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 #define pr_fmt(fmt) "TCP: " fmt
 249
 250 #include <linux/kernel.h>
 251 #include <linux/module.h>
 252 #include <linux/types.h>
 253 #include <linux/fcntl.h>
 254 #include <linux/poll.h>
 255 #include <linux/init.h>
 256 #include <linux/fs.h>
 257 #include <linux/skbuff.h>
 258 #include <linux/scatterlist.h>
 259 #include <linux/splice.h>
 260 #include <linux/net.h>
 261 #include <linux/socket.h>
 262 #include <linux/random.h>
 263 #include <linux/bootmem.h>
 264 #include <linux/highmem.h>
 265 #include <linux/swap.h>
 266 #include <linux/cache.h>
 267 #include <linux/err.h>
 268 #include <linux/crypto.h>
 269 #include <linux/time.h>
 270 #include <linux/slab.h>
 271
 272 #include <net/icmp.h>
 273 #include <net/tcp.h>
 274 #include <net/xfrm.h>
 275 #include <net/ip.h>
 276 #include <net/netdma.h>
 277 #include <net/sock.h>
 278
 279 #include <asm/uaccess.h>
 280 #include <asm/ioctls.h>
 281
 282 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
 283
 284 struct percpu_counter tcp_orphan_count;
 285 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 286
 287 int sysctl_tcp_wmem[3] __read_mostly;
 288 int sysctl_tcp_rmem[3] __read_mostly;
 289
 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                         struct sk_buff *skb;
 485
 486                         answ = tp->rcv_nxt - tp->copied_seq;
 487
 488                         /* Subtract 1, if FIN is in queue. */
 489                         skb = skb_peek_tail(&sk->sk_receive_queue);
 490                         if (answ && skb)
 491                                 answ -= tcp_hdr(skb)->fin;
 492                 } else
 493                         answ = tp->urg_seq - tp->copied_seq;
 494                 release_sock(sk);
 495                 break;
 496         case SIOCATMARK:
 497                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 498                 break;
 499         case SIOCOUTQ:
 500                 if (sk->sk_state == TCP_LISTEN)
 501                         return -EINVAL;
 502
 503                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 504                         answ = 0;
 505                 else
 506                         answ = tp->write_seq - tp->snd_una;
 507                 break;
 508         case SIOCOUTQNSD:
 509                 if (sk->sk_state == TCP_LISTEN)
 510                         return -EINVAL;
 511
 512                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 513                         answ = 0;
 514                 else
 515                         answ = tp->write_seq - tp->snd_nxt;
 516                 break;
 517         default:
 518                 return -ENOIOCTLCMD;
 519         }
 520
 521         return put_user(answ, (int __user *)arg);
 522 }
 523 EXPORT_SYMBOL(tcp_ioctl);
 524
 525 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 526 {
 527         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 528         tp->pushed_seq = tp->write_seq;
 529 }
 530
 531 static inline int forced_push(const struct tcp_sock *tp)
 532 {
 533         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 534 }
 535
 536 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
 537 {
 538         struct tcp_sock *tp = tcp_sk(sk);
 539         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 540
 541         skb->csum    = 0;
 542         tcb->seq     = tcb->end_seq = tp->write_seq;
 543         tcb->tcp_flags = TCPHDR_ACK;
 544         tcb->sacked  = 0;
 545         skb_header_release(skb);
 546         tcp_add_write_queue_tail(sk, skb);
 547         sk->sk_wmem_queued += skb->truesize;
 548         sk_mem_charge(sk, skb->truesize);
 549         if (tp->nonagle & TCP_NAGLE_PUSH)
 550                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 551 }
 552
 553 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
 554 {
 555         if (flags & MSG_OOB)
 556                 tp->snd_up = tp->write_seq;
 557 }
 558
 559 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
 560                             int nonagle)
 561 {
 562         if (tcp_send_head(sk)) {
 563                 struct tcp_sock *tp = tcp_sk(sk);
 564
 565                 if (!(flags & MSG_MORE) || forced_push(tp))
 566                         tcp_mark_push(tp, tcp_write_queue_tail(sk));
 567
 568                 tcp_mark_urg(tp, flags);
 569                 __tcp_push_pending_frames(sk, mss_now,
 570                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
 571         }
 572 }
 573
 574 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 575                                 unsigned int offset, size_t len)
 576 {
 577         struct tcp_splice_state *tss = rd_desc->arg.data;
 578         int ret;
 579
 580         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
 581                               tss->flags);
 582         if (ret > 0)
 583                 rd_desc->count -= ret;
 584         return ret;
 585 }
 586
 587 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 588 {
 589         /* Store TCP splice context information in read_descriptor_t. */
 590         read_descriptor_t rd_desc = {
 591                 .arg.data = tss,
 592                 .count    = tss->len,
 593         };
 594
 595         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 596 }
 597
 598 /**
 599  *  tcp_splice_read - splice data from TCP socket to a pipe
 600  * @sock:       socket to splice from
 601  * @ppos:       position (not valid)
 602  * @pipe:       pipe to splice to
 603  * @len:        number of bytes to splice
 604  * @flags:      splice modifier flags
 605  *
 606  * Description:
 607  *    Will read pages from given socket and fill them into a pipe.
 608  *
 609  **/
 610 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 611                         struct pipe_inode_info *pipe, size_t len,
 612                         unsigned int flags)
 613 {
 614         struct sock *sk = sock->sk;
 615         struct tcp_splice_state tss = {
 616                 .pipe = pipe,
 617                 .len = len,
 618                 .flags = flags,
 619         };
 620         long timeo;
 621         ssize_t spliced;
 622         int ret;
 623
 624         sock_rps_record_flow(sk);
 625         /*
 626          * We can't seek on a socket input
 627          */
 628         if (unlikely(*ppos))
 629                 return -ESPIPE;
 630
 631         ret = spliced = 0;
 632
 633         lock_sock(sk);
 634
 635         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 636         while (tss.len) {
 637                 ret = __tcp_splice_read(sk, &tss);
 638                 if (ret < 0)
 639                         break;
 640                 else if (!ret) {
 641                         if (spliced)
 642                                 break;
 643                         if (sock_flag(sk, SOCK_DONE))
 644                                 break;
 645                         if (sk->sk_err) {
 646                                 ret = sock_error(sk);
 647                                 break;
 648                         }
 649                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 650                                 break;
 651                         if (sk->sk_state == TCP_CLOSE) {
 652                                 /*
 653                                  * This occurs when user tries to read
 654                                  * from never connected socket.
 655                                  */
 656                                 if (!sock_flag(sk, SOCK_DONE))
 657                                         ret = -ENOTCONN;
 658                                 break;
 659                         }
 660                         if (!timeo) {
 661                                 ret = -EAGAIN;
 662                                 break;
 663                         }
 664                         sk_wait_data(sk, &timeo);
 665                         if (signal_pending(current)) {
 666                                 ret = sock_intr_errno(timeo);
 667                                 break;
 668                         }
 669                         continue;
 670                 }
 671                 tss.len -= ret;
 672                 spliced += ret;
 673
 674                 if (!timeo)
 675                         break;
 676                 release_sock(sk);
 677                 lock_sock(sk);
 678
 679                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 680                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 681                     signal_pending(current))
 682                         break;
 683         }
 684
 685         release_sock(sk);
 686
 687         if (spliced)
 688                 return spliced;
 689
 690         return ret;
 691 }
 692 EXPORT_SYMBOL(tcp_splice_read);
 693
 694 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
 695 {
 696         struct sk_buff *skb;
 697
 698         /* The TCP header must be at least 32-bit aligned.  */
 699         size = ALIGN(size, 4);
 700
 701         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 702         if (skb) {
 703                 if (sk_wmem_schedule(sk, skb->truesize)) {
 704                         skb_reserve(skb, sk->sk_prot->max_header);
 705                         /*
 706                          * Make sure that we have exactly size bytes
 707                          * available to the caller, no more, no less.
