net/ipv4/tcp.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the  BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              Implementation of the Transmission Control Protocol(TCP).
   7  *
   8  * Authors:     Ross Biro
   9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
  11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
  12  *              Florian La Roche, <flla@stud.uni-sb.de>
  13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
  15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
  16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
  17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  18  *              Jorge Cwik, <jorge@laser.satlink.net>
  19  *
  20  * Fixes:
  21  *              Alan Cox        :       Numerous verify_area() calls
  22  *              Alan Cox        :       Set the ACK bit on a reset
  23  *              Alan Cox        :       Stopped it crashing if it closed while
  24  *                                      sk->inuse=1 and was trying to connect
  25  *                                      (tcp_err()).
  26  *              Alan Cox        :       All icmp error handling was broken
  27  *                                      pointers passed where wrong and the
  28  *                                      socket was looked up backwards. Nobody
  29  *                                      tested any icmp error code obviously.
  30  *              Alan Cox        :       tcp_err() now handled properly. It
  31  *                                      wakes people on errors. poll
  32  *                                      behaves and the icmp error race
  33  *                                      has gone by moving it into sock.c
  34  *              Alan Cox        :       tcp_send_reset() fixed to work for
  35  *                                      everything not just packets for
  36  *                                      unknown sockets.
  37  *              Alan Cox        :       tcp option processing.
  38  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
  39  *                                      syn rule wrong]
  40  *              Herp Rosmanith  :       More reset fixes
  41  *              Alan Cox        :       No longer acks invalid rst frames.
  42  *                                      Acking any kind of RST is right out.
  43  *              Alan Cox        :       Sets an ignore me flag on an rst
  44  *                                      receive otherwise odd bits of prattle
  45  *                                      escape still
  46  *              Alan Cox        :       Fixed another acking RST frame bug.
  47  *                                      Should stop LAN workplace lockups.
  48  *              Alan Cox        :       Some tidyups using the new skb list
  49  *                                      facilities
  50  *              Alan Cox        :       sk->keepopen now seems to work
  51  *              Alan Cox        :       Pulls options out correctly on accepts
  52  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
  53  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
  54  *                                      bit to skb ops.
  55  *              Alan Cox        :       Tidied tcp_data to avoid a potential
  56  *                                      nasty.
  57  *              Alan Cox        :       Added some better commenting, as the
  58  *                                      tcp is hard to follow
  59  *              Alan Cox        :       Removed incorrect check for 20 * psh
  60  *      Michael O'Reilly        :       ack < copied bug fix.
  61  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
  62  *              Alan Cox        :       FIN with no memory -> CRASH
  63  *              Alan Cox        :       Added socket option proto entries.
  64  *                                      Also added awareness of them to accept.
  65  *              Alan Cox        :       Added TCP options (SOL_TCP)
  66  *              Alan Cox        :       Switched wakeup calls to callbacks,
  67  *                                      so the kernel can layer network
  68  *                                      sockets.
  69  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
  70  *              Alan Cox        :       Handle FIN (more) properly (we hope).
  71  *              Alan Cox        :       RST frames sent on unsynchronised
  72  *                                      state ack error.
  73  *              Alan Cox        :       Put in missing check for SYN bit.
  74  *              Alan Cox        :       Added tcp_select_window() aka NET2E
  75  *                                      window non shrink trick.
  76  *              Alan Cox        :       Added a couple of small NET2E timer
  77  *                                      fixes
  78  *              Charles Hedrick :       TCP fixes
  79  *              Toomas Tamm     :       TCP window fixes
  80  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
  81  *              Charles Hedrick :       Rewrote most of it to actually work
  82  *              Linus           :       Rewrote tcp_read() and URG handling
  83  *                                      completely
  84  *              Gerhard Koerting:       Fixed some missing timer handling
  85  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
  86  *              Gerhard Koerting:       PC/TCP workarounds
  87  *              Adam Caldwell   :       Assorted timer/timing errors
  88  *              Matthew Dillon  :       Fixed another RST bug
  89  *              Alan Cox        :       Move to kernel side addressing changes.
  90  *              Alan Cox        :       Beginning work on TCP fastpathing
  91  *                                      (not yet usable)
  92  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
  93  *              Alan Cox        :       TCP fast path debugging
  94  *              Alan Cox        :       Window clamping
  95  *              Michael Riepe   :       Bug in tcp_check()
  96  *              Matt Dillon     :       More TCP improvements and RST bug fixes
  97  *              Matt Dillon     :       Yet more small nasties remove from the
  98  *                                      TCP code (Be very nice to this man if
  99  *                                      tcp finally works 100%) 8)
 100  *              Alan Cox        :       BSD accept semantics.
 101  *              Alan Cox        :       Reset on closedown bug.
 102  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
 103  *              Michael Pall    :       Handle poll() after URG properly in
 104  *                                      all cases.
 105  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
 106  *                                      (multi URG PUSH broke rlogin).
 107  *              Michael Pall    :       Fix the multi URG PUSH problem in
 108  *                                      tcp_readable(), poll() after URG
 109  *                                      works now.
 110  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
 111  *                                      BSD api.
 112  *              Alan Cox        :       Changed the semantics of sk->socket to
 113  *                                      fix a race and a signal problem with
 114  *                                      accept() and async I/O.
 115  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
 116  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
 117  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
 118  *                                      clients/servers which listen in on
 119  *                                      fixed ports.
 120  *              Alan Cox        :       Cleaned the above up and shrank it to
 121  *                                      a sensible code size.
 122  *              Alan Cox        :       Self connect lockup fix.
 123  *              Alan Cox        :       No connect to multicast.
 124  *              Ross Biro       :       Close unaccepted children on master
 125  *                                      socket close.
 126  *              Alan Cox        :       Reset tracing code.
 127  *              Alan Cox        :       Spurious resets on shutdown.
 128  *              Alan Cox        :       Giant 15 minute/60 second timer error
 129  *              Alan Cox        :       Small whoops in polling before an
 130  *                                      accept.
 131  *              Alan Cox        :       Kept the state trace facility since
 132  *                                      it's handy for debugging.
 133  *              Alan Cox        :       More reset handler fixes.
 134  *              Alan Cox        :       Started rewriting the code based on
 135  *                                      the RFC's for other useful protocol
 136  *                                      references see: Comer, KA9Q NOS, and
 137  *                                      for a reference on the difference
 138  *                                      between specifications and how BSD
 139  *                                      works see the 4.4lite source.
 140  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
 141  *                                      close.
 142  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
 143  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
 144  *              Alan Cox        :       Reimplemented timers as per the RFC
 145  *                                      and using multiple timers for sanity.
 146  *              Alan Cox        :       Small bug fixes, and a lot of new
 147  *                                      comments.
 148  *              Alan Cox        :       Fixed dual reader crash by locking
 149  *                                      the buffers (much like datagram.c)
 150  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
 151  *                                      now gets fed up of retrying without
 152  *                                      (even a no space) answer.
 153  *              Alan Cox        :       Extracted closing code better
 154  *              Alan Cox        :       Fixed the closing state machine to
 155  *                                      resemble the RFC.
 156  *              Alan Cox        :       More 'per spec' fixes.
 157  *              Jorge Cwik      :       Even faster checksumming.
 158  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
 159  *                                      only frames. At least one pc tcp stack
 160  *                                      generates them.
 161  *              Alan Cox        :       Cache last socket.
 162  *              Alan Cox        :       Per route irtt.
 163  *              Matt Day        :       poll()->select() match BSD precisely on error
 164  *              Alan Cox        :       New buffers
 165  *              Marc Tamsky     :       Various sk->prot->retransmits and
 166  *                                      sk->retransmits misupdating fixed.
 167  *                                      Fixed tcp_write_timeout: stuck close,
 168  *                                      and TCP syn retries gets used now.
 169  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
 170  *                                      ack if state is TCP_CLOSED.
 171  *              Alan Cox        :       Look up device on a retransmit - routes may
 172  *                                      change. Doesn't yet cope with MSS shrink right
 173  *                                      but it's a start!
 174  *              Marc Tamsky     :       Closing in closing fixes.
 175  *              Mike Shaver     :       RFC1122 verifications.
 176  *              Alan Cox        :       rcv_saddr errors.
 177  *              Alan Cox        :       Block double connect().
 178  *              Alan Cox        :       Small hooks for enSKIP.
 179  *              Alexey Kuznetsov:       Path MTU discovery.
 180  *              Alan Cox        :       Support soft errors.
 181  *              Alan Cox        :       Fix MTU discovery pathological case
 182  *                                      when the remote claims no mtu!
 183  *              Marc Tamsky     :       TCP_CLOSE fix.
 184  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
 185  *                                      window but wrong (fixes NT lpd problems)
 186  *              Pedro Roque     :       Better TCP window handling, delayed ack.
 187  *              Joerg Reuter    :       No modification of locked buffers in
 188  *                                      tcp_do_retransmit()
 189  *              Eric Schenk     :       Changed receiver side silly window
 190  *                                      avoidance algorithm to BSD style
 191  *                                      algorithm. This doubles throughput
 192  *                                      against machines running Solaris,
 193  *                                      and seems to result in general
 194  *                                      improvement.
 195  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
 196  *      Willy Konynenberg       :       Transparent proxying support.
 197  *      Mike McLagan            :       Routing by source
 198  *              Keith Owens     :       Do proper merging with partial SKB's in
 199  *                                      tcp_do_sendmsg to avoid burstiness.
 200  *              Eric Schenk     :       Fix fast close down bug with
 201  *                                      shutdown() followed by close().
 202  *              Andi Kleen      :       Make poll agree with SIGIO
 203  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
 204  *                                      lingertime == 0 (RFC 793 ABORT Call)
 205  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
 206  *                                      csum_and_copy_from_user() if possible.
 207  *
 208  *              This program is free software; you can redistribute it and/or
 209  *              modify it under the terms of the GNU General Public License
 210  *              as published by the Free Software Foundation; either version
 211  *              2 of the License, or(at your option) any later version.
 212  *
 213  * Description of States:
 214  *
 215  *      TCP_SYN_SENT            sent a connection request, waiting for ack
 216  *
 217  *      TCP_SYN_RECV            received a connection request, sent ack,
 218  *                              waiting for final ack in three-way handshake.
 219  *
 220  *      TCP_ESTABLISHED         connection established
 221  *
 222  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
 223  *                              transmission of remaining buffered data
 224  *
 225  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
 226  *                              to shutdown
 227  *
 228  *      TCP_CLOSING             both sides have shutdown but we still have
 229  *                              data we have to finish sending
 230  *
 231  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
 232  *                              closed, can only be entered from FIN_WAIT2
 233  *                              or CLOSING.  Required because the other end
 234  *                              may not have gotten our last ACK causing it
 235  *                              to retransmit the data packet (which we ignore)
 236  *
 237  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
 238  *                              us to finish writing our data and to shutdown
 239  *                              (we have to close() to move on to LAST_ACK)
 240  *
 241  *      TCP_LAST_ACK            out side has shutdown after remote has
 242  *                              shutdown.  There may still be data in our
 243  *                              buffer that we have to finish sending
 244  *
 245  *      TCP_CLOSE               socket is finished
 246  */
 247
 248 #include <linux/kernel.h>
 249 #include <linux/module.h>
 250 #include <linux/types.h>
 251 #include <linux/fcntl.h>
 252 #include <linux/poll.h>
 253 #include <linux/init.h>
 254 #include <linux/fs.h>
 255 #include <linux/skbuff.h>
 256 #include <linux/scatterlist.h>
 257 #include <linux/splice.h>
 258 #include <linux/net.h>
 259 #include <linux/socket.h>
 260 #include <linux/random.h>
 261 #include <linux/bootmem.h>
 262 #include <linux/highmem.h>
 263 #include <linux/swap.h>
 264 #include <linux/cache.h>
 265 #include <linux/err.h>
 266 #include <linux/crypto.h>
 267 #include <linux/time.h>
 268 #include <linux/slab.h>
 269 #include <linux/uid_stat.h>
 270
 271 #include <net/icmp.h>
 272 #include <net/tcp.h>
 273 #include <net/xfrm.h>
 274 #include <net/ip.h>
 275 #include <net/ip6_route.h>
 276 #include <net/ipv6.h>
 277 #include <net/transp_v6.h>
 278 #include <net/netdma.h>
 279 #include <net/sock.h>
 280
 281 #include <asm/uaccess.h>
 282 #include <asm/ioctls.h>
 283
 284 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
 285
 286 struct percpu_counter tcp_orphan_count;
 287 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 288
 289 long sysctl_tcp_mem[3] __read_mostly;
 290 int sysctl_tcp_wmem[3] __read_mostly;
 291 int sysctl_tcp_rmem[3] __read_mostly;
 292
 293 EXPORT_SYMBOL(sysctl_tcp_mem);
 294 EXPORT_SYMBOL(sysctl_tcp_rmem);
 295 EXPORT_SYMBOL(sysctl_tcp_wmem);
 296
 297 atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
 298 EXPORT_SYMBOL(tcp_memory_allocated);
 299
 300 /*
 301  * Current number of TCP sockets.
 302  */
 303 struct percpu_counter tcp_sockets_allocated;
 304 EXPORT_SYMBOL(tcp_sockets_allocated);
 305
 306 /*
 307  * TCP splice context
 308  */
 309 struct tcp_splice_state {
 310         struct pipe_inode_info *pipe;
 311         size_t len;
 312         unsigned int flags;
 313 };
 314
 315 /*
 316  * Pressure flag: try to collapse.
 317  * Technical note: it is used by multiple contexts non atomically.
 318  * All the __sk_mem_schedule() is of this nature: accounting
 319  * is strict, actions are advisory and have some latency.
