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/inet_common.h>
 274 #include <net/tcp.h>
 275 #include <net/xfrm.h>
 276 #include <net/ip.h>
 277 #include <net/sock.h>
 278
 279 #include <asm/uaccess.h>
 280 #include <asm/ioctls.h>
 281 #include <net/busy_poll.h>
 282
 283 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
 284
 285 int sysctl_tcp_min_tso_segs __read_mostly = 2;
 286
 287 int sysctl_tcp_autocorking __read_mostly = 1;
 288
 289 struct percpu_counter tcp_orphan_count;
 290 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 291
 292 long sysctl_tcp_mem[3] __read_mostly;
 293 int sysctl_tcp_wmem[3] __read_mostly;
 294 int sysctl_tcp_rmem[3] __read_mostly;
 295
 296 EXPORT_SYMBOL(sysctl_tcp_mem);
 297 EXPORT_SYMBOL(sysctl_tcp_rmem);
 298 EXPORT_SYMBOL(sysctl_tcp_wmem);
 299
 300 atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
 301 EXPORT_SYMBOL(tcp_memory_allocated);
 302
 303 /*
 304  * Current number of TCP sockets.
 305  */
 306 struct percpu_counter tcp_sockets_allocated;
 307 EXPORT_SYMBOL(tcp_sockets_allocated);
 308
 309 /*
 310  * TCP splice context
 311  */
 312 struct tcp_splice_state {
 313         struct pipe_inode_info *pipe;
 314         size_t len;
 315         unsigned int flags;
 316 };
 317
 318 /*
 319  * Pressure flag: try to collapse.
 320  * Technical note: it is used by multiple contexts non atomically.
 321  * All the __sk_mem_schedule() is of this nature: accounting
 322  * is strict, actions are advisory and have some latency.
 323  */
 324 int tcp_memory_pressure __read_mostly;
 325 EXPORT_SYMBOL(tcp_memory_pressure);
 326
 327 void tcp_enter_memory_pressure(struct sock *sk)
 328 {
 329         if (!tcp_memory_pressure) {
 330                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
 331                 tcp_memory_pressure = 1;
 332         }
 333 }
 334 EXPORT_SYMBOL(tcp_enter_memory_pressure);
 335
 336 /* Convert seconds to retransmits based on initial and max timeout */
 337 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
 338 {
 339         u8 res = 0;
 340
 341         if (seconds > 0) {
 342                 int period = timeout;
 343
 344                 res = 1;
 345                 while (seconds > period && res < 255) {
 346                         res++;
 347                         timeout <<= 1;
 348                         if (timeout > rto_max)
 349                                 timeout = rto_max;
 350                         period += timeout;
 351                 }
 352         }
 353         return res;
 354 }
 355
 356 /* Convert retransmits to seconds based on initial and max timeout */
 357 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
 358 {
 359         int period = 0;
 360
 361         if (retrans > 0) {
 362                 period = timeout;
 363                 while (--retrans) {
 364                         timeout <<= 1;
 365                         if (timeout > rto_max)
 366                                 timeout = rto_max;
 367                         period += timeout;
 368                 }
 369         }
 370         return period;
 371 }
 372
 373 /* Address-family independent initialization for a tcp_sock.
 374  *
 375  * NOTE: A lot of things set to zero explicitly by call to
 376  *       sk_alloc() so need not be done here.
 377  */
 378 void tcp_init_sock(struct sock *sk)
 379 {
 380         struct inet_connection_sock *icsk = inet_csk(sk);
 381         struct tcp_sock *tp = tcp_sk(sk);
 382
 383         __skb_queue_head_init(&tp->out_of_order_queue);
 384         tcp_init_xmit_timers(sk);
 385         tcp_prequeue_init(tp);
 386         INIT_LIST_HEAD(&tp->tsq_node);
 387
 388         icsk->icsk_rto = TCP_TIMEOUT_INIT;
 389         tp->mdev = TCP_TIMEOUT_INIT;
 390
 391         /* So many TCP implementations out there (incorrectly) count the
 392          * initial SYN frame in their delayed-ACK and congestion control
 393          * algorithms that we must have the following bandaid to talk
 394          * efficiently to them.  -DaveM
 395          */
 396         tp->snd_cwnd = TCP_INIT_CWND;
 397
 398         /* See draft-stevens-tcpca-spec-01 for discussion of the
 399          * initialization of these values.
 400          */
 401         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 402         tp->snd_cwnd_clamp = ~0;
 403         tp->mss_cache = TCP_MSS_DEFAULT;
 404
 405         tp->reordering = sysctl_tcp_reordering;
 406         tcp_enable_early_retrans(tp);
 407         icsk->icsk_ca_ops = &tcp_init_congestion_ops;
 408
 409         tp->tsoffset = 0;
 410
 411         sk->sk_state = TCP_CLOSE;
 412
 413         sk->sk_write_space = sk_stream_write_space;
 414         sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
 415
 416         icsk->icsk_sync_mss = tcp_sync_mss;
 417
 418         sk->sk_sndbuf = sysctl_tcp_wmem[1];
 419         sk->sk_rcvbuf = sysctl_tcp_rmem[1];
 420
 421         local_bh_disable();
 422         sock_update_memcg(sk);
 423         sk_sockets_allocated_inc(sk);
 424         local_bh_enable();
 425 }
 426 EXPORT_SYMBOL(tcp_init_sock);
 427
 428 /*
 429  *      Wait for a TCP event.
 430  *
 431  *      Note that we don't need to lock the socket, as the upper poll layers
 432  *      take care of normal races (between the test and the event) and we don't
 433  *      go look at any of the socket buffers directly.
 434  */
 435 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 436 {
 437         unsigned int mask;
 438         struct sock *sk = sock->sk;
 439         const struct tcp_sock *tp = tcp_sk(sk);
 440
 441         sock_rps_record_flow(sk);
 442
 443         sock_poll_wait(file, sk_sleep(sk), wait);
 444         if (sk->sk_state == TCP_LISTEN)
 445                 return inet_csk_listen_poll(sk);
 446
 447         /* Socket is not locked. We are protected from async events
 448          * by poll logic and correct handling of state changes
 449          * made by other threads is impossible in any case.
 450          */
 451
 452         mask = 0;
 453
 454         /*
 455          * POLLHUP is certainly not done right. But poll() doesn't
 456          * have a notion of HUP in just one direction, and for a
 457          * socket the read side is more interesting.
 458          *
 459          * Some poll() documentation says that POLLHUP is incompatible
 460          * with the POLLOUT/POLLWR flags, so somebody should check this
 461          * all. But careful, it tends to be safer to return too many
 462          * bits than too few, and you can easily break real applications
 463          * if you don't tell them that something has hung up!
 464          *
 465          * Check-me.
 466          *
 467          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
 468          * our fs/select.c). It means that after we received EOF,
 469          * poll always returns immediately, making impossible poll() on write()
 470          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
 471          * if and only if shutdown has been made in both directions.
 472          * Actually, it is interesting to look how Solaris and DUX
 473          * solve this dilemma. I would prefer, if POLLHUP were maskable,
 474          * then we could set it on SND_SHUTDOWN. BTW examples given
 475          * in Stevens' books assume exactly this behaviour, it explains
 476          * why POLLHUP is incompatible with POLLOUT.    --ANK
 477          *
 478          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 479          * blocking on fresh not-connected or disconnected socket. --ANK
 480          */
 481         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
 482                 mask |= POLLHUP;
 483         if (sk->sk_shutdown & RCV_SHUTDOWN)
 484                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
 485
 486         /* Connected or passive Fast Open socket? */
 487         if (sk->sk_state != TCP_SYN_SENT &&
 488             (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk != NULL)) {
 489                 int target = sock_rcvlowat(sk, 0, INT_MAX);
 490
 491                 if (tp->urg_seq == tp->copied_seq &&
 492                     !sock_flag(sk, SOCK_URGINLINE) &&
 493                     tp->urg_data)
 494                         target++;
 495
 496                 /* Potential race condition. If read of tp below will
 497                  * escape above sk->sk_state, we can be illegally awaken
 498                  * in SYN_* states. */
 499                 if (tp->rcv_nxt - tp->copied_seq >= target)
 500                         mask |= POLLIN | POLLRDNORM;
 501
 502                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 503                         if (sk_stream_is_writeable(sk)) {
 504                                 mask |= POLLOUT | POLLWRNORM;
 505                         } else {  /* send SIGIO later */
 506                                 set_bit(SOCK_ASYNC_NOSPACE,
 507                                         &sk->sk_socket->flags);
 508                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 509
 510                                 /* Race breaker. If space is freed after
 511                                  * wspace test but before the flags are set,
 512                                  * IO signal will be lost.
