrets[0],
3123                        COOKIE_WORKSPACE_WORDS);
3124                 rcu_read_unlock_bh();
3125         }
3126         return 0;
3127 }
3128 EXPORT_SYMBOL(tcp_cookie_generator);
3129
3130 void tcp_done(struct sock *sk)
3131 {
3132         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3133                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3134
3135         tcp_set_state(sk, TCP_CLOSE);
3136         tcp_clear_xmit_timers(sk);
3137
3138         sk->sk_shutdown = SHUTDOWN_MASK;
3139
3140         if (!sock_flag(sk, SOCK_DEAD))
3141                 sk->sk_state_change(sk);
3142         else
3143                 inet_csk_destroy_sock(sk);
3144 }
3145 EXPORT_SYMBOL_GPL(tcp_done);
3146
3147 extern struct tcp_congestion_ops tcp_reno;
3148
3149 static __initdata unsigned long thash_entries;
3150 static int __init set_thash_entries(char *str)
3151 {
3152         if (!str)
3153                 return 0;
3154         thash_entries = simple_strtoul(str, &str, 0);
3155         return 1;
3156 }
3157 __setup("thash_entries=", set_thash_entries);
3158
3159 void __init tcp_init(void)
3160 {
3161         struct sk_buff *skb = NULL;
3162         unsigned long limit;
3163         int i, max_rshare, max_wshare, cnt;
3164         unsigned long jiffy = jiffies;
3165
3166         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3167
3168         percpu_counter_init(&tcp_sockets_allocated, 0);
3169         percpu_counter_init(&tcp_orphan_count, 0);
3170         tcp_hashinfo.bind_bucket_cachep =
3171                 kmem_cache_create("tcp_bind_bucket",
3172                                   sizeof(struct inet_bind_bucket), 0,
3173                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3174
3175         /* Size and allocate the main established and bind bucket
3176          * hash tables.
3177          *
3178          * The methodology is similar to that of the buffer cache.
3179          */
3180         tcp_hashinfo.ehash =
3181                 alloc_large_system_hash("TCP established",
3182                                         sizeof(struct inet_ehash_bucket),
3183                                         thash_entries,
3184                                         (totalram_pages >= 128 * 1024) ?
3185                                         13 : 15,
3186                                         0,
3187                                         NULL,
3188                                         &tcp_hashinfo.ehash_mask,
3189                                         thash_entries ? 0 : 512 * 1024);
3190         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3191                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3192                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3193         }
3194         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3195                 panic("TCP: failed to alloc ehash_locks");
3196         tcp_hashinfo.bhash =
3197                 alloc_large_system_hash("TCP bind",
3198                                         sizeof(struct inet_bind_hashbucket),
3199                                         tcp_hashinfo.ehash_mask + 1,
3200                                         (totalram_pages >= 128 * 1024) ?
3201                                         13 : 15,
3202                                         0,
3203                                         &tcp_hashinfo.bhash_size,
3204                                         NULL,
3205                                         64 * 1024);
3206         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
3207         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3208                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3209                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3210         }
3211
3212
3213         cnt = tcp_hashinfo.ehash_mask + 1;
3214
3215         tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3216         sysctl_tcp_max_orphans = cnt / 2;
3217         sysctl_max_syn_backlog = max(128, cnt / 256);
3218
3219         limit = nr_free_buffer_pages() / 8;
3220         limit = max(limit, 128UL);
3221         sysctl_tcp_mem[0] = limit / 4 * 3;
3222         sysctl_tcp_mem[1] = limit;
3223         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3224
3225         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3226         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
3227         max_wshare = min(4UL*1024*1024, limit);
3228         max_rshare = min(6UL*1024*1024, limit);
3229
3230         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3231         sysctl_tcp_wmem[1] = 16*1024;
3232         sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3233
3234         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3235         sysctl_tcp_rmem[1] = 87380;
3236         sysctl_tcp_rmem[2] = max(87380, max_rshare);
3237
3238         printk(KERN_INFO "TCP: Hash tables configured "
3239                "(established %u bind %u)\n",
3240                tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3241
3242         tcp_register_congestion_control(&tcp_reno);
3243
3244         memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3245         memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3246         tcp_secret_one.expires = jiffy; /* past due */
3247         tcp_secret_two.expires = jiffy; /* past due */
3248         tcp_secret_generating = &tcp_secret_one;
3249         tcp_secret_primary = &tcp_secret_one;
3250         tcp_secret_retiring = &tcp_secret_two;
3251         tcp_secret_secondary = &tcp_secret_two;
3252 }