net/ipv4/tcp.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the  BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              Implementation of the Transmission Control Protocol(TCP).
   7  *
   8  * Authors:     Ross Biro
   9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
  11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
  12  *              Florian La Roche, <flla@stud.uni-sb.de>
  13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
  15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
  16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
  17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  18  *              Jorge Cwik, <jorge@laser.satlink.net>
  19  *
  20  * Fixes:
  21  *              Alan Cox        :       Numerous verify_area() calls
  22  *              Alan Cox        :       Set the ACK bit on a reset
  23  *              Alan Cox        :       Stopped it crashing if it closed while
  24  *                                      sk->inuse=1 and was trying to connect
  25  *                                      (tcp_err()).
  26  *              Alan Cox        :       All icmp error handling was broken
  27  *                                      pointers passed where wrong and the
  28  *                                      socket was looked up backwards. Nobody
  29  *                                      tested any icmp error code obviously.
  30  *              Alan Cox        :       tcp_err() now handled properly. It
  31  *                                      wakes people on errors. poll
  32  *                                      behaves and the icmp error race
  33  *                                      has gone by moving it into sock.c
  34  *              Alan Cox        :       tcp_send_reset() fixed to work for
  35  *                                      everything not just packets for
  36  *                                      unknown sockets.
  37  *              Alan Cox        :       tcp option processing.
  38  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
  39  *                                      syn rule wrong]
  40  *              Herp Rosmanith  :       More reset fixes
  41  *              Alan Cox        :       No longer acks invalid rst frames.
  42  *                                      Acking any kind of RST is right out.
  43  *              Alan Cox        :       Sets an ignore me flag on an rst
  44  *                                      receive otherwise odd bits of prattle
  45  *                                      escape still
  46  *              Alan Cox        :       Fixed another acking RST frame bug.
  47  *                                      Should stop LAN workplace lockups.
  48  *              Alan Cox        :       Some tidyups using the new skb list
  49  *                                      facilities
  50  *              Alan Cox        :       sk->keepopen now seems to work
  51  *              Alan Cox        :       Pulls options out correctly on accepts
  52  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
  53  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
  54  *                                      bit to skb ops.
  55  *              Alan Cox        :       Tidied tcp_data to avoid a potential
  56  *                                      nasty.
  57  *              Alan Cox        :       Added some better commenting, as the
  58  *                                      tcp is hard to follow
  59  *              Alan Cox        :       Removed incorrect check for 20 * psh
  60  *      Michael O'Reilly        :       ack < copied bug fix.
  61  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
  62  *              Alan Cox        :       FIN with no memory -> CRASH
  63  *              Alan Cox        :       Added socket option proto entries.
  64  *                                      Also added awareness of them to accept.
  65  *              Alan Cox        :       Added TCP options (SOL_TCP)
  66  *              Alan Cox        :       Switched wakeup calls to callbacks,
  67  *                                      so the kernel can layer network
  68  *                                      sockets.
  69  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
  70  *              Alan Cox        :       Handle FIN (more) properly (we hope).
  71  *              Alan Cox        :       RST frames sent on unsynchronised
  72  *                                      state ack error.
  73  *              Alan Cox        :       Put in missing check for SYN bit.
  74  *              Alan Cox        :       Added tcp_select_window() aka NET2E
  75  *                                      window non shrink trick.
  76  *              Alan Cox        :       Added a couple of small NET2E timer
  77  *                                      fixes
  78  *              Charles Hedrick :       TCP fixes
  79  *              Toomas Tamm     :       TCP window fixes
  80  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
  81  *              Charles Hedrick :       Rewrote most of it to actually work
  82  *              Linus           :       Rewrote tcp_read() and URG handling
  83  *                                      completely
  84  *              Gerhard Koerting:       Fixed some missing timer handling
  85  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
  86  *              Gerhard Koerting:       PC/TCP workarounds
  87  *              Adam Caldwell   :       Assorted timer/timing errors
  88  *              Matthew Dillon  :       Fixed another RST bug
  89  *              Alan Cox        :       Move to kernel side addressing changes.
  90  *              Alan Cox        :       Beginning work on TCP fastpathing
  91  *                                      (not yet usable)
  92  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
  93  *              Alan Cox        :       TCP fast path debugging
  94  *              Alan Cox        :       Window clamping
  95  *              Michael Riepe   :       Bug in tcp_check()
  96  *              Matt Dillon     :       More TCP improvements and RST bug fixes
  97  *              Matt Dillon     :       Yet more small nasties remove from the
  98  *                                      TCP code (Be very nice to this man if
  99  *                                      tcp finally works 100%) 8)
 100  *              Alan Cox        :       BSD accept semantics.
 101  *              Alan Cox        :       Reset on closedown bug.
 102  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
 103  *              Michael Pall    :       Handle poll() after URG properly in
 104  *                                      all cases.
 105  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
 106  *                                      (multi URG PUSH broke rlogin).
 107  *              Michael Pall    :       Fix the multi URG PUSH problem in
 108  *                                      tcp_readable(), poll() after URG
 109  *                                      works now.
 110  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
 111  *                                      BSD api.
 112  *              Alan Cox        :       Changed the semantics of sk->socket to
 113  *                                      fix a race and a signal problem with
 114  *                                      accept() and async I/O.
 115  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
 116  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
 117  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
 118  *                                      clients/servers which listen in on
 119  *                                      fixed ports.
 120  *              Alan Cox        :       Cleaned the above up and shrank it to
 121  *                                      a sensible code size.
 122  *              Alan Cox        :       Self connect lockup fix.
 123  *              Alan Cox        :       No connect to multicast.
 124  *              Ross Biro       :       Close unaccepted children on master
 125  *                                      socket close.
 126  *              Alan Cox        :       Reset tracing code.
 127  *              Alan Cox        :       Spurious resets on shutdown.
 128  *              Alan Cox        :       Giant 15 minute/60 second timer error
 129  *              Alan Cox        :       Small whoops in polling before an
 130  *                                      accept.
 131  *              Alan Cox        :       Kept the state trace facility since
 132  *                                      it's handy for debugging.
 133  *              Alan Cox        :       More reset handler fixes.
 134  *              Alan Cox        :       Started rewriting the code based on
 135  *                                      the RFC's for other useful protocol
 136  *                                      references see: Comer, KA9Q NOS, and
 137  *                                      for a reference on the difference
 138  *                                      between specifications and how BSD
 139  *                                      works see the 4.4lite source.
 140  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
 141  *                                      close.
 142  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
 143  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
 144  *              Alan Cox        :       Reimplemented timers as per the RFC
 145  *                                      and using multiple timers for sanity.
 146  *              Alan Cox        :       Small bug fixes, and a lot of new
 147  *                                      comments.
 148  *              Alan Cox        :       Fixed dual reader crash by locking
 149  *                                      the buffers (much like datagram.c)
 150  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
 151  *                                      now gets fed up of retrying without
 152  *                                      (even a no space) answer.
 153  *              Alan Cox        :       Extracted closing code better
 154  *              Alan Cox        :       Fixed the closing state machine to
 155  *                                      resemble the RFC.
 156  *              Alan Cox        :       More 'per spec' fixes.
 157  *              Jorge Cwik      :       Even faster checksumming.
 158  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
 159  *                                      only frames. At least one pc tcp stack
 160  *                                      generates them.
 161  *              Alan Cox        :       Cache last socket.
 162  *              Alan Cox        :       Per route irtt.
 163  *              Matt Day        :       poll()->select() match BSD precisely on error
 164  *              Alan Cox        :       New buffers
 165  *              Marc Tamsky     :       Various sk->prot->retransmits and
 166  *                                      sk->retransmits misupdating fixed.
 167  *                                      Fixed tcp_write_timeout: stuck close,
 168  *                                      and TCP syn retries gets used now.
 169  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
 170  *                                      ack if state is TCP_CLOSED.
 171  *              Alan Cox        :       Look up device on a retransmit - routes may
 172  *                                      change. Doesn't yet cope with MSS shrink right
 173  *                                      but it's a start!
 174  *              Marc Tamsky     :       Closing in closing fixes.
 175  *              Mike Shaver     :       RFC1122 verifications.
 176  *              Alan Cox        :       rcv_saddr errors.
 177  *              Alan Cox        :       Block double connect().
 178  *              Alan Cox        :       Small hooks for enSKIP.
 179  *              Alexey Kuznetsov:       Path MTU discovery.
 180  *              Alan Cox        :       Support soft errors.
 181  *              Alan Cox        :       Fix MTU discovery pathological case
 182  *                                      when the remote claims no mtu!
 183  *              Marc Tamsky     :       TCP_CLOSE fix.
 184  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
 185  *                                      window but wrong (fixes NT lpd problems)
 186  *              Pedro Roque     :       Better TCP window handling, delayed ack.
 187  *              Joerg Reuter    :       No modification of locked buffers in
 188  *                                      tcp_do_retransmit()
 189  *              Eric Schenk     :       Changed receiver side silly window
 190  *                                      avoidance algorithm to BSD style
 191  *                                      algorithm. This doubles throughput
 192  *                                      against machines running Solaris,
 193  *                                      and seems to result in general
 194  *                                      improvement.
 195  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
 196  *      Willy Konynenberg       :       Transparent proxying support.
 197  *      Mike McLagan            :       Routing by source
 198  *              Keith Owens     :       Do proper merging with partial SKB's in
 199  *                                      tcp_do_sendmsg to avoid burstiness.
 200  *              Eric Schenk     :       Fix fast close down bug with
 201  *                                      shutdown() followed by close().
 202  *              Andi Kleen      :       Make poll agree with SIGIO
 203  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
 204  *                                      lingertime == 0 (RFC 793 ABORT Call)
 205  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
 206  *                                      csum_and_copy_from_user() if possible.
 207  *
 208  *              This program is free software; you can redistribute it and/or
 209  *              modify it under the terms of the GNU General Public License
 210  *              as published by the Free Software Foundation; either version
 211  *              2 of the License, or(at your option) any later version.
 212  *
 213  * Description of States:
 214  *
 215  *      TCP_SYN_SENT            sent a connection request, waiting for ack
 216  *
 217  *      TCP_SYN_RECV            received a connection request, sent ack,
 218  *                              waiting for final ack in three-way handshake.
 219  *
 220  *      TCP_ESTABLISHED         connection established
 221  *
 222  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
 223  *                              transmission of remaining buffered data
 224  *
 225  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
 226  *                              to shutdown
 227  *
 228  *      TCP_CLOSING             both sides have shutdown but we still have
 229  *                              data we have to finish sending
 230  *
 231  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
 232  *                              closed, can only be entered from FIN_WAIT2
 233  *                              or CLOSING.  Required because the other end
 234  *                              may not have gotten our last ACK causing it
 235  *                              to retransmit the data packet (which we ignore)
 236  *
 237  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
 238  *                              us to finish writing our data and to shutdown
 239  *                              (we have to close() to move on to LAST_ACK)
 240  *
 241  *      TCP_LAST_ACK            out side has shutdown after remote has
 242  *                              shutdown.  There may still be data in our
 243  *                              buffer that we have to finish sending
 244  *
 245  *      TCP_CLOSE               socket is finished
 246  */
 247
 248 #define pr_fmt(fmt) "TCP: " fmt
 249
 250 #include <linux/kernel.h>
 251 #include <linux/module.h>
 252 #include <linux/types.h>
 253 #include <linux/fcntl.h>
 254 #include <linux/poll.h>
 255 #include <linux/init.h>
 256 #include <linux/fs.h>
 257 #include <linux/skbuff.h>
 258 #include <linux/scatterlist.h>
 259 #include <linux/splice.h>
 260 #include <linux/net.h>
 261 #include <linux/socket.h>
 262 #include <linux/random.h>
 263 #include <linux/bootmem.h>
 264 #include <linux/highmem.h>
 265 #include <linux/swap.h>
 266 #include <linux/cache.h>
 267 #include <linux/err.h>
 268 #include <linux/crypto.h>
 269 #include <linux/time.h>
 270 #include <linux/slab.h>
 271
 272 #include <net/icmp.h>
 273 #include <net/tcp.h>
 274 #include <net/xfrm.h>
 275 #include <net/ip.h>
 276 #include <net/netdma.h>
 277 #include <net/sock.h>
 278
 279 #include <asm/uaccess.h>
 280 #include <asm/ioctls.h>
 281
 282 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
 283
 284 struct percpu_counter tcp_orphan_count;
 285 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 286
 287 int sysctl_tcp_wmem[3] __read_mostly;
 288 int sysctl_tcp_rmem[3] __read_mostly;
 289
 290 EXPORT_SYMBOL(sysctl_tcp_rmem);
 291 EXPORT_SYMBOL(sysctl_tcp_wmem);
 292
 293 atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
 294 EXPORT_SYMBOL(tcp_memory_allocated);
 295
 296 /*
 297  * Current number of TCP sockets.
 298  */
 299 struct percpu_counter tcp_sockets_allocated;
 300 EXPORT_SYMBOL(tcp_sockets_allocated);
 301
 302 /*
 303  * TCP splice context
 304  */
 305 struct tcp_splice_state {
 306         struct pipe_inode_info *pipe;
 307         size_t len;
 308         unsigned int flags;
 309 };
 310
 311 /*
 312  * Pressure flag: try to collapse.
 313  * Technical note: it is used by multiple contexts non atomically.
 314  * All the __sk_mem_schedule() is of this nature: accounting
 315  * is strict, actions are advisory and have some latency.
 316  */
 317 int tcp_memory_pressure __read_mostly;
 318 EXPORT_SYMBOL(tcp_memory_pressure);
 319
 320 void tcp_enter_memory_pressure(struct sock *sk)
 321 {
 322         if (!tcp_memory_pressure) {
 323                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
 324                 tcp_memory_pressure = 1;
 325         }
 326 }
 327 EXPORT_SYMBOL(tcp_enter_memory_pressure);
 328
 329 /* Convert seconds to retransmits based on initial and max timeout */
 330 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
 331 {
 332         u8 res = 0;
 333
 334         if (seconds > 0) {
 335                 int period = timeout;
 336
 337                 res = 1;
 338                 while (seconds > period && res < 255) {
 339                         res++;
 340                         timeout <<= 1;
 341                         if (timeout > rto_max)
 342                                 timeout = rto_max;
 343                         period += timeout;
 344                 }
 345         }
 346         return res;
 347 }
 348
 349 /* Convert retransmits to seconds based on initial and max timeout */
 350 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
 351 {
 352         int period = 0;
 353
 354         if (retrans > 0) {
 355                 period = timeout;
 356                 while (--retrans) {
 357                         timeout <<= 1;
 358                         if (timeout > rto_max)
 359                                 timeout = rto_max;
 360                         period += timeout;
 361                 }
 362         }
 363         return period;
 364 }
 365
 366 /*
 367  *      Wait for a TCP event.