 513                                  */
 514                                 if (sk_stream_is_writeable(sk))
 515                                         mask |= POLLOUT | POLLWRNORM;
 516                         }
 517                 } else
 518                         mask |= POLLOUT | POLLWRNORM;
 519
 520                 if (tp->urg_data & TCP_URG_VALID)
 521                         mask |= POLLPRI;
 522         }
 523         /* This barrier is coupled with smp_wmb() in tcp_reset() */
 524         smp_rmb();
 525         if (sk->sk_err)
 526                 mask |= POLLERR;
 527
 528         return mask;
 529 }
 530 EXPORT_SYMBOL(tcp_poll);
 531
 532 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 533 {
 534         struct tcp_sock *tp = tcp_sk(sk);
 535         int answ;
 536         bool slow;
 537
 538         switch (cmd) {
 539         case SIOCINQ:
 540                 if (sk->sk_state == TCP_LISTEN)
 541                         return -EINVAL;
 542
 543                 slow = lock_sock_fast(sk);
 544                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 545                         answ = 0;
 546                 else if (sock_flag(sk, SOCK_URGINLINE) ||
 547                          !tp->urg_data ||
 548                          before(tp->urg_seq, tp->copied_seq) ||
 549                          !before(tp->urg_seq, tp->rcv_nxt)) {
 550
 551                         answ = tp->rcv_nxt - tp->copied_seq;
 552
 553                         /* Subtract 1, if FIN was received */
 554                         if (answ && sock_flag(sk, SOCK_DONE))
 555                                 answ--;
 556                 } else
 557                         answ = tp->urg_seq - tp->copied_seq;
 558                 unlock_sock_fast(sk, slow);
 559                 break;
 560         case SIOCATMARK:
 561                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 562                 break;
 563         case SIOCOUTQ:
 564                 if (sk->sk_state == TCP_LISTEN)
 565                         return -EINVAL;
 566
 567                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 568                         answ = 0;
 569                 else
 570                         answ = tp->write_seq - tp->snd_una;
 571                 break;
 572         case SIOCOUTQNSD:
 573                 if (sk->sk_state == TCP_LISTEN)
 574                         return -EINVAL;
 575
 576                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 577                         answ = 0;
 578                 else
 579                         answ = tp->write_seq - tp->snd_nxt;
 580                 break;
 581         default:
 582                 return -ENOIOCTLCMD;
 583         }
 584
 585         return put_user(answ, (int __user *)arg);
 586 }
 587 EXPORT_SYMBOL(tcp_ioctl);
 588
 589 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 590 {
 591         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 592         tp->pushed_seq = tp->write_seq;
 593 }
 594
 595 static inline bool forced_push(const struct tcp_sock *tp)
 596 {
 597         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 598 }
 599
 600 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
 601 {
 602         struct tcp_sock *tp = tcp_sk(sk);
 603         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 604
 605         skb->csum    = 0;
 606         tcb->seq     = tcb->end_seq = tp->write_seq;
 607         tcb->tcp_flags = TCPHDR_ACK;
 608         tcb->sacked  = 0;
 609         skb_header_release(skb);
 610         tcp_add_write_queue_tail(sk, skb);
 611         sk->sk_wmem_queued += skb->truesize;
 612         sk_mem_charge(sk, skb->truesize);
 613         if (tp->nonagle & TCP_NAGLE_PUSH)
 614                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 615 }
 616
 617 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
 618 {
 619         if (flags & MSG_OOB)
 620                 tp->snd_up = tp->write_seq;
 621 }
 622
 623 /* If a not yet filled skb is pushed, do not send it if
 624  * we have data packets in Qdisc or NIC queues :
 625  * Because TX completion will happen shortly, it gives a chance
 626  * to coalesce future sendmsg() payload into this skb, without
 627  * need for a timer, and with no latency trade off.
 628  * As packets containing data payload have a bigger truesize
 629  * than pure acks (dataless) packets, the last checks prevent
 630  * autocorking if we only have an ACK in Qdisc/NIC queues,
 631  * or if TX completion was delayed after we processed ACK packet.
 632  */
 633 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
 634                                 int size_goal)
 635 {
 636         return skb->len < size_goal &&
 637                sysctl_tcp_autocorking &&
 638                skb != tcp_write_queue_head(sk) &&
 639                atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
 640 }
 641
 642 static void tcp_push(struct sock *sk, int flags, int mss_now,
 643                      int nonagle, int size_goal)
 644 {
 645         struct tcp_sock *tp = tcp_sk(sk);
 646         struct sk_buff *skb;
 647
 648         if (!tcp_send_head(sk))
 649                 return;
 650
 651         skb = tcp_write_queue_tail(sk);
 652         if (!(flags & MSG_MORE) || forced_push(tp))
 653                 tcp_mark_push(tp, skb);
 654
 655         tcp_mark_urg(tp, flags);
 656
 657         if (tcp_should_autocork(sk, skb, size_goal)) {
 658
 659                 /* avoid atomic op if TSQ_THROTTLED bit is already set */
 660                 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
 661                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
 662                         set_bit(TSQ_THROTTLED, &tp->tsq_flags);
 663                 }
 664                 /* It is possible TX completion already happened
 665                  * before we set TSQ_THROTTLED.
 666                  */
 667                 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
 668                         return;
 669         }
 670
 671         if (flags & MSG_MORE)
 672                 nonagle = TCP_NAGLE_CORK;
 673
 674         __tcp_push_pending_frames(sk, mss_now, nonagle);
 675 }
 676
 677 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 678                                 unsigned int offset, size_t len)
 679 {
 680         struct tcp_splice_state *tss = rd_desc->arg.data;
 681         int ret;
 682
 683         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
 684                               tss->flags);
 685         if (ret > 0)
 686                 rd_desc->count -= ret;
 687         return ret;
 688 }
 689
 690 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 691 {
 692         /* Store TCP splice context information in read_descriptor_t. */
 693         read_descriptor_t rd_desc = {
 694                 .arg.data = tss,
 695                 .count    = tss->len,
 696         };
 697
 698         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 699 }
 700
 701 /**
 702  *  tcp_splice_read - splice data from TCP socket to a pipe
 703  * @sock:       socket to splice from
 704  * @ppos:       position (not valid)
 705  * @pipe:       pipe to splice to
 706  * @len:        number of bytes to splice
 707  * @flags:      splice modifier flags
 708  *
 709  * Description:
 710  *    Will read pages from given socket and fill them into a pipe.
 711  *
 712  **/
 713 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 714                         struct pipe_inode_info *pipe, size_t len,
 715                         unsigned int flags)
 716 {
 717         struct sock *sk = sock->sk;
 718         struct tcp_splice_state tss = {
 719                 .pipe = pipe,
 720                 .len = len,
 721                 .flags = flags,
 722         };
 723         long timeo;
 724         ssize_t spliced;
 725         int ret;
 726
 727         sock_rps_record_flow(sk);
 728         /*
 729          * We can't seek on a socket input
 730          */
 731         if (unlikely(*ppos))
 732                 return -ESPIPE;
 733
 734         ret = spliced = 0;
 735
 736         lock_sock(sk);
 737
 738         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 739         while (tss.len) {
 740                 ret = __tcp_splice_read(sk, &tss);
 741                 if (ret < 0)
 742                         break;
 743                 else if (!ret) {
 744                         if (spliced)
 745                                 break;
 746                         if (sock_flag(sk, SOCK_DONE))
 747                                 break;
 748                         if (sk->sk_err) {
 749                                 ret = sock_error(sk);
 750                                 break;
 751                         }
 752                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 753                                 break;
 754                         if (sk->sk_state == TCP_CLOSE) {
 755                                 /*
 756                                  * This occurs when user tries to read
 757                                  * from never connected socket.