 708                          */
 709                         skb->avail_size = size;
 710                         return skb;
 711                 }
 712                 __kfree_skb(skb);
 713         } else {
 714                 sk->sk_prot->enter_memory_pressure(sk);
 715                 sk_stream_moderate_sndbuf(sk);
 716         }
 717         return NULL;
 718 }
 719
 720 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 721                                        int large_allowed)
 722 {
 723         struct tcp_sock *tp = tcp_sk(sk);
 724         u32 xmit_size_goal, old_size_goal;
 725
 726         xmit_size_goal = mss_now;
 727
 728         if (large_allowed && sk_can_gso(sk)) {
 729                 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
 730                                   inet_csk(sk)->icsk_af_ops->net_header_len -
 731                                   inet_csk(sk)->icsk_ext_hdr_len -
 732                                   tp->tcp_header_len);
 733
 734                 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
 735
 736                 /* We try hard to avoid divides here */
 737                 old_size_goal = tp->xmit_size_goal_segs * mss_now;
 738
 739                 if (likely(old_size_goal <= xmit_size_goal &&
 740                            old_size_goal + mss_now > xmit_size_goal)) {
 741                         xmit_size_goal = old_size_goal;
 742                 } else {
 743                         tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
 744                         xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
 745                 }
 746         }
 747
 748         return max(xmit_size_goal, mss_now);
 749 }
 750
 751 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 752 {
 753         int mss_now;
 754
 755         mss_now = tcp_current_mss(sk);
 756         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 757
 758         return mss_now;
 759 }
 760
 761 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
 762                          size_t psize, int flags)
 763 {
 764         struct tcp_sock *tp = tcp_sk(sk);
 765         int mss_now, size_goal;
 766         int err;
 767         ssize_t copied;
 768         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 769
 770         /* Wait for a connection to finish. */
 771         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 772                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 773                         goto out_err;
 774
 775         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 776
 777         mss_now = tcp_send_mss(sk, &size_goal, flags);
 778         copied = 0;
 779
 780         err = -EPIPE;
 781         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 782                 goto out_err;
 783
 784         while (psize > 0) {
 785                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 786                 struct page *page = pages[poffset / PAGE_SIZE];
 787                 int copy, i, can_coalesce;
 788                 int offset = poffset % PAGE_SIZE;
 789                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
 790
 791                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
 792 new_segment:
 793                         if (!sk_stream_memory_free(sk))
 794                                 goto wait_for_sndbuf;
 795
 796                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
 797                         if (!skb)
 798                                 goto wait_for_memory;
 799
 800                         skb_entail(sk, skb);
 801                         copy = size_goal;
 802                 }
 803
 804                 if (copy > size)
 805                         copy = size;
 806
 807                 i = skb_shinfo(skb)->nr_frags;
 808                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 809                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
 810                         tcp_mark_push(tp, skb);
 811                         goto new_segment;
 812                 }
 813                 if (!sk_wmem_schedule(sk, copy))
 814                         goto wait_for_memory;
 815
 816                 if (can_coalesce) {
 817                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
 818                 } else {
 819                         get_page(page);
 820                         skb_fill_page_desc(skb, i, page, offset, copy);
 821                 }
 822
 823                 skb->len += copy;
 824                 skb->data_len += copy;
 825                 skb->truesize += copy;
 826                 sk->sk_wmem_queued += copy;
 827                 sk_mem_charge(sk, copy);
 828                 skb->ip_summed = CHECKSUM_PARTIAL;
 829                 tp->write_seq += copy;
 830                 TCP_SKB_CB(skb)->end_seq += copy;
 831                 skb_shinfo(skb)->gso_segs = 0;
 832
 833                 if (!copied)
 834                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
 835
 836                 copied += copy;
 837                 poffset += copy;
 838                 if (!(psize -= copy))
 839                         goto out;
 840
 841                 if (skb->len < size_goal || (flags & MSG_OOB))
 842                         continue;
 843
 844                 if (forced_push(tp)) {
 845                         tcp_mark_push(tp, skb);
 846                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
 847                 } else if (skb == tcp_send_head(sk))
 848                         tcp_push_one(sk, mss_now);
 849                 continue;
 850
 851 wait_for_sndbuf:
 852                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 853 wait_for_memory:
 854                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
 855
 856                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 857                         goto do_error;
 858
 859                 mss_now = tcp_send_mss(sk, &size_goal, flags);
 860         }
 861
 862 out:
 863         if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
 864                 tcp_push(sk, flags, mss_now, tp->nonagle);
 865         return copied;
 866
 867 do_error:
 868         if (copied)
 869                 goto out;
 870 out_err:
 871         return sk_stream_error(sk, flags, err);
 872 }
 873
 874 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
 875                  size_t size, int flags)
 876 {
 877         ssize_t res;
 878
 879         if (!(sk->sk_route_caps & NETIF_F_SG) ||
 880             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
 881                 return sock_no_sendpage(sk->sk_socket, page, offset, size,
 882                                         flags);
 883
 884         lock_sock(sk);
 885         res = do_tcp_sendpages(sk, &page, offset, size, flags);
 886         release_sock(sk);
 887         return res;
 888 }
 889 EXPORT_SYMBOL(tcp_sendpage);
 890
 891 static inline int select_size(const struct sock *sk, bool sg)
 892 {
 893         const struct tcp_sock *tp = tcp_sk(sk);
 894         int tmp = tp->mss_cache;
 895
 896         if (sg) {
 897                 if (sk_can_gso(sk)) {
 898                         /* Small frames wont use a full page:
 899                          * Payload will immediately follow tcp header.
 900                          */
 901                         tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
 902                 } else {
 903                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
 904
 905                         if (tmp >= pgbreak &&
 906                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
 907                                 tmp = pgbreak;
 908                 }
 909         }
 910
 911         return tmp;
 912 }
 913
 914 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
 915                 size_t size)
 916 {
 917         struct iovec *iov;
 918         struct tcp_sock *tp = tcp_sk(sk);
 919         struct sk_buff *skb;
 920         int iovlen, flags, err, copied;
 921         int mss_now, size_goal;
 922         bool sg;
 923         long timeo;
 924
 925         lock_sock(sk);
 926
 927         flags = msg->msg_flags;
 928         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 929
 930         /* Wait for a connection to finish. */
 931         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 932                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 933                         goto out_err;
 934
 935         /* This should be in poll */
 936         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 937
 938         mss_now = tcp_send_mss(sk, &size_goal, flags);
 939
 940         /* Ok commence sending. */
 941         iovlen = msg->msg_iovlen;
 942         iov = msg->msg_iov;
 943         copied = 0;
 944
 945         err = -EPIPE;
 946         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 947                 goto out_err;
 948
 949         sg = !!(sk->sk_route_caps & NETIF_F_SG);
 950
 951         while (--iovlen >= 0) {
 952                 size_t seglen = iov->iov_len;
 953                 unsigned char __user *from = iov->iov_base;
 954
 955                 iov++;
 956
 957                 while (seglen > 0) {
 958                         int copy = 0;
 959                         int max = size_goal;
 960
 961                         skb = tcp_write_queue_tail(sk);
 962                         if (tcp_send_head(sk)) {
 963                                 if (skb->ip_summed == CHECKSUM_NONE)
 964                                         max = mss_now;
 965                                 copy = max - skb->len;
 966                         }
 967
 968                         if (copy <= 0) {
 969 new_segment:
 970                                 /* Allocate new segment. If the interface is SG,
 971                                  * allocate skb fitting to single page.
 972                                  */
 973                                 if (!sk_stream_memory_free(sk))
 974                                         goto wait_for_sndbuf;
 975
 976                                 skb = sk_stream_alloc_skb(sk,
 977                                                           select_size(sk, sg),
 978                                                           sk->sk_allocation);
 979                                 if (!skb)
 980                                         goto wait_for_memory;
 981
 982                                 /*
 983                                  * Check whether we can use HW checksum.
 984                                  */
 985                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
 986                                         skb->ip_summed = CHECKSUM_PARTIAL;
 987
 988                                 skb_entail(sk, skb);
 989                                 copy = size_goal;
 990                                 max = size_goal;
 991                         }
 992
 993                         /* Try to append data to the end of skb. */
 994                         if (copy > seglen)
 995                                 copy = seglen;
 996
 997                         /* Where to copy to? */
 998                         if (skb_availroom(skb) > 0) {
 999                                 /* We have some space in skb head. Superb! */
1000                                 copy = min_t(int, copy, skb_availroom(skb));
1001                                 err = skb_add_data_nocache(sk, skb, from, copy);
1002                                 if (err)
1003                                         goto do_fault;
1004                         } else {
1005                                 int merge = 0;
1006                                 int i = skb_shinfo(skb)->nr_frags;
1007                                 struct page *page = sk->sk_sndmsg_page;
1008                                 int off;
1009
1010                                 if (page && page_count(page) == 1)
1011                                         sk->sk_sndmsg_off = 0;
1012
1013                                 off = sk->sk_sndmsg_off;
1014
1015                                 if (skb_can_coalesce(skb, i, page, off) &&
1016                                     off != PAGE_SIZE) {
1017                                         /* We can extend the last page
1018                                          * fragment. */
1019                                         merge = 1;
1020                                 } else if (i == MAX_SKB_FRAGS || !sg) {
1021                                         /* Need to add new fragment and cannot
1022                                          * do this because interface is non-SG,
1023                                          * or because all the page slots are
1024                                          * busy. */
1025                                         tcp_mark_push(tp, skb);
1026                                         goto new_segment;
1027                                 } else if (page) {
1028                                         if (off == PAGE_SIZE) {
1029                                                 put_page(page);
1030                                                 sk->sk_sndmsg_page = page = NULL;
1031                                                 off = 0;
1032                                         }
1033                                 } else
1034                                         off = 0;
1035
1036                                 if (copy > PAGE_SIZE - off)
1037                                         copy = PAGE_SIZE - off;
1038
1039                                 if (!sk_wmem_schedule(sk, copy))
1040                                         goto wait_for_memory;
1041
1042                                 if (!page) {
1043                                         /* Allocate new cache page. */
1044                                         if (!(page = sk_stream_alloc_page(sk)))
1045                                                 goto wait_for_memory;
1046                                 }
1047
1048                                 /* Time to copy data. We are close to
1049                                  * the end! */
1050                                 err = skb_copy_to_page_nocache(sk, from, skb,
1051                                                                page, off, copy);
1052                                 if (err) {
1053                                         /* If this page was new, give it to the
1054                                          * socket so it does not get leaked.
1055                                          */
1056                                         if (!sk->sk_sndmsg_page) {
1057                                                 sk->sk_sndmsg_page = page;
1058                                                 sk->sk_sndmsg_off = 0;
1059                                         }
1060                                         goto do_error;
1061                                 }
1062
1063                                 /* Update the skb. */
1064                                 if (merge) {
1065                                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1066                                 } else {
1067                                         skb_fill_page_desc(skb, i, page, off, copy);
1068                                         if (sk->sk_sndmsg_page) {
1069                                                 get_page(page);
1070                                         } else if (off + copy < PAGE_SIZE) {
1071                                                 get_page(page);
1072                                                 sk->sk_sndmsg_page = page;
1073                                         }
1074                                 }
1075
1076                                 sk->sk_sndmsg_off = off + copy;
1077                         }
1078
1079                         if (!copied)
1080                                 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1081
1082                         tp->write_seq += copy;
1083                         TCP_SKB_CB(skb)->end_seq += copy;
1084                         skb_shinfo(skb)->gso_segs = 0;
1085
1086                         from += copy;
1087                         copied += copy;
1088                         if ((seglen -= copy) == 0 && iovlen == 0)
1089                                 goto out;
1090
1091                         if (skb->len < max || (flags & MSG_OOB))
1092                                 continue;
1093
1094                         if (forced_push(tp)) {
1095                                 tcp_mark_push(tp, skb);
1096                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1097                         } else if (skb == tcp_send_head(sk))
1098                                 tcp_push_one(sk, mss_now);
1099                         continue;
1100
1101 wait_for_sndbuf:
1102                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1103 wait_for_memory:
1104                         if (copied)
1105                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1106
1107                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1108                                 goto do_error;
1109
1110                         mss_now = tcp_send_mss(sk, &size_goal, flags);
1111                 }
1112         }
1113
1114 out:
1115         if (copied)
1116                 tcp_push(sk, flags, mss_now, tp->nonagle);
1117         release_sock(sk);
1118         return copied;
1119
1120 do_fault:
1121         if (!skb->len) {
1122                 tcp_unlink_write_queue(skb, sk);
1123                 /* It is the one place in all of TCP, except connection
1124                  * reset, where we can be unlinking the send_head.