 368  *
 369  *      Note that we don't need to lock the socket, as the upper poll layers
 370  *      take care of normal races (between the test and the event) and we don't
 371  *      go look at any of the socket buffers directly.
 372  */
 373 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 374 {
 375         unsigned int mask;
 376         struct sock *sk = sock->sk;
 377         const struct tcp_sock *tp = tcp_sk(sk);
 378
 379         sock_poll_wait(file, sk_sleep(sk), wait);
 380         if (sk->sk_state == TCP_LISTEN)
 381                 return inet_csk_listen_poll(sk);
 382
 383         /* Socket is not locked. We are protected from async events
 384          * by poll logic and correct handling of state changes
 385          * made by other threads is impossible in any case.
 386          */
 387
 388         mask = 0;
 389
 390         /*
 391          * POLLHUP is certainly not done right. But poll() doesn't
 392          * have a notion of HUP in just one direction, and for a
 393          * socket the read side is more interesting.
 394          *
 395          * Some poll() documentation says that POLLHUP is incompatible
 396          * with the POLLOUT/POLLWR flags, so somebody should check this
 397          * all. But careful, it tends to be safer to return too many
 398          * bits than too few, and you can easily break real applications
 399          * if you don't tell them that something has hung up!
 400          *
 401          * Check-me.
 402          *
 403          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
 404          * our fs/select.c). It means that after we received EOF,
 405          * poll always returns immediately, making impossible poll() on write()
 406          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
 407          * if and only if shutdown has been made in both directions.
 408          * Actually, it is interesting to look how Solaris and DUX
 409          * solve this dilemma. I would prefer, if POLLHUP were maskable,
 410          * then we could set it on SND_SHUTDOWN. BTW examples given
 411          * in Stevens' books assume exactly this behaviour, it explains
 412          * why POLLHUP is incompatible with POLLOUT.    --ANK
 413          *
 414          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 415          * blocking on fresh not-connected or disconnected socket. --ANK
 416          */
 417         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
 418                 mask |= POLLHUP;
 419         if (sk->sk_shutdown & RCV_SHUTDOWN)
 420                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
 421
 422         /* Connected? */
 423         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
 424                 int target = sock_rcvlowat(sk, 0, INT_MAX);
 425
 426                 if (tp->urg_seq == tp->copied_seq &&
 427                     !sock_flag(sk, SOCK_URGINLINE) &&
 428                     tp->urg_data)
 429                         target++;
 430
 431                 /* Potential race condition. If read of tp below will
 432                  * escape above sk->sk_state, we can be illegally awaken
 433                  * in SYN_* states. */
 434                 if (tp->rcv_nxt - tp->copied_seq >= target)
 435                         mask |= POLLIN | POLLRDNORM;
 436
 437                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 438                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
 439                                 mask |= POLLOUT | POLLWRNORM;
 440                         } else {  /* send SIGIO later */
 441                                 set_bit(SOCK_ASYNC_NOSPACE,
 442                                         &sk->sk_socket->flags);
 443                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 444
 445                                 /* Race breaker. If space is freed after
 446                                  * wspace test but before the flags are set,
 447                                  * IO signal will be lost.
 448                                  */
 449                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
 450                                         mask |= POLLOUT | POLLWRNORM;
 451                         }
 452                 } else
 453                         mask |= POLLOUT | POLLWRNORM;
 454
 455                 if (tp->urg_data & TCP_URG_VALID)
 456                         mask |= POLLPRI;
 457         }
 458         /* This barrier is coupled with smp_wmb() in tcp_reset() */
 459         smp_rmb();
 460         if (sk->sk_err)
 461                 mask |= POLLERR;
 462
 463         return mask;
 464 }
 465 EXPORT_SYMBOL(tcp_poll);
 466
 467 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 468 {
 469         struct tcp_sock *tp = tcp_sk(sk);
 470         int answ;
 471
 472         switch (cmd) {
 473         case SIOCINQ:
 474                 if (sk->sk_state == TCP_LISTEN)
 475                         return -EINVAL;
 476
 477                 lock_sock(sk);
 478                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 479                         answ = 0;
 480                 else if (sock_flag(sk, SOCK_URGINLINE) ||
 481                          !tp->urg_data ||
 482                          before(tp->urg_seq, tp->copied_seq) ||
 483                          !before(tp->urg_seq, tp->rcv_nxt)) {
 484                         struct sk_buff *skb;
 485
 486                         answ = tp->rcv_nxt - tp->copied_seq;
 487
 488                         /* Subtract 1, if FIN is in queue. */
 489                         skb = skb_peek_tail(&sk->sk_receive_queue);
 490                         if (answ && skb)
 491                                 answ -= tcp_hdr(skb)->fin;
 492                 } else
 493                         answ = tp->urg_seq - tp->copied_seq;
 494                 release_sock(sk);
 495                 break;
 496         case SIOCATMARK:
 497                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 498                 break;
 499         case SIOCOUTQ:
 500                 if (sk->sk_state == TCP_LISTEN)
 501                         return -EINVAL;
 502
 503                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 504                         answ = 0;
 505                 else
 506                         answ = tp->write_seq - tp->snd_una;
 507                 break;
 508         case SIOCOUTQNSD:
 509                 if (sk->sk_state == TCP_LISTEN)
 510                         return -EINVAL;
 511
 512                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 513                         answ = 0;
 514                 else
 515                         answ = tp->write_seq - tp->snd_nxt;
 516                 break;
 517         default:
 518                 return -ENOIOCTLCMD;
 519         }
 520
 521         return put_user(answ, (int __user *)arg);
 522 }
 523 EXPORT_SYMBOL(tcp_ioctl);
 524
 525 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 526 {
 527         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 528         tp->pushed_seq = tp->write_seq;
 529 }
 530
 531 static inline int forced_push(const struct tcp_sock *tp)
 532 {
 533         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 534 }
 535
 536 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
 537 {
 538         struct tcp_sock *tp = tcp_sk(sk);
 539         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 540
 541         skb->csum    = 0;
 542         tcb->seq     = tcb->end_seq = tp->write_seq;
 543         tcb->tcp_flags = TCPHDR_ACK;
 544         tcb->sacked  = 0;
 545         skb_header_release(skb);
 546         tcp_add_write_queue_tail(sk, skb);
 547         sk->sk_wmem_queued += skb->truesize;
 548         sk_mem_charge(sk, skb->truesize);
 549         if (tp->nonagle & TCP_NAGLE_PUSH)
 550                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 551 }
 552
 553 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
 554 {
 555         if (flags & MSG_OOB)
 556                 tp->snd_up = tp->write_seq;
 557 }
 558
 559 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
 560                             int nonagle)
 561 {
 562         if (tcp_send_head(sk)) {
 563                 struct tcp_sock *tp = tcp_sk(sk);
 564
 565                 if (!(flags & MSG_MORE) || forced_push(tp))
 566                         tcp_mark_push(tp, tcp_write_queue_tail(sk));
 567
 568                 tcp_mark_urg(tp, flags);
 569                 __tcp_push_pending_frames(sk, mss_now,
 570                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
 571         }
 572 }
 573
 574 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 575                                 unsigned int offset, size_t len)
 576 {
 577         struct tcp_splice_state *tss = rd_desc->arg.data;
 578         int ret;
 579
 580         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
 581                               tss->flags);
 582         if (ret > 0)
 583                 rd_desc->count -= ret;
 584         return ret;
 585 }
 586
 587 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 588 {
 589         /* Store TCP splice context information in read_descriptor_t. */
 590         read_descriptor_t rd_desc = {
 591                 .arg.data = tss,
 592                 .count    = tss->len,
 593         };
 594
 595         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 596 }
 597
 598 /**
 599  *  tcp_splice_read - splice data from TCP socket to a pipe
 600  * @sock:       socket to splice from
 601  * @ppos:       position (not valid)
 602  * @pipe:       pipe to splice to
 603  * @len:        number of bytes to splice
 604  * @flags:      splice modifier flags
 605  *
 606  * Description:
 607  *    Will read pages from given socket and fill them into a pipe.
 608  *
 609  **/
 610 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 611                         struct pipe_inode_info *pipe, size_t len,
 612                         unsigned int flags)
 613 {
 614         struct sock *sk = sock->sk;
 615         struct tcp_splice_state tss = {
 616                 .pipe = pipe,
 617                 .len = len,
 618                 .flags = flags,
 619         };
 620         long timeo;
 621         ssize_t spliced;
 622         int ret;
 623
 624         sock_rps_record_flow(sk);
 625         /*
 626          * We can't seek on a socket input
 627          */
 628         if (unlikely(*ppos))
 629                 return -ESPIPE;
 630
 631         ret = spliced = 0;
 632
 633         lock_sock(sk);
 634
 635         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 636         while (tss.len) {
 637                 ret = __tcp_splice_read(sk, &tss);
 638                 if (ret < 0)
 639                         break;
 640                 else if (!ret) {
 641                         if (spliced)
 642                                 break;
 643                         if (sock_flag(sk, SOCK_DONE))
 644                                 break;
 645                         if (sk->sk_err) {
 646                                 ret = sock_error(sk);
 647                                 break;
 648                         }
 649                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 650                                 break;
 651                         if (sk->sk_state == TCP_CLOSE) {
 652                                 /*
 653                                  * This occurs when user tries to read
 654                                  * from never connected socket.
 655                                  */
 656                                 if (!sock_flag(sk, SOCK_DONE))
 657                                         ret = -ENOTCONN;
 658                                 break;
 659                         }
 660                         if (!timeo) {
 661                                 ret = -EAGAIN;
 662                                 break;
 663                         }
 664                         sk_wait_data(sk, &timeo);
 665                         if (signal_pending(current)) {
 666                                 ret = sock_intr_errno(timeo);
 667                                 break;
 668                         }
 669                         continue;
 670                 }
 671                 tss.len -= ret;
 672                 spliced += ret;
 673
 674                 if (!timeo)
 675                         break;
 676                 release_sock(sk);
 677                 lock_sock(sk);
 678
 679                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 680                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 681                     signal_pending(current))
 682                         break;
 683         }
 684
 685         release_sock(sk);
 686
 687         if (spliced)
 688                 return spliced;
 689
 690         return ret;
 691 }
 692 EXPORT_SYMBOL(tcp_splice_read);
 693
 694 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
 695 {
 696         struct sk_buff *skb;
 697
 698         /* The TCP header must be at least 32-bit aligned.  */
 699         size = ALIGN(size, 4);
 700
 701         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 702         if (skb) {
 703                 if (sk_wmem_schedule(sk, skb->truesize)) {
 704                         skb_reserve(skb, sk->sk_prot->max_header);
 705                         /*
 706                          * Make sure that we have exactly size bytes
 707                          * available to the caller, no more, no less.
 708                          */
 709                         skb->avail_size = size;
 710                         return skb;
 711                 }
 712                 __kfree_skb(skb);
 713         } else {
 714                 sk->sk_prot->enter_memory_pressure(sk);
 715                 sk_stream_moderate_sndbuf(sk);
 716         }
 717         return NULL;
 718 }
 719
 720 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 721                                        int large_allowed)
 722 {
 723         struct tcp_sock *tp = tcp_sk(sk);
 724         u32 xmit_size_goal, old_size_goal;
 725
 726         xmit_size_goal = mss_now;
 727
 728         if (large_allowed && sk_can_gso(sk)) {
 729                 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
 730                                   inet_csk(sk)->icsk_af_ops->net_header_len -
 731                                   inet_csk(sk)->icsk_ext_hdr_len -
 732                                   tp->tcp_header_len);
 733
 734                 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
 735
 736                 /* We try hard to avoid divides here */
 737                 old_size_goal = tp->xmit_size_goal_segs * mss_now;
 738
 739                 if (likely(old_size_goal <= xmit_size_goal &&
 740                            old_size_goal + mss_now > xmit_size_goal)) {
 741                         xmit_size_goal = old_size_goal;
 742                 } else {
 743                         tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
 744                         xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
 745                 }
 746         }
 747
 748         return max(xmit_size_goal, mss_now);
 749 }
 750
 751 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 752 {
 753         int mss_now;
 754
 755         mss_now = tcp_current_mss(sk);
 756         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 757
 758         return mss_now;
 759 }
 760
 761 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
 762                          size_t psize, int flags)
 763 {
 764         struct tcp_sock *tp = tcp_sk(sk);
 765         int mss_now, size_goal;
 766         int err;
 767         ssize_t copied;
 768         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 769
 770         /* Wait for a connection to finish. */
 771         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 772                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 773                         goto out_err;
 774
 775         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 776
 777         mss_now = tcp_send_mss(sk, &size_goal, flags);
 778         copied = 0;
 779
 780         err = -EPIPE;
 781         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 782                 goto out_err;
 783
 784         while (psize > 0) {
 785                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 786                 struct page *page = pages[poffset / PAGE_SIZE];
 787                 int copy, i, can_coalesce;
 788                 int offset = poffset % PAGE_SIZE;
 789                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
 790
 791                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
 792 new_segment:
 793                         if (!sk_stream_memory_free(sk))
 794                                 goto wait_for_sndbuf;
 795
 796                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
 797                         if (!skb)
 798                                 goto wait_for_memory;
 799
 800                         skb_entail(sk, skb);
 801                         copy = size_goal;
 802                 }
 803
 804                 if (copy > size)
 805                         copy = size;
 806
 807                 i = skb_shinfo(skb)->nr_frags;
 808                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 809                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
 810                         tcp_mark_push(tp, skb);
 811                         goto new_segment;
 812                 }
 813                 if (!sk_wmem_schedule(sk, copy))
 814                         goto wait_for_memory;
 815
 816                 if (can_coalesce) {
 817                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
 818                 } else {
 819                         get_page(page);
 820                         skb_fill_page_desc(skb, i, page, offset, copy);
 821                 }
 822
 823                 skb->len += copy;
 824                 skb->data_len += copy;
 825                 skb->truesize += copy;
 826                 sk->sk_wmem_queued += copy;
 827                 sk_mem_charge(sk, copy);
 828                 skb->ip_summed = CHECKSUM_PARTIAL;
 829                 tp->write_seq += copy;
 830                 TCP_SKB_CB(skb)->end_seq += copy;
 831                 skb_shinfo(skb)->gso_segs = 0;
 832
 833                 if (!copied)
 834                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
 835
 836                 copied += copy;
 837                 poffset += copy;
 838                 if (!(psize -= copy))
 839                         goto out;
 840
 841                 if (skb->len < size_goal || (flags & MSG_OOB))
 842                         continue;
 843
 844                 if (forced_push(tp)) {
 845                         tcp_mark_push(tp, skb);
 846                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
 847                 } else if (skb == tcp_send_head(sk))
 848                         tcp_push_one(sk, mss_now);
 849                 continue;
 850
 851 wait_for_sndbuf:
 852                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 853 wait_for_memory:
 854                 if (copied)
 855                         tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
 856
 857                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 858                         goto do_error;
 859
 860                 mss_now = tcp_send_mss(sk, &size_goal, flags);
 861         }
 862
 863 out:
 864         if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
 865                 tcp_push(sk, flags, mss_now, tp->nonagle);
 866         return copied;
 867
 868 do_error:
 869         if (copied)
 870                 goto out;
 871 out_err:
 872         return sk_stream_error(sk, flags, err);
 873 }
 874
 875 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
 876                  size_t size, int flags)
 877 {
 878         ssize_t res;
 879
 880         if (!(sk->sk_route_caps & NETIF_F_SG) ||
 881             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
 882                 return sock_no_sendpage(sk->sk_socket, page, offset, size,
 883                                         flags);
 884
 885         lock_sock(sk);
 886         res = do_tcp_sendpages(sk, &page, offset, size, flags);
 887         release_sock(sk);
 888         return res;
 889 }
 890 EXPORT_SYMBOL(tcp_sendpage);
 891
 892 static inline int select_size(const struct sock *sk, bool sg)
 893 {
 894         const struct tcp_sock *tp = tcp_sk(sk);
 895         int tmp = tp->mss_cache;
 896
 897         if (sg) {
 898                 if (sk_can_gso(sk)) {
 899                         /* Small frames wont use a full page:
 900                          * Payload will immediately follow tcp header.