 758                                  */
 759                                 if (!sock_flag(sk, SOCK_DONE))
 760                                         ret = -ENOTCONN;
 761                                 break;
 762                         }
 763                         if (!timeo) {
 764                                 ret = -EAGAIN;
 765                                 break;
 766                         }
 767                         sk_wait_data(sk, &timeo);
 768                         if (signal_pending(current)) {
 769                                 ret = sock_intr_errno(timeo);
 770                                 break;
 771                         }
 772                         continue;
 773                 }
 774                 tss.len -= ret;
 775                 spliced += ret;
 776
 777                 if (!timeo)
 778                         break;
 779                 release_sock(sk);
 780                 lock_sock(sk);
 781
 782                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 783                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 784                     signal_pending(current))
 785                         break;
 786         }
 787
 788         release_sock(sk);
 789
 790         if (spliced)
 791                 return spliced;
 792
 793         return ret;
 794 }
 795 EXPORT_SYMBOL(tcp_splice_read);
 796
 797 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
 798 {
 799         struct sk_buff *skb;
 800
 801         /* The TCP header must be at least 32-bit aligned.  */
 802         size = ALIGN(size, 4);
 803
 804         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 805         if (skb) {
 806                 if (sk_wmem_schedule(sk, skb->truesize)) {
 807                         skb_reserve(skb, sk->sk_prot->max_header);
 808                         /*
 809                          * Make sure that we have exactly size bytes
 810                          * available to the caller, no more, no less.
 811                          */
 812                         skb->reserved_tailroom = skb->end - skb->tail - size;
 813                         return skb;
 814                 }
 815                 __kfree_skb(skb);
 816         } else {
 817                 sk->sk_prot->enter_memory_pressure(sk);
 818                 sk_stream_moderate_sndbuf(sk);
 819         }
 820         return NULL;
 821 }
 822
 823 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 824                                        int large_allowed)
 825 {
 826         struct tcp_sock *tp = tcp_sk(sk);
 827         u32 xmit_size_goal, old_size_goal;
 828
 829         xmit_size_goal = mss_now;
 830
 831         if (large_allowed && sk_can_gso(sk)) {
 832                 u32 gso_size, hlen;
 833
 834                 /* Maybe we should/could use sk->sk_prot->max_header here ? */
 835                 hlen = inet_csk(sk)->icsk_af_ops->net_header_len +
 836                        inet_csk(sk)->icsk_ext_hdr_len +
 837                        tp->tcp_header_len;
 838
 839                 /* Goal is to send at least one packet per ms,
 840                  * not one big TSO packet every 100 ms.
 841                  * This preserves ACK clocking and is consistent
 842                  * with tcp_tso_should_defer() heuristic.
 843                  */
 844                 gso_size = sk->sk_pacing_rate / (2 * MSEC_PER_SEC);
 845                 gso_size = max_t(u32, gso_size,
 846                                  sysctl_tcp_min_tso_segs * mss_now);
 847
 848                 xmit_size_goal = min_t(u32, gso_size,
 849                                        sk->sk_gso_max_size - 1 - hlen);
 850
 851                 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
 852
 853                 /* We try hard to avoid divides here */
 854                 old_size_goal = tp->xmit_size_goal_segs * mss_now;
 855
 856                 if (likely(old_size_goal <= xmit_size_goal &&
 857                            old_size_goal + mss_now > xmit_size_goal)) {
 858                         xmit_size_goal = old_size_goal;
 859                 } else {
 860                         tp->xmit_size_goal_segs =
 861                                 min_t(u16, xmit_size_goal / mss_now,
 862                                       sk->sk_gso_max_segs);
 863                         xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
 864                 }
 865         }
 866
 867         return max(xmit_size_goal, mss_now);
 868 }
 869
 870 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 871 {
 872         int mss_now;
 873
 874         mss_now = tcp_current_mss(sk);
 875         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 876
 877         return mss_now;
 878 }
 879
 880 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
 881                                 size_t size, int flags)
 882 {
 883         struct tcp_sock *tp = tcp_sk(sk);
 884         int mss_now, size_goal;
 885         int err;
 886         ssize_t copied;
 887         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 888
 889         /* Wait for a connection to finish. One exception is TCP Fast Open
 890          * (passive side) where data is allowed to be sent before a connection
 891          * is fully established.
 892          */
 893         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
 894             !tcp_passive_fastopen(sk)) {
 895                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 896                         goto out_err;
 897         }
 898
 899         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 900
 901         mss_now = tcp_send_mss(sk, &size_goal, flags);
 902         copied = 0;
 903
 904         err = -EPIPE;
 905         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 906                 goto out_err;
 907
 908         while (size > 0) {
 909                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 910                 int copy, i;
 911                 bool can_coalesce;
 912
 913                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
 914 new_segment:
 915                         if (!sk_stream_memory_free(sk))
 916                                 goto wait_for_sndbuf;
 917
 918                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
 919                         if (!skb)
 920                                 goto wait_for_memory;
 921
 922                         skb_entail(sk, skb);
 923                         copy = size_goal;
 924                 }
 925
 926                 if (copy > size)
 927                         copy = size;
 928
 929                 i = skb_shinfo(skb)->nr_frags;
 930                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 931                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
 932                         tcp_mark_push(tp, skb);
 933                         goto new_segment;
 934                 }
 935                 if (!sk_wmem_schedule(sk, copy))
 936                         goto wait_for_memory;
 937
 938                 if (can_coalesce) {
 939                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
 940                 } else {
 941                         get_page(page);
 942                         skb_fill_page_desc(skb, i, page, offset, copy);
 943                 }
 944                 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
 945
 946                 skb->len += copy;
 947                 skb->data_len += copy;
 948                 skb->truesize += copy;
 949                 sk->sk_wmem_queued += copy;
 950                 sk_mem_charge(sk, copy);
 951                 skb->ip_summed = CHECKSUM_PARTIAL;
 952                 tp->write_seq += copy;
 953                 TCP_SKB_CB(skb)->end_seq += copy;
 954                 skb_shinfo(skb)->gso_segs = 0;
 955
 956                 if (!copied)
 957                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
 958
 959                 copied += copy;
 960                 offset += copy;
 961                 if (!(size -= copy))
 962                         goto out;
 963
 964                 if (skb->len < size_goal || (flags & MSG_OOB))
 965                         continue;
 966
 967                 if (forced_push(tp)) {
 968                         tcp_mark_push(tp, skb);
 969                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
 970                 } else if (skb == tcp_send_head(sk))
 971                         tcp_push_one(sk, mss_now);
 972                 continue;
 973
 974 wait_for_sndbuf:
 975                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 976 wait_for_memory:
 977                 tcp_push(sk, flags & ~MSG_MORE, mss_now,
 978                          TCP_NAGLE_PUSH, size_goal);
 979
 980                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 981                         goto do_error;
 982
 983                 mss_now = tcp_send_mss(sk, &size_goal, flags);
 984         }
 985
 986 out:
 987         if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
 988                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
 989         return copied;
 990
 991 do_error:
 992         if (copied)
 993                 goto out;
 994 out_err:
 995         return sk_stream_error(sk, flags, err);
 996 }
 997
 998 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
 999                  size_t size, int flags)
1000 {
1001         ssize_t res;
1002
1003         if (!(sk->sk_route_caps & NETIF_F_SG) ||
1004             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
1005                 return sock_no_sendpage(sk->sk_socket, page, offset, size,
1006                                         flags);
1007
1008         lock_sock(sk);
1009         res = do_tcp_sendpages(sk, page, offset, size, flags);
1010         release_sock(sk);
1011         return res;
1012 }
1013 EXPORT_SYMBOL(tcp_sendpage);
1014
1015 static inline int select_size(const struct sock *sk, bool sg)
1016 {
1017         const struct tcp_sock *tp = tcp_sk(sk);
1018         int tmp = tp->mss_cache;
1019
1020         if (sg) {
1021                 if (sk_can_gso(sk)) {
1022                         /* Small frames wont use a full page:
1023                          * Payload will immediately follow tcp header.