1125                  */
1126                 tcp_check_send_head(sk, skb);
1127                 sk_wmem_free_skb(sk, skb);
1128         }
1129
1130 do_error:
1131         if (copied)
1132                 goto out;
1133 out_err:
1134         err = sk_stream_error(sk, flags, err);
1135         release_sock(sk);
1136         return err;
1137 }
1138 EXPORT_SYMBOL(tcp_sendmsg);
1139
1140 /*
1141  *      Handle reading urgent data. BSD has very simple semantics for
1142  *      this, no blocking and very strange errors 8)
1143  */
1144
1145 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1146 {
1147         struct tcp_sock *tp = tcp_sk(sk);
1148
1149         /* No URG data to read. */
1150         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1151             tp->urg_data == TCP_URG_READ)
1152                 return -EINVAL; /* Yes this is right ! */
1153
1154         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1155                 return -ENOTCONN;
1156
1157         if (tp->urg_data & TCP_URG_VALID) {
1158                 int err = 0;
1159                 char c = tp->urg_data;
1160
1161                 if (!(flags & MSG_PEEK))
1162                         tp->urg_data = TCP_URG_READ;
1163
1164                 /* Read urgent data. */
1165                 msg->msg_flags |= MSG_OOB;
1166
1167                 if (len > 0) {
1168                         if (!(flags & MSG_TRUNC))
1169                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1170                         len = 1;
1171                 } else
1172                         msg->msg_flags |= MSG_TRUNC;
1173
1174                 return err ? -EFAULT : len;
1175         }
1176
1177         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1178                 return 0;
1179
1180         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1181          * the available implementations agree in this case:
1182          * this call should never block, independent of the
1183          * blocking state of the socket.
1184          * Mike <pall@rz.uni-karlsruhe.de>
1185          */
1186         return -EAGAIN;
1187 }
1188
1189 /* Clean up the receive buffer for full frames taken by the user,
1190  * then send an ACK if necessary.  COPIED is the number of bytes
1191  * tcp_recvmsg has given to the user so far, it speeds up the
1192  * calculation of whether or not we must ACK for the sake of
1193  * a window update.
1194  */
1195 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1196 {
1197         struct tcp_sock *tp = tcp_sk(sk);
1198         int time_to_ack = 0;
1199
1200         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1201
1202         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1203              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1204              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1205
1206         if (inet_csk_ack_scheduled(sk)) {
1207                 const struct inet_connection_sock *icsk = inet_csk(sk);
1208                    /* Delayed ACKs frequently hit locked sockets during bulk
1209                     * receive. */
1210                 if (icsk->icsk_ack.blocked ||
1211                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1212                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1213                     /*
1214                      * If this read emptied read buffer, we send ACK, if
1215                      * connection is not bidirectional, user drained
1216                      * receive buffer and there was a small segment
1217                      * in queue.
1218                      */
1219                     (copied > 0 &&
1220                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1221                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1222                        !icsk->icsk_ack.pingpong)) &&
1223                       !atomic_read(&sk->sk_rmem_alloc)))
1224                         time_to_ack = 1;
1225         }
1226
1227         /* We send an ACK if we can now advertise a non-zero window
1228          * which has been raised "significantly".
1229          *
1230          * Even if window raised up to infinity, do not send window open ACK
1231          * in states, where we will not receive more. It is useless.
1232          */
1233         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1234                 __u32 rcv_window_now = tcp_receive_window(tp);
1235
1236                 /* Optimize, __tcp_select_window() is not cheap. */
1237                 if (2*rcv_window_now <= tp->window_clamp) {
1238                         __u32 new_window = __tcp_select_window(sk);
1239
1240                         /* Send ACK now, if this read freed lots of space
1241                          * in our buffer. Certainly, new_window is new window.
1242                          * We can advertise it now, if it is not less than current one.
1243                          * "Lots" means "at least twice" here.
1244                          */
1245                         if (new_window && new_window >= 2 * rcv_window_now)
1246                                 time_to_ack = 1;
1247                 }
1248         }
1249         if (time_to_ack)
1250                 tcp_send_ack(sk);
1251 }
1252
1253 static void tcp_prequeue_process(struct sock *sk)
1254 {
1255         struct sk_buff *skb;
1256         struct tcp_sock *tp = tcp_sk(sk);
1257
1258         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1259
1260         /* RX process wants to run with disabled BHs, though it is not
1261          * necessary */
1262         local_bh_disable();
1263         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1264                 sk_backlog_rcv(sk, skb);
1265         local_bh_enable();
1266
1267         /* Clear memory counter. */
1268         tp->ucopy.memory = 0;
1269 }
1270
1271 #ifdef CONFIG_NET_DMA
1272 static void tcp_service_net_dma(struct sock *sk, bool wait)
1273 {
1274         dma_cookie_t done, used;
1275         dma_cookie_t last_issued;
1276         struct tcp_sock *tp = tcp_sk(sk);
1277
1278         if (!tp->ucopy.dma_chan)
1279                 return;
1280
1281         last_issued = tp->ucopy.dma_cookie;
1282         dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1283
1284         do {
1285                 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1286                                               last_issued, &done,
1287                                               &used) == DMA_SUCCESS) {
1288                         /* Safe to free early-copied skbs now */
1289                         __skb_queue_purge(&sk->sk_async_wait_queue);
1290                         break;
1291                 } else {
1292                         struct sk_buff *skb;
1293                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1294                                (dma_async_is_complete(skb->dma_cookie, done,
1295                                                       used) == DMA_SUCCESS)) {
1296                                 __skb_dequeue(&sk->sk_async_wait_queue);
1297                                 kfree_skb(skb);
1298                         }
1299                 }
1300         } while (wait);
1301 }
1302 #endif
1303
1304 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1305 {
1306         struct sk_buff *skb;
1307         u32 offset;
1308
1309         skb_queue_walk(&sk->sk_receive_queue, skb) {
1310                 offset = seq - TCP_SKB_CB(skb)->seq;
1311                 if (tcp_hdr(skb)->syn)
1312                         offset--;
1313                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1314                         *off = offset;
1315                         return skb;
1316                 }
1317         }
1318         return NULL;
1319 }
1320
1321 /*
1322  * This routine provides an alternative to tcp_recvmsg() for routines
1323  * that would like to handle copying from skbuffs directly in 'sendfile'
1324  * fashion.
1325  * Note:
1326  *      - It is assumed that the socket was locked by the caller.
1327  *      - The routine does not block.
1328  *      - At present, there is no support for reading OOB data
1329  *        or for 'peeking' the socket using this routine
1330  *        (although both would be easy to implement).
1331  */
1332 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1333                   sk_read_actor_t recv_actor)
1334 {
1335         struct sk_buff *skb;
1336         struct tcp_sock *tp = tcp_sk(sk);
1337         u32 seq = tp->copied_seq;
1338         u32 offset;
1339         int copied = 0;
1340
1341         if (sk->sk_state == TCP_LISTEN)
1342                 return -ENOTCONN;
1343         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1344                 if (offset < skb->len) {
1345                         int used;
1346                         size_t len;
1347
1348                         len = skb->len - offset;
1349                         /* Stop reading if we hit a patch of urgent data */
1350                         if (tp->urg_data) {
1351                                 u32 urg_offset = tp->urg_seq - seq;
1352                                 if (urg_offset < len)
1353                                         len = urg_offset;
1354                                 if (!len)
1355                                         break;
1356                         }
1357                         used = recv_actor(desc, skb, offset, len);
1358                         if (used < 0) {
1359                                 if (!copied)
1360                                         copied = used;
1361                                 break;
1362                         } else if (used <= len) {
1363                                 seq += used;
1364                                 copied += used;
1365                                 offset += used;
1366                         }
1367                         /*
1368                          * If recv_actor drops the lock (e.g. TCP splice
1369                          * receive) the skb pointer might be invalid when
1370                          * getting here: tcp_collapse might have deleted it
1371                          * while aggregating skbs from the socket queue.
1372                          */
1373                         skb = tcp_recv_skb(sk, seq-1, &offset);
1374                         if (!skb || (offset+1 != skb->len))
1375                                 break;
1376                 }
1377                 if (tcp_hdr(skb)->fin) {
1378                         sk_eat_skb(sk, skb, 0);
1379                         ++seq;
1380                         break;
1381                 }
1382                 sk_eat_skb(sk, skb, 0);
1383                 if (!desc->count)
1384                         break;
1385                 tp->copied_seq = seq;
1386         }
1387         tp->copied_seq = seq;
1388
1389         tcp_rcv_space_adjust(sk);
1390
1391         /* Clean up data we have read: This will do ACK frames. */
1392         if (copied > 0)
1393                 tcp_cleanup_rbuf(sk, copied);
1394         return copied;
1395 }
1396 EXPORT_SYMBOL(tcp_read_sock);
1397
1398 /*
1399  *      This routine copies from a sock struct into the user buffer.
1400  *
1401  *      Technical note: in 2.3 we work on _locked_ socket, so that
1402  *      tricks with *seq access order and skb->users are not required.
1403  *      Probably, code can be easily improved even more.
1404  */
1405
1406 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1407                 size_t len, int nonblock, int flags, int *addr_len)
1408 {
1409         struct tcp_sock *tp = tcp_sk(sk);
1410         int copied = 0;
1411         u32 peek_seq;
1412         u32 *seq;
1413         unsigned long used;
1414         int err;
1415         int target;             /* Read at least this many bytes */
1416         long timeo;
1417         struct task_struct *user_recv = NULL;
1418         int copied_early = 0;
1419         struct sk_buff *skb;
1420         u32 urg_hole = 0;
1421
1422         lock_sock(sk);
1423
1424         err = -ENOTCONN;
1425         if (sk->sk_state == TCP_LISTEN)
1426                 goto out;
1427
1428         timeo = sock_rcvtimeo(sk, nonblock);
1429
1430         /* Urgent data needs to be handled specially. */
1431         if (flags & MSG_OOB)
1432                 goto recv_urg;
1433
1434         seq = &tp->copied_seq;
1435         if (flags & MSG_PEEK) {
1436                 peek_seq = tp->copied_seq;
1437                 seq = &peek_seq;
1438         }
1439
1440         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1441
1442 #ifdef CONFIG_NET_DMA
1443         tp->ucopy.dma_chan = NULL;
1444         preempt_disable();
1445         skb = skb_peek_tail(&sk->sk_receive_queue);
1446         {
1447                 int available = 0;
1448
1449                 if (skb)
1450                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1451                 if ((available < target) &&
1452                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1453                     !sysctl_tcp_low_latency &&
1454                     net_dma_find_channel()) {
1455                         preempt_enable_no_resched();
1456                         tp->ucopy.pinned_list =
1457                                         dma_pin_iovec_pages(msg->msg_iov, len);
1458                 } else {
1459                         preempt_enable_no_resched();
1460                 }
1461         }
1462 #endif
1463
1464         do {
1465                 u32 offset;
1466
1467                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1468                 if (tp->urg_data && tp->urg_seq == *seq) {
1469                         if (copied)
1470                                 break;
1471                         if (signal_pending(current)) {
1472                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1473                                 break;
1474                         }
1475                 }
1476
1477                 /* Next get a buffer. */
1478
1479                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1480                         /* Now that we have two receive queues this
1481                          * shouldn't happen.