 901                          */
 902                         tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
 903                 } else {
 904                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
 905
 906                         if (tmp >= pgbreak &&
 907                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
 908                                 tmp = pgbreak;
 909                 }
 910         }
 911
 912         return tmp;
 913 }
 914
 915 static int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
 916 {
 917         struct sk_buff *skb;
 918         struct tcp_skb_cb *cb;
 919         struct tcphdr *th;
 920
 921         skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
 922         if (!skb)
 923                 goto err;
 924
 925         th = (struct tcphdr *)skb_put(skb, sizeof(*th));
 926         skb_reset_transport_header(skb);
 927         memset(th, 0, sizeof(*th));
 928
 929         if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
 930                 goto err_free;
 931
 932         cb = TCP_SKB_CB(skb);
 933
 934         TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
 935         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
 936         TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
 937
 938         tcp_queue_rcv(sk, skb, sizeof(*th));
 939
 940         return size;
 941
 942 err_free:
 943         kfree_skb(skb);
 944 err:
 945         return -ENOMEM;
 946 }
 947
 948 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
 949                 size_t size)
 950 {
 951         struct iovec *iov;
 952         struct tcp_sock *tp = tcp_sk(sk);
 953         struct sk_buff *skb;
 954         int iovlen, flags, err, copied;
 955         int mss_now = 0, size_goal;
 956         bool sg;
 957         long timeo;
 958
 959         lock_sock(sk);
 960
 961         flags = msg->msg_flags;
 962         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 963
 964         /* Wait for a connection to finish. */
 965         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 966                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 967                         goto out_err;
 968
 969         if (unlikely(tp->repair)) {
 970                 if (tp->repair_queue == TCP_RECV_QUEUE) {
 971                         copied = tcp_send_rcvq(sk, msg, size);
 972                         goto out;
 973                 }
 974
 975                 err = -EINVAL;
 976                 if (tp->repair_queue == TCP_NO_QUEUE)
 977                         goto out_err;
 978
 979                 /* 'common' sending to sendq */
 980         }
 981
 982         /* This should be in poll */
 983         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 984
 985         mss_now = tcp_send_mss(sk, &size_goal, flags);
 986
 987         /* Ok commence sending. */
 988         iovlen = msg->msg_iovlen;
 989         iov = msg->msg_iov;
 990         copied = 0;
 991
 992         err = -EPIPE;
 993         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 994                 goto out_err;
 995
 996         sg = !!(sk->sk_route_caps & NETIF_F_SG);
 997
 998         while (--iovlen >= 0) {
 999                 size_t seglen = iov->iov_len;
1000                 unsigned char __user *from = iov->iov_base;
1001
1002                 iov++;
1003
1004                 while (seglen > 0) {
1005                         int copy = 0;
1006                         int max = size_goal;
1007
1008                         skb = tcp_write_queue_tail(sk);
1009                         if (tcp_send_head(sk)) {
1010                                 if (skb->ip_summed == CHECKSUM_NONE)
1011                                         max = mss_now;
1012                                 copy = max - skb->len;
1013                         }
1014
1015                         if (copy <= 0) {
1016 new_segment:
1017                                 /* Allocate new segment. If the interface is SG,
1018                                  * allocate skb fitting to single page.
1019                                  */
1020                                 if (!sk_stream_memory_free(sk))
1021                                         goto wait_for_sndbuf;
1022
1023                                 skb = sk_stream_alloc_skb(sk,
1024                                                           select_size(sk, sg),
1025                                                           sk->sk_allocation);
1026                                 if (!skb)
1027                                         goto wait_for_memory;
1028
1029                                 /*
1030                                  * Check whether we can use HW checksum.
1031                                  */
1032                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1033                                         skb->ip_summed = CHECKSUM_PARTIAL;
1034
1035                                 skb_entail(sk, skb);
1036                                 copy = size_goal;
1037                                 max = size_goal;
1038                         }
1039
1040                         /* Try to append data to the end of skb. */
1041                         if (copy > seglen)
1042                                 copy = seglen;
1043
1044                         /* Where to copy to? */
1045                         if (skb_availroom(skb) > 0) {
1046                                 /* We have some space in skb head. Superb! */
1047                                 copy = min_t(int, copy, skb_availroom(skb));
1048                                 err = skb_add_data_nocache(sk, skb, from, copy);
1049                                 if (err)
1050                                         goto do_fault;
1051                         } else {
1052                                 int merge = 0;
1053                                 int i = skb_shinfo(skb)->nr_frags;
1054                                 struct page *page = sk->sk_sndmsg_page;
1055                                 int off;
1056
1057                                 if (page && page_count(page) == 1)
1058                                         sk->sk_sndmsg_off = 0;
1059
1060                                 off = sk->sk_sndmsg_off;
1061
1062                                 if (skb_can_coalesce(skb, i, page, off) &&
1063                                     off != PAGE_SIZE) {
1064                                         /* We can extend the last page
1065                                          * fragment. */
1066                                         merge = 1;
1067                                 } else if (i == MAX_SKB_FRAGS || !sg) {
1068                                         /* Need to add new fragment and cannot
1069                                          * do this because interface is non-SG,
1070                                          * or because all the page slots are
1071                                          * busy. */
1072                                         tcp_mark_push(tp, skb);
1073                                         goto new_segment;
1074                                 } else if (page) {
1075                                         if (off == PAGE_SIZE) {
1076                                                 put_page(page);
1077                                                 sk->sk_sndmsg_page = page = NULL;
1078                                                 off = 0;
1079                                         }
1080                                 } else
1081                                         off = 0;
1082
1083                                 if (copy > PAGE_SIZE - off)
1084                                         copy = PAGE_SIZE - off;
1085
1086                                 if (!sk_wmem_schedule(sk, copy))
1087                                         goto wait_for_memory;
1088
1089                                 if (!page) {
1090                                         /* Allocate new cache page. */
1091                                         if (!(page = sk_stream_alloc_page(sk)))
1092                                                 goto wait_for_memory;
1093                                 }
1094
1095                                 /* Time to copy data. We are close to
1096                                  * the end! */
1097                                 err = skb_copy_to_page_nocache(sk, from, skb,
1098                                                                page, off, copy);
1099                                 if (err) {
1100                                         /* If this page was new, give it to the
1101                                          * socket so it does not get leaked.
1102                                          */
1103                                         if (!sk->sk_sndmsg_page) {
1104                                                 sk->sk_sndmsg_page = page;
1105                                                 sk->sk_sndmsg_off = 0;
1106                                         }
1107                                         goto do_error;
1108                                 }
1109
1110                                 /* Update the skb. */
1111                                 if (merge) {
1112                                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1113                                 } else {
1114                                         skb_fill_page_desc(skb, i, page, off, copy);
1115                                         if (sk->sk_sndmsg_page) {
1116                                                 get_page(page);
1117                                         } else if (off + copy < PAGE_SIZE) {
1118                                                 get_page(page);
1119                                                 sk->sk_sndmsg_page = page;
1120                                         }
1121                                 }
1122
1123                                 sk->sk_sndmsg_off = off + copy;
1124                         }
1125
1126                         if (!copied)
1127                                 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1128
1129                         tp->write_seq += copy;
1130                         TCP_SKB_CB(skb)->end_seq += copy;
1131                         skb_shinfo(skb)->gso_segs = 0;
1132
1133                         from += copy;
1134                         copied += copy;
1135                         if ((seglen -= copy) == 0 && iovlen == 0)
1136                                 goto out;
1137
1138                         if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1139                                 continue;
1140
1141                         if (forced_push(tp)) {
1142                                 tcp_mark_push(tp, skb);
1143                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1144                         } else if (skb == tcp_send_head(sk))
1145                                 tcp_push_one(sk, mss_now);
1146                         continue;
1147
1148 wait_for_sndbuf:
1149                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1150 wait_for_memory:
1151                         if (copied && likely(!tp->repair))
1152                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1153
1154                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1155                                 goto do_error;
1156
1157                         mss_now = tcp_send_mss(sk, &size_goal, flags);
1158                 }
1159         }
1160
1161 out:
1162         if (copied && likely(!tp->repair))
1163                 tcp_push(sk, flags, mss_now, tp->nonagle);
1164         release_sock(sk);
1165         return copied;
1166
1167 do_fault:
1168         if (!skb->len) {
1169                 tcp_unlink_write_queue(skb, sk);
1170                 /* It is the one place in all of TCP, except connection
1171                  * reset, where we can be unlinking the send_head.
1172                  */
1173                 tcp_check_send_head(sk, skb);
1174                 sk_wmem_free_skb(sk, skb);
1175         }
1176
1177 do_error:
1178         if (copied)
1179                 goto out;
1180 out_err:
1181         err = sk_stream_error(sk, flags, err);
1182         release_sock(sk);
1183         return err;
1184 }
1185 EXPORT_SYMBOL(tcp_sendmsg);
1186
1187 /*
1188  *      Handle reading urgent data. BSD has very simple semantics for
1189  *      this, no blocking and very strange errors 8)
1190  */
1191
1192 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1193 {
1194         struct tcp_sock *tp = tcp_sk(sk);
1195
1196         /* No URG data to read. */
1197         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1198             tp->urg_data == TCP_URG_READ)
1199                 return -EINVAL; /* Yes this is right ! */
1200
1201         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1202                 return -ENOTCONN;
1203
1204         if (tp->urg_data & TCP_URG_VALID) {
1205                 int err = 0;
1206                 char c = tp->urg_data;
1207
1208                 if (!(flags & MSG_PEEK))
1209                         tp->urg_data = TCP_URG_READ;
1210
1211                 /* Read urgent data. */
1212                 msg->msg_flags |= MSG_OOB;
1213
1214                 if (len > 0) {
1215                         if (!(flags & MSG_TRUNC))
1216                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1217                         len = 1;
1218                 } else
1219                         msg->msg_flags |= MSG_TRUNC;
1220
1221                 return err ? -EFAULT : len;
1222         }
1223
1224         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1225                 return 0;
1226
1227         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1228          * the available implementations agree in this case:
1229          * this call should never block, independent of the
1230          * blocking state of the socket.
1231          * Mike <pall@rz.uni-karlsruhe.de>
1232          */
1233         return -EAGAIN;
1234 }
1235
1236 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1237 {
1238         struct sk_buff *skb;
1239         int copied = 0, err = 0;
1240
1241         /* XXX -- need to support SO_PEEK_OFF */
1242
1243         skb_queue_walk(&sk->sk_write_queue, skb) {
1244                 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1245                 if (err)
1246                         break;
1247
1248                 copied += skb->len;
1249         }
1250
1251         return err ?: copied;
1252 }
1253
1254 /* Clean up the receive buffer for full frames taken by the user,
1255  * then send an ACK if necessary.  COPIED is the number of bytes
1256  * tcp_recvmsg has given to the user so far, it speeds up the
1257  * calculation of whether or not we must ACK for the sake of
1258  * a window update.
1259  */
1260 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1261 {
1262         struct tcp_sock *tp = tcp_sk(sk);
1263         int time_to_ack = 0;
1264
1265         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1266
1267         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1268              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1269              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1270
1271         if (inet_csk_ack_scheduled(sk)) {
1272                 const struct inet_connection_sock *icsk = inet_csk(sk);
1273                    /* Delayed ACKs frequently hit locked sockets during bulk
1274                     * receive. */
1275                 if (icsk->icsk_ack.blocked ||
1276                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1277                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1278                     /*
1279                      * If this read emptied read buffer, we send ACK, if
1280                      * connection is not bidirectional, user drained
1281                      * receive buffer and there was a small segment
1282                      * in queue.
1283                      */
1284                     (copied > 0 &&
1285                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1286                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1287                        !icsk->icsk_ack.pingpong)) &&
1288                       !atomic_read(&sk->sk_rmem_alloc)))
1289                         time_to_ack = 1;
1290         }
1291
1292         /* We send an ACK if we can now advertise a non-zero window
1293          * which has been raised "significantly".
1294          *
1295          * Even if window raised up to infinity, do not send window open ACK
1296          * in states, where we will not receive more. It is useless.