1024                          */
1025                         tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1026                 } else {
1027                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1028
1029                         if (tmp >= pgbreak &&
1030                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1031                                 tmp = pgbreak;
1032                 }
1033         }
1034
1035         return tmp;
1036 }
1037
1038 void tcp_free_fastopen_req(struct tcp_sock *tp)
1039 {
1040         if (tp->fastopen_req != NULL) {
1041                 kfree(tp->fastopen_req);
1042                 tp->fastopen_req = NULL;
1043         }
1044 }
1045
1046 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1047                                 int *copied, size_t size)
1048 {
1049         struct tcp_sock *tp = tcp_sk(sk);
1050         int err, flags;
1051
1052         if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1053                 return -EOPNOTSUPP;
1054         if (tp->fastopen_req != NULL)
1055                 return -EALREADY; /* Another Fast Open is in progress */
1056
1057         tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1058                                    sk->sk_allocation);
1059         if (unlikely(tp->fastopen_req == NULL))
1060                 return -ENOBUFS;
1061         tp->fastopen_req->data = msg;
1062         tp->fastopen_req->size = size;
1063
1064         flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1065         err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1066                                     msg->msg_namelen, flags);
1067         *copied = tp->fastopen_req->copied;
1068         tcp_free_fastopen_req(tp);
1069         return err;
1070 }
1071
1072 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1073                 size_t size)
1074 {
1075         struct iovec *iov;
1076         struct tcp_sock *tp = tcp_sk(sk);
1077         struct sk_buff *skb;
1078         int iovlen, flags, err, copied = 0;
1079         int mss_now = 0, size_goal, copied_syn = 0, offset = 0;
1080         bool sg;
1081         long timeo;
1082
1083         lock_sock(sk);
1084
1085         flags = msg->msg_flags;
1086         if (flags & MSG_FASTOPEN) {
1087                 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1088                 if (err == -EINPROGRESS && copied_syn > 0)
1089                         goto out;
1090                 else if (err)
1091                         goto out_err;
1092                 offset = copied_syn;
1093         }
1094
1095         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1096
1097         /* Wait for a connection to finish. One exception is TCP Fast Open
1098          * (passive side) where data is allowed to be sent before a connection
1099          * is fully established.
1100          */
1101         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1102             !tcp_passive_fastopen(sk)) {
1103                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1104                         goto do_error;
1105         }
1106
1107         if (unlikely(tp->repair)) {
1108                 if (tp->repair_queue == TCP_RECV_QUEUE) {
1109                         copied = tcp_send_rcvq(sk, msg, size);
1110                         goto out;
1111                 }
1112
1113                 err = -EINVAL;
1114                 if (tp->repair_queue == TCP_NO_QUEUE)
1115                         goto out_err;
1116
1117                 /* 'common' sending to sendq */
1118         }
1119
1120         /* This should be in poll */
1121         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1122
1123         mss_now = tcp_send_mss(sk, &size_goal, flags);
1124
1125         /* Ok commence sending. */
1126         iovlen = msg->msg_iovlen;
1127         iov = msg->msg_iov;
1128         copied = 0;
1129
1130         err = -EPIPE;
1131         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1132                 goto out_err;
1133
1134         sg = !!(sk->sk_route_caps & NETIF_F_SG);
1135
1136         while (--iovlen >= 0) {
1137                 size_t seglen = iov->iov_len;
1138                 unsigned char __user *from = iov->iov_base;
1139
1140                 iov++;
1141                 if (unlikely(offset > 0)) {  /* Skip bytes copied in SYN */
1142                         if (offset >= seglen) {
1143                                 offset -= seglen;
1144                                 continue;
1145                         }
1146                         seglen -= offset;
1147                         from += offset;
1148                         offset = 0;
1149                 }
1150
1151                 while (seglen > 0) {
1152                         int copy = 0;
1153                         int max = size_goal;
1154
1155                         skb = tcp_write_queue_tail(sk);
1156                         if (tcp_send_head(sk)) {
1157                                 if (skb->ip_summed == CHECKSUM_NONE)
1158                                         max = mss_now;
1159                                 copy = max - skb->len;
1160                         }
1161
1162                         if (copy <= 0) {
1163 new_segment:
1164                                 /* Allocate new segment. If the interface is SG,
1165                                  * allocate skb fitting to single page.
1166                                  */
1167                                 if (!sk_stream_memory_free(sk))
1168                                         goto wait_for_sndbuf;
1169
1170                                 skb = sk_stream_alloc_skb(sk,
1171                                                           select_size(sk, sg),
1172                                                           sk->sk_allocation);
1173                                 if (!skb)
1174                                         goto wait_for_memory;
1175
1176                                 /*
1177                                  * All packets are restored as if they have
1178                                  * already been sent.
1179                                  */
1180                                 if (tp->repair)
1181                                         TCP_SKB_CB(skb)->when = tcp_time_stamp;
1182
1183                                 /*
1184                                  * Check whether we can use HW checksum.
1185                                  */
1186                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1187                                         skb->ip_summed = CHECKSUM_PARTIAL;
1188
1189                                 skb_entail(sk, skb);
1190                                 copy = size_goal;
1191                                 max = size_goal;
1192                         }
1193
1194                         /* Try to append data to the end of skb. */
1195                         if (copy > seglen)
1196                                 copy = seglen;
1197
1198                         /* Where to copy to? */
1199                         if (skb_availroom(skb) > 0) {
1200                                 /* We have some space in skb head. Superb! */
1201                                 copy = min_t(int, copy, skb_availroom(skb));
1202                                 err = skb_add_data_nocache(sk, skb, from, copy);
1203                                 if (err)
1204                                         goto do_fault;
1205                         } else {
1206                                 bool merge = true;
1207                                 int i = skb_shinfo(skb)->nr_frags;
1208                                 struct page_frag *pfrag = sk_page_frag(sk);
1209
1210                                 if (!sk_page_frag_refill(sk, pfrag))
1211                                         goto wait_for_memory;
1212
1213                                 if (!skb_can_coalesce(skb, i, pfrag->page,
1214                                                       pfrag->offset)) {
1215                                         if (i == MAX_SKB_FRAGS || !sg) {
1216                                                 tcp_mark_push(tp, skb);
1217                                                 goto new_segment;
1218                                         }
1219                                         merge = false;
1220                                 }
1221
1222                                 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1223
1224                                 if (!sk_wmem_schedule(sk, copy))
1225                                         goto wait_for_memory;
1226
1227                                 err = skb_copy_to_page_nocache(sk, from, skb,
1228                                                                pfrag->page,
1229                                                                pfrag->offset,
1230                                                                copy);
1231                                 if (err)
1232                                         goto do_error;
1233
1234                                 /* Update the skb. */
1235                                 if (merge) {
1236                                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1237                                 } else {
1238                                         skb_fill_page_desc(skb, i, pfrag->page,
1239                                                            pfrag->offset, copy);
1240                                         get_page(pfrag->page);
1241                                 }
1242                                 pfrag->offset += copy;
1243                         }
1244
1245                         if (!copied)
1246                                 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1247
1248                         tp->write_seq += copy;
1249                         TCP_SKB_CB(skb)->end_seq += copy;
1250                         skb_shinfo(skb)->gso_segs = 0;
1251
1252                         from += copy;
1253                         copied += copy;
1254                         if ((seglen -= copy) == 0 && iovlen == 0)
1255                                 goto out;
1256
1257                         if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1258                                 continue;
1259
1260                         if (forced_push(tp)) {
1261                                 tcp_mark_push(tp, skb);
1262                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1263                         } else if (skb == tcp_send_head(sk))
1264                                 tcp_push_one(sk, mss_now);
1265                         continue;
1266
1267 wait_for_sndbuf:
1268                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1269 wait_for_memory:
1270                         if (copied)
1271                                 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1272                                          TCP_NAGLE_PUSH, size_goal);
1273
1274                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1275                                 goto do_error;
1276
1277                         mss_now = tcp_send_mss(sk, &size_goal, flags);
1278                 }
1279         }
1280
1281 out:
1282         if (copied)
1283                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1284         release_sock(sk);
1285         return copied + copied_syn;
1286
1287 do_fault:
1288         if (!skb->len) {
1289                 tcp_unlink_write_queue(skb, sk);
1290                 /* It is the one place in all of TCP, except connection
1291                  * reset, where we can be unlinking the send_head.