1482                          */
1483                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1484                                  "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1485                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1486                                  flags))
1487                                 break;
1488
1489                         offset = *seq - TCP_SKB_CB(skb)->seq;
1490                         if (tcp_hdr(skb)->syn)
1491                                 offset--;
1492                         if (offset < skb->len)
1493                                 goto found_ok_skb;
1494                         if (tcp_hdr(skb)->fin)
1495                                 goto found_fin_ok;
1496                         WARN(!(flags & MSG_PEEK),
1497                              "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1498                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1499                 }
1500
1501                 /* Well, if we have backlog, try to process it now yet. */
1502
1503                 if (copied >= target && !sk->sk_backlog.tail)
1504                         break;
1505
1506                 if (copied) {
1507                         if (sk->sk_err ||
1508                             sk->sk_state == TCP_CLOSE ||
1509                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1510                             !timeo ||
1511                             signal_pending(current))
1512                                 break;
1513                 } else {
1514                         if (sock_flag(sk, SOCK_DONE))
1515                                 break;
1516
1517                         if (sk->sk_err) {
1518                                 copied = sock_error(sk);
1519                                 break;
1520                         }
1521
1522                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1523                                 break;
1524
1525                         if (sk->sk_state == TCP_CLOSE) {
1526                                 if (!sock_flag(sk, SOCK_DONE)) {
1527                                         /* This occurs when user tries to read
1528                                          * from never connected socket.
1529                                          */
1530                                         copied = -ENOTCONN;
1531                                         break;
1532                                 }
1533                                 break;
1534                         }
1535
1536                         if (!timeo) {
1537                                 copied = -EAGAIN;
1538                                 break;
1539                         }
1540
1541                         if (signal_pending(current)) {
1542                                 copied = sock_intr_errno(timeo);
1543                                 break;
1544                         }
1545                 }
1546
1547                 tcp_cleanup_rbuf(sk, copied);
1548
1549                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1550                         /* Install new reader */
1551                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1552                                 user_recv = current;
1553                                 tp->ucopy.task = user_recv;
1554                                 tp->ucopy.iov = msg->msg_iov;
1555                         }
1556
1557                         tp->ucopy.len = len;
1558
1559                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1560                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1561
1562                         /* Ugly... If prequeue is not empty, we have to
1563                          * process it before releasing socket, otherwise
1564                          * order will be broken at second iteration.
1565                          * More elegant solution is required!!!
1566                          *
1567                          * Look: we have the following (pseudo)queues:
1568                          *
1569                          * 1. packets in flight
1570                          * 2. backlog
1571                          * 3. prequeue
1572                          * 4. receive_queue
1573                          *
1574                          * Each queue can be processed only if the next ones
1575                          * are empty. At this point we have empty receive_queue.
1576                          * But prequeue _can_ be not empty after 2nd iteration,
1577                          * when we jumped to start of loop because backlog
1578                          * processing added something to receive_queue.
1579                          * We cannot release_sock(), because backlog contains
1580                          * packets arrived _after_ prequeued ones.
1581                          *
1582                          * Shortly, algorithm is clear --- to process all
1583                          * the queues in order. We could make it more directly,
1584                          * requeueing packets from backlog to prequeue, if
1585                          * is not empty. It is more elegant, but eats cycles,
1586                          * unfortunately.
1587                          */
1588                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1589                                 goto do_prequeue;
1590
1591                         /* __ Set realtime policy in scheduler __ */
1592                 }
1593
1594 #ifdef CONFIG_NET_DMA
1595                 if (tp->ucopy.dma_chan)
1596                         dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1597 #endif
1598                 if (copied >= target) {
1599                         /* Do not sleep, just process backlog. */
1600                         release_sock(sk);
1601                         lock_sock(sk);
1602                 } else
1603                         sk_wait_data(sk, &timeo);
1604
1605 #ifdef CONFIG_NET_DMA
1606                 tcp_service_net_dma(sk, false);  /* Don't block */
1607                 tp->ucopy.wakeup = 0;
1608 #endif
1609
1610                 if (user_recv) {
1611                         int chunk;
1612
1613                         /* __ Restore normal policy in scheduler __ */
1614
1615                         if ((chunk = len - tp->ucopy.len) != 0) {
1616                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1617                                 len -= chunk;
1618                                 copied += chunk;
1619                         }
1620
1621                         if (tp->rcv_nxt == tp->copied_seq &&
1622                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1623 do_prequeue:
1624                                 tcp_prequeue_process(sk);
1625
1626                                 if ((chunk = len - tp->ucopy.len) != 0) {
1627                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1628                                         len -= chunk;
1629                                         copied += chunk;
1630                                 }
1631                         }
1632                 }
1633                 if ((flags & MSG_PEEK) &&
1634                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1635                         if (net_ratelimit())
1636                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1637                                        current->comm, task_pid_nr(current));
1638                         peek_seq = tp->copied_seq;
1639                 }
1640                 continue;
1641
1642         found_ok_skb:
1643                 /* Ok so how much can we use? */
1644                 used = skb->len - offset;
1645                 if (len < used)
1646                         used = len;
1647
1648                 /* Do we have urgent data here? */
1649                 if (tp->urg_data) {
1650                         u32 urg_offset = tp->urg_seq - *seq;
1651                         if (urg_offset < used) {
1652                                 if (!urg_offset) {
1653                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1654                                                 ++*seq;
1655                                                 urg_hole++;
1656                                                 offset++;
1657                                                 used--;
1658                                                 if (!used)
1659                                                         goto skip_copy;
1660                                         }
1661                                 } else
1662                                         used = urg_offset;
1663                         }
1664                 }
1665
1666                 if (!(flags & MSG_TRUNC)) {
1667 #ifdef CONFIG_NET_DMA
1668                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1669                                 tp->ucopy.dma_chan = net_dma_find_channel();
1670
1671                         if (tp->ucopy.dma_chan) {
1672                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1673                                         tp->ucopy.dma_chan, skb, offset,
1674                                         msg->msg_iov, used,
1675                                         tp->ucopy.pinned_list);
1676
1677                                 if (tp->ucopy.dma_cookie < 0) {
1678
1679                                         pr_alert("%s: dma_cookie < 0\n",
1680                                                  __func__);
1681
1682                                         /* Exception. Bailout! */
1683                                         if (!copied)
1684                                                 copied = -EFAULT;
1685                                         break;
1686                                 }
1687
1688                                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1689
1690                                 if ((offset + used) == skb->len)
1691                                         copied_early = 1;
1692
1693                         } else
1694 #endif
1695                         {
1696                                 err = skb_copy_datagram_iovec(skb, offset,
1697                                                 msg->msg_iov, used);
1698                                 if (err) {
1699                                         /* Exception. Bailout! */
1700                                         if (!copied)
1701                                                 copied = -EFAULT;
1702                                         break;
1703                                 }
1704                         }
1705                 }
1706
1707                 *seq += used;
1708                 copied += used;
1709                 len -= used;
1710
1711                 tcp_rcv_space_adjust(sk);
1712
1713 skip_copy:
1714                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1715                         tp->urg_data = 0;
1716                         tcp_fast_path_check(sk);
1717                 }
1718                 if (used + offset < skb->len)
1719                         continue;
1720
1721                 if (tcp_hdr(skb)->fin)
1722                         goto found_fin_ok;
1723                 if (!(flags & MSG_PEEK)) {
1724                         sk_eat_skb(sk, skb, copied_early);
1725                         copied_early = 0;
1726                 }
1727                 continue;
1728
1729         found_fin_ok:
1730                 /* Process the FIN. */
1731                 ++*seq;
1732                 if (!(flags & MSG_PEEK)) {
1733                         sk_eat_skb(sk, skb, copied_early);
1734                         copied_early = 0;
1735                 }
1736                 break;
1737         } while (len > 0);
1738
1739         if (user_recv) {
1740                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1741                         int chunk;
1742
1743                         tp->ucopy.len = copied > 0 ? len : 0;
1744
1745                         tcp_prequeue_process(sk);
1746
1747                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1748                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1749                                 len -= chunk;
1750                                 copied += chunk;
1751                         }
1752                 }
1753
1754                 tp->ucopy.task = NULL;
1755                 tp->ucopy.len = 0;
1756         }
1757
1758 #ifdef CONFIG_NET_DMA
1759         tcp_service_net_dma(sk, true);  /* Wait for queue to drain */
1760         tp->ucopy.dma_chan = NULL;
1761
1762         if (tp->ucopy.pinned_list) {
1763                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1764                 tp->ucopy.pinned_list = NULL;
1765         }
1766 #endif
1767
1768         /* According to UNIX98, msg_name/msg_namelen are ignored
1769          * on connected socket. I was just happy when found this 8) --ANK
1770          */
1771
1772         /* Clean up data we have read: This will do ACK frames. */
1773         tcp_cleanup_rbuf(sk, copied);
1774
1775         release_sock(sk);
1776         return copied;
1777
1778 out:
1779         release_sock(sk);
1780         return err;
1781
1782 recv_urg:
1783         err = tcp_recv_urg(sk, msg, len, flags);
1784         goto out;
1785 }
1786 EXPORT_SYMBOL(tcp_recvmsg);
1787
1788 void tcp_set_state(struct sock *sk, int state)
1789 {
1790         int oldstate = sk->sk_state;
1791
1792         switch (state) {
1793         case TCP_ESTABLISHED:
1794                 if (oldstate != TCP_ESTABLISHED)
1795                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1796                 break;
1797
1798         case TCP_CLOSE:
1799                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1800                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1801
1802                 sk->sk_prot->unhash(sk);
1803                 if (inet_csk(sk)->icsk_bind_hash &&
1804                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1805                         inet_put_port(sk);
1806                 /* fall through */
1807         default:
1808                 if (oldstate == TCP_ESTABLISHED)
1809                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1810         }
1811
1812         /* Change state AFTER socket is unhashed to avoid closed
1813          * socket sitting in hash tables.
1814          */
1815         sk->sk_state = state;
1816
1817 #ifdef STATE_TRACE
1818         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1819 #endif
1820 }
1821 EXPORT_SYMBOL_GPL(tcp_set_state);
1822
1823 /*
1824  *      State processing on a close. This implements the state shift for
1825  *      sending our FIN frame. Note that we only send a FIN for some
1826  *      states. A shutdown() may have already sent the FIN, or we may be
1827  *      closed.