1297          */
1298         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1299                 __u32 rcv_window_now = tcp_receive_window(tp);
1300
1301                 /* Optimize, __tcp_select_window() is not cheap. */
1302                 if (2*rcv_window_now <= tp->window_clamp) {
1303                         __u32 new_window = __tcp_select_window(sk);
1304
1305                         /* Send ACK now, if this read freed lots of space
1306                          * in our buffer. Certainly, new_window is new window.
1307                          * We can advertise it now, if it is not less than current one.
1308                          * "Lots" means "at least twice" here.
1309                          */
1310                         if (new_window && new_window >= 2 * rcv_window_now)
1311                                 time_to_ack = 1;
1312                 }
1313         }
1314         if (time_to_ack)
1315                 tcp_send_ack(sk);
1316 }
1317
1318 static void tcp_prequeue_process(struct sock *sk)
1319 {
1320         struct sk_buff *skb;
1321         struct tcp_sock *tp = tcp_sk(sk);
1322
1323         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1324
1325         /* RX process wants to run with disabled BHs, though it is not
1326          * necessary */
1327         local_bh_disable();
1328         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1329                 sk_backlog_rcv(sk, skb);
1330         local_bh_enable();
1331
1332         /* Clear memory counter. */
1333         tp->ucopy.memory = 0;
1334 }
1335
1336 #ifdef CONFIG_NET_DMA
1337 static void tcp_service_net_dma(struct sock *sk, bool wait)
1338 {
1339         dma_cookie_t done, used;
1340         dma_cookie_t last_issued;
1341         struct tcp_sock *tp = tcp_sk(sk);
1342
1343         if (!tp->ucopy.dma_chan)
1344                 return;
1345
1346         last_issued = tp->ucopy.dma_cookie;
1347         dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1348
1349         do {
1350                 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1351                                               last_issued, &done,
1352                                               &used) == DMA_SUCCESS) {
1353                         /* Safe to free early-copied skbs now */
1354                         __skb_queue_purge(&sk->sk_async_wait_queue);
1355                         break;
1356                 } else {
1357                         struct sk_buff *skb;
1358                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1359                                (dma_async_is_complete(skb->dma_cookie, done,
1360                                                       used) == DMA_SUCCESS)) {
1361                                 __skb_dequeue(&sk->sk_async_wait_queue);
1362                                 kfree_skb(skb);
1363                         }
1364                 }
1365         } while (wait);
1366 }
1367 #endif
1368
1369 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1370 {
1371         struct sk_buff *skb;
1372         u32 offset;
1373
1374         skb_queue_walk(&sk->sk_receive_queue, skb) {
1375                 offset = seq - TCP_SKB_CB(skb)->seq;
1376                 if (tcp_hdr(skb)->syn)
1377                         offset--;
1378                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1379                         *off = offset;
1380                         return skb;
1381                 }
1382         }
1383         return NULL;
1384 }
1385
1386 /*
1387  * This routine provides an alternative to tcp_recvmsg() for routines
1388  * that would like to handle copying from skbuffs directly in 'sendfile'
1389  * fashion.
1390  * Note:
1391  *      - It is assumed that the socket was locked by the caller.
1392  *      - The routine does not block.
1393  *      - At present, there is no support for reading OOB data
1394  *        or for 'peeking' the socket using this routine
1395  *        (although both would be easy to implement).
1396  */
1397 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1398                   sk_read_actor_t recv_actor)
1399 {
1400         struct sk_buff *skb;
1401         struct tcp_sock *tp = tcp_sk(sk);
1402         u32 seq = tp->copied_seq;
1403         u32 offset;
1404         int copied = 0;
1405
1406         if (sk->sk_state == TCP_LISTEN)
1407                 return -ENOTCONN;
1408         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1409                 if (offset < skb->len) {
1410                         int used;
1411                         size_t len;
1412
1413                         len = skb->len - offset;
1414                         /* Stop reading if we hit a patch of urgent data */
1415                         if (tp->urg_data) {
1416                                 u32 urg_offset = tp->urg_seq - seq;
1417                                 if (urg_offset < len)
1418                                         len = urg_offset;
1419                                 if (!len)
1420                                         break;
1421                         }
1422                         used = recv_actor(desc, skb, offset, len);
1423                         if (used < 0) {
1424                                 if (!copied)
1425                                         copied = used;
1426                                 break;
1427                         } else if (used <= len) {
1428                                 seq += used;
1429                                 copied += used;
1430                                 offset += used;
1431                         }
1432                         /*
1433                          * If recv_actor drops the lock (e.g. TCP splice
1434                          * receive) the skb pointer might be invalid when
1435                          * getting here: tcp_collapse might have deleted it
1436                          * while aggregating skbs from the socket queue.
1437                          */
1438                         skb = tcp_recv_skb(sk, seq-1, &offset);
1439                         if (!skb || (offset+1 != skb->len))
1440                                 break;
1441                 }
1442                 if (tcp_hdr(skb)->fin) {
1443                         sk_eat_skb(sk, skb, 0);
1444                         ++seq;
1445                         break;
1446                 }
1447                 sk_eat_skb(sk, skb, 0);
1448                 if (!desc->count)
1449                         break;
1450                 tp->copied_seq = seq;
1451         }
1452         tp->copied_seq = seq;
1453
1454         tcp_rcv_space_adjust(sk);
1455
1456         /* Clean up data we have read: This will do ACK frames. */
1457         if (copied > 0)
1458                 tcp_cleanup_rbuf(sk, copied);
1459         return copied;
1460 }
1461 EXPORT_SYMBOL(tcp_read_sock);
1462
1463 /*
1464  *      This routine copies from a sock struct into the user buffer.
1465  *
1466  *      Technical note: in 2.3 we work on _locked_ socket, so that
1467  *      tricks with *seq access order and skb->users are not required.
1468  *      Probably, code can be easily improved even more.
1469  */
1470
1471 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1472                 size_t len, int nonblock, int flags, int *addr_len)
1473 {
1474         struct tcp_sock *tp = tcp_sk(sk);
1475         int copied = 0;
1476         u32 peek_seq;
1477         u32 *seq;
1478         unsigned long used;
1479         int err;
1480         int target;             /* Read at least this many bytes */
1481         long timeo;
1482         struct task_struct *user_recv = NULL;
1483         int copied_early = 0;
1484         struct sk_buff *skb;
1485         u32 urg_hole = 0;
1486
1487         lock_sock(sk);
1488
1489         err = -ENOTCONN;
1490         if (sk->sk_state == TCP_LISTEN)
1491                 goto out;
1492
1493         timeo = sock_rcvtimeo(sk, nonblock);
1494
1495         /* Urgent data needs to be handled specially. */
1496         if (flags & MSG_OOB)
1497                 goto recv_urg;
1498
1499         if (unlikely(tp->repair)) {
1500                 err = -EPERM;
1501                 if (!(flags & MSG_PEEK))
1502                         goto out;
1503
1504                 if (tp->repair_queue == TCP_SEND_QUEUE)
1505                         goto recv_sndq;
1506
1507                 err = -EINVAL;
1508                 if (tp->repair_queue == TCP_NO_QUEUE)
1509                         goto out;
1510
1511                 /* 'common' recv queue MSG_PEEK-ing */
1512         }
1513
1514         seq = &tp->copied_seq;
1515         if (flags & MSG_PEEK) {
1516                 peek_seq = tp->copied_seq;
1517                 seq = &peek_seq;
1518         }
1519
1520         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1521
1522 #ifdef CONFIG_NET_DMA
1523         tp->ucopy.dma_chan = NULL;
1524         preempt_disable();
1525         skb = skb_peek_tail(&sk->sk_receive_queue);
1526         {
1527                 int available = 0;
1528
1529                 if (skb)
1530                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1531                 if ((available < target) &&
1532                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1533                     !sysctl_tcp_low_latency &&
1534                     net_dma_find_channel()) {
1535                         preempt_enable_no_resched();
1536                         tp->ucopy.pinned_list =
1537                                         dma_pin_iovec_pages(msg->msg_iov, len);
1538                 } else {
1539                         preempt_enable_no_resched();
1540                 }
1541         }
1542 #endif
1543
1544         do {
1545                 u32 offset;
1546
1547                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1548                 if (tp->urg_data && tp->urg_seq == *seq) {
1549                         if (copied)
1550                                 break;
1551                         if (signal_pending(current)) {
1552                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1553                                 break;
1554                         }
1555                 }
1556
1557                 /* Next get a buffer. */
1558
1559                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1560                         /* Now that we have two receive queues this
1561                          * shouldn't happen.
1562                          */
1563                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1564                                  "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1565                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1566                                  flags))
1567                                 break;
1568
1569                         offset = *seq - TCP_SKB_CB(skb)->seq;
1570                         if (tcp_hdr(skb)->syn)
1571                                 offset--;
1572                         if (offset < skb->len)
1573                                 goto found_ok_skb;
1574                         if (tcp_hdr(skb)->fin)
1575                                 goto found_fin_ok;
1576                         WARN(!(flags & MSG_PEEK),
1577                              "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1578                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1579                 }
1580
1581                 /* Well, if we have backlog, try to process it now yet. */
1582
1583                 if (copied >= target && !sk->sk_backlog.tail)
1584                         break;
1585
1586                 if (copied) {
1587                         if (sk->sk_err ||
1588                             sk->sk_state == TCP_CLOSE ||
1589                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1590                             !timeo ||
1591                             signal_pending(current))
1592                                 break;
1593                 } else {
1594                         if (sock_flag(sk, SOCK_DONE))
1595                                 break;
1596
1597                         if (sk->sk_err) {
1598                                 copied = sock_error(sk);
1599                                 break;
1600                         }
1601
1602                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1603                                 break;
1604
1605                         if (sk->sk_state == TCP_CLOSE) {
1606                                 if (!sock_flag(sk, SOCK_DONE)) {
1607                                         /* This occurs when user tries to read
1608                                          * from never connected socket.
1609                                          */
1610                                         copied = -ENOTCONN;
1611                                         break;
1612                                 }
1613                                 break;
1614                         }
1615
1616                         if (!timeo) {
1617                                 copied = -EAGAIN;
1618                                 break;
1619                         }
1620
1621                         if (signal_pending(current)) {
1622                                 copied = sock_intr_errno(timeo);
1623                                 break;
1624                         }
1625                 }
1626
1627                 tcp_cleanup_rbuf(sk, copied);
1628
1629                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1630                         /* Install new reader */
1631                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1632                                 user_recv = current;
1633                                 tp->ucopy.task = user_recv;
1634                                 tp->ucopy.iov = msg->msg_iov;
1635                         }
1636
1637                         tp->ucopy.len = len;
1638
1639                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1640                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1641
1642                         /* Ugly... If prequeue is not empty, we have to
1643                          * process it before releasing socket, otherwise
1644                          * order will be broken at second iteration.
1645                          * More elegant solution is required!!!
1646                          *
1647                          * Look: we have the following (pseudo)queues:
1648                          *
1649                          * 1. packets in flight
1650                          * 2. backlog
1651                          * 3. prequeue
1652                          * 4. receive_queue
1653                          *
1654                          * Each queue can be processed only if the next ones
1655                          * are empty. At this point we have empty receive_queue.
1656                          * But prequeue _can_ be not empty after 2nd iteration,
1657                          * when we jumped to start of loop because backlog
1658                          * processing added something to receive_queue.
1659                          * We cannot release_sock(), because backlog contains
1660                          * packets arrived _after_ prequeued ones.
1661                          *
1662                          * Shortly, algorithm is clear --- to process all
1663                          * the queues in order. We could make it more directly,
1664                          * requeueing packets from backlog to prequeue, if
1665                          * is not empty. It is more elegant, but eats cycles,
1666                          * unfortunately.