1292                  */
1293                 tcp_check_send_head(sk, skb);
1294                 sk_wmem_free_skb(sk, skb);
1295         }
1296
1297 do_error:
1298         if (copied + copied_syn)
1299                 goto out;
1300 out_err:
1301         err = sk_stream_error(sk, flags, err);
1302         release_sock(sk);
1303         return err;
1304 }
1305 EXPORT_SYMBOL(tcp_sendmsg);
1306
1307 /*
1308  *      Handle reading urgent data. BSD has very simple semantics for
1309  *      this, no blocking and very strange errors 8)
1310  */
1311
1312 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1313 {
1314         struct tcp_sock *tp = tcp_sk(sk);
1315
1316         /* No URG data to read. */
1317         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1318             tp->urg_data == TCP_URG_READ)
1319                 return -EINVAL; /* Yes this is right ! */
1320
1321         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1322                 return -ENOTCONN;
1323
1324         if (tp->urg_data & TCP_URG_VALID) {
1325                 int err = 0;
1326                 char c = tp->urg_data;
1327
1328                 if (!(flags & MSG_PEEK))
1329                         tp->urg_data = TCP_URG_READ;
1330
1331                 /* Read urgent data. */
1332                 msg->msg_flags |= MSG_OOB;
1333
1334                 if (len > 0) {
1335                         if (!(flags & MSG_TRUNC))
1336                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1337                         len = 1;
1338                 } else
1339                         msg->msg_flags |= MSG_TRUNC;
1340
1341                 return err ? -EFAULT : len;
1342         }
1343
1344         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1345                 return 0;
1346
1347         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1348          * the available implementations agree in this case:
1349          * this call should never block, independent of the
1350          * blocking state of the socket.
1351          * Mike <pall@rz.uni-karlsruhe.de>
1352          */
1353         return -EAGAIN;
1354 }
1355
1356 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1357 {
1358         struct sk_buff *skb;
1359         int copied = 0, err = 0;
1360
1361         /* XXX -- need to support SO_PEEK_OFF */
1362
1363         skb_queue_walk(&sk->sk_write_queue, skb) {
1364                 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1365                 if (err)
1366                         break;
1367
1368                 copied += skb->len;
1369         }
1370
1371         return err ?: copied;
1372 }
1373
1374 /* Clean up the receive buffer for full frames taken by the user,
1375  * then send an ACK if necessary.  COPIED is the number of bytes
1376  * tcp_recvmsg has given to the user so far, it speeds up the
1377  * calculation of whether or not we must ACK for the sake of
1378  * a window update.
1379  */
1380 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1381 {
1382         struct tcp_sock *tp = tcp_sk(sk);
1383         bool time_to_ack = false;
1384
1385         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1386
1387         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1388              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1389              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1390
1391         if (inet_csk_ack_scheduled(sk)) {
1392                 const struct inet_connection_sock *icsk = inet_csk(sk);
1393                    /* Delayed ACKs frequently hit locked sockets during bulk
1394                     * receive. */
1395                 if (icsk->icsk_ack.blocked ||
1396                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1397                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1398                     /*
1399                      * If this read emptied read buffer, we send ACK, if
1400                      * connection is not bidirectional, user drained
1401                      * receive buffer and there was a small segment
1402                      * in queue.
1403                      */
1404                     (copied > 0 &&
1405                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1406                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1407                        !icsk->icsk_ack.pingpong)) &&
1408                       !atomic_read(&sk->sk_rmem_alloc)))
1409                         time_to_ack = true;
1410         }
1411
1412         /* We send an ACK if we can now advertise a non-zero window
1413          * which has been raised "significantly".
1414          *
1415          * Even if window raised up to infinity, do not send window open ACK
1416          * in states, where we will not receive more. It is useless.
1417          */
1418         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1419                 __u32 rcv_window_now = tcp_receive_window(tp);
1420
1421                 /* Optimize, __tcp_select_window() is not cheap. */
1422                 if (2*rcv_window_now <= tp->window_clamp) {
1423                         __u32 new_window = __tcp_select_window(sk);
1424
1425                         /* Send ACK now, if this read freed lots of space
1426                          * in our buffer. Certainly, new_window is new window.
1427                          * We can advertise it now, if it is not less than current one.
1428                          * "Lots" means "at least twice" here.
1429                          */
1430                         if (new_window && new_window >= 2 * rcv_window_now)
1431                                 time_to_ack = true;
1432                 }
1433         }
1434         if (time_to_ack)
1435                 tcp_send_ack(sk);
1436 }
1437
1438 static void tcp_prequeue_process(struct sock *sk)
1439 {
1440         struct sk_buff *skb;
1441         struct tcp_sock *tp = tcp_sk(sk);
1442
1443         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1444
1445         /* RX process wants to run with disabled BHs, though it is not
1446          * necessary */
1447         local_bh_disable();
1448         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1449                 sk_backlog_rcv(sk, skb);
1450         local_bh_enable();
1451
1452         /* Clear memory counter. */
1453         tp->ucopy.memory = 0;
1454 }
1455
1456 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1457 {
1458         struct sk_buff *skb;
1459         u32 offset;
1460
1461         while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1462                 offset = seq - TCP_SKB_CB(skb)->seq;
1463                 if (tcp_hdr(skb)->syn)
1464                         offset--;
1465                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1466                         *off = offset;
1467                         return skb;
1468                 }
1469                 /* This looks weird, but this can happen if TCP collapsing
1470                  * splitted a fat GRO packet, while we released socket lock
1471                  * in skb_splice_bits()
1472                  */
1473                 sk_eat_skb(sk, skb);
1474         }
1475         return NULL;
1476 }
1477
1478 /*
1479  * This routine provides an alternative to tcp_recvmsg() for routines
1480  * that would like to handle copying from skbuffs directly in 'sendfile'
1481  * fashion.
1482  * Note:
1483  *      - It is assumed that the socket was locked by the caller.
1484  *      - The routine does not block.
1485  *      - At present, there is no support for reading OOB data
1486  *        or for 'peeking' the socket using this routine
1487  *        (although both would be easy to implement).
1488  */
1489 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1490                   sk_read_actor_t recv_actor)
1491 {
1492         struct sk_buff *skb;
1493         struct tcp_sock *tp = tcp_sk(sk);
1494         u32 seq = tp->copied_seq;
1495         u32 offset;
1496         int copied = 0;
1497
1498         if (sk->sk_state == TCP_LISTEN)
1499                 return -ENOTCONN;
1500         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1501                 if (offset < skb->len) {
1502                         int used;
1503                         size_t len;
1504
1505                         len = skb->len - offset;
1506                         /* Stop reading if we hit a patch of urgent data */
1507                         if (tp->urg_data) {
1508                                 u32 urg_offset = tp->urg_seq - seq;
1509                                 if (urg_offset < len)
1510                                         len = urg_offset;
1511                                 if (!len)
1512                                         break;
1513                         }
1514                         used = recv_actor(desc, skb, offset, len);
1515                         if (used <= 0) {
1516                                 if (!copied)
1517                                         copied = used;
1518                                 break;
1519                         } else if (used <= len) {
1520                                 seq += used;
1521                                 copied += used;
1522                                 offset += used;
1523                         }
1524                         /* If recv_actor drops the lock (e.g. TCP splice
1525                          * receive) the skb pointer might be invalid when
1526                          * getting here: tcp_collapse might have deleted it
1527                          * while aggregating skbs from the socket queue.