1828  */
1829
1830 static const unsigned char new_state[16] = {
1831   /* current state:        new state:      action:      */
1832   /* (Invalid)          */ TCP_CLOSE,
1833   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1834   /* TCP_SYN_SENT       */ TCP_CLOSE,
1835   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1836   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1837   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1838   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1839   /* TCP_CLOSE          */ TCP_CLOSE,
1840   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1841   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1842   /* TCP_LISTEN         */ TCP_CLOSE,
1843   /* TCP_CLOSING        */ TCP_CLOSING,
1844 };
1845
1846 static int tcp_close_state(struct sock *sk)
1847 {
1848         int next = (int)new_state[sk->sk_state];
1849         int ns = next & TCP_STATE_MASK;
1850
1851         tcp_set_state(sk, ns);
1852
1853         return next & TCP_ACTION_FIN;
1854 }
1855
1856 /*
1857  *      Shutdown the sending side of a connection. Much like close except
1858  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1859  */
1860
1861 void tcp_shutdown(struct sock *sk, int how)
1862 {
1863         /*      We need to grab some memory, and put together a FIN,
1864          *      and then put it into the queue to be sent.
1865          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1866          */
1867         if (!(how & SEND_SHUTDOWN))
1868                 return;
1869
1870         /* If we've already sent a FIN, or it's a closed state, skip this. */
1871         if ((1 << sk->sk_state) &
1872             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1873              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1874                 /* Clear out any half completed packets.  FIN if needed. */
1875                 if (tcp_close_state(sk))
1876                         tcp_send_fin(sk);
1877         }
1878 }
1879 EXPORT_SYMBOL(tcp_shutdown);
1880
1881 bool tcp_check_oom(struct sock *sk, int shift)
1882 {
1883         bool too_many_orphans, out_of_socket_memory;
1884
1885         too_many_orphans = tcp_too_many_orphans(sk, shift);
1886         out_of_socket_memory = tcp_out_of_memory(sk);
1887
1888         if (too_many_orphans && net_ratelimit())
1889                 pr_info("too many orphaned sockets\n");
1890         if (out_of_socket_memory && net_ratelimit())
1891                 pr_info("out of memory -- consider tuning tcp_mem\n");
1892         return too_many_orphans || out_of_socket_memory;
1893 }
1894
1895 void tcp_close(struct sock *sk, long timeout)
1896 {
1897         struct sk_buff *skb;
1898         int data_was_unread = 0;
1899         int state;
1900
1901         lock_sock(sk);
1902         sk->sk_shutdown = SHUTDOWN_MASK;
1903
1904         if (sk->sk_state == TCP_LISTEN) {
1905                 tcp_set_state(sk, TCP_CLOSE);
1906
1907                 /* Special case. */
1908                 inet_csk_listen_stop(sk);
1909
1910                 goto adjudge_to_death;
1911         }
1912
1913         /*  We need to flush the recv. buffs.  We do this only on the
1914          *  descriptor close, not protocol-sourced closes, because the
1915          *  reader process may not have drained the data yet!
1916          */
1917         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1918                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1919                           tcp_hdr(skb)->fin;
1920                 data_was_unread += len;
1921                 __kfree_skb(skb);
1922         }
1923
1924         sk_mem_reclaim(sk);
1925
1926         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
1927         if (sk->sk_state == TCP_CLOSE)
1928                 goto adjudge_to_death;
1929
1930         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1931          * data was lost. To witness the awful effects of the old behavior of
1932          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1933          * GET in an FTP client, suspend the process, wait for the client to
1934          * advertise a zero window, then kill -9 the FTP client, wheee...
1935          * Note: timeout is always zero in such a case.
1936          */
1937         if (data_was_unread) {
1938                 /* Unread data was tossed, zap the connection. */
1939                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1940                 tcp_set_state(sk, TCP_CLOSE);
1941                 tcp_send_active_reset(sk, sk->sk_allocation);
1942         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1943                 /* Check zero linger _after_ checking for unread data. */
1944                 sk->sk_prot->disconnect(sk, 0);
1945                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1946         } else if (tcp_close_state(sk)) {
1947                 /* We FIN if the application ate all the data before
1948                  * zapping the connection.
1949                  */
1950
1951                 /* RED-PEN. Formally speaking, we have broken TCP state
1952                  * machine. State transitions:
1953                  *
1954                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1955                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1956                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1957                  *
1958                  * are legal only when FIN has been sent (i.e. in window),
1959                  * rather than queued out of window. Purists blame.
1960                  *
1961                  * F.e. "RFC state" is ESTABLISHED,
1962                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1963                  *
1964                  * The visible declinations are that sometimes
1965                  * we enter time-wait state, when it is not required really
1966                  * (harmless), do not send active resets, when they are
1967                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1968                  * they look as CLOSING or LAST_ACK for Linux)
1969                  * Probably, I missed some more holelets.
1970                  *                                              --ANK
1971                  */
1972                 tcp_send_fin(sk);
1973         }
1974
1975         sk_stream_wait_close(sk, timeout);
1976
1977 adjudge_to_death:
1978         state = sk->sk_state;
1979         sock_hold(sk);
1980         sock_orphan(sk);
1981
1982         /* It is the last release_sock in its life. It will remove backlog. */
1983         release_sock(sk);
1984
1985
1986         /* Now socket is owned by kernel and we acquire BH lock
1987            to finish close. No need to check for user refs.
1988          */
1989         local_bh_disable();
1990         bh_lock_sock(sk);
1991         WARN_ON(sock_owned_by_user(sk));
1992
1993         percpu_counter_inc(sk->sk_prot->orphan_count);
1994
1995         /* Have we already been destroyed by a softirq or backlog? */
1996         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1997                 goto out;
1998
1999         /*      This is a (useful) BSD violating of the RFC. There is a
2000          *      problem with TCP as specified in that the other end could
2001          *      keep a socket open forever with no application left this end.
2002          *      We use a 3 minute timeout (about the same as BSD) then kill
2003          *      our end. If they send after that then tough - BUT: long enough
2004          *      that we won't make the old 4*rto = almost no time - whoops
2005          *      reset mistake.
2006          *
2007          *      Nope, it was not mistake. It is really desired behaviour
2008          *      f.e. on http servers, when such sockets are useless, but
2009          *      consume significant resources. Let's do it with special
2010          *      linger2 option.                                 --ANK
2011          */
2012
2013         if (sk->sk_state == TCP_FIN_WAIT2) {
2014                 struct tcp_sock *tp = tcp_sk(sk);
2015                 if (tp->linger2 < 0) {
2016                         tcp_set_state(sk, TCP_CLOSE);
2017                         tcp_send_active_reset(sk, GFP_ATOMIC);
2018                         NET_INC_STATS_BH(sock_net(sk),
2019                                         LINUX_MIB_TCPABORTONLINGER);
2020                 } else {
2021                         const int tmo = tcp_fin_time(sk);
2022
2023                         if (tmo > TCP_TIMEWAIT_LEN) {
2024                                 inet_csk_reset_keepalive_timer(sk,
2025                                                 tmo - TCP_TIMEWAIT_LEN);
2026                         } else {
2027                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2028                                 goto out;
2029                         }
2030                 }
2031         }
2032         if (sk->sk_state != TCP_CLOSE) {
2033                 sk_mem_reclaim(sk);
2034                 if (tcp_check_oom(sk, 0)) {
2035                         tcp_set_state(sk, TCP_CLOSE);
2036                         tcp_send_active_reset(sk, GFP_ATOMIC);
2037                         NET_INC_STATS_BH(sock_net(sk),
2038                                         LINUX_MIB_TCPABORTONMEMORY);
2039                 }
2040         }
2041
2042         if (sk->sk_state == TCP_CLOSE)
2043                 inet_csk_destroy_sock(sk);
2044         /* Otherwise, socket is reprieved until protocol close. */
2045
2046 out:
2047         bh_unlock_sock(sk);
2048         local_bh_enable();
2049         sock_put(sk);
2050 }
2051 EXPORT_SYMBOL(tcp_close);
2052
2053 /* These states need RST on ABORT according to RFC793 */
2054
2055 static inline int tcp_need_reset(int state)
2056 {
2057         return (1 << state) &
2058                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2059                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2060 }
2061
2062 int tcp_disconnect(struct sock *sk, int flags)
2063 {
2064         struct inet_sock *inet = inet_sk(sk);
2065         struct inet_connection_sock *icsk = inet_csk(sk);
2066         struct tcp_sock *tp = tcp_sk(sk);
2067         int err = 0;
2068         int old_state = sk->sk_state;
2069
2070         if (old_state != TCP_CLOSE)
2071                 tcp_set_state(sk, TCP_CLOSE);
2072
2073         /* ABORT function of RFC793 */
2074         if (old_state == TCP_LISTEN) {
2075                 inet_csk_listen_stop(sk);
2076         } else if (tcp_need_reset(old_state) ||
2077                    (tp->snd_nxt != tp->write_seq &&
2078                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2079                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2080                  * states
2081                  */
2082                 tcp_send_active_reset(sk, gfp_any());
2083                 sk->sk_err = ECONNRESET;
2084         } else if (old_state == TCP_SYN_SENT)
2085                 sk->sk_err = ECONNRESET;
2086
2087         tcp_clear_xmit_timers(sk);
2088         __skb_queue_purge(&sk->sk_receive_queue);
2089         tcp_write_queue_purge(sk);
2090         __skb_queue_purge(&tp->out_of_order_queue);
2091 #ifdef CONFIG_NET_DMA
2092         __skb_queue_purge(&sk->sk_async_wait_queue);
2093 #endif
2094
2095         inet->inet_dport = 0;
2096
2097         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2098                 inet_reset_saddr(sk);
2099
2100         sk->sk_shutdown = 0;
2101         sock_reset_flag(sk, SOCK_DONE);
2102         tp->srtt = 0;
2103         if ((tp->write_seq += tp->max_window + 2) == 0)
2104                 tp->write_seq = 1;
2105         icsk->icsk_backoff = 0;
2106         tp->snd_cwnd = 2;
2107         icsk->icsk_probes_out = 0;
2108         tp->packets_out = 0;
2109         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2110         tp->snd_cwnd_cnt = 0;
2111         tp->bytes_acked = 0;
2112         tp->window_clamp = 0;
2113         tcp_set_ca_state(sk, TCP_CA_Open);
2114         tcp_clear_retrans(tp);
2115         inet_csk_delack_init(sk);
2116         tcp_init_send_head(sk);
2117         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2118         __sk_dst_reset(sk);
2119
2120         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2121
2122         sk->sk_error_report(sk);
2123         return err;
2124 }
2125 EXPORT_SYMBOL(tcp_disconnect);
2126
2127 /*
2128  *      Socket option code for TCP.