 320  */
 321 int tcp_memory_pressure __read_mostly;
 322 EXPORT_SYMBOL(tcp_memory_pressure);
 323
 324 void tcp_enter_memory_pressure(struct sock *sk)
 325 {
 326         if (!tcp_memory_pressure) {
 327                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
 328                 tcp_memory_pressure = 1;
 329         }
 330 }
 331 EXPORT_SYMBOL(tcp_enter_memory_pressure);
 332
 333 /* Convert seconds to retransmits based on initial and max timeout */
 334 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
 335 {
 336         u8 res = 0;
 337
 338         if (seconds > 0) {
 339                 int period = timeout;
 340
 341                 res = 1;
 342                 while (seconds > period && res < 255) {
 343                         res++;
 344                         timeout <<= 1;
 345                         if (timeout > rto_max)
 346                                 timeout = rto_max;
 347                         period += timeout;
 348                 }
 349         }
 350         return res;
 351 }
 352
 353 /* Convert retransmits to seconds based on initial and max timeout */
 354 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
 355 {
 356         int period = 0;
 357
 358         if (retrans > 0) {
 359                 period = timeout;
 360                 while (--retrans) {
 361                         timeout <<= 1;
 362                         if (timeout > rto_max)
 363                                 timeout = rto_max;
 364                         period += timeout;
 365                 }
 366         }
 367         return period;
 368 }
 369
 370 /*
 371  *      Wait for a TCP event.
 372  *
 373  *      Note that we don't need to lock the socket, as the upper poll layers
 374  *      take care of normal races (between the test and the event) and we don't
 375  *      go look at any of the socket buffers directly.
 376  */
 377 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 378 {
 379         unsigned int mask;
 380         struct sock *sk = sock->sk;
 381         struct tcp_sock *tp = tcp_sk(sk);
 382
 383         sock_poll_wait(file, sk_sleep(sk), wait);
 384         if (sk->sk_state == TCP_LISTEN)
 385                 return inet_csk_listen_poll(sk);
 386
 387         /* Socket is not locked. We are protected from async events
 388          * by poll logic and correct handling of state changes
 389          * made by other threads is impossible in any case.
 390          */
 391
 392         mask = 0;
 393
 394         /*
 395          * POLLHUP is certainly not done right. But poll() doesn't
 396          * have a notion of HUP in just one direction, and for a
 397          * socket the read side is more interesting.
 398          *
 399          * Some poll() documentation says that POLLHUP is incompatible
 400          * with the POLLOUT/POLLWR flags, so somebody should check this
 401          * all. But careful, it tends to be safer to return too many
 402          * bits than too few, and you can easily break real applications
 403          * if you don't tell them that something has hung up!
 404          *
 405          * Check-me.
 406          *
 407          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
 408          * our fs/select.c). It means that after we received EOF,
 409          * poll always returns immediately, making impossible poll() on write()
 410          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
 411          * if and only if shutdown has been made in both directions.
 412          * Actually, it is interesting to look how Solaris and DUX
 413          * solve this dilemma. I would prefer, if POLLHUP were maskable,
 414          * then we could set it on SND_SHUTDOWN. BTW examples given
 415          * in Stevens' books assume exactly this behaviour, it explains
 416          * why POLLHUP is incompatible with POLLOUT.    --ANK
 417          *
 418          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 419          * blocking on fresh not-connected or disconnected socket. --ANK
 420          */
 421         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
 422                 mask |= POLLHUP;
 423         if (sk->sk_shutdown & RCV_SHUTDOWN)
 424                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
 425
 426         /* Connected? */
 427         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
 428                 int target = sock_rcvlowat(sk, 0, INT_MAX);
 429
 430                 if (tp->urg_seq == tp->copied_seq &&
 431                     !sock_flag(sk, SOCK_URGINLINE) &&
 432                     tp->urg_data)
 433                         target++;
 434
 435                 /* Potential race condition. If read of tp below will
 436                  * escape above sk->sk_state, we can be illegally awaken
 437                  * in SYN_* states. */
 438                 if (tp->rcv_nxt - tp->copied_seq >= target)
 439                         mask |= POLLIN | POLLRDNORM;
 440
 441                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 442                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
 443                                 mask |= POLLOUT | POLLWRNORM;
 444                         } else {  /* send SIGIO later */
 445                                 set_bit(SOCK_ASYNC_NOSPACE,
 446                                         &sk->sk_socket->flags);
 447                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 448
 449                                 /* Race breaker. If space is freed after
 450                                  * wspace test but before the flags are set,
 451                                  * IO signal will be lost.
 452                                  */
 453                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
 454                                         mask |= POLLOUT | POLLWRNORM;
 455                         }
 456                 } else
 457                         mask |= POLLOUT | POLLWRNORM;
 458
 459                 if (tp->urg_data & TCP_URG_VALID)
 460                         mask |= POLLPRI;
 461         }
 462         /* This barrier is coupled with smp_wmb() in tcp_reset() */
 463         smp_rmb();
 464         if (sk->sk_err)
 465                 mask |= POLLERR;
 466
 467         return mask;
 468 }
 469 EXPORT_SYMBOL(tcp_poll);
 470
 471 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 472 {
 473         struct tcp_sock *tp = tcp_sk(sk);
 474         int answ;
 475
 476         switch (cmd) {
 477         case SIOCINQ:
 478                 if (sk->sk_state == TCP_LISTEN)
 479                         return -EINVAL;
 480
 481                 lock_sock(sk);
 482                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 483                         answ = 0;
 484                 else if (sock_flag(sk, SOCK_URGINLINE) ||
 485                          !tp->urg_data ||
 486                          before(tp->urg_seq, tp->copied_seq) ||
 487                          !before(tp->urg_seq, tp->rcv_nxt)) {
 488                         struct sk_buff *skb;
 489
 490                         answ = tp->rcv_nxt - tp->copied_seq;
 491
 492                         /* Subtract 1, if FIN is in queue. */
 493                         skb = skb_peek_tail(&sk->sk_receive_queue);
 494                         if (answ && skb)
 495                                 answ -= tcp_hdr(skb)->fin;
 496                 } else
 497                         answ = tp->urg_seq - tp->copied_seq;
 498                 release_sock(sk);
 499                 break;
 500         case SIOCATMARK:
 501                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 502                 break;
 503         case SIOCOUTQ:
 504                 if (sk->sk_state == TCP_LISTEN)
 505                         return -EINVAL;
 506
 507                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 508                         answ = 0;
 509                 else
 510                         answ = tp->write_seq - tp->snd_una;
 511                 break;
 512         case SIOCOUTQNSD:
 513                 if (sk->sk_state == TCP_LISTEN)
 514                         return -EINVAL;
 515
 516                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 517                         answ = 0;
 518                 else
 519                         answ = tp->write_seq - tp->snd_nxt;
 520                 break;
 521         default:
 522                 return -ENOIOCTLCMD;
 523         }
 524
 525         return put_user(answ, (int __user *)arg);
 526 }
 527 EXPORT_SYMBOL(tcp_ioctl);
 528
 529 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 530 {
 531         TCP_SKB_CB(skb)->flags |= TCPHDR_PSH;
 532         tp->pushed_seq = tp->write_seq;
 533 }
 534
 535 static inline int forced_push(struct tcp_sock *tp)
 536 {
 537         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 538 }
 539
 540 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
 541 {
 542         struct tcp_sock *tp = tcp_sk(sk);
 543         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 544
 545         skb->csum    = 0;
 546         tcb->seq     = tcb->end_seq = tp->write_seq;
 547         tcb->flags   = TCPHDR_ACK;
 548         tcb->sacked  = 0;
 549         skb_header_release(skb);
 550         tcp_add_write_queue_tail(sk, skb);
 551         sk->sk_wmem_queued += skb->truesize;
 552         sk_mem_charge(sk, skb->truesize);
 553         if (tp->nonagle & TCP_NAGLE_PUSH)
 554                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 555 }
 556
 557 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
 558 {
 559         if (flags & MSG_OOB)
 560                 tp->snd_up = tp->write_seq;
 561 }
 562
 563 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
 564                             int nonagle)
 565 {
 566         if (tcp_send_head(sk)) {
 567                 struct tcp_sock *tp = tcp_sk(sk);
 568
 569                 if (!(flags & MSG_MORE) || forced_push(tp))
 570                         tcp_mark_push(tp, tcp_write_queue_tail(sk));
 571
 572                 tcp_mark_urg(tp, flags);
 573                 __tcp_push_pending_frames(sk, mss_now,
 574                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
 575         }
 576 }
 577
 578 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 579                                 unsigned int offset, size_t len)
 580 {
 581         struct tcp_splice_state *tss = rd_desc->arg.data;
 582         int ret;
 583
 584         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
 585                               tss->flags);
 586         if (ret > 0)
 587                 rd_desc->count -= ret;
 588         return ret;
 589 }
 590
 591 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 592 {
 593         /* Store TCP splice context information in read_descriptor_t. */
 594         read_descriptor_t rd_desc = {
 595                 .arg.data = tss,
 596                 .count    = tss->len,
 597         };
 598
 599         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 600 }
 601
 602 /**
 603  *  tcp_splice_read - splice data from TCP socket to a pipe
 604  * @sock:       socket to splice from
 605  * @ppos:       position (not valid)
 606  * @pipe:       pipe to splice to
 607  * @len:        number of bytes to splice
 608  * @flags:      splice modifier flags
 609  *
 610  * Description:
 611  *    Will read pages from given socket and fill them into a pipe.
 612  *
 613  **/
 614 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 615                         struct pipe_inode_info *pipe, size_t len,
 616                         unsigned int flags)
 617 {
 618         struct sock *sk = sock->sk;
 619         struct tcp_splice_state tss = {
 620                 .pipe = pipe,
 621                 .len = len,
 622                 .flags = flags,
 623         };
 624         long timeo;
 625         ssize_t spliced;
 626         int ret;
 627
 628         sock_rps_record_flow(sk);
 629         /*
 630          * We can't seek on a socket input
 631          */
 632         if (unlikely(*ppos))
 633                 return -ESPIPE;
 634
 635         ret = spliced = 0;
 636
 637         lock_sock(sk);
 638
 639         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 640         while (tss.len) {
 641                 ret = __tcp_splice_read(sk, &tss);
 642                 if (ret < 0)
 643                         break;
 644                 else if (!ret) {
 645                         if (spliced)
 646                                 break;
 647                         if (sock_flag(sk, SOCK_DONE))
 648                                 break;
 649                         if (sk->sk_err) {
 650                                 ret = sock_error(sk);
 651                                 break;
 652                         }
 653                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 654                                 break;
 655                         if (sk->sk_state == TCP_CLOSE) {
 656                                 /*
 657                                  * This occurs when user tries to read
 658                                  * from never connected socket.
 659                                  */
 660                                 if (!sock_flag(sk, SOCK_DONE))
 661                                         ret = -ENOTCONN;
 662                                 break;
 663                         }
 664                         if (!timeo) {
 665                                 ret = -EAGAIN;
 666                                 break;
 667                         }
 668                         sk_wait_data(sk, &timeo);
 669                         if (signal_pending(current)) {
 670                                 ret = sock_intr_errno(timeo);
 671                                 break;
 672                         }
 673                         continue;
 674                 }
 675                 tss.len -= ret;
 676                 spliced += ret;
 677
 678                 if (!timeo)
 679                         break;
 680                 release_sock(sk);
 681                 lock_sock(sk);
 682
 683                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 684                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 685                     signal_pending(current))
 686                         break;
 687         }
 688
 689         release_sock(sk);
 690
 691         if (spliced)
 692                 return spliced;
 693
 694         return ret;
 695 }
 696 EXPORT_SYMBOL(tcp_splice_read);
 697
 698 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
 699 {
 700         struct sk_buff *skb;
 701
 702         /* The TCP header must be at least 32-bit aligned.  */
 703         size = ALIGN(size, 4);
 704
 705         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 706         if (skb) {
 707                 if (sk_wmem_schedule(sk, skb->truesize)) {
 708                         /*
 709                          * Make sure that we have exactly size bytes
 710                          * available to the caller, no more, no less.