1528                          */
1529                         skb = tcp_recv_skb(sk, seq - 1, &offset);
1530                         if (!skb)
1531                                 break;
1532                         /* TCP coalescing might have appended data to the skb.
1533                          * Try to splice more frags
1534                          */
1535                         if (offset + 1 != skb->len)
1536                                 continue;
1537                 }
1538                 if (tcp_hdr(skb)->fin) {
1539                         sk_eat_skb(sk, skb);
1540                         ++seq;
1541                         break;
1542                 }
1543                 sk_eat_skb(sk, skb);
1544                 if (!desc->count)
1545                         break;
1546                 tp->copied_seq = seq;
1547         }
1548         tp->copied_seq = seq;
1549
1550         tcp_rcv_space_adjust(sk);
1551
1552         /* Clean up data we have read: This will do ACK frames. */
1553         if (copied > 0) {
1554                 tcp_recv_skb(sk, seq, &offset);
1555                 tcp_cleanup_rbuf(sk, copied);
1556         }
1557         return copied;
1558 }
1559 EXPORT_SYMBOL(tcp_read_sock);
1560
1561 /*
1562  *      This routine copies from a sock struct into the user buffer.
1563  *
1564  *      Technical note: in 2.3 we work on _locked_ socket, so that
1565  *      tricks with *seq access order and skb->users are not required.
1566  *      Probably, code can be easily improved even more.
1567  */
1568
1569 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1570                 size_t len, int nonblock, int flags, int *addr_len)
1571 {
1572         struct tcp_sock *tp = tcp_sk(sk);
1573         int copied = 0;
1574         u32 peek_seq;
1575         u32 *seq;
1576         unsigned long used;
1577         int err;
1578         int target;             /* Read at least this many bytes */
1579         long timeo;
1580         struct task_struct *user_recv = NULL;
1581         struct sk_buff *skb;
1582         u32 urg_hole = 0;
1583
1584         if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1585             (sk->sk_state == TCP_ESTABLISHED))
1586                 sk_busy_loop(sk, nonblock);
1587
1588         lock_sock(sk);
1589
1590         err = -ENOTCONN;
1591         if (sk->sk_state == TCP_LISTEN)
1592                 goto out;
1593
1594         timeo = sock_rcvtimeo(sk, nonblock);
1595
1596         /* Urgent data needs to be handled specially. */
1597         if (flags & MSG_OOB)
1598                 goto recv_urg;
1599
1600         if (unlikely(tp->repair)) {
1601                 err = -EPERM;
1602                 if (!(flags & MSG_PEEK))
1603                         goto out;
1604
1605                 if (tp->repair_queue == TCP_SEND_QUEUE)
1606                         goto recv_sndq;
1607
1608                 err = -EINVAL;
1609                 if (tp->repair_queue == TCP_NO_QUEUE)
1610                         goto out;
1611
1612                 /* 'common' recv queue MSG_PEEK-ing */
1613         }
1614
1615         seq = &tp->copied_seq;
1616         if (flags & MSG_PEEK) {
1617                 peek_seq = tp->copied_seq;
1618                 seq = &peek_seq;
1619         }
1620
1621         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1622
1623         do {
1624                 u32 offset;
1625
1626                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1627                 if (tp->urg_data && tp->urg_seq == *seq) {
1628                         if (copied)
1629                                 break;
1630                         if (signal_pending(current)) {
1631                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1632                                 break;
1633                         }
1634                 }
1635
1636                 /* Next get a buffer. */
1637
1638                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1639                         /* Now that we have two receive queues this
1640                          * shouldn't happen.
1641                          */
1642                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1643                                  "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1644                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1645                                  flags))
1646                                 break;
1647
1648                         offset = *seq - TCP_SKB_CB(skb)->seq;
1649                         if (tcp_hdr(skb)->syn)
1650                                 offset--;
1651                         if (offset < skb->len)
1652                                 goto found_ok_skb;
1653                         if (tcp_hdr(skb)->fin)
1654                                 goto found_fin_ok;
1655                         WARN(!(flags & MSG_PEEK),
1656                              "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1657                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1658                 }
1659
1660                 /* Well, if we have backlog, try to process it now yet. */
1661
1662                 if (copied >= target && !sk->sk_backlog.tail)
1663                         break;
1664
1665                 if (copied) {
1666                         if (sk->sk_err ||
1667                             sk->sk_state == TCP_CLOSE ||
1668                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1669                             !timeo ||
1670                             signal_pending(current))
1671                                 break;
1672                 } else {
1673                         if (sock_flag(sk, SOCK_DONE))
1674                                 break;
1675
1676                         if (sk->sk_err) {
1677                                 copied = sock_error(sk);
1678                                 break;
1679                         }
1680
1681                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1682                                 break;
1683
1684                         if (sk->sk_state == TCP_CLOSE) {
1685                                 if (!sock_flag(sk, SOCK_DONE)) {
1686                                         /* This occurs when user tries to read
1687                                          * from never connected socket.
1688                                          */
1689                                         copied = -ENOTCONN;
1690                                         break;
1691                                 }
1692                                 break;
1693                         }
1694
1695                         if (!timeo) {
1696                                 copied = -EAGAIN;
1697                                 break;
1698                         }
1699
1700                         if (signal_pending(current)) {
1701                                 copied = sock_intr_errno(timeo);
1702                                 break;
1703                         }
1704                 }
1705
1706                 tcp_cleanup_rbuf(sk, copied);
1707
1708                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1709                         /* Install new reader */
1710                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1711                                 user_recv = current;
1712                                 tp->ucopy.task = user_recv;
1713                                 tp->ucopy.iov = msg->msg_iov;
1714                         }
1715
1716                         tp->ucopy.len = len;
1717
1718                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1719                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1720
1721                         /* Ugly... If prequeue is not empty, we have to
1722                          * process it before releasing socket, otherwise
1723                          * order will be broken at second iteration.
1724                          * More elegant solution is required!!!
1725                          *
1726                          * Look: we have the following (pseudo)queues:
1727                          *
1728                          * 1. packets in flight
1729                          * 2. backlog
1730                          * 3. prequeue
1731                          * 4. receive_queue
1732                          *
1733                          * Each queue can be processed only if the next ones
1734                          * are empty. At this point we have empty receive_queue.
1735                          * But prequeue _can_ be not empty after 2nd iteration,
1736                          * when we jumped to start of loop because backlog
1737                          * processing added something to receive_queue.
1738                          * We cannot release_sock(), because backlog contains
1739                          * packets arrived _after_ prequeued ones.
1740                          *
1741                          * Shortly, algorithm is clear --- to process all
1742                          * the queues in order. We could make it more directly,
1743                          * requeueing packets from backlog to prequeue, if
1744                          * is not empty. It is more elegant, but eats cycles,
1745                          * unfortunately.