2129  */
2130 static int do_tcp_setsockopt(struct sock *sk, int level,
2131                 int optname, char __user *optval, unsigned int optlen)
2132 {
2133         struct tcp_sock *tp = tcp_sk(sk);
2134         struct inet_connection_sock *icsk = inet_csk(sk);
2135         int val;
2136         int err = 0;
2137
2138         /* These are data/string values, all the others are ints */
2139         switch (optname) {
2140         case TCP_CONGESTION: {
2141                 char name[TCP_CA_NAME_MAX];
2142
2143                 if (optlen < 1)
2144                         return -EINVAL;
2145
2146                 val = strncpy_from_user(name, optval,
2147                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2148                 if (val < 0)
2149                         return -EFAULT;
2150                 name[val] = 0;
2151
2152                 lock_sock(sk);
2153                 err = tcp_set_congestion_control(sk, name);
2154                 release_sock(sk);
2155                 return err;
2156         }
2157         case TCP_COOKIE_TRANSACTIONS: {
2158                 struct tcp_cookie_transactions ctd;
2159                 struct tcp_cookie_values *cvp = NULL;
2160
2161                 if (sizeof(ctd) > optlen)
2162                         return -EINVAL;
2163                 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2164                         return -EFAULT;
2165
2166                 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2167                     ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2168                         return -EINVAL;
2169
2170                 if (ctd.tcpct_cookie_desired == 0) {
2171                         /* default to global value */
2172                 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2173                            ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2174                            ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2175                         return -EINVAL;
2176                 }
2177
2178                 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2179                         /* Supercedes all other values */
2180                         lock_sock(sk);
2181                         if (tp->cookie_values != NULL) {
2182                                 kref_put(&tp->cookie_values->kref,
2183                                          tcp_cookie_values_release);
2184                                 tp->cookie_values = NULL;
2185                         }
2186                         tp->rx_opt.cookie_in_always = 0; /* false */
2187                         tp->rx_opt.cookie_out_never = 1; /* true */
2188                         release_sock(sk);
2189                         return err;
2190                 }
2191
2192                 /* Allocate ancillary memory before locking.
2193                  */
2194                 if (ctd.tcpct_used > 0 ||
2195                     (tp->cookie_values == NULL &&
2196                      (sysctl_tcp_cookie_size > 0 ||
2197                       ctd.tcpct_cookie_desired > 0 ||
2198                       ctd.tcpct_s_data_desired > 0))) {
2199                         cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2200                                       GFP_KERNEL);
2201                         if (cvp == NULL)
2202                                 return -ENOMEM;
2203
2204                         kref_init(&cvp->kref);
2205                 }
2206                 lock_sock(sk);
2207                 tp->rx_opt.cookie_in_always =
2208                         (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2209                 tp->rx_opt.cookie_out_never = 0; /* false */
2210
2211                 if (tp->cookie_values != NULL) {
2212                         if (cvp != NULL) {
2213                                 /* Changed values are recorded by a changed
2214                                  * pointer, ensuring the cookie will differ,
2215                                  * without separately hashing each value later.
2216                                  */
2217                                 kref_put(&tp->cookie_values->kref,
2218                                          tcp_cookie_values_release);
2219                         } else {
2220                                 cvp = tp->cookie_values;
2221                         }
2222                 }
2223
2224                 if (cvp != NULL) {
2225                         cvp->cookie_desired = ctd.tcpct_cookie_desired;
2226
2227                         if (ctd.tcpct_used > 0) {
2228                                 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2229                                        ctd.tcpct_used);
2230                                 cvp->s_data_desired = ctd.tcpct_used;
2231                                 cvp->s_data_constant = 1; /* true */
2232                         } else {
2233                                 /* No constant payload data. */
2234                                 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2235                                 cvp->s_data_constant = 0; /* false */
2236                         }
2237
2238                         tp->cookie_values = cvp;
2239                 }
2240                 release_sock(sk);
2241                 return err;
2242         }
2243         default:
2244                 /* fallthru */
2245                 break;
2246         }
2247
2248         if (optlen < sizeof(int))
2249                 return -EINVAL;
2250
2251         if (get_user(val, (int __user *)optval))
2252                 return -EFAULT;
2253
2254         lock_sock(sk);
2255
2256         switch (optname) {
2257         case TCP_MAXSEG:
2258                 /* Values greater than interface MTU won't take effect. However
2259                  * at the point when this call is done we typically don't yet
2260                  * know which interface is going to be used */
2261                 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2262                         err = -EINVAL;
2263                         break;
2264                 }
2265                 tp->rx_opt.user_mss = val;
2266                 break;
2267
2268         case TCP_NODELAY:
2269                 if (val) {
2270                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2271                          * this option on corked socket is remembered, but
2272                          * it is not activated until cork is cleared.
2273                          *
2274                          * However, when TCP_NODELAY is set we make
2275                          * an explicit push, which overrides even TCP_CORK
2276                          * for currently queued segments.
2277                          */
2278                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2279                         tcp_push_pending_frames(sk);
2280                 } else {
2281                         tp->nonagle &= ~TCP_NAGLE_OFF;
2282                 }
2283                 break;
2284
2285         case TCP_THIN_LINEAR_TIMEOUTS:
2286                 if (val < 0 || val > 1)
2287                         err = -EINVAL;
2288                 else
2289                         tp->thin_lto = val;
2290                 break;
2291
2292         case TCP_THIN_DUPACK:
2293                 if (val < 0 || val > 1)
2294                         err = -EINVAL;
2295                 else
2296                         tp->thin_dupack = val;
2297                 break;
2298
2299         case TCP_CORK:
2300                 /* When set indicates to always queue non-full frames.
2301                  * Later the user clears this option and we transmit
2302                  * any pending partial frames in the queue.  This is
2303                  * meant to be used alongside sendfile() to get properly
2304                  * filled frames when the user (for example) must write
2305                  * out headers with a write() call first and then use
2306                  * sendfile to send out the data parts.
2307                  *
2308                  * TCP_CORK can be set together with TCP_NODELAY and it is
2309                  * stronger than TCP_NODELAY.
2310                  */
2311                 if (val) {
2312                         tp->nonagle |= TCP_NAGLE_CORK;
2313                 } else {
2314                         tp->nonagle &= ~TCP_NAGLE_CORK;
2315                         if (tp->nonagle&TCP_NAGLE_OFF)
2316                                 tp->nonagle |= TCP_NAGLE_PUSH;
2317                         tcp_push_pending_frames(sk);
2318                 }
2319                 break;
2320
2321         case TCP_KEEPIDLE:
2322                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2323                         err = -EINVAL;
2324                 else {
2325                         tp->keepalive_time = val * HZ;
2326                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2327                             !((1 << sk->sk_state) &
2328                               (TCPF_CLOSE | TCPF_LISTEN))) {
2329                                 u32 elapsed = keepalive_time_elapsed(tp);
2330                                 if (tp->keepalive_time > elapsed)
2331                                         elapsed = tp->keepalive_time - elapsed;
2332                                 else
2333                                         elapsed = 0;
2334                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2335                         }
2336                 }
2337                 break;
2338         case TCP_KEEPINTVL:
2339                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2340                         err = -EINVAL;
2341                 else
2342                         tp->keepalive_intvl = val * HZ;
2343                 break;
2344         case TCP_KEEPCNT:
2345                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2346                         err = -EINVAL;
2347                 else
2348                         tp->keepalive_probes = val;
2349                 break;
2350         case TCP_SYNCNT:
2351                 if (val < 1 || val > MAX_TCP_SYNCNT)
2352                         err = -EINVAL;
2353                 else
2354                         icsk->icsk_syn_retries = val;
2355                 break;
2356
2357         case TCP_LINGER2:
2358                 if (val < 0)
2359                         tp->linger2 = -1;
2360                 else if (val > sysctl_tcp_fin_timeout / HZ)
2361                         tp->linger2 = 0;
2362                 else
2363                         tp->linger2 = val * HZ;
2364                 break;
2365
2366         case TCP_DEFER_ACCEPT:
2367                 /* Translate value in seconds to number of retransmits */
2368                 icsk->icsk_accept_queue.rskq_defer_accept =
2369                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2370                                         TCP_RTO_MAX / HZ);
2371                 break;
2372
2373         case TCP_WINDOW_CLAMP:
2374                 if (!val) {
2375                         if (sk->sk_state != TCP_CLOSE) {
2376                                 err = -EINVAL;
2377                                 break;
2378                         }
2379                         tp->window_clamp = 0;
2380                 } else
2381                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2382                                                 SOCK_MIN_RCVBUF / 2 : val;
2383                 break;
2384
2385         case TCP_QUICKACK:
2386                 if (!val) {
2387                         icsk->icsk_ack.pingpong = 1;
2388                 } else {
2389                         icsk->icsk_ack.pingpong = 0;
2390                         if ((1 << sk->sk_state) &
2391                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2392                             inet_csk_ack_scheduled(sk)) {
2393                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2394                                 tcp_cleanup_rbuf(sk, 1);
2395                                 if (!(val & 1))
2396                                         icsk->icsk_ack.pingpong = 1;
2397                         }
2398                 }
2399                 break;
2400
2401 #ifdef CONFIG_TCP_MD5SIG
2402         case TCP_MD5SIG:
2403                 /* Read the IP->Key mappings from userspace */
2404                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2405                 break;
2406 #endif
2407         case TCP_USER_TIMEOUT:
2408                 /* Cap the max timeout in ms TCP will retry/retrans
2409                  * before giving up and aborting (ETIMEDOUT) a connection.