1667                          */
1668                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1669                                 goto do_prequeue;
1670
1671                         /* __ Set realtime policy in scheduler __ */
1672                 }
1673
1674 #ifdef CONFIG_NET_DMA
1675                 if (tp->ucopy.dma_chan)
1676                         dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1677 #endif
1678                 if (copied >= target) {
1679                         /* Do not sleep, just process backlog. */
1680                         release_sock(sk);
1681                         lock_sock(sk);
1682                 } else
1683                         sk_wait_data(sk, &timeo);
1684
1685 #ifdef CONFIG_NET_DMA
1686                 tcp_service_net_dma(sk, false);  /* Don't block */
1687                 tp->ucopy.wakeup = 0;
1688 #endif
1689
1690                 if (user_recv) {
1691                         int chunk;
1692
1693                         /* __ Restore normal policy in scheduler __ */
1694
1695                         if ((chunk = len - tp->ucopy.len) != 0) {
1696                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1697                                 len -= chunk;
1698                                 copied += chunk;
1699                         }
1700
1701                         if (tp->rcv_nxt == tp->copied_seq &&
1702                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1703 do_prequeue:
1704                                 tcp_prequeue_process(sk);
1705
1706                                 if ((chunk = len - tp->ucopy.len) != 0) {
1707                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1708                                         len -= chunk;
1709                                         copied += chunk;
1710                                 }
1711                         }
1712                 }
1713                 if ((flags & MSG_PEEK) &&
1714                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1715                         if (net_ratelimit())
1716                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1717                                        current->comm, task_pid_nr(current));
1718                         peek_seq = tp->copied_seq;
1719                 }
1720                 continue;
1721
1722         found_ok_skb:
1723                 /* Ok so how much can we use? */
1724                 used = skb->len - offset;
1725                 if (len < used)
1726                         used = len;
1727
1728                 /* Do we have urgent data here? */
1729                 if (tp->urg_data) {
1730                         u32 urg_offset = tp->urg_seq - *seq;
1731                         if (urg_offset < used) {
1732                                 if (!urg_offset) {
1733                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1734                                                 ++*seq;
1735                                                 urg_hole++;
1736                                                 offset++;
1737                                                 used--;
1738                                                 if (!used)
1739                                                         goto skip_copy;
1740                                         }
1741                                 } else
1742                                         used = urg_offset;
1743                         }
1744                 }
1745
1746                 if (!(flags & MSG_TRUNC)) {
1747 #ifdef CONFIG_NET_DMA
1748                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1749                                 tp->ucopy.dma_chan = net_dma_find_channel();
1750
1751                         if (tp->ucopy.dma_chan) {
1752                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1753                                         tp->ucopy.dma_chan, skb, offset,
1754                                         msg->msg_iov, used,
1755                                         tp->ucopy.pinned_list);
1756
1757                                 if (tp->ucopy.dma_cookie < 0) {
1758
1759                                         pr_alert("%s: dma_cookie < 0\n",
1760                                                  __func__);
1761
1762                                         /* Exception. Bailout! */
1763                                         if (!copied)
1764                                                 copied = -EFAULT;
1765                                         break;
1766                                 }
1767
1768                                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1769
1770                                 if ((offset + used) == skb->len)
1771                                         copied_early = 1;
1772
1773                         } else
1774 #endif
1775                         {
1776                                 err = skb_copy_datagram_iovec(skb, offset,
1777                                                 msg->msg_iov, used);
1778                                 if (err) {
1779                                         /* Exception. Bailout! */
1780                                         if (!copied)
1781                                                 copied = -EFAULT;
1782                                         break;
1783                                 }
1784                         }
1785                 }
1786
1787                 *seq += used;
1788                 copied += used;
1789                 len -= used;
1790
1791                 tcp_rcv_space_adjust(sk);
1792
1793 skip_copy:
1794                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1795                         tp->urg_data = 0;
1796                         tcp_fast_path_check(sk);
1797                 }
1798                 if (used + offset < skb->len)
1799                         continue;
1800
1801                 if (tcp_hdr(skb)->fin)
1802                         goto found_fin_ok;
1803                 if (!(flags & MSG_PEEK)) {
1804                         sk_eat_skb(sk, skb, copied_early);
1805                         copied_early = 0;
1806                 }
1807                 continue;
1808
1809         found_fin_ok:
1810                 /* Process the FIN. */
1811                 ++*seq;
1812                 if (!(flags & MSG_PEEK)) {
1813                         sk_eat_skb(sk, skb, copied_early);
1814                         copied_early = 0;
1815                 }
1816                 break;
1817         } while (len > 0);
1818
1819         if (user_recv) {
1820                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1821                         int chunk;
1822
1823                         tp->ucopy.len = copied > 0 ? len : 0;
1824
1825                         tcp_prequeue_process(sk);
1826
1827                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1828                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1829                                 len -= chunk;
1830                                 copied += chunk;
1831                         }
1832                 }
1833
1834                 tp->ucopy.task = NULL;
1835                 tp->ucopy.len = 0;
1836         }
1837
1838 #ifdef CONFIG_NET_DMA
1839         tcp_service_net_dma(sk, true);  /* Wait for queue to drain */
1840         tp->ucopy.dma_chan = NULL;
1841
1842         if (tp->ucopy.pinned_list) {
1843                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1844                 tp->ucopy.pinned_list = NULL;
1845         }
1846 #endif
1847
1848         /* According to UNIX98, msg_name/msg_namelen are ignored
1849          * on connected socket. I was just happy when found this 8) --ANK
1850          */
1851
1852         /* Clean up data we have read: This will do ACK frames. */
1853         tcp_cleanup_rbuf(sk, copied);
1854
1855         release_sock(sk);
1856         return copied;
1857
1858 out:
1859         release_sock(sk);
1860         return err;
1861
1862 recv_urg:
1863         err = tcp_recv_urg(sk, msg, len, flags);
1864         goto out;
1865
1866 recv_sndq:
1867         err = tcp_peek_sndq(sk, msg, len);
1868         goto out;
1869 }
1870 EXPORT_SYMBOL(tcp_recvmsg);
1871
1872 void tcp_set_state(struct sock *sk, int state)
1873 {
1874         int oldstate = sk->sk_state;
1875
1876         switch (state) {
1877         case TCP_ESTABLISHED:
1878                 if (oldstate != TCP_ESTABLISHED)
1879                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1880                 break;
1881
1882         case TCP_CLOSE:
1883                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1884                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1885
1886                 sk->sk_prot->unhash(sk);
1887                 if (inet_csk(sk)->icsk_bind_hash &&
1888                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1889                         inet_put_port(sk);
1890                 /* fall through */
1891         default:
1892                 if (oldstate == TCP_ESTABLISHED)
1893                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1894         }
1895
1896         /* Change state AFTER socket is unhashed to avoid closed
1897          * socket sitting in hash tables.
1898          */
1899         sk->sk_state = state;
1900
1901 #ifdef STATE_TRACE
1902         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1903 #endif
1904 }
1905 EXPORT_SYMBOL_GPL(tcp_set_state);
1906
1907 /*
1908  *      State processing on a close. This implements the state shift for
1909  *      sending our FIN frame. Note that we only send a FIN for some
1910  *      states. A shutdown() may have already sent the FIN, or we may be
1911  *      closed.
1912  */
1913
1914 static const unsigned char new_state[16] = {
1915   /* current state:        new state:      action:      */
1916   /* (Invalid)          */ TCP_CLOSE,
1917   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1918   /* TCP_SYN_SENT       */ TCP_CLOSE,
1919   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1920   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1921   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1922   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1923   /* TCP_CLOSE          */ TCP_CLOSE,
1924   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1925   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1926   /* TCP_LISTEN         */ TCP_CLOSE,
1927   /* TCP_CLOSING        */ TCP_CLOSING,
1928 };
1929
1930 static int tcp_close_state(struct sock *sk)
1931 {
1932         int next = (int)new_state[sk->sk_state];
1933         int ns = next & TCP_STATE_MASK;
1934
1935         tcp_set_state(sk, ns);
1936
1937         return next & TCP_ACTION_FIN;
1938 }
1939
1940 /*
1941  *      Shutdown the sending side of a connection. Much like close except
1942  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1943  */
1944
1945 void tcp_shutdown(struct sock *sk, int how)
1946 {
1947         /*      We need to grab some memory, and put together a FIN,
1948          *      and then put it into the queue to be sent.
1949          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1950          */
1951         if (!(how & SEND_SHUTDOWN))
1952                 return;
1953
1954         /* If we've already sent a FIN, or it's a closed state, skip this. */
1955         if ((1 << sk->sk_state) &
1956             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1957              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1958                 /* Clear out any half completed packets.  FIN if needed. */
1959                 if (tcp_close_state(sk))
1960                         tcp_send_fin(sk);
1961         }
1962 }
1963 EXPORT_SYMBOL(tcp_shutdown);
1964
1965 bool tcp_check_oom(struct sock *sk, int shift)
1966 {
1967         bool too_many_orphans, out_of_socket_memory;
1968
1969         too_many_orphans = tcp_too_many_orphans(sk, shift);
1970         out_of_socket_memory = tcp_out_of_memory(sk);
1971
1972         if (too_many_orphans && net_ratelimit())
1973                 pr_info("too many orphaned sockets\n");
1974         if (out_of_socket_memory && net_ratelimit())
1975                 pr_info("out of memory -- consider tuning tcp_mem\n");
1976         return too_many_orphans || out_of_socket_memory;
1977 }
1978
1979 void tcp_close(struct sock *sk, long timeout)
1980 {
1981         struct sk_buff *skb;
1982         int data_was_unread = 0;
1983         int state;
1984
1985         lock_sock(sk);
1986         sk->sk_shutdown = SHUTDOWN_MASK;
1987
1988         if (sk->sk_state == TCP_LISTEN) {
1989                 tcp_set_state(sk, TCP_CLOSE);
1990
1991                 /* Special case. */
1992                 inet_csk_listen_stop(sk);
1993
1994                 goto adjudge_to_death;
1995         }
1996
1997         /*  We need to flush the recv. buffs.  We do this only on the
1998          *  descriptor close, not protocol-sourced closes, because the
1999          *  reader process may not have drained the data yet!
2000          */
2001         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2002                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2003                           tcp_hdr(skb)->fin;
2004                 data_was_unread += len;
2005                 __kfree_skb(skb);
2006         }
2007
2008         sk_mem_reclaim(sk);
2009
2010         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2011         if (sk->sk_state == TCP_CLOSE)
2012                 goto adjudge_to_death;
2013
2014         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2015          * data was lost. To witness the awful effects of the old behavior of
2016          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2017          * GET in an FTP client, suspend the process, wait for the client to
2018          * advertise a zero window, then kill -9 the FTP client, wheee...
2019          * Note: timeout is always zero in such a case.
2020          */
2021         if (unlikely(tcp_sk(sk)->repair)) {
2022                 sk->sk_prot->disconnect(sk, 0);
2023         } else if (data_was_unread) {
2024                 /* Unread data was tossed, zap the connection. */
2025                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2026                 tcp_set_state(sk, TCP_CLOSE);
2027                 tcp_send_active_reset(sk, sk->sk_allocation);
2028         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2029                 /* Check zero linger _after_ checking for unread data. */
2030                 sk->sk_prot->disconnect(sk, 0);
2031                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2032         } else if (tcp_close_state(sk)) {
2033                 /* We FIN if the application ate all the data before
2034                  * zapping the connection.
2035                  */
2036
2037                 /* RED-PEN. Formally speaking, we have broken TCP state
2038                  * machine. State transitions:
2039                  *
2040                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2041                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2042                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2043                  *
2044                  * are legal only when FIN has been sent (i.e. in window),
2045                  * rather than queued out of window. Purists blame.
2046                  *
2047                  * F.e. "RFC state" is ESTABLISHED,
2048                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2049                  *
2050                  * The visible declinations are that sometimes
2051                  * we enter time-wait state, when it is not required really
2052                  * (harmless), do not send active resets, when they are
2053                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2054                  * they look as CLOSING or LAST_ACK for Linux)
2055                  * Probably, I missed some more holelets.
2056                  *                                              --ANK
2057                  */
2058                 tcp_send_fin(sk);
2059         }
2060
2061         sk_stream_wait_close(sk, timeout);
2062
2063 adjudge_to_death:
2064         state = sk->sk_state;
2065         sock_hold(sk);
2066         sock_orphan(sk);
2067
2068         /* It is the last release_sock in its life. It will remove backlog. */
2069         release_sock(sk);
2070
2071
2072         /* Now socket is owned by kernel and we acquire BH lock
2073            to finish close. No need to check for user refs.
2074          */
2075         local_bh_disable();
2076         bh_lock_sock(sk);
2077         WARN_ON(sock_owned_by_user(sk));
2078
2079         percpu_counter_inc(sk->sk_prot->orphan_count);
2080
2081         /* Have we already been destroyed by a softirq or backlog? */
2082         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2083                 goto out;
2084
2085         /*      This is a (useful) BSD violating of the RFC. There is a
2086          *      problem with TCP as specified in that the other end could
2087          *      keep a socket open forever with no application left this end.
2088          *      We use a 3 minute timeout (about the same as BSD) then kill
2089          *      our end. If they send after that then tough - BUT: long enough
2090          *      that we won't make the old 4*rto = almost no time - whoops
2091          *      reset mistake.
2092          *
2093          *      Nope, it was not mistake. It is really desired behaviour
2094          *      f.e. on http servers, when such sockets are useless, but
2095          *      consume significant resources. Let's do it with special
2096          *      linger2 option.                                 --ANK
2097          */
2098
2099         if (sk->sk_state == TCP_FIN_WAIT2) {
2100                 struct tcp_sock *tp = tcp_sk(sk);
2101                 if (tp->linger2 < 0) {
2102                         tcp_set_state(sk, TCP_CLOSE);
2103                         tcp_send_active_reset(sk, GFP_ATOMIC);
2104                         NET_INC_STATS_BH(sock_net(sk),
2105                                         LINUX_MIB_TCPABORTONLINGER);
2106                 } else {
2107                         const int tmo = tcp_fin_time(sk);
2108
2109                         if (tmo > TCP_TIMEWAIT_LEN) {
2110                                 inet_csk_reset_keepalive_timer(sk,
2111                                                 tmo - TCP_TIMEWAIT_LEN);
2112                         } else {
2113                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2114                                 goto out;
2115                         }
2116                 }
2117         }
2118         if (sk->sk_state != TCP_CLOSE) {
2119                 sk_mem_reclaim(sk);
2120                 if (tcp_check_oom(sk, 0)) {
2121                         tcp_set_state(sk, TCP_CLOSE);
2122                         tcp_send_active_reset(sk, GFP_ATOMIC);
2123                         NET_INC_STATS_BH(sock_net(sk),
2124                                         LINUX_MIB_TCPABORTONMEMORY);
2125                 }
2126         }
2127
2128         if (sk->sk_state == TCP_CLOSE)
2129                 inet_csk_destroy_sock(sk);
2130         /* Otherwise, socket is reprieved until protocol close. */
2131
2132 out:
2133         bh_unlock_sock(sk);
2134         local_bh_enable();
2135         sock_put(sk);
2136 }
2137 EXPORT_SYMBOL(tcp_close);
2138
2139 /* These states need RST on ABORT according to RFC793 */
2140
2141 static inline int tcp_need_reset(int state)
2142 {
2143         return (1 << state) &
2144                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2145                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2146 }
2147
2148 int tcp_disconnect(struct sock *sk, int flags)
2149 {
2150         struct inet_sock *inet = inet_sk(sk);
2151         struct inet_connection_sock *icsk = inet_csk(sk);
2152         struct tcp_sock *tp = tcp_sk(sk);
2153         int err = 0;
2154         int old_state = sk->sk_state;
2155
2156         if (old_state != TCP_CLOSE)
2157                 tcp_set_state(sk, TCP_CLOSE);
2158
2159         /* ABORT function of RFC793 */
2160         if (old_state == TCP_LISTEN) {
2161                 inet_csk_listen_stop(sk);
2162         } else if (unlikely(tp->repair)) {
2163                 sk->sk_err = ECONNABORTED;
2164         } else if (tcp_need_reset(old_state) ||
2165                    (tp->snd_nxt != tp->write_seq &&
2166                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2167                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2168                  * states
2169                  */
2170                 tcp_send_active_reset(sk, gfp_any());
2171                 sk->sk_err = ECONNRESET;
2172         } else if (old_state == TCP_SYN_SENT)
2173                 sk->sk_err = ECONNRESET;
2174
2175         tcp_clear_xmit_timers(sk);
2176         __skb_queue_purge(&sk->sk_receive_queue);
2177         tcp_write_queue_purge(sk);
2178         __skb_queue_purge(&tp->out_of_order_queue);
2179 #ifdef CONFIG_NET_DMA
2180         __skb_queue_purge(&sk->sk_async_wait_queue);
2181 #endif
2182
2183         inet->inet_dport = 0;
2184
2185         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2186                 inet_reset_saddr(sk);
2187
2188         sk->sk_shutdown = 0;
2189         sock_reset_flag(sk, SOCK_DONE);
2190         tp->srtt = 0;
2191         if ((tp->write_seq += tp->max_window + 2) == 0)
2192                 tp->write_seq = 1;
2193         icsk->icsk_backoff = 0;
2194         tp->snd_cwnd = 2;
2195         icsk->icsk_probes_out = 0;
2196         tp->packets_out = 0;
2197         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2198         tp->snd_cwnd_cnt = 0;
2199         tp->bytes_acked = 0;
2200         tp->window_clamp = 0;
2201         tcp_set_ca_state(sk, TCP_CA_Open);
2202         tcp_clear_retrans(tp);
2203         inet_csk_delack_init(sk);
2204         tcp_init_send_head(sk);
2205         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2206         __sk_dst_reset(sk);
2207
2208         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2209
2210         sk->sk_error_report(sk);
2211         return err;
2212 }
2213 EXPORT_SYMBOL(tcp_disconnect);
2214
2215 static inline int tcp_can_repair_sock(struct sock *sk)
2216 {
2217         return capable(CAP_NET_ADMIN) &&
2218                 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2219 }
2220
2221 /*
2222  *      Socket option code for TCP.