1746                          */
1747                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1748                                 goto do_prequeue;
1749
1750                         /* __ Set realtime policy in scheduler __ */
1751                 }
1752
1753                 if (copied >= target) {
1754                         /* Do not sleep, just process backlog. */
1755                         release_sock(sk);
1756                         lock_sock(sk);
1757                 } else
1758                         sk_wait_data(sk, &timeo);
1759
1760                 if (user_recv) {
1761                         int chunk;
1762
1763                         /* __ Restore normal policy in scheduler __ */
1764
1765                         if ((chunk = len - tp->ucopy.len) != 0) {
1766                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1767                                 len -= chunk;
1768                                 copied += chunk;
1769                         }
1770
1771                         if (tp->rcv_nxt == tp->copied_seq &&
1772                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1773 do_prequeue:
1774                                 tcp_prequeue_process(sk);
1775
1776                                 if ((chunk = len - tp->ucopy.len) != 0) {
1777                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1778                                         len -= chunk;
1779                                         copied += chunk;
1780                                 }
1781                         }
1782                 }
1783                 if ((flags & MSG_PEEK) &&
1784                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1785                         net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1786                                             current->comm,
1787                                             task_pid_nr(current));
1788                         peek_seq = tp->copied_seq;
1789                 }
1790                 continue;
1791
1792         found_ok_skb:
1793                 /* Ok so how much can we use? */
1794                 used = skb->len - offset;
1795                 if (len < used)
1796                         used = len;
1797
1798                 /* Do we have urgent data here? */
1799                 if (tp->urg_data) {
1800                         u32 urg_offset = tp->urg_seq - *seq;
1801                         if (urg_offset < used) {
1802                                 if (!urg_offset) {
1803                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1804                                                 ++*seq;
1805                                                 urg_hole++;
1806                                                 offset++;
1807                                                 used--;
1808                                                 if (!used)
1809                                                         goto skip_copy;
1810                                         }
1811                                 } else
1812                                         used = urg_offset;
1813                         }
1814                 }
1815
1816                 if (!(flags & MSG_TRUNC)) {
1817                         err = skb_copy_datagram_iovec(skb, offset,
1818                                                       msg->msg_iov, used);
1819                         if (err) {
1820                                 /* Exception. Bailout! */
1821                                 if (!copied)
1822                                         copied = -EFAULT;
1823                                 break;
1824                         }
1825                 }
1826
1827                 *seq += used;
1828                 copied += used;
1829                 len -= used;
1830
1831                 tcp_rcv_space_adjust(sk);
1832
1833 skip_copy:
1834                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1835                         tp->urg_data = 0;
1836                         tcp_fast_path_check(sk);
1837                 }
1838                 if (used + offset < skb->len)
1839                         continue;
1840
1841                 if (tcp_hdr(skb)->fin)
1842                         goto found_fin_ok;
1843                 if (!(flags & MSG_PEEK))
1844                         sk_eat_skb(sk, skb);
1845                 continue;
1846
1847         found_fin_ok:
1848                 /* Process the FIN. */
1849                 ++*seq;
1850                 if (!(flags & MSG_PEEK))
1851                         sk_eat_skb(sk, skb);
1852                 break;
1853         } while (len > 0);
1854
1855         if (user_recv) {
1856                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1857                         int chunk;
1858
1859                         tp->ucopy.len = copied > 0 ? len : 0;
1860
1861                         tcp_prequeue_process(sk);
1862
1863                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1864                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1865                                 len -= chunk;
1866                                 copied += chunk;
1867                         }
1868                 }
1869
1870                 tp->ucopy.task = NULL;
1871                 tp->ucopy.len = 0;
1872         }
1873
1874         /* According to UNIX98, msg_name/msg_namelen are ignored
1875          * on connected socket. I was just happy when found this 8) --ANK
1876          */
1877
1878         /* Clean up data we have read: This will do ACK frames. */
1879         tcp_cleanup_rbuf(sk, copied);
1880
1881         release_sock(sk);
1882         return copied;
1883
1884 out:
1885         release_sock(sk);
1886         return err;
1887
1888 recv_urg:
1889         err = tcp_recv_urg(sk, msg, len, flags);
1890         goto out;
1891
1892 recv_sndq:
1893         err = tcp_peek_sndq(sk, msg, len);
1894         goto out;
1895 }
1896 EXPORT_SYMBOL(tcp_recvmsg);
1897
1898 void tcp_set_state(struct sock *sk, int state)
1899 {
1900         int oldstate = sk->sk_state;
1901
1902         switch (state) {
1903         case TCP_ESTABLISHED:
1904                 if (oldstate != TCP_ESTABLISHED)
1905                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1906                 break;
1907
1908         case TCP_CLOSE:
1909                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1910                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1911
1912                 sk->sk_prot->unhash(sk);
1913                 if (inet_csk(sk)->icsk_bind_hash &&
1914                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1915                         inet_put_port(sk);
1916                 /* fall through */
1917         default:
1918                 if (oldstate == TCP_ESTABLISHED)
1919                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1920         }
1921
1922         /* Change state AFTER socket is unhashed to avoid closed
1923          * socket sitting in hash tables.
1924          */
1925         sk->sk_state = state;
1926
1927 #ifdef STATE_TRACE
1928         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1929 #endif
1930 }
1931 EXPORT_SYMBOL_GPL(tcp_set_state);
1932
1933 /*
1934  *      State processing on a close. This implements the state shift for
1935  *      sending our FIN frame. Note that we only send a FIN for some
1936  *      states. A shutdown() may have already sent the FIN, or we may be
1937  *      closed.
1938  */
1939
1940 static const unsigned char new_state[16] = {
1941   /* current state:        new state:      action:      */
1942   /* (Invalid)          */ TCP_CLOSE,
1943   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1944   /* TCP_SYN_SENT       */ TCP_CLOSE,
1945   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1946   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1947   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1948   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1949   /* TCP_CLOSE          */ TCP_CLOSE,
1950   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1951   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1952   /* TCP_LISTEN         */ TCP_CLOSE,
1953   /* TCP_CLOSING        */ TCP_CLOSING,
1954 };
1955
1956 static int tcp_close_state(struct sock *sk)
1957 {
1958         int next = (int)new_state[sk->sk_state];
1959         int ns = next & TCP_STATE_MASK;
1960
1961         tcp_set_state(sk, ns);
1962
1963         return next & TCP_ACTION_FIN;
1964 }
1965
1966 /*
1967  *      Shutdown the sending side of a connection. Much like close except
1968  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1969  */
1970
1971 void tcp_shutdown(struct sock *sk, int how)
1972 {
1973         /*      We need to grab some memory, and put together a FIN,
1974          *      and then put it into the queue to be sent.
1975          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1976          */
1977         if (!(how & SEND_SHUTDOWN))
1978                 return;
1979
1980         /* If we've already sent a FIN, or it's a closed state, skip this. */
1981         if ((1 << sk->sk_state) &
1982             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1983              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1984                 /* Clear out any half completed packets.  FIN if needed. */
1985                 if (tcp_close_state(sk))
1986                         tcp_send_fin(sk);
1987         }
1988 }
1989 EXPORT_SYMBOL(tcp_shutdown);
1990
1991 bool tcp_check_oom(struct sock *sk, int shift)
1992 {
1993         bool too_many_orphans, out_of_socket_memory;
1994
1995         too_many_orphans = tcp_too_many_orphans(sk, shift);
1996         out_of_socket_memory = tcp_out_of_memory(sk);
1997
1998         if (too_many_orphans)
1999                 net_info_ratelimited("too many orphaned sockets\n");
2000         if (out_of_socket_memory)
2001                 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2002         return too_many_orphans || out_of_socket_memory;
2003 }
2004
2005 void tcp_close(struct sock *sk, long timeout)
2006 {
2007         struct sk_buff *skb;
2008         int data_was_unread = 0;
2009         int state;
2010
2011         lock_sock(sk);
2012         sk->sk_shutdown = SHUTDOWN_MASK;
2013
2014         if (sk->sk_state == TCP_LISTEN) {
2015                 tcp_set_state(sk, TCP_CLOSE);
2016
2017                 /* Special case. */
2018                 inet_csk_listen_stop(sk);
2019
2020                 goto adjudge_to_death;
2021         }
2022
2023         /*  We need to flush the recv. buffs.  We do this only on the
2024          *  descriptor close, not protocol-sourced closes, because the
2025          *  reader process may not have drained the data yet!
2026          */
2027         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2028                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2029                           tcp_hdr(skb)->fin;
2030                 data_was_unread += len;
2031                 __kfree_skb(skb);
2032         }
2033
2034         sk_mem_reclaim(sk);
2035
2036         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2037         if (sk->sk_state == TCP_CLOSE)
2038                 goto adjudge_to_death;
2039
2040         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2041          * data was lost. To witness the awful effects of the old behavior of
2042          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2043          * GET in an FTP client, suspend the process, wait for the client to
2044          * advertise a zero window, then kill -9 the FTP client, wheee...
2045          * Note: timeout is always zero in such a case.