2410                  */
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,
2673         netdev_features_t features)
2674 {
2675         struct sk_buff *segs = ERR_PTR(-EINVAL);
2676         struct tcphdr *th;
2677         unsigned thlen;
2678         unsigned int seq;
2679         __be32 delta;
2680         unsigned int oldlen;
2681         unsigned int mss;
2682
2683         if (!pskb_may_pull(skb, sizeof(*th)))
2684                 goto out;
2685
2686         th = tcp_hdr(skb);
2687         thlen = th->doff * 4;
2688         if (thlen < sizeof(*th))
2689                 goto out;
2690
2691         if (!pskb_may_pull(skb, thlen))
2692                 goto out;
2693
2694         oldlen = (u16)~skb->len;
2695         __skb_pull(skb, thlen);
2696
2697         mss = skb_shinfo(skb)->gso_size;
2698         if (unlikely(skb->len <= mss))
2699                 goto out;
2700
2701         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2702                 /* Packet is from an untrusted source, reset gso_segs. */
2703                 int type = skb_shinfo(skb)->gso_type;
2704
2705                 if (unlikely(type &
2706                              ~(SKB_GSO_TCPV4 |
2707                                SKB_GSO_DODGY |
2708                                SKB_GSO_TCP_ECN |
2709                                SKB_GSO_TCPV6 |
2710                                0) ||
2711                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2712                         goto out;
2713
2714                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2715
2716                 segs = NULL;
2717                 goto out;
2718         }
2719
2720         segs = skb_segment(skb, features);
2721         if (IS_ERR(segs))
2722                 goto out;
2723
2724         delta = htonl(oldlen + (thlen + mss));
2725
2726         skb = segs;
2727         th = tcp_hdr(skb);
2728         seq = ntohl(th->seq);
2729
2730         do {
2731                 th->fin = th->psh = 0;
2732
2733                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2734                                        (__force u32)delta));
2735                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2736                         th->check =
2737                              csum_fold(csum_partial(skb_transport_header(skb),
2738                                                     thlen, skb->csum));
2739
2740                 seq += mss;
2741                 skb = skb->next;
2742                 th = tcp_hdr(skb);
2743
2744                 th->seq = htonl(seq);
2745                 th->cwr = 0;
2746         } while (skb->next);
2747
2748         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2749                       skb->data_len);
2750         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2751                                 (__force u32)delta));
2752         if (skb->ip_summed != CHECKSUM_PARTIAL)
2753                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2754                                                    thlen, skb->csum));
2755
2756 out:
2757         return segs;
2758 }
2759 EXPORT_SYMBOL(tcp_tso_segment);
2760
2761 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2762 {
2763         struct sk_buff **pp = NULL;
2764         struct sk_buff *p;
2765         struct tcphdr *th;
2766         struct tcphdr *th2;
2767         unsigned int len;
2768         unsigned int thlen;
2769         __be32 flags;
2770         unsigned int mss = 1;
2771         unsigned int hlen;
2772         unsigned int off;
2773         int flush = 1;
2774         int i;
2775
2776         off = skb_gro_offset(skb);
2777         hlen = off + sizeof(*th);
2778         th = skb_gro_header_fast(skb, off);
2779         if (skb_gro_header_hard(skb, hlen)) {
2780                 th = skb_gro_header_slow(skb, hlen, off);
2781                 if (unlikely(!th))
2782                         goto out;
2783         }
2784
2785         thlen = th->doff * 4;
2786         if (thlen < sizeof(*th))
2787                 goto out;
2788
2789         hlen = off + thlen;
2790         if (skb_gro_header_hard(skb, hlen)) {
2791                 th = skb_gro_header_slow(skb, hlen, off);
2792                 if (unlikely(!th))
2793                         goto out;
2794         }
2795
2796         skb_gro_pull(skb, thlen);
2797
2798         len = skb_gro_len(skb);
2799         flags = tcp_flag_word(th);
2800
2801         for (; (p = *head); head = &p->next) {
2802                 if (!NAPI_GRO_CB(p)->same_flow)
2803                         continue;
2804
2805                 th2 = tcp_hdr(p);
2806
2807                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2808                         NAPI_GRO_CB(p)->same_flow = 0;
2809                         continue;
2810                 }
2811
2812                 goto found;
2813         }
2814
2815         goto out_check_final;
2816
2817 found:
2818         flush = NAPI_GRO_CB(p)->flush;
2819         flush |= (__force int)(flags & TCP_FLAG_CWR);
2820         flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
2821                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
2822         flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
2823         for (i = sizeof(*th); i < thlen; i += 4)
2824                 flush |= *(u32 *)((u8 *)th + i) ^
2825                          *(u32 *)((u8 *)th2 + i);
2826
2827         mss = skb_shinfo(p)->gso_size;
2828
2829         flush |= (len - 1) >= mss;
2830         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2831
2832         if (flush || skb_gro_receive(head, skb)) {
2833                 mss = 1;
2834                 goto out_check_final;
2835         }
2836
2837         p = *head;
2838         th2 = tcp_hdr(p);
2839         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2840
2841 out_check_final:
2842         flush = len < mss;
2843         flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
2844                                         TCP_FLAG_RST | TCP_FLAG_SYN |
2845                                         TCP_FLAG_FIN));
2846
2847         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2848                 pp = head;
2849
2850 out:
2851         NAPI_GRO_CB(skb)->flush |= flush;
2852
2853         return pp;
2854 }
2855 EXPORT_SYMBOL(tcp_gro_receive);
2856
2857 int tcp_gro_complete(struct sk_buff *skb)
2858 {
2859         struct tcphdr *th = tcp_hdr(skb);
2860
2861         skb->csum_start = skb_transport_header(skb) - skb->head;
2862         skb->csum_offset = offsetof(struct tcphdr, check);
2863         skb->ip_summed = CHECKSUM_PARTIAL;
2864
2865         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2866
2867         if (th->cwr)
2868                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2869
2870         return 0;
2871 }
2872 EXPORT_SYMBOL(tcp_gro_complete);
2873
2874 #ifdef CONFIG_TCP_MD5SIG
2875 static unsigned long tcp_md5sig_users;
2876 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool;
2877 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2878
2879 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
2880 {
2881         int cpu;
2882
2883         for_each_possible_cpu(cpu) {
2884                 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
2885
2886                 if (p->md5_desc.tfm)
2887                         crypto_free_hash(p->md5_desc.tfm);
2888         }
2889         free_percpu(pool);
2890 }
2891
2892 void tcp_free_md5sig_pool(void)
2893 {
2894         struct tcp_md5sig_pool __percpu *pool = NULL;
2895
2896         spin_lock_bh(&tcp_md5sig_pool_lock);
2897         if (--tcp_md5sig_users == 0) {
2898                 pool = tcp_md5sig_pool;
2899                 tcp_md5sig_pool = NULL;
2900         }
2901         spin_unlock_bh(&tcp_md5sig_pool_lock);
2902         if (pool)
2903                 __tcp_free_md5sig_pool(pool);
2904 }
2905 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2906
2907 static struct tcp_md5sig_pool __percpu *
2908 __tcp_alloc_md5sig_pool(struct sock *sk)
2909 {
2910         int cpu;
2911         struct tcp_md5sig_pool __percpu *pool;
2912
2913         pool = alloc_percpu(struct tcp_md5sig_pool);
2914         if (!pool)
2915                 return NULL;
2916
2917         for_each_possible_cpu(cpu) {
2918                 struct crypto_hash *hash;
2919
2920                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2921                 if (!hash || IS_ERR(hash))
2922                         goto out_free;
2923
2924                 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
2925         }
2926         return pool;
2927 out_free:
2928         __tcp_free_md5sig_pool(pool);
2929         return NULL;
2930 }
2931
2932 struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
2933 {
2934         struct tcp_md5sig_pool __percpu *pool;
2935         int alloc = 0;
2936
2937 retry:
2938         spin_lock_bh(&tcp_md5sig_pool_lock);
2939         pool = tcp_md5sig_pool;
2940         if (tcp_md5sig_users++ == 0) {
2941                 alloc = 1;
2942                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2943         } else if (!pool) {
2944                 tcp_md5sig_users--;
2945                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2946                 cpu_relax();
2947                 goto retry;
2948         } else
2949                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2950
2951         if (alloc) {
2952                 /* we cannot hold spinlock here because this may sleep. */
2953                 struct tcp_md5sig_pool __percpu *p;
2954
2955                 p = __tcp_alloc_md5sig_pool(sk);
2956                 spin_lock_bh(&tcp_md5sig_pool_lock);
2957                 if (!p) {
2958                         tcp_md5sig_users--;
2959                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2960                         return NULL;
2961                 }
2962                 pool = tcp_md5sig_pool;
2963                 if (pool) {
2964                         /* oops, it has already been assigned. */
2965                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2966                         __tcp_free_md5sig_pool(p);
2967                 } else {
2968                         tcp_md5sig_pool = pool = p;
2969                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2970                 }
2971         }
2972         return pool;
2973 }
2974 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2975
2976
2977 /**
2978  *      tcp_get_md5sig_pool - get md5sig_pool for this user
2979  *
2980  *      We use percpu structure, so if we succeed, we exit with preemption
2981  *      and BH disabled, to make sure another thread or softirq handling
2982  *      wont try to get same context.
2983  */
2984 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2985 {
2986         struct tcp_md5sig_pool __percpu *p;
2987
2988         local_bh_disable();
2989
2990         spin_lock(&tcp_md5sig_pool_lock);
2991         p = tcp_md5sig_pool;
2992         if (p)
2993                 tcp_md5sig_users++;
2994         spin_unlock(&tcp_md5sig_pool_lock);
2995
2996         if (p)
2997                 return this_cpu_ptr(p);
2998
2999         local_bh_enable();
3000         return NULL;
3001 }
3002 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3003
3004 void tcp_put_md5sig_pool(void)
3005 {
3006         local_bh_enable();
3007         tcp_free_md5sig_pool();
3008 }
3009 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3010
3011 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3012                         const struct tcphdr *th)
3013 {
3014         struct scatterlist sg;
3015         struct tcphdr hdr;
3016         int err;
3017
3018         /* We are not allowed to change tcphdr, make a local copy */
3019         memcpy(&hdr, th, sizeof(hdr));
3020         hdr.check = 0;
3021
3022         /* options aren't included in the hash */
3023         sg_init_one(&sg, &hdr, sizeof(hdr));
3024         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3025         return err;
3026 }
3027 EXPORT_SYMBOL(tcp_md5_hash_header);
3028
3029 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3030                           const struct sk_buff *skb, unsigned int header_len)
3031 {
3032         struct scatterlist sg;
3033         const struct tcphdr *tp = tcp_hdr(skb);
3034         struct hash_desc *desc = &hp->md5_desc;
3035         unsigned i;
3036         const unsigned head_data_len = skb_headlen(skb) > header_len ?