2223  */
2224 static int do_tcp_setsockopt(struct sock *sk, int level,
2225                 int optname, char __user *optval, unsigned int optlen)
2226 {
2227         struct tcp_sock *tp = tcp_sk(sk);
2228         struct inet_connection_sock *icsk = inet_csk(sk);
2229         int val;
2230         int err = 0;
2231
2232         /* These are data/string values, all the others are ints */
2233         switch (optname) {
2234         case TCP_CONGESTION: {
2235                 char name[TCP_CA_NAME_MAX];
2236
2237                 if (optlen < 1)
2238                         return -EINVAL;
2239
2240                 val = strncpy_from_user(name, optval,
2241                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2242                 if (val < 0)
2243                         return -EFAULT;
2244                 name[val] = 0;
2245
2246                 lock_sock(sk);
2247                 err = tcp_set_congestion_control(sk, name);
2248                 release_sock(sk);
2249                 return err;
2250         }
2251         case TCP_COOKIE_TRANSACTIONS: {
2252                 struct tcp_cookie_transactions ctd;
2253                 struct tcp_cookie_values *cvp = NULL;
2254
2255                 if (sizeof(ctd) > optlen)
2256                         return -EINVAL;
2257                 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2258                         return -EFAULT;
2259
2260                 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2261                     ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2262                         return -EINVAL;
2263
2264                 if (ctd.tcpct_cookie_desired == 0) {
2265                         /* default to global value */
2266                 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2267                            ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2268                            ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2269                         return -EINVAL;
2270                 }
2271
2272                 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2273                         /* Supercedes all other values */
2274                         lock_sock(sk);
2275                         if (tp->cookie_values != NULL) {
2276                                 kref_put(&tp->cookie_values->kref,
2277                                          tcp_cookie_values_release);
2278                                 tp->cookie_values = NULL;
2279                         }
2280                         tp->rx_opt.cookie_in_always = 0; /* false */
2281                         tp->rx_opt.cookie_out_never = 1; /* true */
2282                         release_sock(sk);
2283                         return err;
2284                 }
2285
2286                 /* Allocate ancillary memory before locking.
2287                  */
2288                 if (ctd.tcpct_used > 0 ||
2289                     (tp->cookie_values == NULL &&
2290                      (sysctl_tcp_cookie_size > 0 ||
2291                       ctd.tcpct_cookie_desired > 0 ||
2292                       ctd.tcpct_s_data_desired > 0))) {
2293                         cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2294                                       GFP_KERNEL);
2295                         if (cvp == NULL)
2296                                 return -ENOMEM;
2297
2298                         kref_init(&cvp->kref);
2299                 }
2300                 lock_sock(sk);
2301                 tp->rx_opt.cookie_in_always =
2302                         (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2303                 tp->rx_opt.cookie_out_never = 0; /* false */
2304
2305                 if (tp->cookie_values != NULL) {
2306                         if (cvp != NULL) {
2307                                 /* Changed values are recorded by a changed
2308                                  * pointer, ensuring the cookie will differ,
2309                                  * without separately hashing each value later.
2310                                  */
2311                                 kref_put(&tp->cookie_values->kref,
2312                                          tcp_cookie_values_release);
2313                         } else {
2314                                 cvp = tp->cookie_values;
2315                         }
2316                 }
2317
2318                 if (cvp != NULL) {
2319                         cvp->cookie_desired = ctd.tcpct_cookie_desired;
2320
2321                         if (ctd.tcpct_used > 0) {
2322                                 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2323                                        ctd.tcpct_used);
2324                                 cvp->s_data_desired = ctd.tcpct_used;
2325                                 cvp->s_data_constant = 1; /* true */
2326                         } else {
2327                                 /* No constant payload data. */
2328                                 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2329                                 cvp->s_data_constant = 0; /* false */
2330                         }
2331
2332                         tp->cookie_values = cvp;
2333                 }
2334                 release_sock(sk);
2335                 return err;
2336         }
2337         default:
2338                 /* fallthru */
2339                 break;
2340         }
2341
2342         if (optlen < sizeof(int))
2343                 return -EINVAL;
2344
2345         if (get_user(val, (int __user *)optval))
2346                 return -EFAULT;
2347
2348         lock_sock(sk);
2349
2350         switch (optname) {
2351         case TCP_MAXSEG:
2352                 /* Values greater than interface MTU won't take effect. However
2353                  * at the point when this call is done we typically don't yet
2354                  * know which interface is going to be used */
2355                 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2356                         err = -EINVAL;
2357                         break;
2358                 }
2359                 tp->rx_opt.user_mss = val;
2360                 break;
2361
2362         case TCP_NODELAY:
2363                 if (val) {
2364                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2365                          * this option on corked socket is remembered, but
2366                          * it is not activated until cork is cleared.
2367                          *
2368                          * However, when TCP_NODELAY is set we make
2369                          * an explicit push, which overrides even TCP_CORK
2370                          * for currently queued segments.
2371                          */
2372                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2373                         tcp_push_pending_frames(sk);
2374                 } else {
2375                         tp->nonagle &= ~TCP_NAGLE_OFF;
2376                 }
2377                 break;
2378
2379         case TCP_THIN_LINEAR_TIMEOUTS:
2380                 if (val < 0 || val > 1)
2381                         err = -EINVAL;
2382                 else
2383                         tp->thin_lto = val;
2384                 break;
2385
2386         case TCP_THIN_DUPACK:
2387                 if (val < 0 || val > 1)
2388                         err = -EINVAL;
2389                 else
2390                         tp->thin_dupack = val;
2391                 break;
2392
2393         case TCP_REPAIR:
2394                 if (!tcp_can_repair_sock(sk))
2395                         err = -EPERM;
2396                 else if (val == 1) {
2397                         tp->repair = 1;
2398                         sk->sk_reuse = SK_FORCE_REUSE;
2399                         tp->repair_queue = TCP_NO_QUEUE;
2400                 } else if (val == 0) {
2401                         tp->repair = 0;
2402                         sk->sk_reuse = SK_NO_REUSE;
2403                         tcp_send_window_probe(sk);
2404                 } else
2405                         err = -EINVAL;
2406
2407                 break;
2408
2409         case TCP_REPAIR_QUEUE:
2410                 if (!tp->repair)
2411                         err = -EPERM;
2412                 else if (val < TCP_QUEUES_NR)
2413                         tp->repair_queue = val;
2414                 else
2415                         err = -EINVAL;
2416                 break;
2417
2418         case TCP_QUEUE_SEQ:
2419                 if (sk->sk_state != TCP_CLOSE)
2420                         err = -EPERM;
2421                 else if (tp->repair_queue == TCP_SEND_QUEUE)
2422                         tp->write_seq = val;
2423                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2424                         tp->rcv_nxt = val;
2425                 else
2426                         err = -EINVAL;
2427                 break;
2428
2429         case TCP_CORK:
2430                 /* When set indicates to always queue non-full frames.
2431                  * Later the user clears this option and we transmit
2432                  * any pending partial frames in the queue.  This is
2433                  * meant to be used alongside sendfile() to get properly
2434                  * filled frames when the user (for example) must write
2435                  * out headers with a write() call first and then use
2436                  * sendfile to send out the data parts.
2437                  *
2438                  * TCP_CORK can be set together with TCP_NODELAY and it is
2439                  * stronger than TCP_NODELAY.
2440                  */
2441                 if (val) {
2442                         tp->nonagle |= TCP_NAGLE_CORK;
2443                 } else {
2444                         tp->nonagle &= ~TCP_NAGLE_CORK;
2445                         if (tp->nonagle&TCP_NAGLE_OFF)
2446                                 tp->nonagle |= TCP_NAGLE_PUSH;
2447                         tcp_push_pending_frames(sk);
2448                 }
2449                 break;
2450
2451         case TCP_KEEPIDLE:
2452                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2453                         err = -EINVAL;
2454                 else {
2455                         tp->keepalive_time = val * HZ;
2456                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2457                             !((1 << sk->sk_state) &
2458                               (TCPF_CLOSE | TCPF_LISTEN))) {
2459                                 u32 elapsed = keepalive_time_elapsed(tp);
2460                                 if (tp->keepalive_time > elapsed)
2461                                         elapsed = tp->keepalive_time - elapsed;
2462                                 else
2463                                         elapsed = 0;
2464                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2465                         }
2466                 }
2467                 break;
2468         case TCP_KEEPINTVL:
2469                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2470                         err = -EINVAL;
2471                 else
2472                         tp->keepalive_intvl = val * HZ;
2473                 break;
2474         case TCP_KEEPCNT:
2475                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2476                         err = -EINVAL;
2477                 else
2478                         tp->keepalive_probes = val;
2479                 break;
2480         case TCP_SYNCNT:
2481                 if (val < 1 || val > MAX_TCP_SYNCNT)
2482                         err = -EINVAL;
2483                 else
2484                         icsk->icsk_syn_retries = val;
2485                 break;
2486
2487         case TCP_LINGER2:
2488                 if (val < 0)
2489                         tp->linger2 = -1;
2490                 else if (val > sysctl_tcp_fin_timeout / HZ)
2491                         tp->linger2 = 0;
2492                 else
2493                         tp->linger2 = val * HZ;
2494                 break;
2495
2496         case TCP_DEFER_ACCEPT:
2497                 /* Translate value in seconds to number of retransmits */
2498                 icsk->icsk_accept_queue.rskq_defer_accept =
2499                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2500                                         TCP_RTO_MAX / HZ);
2501                 break;
2502
2503         case TCP_WINDOW_CLAMP:
2504                 if (!val) {
2505                         if (sk->sk_state != TCP_CLOSE) {
2506                                 err = -EINVAL;
2507                                 break;
2508                         }
2509                         tp->window_clamp = 0;
2510                 } else
2511                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2512                                                 SOCK_MIN_RCVBUF / 2 : val;
2513                 break;
2514
2515         case TCP_QUICKACK:
2516                 if (!val) {
2517                         icsk->icsk_ack.pingpong = 1;
2518                 } else {
2519                         icsk->icsk_ack.pingpong = 0;
2520                         if ((1 << sk->sk_state) &
2521                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2522                             inet_csk_ack_scheduled(sk)) {
2523                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2524                                 tcp_cleanup_rbuf(sk, 1);
2525                                 if (!(val & 1))
2526                                         icsk->icsk_ack.pingpong = 1;
2527                         }
2528                 }
2529                 break;
2530
2531 #ifdef CONFIG_TCP_MD5SIG
2532         case TCP_MD5SIG:
2533                 /* Read the IP->Key mappings from userspace */
2534                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2535                 break;
2536 #endif
2537         case TCP_USER_TIMEOUT:
2538                 /* Cap the max timeout in ms TCP will retry/retrans
2539                  * before giving up and aborting (ETIMEDOUT) a connection.