 711                          */
 712                         skb_reserve(skb, skb_tailroom(skb) - size);
 713                         return skb;
 714                 }
 715                 __kfree_skb(skb);
 716         } else {
 717                 sk->sk_prot->enter_memory_pressure(sk);
 718                 sk_stream_moderate_sndbuf(sk);
 719         }
 720         return NULL;
 721 }
 722
 723 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 724                                        int large_allowed)
 725 {
 726         struct tcp_sock *tp = tcp_sk(sk);
 727         u32 xmit_size_goal, old_size_goal;
 728
 729         xmit_size_goal = mss_now;
 730
 731         if (large_allowed && sk_can_gso(sk)) {
 732                 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
 733                                   inet_csk(sk)->icsk_af_ops->net_header_len -
 734                                   inet_csk(sk)->icsk_ext_hdr_len -
 735                                   tp->tcp_header_len);
 736
 737                 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
 738
 739                 /* We try hard to avoid divides here */
 740                 old_size_goal = tp->xmit_size_goal_segs * mss_now;
 741
 742                 if (likely(old_size_goal <= xmit_size_goal &&
 743                            old_size_goal + mss_now > xmit_size_goal)) {
 744                         xmit_size_goal = old_size_goal;
 745                 } else {
 746                         tp->xmit_size_goal_segs =
 747                                 min_t(u16, xmit_size_goal / mss_now,
 748                                       sk->sk_gso_max_segs);
 749                         xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
 750                 }
 751         }
 752
 753         return max(xmit_size_goal, mss_now);
 754 }
 755
 756 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 757 {
 758         int mss_now;
 759
 760         mss_now = tcp_current_mss(sk);
 761         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 762
 763         return mss_now;
 764 }
 765
 766 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
 767                          size_t psize, int flags)
 768 {
 769         struct tcp_sock *tp = tcp_sk(sk);
 770         int mss_now, size_goal;
 771         int err;
 772         ssize_t copied;
 773         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 774
 775         /* Wait for a connection to finish. */
 776         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 777                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 778                         goto out_err;
 779
 780         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 781
 782         mss_now = tcp_send_mss(sk, &size_goal, flags);
 783         copied = 0;
 784
 785         err = -EPIPE;
 786         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 787                 goto out_err;
 788
 789         while (psize > 0) {
 790                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 791                 struct page *page = pages[poffset / PAGE_SIZE];
 792                 int copy, i, can_coalesce;
 793                 int offset = poffset % PAGE_SIZE;
 794                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
 795
 796                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
 797 new_segment:
 798                         if (!sk_stream_memory_free(sk))
 799                                 goto wait_for_sndbuf;
 800
 801                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
 802                         if (!skb)
 803                                 goto wait_for_memory;
 804
 805                         skb_entail(sk, skb);
 806                         copy = size_goal;
 807                 }
 808
 809                 if (copy > size)
 810                         copy = size;
 811
 812                 i = skb_shinfo(skb)->nr_frags;
 813                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 814                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
 815                         tcp_mark_push(tp, skb);
 816                         goto new_segment;
 817                 }
 818                 if (!sk_wmem_schedule(sk, copy))
 819                         goto wait_for_memory;
 820
 821                 if (can_coalesce) {
 822                         skb_shinfo(skb)->frags[i - 1].size += copy;
 823                 } else {
 824                         get_page(page);
 825                         skb_fill_page_desc(skb, i, page, offset, copy);
 826                 }
 827
 828                 skb->len += copy;
 829                 skb->data_len += copy;
 830                 skb->truesize += copy;
 831                 sk->sk_wmem_queued += copy;
 832                 sk_mem_charge(sk, copy);
 833                 skb->ip_summed = CHECKSUM_PARTIAL;
 834                 tp->write_seq += copy;
 835                 TCP_SKB_CB(skb)->end_seq += copy;
 836                 skb_shinfo(skb)->gso_segs = 0;
 837
 838                 if (!copied)
 839                         TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH;
 840
 841                 copied += copy;
 842                 poffset += copy;
 843                 if (!(psize -= copy))
 844                         goto out;
 845
 846                 if (skb->len < size_goal || (flags & MSG_OOB))
 847                         continue;
 848
 849                 if (forced_push(tp)) {
 850                         tcp_mark_push(tp, skb);
 851                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
 852                 } else if (skb == tcp_send_head(sk))
 853                         tcp_push_one(sk, mss_now);
 854                 continue;
 855
 856 wait_for_sndbuf:
 857                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 858 wait_for_memory:
 859                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
 860
 861                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 862                         goto do_error;
 863
 864                 mss_now = tcp_send_mss(sk, &size_goal, flags);
 865         }
 866
 867 out:
 868         if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
 869                 tcp_push(sk, flags, mss_now, tp->nonagle);
 870         return copied;
 871
 872 do_error:
 873         if (copied)
 874                 goto out;
 875 out_err:
 876         return sk_stream_error(sk, flags, err);
 877 }
 878
 879 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
 880                  size_t size, int flags)
 881 {
 882         ssize_t res;
 883
 884         if (!(sk->sk_route_caps & NETIF_F_SG) ||
 885             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
 886                 return sock_no_sendpage(sk->sk_socket, page, offset, size,
 887                                         flags);
 888
 889         lock_sock(sk);
 890         res = do_tcp_sendpages(sk, &page, offset, size, flags);
 891         release_sock(sk);
 892         return res;
 893 }
 894 EXPORT_SYMBOL(tcp_sendpage);
 895
 896 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
 897 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
 898
 899 static inline int select_size(struct sock *sk, int sg)
 900 {
 901         struct tcp_sock *tp = tcp_sk(sk);
 902         int tmp = tp->mss_cache;
 903
 904         if (sg) {
 905                 if (sk_can_gso(sk))
 906                         tmp = 0;
 907                 else {
 908                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
 909
 910                         if (tmp >= pgbreak &&
 911                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
 912                                 tmp = pgbreak;
 913                 }
 914         }
 915
 916         return tmp;
 917 }
 918
 919 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
 920                 size_t size)
 921 {
 922         struct iovec *iov;
 923         struct tcp_sock *tp = tcp_sk(sk);
 924         struct sk_buff *skb;
 925         int iovlen, flags;
 926         int mss_now, size_goal;
 927         int sg, err, copied;
 928         long timeo;
 929
 930         lock_sock(sk);
 931
 932         flags = msg->msg_flags;
 933         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 934
 935         /* Wait for a connection to finish. */
 936         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 937                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 938                         goto out_err;
 939
 940         /* This should be in poll */
 941         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 942
 943         mss_now = tcp_send_mss(sk, &size_goal, flags);
 944
 945         /* Ok commence sending. */
 946         iovlen = msg->msg_iovlen;
 947         iov = msg->msg_iov;
 948         copied = 0;
 949
 950         err = -EPIPE;
 951         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 952                 goto out_err;
 953
 954         sg = sk->sk_route_caps & NETIF_F_SG;
 955
 956         while (--iovlen >= 0) {
 957                 size_t seglen = iov->iov_len;
 958                 unsigned char __user *from = iov->iov_base;
 959
 960                 iov++;
 961
 962                 while (seglen > 0) {
 963                         int copy = 0;
 964                         int max = size_goal;
 965
 966                         skb = tcp_write_queue_tail(sk);
 967                         if (tcp_send_head(sk)) {
 968                                 if (skb->ip_summed == CHECKSUM_NONE)
 969                                         max = mss_now;
 970                                 copy = max - skb->len;
 971                         }
 972
 973                         if (copy <= 0) {
 974 new_segment:
 975                                 /* Allocate new segment. If the interface is SG,
 976                                  * allocate skb fitting to single page.
 977                                  */
 978                                 if (!sk_stream_memory_free(sk))
 979                                         goto wait_for_sndbuf;
 980
 981                                 skb = sk_stream_alloc_skb(sk,
 982                                                           select_size(sk, sg),
 983                                                           sk->sk_allocation);
 984                                 if (!skb)
 985                                         goto wait_for_memory;
 986
 987                                 /*
 988                                  * Check whether we can use HW checksum.
 989                                  */
 990                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
 991                                         skb->ip_summed = CHECKSUM_PARTIAL;
 992
 993                                 skb_entail(sk, skb);
 994                                 copy = size_goal;
 995                                 max = size_goal;
 996                         }
 997
 998                         /* Try to append data to the end of skb. */
 999                         if (copy > seglen)
1000                                 copy = seglen;
1001
1002                         /* Where to copy to? */
1003                         if (skb_tailroom(skb) > 0) {
1004                                 /* We have some space in skb head. Superb! */
1005                                 if (copy > skb_tailroom(skb))
1006                                         copy = skb_tailroom(skb);
1007                                 err = skb_add_data_nocache(sk, skb, from, copy);
1008                                 if (err)
1009                                         goto do_fault;
1010                         } else {
1011                                 int merge = 0;
1012                                 int i = skb_shinfo(skb)->nr_frags;
1013                                 struct page *page = TCP_PAGE(sk);
1014                                 int off = TCP_OFF(sk);
1015
1016                                 if (skb_can_coalesce(skb, i, page, off) &&
1017                                     off != PAGE_SIZE) {
1018                                         /* We can extend the last page
1019                                          * fragment. */
1020                                         merge = 1;
1021                                 } else if (i == MAX_SKB_FRAGS || !sg) {
1022                                         /* Need to add new fragment and cannot
1023                                          * do this because interface is non-SG,
1024                                          * or because all the page slots are
1025                                          * busy. */
1026                                         tcp_mark_push(tp, skb);
1027                                         goto new_segment;
1028                                 } else if (page) {
1029                                         if (off == PAGE_SIZE) {
1030                                                 put_page(page);
1031                                                 TCP_PAGE(sk) = page = NULL;
1032                                                 off = 0;
1033                                         }
1034                                 } else
1035                                         off = 0;
1036
1037                                 if (copy > PAGE_SIZE - off)
1038                                         copy = PAGE_SIZE - off;
1039
1040                                 if (!sk_wmem_schedule(sk, copy))
1041                                         goto wait_for_memory;
1042
1043                                 if (!page) {
1044                                         /* Allocate new cache page. */
1045                                         if (!(page = sk_stream_alloc_page(sk)))
1046                                                 goto wait_for_memory;
1047                                 }
1048
1049                                 /* Time to copy data. We are close to
1050                                  * the end! */
1051                                 err = skb_copy_to_page_nocache(sk, from, skb,
1052                                                                page, off, copy);
1053                                 if (err) {
1054                                         /* If this page was new, give it to the
1055                                          * socket so it does not get leaked.
1056                                          */
1057                                         if (!TCP_PAGE(sk)) {
1058                                                 TCP_PAGE(sk) = page;
1059                                                 TCP_OFF(sk) = 0;
1060                                         }
1061                                         goto do_error;
1062                                 }
1063
1064                                 /* Update the skb. */
1065                                 if (merge) {
1066                                         skb_shinfo(skb)->frags[i - 1].size +=
1067                                                                         copy;
1068                                 } else {
1069                                         skb_fill_page_desc(skb, i, page, off, copy);
1070                                         if (TCP_PAGE(sk)) {
1071                                                 get_page(page);
1072                                         } else if (off + copy < PAGE_SIZE) {
1073                                                 get_page(page);
1074                                                 TCP_PAGE(sk) = page;
1075                                         }
1076                                 }
1077
1078                                 TCP_OFF(sk) = off + copy;
1079                         }
1080
1081                         if (!copied)
1082                                 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH;
1083
1084                         tp->write_seq += copy;
1085                         TCP_SKB_CB(skb)->end_seq += copy;
1086                         skb_shinfo(skb)->gso_segs = 0;
1087
1088                         from += copy;
1089                         copied += copy;
1090                         if ((seglen -= copy) == 0 && iovlen == 0)
1091                                 goto out;
1092
1093                         if (skb->len < max || (flags & MSG_OOB))
1094                                 continue;
1095
1096                         if (forced_push(tp)) {
1097                                 tcp_mark_push(tp, skb);
1098                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1099                         } else if (skb == tcp_send_head(sk))
1100                                 tcp_push_one(sk, mss_now);
1101                         continue;
1102
1103 wait_for_sndbuf:
1104                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1105 wait_for_memory:
1106                         if (copied)
1107                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1108
1109                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1110                                 goto do_error;
1111
1112                         mss_now = tcp_send_mss(sk, &size_goal, flags);
1113                 }
1114         }
1115
1116 out:
1117         if (copied)
1118                 tcp_push(sk, flags, mss_now, tp->nonagle);
1119         release_sock(sk);
1120
1121         if (copied > 0)
1122                 uid_stat_tcp_snd(current_uid(), copied);
1123         return copied;
1124
1125 do_fault:
1126         if (!skb->len) {
1127                 tcp_unlink_write_queue(skb, sk);
1128                 /* It is the one place in all of TCP, except connection
1129                  * reset, where we can be unlinking the send_head.
1130                  */
1131                 tcp_check_send_head(sk, skb);
1132                 sk_wmem_free_skb(sk, skb);
1133         }
1134
1135 do_error:
1136         if (copied)
1137                 goto out;
1138 out_err:
1139         err = sk_stream_error(sk, flags, err);
1140         release_sock(sk);
1141         return err;
1142 }
1143 EXPORT_SYMBOL(tcp_sendmsg);
1144
1145 /*
1146  *      Handle reading urgent data. BSD has very simple semantics for
1147  *      this, no blocking and very strange errors 8)
1148  */
1149
1150 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1151 {
1152         struct tcp_sock *tp = tcp_sk(sk);
1153
1154         /* No URG data to read. */
1155         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1156             tp->urg_data == TCP_URG_READ)
1157                 return -EINVAL; /* Yes this is right ! */
1158
1159         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1160                 return -ENOTCONN;
1161
1162         if (tp->urg_data & TCP_URG_VALID) {
1163                 int err = 0;
1164                 char c = tp->urg_data;
1165
1166                 if (!(flags & MSG_PEEK))
1167                         tp->urg_data = TCP_URG_READ;
1168
1169                 /* Read urgent data. */
1170                 msg->msg_flags |= MSG_OOB;
1171
1172                 if (len > 0) {
1173                         if (!(flags & MSG_TRUNC))
1174                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1175                         len = 1;
1176                 } else
1177                         msg->msg_flags |= MSG_TRUNC;
1178
1179                 return err ? -EFAULT : len;
1180         }
1181
1182         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1183                 return 0;
1184
1185         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1186          * the available implementations agree in this case:
1187          * this call should never block, independent of the
1188          * blocking state of the socket.
1189          * Mike <pall@rz.uni-karlsruhe.de>
1190          */
1191         return -EAGAIN;
1192 }
1193
1194 /* Clean up the receive buffer for full frames taken by the user,
1195  * then send an ACK if necessary.  COPIED is the number of bytes
1196  * tcp_recvmsg has given to the user so far, it speeds up the
1197  * calculation of whether or not we must ACK for the sake of
1198  * a window update.
1199  */
1200 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1201 {
1202         struct tcp_sock *tp = tcp_sk(sk);
1203         int time_to_ack = 0;
1204
1205 #if TCP_DEBUG
1206         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1207
1208         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1209              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1210              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1211 #endif
1212
1213         if (inet_csk_ack_scheduled(sk)) {
1214                 const struct inet_connection_sock *icsk = inet_csk(sk);
1215                    /* Delayed ACKs frequently hit locked sockets during bulk
1216                     * receive. */
1217                 if (icsk->icsk_ack.blocked ||
1218                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1219                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1220                     /*
1221                      * If this read emptied read buffer, we send ACK, if
1222                      * connection is not bidirectional, user drained
1223                      * receive buffer and there was a small segment
1224                      * in queue.
1225                      */
1226                     (copied > 0 &&
1227                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1228                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1229                        !icsk->icsk_ack.pingpong)) &&
1230                       !atomic_read(&sk->sk_rmem_alloc)))
1231                         time_to_ack = 1;
1232         }
1233
1234         /* We send an ACK if we can now advertise a non-zero window
1235          * which has been raised "significantly".
1236          *
1237          * Even if window raised up to infinity, do not send window open ACK
1238          * in states, where we will not receive more. It is useless.
1239          */
1240         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1241                 __u32 rcv_window_now = tcp_receive_window(tp);
1242
1243                 /* Optimize, __tcp_select_window() is not cheap. */
1244                 if (2*rcv_window_now <= tp->window_clamp) {
1245                         __u32 new_window = __tcp_select_window(sk);
1246
1247                         /* Send ACK now, if this read freed lots of space
1248                          * in our buffer. Certainly, new_window is new window.
1249                          * We can advertise it now, if it is not less than current one.
1250                          * "Lots" means "at least twice" here.