3037                                        skb_headlen(skb) - header_len : 0;
3038         const struct skb_shared_info *shi = skb_shinfo(skb);
3039         struct sk_buff *frag_iter;
3040
3041         sg_init_table(&sg, 1);
3042
3043         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3044         if (crypto_hash_update(desc, &sg, head_data_len))
3045                 return 1;
3046
3047         for (i = 0; i < shi->nr_frags; ++i) {
3048                 const struct skb_frag_struct *f = &shi->frags[i];
3049                 struct page *page = skb_frag_page(f);
3050                 sg_set_page(&sg, page, skb_frag_size(f), f->page_offset);
3051                 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3052                         return 1;
3053         }
3054
3055         skb_walk_frags(skb, frag_iter)
3056                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3057                         return 1;
3058
3059         return 0;
3060 }
3061 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3062
3063 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3064 {
3065         struct scatterlist sg;
3066
3067         sg_init_one(&sg, key->key, key->keylen);
3068         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3069 }
3070 EXPORT_SYMBOL(tcp_md5_hash_key);
3071
3072 #endif
3073
3074 /**
3075  * Each Responder maintains up to two secret values concurrently for
3076  * efficient secret rollover.  Each secret value has 4 states:
3077  *
3078  * Generating.  (tcp_secret_generating != tcp_secret_primary)
3079  *    Generates new Responder-Cookies, but not yet used for primary
3080  *    verification.  This is a short-term state, typically lasting only
3081  *    one round trip time (RTT).
3082  *
3083  * Primary.  (tcp_secret_generating == tcp_secret_primary)
3084  *    Used both for generation and primary verification.
3085  *
3086  * Retiring.  (tcp_secret_retiring != tcp_secret_secondary)
3087  *    Used for verification, until the first failure that can be
3088  *    verified by the newer Generating secret.  At that time, this
3089  *    cookie's state is changed to Secondary, and the Generating
3090  *    cookie's state is changed to Primary.  This is a short-term state,
3091  *    typically lasting only one round trip time (RTT).
3092  *
3093  * Secondary.  (tcp_secret_retiring == tcp_secret_secondary)
3094  *    Used for secondary verification, after primary verification
3095  *    failures.  This state lasts no more than twice the Maximum Segment
3096  *    Lifetime (2MSL).  Then, the secret is discarded.
3097  */
3098 struct tcp_cookie_secret {
3099         /* The secret is divided into two parts.  The digest part is the
3100          * equivalent of previously hashing a secret and saving the state,
3101          * and serves as an initialization vector (IV).  The message part
3102          * serves as the trailing secret.
3103          */
3104         u32                             secrets[COOKIE_WORKSPACE_WORDS];
3105         unsigned long                   expires;
3106 };
3107
3108 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3109 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3110 #define TCP_SECRET_LIFE (HZ * 600)
3111
3112 static struct tcp_cookie_secret tcp_secret_one;
3113 static struct tcp_cookie_secret tcp_secret_two;
3114
3115 /* Essentially a circular list, without dynamic allocation. */
3116 static struct tcp_cookie_secret *tcp_secret_generating;
3117 static struct tcp_cookie_secret *tcp_secret_primary;
3118 static struct tcp_cookie_secret *tcp_secret_retiring;
3119 static struct tcp_cookie_secret *tcp_secret_secondary;
3120
3121 static DEFINE_SPINLOCK(tcp_secret_locker);
3122
3123 /* Select a pseudo-random word in the cookie workspace.
3124  */
3125 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3126 {
3127         return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3128 }
3129
3130 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3131  * Called in softirq context.
3132  * Returns: 0 for success.
3133  */
3134 int tcp_cookie_generator(u32 *bakery)
3135 {
3136         unsigned long jiffy = jiffies;
3137
3138         if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3139                 spin_lock_bh(&tcp_secret_locker);
3140                 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3141                         /* refreshed by another */
3142                         memcpy(bakery,
3143                                &tcp_secret_generating->secrets[0],
3144                                COOKIE_WORKSPACE_WORDS);
3145                 } else {
3146                         /* still needs refreshing */
3147                         get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3148
3149                         /* The first time, paranoia assumes that the
3150                          * randomization function isn't as strong.  But,
3151                          * this secret initialization is delayed until
3152                          * the last possible moment (packet arrival).
3153                          * Although that time is observable, it is
3154                          * unpredictably variable.  Mash in the most
3155                          * volatile clock bits available, and expire the
3156                          * secret extra quickly.
3157                          */
3158                         if (unlikely(tcp_secret_primary->expires ==
3159                                      tcp_secret_secondary->expires)) {
3160                                 struct timespec tv;
3161
3162                                 getnstimeofday(&tv);
3163                                 bakery[COOKIE_DIGEST_WORDS+0] ^=
3164                                         (u32)tv.tv_nsec;
3165
3166                                 tcp_secret_secondary->expires = jiffy
3167                                         + TCP_SECRET_1MSL
3168                                         + (0x0f & tcp_cookie_work(bakery, 0));
3169                         } else {
3170                                 tcp_secret_secondary->expires = jiffy
3171                                         + TCP_SECRET_LIFE
3172                                         + (0xff & tcp_cookie_work(bakery, 1));
3173                                 tcp_secret_primary->expires = jiffy
3174                                         + TCP_SECRET_2MSL
3175                                         + (0x1f & tcp_cookie_work(bakery, 2));
3176                         }
3177                         memcpy(&tcp_secret_secondary->secrets[0],
3178                                bakery, COOKIE_WORKSPACE_WORDS);
3179
3180                         rcu_assign_pointer(tcp_secret_generating,
3181                                            tcp_secret_secondary);
3182                         rcu_assign_pointer(tcp_secret_retiring,
3183                                            tcp_secret_primary);
3184                         /*
3185                          * Neither call_rcu() nor synchronize_rcu() needed.
3186                          * Retiring data is not freed.  It is replaced after
3187                          * further (locked) pointer updates, and a quiet time
3188                          * (minimum 1MSL, maximum LIFE - 2MSL).
3189                          */
3190                 }
3191                 spin_unlock_bh(&tcp_secret_locker);
3192         } else {
3193                 rcu_read_lock_bh();
3194                 memcpy(bakery,
3195                        &rcu_dereference(tcp_secret_generating)->secrets[0],
3196                        COOKIE_WORKSPACE_WORDS);
3197                 rcu_read_unlock_bh();
3198         }
3199         return 0;
3200 }
3201 EXPORT_SYMBOL(tcp_cookie_generator);
3202
3203 void tcp_done(struct sock *sk)
3204 {
3205         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3206                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3207
3208         tcp_set_state(sk, TCP_CLOSE);
3209         tcp_clear_xmit_timers(sk);
3210
3211         sk->sk_shutdown = SHUTDOWN_MASK;
3212
3213         if (!sock_flag(sk, SOCK_DEAD))
3214                 sk->sk_state_change(sk);
3215         else
3216                 inet_csk_destroy_sock(sk);
3217 }
3218 EXPORT_SYMBOL_GPL(tcp_done);
3219
3220 extern struct tcp_congestion_ops tcp_reno;
3221
3222 static __initdata unsigned long thash_entries;
3223 static int __init set_thash_entries(char *str)
3224 {
3225         if (!str)
3226                 return 0;
3227         thash_entries = simple_strtoul(str, &str, 0);
3228         return 1;
3229 }
3230 __setup("thash_entries=", set_thash_entries);
3231
3232 void tcp_init_mem(struct net *net)
3233 {
3234         unsigned long limit = nr_free_buffer_pages() / 8;
3235         limit = max(limit, 128UL);
3236         net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3;
3237         net->ipv4.sysctl_tcp_mem[1] = limit;
3238         net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2;
3239 }
3240
3241 void __init tcp_init(void)
3242 {
3243         struct sk_buff *skb = NULL;
3244         unsigned long limit;
3245         int max_rshare, max_wshare, cnt;
3246         unsigned int i;
3247         unsigned long jiffy = jiffies;
3248
3249         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3250
3251         percpu_counter_init(&tcp_sockets_allocated, 0);
3252         percpu_counter_init(&tcp_orphan_count, 0);
3253         tcp_hashinfo.bind_bucket_cachep =
3254                 kmem_cache_create("tcp_bind_bucket",
3255                                   sizeof(struct inet_bind_bucket), 0,
3256                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3257
3258         /* Size and allocate the main established and bind bucket
3259          * hash tables.
3260          *
3261          * The methodology is similar to that of the buffer cache.
3262          */
3263         tcp_hashinfo.ehash =
3264                 alloc_large_system_hash("TCP established",
3265                                         sizeof(struct inet_ehash_bucket),
3266                                         thash_entries,
3267                                         (totalram_pages >= 128 * 1024) ?
3268                                         13 : 15,
3269                                         0,
3270                                         NULL,
3271                                         &tcp_hashinfo.ehash_mask,
3272                                         thash_entries ? 0 : 512 * 1024);
3273         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3274                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3275                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3276         }
3277         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3278                 panic("TCP: failed to alloc ehash_locks");
3279         tcp_hashinfo.bhash =
3280                 alloc_large_system_hash("TCP bind",
3281                                         sizeof(struct inet_bind_hashbucket),
3282                                         tcp_hashinfo.ehash_mask + 1,
3283                                         (totalram_pages >= 128 * 1024) ?
3284                                         13 : 15,
3285                                         0,
3286                                         &tcp_hashinfo.bhash_size,
3287                                         NULL,
3288                                         64 * 1024);
3289         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3290         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3291                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3292                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3293         }
3294
3295
3296         cnt = tcp_hashinfo.ehash_mask + 1;
3297
3298         tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3299         sysctl_tcp_max_orphans = cnt / 2;
3300         sysctl_max_syn_backlog = max(128, cnt / 256);
3301
3302         tcp_init_mem(&init_net);
3303         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3304         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3305         max_wshare = min(4UL*1024*1024, limit);
3306         max_rshare = min(6UL*1024*1024, limit);
3307
3308         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3309         sysctl_tcp_wmem[1] = 16*1024;
3310         sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3311
3312         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3313         sysctl_tcp_rmem[1] = 87380;
3314         sysctl_tcp_rmem[2] = max(87380, max_rshare);
3315
3316         pr_info("Hash tables configured (established %u bind %u)\n",
3317                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3318
3319         tcp_register_congestion_control(&tcp_reno);
3320
3321         memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3322         memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3323         tcp_secret_one.expires = jiffy; /* past due */
3324         tcp_secret_two.expires = jiffy; /* past due */
3325         tcp_secret_generating = &tcp_secret_one;
3326         tcp_secret_primary = &tcp_secret_one;
3327         tcp_secret_retiring = &tcp_secret_two;
3328         tcp_secret_secondary = &tcp_secret_two;
3329 }