2540                  */
2541                 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2542                 break;
2543         default:
2544                 err = -ENOPROTOOPT;
2545                 break;
2546         }
2547
2548         release_sock(sk);
2549         return err;
2550 }
2551
2552 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2553                    unsigned int optlen)
2554 {
2555         const struct inet_connection_sock *icsk = inet_csk(sk);
2556
2557         if (level != SOL_TCP)
2558                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2559                                                      optval, optlen);
2560         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2561 }
2562 EXPORT_SYMBOL(tcp_setsockopt);
2563
2564 #ifdef CONFIG_COMPAT
2565 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2566                           char __user *optval, unsigned int optlen)
2567 {
2568         if (level != SOL_TCP)
2569                 return inet_csk_compat_setsockopt(sk, level, optname,
2570                                                   optval, optlen);
2571         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2572 }
2573 EXPORT_SYMBOL(compat_tcp_setsockopt);
2574 #endif
2575
2576 /* Return information about state of tcp endpoint in API format. */
2577 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2578 {
2579         const struct tcp_sock *tp = tcp_sk(sk);
2580         const struct inet_connection_sock *icsk = inet_csk(sk);
2581         u32 now = tcp_time_stamp;
2582
2583         memset(info, 0, sizeof(*info));
2584
2585         info->tcpi_state = sk->sk_state;
2586         info->tcpi_ca_state = icsk->icsk_ca_state;
2587         info->tcpi_retransmits = icsk->icsk_retransmits;
2588         info->tcpi_probes = icsk->icsk_probes_out;
2589         info->tcpi_backoff = icsk->icsk_backoff;
2590
2591         if (tp->rx_opt.tstamp_ok)
2592                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2593         if (tcp_is_sack(tp))
2594                 info->tcpi_options |= TCPI_OPT_SACK;
2595         if (tp->rx_opt.wscale_ok) {
2596                 info->tcpi_options |= TCPI_OPT_WSCALE;
2597                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2598                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2599         }
2600
2601         if (tp->ecn_flags & TCP_ECN_OK)
2602                 info->tcpi_options |= TCPI_OPT_ECN;
2603         if (tp->ecn_flags & TCP_ECN_SEEN)
2604                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2605
2606         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2607         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2608         info->tcpi_snd_mss = tp->mss_cache;
2609         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2610
2611         if (sk->sk_state == TCP_LISTEN) {
2612                 info->tcpi_unacked = sk->sk_ack_backlog;
2613                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2614         } else {
2615                 info->tcpi_unacked = tp->packets_out;
2616                 info->tcpi_sacked = tp->sacked_out;
2617         }
2618         info->tcpi_lost = tp->lost_out;
2619         info->tcpi_retrans = tp->retrans_out;
2620         info->tcpi_fackets = tp->fackets_out;
2621
2622         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2623         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2624         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2625
2626         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2627         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2628         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2629         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2630         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2631         info->tcpi_snd_cwnd = tp->snd_cwnd;
2632         info->tcpi_advmss = tp->advmss;
2633         info->tcpi_reordering = tp->reordering;
2634
2635         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2636         info->tcpi_rcv_space = tp->rcvq_space.space;
2637
2638         info->tcpi_total_retrans = tp->total_retrans;
2639 }
2640 EXPORT_SYMBOL_GPL(tcp_get_info);
2641
2642 static int do_tcp_getsockopt(struct sock *sk, int level,
2643                 int optname, char __user *optval, int __user *optlen)
2644 {
2645         struct inet_connection_sock *icsk = inet_csk(sk);
2646         struct tcp_sock *tp = tcp_sk(sk);
2647         int val, len;
2648
2649         if (get_user(len, optlen))
2650                 return -EFAULT;
2651
2652         len = min_t(unsigned int, len, sizeof(int));
2653
2654         if (len < 0)
2655                 return -EINVAL;
2656
2657         switch (optname) {
2658         case TCP_MAXSEG:
2659                 val = tp->mss_cache;
2660                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2661                         val = tp->rx_opt.user_mss;
2662                 if (tp->repair)
2663                         val = tp->rx_opt.mss_clamp;
2664                 break;
2665         case TCP_NODELAY:
2666                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2667                 break;
2668         case TCP_CORK:
2669                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2670                 break;
2671         case TCP_KEEPIDLE:
2672                 val = keepalive_time_when(tp) / HZ;
2673                 break;
2674         case TCP_KEEPINTVL:
2675                 val = keepalive_intvl_when(tp) / HZ;
2676                 break;
2677         case TCP_KEEPCNT:
2678                 val = keepalive_probes(tp);
2679                 break;
2680         case TCP_SYNCNT:
2681                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2682                 break;
2683         case TCP_LINGER2:
2684                 val = tp->linger2;
2685                 if (val >= 0)
2686                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2687                 break;
2688         case TCP_DEFER_ACCEPT:
2689                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2690                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2691                 break;
2692         case TCP_WINDOW_CLAMP:
2693                 val = tp->window_clamp;
2694                 break;
2695         case TCP_INFO: {
2696                 struct tcp_info info;
2697
2698                 if (get_user(len, optlen))
2699                         return -EFAULT;
2700
2701                 tcp_get_info(sk, &info);
2702
2703                 len = min_t(unsigned int, len, sizeof(info));
2704                 if (put_user(len, optlen))
2705                         return -EFAULT;
2706                 if (copy_to_user(optval, &info, len))
2707                         return -EFAULT;
2708                 return 0;
2709         }
2710         case TCP_QUICKACK:
2711                 val = !icsk->icsk_ack.pingpong;
2712                 break;
2713
2714         case TCP_CONGESTION:
2715                 if (get_user(len, optlen))
2716                         return -EFAULT;
2717                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2718                 if (put_user(len, optlen))
2719                         return -EFAULT;
2720                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2721                         return -EFAULT;
2722                 return 0;
2723
2724         case TCP_COOKIE_TRANSACTIONS: {
2725                 struct tcp_cookie_transactions ctd;
2726                 struct tcp_cookie_values *cvp = tp->cookie_values;
2727
2728                 if (get_user(len, optlen))
2729                         return -EFAULT;
2730                 if (len < sizeof(ctd))
2731                         return -EINVAL;
2732
2733                 memset(&ctd, 0, sizeof(ctd));
2734                 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2735                                    TCP_COOKIE_IN_ALWAYS : 0)
2736                                 | (tp->rx_opt.cookie_out_never ?
2737                                    TCP_COOKIE_OUT_NEVER : 0);
2738
2739                 if (cvp != NULL) {
2740                         ctd.tcpct_flags |= (cvp->s_data_in ?
2741                                             TCP_S_DATA_IN : 0)
2742                                          | (cvp->s_data_out ?
2743                                             TCP_S_DATA_OUT : 0);
2744
2745                         ctd.tcpct_cookie_desired = cvp->cookie_desired;
2746                         ctd.tcpct_s_data_desired = cvp->s_data_desired;
2747
2748                         memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2749                                cvp->cookie_pair_size);
2750                         ctd.tcpct_used = cvp->cookie_pair_size;
2751                 }
2752
2753                 if (put_user(sizeof(ctd), optlen))
2754                         return -EFAULT;
2755                 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2756                         return -EFAULT;
2757                 return 0;
2758         }
2759         case TCP_THIN_LINEAR_TIMEOUTS:
2760                 val = tp->thin_lto;
2761                 break;
2762         case TCP_THIN_DUPACK:
2763                 val = tp->thin_dupack;
2764                 break;
2765
2766         case TCP_REPAIR:
2767                 val = tp->repair;
2768                 break;
2769
2770         case TCP_REPAIR_QUEUE:
2771                 if (tp->repair)
2772                         val = tp->repair_queue;
2773                 else
2774                         return -EINVAL;
2775                 break;
2776
2777         case TCP_QUEUE_SEQ:
2778                 if (tp->repair_queue == TCP_SEND_QUEUE)
2779                         val = tp->write_seq;
2780                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2781                         val = tp->rcv_nxt;
2782                 else
2783                         return -EINVAL;
2784                 break;
2785
2786         case TCP_USER_TIMEOUT:
2787                 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2788                 break;
2789         default:
2790                 return -ENOPROTOOPT;
2791         }
2792
2793         if (put_user(len, optlen))
2794                 return -EFAULT;
2795         if (copy_to_user(optval, &val, len))
2796                 return -EFAULT;
2797         return 0;
2798 }
2799
2800 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2801                    int __user *optlen)
2802 {
2803         struct inet_connection_sock *icsk = inet_csk(sk);
2804
2805         if (level != SOL_TCP)
2806                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2807                                                      optval, optlen);
2808         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2809 }
2810 EXPORT_SYMBOL(tcp_getsockopt);
2811
2812 #ifdef CONFIG_COMPAT
2813 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2814                           char __user *optval, int __user *optlen)
2815 {
2816         if (level != SOL_TCP)
2817                 return inet_csk_compat_getsockopt(sk, level, optname,
2818                                                   optval, optlen);
2819         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2820 }
2821 EXPORT_SYMBOL(compat_tcp_getsockopt);
2822 #endif
2823
2824 struct sk_buff *tcp_tso_segment(struct sk_buff *skb,
2825         netdev_features_t features)
2826 {
2827         struct sk_buff *segs = ERR_PTR(-EINVAL);
2828         struct tcphdr *th;
2829         unsigned int thlen;
2830         unsigned int seq;
2831         __be32 delta;
2832         unsigned int oldlen;
2833         unsigned int mss;
2834
2835         if (!pskb_may_pull(skb, sizeof(*th)))
2836                 goto out;
2837
2838         th = tcp_hdr(skb);
2839         thlen = th->doff * 4;
2840         if (thlen < sizeof(*th))
2841                 goto out;
2842
2843         if (!pskb_may_pull(skb, thlen))
2844                 goto out;
2845
2846         oldlen = (u16)~skb->len;
2847         __skb_pull(skb, thlen);
2848
2849         mss = skb_shinfo(skb)->gso_size;
2850         if (unlikely(skb->len <= mss))
2851                 goto out;
2852
2853         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2854                 /* Packet is from an untrusted source, reset gso_segs. */
2855                 int type = skb_shinfo(skb)->gso_type;
2856
2857                 if (unlikely(type &
2858                              ~(SKB_GSO_TCPV4 |
2859                                SKB_GSO_DODGY |
2860                                SKB_GSO_TCP_ECN |
2861                                SKB_GSO_TCPV6 |
2862                                0) ||
2863                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2864                         goto out;
2865
2866                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2867
2868                 segs = NULL;
2869                 goto out;
2870         }
2871
2872         segs = skb_segment(skb, features);
2873         if (IS_ERR(segs))
2874                 goto out;
2875
2876         delta = htonl(oldlen + (thlen + mss));
2877
2878         skb = segs;
2879         th = tcp_hdr(skb);
2880         seq = ntohl(th->seq);
2881
2882         do {
2883                 th->fin = th->psh = 0;
2884
2885                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2886                                        (__force u32)delta));
2887                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2888                         th->check =
2889                              csum_fold(csum_partial(skb_transport_header(skb),
2890                                                     thlen, skb->csum));
2891
2892                 seq += mss;
2893                 skb = skb->next;
2894                 th = tcp_hdr(skb);
2895
2896                 th->seq = htonl(seq);
2897                 th->cwr = 0;
2898         } while (skb->next);
2899
2900         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2901                       skb->data_len);
2902         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2903                                 (__force u32)delta));
2904         if (skb->ip_summed != CHECKSUM_PARTIAL)
2905                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2906                                                    thlen, skb->csum));
2907
2908 out:
2909         return segs;
2910 }
2911 EXPORT_SYMBOL(tcp_tso_segment);
2912
2913 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2914 {
2915         struct sk_buff **pp = NULL;
2916         struct sk_buff *p;
2917         struct tcphdr *th;
2918         struct tcphdr *th2;
2919         unsigned int len;
2920         unsigned int thlen;
2921         __be32 flags;
2922         unsigned int mss = 1;
2923         unsigned int hlen;
2924         unsigned int off;
2925         int flush = 1;
2926         int i;
2927
2928         off = skb_gro_offset(skb);
2929         hlen = off + sizeof(*th);
2930         th = skb_gro_header_fast(skb, off);
2931         if (skb_gro_header_hard(skb, hlen)) {
2932                 th = skb_gro_header_slow(skb, hlen, off);
2933                 if (unlikely(!th))
2934                         goto out;
2935         }
2936
2937         thlen = th->doff * 4;
2938         if (thlen < sizeof(*th))
2939                 goto out;
2940
2941         hlen = off + thlen;
2942         if (skb_gro_header_hard(skb, hlen)) {
2943                 th = skb_gro_header_slow(skb, hlen, off);
2944                 if (unlikely(!th))
2945                         goto out;
2946         }
2947
2948         skb_gro_pull(skb, thlen);
2949
2950         len = skb_gro_len(skb);
2951         flags = tcp_flag_word(th);
2952
2953         for (; (p = *head); head = &p->next) {
2954                 if (!NAPI_GRO_CB(p)->same_flow)
2955                         continue;
2956
2957                 th2 = tcp_hdr(p);
2958
2959                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2960                         NAPI_GRO_CB(p)->same_flow = 0;
2961                         continue;
2962                 }
2963
2964                 goto found;
2965         }
2966
2967         goto out_check_final;
2968
2969 found:
2970         flush = NAPI_GRO_CB(p)->flush;
2971         flush |= (__force int)(flags & TCP_FLAG_CWR);
2972         flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
2973                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
2974         flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
2975         for (i = sizeof(*th); i < thlen; i += 4)
2976                 flush |= *(u32 *)((u8 *)th + i) ^
2977                          *(u32 *)((u8 *)th2 + i);
2978
2979         mss = skb_shinfo(p)->gso_size;
2980
2981         flush |= (len - 1) >= mss;
2982         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2983
2984         if (flush || skb_gro_receive(head, skb)) {
2985                 mss = 1;
2986                 goto out_check_final;
2987         }
2988
2989         p = *head;
2990         th2 = tcp_hdr(p);
2991         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2992
2993 out_check_final:
2994         flush = len < mss;
2995         flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
2996                                         TCP_FLAG_RST | TCP_FLAG_SYN |
2997                                         TCP_FLAG_FIN));
2998
2999         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
3000                 pp = head;
3001
3002 out:
3003         NAPI_GRO_CB(skb)->flush |= flush;
3004
3005         return pp;
3006 }
3007 EXPORT_SYMBOL(tcp_gro_receive);
3008
3009 int tcp_gro_complete(struct sk_buff *skb)
3010 {
3011         struct tcphdr *th = tcp_hdr(skb);
3012
3013         skb->csum_start = skb_transport_header(skb) - skb->head;
3014         skb->csum_offset = offsetof(struct tcphdr, check);
3015         skb->ip_summed = CHECKSUM_PARTIAL;
3016
3017         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
3018
3019         if (th->cwr)
3020                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
3021
3022         return 0;
3023 }
3024 EXPORT_SYMBOL(tcp_gro_complete);
3025
3026 #ifdef CONFIG_TCP_MD5SIG
3027 static unsigned long tcp_md5sig_users;
3028 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool;
3029 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
3030
3031 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
3032 {
3033         int cpu;
3034
3035         for_each_possible_cpu(cpu) {
3036                 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
3037
3038                 if (p->md5_desc.tfm)
3039                         crypto_free_hash(p->md5_desc.tfm);
3040         }
3041         free_percpu(pool);
3042 }
3043
3044 void tcp_free_md5sig_pool(void)
3045 {
3046         struct tcp_md5sig_pool __percpu *pool = NULL;
3047
3048         spin_lock_bh(&tcp_md5sig_pool_lock);
3049         if (--tcp_md5sig_users == 0) {
3050                 pool = tcp_md5sig_pool;
3051                 tcp_md5sig_pool = NULL;
3052         }
3053         spin_unlock_bh(&tcp_md5sig_pool_lock);
3054         if (pool)
3055                 __tcp_free_md5sig_pool(pool);
3056 }
3057 EXPORT_SYMBOL(tcp_free_md5sig_pool);
3058
3059 static struct tcp_md5sig_pool __percpu *
3060 __tcp_alloc_md5sig_pool(struct sock *sk)
3061 {
3062         int cpu;
3063         struct tcp_md5sig_pool __percpu *pool;
3064
3065         pool = alloc_percpu(struct tcp_md5sig_pool);
3066         if (!pool)
3067                 return NULL;
3068
3069         for_each_possible_cpu(cpu) {
3070                 struct crypto_hash *hash;
3071
3072                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
3073                 if (!hash || IS_ERR(hash))
3074                         goto out_free;
3075
3076                 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
3077         }
3078         return pool;
3079 out_free:
3080         __tcp_free_md5sig_pool(pool);
3081         return NULL;
3082 }
3083
3084 struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
3085 {
3086         struct tcp_md5sig_pool __percpu *pool;
3087         int alloc = 0;
3088
3089 retry:
3090         spin_lock_bh(&tcp_md5sig_pool_lock);
3091         pool = tcp_md5sig_pool;
3092         if (tcp_md5sig_users++ == 0) {
3093                 alloc = 1;
3094                 spin_unlock_bh(&tcp_md5sig_pool_lock);
3095         } else if (!pool) {
3096                 tcp_md5sig_users--;
3097                 spin_unlock_bh(&tcp_md5sig_pool_lock);
3098                 cpu_relax();
3099                 goto retry;
3100         } else
3101                 spin_unlock_bh(&tcp_md5sig_pool_lock);
3102
3103         if (alloc) {
3104                 /* we cannot hold spinlock here because this may sleep. */
3105                 struct tcp_md5sig_pool __percpu *p;
3106
3107                 p = __tcp_alloc_md5sig_pool(sk);
3108                 spin_lock_bh(&tcp_md5sig_pool_lock);
3109                 if (!p) {
3110                         tcp_md5sig_users--;
3111                         spin_unlock_bh(&tcp_md5sig_pool_lock);
3112                         return NULL;
3113                 }
3114                 pool = tcp_md5sig_pool;
3115                 if (pool) {
3116                         /* oops, it has already been assigned. */
3117                         spin_unlock_bh(&tcp_md5sig_pool_lock);
3118                         __tcp_free_md5sig_pool(p);
3119                 } else {
3120                         tcp_md5sig_pool = pool = p;
3121                         spin_unlock_bh(&tcp_md5sig_pool_lock);
3122                 }
3123         }
3124         return pool;
3125 }
3126 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3127
3128
3129 /**
3130  *      tcp_get_md5sig_pool - get md5sig_pool for this user
3131  *
3132  *      We use percpu structure, so if we succeed, we exit with preemption
3133  *      and BH disabled, to make sure another thread or softirq handling
3134  *      wont try to get same context.