1251                          */
1252                         if (new_window && new_window >= 2 * rcv_window_now)
1253                                 time_to_ack = 1;
1254                 }
1255         }
1256         if (time_to_ack)
1257                 tcp_send_ack(sk);
1258 }
1259
1260 static void tcp_prequeue_process(struct sock *sk)
1261 {
1262         struct sk_buff *skb;
1263         struct tcp_sock *tp = tcp_sk(sk);
1264
1265         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1266
1267         /* RX process wants to run with disabled BHs, though it is not
1268          * necessary */
1269         local_bh_disable();
1270         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1271                 sk_backlog_rcv(sk, skb);
1272         local_bh_enable();
1273
1274         /* Clear memory counter. */
1275         tp->ucopy.memory = 0;
1276 }
1277
1278 #ifdef CONFIG_NET_DMA
1279 static void tcp_service_net_dma(struct sock *sk, bool wait)
1280 {
1281         dma_cookie_t done, used;
1282         dma_cookie_t last_issued;
1283         struct tcp_sock *tp = tcp_sk(sk);
1284
1285         if (!tp->ucopy.dma_chan)
1286                 return;
1287
1288         last_issued = tp->ucopy.dma_cookie;
1289         dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1290
1291         do {
1292                 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1293                                               last_issued, &done,
1294                                               &used) == DMA_SUCCESS) {
1295                         /* Safe to free early-copied skbs now */
1296                         __skb_queue_purge(&sk->sk_async_wait_queue);
1297                         break;
1298                 } else {
1299                         struct sk_buff *skb;
1300                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1301                                (dma_async_is_complete(skb->dma_cookie, done,
1302                                                       used) == DMA_SUCCESS)) {
1303                                 __skb_dequeue(&sk->sk_async_wait_queue);
1304                                 kfree_skb(skb);
1305                         }
1306                 }
1307         } while (wait);
1308 }
1309 #endif
1310
1311 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1312 {
1313         struct sk_buff *skb;
1314         u32 offset;
1315
1316         skb_queue_walk(&sk->sk_receive_queue, skb) {
1317                 offset = seq - TCP_SKB_CB(skb)->seq;
1318                 if (tcp_hdr(skb)->syn)
1319                         offset--;
1320                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1321                         *off = offset;
1322                         return skb;
1323                 }
1324         }
1325         return NULL;
1326 }
1327
1328 /*
1329  * This routine provides an alternative to tcp_recvmsg() for routines
1330  * that would like to handle copying from skbuffs directly in 'sendfile'
1331  * fashion.
1332  * Note:
1333  *      - It is assumed that the socket was locked by the caller.
1334  *      - The routine does not block.
1335  *      - At present, there is no support for reading OOB data
1336  *        or for 'peeking' the socket using this routine
1337  *        (although both would be easy to implement).
1338  */
1339 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1340                   sk_read_actor_t recv_actor)
1341 {
1342         struct sk_buff *skb;
1343         struct tcp_sock *tp = tcp_sk(sk);
1344         u32 seq = tp->copied_seq;
1345         u32 offset;
1346         int copied = 0;
1347
1348         if (sk->sk_state == TCP_LISTEN)
1349                 return -ENOTCONN;
1350         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1351                 if (offset < skb->len) {
1352                         int used;
1353                         size_t len;
1354
1355                         len = skb->len - offset;
1356                         /* Stop reading if we hit a patch of urgent data */
1357                         if (tp->urg_data) {
1358                                 u32 urg_offset = tp->urg_seq - seq;
1359                                 if (urg_offset < len)
1360                                         len = urg_offset;
1361                                 if (!len)
1362                                         break;
1363                         }
1364                         used = recv_actor(desc, skb, offset, len);
1365                         if (used < 0) {
1366                                 if (!copied)
1367                                         copied = used;
1368                                 break;
1369                         } else if (used <= len) {
1370                                 seq += used;
1371                                 copied += used;
1372                                 offset += used;
1373                         }
1374                         /*
1375                          * If recv_actor drops the lock (e.g. TCP splice
1376                          * receive) the skb pointer might be invalid when
1377                          * getting here: tcp_collapse might have deleted it
1378                          * while aggregating skbs from the socket queue.
1379                          */
1380                         skb = tcp_recv_skb(sk, seq-1, &offset);
1381                         if (!skb || (offset+1 != skb->len))
1382                                 break;
1383                 }
1384                 if (tcp_hdr(skb)->fin) {
1385                         sk_eat_skb(sk, skb, 0);
1386                         ++seq;
1387                         break;
1388                 }
1389                 sk_eat_skb(sk, skb, 0);
1390                 if (!desc->count)
1391                         break;
1392                 tp->copied_seq = seq;
1393         }
1394         tp->copied_seq = seq;
1395
1396         tcp_rcv_space_adjust(sk);
1397
1398         /* Clean up data we have read: This will do ACK frames. */
1399         if (copied > 0) {
1400                 tcp_cleanup_rbuf(sk, copied);
1401                 uid_stat_tcp_rcv(current_uid(), copied);
1402         }
1403
1404         return copied;
1405 }
1406 EXPORT_SYMBOL(tcp_read_sock);
1407
1408 /*
1409  *      This routine copies from a sock struct into the user buffer.
1410  *
1411  *      Technical note: in 2.3 we work on _locked_ socket, so that
1412  *      tricks with *seq access order and skb->users are not required.
1413  *      Probably, code can be easily improved even more.
1414  */
1415
1416 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1417                 size_t len, int nonblock, int flags, int *addr_len)
1418 {
1419         struct tcp_sock *tp = tcp_sk(sk);
1420         int copied = 0;
1421         u32 peek_seq;
1422         u32 *seq;
1423         unsigned long used;
1424         int err;
1425         int target;             /* Read at least this many bytes */
1426         long timeo;
1427         struct task_struct *user_recv = NULL;
1428         int copied_early = 0;
1429         struct sk_buff *skb;
1430         u32 urg_hole = 0;
1431
1432         lock_sock(sk);
1433
1434         err = -ENOTCONN;
1435         if (sk->sk_state == TCP_LISTEN)
1436                 goto out;
1437
1438         timeo = sock_rcvtimeo(sk, nonblock);
1439
1440         /* Urgent data needs to be handled specially. */
1441         if (flags & MSG_OOB)
1442                 goto recv_urg;
1443
1444         seq = &tp->copied_seq;
1445         if (flags & MSG_PEEK) {
1446                 peek_seq = tp->copied_seq;
1447                 seq = &peek_seq;
1448         }
1449
1450         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1451
1452 #ifdef CONFIG_NET_DMA
1453         tp->ucopy.dma_chan = NULL;
1454         preempt_disable();
1455         skb = skb_peek_tail(&sk->sk_receive_queue);
1456         {
1457                 int available = 0;
1458
1459                 if (skb)
1460                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1461                 if ((available < target) &&
1462                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1463                     !sysctl_tcp_low_latency &&
1464                     dma_find_channel(DMA_MEMCPY)) {
1465                         preempt_enable_no_resched();
1466                         tp->ucopy.pinned_list =
1467                                         dma_pin_iovec_pages(msg->msg_iov, len);
1468                 } else {
1469                         preempt_enable_no_resched();
1470                 }
1471         }
1472 #endif
1473
1474         do {
1475                 u32 offset;
1476
1477                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1478                 if (tp->urg_data && tp->urg_seq == *seq) {
1479                         if (copied)
1480                                 break;
1481                         if (signal_pending(current)) {
1482                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1483                                 break;
1484                         }
1485                 }
1486
1487                 /* Next get a buffer. */
1488
1489                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1490                         /* Now that we have two receive queues this
1491                          * shouldn't happen.
1492                          */
1493                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1494                                  "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1495                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1496                                  flags))
1497                                 break;
1498
1499                         offset = *seq - TCP_SKB_CB(skb)->seq;
1500                         if (tcp_hdr(skb)->syn)
1501                                 offset--;
1502                         if (offset < skb->len)
1503                                 goto found_ok_skb;
1504                         if (tcp_hdr(skb)->fin)
1505                                 goto found_fin_ok;
1506                         WARN(!(flags & MSG_PEEK),
1507                              "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1508                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1509                 }
1510
1511                 /* Well, if we have backlog, try to process it now yet. */
1512
1513                 if (copied >= target && !sk->sk_backlog.tail)
1514                         break;
1515
1516                 if (copied) {
1517                         if (sk->sk_err ||
1518                             sk->sk_state == TCP_CLOSE ||
1519                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1520                             !timeo ||
1521                             signal_pending(current))
1522                                 break;
1523                 } else {
1524                         if (sock_flag(sk, SOCK_DONE))
1525                                 break;
1526
1527                         if (sk->sk_err) {
1528                                 copied = sock_error(sk);
1529                                 break;
1530                         }
1531
1532                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1533                                 break;
1534
1535                         if (sk->sk_state == TCP_CLOSE) {
1536                                 if (!sock_flag(sk, SOCK_DONE)) {
1537                                         /* This occurs when user tries to read
1538                                          * from never connected socket.
1539                                          */
1540                                         copied = -ENOTCONN;
1541                                         break;
1542                                 }
1543                                 break;
1544                         }
1545
1546                         if (!timeo) {
1547                                 copied = -EAGAIN;
1548                                 break;
1549                         }
1550
1551                         if (signal_pending(current)) {
1552                                 copied = sock_intr_errno(timeo);
1553                                 break;
1554                         }
1555                 }
1556
1557                 tcp_cleanup_rbuf(sk, copied);
1558
1559                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1560                         /* Install new reader */
1561                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1562                                 user_recv = current;
1563                                 tp->ucopy.task = user_recv;
1564                                 tp->ucopy.iov = msg->msg_iov;
1565                         }
1566
1567                         tp->ucopy.len = len;
1568
1569                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1570                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1571
1572                         /* Ugly... If prequeue is not empty, we have to
1573                          * process it before releasing socket, otherwise
1574                          * order will be broken at second iteration.
1575                          * More elegant solution is required!!!
1576                          *
1577                          * Look: we have the following (pseudo)queues:
1578                          *
1579                          * 1. packets in flight
1580                          * 2. backlog
1581                          * 3. prequeue
1582                          * 4. receive_queue
1583                          *
1584                          * Each queue can be processed only if the next ones
1585                          * are empty. At this point we have empty receive_queue.
1586                          * But prequeue _can_ be not empty after 2nd iteration,
1587                          * when we jumped to start of loop because backlog
1588                          * processing added something to receive_queue.
1589                          * We cannot release_sock(), because backlog contains
1590                          * packets arrived _after_ prequeued ones.
1591                          *
1592                          * Shortly, algorithm is clear --- to process all
1593                          * the queues in order. We could make it more directly,
1594                          * requeueing packets from backlog to prequeue, if
1595                          * is not empty. It is more elegant, but eats cycles,
1596                          * unfortunately.
1597                          */
1598                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1599                                 goto do_prequeue;
1600
1601                         /* __ Set realtime policy in scheduler __ */
1602                 }
1603
1604 #ifdef CONFIG_NET_DMA
1605                 if (tp->ucopy.dma_chan) {
1606                         if (tp->rcv_wnd == 0 &&
1607                             !skb_queue_empty(&sk->sk_async_wait_queue)) {
1608                                 tcp_service_net_dma(sk, true);
1609                                 tcp_cleanup_rbuf(sk, copied);
1610                         } else
1611                                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1612                 }
1613 #endif
1614                 if (copied >= target) {
1615                         /* Do not sleep, just process backlog. */
1616                         release_sock(sk);
1617                         lock_sock(sk);
1618                 } else
1619                         sk_wait_data(sk, &timeo);
1620
1621 #ifdef CONFIG_NET_DMA
1622                 tcp_service_net_dma(sk, false);  /* Don't block */
1623                 tp->ucopy.wakeup = 0;
1624 #endif
1625
1626                 if (user_recv) {
1627                         int chunk;
1628
1629                         /* __ Restore normal policy in scheduler __ */
1630
1631                         if ((chunk = len - tp->ucopy.len) != 0) {
1632                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1633                                 len -= chunk;
1634                                 copied += chunk;
1635                         }
1636
1637                         if (tp->rcv_nxt == tp->copied_seq &&
1638                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1639 do_prequeue:
1640                                 tcp_prequeue_process(sk);
1641
1642                                 if ((chunk = len - tp->ucopy.len) != 0) {
1643                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1644                                         len -= chunk;
1645                                         copied += chunk;
1646                                 }
1647                         }
1648                 }
1649                 if ((flags & MSG_PEEK) &&
1650                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1651                         if (net_ratelimit())
1652                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1653                                        current->comm, task_pid_nr(current));
1654                         peek_seq = tp->copied_seq;
1655                 }
1656                 continue;
1657
1658         found_ok_skb:
1659                 /* Ok so how much can we use? */
1660                 used = skb->len - offset;
1661                 if (len < used)
1662                         used = len;
1663
1664                 /* Do we have urgent data here? */
1665                 if (tp->urg_data) {
1666                         u32 urg_offset = tp->urg_seq - *seq;
1667                         if (urg_offset < used) {
1668                                 if (!urg_offset) {
1669                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1670                                                 ++*seq;
1671                                                 urg_hole++;
1672                                                 offset++;
1673                                                 used--;
1674                                                 if (!used)
1675                                                         goto skip_copy;
1676                                         }
1677                                 } else
1678                                         used = urg_offset;
1679                         }
1680                 }
1681
1682                 if (!(flags & MSG_TRUNC)) {
1683 #ifdef CONFIG_NET_DMA
1684                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1685                                 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1686
1687                         if (tp->ucopy.dma_chan) {
1688                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1689                                         tp->ucopy.dma_chan, skb, offset,
1690                                         msg->msg_iov, used,
1691                                         tp->ucopy.pinned_list);
1692
1693                                 if (tp->ucopy.dma_cookie < 0) {
1694
1695                                         printk(KERN_ALERT "dma_cookie < 0\n");
1696
1697                                         /* Exception. Bailout! */
1698                                         if (!copied)
1699                                                 copied = -EFAULT;
1700                                         break;
1701                                 }
1702
1703                                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1704
1705                                 if ((offset + used) == skb->len)
1706                                         copied_early = 1;
1707
1708                         } else
1709 #endif
1710                         {
1711                                 err = skb_copy_datagram_iovec(skb, offset,
1712                                                 msg->msg_iov, used);
1713                                 if (err) {
1714                                         /* Exception. Bailout! */
1715                                         if (!copied)
1716                                                 copied = -EFAULT;
1717                                         break;
1718                                 }
1719                         }
1720                 }
1721
1722                 *seq += used;
1723                 copied += used;
1724                 len -= used;
1725
1726                 tcp_rcv_space_adjust(sk);
1727
1728 skip_copy:
1729                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1730                         tp->urg_data = 0;
1731                         tcp_fast_path_check(sk);
1732                 }
1733                 if (used + offset < skb->len)
1734                         continue;
1735
1736                 if (tcp_hdr(skb)->fin)
1737                         goto found_fin_ok;
1738                 if (!(flags & MSG_PEEK)) {
1739                         sk_eat_skb(sk, skb, copied_early);
1740                         copied_early = 0;
1741                 }
1742                 continue;
1743
1744         found_fin_ok:
1745                 /* Process the FIN. */
1746                 ++*seq;
1747                 if (!(flags & MSG_PEEK)) {
1748                         sk_eat_skb(sk, skb, copied_early);
1749                         copied_early = 0;
1750                 }
1751                 break;
1752         } while (len > 0);
1753
1754         if (user_recv) {
1755                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1756                         int chunk;
1757
1758                         tp->ucopy.len = copied > 0 ? len : 0;
1759
1760                         tcp_prequeue_process(sk);
1761
1762                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1763                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1764                                 len -= chunk;
1765                                 copied += chunk;
1766                         }
1767                 }
1768
1769                 tp->ucopy.task = NULL;
1770                 tp->ucopy.len = 0;
1771         }
1772
1773 #ifdef CONFIG_NET_DMA
1774         tcp_service_net_dma(sk, true);  /* Wait for queue to drain */
1775         tp->ucopy.dma_chan = NULL;
1776
1777         if (tp->ucopy.pinned_list) {
1778                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1779                 tp->ucopy.pinned_list = NULL;
1780         }
1781 #endif
1782
1783         /* According to UNIX98, msg_name/msg_namelen are ignored
1784          * on connected socket. I was just happy when found this 8) --ANK
1785          */
1786
1787         /* Clean up data we have read: This will do ACK frames. */
1788         tcp_cleanup_rbuf(sk, copied);
1789
1790         release_sock(sk);
1791
1792         if (copied > 0)
1793                 uid_stat_tcp_rcv(current_uid(), copied);
1794         return copied;
1795
1796 out:
1797         release_sock(sk);
1798         return err;
1799
1800 recv_urg:
1801         err = tcp_recv_urg(sk, msg, len, flags);
1802         if (err > 0)
1803                 uid_stat_tcp_rcv(current_uid(), err);
1804         goto out;
1805 }
1806 EXPORT_SYMBOL(tcp_recvmsg);
1807
1808 void tcp_set_state(struct sock *sk, int state)
1809 {
1810         int oldstate = sk->sk_state;
1811
1812         switch (state) {
1813         case TCP_ESTABLISHED:
1814                 if (oldstate != TCP_ESTABLISHED)
1815                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1816                 break;
1817
1818         case TCP_CLOSE:
1819                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1820                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1821
1822                 sk->sk_prot->unhash(sk);
1823                 if (inet_csk(sk)->icsk_bind_hash &&
1824                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1825                         inet_put_port(sk);
1826                 /* fall through */
1827         default:
1828                 if (oldstate == TCP_ESTABLISHED)
1829                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1830         }
1831
1832         /* Change state AFTER socket is unhashed to avoid closed
1833          * socket sitting in hash tables.
1834          */
1835         sk->sk_state = state;
1836
1837 #ifdef STATE_TRACE
1838         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1839 #endif
1840 }
1841 EXPORT_SYMBOL_GPL(tcp_set_state);
1842
1843 /*
1844  *      State processing on a close. This implements the state shift for
1845  *      sending our FIN frame. Note that we only send a FIN for some
1846  *      states. A shutdown() may have already sent the FIN, or we may be
1847  *      closed.
1848  */
1849
1850 static const unsigned char new_state[16] = {
1851   /* current state:        new state:      action:      */
1852   /* (Invalid)          */ TCP_CLOSE,
1853   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1854   /* TCP_SYN_SENT       */ TCP_CLOSE,
1855   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1856   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1857   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1858   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1859   /* TCP_CLOSE          */ TCP_CLOSE,
1860   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1861   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1862   /* TCP_LISTEN         */ TCP_CLOSE,
1863   /* TCP_CLOSING        */ TCP_CLOSING,
1864 };
1865
1866 static int tcp_close_state(struct sock *sk)
1867 {
1868         int next = (int)new_state[sk->sk_state];
1869         int ns = next & TCP_STATE_MASK;
1870
1871         tcp_set_state(sk, ns);
1872
1873         return next & TCP_ACTION_FIN;
1874 }
1875
1876 /*
1877  *      Shutdown the sending side of a connection. Much like close except
1878  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1879  */
1880
1881 void tcp_shutdown(struct sock *sk, int how)
1882 {
1883         /*      We need to grab some memory, and put together a FIN,
1884          *      and then put it into the queue to be sent.
1885          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1886          */
1887         if (!(how & SEND_SHUTDOWN))
1888                 return;
1889
1890         /* If we've already sent a FIN, or it's a closed state, skip this. */
1891         if ((1 << sk->sk_state) &
1892             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1893              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1894                 /* Clear out any half completed packets.  FIN if needed. */
1895                 if (tcp_close_state(sk))
1896                         tcp_send_fin(sk);
1897         }
1898 }
1899 EXPORT_SYMBOL(tcp_shutdown);
1900
1901 void tcp_close(struct sock *sk, long timeout)
1902 {
1903         struct sk_buff *skb;
1904         int data_was_unread = 0;
1905         int state;
1906
1907         lock_sock(sk);
1908         sk->sk_shutdown = SHUTDOWN_MASK;
1909
1910         if (sk->sk_state == TCP_LISTEN) {
1911                 tcp_set_state(sk, TCP_CLOSE);
1912
1913                 /* Special case. */
1914                 inet_csk_listen_stop(sk);
1915
1916                 goto adjudge_to_death;
1917         }
1918
1919         /*  We need to flush the recv. buffs.  We do this only on the
1920          *  descriptor close, not protocol-sourced closes, because the
1921          *  reader process may not have drained the data yet!
1922          */
1923         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1924                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1925                           tcp_hdr(skb)->fin;
1926                 data_was_unread += len;
1927                 __kfree_skb(skb);
1928         }
1929
1930         sk_mem_reclaim(sk);
1931
1932         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
1933         if (sk->sk_state == TCP_CLOSE)
1934                 goto adjudge_to_death;
1935
1936         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1937          * data was lost. To witness the awful effects of the old behavior of
1938          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1939          * GET in an FTP client, suspend the process, wait for the client to
1940          * advertise a zero window, then kill -9 the FTP client, wheee...
1941          * Note: timeout is always zero in such a case.
1942          */
1943         if (data_was_unread) {
1944                 /* Unread data was tossed, zap the connection. */
1945                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1946                 tcp_set_state(sk, TCP_CLOSE);
1947                 tcp_send_active_reset(sk, sk->sk_allocation);
1948         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1949                 /* Check zero linger _after_ checking for unread data. */
1950                 sk->sk_prot->disconnect(sk, 0);
1951                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1952         } else if (tcp_close_state(sk)) {
1953                 /* We FIN if the application ate all the data before
1954                  * zapping the connection.
1955                  */
1956
1957                 /* RED-PEN. Formally speaking, we have broken TCP state
1958                  * machine. State transitions:
1959                  *
1960                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1961                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1962                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1963                  *
1964                  * are legal only when FIN has been sent (i.e. in window),
1965                  * rather than queued out of window. Purists blame.
1966                  *
1967                  * F.e. "RFC state" is ESTABLISHED,
1968                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1969                  *
1970                  * The visible declinations are that sometimes
1971                  * we enter time-wait state, when it is not required really
1972                  * (harmless), do not send active resets, when they are
1973                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1974                  * they look as CLOSING or LAST_ACK for Linux)
1975                  * Probably, I missed some more holelets.
1976                  *                                              --ANK
1977                  */
1978                 tcp_send_fin(sk);
1979         }
1980
1981         sk_stream_wait_close(sk, timeout);
1982
1983 adjudge_to_death:
1984         state = sk->sk_state;
1985         sock_hold(sk);
1986         sock_orphan(sk);
1987
1988         /* It is the last release_sock in its life. It will remove backlog. */
1989         release_sock(sk);
1990
1991
1992         /* Now socket is owned by kernel and we acquire BH lock
1993            to finish close. No need to check for user refs.
1994          */
1995         local_bh_disable();
1996         bh_lock_sock(sk);
1997         WARN_ON(sock_owned_by_user(sk));
1998
1999         percpu_counter_inc(sk->sk_prot->orphan_count);
2000
2001         /* Have we already been destroyed by a softirq or backlog? */
2002         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2003                 goto out;
2004
2005         /*      This is a (useful) BSD violating of the RFC. There is a
2006          *      problem with TCP as specified in that the other end could
2007          *      keep a socket open forever with no application left this end.
2008          *      We use a 3 minute timeout (about the same as BSD) then kill
2009          *      our end. If they send after that then tough - BUT: long enough
2010          *      that we won't make the old 4*rto = almost no time - whoops
2011          *      reset mistake.
2012          *
2013          *      Nope, it was not mistake. It is really desired behaviour
2014          *      f.e. on http servers, when such sockets are useless, but
2015          *      consume significant resources. Let's do it with special
2016          *      linger2 option.                                 --ANK
2017          */
2018
2019         if (sk->sk_state == TCP_FIN_WAIT2) {
2020                 struct tcp_sock *tp = tcp_sk(sk);
2021                 if (tp->linger2 < 0) {
2022                         tcp_set_state(sk, TCP_CLOSE);
2023                         tcp_send_active_reset(sk, GFP_ATOMIC);
2024                         NET_INC_STATS_BH(sock_net(sk),
2025                                         LINUX_MIB_TCPABORTONLINGER);
2026                 } else {
2027                         const int tmo = tcp_fin_time(sk);
2028
2029                         if (tmo > TCP_TIMEWAIT_LEN) {
2030                                 inet_csk_reset_keepalive_timer(sk,
2031                                                 tmo - TCP_TIMEWAIT_LEN);
2032                         } else {
2033                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2034                                 goto out;
2035                         }
2036                 }
2037         }
2038         if (sk->sk_state != TCP_CLOSE) {
2039                 sk_mem_reclaim(sk);
2040                 if (tcp_too_many_orphans(sk, 0)) {
2041                         if (net_ratelimit())
2042                                 printk(KERN_INFO "TCP: too many of orphaned "
2043                                        "sockets\n");
2044                         tcp_set_state(sk, TCP_CLOSE);
2045                         tcp_send_active_reset(sk, GFP_ATOMIC);
2046                         NET_INC_STATS_BH(sock_net(sk),
2047                                         LINUX_MIB_TCPABORTONMEMORY);
2048                 }
2049         }
2050
2051         if (sk->sk_state == TCP_CLOSE)
2052                 inet_csk_destroy_sock(sk);
2053         /* Otherwise, socket is reprieved until protocol close. */
2054
2055 out:
2056         bh_unlock_sock(sk);
2057         local_bh_enable();
2058         sock_put(sk);
2059 }
2060 EXPORT_SYMBOL(tcp_close);
2061
2062 /* These states need RST on ABORT according to RFC793 */
2063
2064 static inline int tcp_need_reset(int state)
2065 {
2066         return (1 << state) &
2067                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2068                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2069 }
2070
2071 int tcp_disconnect(struct sock *sk, int flags)
2072 {
2073         struct inet_sock *inet = inet_sk(sk);
2074         struct inet_connection_sock *icsk = inet_csk(sk);
2075         struct tcp_sock *tp = tcp_sk(sk);
2076         int err = 0;
2077         int old_state = sk->sk_state;
2078
2079         if (old_state != TCP_CLOSE)
2080                 tcp_set_state(sk, TCP_CLOSE);
2081
2082         /* ABORT function of RFC793 */
2083         if (old_state == TCP_LISTEN) {
2084                 inet_csk_listen_stop(sk);
2085         } else if (tcp_need_reset(old_state) ||
2086                    (tp->snd_nxt != tp->write_seq &&
2087                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2088                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2089                  * states
2090                  */
2091                 tcp_send_active_reset(sk, gfp_any());
2092                 sk->sk_err = ECONNRESET;
2093         } else if (old_state == TCP_SYN_SENT)
2094                 sk->sk_err = ECONNRESET;
2095
2096         tcp_clear_xmit_timers(sk);
2097         __skb_queue_purge(&sk->sk_receive_queue);
2098         tcp_write_queue_purge(sk);
2099         __skb_queue_purge(&tp->out_of_order_queue);
2100 #ifdef CONFIG_NET_DMA
2101         __skb_queue_purge(&sk->sk_async_wait_queue);
2102 #endif
2103
2104         inet->inet_dport = 0;
2105
2106         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2107                 inet_reset_saddr(sk);
2108
2109         sk->sk_shutdown = 0;
2110         sock_reset_flag(sk, SOCK_DONE);
2111         tp->srtt = 0;
2112         if ((tp->write_seq += tp->max_window + 2) == 0)
2113                 tp->write_seq = 1;
2114         icsk->icsk_backoff = 0;
2115         tp->snd_cwnd = 2;
2116         icsk->icsk_probes_out = 0;
2117         tp->packets_out = 0;
2118         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2119         tp->snd_cwnd_cnt = 0;
2120         tp->bytes_acked = 0;
2121         tp->window_clamp = 0;
2122         tcp_set_ca_state(sk, TCP_CA_Open);
2123         tcp_clear_retrans(tp);
2124         inet_csk_delack_init(sk);
2125         tcp_init_send_head(sk);
2126         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2127         __sk_dst_reset(sk);
2128
2129         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2130
2131         sk->sk_error_report(sk);
2132         return err;
2133 }
2134 EXPORT_SYMBOL(tcp_disconnect);
2135
2136 /*
2137  *      Socket option code for TCP.