3135  */
3136 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3137 {
3138         struct tcp_md5sig_pool __percpu *p;
3139
3140         local_bh_disable();
3141
3142         spin_lock(&tcp_md5sig_pool_lock);
3143         p = tcp_md5sig_pool;
3144         if (p)
3145                 tcp_md5sig_users++;
3146         spin_unlock(&tcp_md5sig_pool_lock);
3147
3148         if (p)
3149                 return this_cpu_ptr(p);
3150
3151         local_bh_enable();
3152         return NULL;
3153 }
3154 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3155
3156 void tcp_put_md5sig_pool(void)
3157 {
3158         local_bh_enable();
3159         tcp_free_md5sig_pool();
3160 }
3161 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3162
3163 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3164                         const struct tcphdr *th)
3165 {
3166         struct scatterlist sg;
3167         struct tcphdr hdr;
3168         int err;
3169
3170         /* We are not allowed to change tcphdr, make a local copy */
3171         memcpy(&hdr, th, sizeof(hdr));
3172         hdr.check = 0;
3173
3174         /* options aren't included in the hash */
3175         sg_init_one(&sg, &hdr, sizeof(hdr));
3176         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3177         return err;
3178 }
3179 EXPORT_SYMBOL(tcp_md5_hash_header);
3180
3181 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3182                           const struct sk_buff *skb, unsigned int header_len)
3183 {
3184         struct scatterlist sg;
3185         const struct tcphdr *tp = tcp_hdr(skb);
3186         struct hash_desc *desc = &hp->md5_desc;
3187         unsigned int i;
3188         const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3189                                            skb_headlen(skb) - header_len : 0;
3190         const struct skb_shared_info *shi = skb_shinfo(skb);
3191         struct sk_buff *frag_iter;
3192
3193         sg_init_table(&sg, 1);
3194
3195         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3196         if (crypto_hash_update(desc, &sg, head_data_len))
3197                 return 1;
3198
3199         for (i = 0; i < shi->nr_frags; ++i) {
3200                 const struct skb_frag_struct *f = &shi->frags[i];
3201                 struct page *page = skb_frag_page(f);
3202                 sg_set_page(&sg, page, skb_frag_size(f), f->page_offset);
3203                 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3204                         return 1;
3205         }
3206
3207         skb_walk_frags(skb, frag_iter)
3208                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3209                         return 1;
3210
3211         return 0;
3212 }
3213 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3214
3215 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3216 {
3217         struct scatterlist sg;
3218
3219         sg_init_one(&sg, key->key, key->keylen);
3220         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3221 }
3222 EXPORT_SYMBOL(tcp_md5_hash_key);
3223
3224 #endif
3225
3226 /**
3227  * Each Responder maintains up to two secret values concurrently for
3228  * efficient secret rollover.  Each secret value has 4 states:
3229  *
3230  * Generating.  (tcp_secret_generating != tcp_secret_primary)
3231  *    Generates new Responder-Cookies, but not yet used for primary
3232  *    verification.  This is a short-term state, typically lasting only
3233  *    one round trip time (RTT).
3234  *
3235  * Primary.  (tcp_secret_generating == tcp_secret_primary)
3236  *    Used both for generation and primary verification.
3237  *
3238  * Retiring.  (tcp_secret_retiring != tcp_secret_secondary)
3239  *    Used for verification, until the first failure that can be
3240  *    verified by the newer Generating secret.  At that time, this
3241  *    cookie's state is changed to Secondary, and the Generating
3242  *    cookie's state is changed to Primary.  This is a short-term state,
3243  *    typically lasting only one round trip time (RTT).
3244  *
3245  * Secondary.  (tcp_secret_retiring == tcp_secret_secondary)
3246  *    Used for secondary verification, after primary verification
3247  *    failures.  This state lasts no more than twice the Maximum Segment
3248  *    Lifetime (2MSL).  Then, the secret is discarded.
3249  */
3250 struct tcp_cookie_secret {
3251         /* The secret is divided into two parts.  The digest part is the
3252          * equivalent of previously hashing a secret and saving the state,
3253          * and serves as an initialization vector (IV).  The message part
3254          * serves as the trailing secret.
3255          */
3256         u32                             secrets[COOKIE_WORKSPACE_WORDS];
3257         unsigned long                   expires;
3258 };
3259
3260 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3261 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3262 #define TCP_SECRET_LIFE (HZ * 600)
3263
3264 static struct tcp_cookie_secret tcp_secret_one;
3265 static struct tcp_cookie_secret tcp_secret_two;
3266
3267 /* Essentially a circular list, without dynamic allocation. */
3268 static struct tcp_cookie_secret *tcp_secret_generating;
3269 static struct tcp_cookie_secret *tcp_secret_primary;
3270 static struct tcp_cookie_secret *tcp_secret_retiring;
3271 static struct tcp_cookie_secret *tcp_secret_secondary;
3272
3273 static DEFINE_SPINLOCK(tcp_secret_locker);
3274
3275 /* Select a pseudo-random word in the cookie workspace.
3276  */
3277 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3278 {
3279         return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3280 }
3281
3282 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3283  * Called in softirq context.
3284  * Returns: 0 for success.
3285  */
3286 int tcp_cookie_generator(u32 *bakery)
3287 {
3288         unsigned long jiffy = jiffies;
3289
3290         if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3291                 spin_lock_bh(&tcp_secret_locker);
3292                 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3293                         /* refreshed by another */
3294                         memcpy(bakery,
3295                                &tcp_secret_generating->secrets[0],
3296                                COOKIE_WORKSPACE_WORDS);
3297                 } else {
3298                         /* still needs refreshing */
3299                         get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3300
3301                         /* The first time, paranoia assumes that the
3302                          * randomization function isn't as strong.  But,
3303                          * this secret initialization is delayed until
3304                          * the last possible moment (packet arrival).
3305                          * Although that time is observable, it is
3306                          * unpredictably variable.  Mash in the most
3307                          * volatile clock bits available, and expire the
3308                          * secret extra quickly.
3309                          */
3310                         if (unlikely(tcp_secret_primary->expires ==
3311                                      tcp_secret_secondary->expires)) {
3312                                 struct timespec tv;
3313
3314                                 getnstimeofday(&tv);
3315                                 bakery[COOKIE_DIGEST_WORDS+0] ^=
3316                                         (u32)tv.tv_nsec;
3317
3318                                 tcp_secret_secondary->expires = jiffy
3319                                         + TCP_SECRET_1MSL
3320                                         + (0x0f & tcp_cookie_work(bakery, 0));
3321                         } else {
3322                                 tcp_secret_secondary->expires = jiffy
3323                                         + TCP_SECRET_LIFE
3324                                         + (0xff & tcp_cookie_work(bakery, 1));
3325                                 tcp_secret_primary->expires = jiffy
3326                                         + TCP_SECRET_2MSL
3327                                         + (0x1f & tcp_cookie_work(bakery, 2));
3328                         }
3329                         memcpy(&tcp_secret_secondary->secrets[0],
3330                                bakery, COOKIE_WORKSPACE_WORDS);
3331
3332                         rcu_assign_pointer(tcp_secret_generating,
3333                                            tcp_secret_secondary);
3334                         rcu_assign_pointer(tcp_secret_retiring,
3335                                            tcp_secret_primary);
3336                         /*
3337                          * Neither call_rcu() nor synchronize_rcu() needed.
3338                          * Retiring data is not freed.  It is replaced after
3339                          * further (locked) pointer updates, and a quiet time
3340                          * (minimum 1MSL, maximum LIFE - 2MSL).
3341                          */
3342                 }
3343                 spin_unlock_bh(&tcp_secret_locker);
3344         } else {
3345                 rcu_read_lock_bh();
3346                 memcpy(bakery,
3347                        &rcu_dereference(tcp_secret_generating)->secrets[0],
3348                        COOKIE_WORKSPACE_WORDS);
3349                 rcu_read_unlock_bh();
3350         }
3351         return 0;
3352 }
3353 EXPORT_SYMBOL(tcp_cookie_generator);
3354
3355 void tcp_done(struct sock *sk)
3356 {
3357         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3358                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3359
3360         tcp_set_state(sk, TCP_CLOSE);
3361         tcp_clear_xmit_timers(sk);
3362
3363         sk->sk_shutdown = SHUTDOWN_MASK;
3364
3365         if (!sock_flag(sk, SOCK_DEAD))
3366                 sk->sk_state_change(sk);
3367         else
3368                 inet_csk_destroy_sock(sk);
3369 }
3370 EXPORT_SYMBOL_GPL(tcp_done);
3371
3372 extern struct tcp_congestion_ops tcp_reno;
3373
3374 static __initdata unsigned long thash_entries;
3375 static int __init set_thash_entries(char *str)
3376 {
3377         if (!str)
3378                 return 0;
3379         thash_entries = simple_strtoul(str, &str, 0);
3380         return 1;
3381 }
3382 __setup("thash_entries=", set_thash_entries);
3383
3384 void tcp_init_mem(struct net *net)
3385 {
3386         unsigned long limit = nr_free_buffer_pages() / 8;
3387         limit = max(limit, 128UL);
3388         net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3;
3389         net->ipv4.sysctl_tcp_mem[1] = limit;
3390         net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2;
3391 }
3392
3393 void __init tcp_init(void)
3394 {
3395         struct sk_buff *skb = NULL;
3396         unsigned long limit;
3397         int max_share, cnt;
3398         unsigned int i;
3399         unsigned long jiffy = jiffies;
3400
3401         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3402
3403         percpu_counter_init(&tcp_sockets_allocated, 0);
3404         percpu_counter_init(&tcp_orphan_count, 0);
3405         tcp_hashinfo.bind_bucket_cachep =
3406                 kmem_cache_create("tcp_bind_bucket",
3407                                   sizeof(struct inet_bind_bucket), 0,
3408                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3409
3410         /* Size and allocate the main established and bind bucket
3411          * hash tables.
3412          *
3413          * The methodology is similar to that of the buffer cache.
3414          */
3415         tcp_hashinfo.ehash =
3416                 alloc_large_system_hash("TCP established",
3417                                         sizeof(struct inet_ehash_bucket),
3418                                         thash_entries,
3419                                         (totalram_pages >= 128 * 1024) ?
3420                                         13 : 15,
3421                                         0,
3422                                         NULL,
3423                                         &tcp_hashinfo.ehash_mask,
3424                                         thash_entries ? 0 : 512 * 1024);
3425         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3426                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3427                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3428         }
3429         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3430                 panic("TCP: failed to alloc ehash_locks");
3431         tcp_hashinfo.bhash =
3432                 alloc_large_system_hash("TCP bind",
3433                                         sizeof(struct inet_bind_hashbucket),
3434                                         tcp_hashinfo.ehash_mask + 1,
3435                                         (totalram_pages >= 128 * 1024) ?
3436                                         13 : 15,
3437                                         0,
3438                                         &tcp_hashinfo.bhash_size,
3439                                         NULL,
3440                                         64 * 1024);
3441         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3442         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3443                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3444                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3445         }
3446
3447
3448         cnt = tcp_hashinfo.ehash_mask + 1;
3449
3450         tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3451         sysctl_tcp_max_orphans = cnt / 2;
3452         sysctl_max_syn_backlog = max(128, cnt / 256);
3453
3454         tcp_init_mem(&init_net);
3455         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3456         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3457         max_share = min(4UL*1024*1024, limit);
3458
3459         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3460         sysctl_tcp_wmem[1] = 16*1024;
3461         sysctl_tcp_wmem[2] = max(64*1024, max_share);
3462
3463         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3464         sysctl_tcp_rmem[1] = 87380;
3465         sysctl_tcp_rmem[2] = max(87380, max_share);
3466
3467         pr_info("Hash tables configured (established %u bind %u)\n",
3468                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3469
3470         tcp_register_congestion_control(&tcp_reno);
3471
3472         memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3473         memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3474         tcp_secret_one.expires = jiffy; /* past due */
3475         tcp_secret_two.expires = jiffy; /* past due */
3476         tcp_secret_generating = &tcp_secret_one;
3477         tcp_secret_primary = &tcp_secret_one;
3478         tcp_secret_retiring = &tcp_secret_two;
3479         tcp_secret_secondary = &tcp_secret_two;
3480 }