2138  */
2139 static int do_tcp_setsockopt(struct sock *sk, int level,
2140                 int optname, char __user *optval, unsigned int optlen)
2141 {
2142         struct tcp_sock *tp = tcp_sk(sk);
2143         struct inet_connection_sock *icsk = inet_csk(sk);
2144         int val;
2145         int err = 0;
2146
2147         /* These are data/string values, all the others are ints */
2148         switch (optname) {
2149         case TCP_CONGESTION: {
2150                 char name[TCP_CA_NAME_MAX];
2151
2152                 if (optlen < 1)
2153                         return -EINVAL;
2154
2155                 val = strncpy_from_user(name, optval,
2156                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2157                 if (val < 0)
2158                         return -EFAULT;
2159                 name[val] = 0;
2160
2161                 lock_sock(sk);
2162                 err = tcp_set_congestion_control(sk, name);
2163                 release_sock(sk);
2164                 return err;
2165         }
2166         case TCP_COOKIE_TRANSACTIONS: {
2167                 struct tcp_cookie_transactions ctd;
2168                 struct tcp_cookie_values *cvp = NULL;
2169
2170                 if (sizeof(ctd) > optlen)
2171                         return -EINVAL;
2172                 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2173                         return -EFAULT;
2174
2175                 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2176                     ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2177                         return -EINVAL;
2178
2179                 if (ctd.tcpct_cookie_desired == 0) {
2180                         /* default to global value */
2181                 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2182                            ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2183                            ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2184                         return -EINVAL;
2185                 }
2186
2187                 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2188                         /* Supercedes all other values */
2189                         lock_sock(sk);
2190                         if (tp->cookie_values != NULL) {
2191                                 kref_put(&tp->cookie_values->kref,
2192                                          tcp_cookie_values_release);
2193                                 tp->cookie_values = NULL;
2194                         }
2195                         tp->rx_opt.cookie_in_always = 0; /* false */
2196                         tp->rx_opt.cookie_out_never = 1; /* true */
2197                         release_sock(sk);
2198                         return err;
2199                 }
2200
2201                 /* Allocate ancillary memory before locking.
2202                  */
2203                 if (ctd.tcpct_used > 0 ||
2204                     (tp->cookie_values == NULL &&
2205                      (sysctl_tcp_cookie_size > 0 ||
2206                       ctd.tcpct_cookie_desired > 0 ||
2207                       ctd.tcpct_s_data_desired > 0))) {
2208                         cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2209                                       GFP_KERNEL);
2210                         if (cvp == NULL)
2211                                 return -ENOMEM;
2212
2213                         kref_init(&cvp->kref);
2214                 }
2215                 lock_sock(sk);
2216                 tp->rx_opt.cookie_in_always =
2217                         (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2218                 tp->rx_opt.cookie_out_never = 0; /* false */
2219
2220                 if (tp->cookie_values != NULL) {
2221                         if (cvp != NULL) {
2222                                 /* Changed values are recorded by a changed
2223                                  * pointer, ensuring the cookie will differ,
2224                                  * without separately hashing each value later.
2225                                  */
2226                                 kref_put(&tp->cookie_values->kref,
2227                                          tcp_cookie_values_release);
2228                         } else {
2229                                 cvp = tp->cookie_values;
2230                         }
2231                 }
2232
2233                 if (cvp != NULL) {
2234                         cvp->cookie_desired = ctd.tcpct_cookie_desired;
2235
2236                         if (ctd.tcpct_used > 0) {
2237                                 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2238                                        ctd.tcpct_used);
2239                                 cvp->s_data_desired = ctd.tcpct_used;
2240                                 cvp->s_data_constant = 1; /* true */
2241                         } else {
2242                                 /* No constant payload data. */
2243                                 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2244                                 cvp->s_data_constant = 0; /* false */
2245                         }
2246
2247                         tp->cookie_values = cvp;
2248                 }
2249                 release_sock(sk);
2250                 return err;
2251         }
2252         default:
2253                 /* fallthru */
2254                 break;
2255         }
2256
2257         if (optlen < sizeof(int))
2258                 return -EINVAL;
2259
2260         if (get_user(val, (int __user *)optval))
2261                 return -EFAULT;
2262
2263         lock_sock(sk);
2264
2265         switch (optname) {
2266         case TCP_MAXSEG:
2267                 /* Values greater than interface MTU won't take effect. However
2268                  * at the point when this call is done we typically don't yet
2269                  * know which interface is going to be used */
2270                 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2271                         err = -EINVAL;
2272                         break;
2273                 }
2274                 tp->rx_opt.user_mss = val;
2275                 break;
2276
2277         case TCP_NODELAY:
2278                 if (val) {
2279                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2280                          * this option on corked socket is remembered, but
2281                          * it is not activated until cork is cleared.
2282                          *
2283                          * However, when TCP_NODELAY is set we make
2284                          * an explicit push, which overrides even TCP_CORK
2285                          * for currently queued segments.
2286                          */
2287                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2288                         tcp_push_pending_frames(sk);
2289                 } else {
2290                         tp->nonagle &= ~TCP_NAGLE_OFF;
2291                 }
2292                 break;
2293
2294         case TCP_THIN_LINEAR_TIMEOUTS:
2295                 if (val < 0 || val > 1)
2296                         err = -EINVAL;
2297                 else
2298                         tp->thin_lto = val;
2299                 break;
2300
2301         case TCP_THIN_DUPACK:
2302                 if (val < 0 || val > 1)
2303                         err = -EINVAL;
2304                 else
2305                         tp->thin_dupack = val;
2306                 break;
2307
2308         case TCP_CORK:
2309                 /* When set indicates to always queue non-full frames.
2310                  * Later the user clears this option and we transmit
2311                  * any pending partial frames in the queue.  This is
2312                  * meant to be used alongside sendfile() to get properly
2313                  * filled frames when the user (for example) must write
2314                  * out headers with a write() call first and then use
2315                  * sendfile to send out the data parts.
2316                  *
2317                  * TCP_CORK can be set together with TCP_NODELAY and it is
2318                  * stronger than TCP_NODELAY.
2319                  */
2320                 if (val) {
2321                         tp->nonagle |= TCP_NAGLE_CORK;
2322                 } else {
2323                         tp->nonagle &= ~TCP_NAGLE_CORK;
2324                         if (tp->nonagle&TCP_NAGLE_OFF)
2325                                 tp->nonagle |= TCP_NAGLE_PUSH;
2326                         tcp_push_pending_frames(sk);
2327                 }
2328                 break;
2329
2330         case TCP_KEEPIDLE:
2331                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2332                         err = -EINVAL;
2333                 else {
2334                         tp->keepalive_time = val * HZ;
2335                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2336                             !((1 << sk->sk_state) &
2337                               (TCPF_CLOSE | TCPF_LISTEN))) {
2338                                 u32 elapsed = keepalive_time_elapsed(tp);
2339                                 if (tp->keepalive_time > elapsed)
2340                                         elapsed = tp->keepalive_time - elapsed;
2341                                 else
2342                                         elapsed = 0;
2343                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2344                         }
2345                 }
2346                 break;
2347         case TCP_KEEPINTVL:
2348                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2349                         err = -EINVAL;
2350                 else
2351                         tp->keepalive_intvl = val * HZ;
2352                 break;
2353         case TCP_KEEPCNT:
2354                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2355                         err = -EINVAL;
2356                 else
2357                         tp->keepalive_probes = val;
2358                 break;
2359         case TCP_SYNCNT:
2360                 if (val < 1 || val > MAX_TCP_SYNCNT)
2361                         err = -EINVAL;
2362                 else
2363                         icsk->icsk_syn_retries = val;
2364                 break;
2365
2366         case TCP_LINGER2:
2367                 if (val < 0)
2368                         tp->linger2 = -1;
2369                 else if (val > sysctl_tcp_fin_timeout / HZ)
2370                         tp->linger2 = 0;
2371                 else
2372                         tp->linger2 = val * HZ;
2373                 break;
2374
2375         case TCP_DEFER_ACCEPT:
2376                 /* Translate value in seconds to number of retransmits */
2377                 icsk->icsk_accept_queue.rskq_defer_accept =
2378                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2379                                         TCP_RTO_MAX / HZ);
2380                 break;
2381
2382         case TCP_WINDOW_CLAMP:
2383                 if (!val) {
2384                         if (sk->sk_state != TCP_CLOSE) {
2385                                 err = -EINVAL;
2386                                 break;
2387                         }
2388                         tp->window_clamp = 0;
2389                 } else
2390                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2391                                                 SOCK_MIN_RCVBUF / 2 : val;
2392                 break;
2393
2394         case TCP_QUICKACK:
2395                 if (!val) {
2396                         icsk->icsk_ack.pingpong = 1;
2397                 } else {
2398                         icsk->icsk_ack.pingpong = 0;
2399                         if ((1 << sk->sk_state) &
2400                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2401                             inet_csk_ack_scheduled(sk)) {
2402                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2403                                 tcp_cleanup_rbuf(sk, 1);
2404                                 if (!(val & 1))
2405                                         icsk->icsk_ack.pingpong = 1;
2406                         }
2407                 }
2408                 break;
2409
2410 #ifdef CONFIG_TCP_MD5SIG
2411         case TCP_MD5SIG:
2412                 /* Read the IP->Key mappings from userspace */
2413                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2414                 break;
2415 #endif
2416         case TCP_USER_TIMEOUT:
2417                 /* Cap the max timeout in ms TCP will retry/retrans
2418                  * before giving up and aborting (ETIMEDOUT) a connection.
2419                  */
2420                 if (val < 0)
2421                         err = -EINVAL;
2422                 else
2423                         icsk->icsk_user_timeout = msecs_to_jiffies(val);
2424                 break;
2425         default:
2426                 err = -ENOPROTOOPT;
2427                 break;
2428         }
2429
2430         release_sock(sk);
2431         return err;
2432 }
2433
2434 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2435                    unsigned int optlen)
2436 {
2437         struct inet_connection_sock *icsk = inet_csk(sk);
2438
2439         if (level != SOL_TCP)
2440                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2441                                                      optval, optlen);
2442         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2443 }
2444 EXPORT_SYMBOL(tcp_setsockopt);
2445
2446 #ifdef CONFIG_COMPAT
2447 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2448                           char __user *optval, unsigned int optlen)
2449 {
2450         if (level != SOL_TCP)
2451                 return inet_csk_compat_setsockopt(sk, level, optname,
2452                                                   optval, optlen);
2453         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2454 }
2455 EXPORT_SYMBOL(compat_tcp_setsockopt);
2456 #endif
2457
2458 /* Return information about state of tcp endpoint in API format. */
2459 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2460 {
2461         struct tcp_sock *tp = tcp_sk(sk);
2462         const struct inet_connection_sock *icsk = inet_csk(sk);
2463         u32 now = tcp_time_stamp;
2464
2465         memset(info, 0, sizeof(*info));
2466
2467         info->tcpi_state = sk->sk_state;
2468         info->tcpi_ca_state = icsk->icsk_ca_state;
2469         info->tcpi_retransmits = icsk->icsk_retransmits;
2470         info->tcpi_probes = icsk->icsk_probes_out;
2471         info->tcpi_backoff = icsk->icsk_backoff;
2472
2473         if (tp->rx_opt.tstamp_ok)
2474                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2475         if (tcp_is_sack(tp))
2476                 info->tcpi_options |= TCPI_OPT_SACK;
2477         if (tp->rx_opt.wscale_ok) {
2478                 info->tcpi_options |= TCPI_OPT_WSCALE;
2479                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2480                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2481         }
2482
2483         if (tp->ecn_flags&TCP_ECN_OK)
2484                 info->tcpi_options |= TCPI_OPT_ECN;
2485
2486         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2487         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2488         info->tcpi_snd_mss = tp->mss_cache;
2489         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2490
2491         if (sk->sk_state == TCP_LISTEN) {
2492                 info->tcpi_unacked = sk->sk_ack_backlog;
2493                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2494         } else {
2495                 info->tcpi_unacked = tp->packets_out;
2496                 info->tcpi_sacked = tp->sacked_out;
2497         }
2498         info->tcpi_lost = tp->lost_out;
2499         info->tcpi_retrans = tp->retrans_out;
2500         info->tcpi_fackets = tp->fackets_out;
2501
2502         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2503         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2504         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2505
2506         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2507         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2508         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2509         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2510         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2511         info->tcpi_snd_cwnd = tp->snd_cwnd;
2512         info->tcpi_advmss = tp->advmss;
2513         info->tcpi_reordering = tp->reordering;
2514
2515         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2516         info->tcpi_rcv_space = tp->rcvq_space.space;
2517
2518         info->tcpi_total_retrans = tp->total_retrans;
2519 }
2520 EXPORT_SYMBOL_GPL(tcp_get_info);
2521
2522 static int do_tcp_getsockopt(struct sock *sk, int level,
2523                 int optname, char __user *optval, int __user *optlen)
2524 {
2525         struct inet_connection_sock *icsk = inet_csk(sk);
2526         struct tcp_sock *tp = tcp_sk(sk);
2527         int val, len;
2528
2529         if (get_user(len, optlen))
2530                 return -EFAULT;
2531
2532         len = min_t(unsigned int, len, sizeof(int));
2533
2534         if (len < 0)
2535                 return -EINVAL;
2536
2537         switch (optname) {
2538         case TCP_MAXSEG:
2539                 val = tp->mss_cache;
2540                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2541                         val = tp->rx_opt.user_mss;
2542                 break;
2543         case TCP_NODELAY:
2544                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2545                 break;
2546         case TCP_CORK:
2547                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2548                 break;
2549         case TCP_KEEPIDLE:
2550                 val = keepalive_time_when(tp) / HZ;
2551                 break;
2552         case TCP_KEEPINTVL:
2553                 val = keepalive_intvl_when(tp) / HZ;
2554                 break;
2555         case TCP_KEEPCNT:
2556                 val = keepalive_probes(tp);
2557                 break;
2558         case TCP_SYNCNT:
2559                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2560                 break;
2561         case TCP_LINGER2:
2562                 val = tp->linger2;
2563                 if (val >= 0)
2564                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2565                 break;
2566         case TCP_DEFER_ACCEPT:
2567                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2568                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2569                 break;
2570         case TCP_WINDOW_CLAMP:
2571                 val = tp->window_clamp;
2572                 break;
2573         case TCP_INFO: {
2574                 struct tcp_info info;
2575
2576                 if (get_user(len, optlen))
2577                         return -EFAULT;
2578
2579                 tcp_get_info(sk, &info);
2580
2581                 len = min_t(unsigned int, len, sizeof(info));
2582                 if (put_user(len, optlen))
2583                         return -EFAULT;
2584                 if (copy_to_user(optval, &info, len))
2585                         return -EFAULT;
2586                 return 0;
2587         }
2588         case TCP_QUICKACK:
2589                 val = !icsk->icsk_ack.pingpong;
2590                 break;
2591
2592         case TCP_CONGESTION:
2593                 if (get_user(len, optlen))
2594                         return -EFAULT;
2595                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2596                 if (put_user(len, optlen))
2597                         return -EFAULT;
2598                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2599                         return -EFAULT;
2600                 return 0;
2601
2602         case TCP_COOKIE_TRANSACTIONS: {
2603                 struct tcp_cookie_transactions ctd;
2604                 struct tcp_cookie_values *cvp = tp->cookie_values;
2605
2606                 if (get_user(len, optlen))
2607                         return -EFAULT;
2608                 if (len < sizeof(ctd))
2609                         return -EINVAL;
2610
2611                 memset(&ctd, 0, sizeof(ctd));
2612                 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2613                                    TCP_COOKIE_IN_ALWAYS : 0)
2614                                 | (tp->rx_opt.cookie_out_never ?
2615                                    TCP_COOKIE_OUT_NEVER : 0);
2616
2617                 if (cvp != NULL) {
2618                         ctd.tcpct_flags |= (cvp->s_data_in ?
2619                                             TCP_S_DATA_IN : 0)
2620                                          | (cvp->s_data_out ?
2621                                             TCP_S_DATA_OUT : 0);
2622
2623                         ctd.tcpct_cookie_desired = cvp->cookie_desired;
2624                         ctd.tcpct_s_data_desired = cvp->s_data_desired;
2625
2626                         memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2627                                cvp->cookie_pair_size);
2628                         ctd.tcpct_used = cvp->cookie_pair_size;
2629                 }
2630
2631                 if (put_user(sizeof(ctd), optlen))
2632                         return -EFAULT;
2633                 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2634                         return -EFAULT;
2635                 return 0;
2636         }
2637         case TCP_THIN_LINEAR_TIMEOUTS:
2638                 val = tp->thin_lto;
2639                 break;
2640         case TCP_THIN_DUPACK:
2641                 val = tp->thin_dupack;
2642                 break;
2643
2644         case TCP_USER_TIMEOUT:
2645                 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2646                 break;
2647         default:
2648                 return -ENOPROTOOPT;
2649         }
2650
2651         if (put_user(len, optlen))
2652                 return -EFAULT;
2653         if (copy_to_user(optval, &val, len))
2654                 return -EFAULT;
2655         return 0;
2656 }
2657
2658 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2659                    int __user *optlen)
2660 {
2661         struct inet_connection_sock *icsk = inet_csk(sk);
2662
2663         if (level != SOL_TCP)
2664                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2665                                                      optval, optlen);
2666         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2667 }
2668 EXPORT_SYMBOL(tcp_getsockopt);
2669
2670 #ifdef CONFIG_COMPAT
2671 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2672                           char __user *optval, int __user *optlen)
2673 {
2674         if (level != SOL_TCP)
2675                 return inet_csk_compat_getsockopt(sk, level, optname,
2676                                                   optval, optlen);
2677         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2678 }
2679 EXPORT_SYMBOL(compat_tcp_getsockopt);
2680 #endif
2681
2682 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, u32 features)
2683 {
2684         struct sk_buff *segs = ERR_PTR(-EINVAL);
2685         struct tcphdr *th;
2686         unsigned thlen;
2687         unsigned int seq;
2688         __be32 delta;
2689         unsigned int oldlen;
2690         unsigned int mss;
2691
2692         if (!pskb_may_pull(skb, sizeof(*th)))
2693                 goto out;
2694
2695         th = tcp_hdr(skb);
2696         thlen = th->doff * 4;
2697         if (thlen < sizeof(*th))
2698                 goto out;
2699
2700         if (!pskb_may_pull(skb, thlen))
2701                 goto out;
2702
2703         oldlen = (u16)~skb->len;
2704         __skb_pull(skb, thlen);
2705
2706         mss = skb_shinfo(skb)->gso_size;
2707         if (unlikely(skb->len <= mss))
2708                 goto out;
2709
2710         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2711                 /* Packet is from an untrusted source, reset gso_segs. */
2712                 int type = skb_shinfo(skb)->gso_type;
2713
2714                 if (unlikely(type &
2715                              ~(SKB_GSO_TCPV4 |
2716                                SKB_GSO_DODGY |
2717                                SKB_GSO_TCP_ECN |
2718                                SKB_GSO_TCPV6 |
2719                                0) ||
2720                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2721                         goto out;
2722
2723                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2724
2725                 segs = NULL;
2726                 goto out;
2727         }
2728
2729         segs = skb_segment(skb, features);
2730         if (IS_ERR(segs))
2731                 goto out;
2732
2733         delta = htonl(oldlen + (thlen + mss));
2734
2735         skb = segs;
2736         th = tcp_hdr(skb);
2737         seq = ntohl(th->seq);
2738
2739         do {
2740                 th->fin = th->psh = 0;
2741
2742                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2743                                        (__force u32)delta));
2744                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2745                         th->check =
2746                              csum_fold(csum_partial(skb_transport_header(skb),
2747                                                     thlen, skb->csum));
2748
2749                 seq += mss;
2750                 skb = skb->next;
2751                 th = tcp_hdr(skb);
2752
2753                 th->seq = htonl(seq);
2754                 th->cwr = 0;
2755         } while (skb->next);
2756
2757         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2758                       skb->data_len);
2759         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2760                                 (__force u32)delta));
2761         if (skb->ip_summed != CHECKSUM_PARTIAL)
2762                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2763                                                    thlen, skb->csum));
2764
2765 out:
2766         return segs;
2767 }
2768 EXPORT_SYMBOL(tcp_tso_segment);
2769
2770 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2771 {
2772         struct sk_buff **pp = NULL;
2773         struct sk_buff *p;
2774         struct tcphdr *th;
2775         struct tcphdr *th2;
2776         unsigned int len;
2777         unsigned int thlen;
2778         __be32 flags;
2779         unsigned int mss = 1;
2780         unsigned int hlen;
2781         unsigned int off;
2782         int flush = 1;
2783         int i;
2784
2785         off = skb_gro_offset(skb);
2786         hlen = off + sizeof(*th);
2787         th = skb_gro_header_fast(skb, off);
2788         if (skb_gro_header_hard(skb, hlen)) {
2789                 th = skb_gro_header_slow(skb, hlen, off);
2790                 if (unlikely(!th))
2791                         goto out;
2792         }
2793
2794         thlen = th->doff * 4;
2795         if (thlen < sizeof(*th))
2796                 goto out;
2797
2798         hlen = off + thlen;
2799         if (skb_gro_header_hard(skb, hlen)) {
2800                 th = skb_gro_header_slow(skb, hlen, off);
2801                 if (unlikely(!th))
2802                         goto out;
2803         }
2804
2805         skb_gro_pull(skb, thlen);
2806
2807         len = skb_gro_len(skb);
2808         flags = tcp_flag_word(th);
2809
2810         for (; (p = *head); head = &p->next) {
2811                 if (!NAPI_GRO_CB(p)->same_flow)
2812                         continue;
2813
2814                 th2 = tcp_hdr(p);
2815
2816                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2817                         NAPI_GRO_CB(p)->same_flow = 0;
2818                         continue;
2819                 }
2820
2821                 goto found;
2822         }
2823
2824         goto out_check_final;
2825
2826 found:
2827         flush = NAPI_GRO_CB(p)->flush;
2828         flush |= (__force int)(flags & TCP_FLAG_CWR);
2829         flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
2830                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
2831         flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
2832         for (i = sizeof(*th); i < thlen; i += 4)
2833                 flush |= *(u32 *)((u8 *)th + i) ^
2834                          *(u32 *)((u8 *)th2 + i);
2835
2836         mss = skb_shinfo(p)->gso_size;
2837
2838         flush |= (len - 1) >= mss;
2839         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2840
2841         if (flush || skb_gro_receive(head, skb)) {
2842                 mss = 1;
2843                 goto out_check_final;
2844         }
2845
2846         p = *head;
2847         th2 = tcp_hdr(p);
2848         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2849
2850 out_check_final:
2851         flush = len < mss;
2852         flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
2853                                         TCP_FLAG_RST | TCP_FLAG_SYN |
2854                                         TCP_FLAG_FIN));
2855
2856         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2857                 pp = head;
2858
2859 out:
2860         NAPI_GRO_CB(skb)->flush |= flush;
2861
2862         return pp;
2863 }
2864 EXPORT_SYMBOL(tcp_gro_receive);
2865
2866 int tcp_gro_complete(struct sk_buff *skb)
2867 {
2868         struct tcphdr *th = tcp_hdr(skb);
2869
2870         skb->csum_start = skb_transport_header(skb) - skb->head;
2871         skb->csum_offset = offsetof(struct tcphdr, check);
2872         skb->ip_summed = CHECKSUM_PARTIAL;
2873
2874         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2875
2876         if (th->cwr)
2877                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2878
2879         return 0;
2880 }
2881 EXPORT_SYMBOL(tcp_gro_complete);
2882
2883 #ifdef CONFIG_TCP_MD5SIG
2884 static unsigned long tcp_md5sig_users;
2885 static struct tcp_md5sig_pool * __percpu *tcp_md5sig_pool;
2886 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2887
2888 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool * __percpu *pool)
2889 {
2890         int cpu;
2891         for_each_possible_cpu(cpu) {
2892                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2893                 if (p) {
2894                         if (p->md5_desc.tfm)
2895                                 crypto_free_hash(p->md5_desc.tfm);
2896                         kfree(p);
2897                 }
2898         }
2899         free_percpu(pool);
2900 }
2901
2902 void tcp_free_md5sig_pool(void)
2903 {
2904         struct tcp_md5sig_pool * __percpu *pool = NULL;
2905
2906         spin_lock_bh(&tcp_md5sig_pool_lock);
2907         if (--tcp_md5sig_users == 0) {
2908                 pool = tcp_md5sig_pool;
2909                 tcp_md5sig_pool = NULL;
2910         }
2911         spin_unlock_bh(&tcp_md5sig_pool_lock);
2912         if (pool)
2913                 __tcp_free_md5sig_pool(pool);
2914 }
2915 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2916
2917 static struct tcp_md5sig_pool * __percpu *
2918 __tcp_alloc_md5sig_pool(struct sock *sk)
2919 {
2920         int cpu;
2921         struct tcp_md5sig_pool * __percpu *pool;
2922
2923         pool = alloc_percpu(struct tcp_md5sig_pool *);
2924         if (!pool)
2925                 return NULL;
2926
2927         for_each_possible_cpu(cpu) {
2928                 struct tcp_md5sig_pool *p;
2929                 struct crypto_hash *hash;
2930
2931                 p = kzalloc(sizeof(*p), sk->sk_allocation);
2932                 if (!p)
2933                         goto out_free;
2934                 *per_cpu_ptr(pool, cpu) = p;
2935
2936                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2937                 if (!hash || IS_ERR(hash))
2938                         goto out_free;
2939
2940                 p->md5_desc.tfm = hash;
2941         }
2942         return pool;
2943 out_free:
2944         __tcp_free_md5sig_pool(pool);
2945         return NULL;
2946 }
2947
2948 struct tcp_md5sig_pool * __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
2949 {
2950         struct tcp_md5sig_pool * __percpu *pool;
2951         int alloc = 0;
2952
2953 retry:
2954         spin_lock_bh(&tcp_md5sig_pool_lock);
2955         pool = tcp_md5sig_pool;
2956         if (tcp_md5sig_users++ == 0) {
2957                 alloc = 1;
2958                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2959         } else if (!pool) {
2960                 tcp_md5sig_users--;
2961                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2962                 cpu_relax();
2963                 goto retry;
2964         } else
2965                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2966
2967         if (alloc) {
2968                 /* we cannot hold spinlock here because this may sleep. */
2969                 struct tcp_md5sig_pool * __percpu *p;
2970
2971                 p = __tcp_alloc_md5sig_pool(sk);
2972                 spin_lock_bh(&tcp_md5sig_pool_lock);
2973                 if (!p) {
2974                         tcp_md5sig_users--;
2975                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2976                         return NULL;
2977                 }
2978                 pool = tcp_md5sig_pool;
2979                 if (pool) {
2980                         /* oops, it has already been assigned. */
2981                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2982                         __tcp_free_md5sig_pool(p);
2983                 } else {
2984                         tcp_md5sig_pool = pool = p;
2985                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2986                 }
2987         }
2988         return pool;
2989 }
2990 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2991
2992
2993 /**
2994  *      tcp_get_md5sig_pool - get md5sig_pool for this user
2995  *
2996  *      We use percpu structure, so if we succeed, we exit with preemption
2997  *      and BH disabled, to make sure another thread or softirq handling
2998  *      wont try to get same context.
2999  */
3000 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3001 {
3002         struct tcp_md5sig_pool * __percpu *p;
3003
3004         local_bh_disable();
3005
3006         spin_lock(&tcp_md5sig_pool_lock);
3007         p = tcp_md5sig_pool;
3008         if (p)
3009                 tcp_md5sig_users++;
3010         spin_unlock(&tcp_md5sig_pool_lock);
3011
3012         if (p)
3013                 return *this_cpu_ptr(p);
3014
3015         local_bh_enable();
3016         return NULL;
3017 }
3018 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3019
3020 void tcp_put_md5sig_pool(void)
3021 {
3022         local_bh_enable();
3023         tcp_free_md5sig_pool();
3024 }
3025 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3026
3027 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3028                         struct tcphdr *th)
3029 {
3030         struct scatterlist sg;
3031         int err;
3032
3033         __sum16 old_checksum = th->check;
3034         th->check = 0;
3035         /* options aren't included in the hash */
3036         sg_init_one(&sg, th, sizeof(struct tcphdr));
3037         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
3038         th->check = old_checksum;
3039         return err;
3040 }
3041 EXPORT_SYMBOL(tcp_md5_hash_header);
3042
3043 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3044                           struct sk_buff *skb, unsigned header_len)
3045 {
3046         struct scatterlist sg;
3047         const struct tcphdr *tp = tcp_hdr(skb);
3048         struct hash_desc *desc = &hp->md5_desc;
3049         unsigned i;
3050         const unsigned head_data_len = skb_headlen(skb) > header_len ?
3051                                        skb_headlen(skb) - header_len : 0;
3052         const struct skb_shared_info *shi = skb_shinfo(skb);
3053         struct sk_buff *frag_iter;
3054
3055         sg_init_table(&sg, 1);
3056
3057         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3058         if (crypto_hash_update(desc, &sg, head_data_len))
3059                 return 1;
3060
3061         for (i = 0; i < shi->nr_frags; ++i) {
3062                 const struct skb_frag_struct *f = &shi->frags[i];
3063                 sg_set_page(&sg, f->page, f->size, f->page_offset);
3064                 if (crypto_hash_update(desc, &sg, f->size))
3065                         return 1;
3066         }
3067
3068         skb_walk_frags(skb, frag_iter)
3069                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3070                         return 1;
3071
3072         return 0;
3073 }
3074 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3075
3076 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
3077 {
3078         struct scatterlist sg;
3079
3080         sg_init_one(&sg, key->key, key->keylen);
3081         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3082 }
3083 EXPORT_SYMBOL(tcp_md5_hash_key);
3084
3085 #endif
3086
3087 /**
3088  * Each Responder maintains up to two secret values concurrently for
3089  * efficient secret rollover.  Each secret value has 4 states:
3090  *
3091  * Generating.  (tcp_secret_generating != tcp_secret_primary)
3092  *    Generates new Responder-Cookies, but not yet used for primary
3093  *    verification.  This is a short-term state, typically lasting only
3094  *    one round trip time (RTT).
3095  *
3096  * Primary.  (tcp_secret_generating == tcp_secret_primary)
3097  *    Used both for generation and primary verification.
3098  *
3099  * Retiring.  (tcp_secret_retiring != tcp_secret_secondary)
3100  *    Used for verification, until the first failure that can be
3101  *    verified by the newer Generating secret.  At that time, this
3102  *    cookie's state is changed to Secondary, and the Generating
3103  *    cookie's state is changed to Primary.  This is a short-term state,
3104  *    typically lasting only one round trip time (RTT).
3105  *
3106  * Secondary.  (tcp_secret_retiring == tcp_secret_secondary)
3107  *    Used for secondary verification, after primary verification
3108  *    failures.  This state lasts no more than twice the Maximum Segment
3109  *    Lifetime (2MSL).  Then, the secret is discarded.
3110  */
3111 struct tcp_cookie_secret {
3112         /* The secret is divided into two parts.  The digest part is the
3113          * equivalent of previously hashing a secret and saving the state,
3114          * and serves as an initialization vector (IV).  The message part
3115          * serves as the trailing secret.
3116          */
3117         u32                             secrets[COOKIE_WORKSPACE_WORDS];
3118         unsigned long                   expires;
3119 };
3120
3121 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3122 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3123 #define TCP_SECRET_LIFE (HZ * 600)
3124
3125 static struct tcp_cookie_secret tcp_secret_one;
3126 static struct tcp_cookie_secret tcp_secret_two;
3127
3128 /* Essentially a circular list, without dynamic allocation. */
3129 static struct tcp_cookie_secret *tcp_secret_generating;
3130 static struct tcp_cookie_secret *tcp_secret_primary;
3131 static struct tcp_cookie_secret *tcp_secret_retiring;
3132 static struct tcp_cookie_secret *tcp_secret_secondary;
3133
3134 static DEFINE_SPINLOCK(tcp_secret_locker);
3135
3136 /* Select a pseudo-random word in the cookie workspace.
3137  */
3138 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3139 {
3140         return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3141 }
3142
3143 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3144  * Called in softirq context.
3145  * Returns: 0 for success.
3146  */
3147 int tcp_cookie_generator(u32 *bakery)
3148 {
3149         unsigned long jiffy = jiffies;
3150
3151         if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3152                 spin_lock_bh(&tcp_secret_locker);
3153                 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3154                         /* refreshed by another */
3155                         memcpy(bakery,
3156                                &tcp_secret_generating->secrets[0],
3157                                COOKIE_WORKSPACE_WORDS);
3158                 } else {
3159                         /* still needs refreshing */
3160                         get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3161
3162                         /* The first time, paranoia assumes that the
3163                          * randomization function isn't as strong.  But,
3164                          * this secret initialization is delayed until
3165                          * the last possible moment (packet arrival).
3166                          * Although that time is observable, it is
3167                          * unpredictably variable.  Mash in the most
3168                          * volatile clock bits available, and expire the
3169                          * secret extra quickly.
3170                          */
3171                         if (unlikely(tcp_secret_primary->expires ==
3172                                      tcp_secret_secondary->expires)) {
3173                                 struct timespec tv;
3174
3175                                 getnstimeofday(&tv);
3176                                 bakery[COOKIE_DIGEST_WORDS+0] ^=
3177                                         (u32)tv.tv_nsec;
3178
3179                                 tcp_secret_secondary->expires = jiffy
3180                                         + TCP_SECRET_1MSL
3181                                         + (0x0f & tcp_cookie_work(bakery, 0));
3182                         } else {
3183                                 tcp_secret_secondary->expires = jiffy
3184                                         + TCP_SECRET_LIFE
3185                                         + (0xff & tcp_cookie_work(bakery, 1));
3186                                 tcp_secret_primary->expires = jiffy
3187                                         + TCP_SECRET_2MSL
3188                                         + (0x1f & tcp_cookie_work(bakery, 2));
3189                         }
3190                         memcpy(&tcp_secret_secondary->secrets[0],
3191                                bakery, COOKIE_WORKSPACE_WORDS);
3192
3193                         rcu_assign_pointer(tcp_secret_generating,
3194                                            tcp_secret_secondary);
3195                         rcu_assign_pointer(tcp_secret_retiring,
3196                                            tcp_secret_primary);
3197                         /*
3198                          * Neither call_rcu() nor synchronize_rcu() needed.
3199                          * Retiring data is not freed.  It is replaced after
3200                          * further (locked) pointer updates, and a quiet time
3201                          * (minimum 1MSL, maximum LIFE - 2MSL).
3202                          */
3203                 }
3204                 spin_unlock_bh(&tcp_secret_locker);
3205         } else {
3206                 rcu_read_lock_bh();
3207                 memcpy(bakery,
3208                        &rcu_dereference(tcp_secret_generating)->secrets[0],
3209                        COOKIE_WORKSPACE_WORDS);
3210                 rcu_read_unlock_bh();
3211         }
3212         return 0;
3213 }
3214 EXPORT_SYMBOL(tcp_cookie_generator);
3215
3216 void tcp_done(struct sock *sk)
3217 {
3218         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3219                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3220
3221         tcp_set_state(sk, TCP_CLOSE);
3222         tcp_clear_xmit_timers(sk);
3223
3224         sk->sk_shutdown = SHUTDOWN_MASK;
3225
3226         if (!sock_flag(sk, SOCK_DEAD))
3227                 sk->sk_state_change(sk);
3228         else
3229                 inet_csk_destroy_sock(sk);
3230 }
3231 EXPORT_SYMBOL_GPL(tcp_done);
3232
3233 extern struct tcp_congestion_ops tcp_reno;
3234
3235 static __initdata unsigned long thash_entries;
3236 static int __init set_thash_entries(char *str)
3237 {
3238         if (!str)
3239                 return 0;
3240         thash_entries = simple_strtoul(str, &str, 0);
3241         return 1;
3242 }
3243 __setup("thash_entries=", set_thash_entries);
3244
3245 void __init tcp_init(void)
3246 {
3247         struct sk_buff *skb = NULL;
3248         unsigned long limit;
3249         int i, max_rshare, max_wshare, cnt;
3250         unsigned long jiffy = jiffies;
3251
3252         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3253
3254         percpu_counter_init(&tcp_sockets_allocated, 0);
3255         percpu_counter_init(&tcp_orphan_count, 0);
3256         tcp_hashinfo.bind_bucket_cachep =
3257                 kmem_cache_create("tcp_bind_bucket",
3258                                   sizeof(struct inet_bind_bucket), 0,
3259                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3260
3261         /* Size and allocate the main established and bind bucket
3262          * hash tables.
3263          *
3264          * The methodology is similar to that of the buffer cache.
3265          */
3266         tcp_hashinfo.ehash =
3267                 alloc_large_system_hash("TCP established",
3268                                         sizeof(struct inet_ehash_bucket),
3269                                         thash_entries,
3270                                         (totalram_pages >= 128 * 1024) ?
3271                                         13 : 15,
3272                                         0,
3273                                         NULL,
3274                                         &tcp_hashinfo.ehash_mask,
3275                                         thash_entries ? 0 : 512 * 1024);
3276         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3277                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3278                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3279         }
3280         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3281                 panic("TCP: failed to alloc ehash_locks");
3282         tcp_hashinfo.bhash =
3283                 alloc_large_system_hash("TCP bind",
3284                                         sizeof(struct inet_bind_hashbucket),
3285                                         tcp_hashinfo.ehash_mask + 1,
3286                                         (totalram_pages >= 128 * 1024) ?
3287                                         13 : 15,
3288                                         0,
3289                                         &tcp_hashinfo.bhash_size,
3290                                         NULL,
3291                                         64 * 1024);
3292         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
3293         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3294                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3295                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3296         }
3297
3298
3299         cnt = tcp_hashinfo.ehash_mask + 1;
3300
3301         tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3302         sysctl_tcp_max_orphans = cnt / 2;
3303         sysctl_max_syn_backlog = max(128, cnt / 256);
3304
3305         limit = nr_free_buffer_pages() / 8;
3306         limit = max(limit, 128UL);
3307         sysctl_tcp_mem[0] = limit / 4 * 3;
3308         sysctl_tcp_mem[1] = limit;
3309         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3310
3311         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3312         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
3313         max_wshare = min(4UL*1024*1024, limit);
3314         max_rshare = min(6UL*1024*1024, limit);
3315
3316         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3317         sysctl_tcp_wmem[1] = 16*1024;
3318         sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3319
3320         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3321         sysctl_tcp_rmem[1] = 87380;
3322         sysctl_tcp_rmem[2] = max(87380, max_rshare);
3323
3324         printk(KERN_INFO "TCP: Hash tables configured "
3325                "(established %u bind %u)\n",
3326                tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3327
3328         tcp_register_congestion_control(&tcp_reno);
3329
3330         memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3331         memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3332         tcp_secret_one.expires = jiffy; /* past due */
3333         tcp_secret_two.expires = jiffy; /* past due */
3334         tcp_secret_generating = &tcp_secret_one;
3335         tcp_secret_primary = &tcp_secret_one;
3336         tcp_secret_retiring = &tcp_secret_two;
3337         tcp_secret_secondary = &tcp_secret_two;
3338 }
3339
3340 static int tcp_is_local(struct net *net, __be32 addr) {
3341         struct rtable *rt;
3342         struct flowi4 fl4 = { .daddr = addr };
3343         rt = ip_route_output_key(net, &fl4);
3344         if (IS_ERR_OR_NULL(rt))
3345                 return 0;
3346         return rt->dst.dev && (rt->dst.dev->flags & IFF_LOOPBACK);
3347 }
3348
3349 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3350 static int tcp_is_local6(struct net *net, struct in6_addr *addr) {
3351         struct rt6_info *rt6 = rt6_lookup(net, addr, addr, 0, 0);
3352         return rt6 && rt6->rt6i_dev && (rt6->rt6i_dev->flags & IFF_LOOPBACK);
3353 }
3354 #endif
3355
3356 /*
3357  * tcp_nuke_addr - destroy all sockets on the given local address
3358  * if local address is the unspecified address (0.0.0.0 or ::), destroy all
3359  * sockets with local addresses that are not configured.
3360  */
3361 int tcp_nuke_addr(struct net *net, struct sockaddr *addr)
3362 {
3363         int family = addr->sa_family;
3364         unsigned int bucket;
3365
3366         struct in_addr *in;
3367 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3368         struct in6_addr *in6;
3369 #endif
3370         if (family == AF_INET) {
3371                 in = &((struct sockaddr_in *)addr)->sin_addr;
3372 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3373         } else if (family == AF_INET6) {
3374                 in6 = &((struct sockaddr_in6 *)addr)->sin6_addr;
3375 #endif
3376         } else {
3377                 return -EAFNOSUPPORT;
3378         }
3379
3380         for (bucket = 0; bucket < tcp_hashinfo.ehash_mask; bucket++) {
3381                 struct hlist_nulls_node *node;
3382                 struct sock *sk;
3383                 spinlock_t *lock = inet_ehash_lockp(&tcp_hashinfo, bucket);
3384
3385 restart:
3386                 spin_lock_bh(lock);
3387                 sk_nulls_for_each(sk, node, &tcp_hashinfo.ehash[bucket].chain) {
3388                         struct inet_sock *inet = inet_sk(sk);
3389
3390                         if (sysctl_ip_dynaddr && sk->sk_state == TCP_SYN_SENT)
3391                                 continue;
3392                         if (sock_flag(sk, SOCK_DEAD))
3393                                 continue;
3394
3395                         if (family == AF_INET) {
3396                                 __be32 s4 = inet->inet_rcv_saddr;
3397                                 if (s4 == LOOPBACK4_IPV6)
3398                                         continue;
3399
3400                                 if (in->s_addr != s4 &&
3401                                     !(in->s_addr == INADDR_ANY &&
3402                                       !tcp_is_local(net, s4)))
3403                                         continue;
3404                         }
3405
3406 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3407                         if (family == AF_INET6) {
3408                                 struct in6_addr *s6;
3409                                 if (!inet->pinet6)
3410                                         continue;
3411
3412                                 s6 = &inet->pinet6->rcv_saddr;
3413                                 if (ipv6_addr_type(s6) == IPV6_ADDR_MAPPED)
3414                                         continue;
3415
3416                                 if (!ipv6_addr_equal(in6, s6) &&
3417                                     !(ipv6_addr_equal(in6, &in6addr_any) &&
3418                                       !tcp_is_local6(net, s6)))
3419                                 continue;
3420                         }
3421 #endif
3422
3423                         sock_hold(sk);
3424                         spin_unlock_bh(lock);
3425
3426                         local_bh_disable();
3427                         bh_lock_sock(sk);
3428                         sk->sk_err = ETIMEDOUT;
3429                         sk->sk_error_report(sk);
3430
3431                         tcp_done(sk);
3432                         bh_unlock_sock(sk);
3433                         local_bh_enable();
3434                         sock_put(sk);
3435
3436                         goto restart;
3437                 }
3438                 spin_unlock_bh(lock);
3439         }
3440
3441         return 0;
3442 }