2046          */
2047         if (unlikely(tcp_sk(sk)->repair)) {
2048                 sk->sk_prot->disconnect(sk, 0);
2049         } else if (data_was_unread) {
2050                 /* Unread data was tossed, zap the connection. */
2051                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2052                 tcp_set_state(sk, TCP_CLOSE);
2053                 tcp_send_active_reset(sk, sk->sk_allocation);
2054         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2055                 /* Check zero linger _after_ checking for unread data. */
2056                 sk->sk_prot->disconnect(sk, 0);
2057                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2058         } else if (tcp_close_state(sk)) {
2059                 /* We FIN if the application ate all the data before
2060                  * zapping the connection.
2061                  */
2062
2063                 /* RED-PEN. Formally speaking, we have broken TCP state
2064                  * machine. State transitions:
2065                  *
2066                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2067                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2068                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2069                  *
2070                  * are legal only when FIN has been sent (i.e. in window),
2071                  * rather than queued out of window. Purists blame.
2072                  *
2073                  * F.e. "RFC state" is ESTABLISHED,
2074                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2075                  *
2076                  * The visible declinations are that sometimes
2077                  * we enter time-wait state, when it is not required really
2078                  * (harmless), do not send active resets, when they are
2079                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2080                  * they look as CLOSING or LAST_ACK for Linux)
2081                  * Probably, I missed some more holelets.
2082                  *                                              --ANK
2083                  * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2084                  * in a single packet! (May consider it later but will
2085                  * probably need API support or TCP_CORK SYN-ACK until
2086                  * data is written and socket is closed.)
2087                  */
2088                 tcp_send_fin(sk);
2089         }
2090
2091         sk_stream_wait_close(sk, timeout);
2092
2093 adjudge_to_death:
2094         state = sk->sk_state;
2095         sock_hold(sk);
2096         sock_orphan(sk);
2097
2098         /* It is the last release_sock in its life. It will remove backlog. */
2099         release_sock(sk);
2100
2101
2102         /* Now socket is owned by kernel and we acquire BH lock
2103            to finish close. No need to check for user refs.
2104          */
2105         local_bh_disable();
2106         bh_lock_sock(sk);
2107         WARN_ON(sock_owned_by_user(sk));
2108
2109         percpu_counter_inc(sk->sk_prot->orphan_count);
2110
2111         /* Have we already been destroyed by a softirq or backlog? */
2112         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2113                 goto out;
2114
2115         /*      This is a (useful) BSD violating of the RFC. There is a
2116          *      problem with TCP as specified in that the other end could
2117          *      keep a socket open forever with no application left this end.
2118          *      We use a 1 minute timeout (about the same as BSD) then kill
2119          *      our end. If they send after that then tough - BUT: long enough
2120          *      that we won't make the old 4*rto = almost no time - whoops
2121          *      reset mistake.
2122          *
2123          *      Nope, it was not mistake. It is really desired behaviour
2124          *      f.e. on http servers, when such sockets are useless, but
2125          *      consume significant resources. Let's do it with special
2126          *      linger2 option.                                 --ANK
2127          */
2128
2129         if (sk->sk_state == TCP_FIN_WAIT2) {
2130                 struct tcp_sock *tp = tcp_sk(sk);
2131                 if (tp->linger2 < 0) {
2132                         tcp_set_state(sk, TCP_CLOSE);
2133                         tcp_send_active_reset(sk, GFP_ATOMIC);
2134                         NET_INC_STATS_BH(sock_net(sk),
2135                                         LINUX_MIB_TCPABORTONLINGER);
2136                 } else {
2137                         const int tmo = tcp_fin_time(sk);
2138
2139                         if (tmo > TCP_TIMEWAIT_LEN) {
2140                                 inet_csk_reset_keepalive_timer(sk,
2141                                                 tmo - TCP_TIMEWAIT_LEN);
2142                         } else {
2143                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2144                                 goto out;
2145                         }
2146                 }
2147         }
2148         if (sk->sk_state != TCP_CLOSE) {
2149                 sk_mem_reclaim(sk);
2150                 if (tcp_check_oom(sk, 0)) {
2151                         tcp_set_state(sk, TCP_CLOSE);
2152                         tcp_send_active_reset(sk, GFP_ATOMIC);
2153                         NET_INC_STATS_BH(sock_net(sk),
2154                                         LINUX_MIB_TCPABORTONMEMORY);
2155                 }
2156         }
2157
2158         if (sk->sk_state == TCP_CLOSE) {
2159                 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2160                 /* We could get here with a non-NULL req if the socket is
2161                  * aborted (e.g., closed with unread data) before 3WHS
2162                  * finishes.
2163                  */
2164                 if (req != NULL)
2165                         reqsk_fastopen_remove(sk, req, false);
2166                 inet_csk_destroy_sock(sk);
2167         }
2168         /* Otherwise, socket is reprieved until protocol close. */
2169
2170 out:
2171         bh_unlock_sock(sk);
2172         local_bh_enable();
2173         sock_put(sk);
2174 }
2175 EXPORT_SYMBOL(tcp_close);
2176
2177 /* These states need RST on ABORT according to RFC793 */
2178
2179 static inline bool tcp_need_reset(int state)
2180 {
2181         return (1 << state) &
2182                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2183                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2184 }
2185
2186 int tcp_disconnect(struct sock *sk, int flags)
2187 {
2188         struct inet_sock *inet = inet_sk(sk);
2189         struct inet_connection_sock *icsk = inet_csk(sk);
2190         struct tcp_sock *tp = tcp_sk(sk);
2191         int err = 0;
2192         int old_state = sk->sk_state;
2193
2194         if (old_state != TCP_CLOSE)
2195                 tcp_set_state(sk, TCP_CLOSE);
2196
2197         /* ABORT function of RFC793 */
2198         if (old_state == TCP_LISTEN) {
2199                 inet_csk_listen_stop(sk);
2200         } else if (unlikely(tp->repair)) {
2201                 sk->sk_err = ECONNABORTED;
2202         } else if (tcp_need_reset(old_state) ||
2203                    (tp->snd_nxt != tp->write_seq &&
2204                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2205                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2206                  * states
2207                  */
2208                 tcp_send_active_reset(sk, gfp_any());
2209                 sk->sk_err = ECONNRESET;
2210         } else if (old_state == TCP_SYN_SENT)
2211                 sk->sk_err = ECONNRESET;
2212
2213         tcp_clear_xmit_timers(sk);
2214         __skb_queue_purge(&sk->sk_receive_queue);
2215         tcp_write_queue_purge(sk);
2216         __skb_queue_purge(&tp->out_of_order_queue);
2217
2218         inet->inet_dport = 0;
2219
2220         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2221                 inet_reset_saddr(sk);
2222
2223         sk->sk_shutdown = 0;
2224         sock_reset_flag(sk, SOCK_DONE);
2225         tp->srtt = 0;
2226         if ((tp->write_seq += tp->max_window + 2) == 0)
2227                 tp->write_seq = 1;
2228         icsk->icsk_backoff = 0;
2229         tp->snd_cwnd = 2;
2230         icsk->icsk_probes_out = 0;
2231         tp->packets_out = 0;
2232         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2233         tp->snd_cwnd_cnt = 0;
2234         tp->window_clamp = 0;
2235         tcp_set_ca_state(sk, TCP_CA_Open);
2236         tcp_clear_retrans(tp);
2237         inet_csk_delack_init(sk);
2238         tcp_init_send_head(sk);
2239         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2240         __sk_dst_reset(sk);
2241
2242         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2243
2244         sk->sk_error_report(sk);
2245         return err;
2246 }
2247 EXPORT_SYMBOL(tcp_disconnect);
2248
2249 void tcp_sock_destruct(struct sock *sk)
2250 {
2251         inet_sock_destruct(sk);
2252
2253         kfree(inet_csk(sk)->icsk_accept_queue.fastopenq);
2254 }
2255
2256 static inline bool tcp_can_repair_sock(const struct sock *sk)
2257 {
2258         return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2259                 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2260 }
2261
2262 static int tcp_repair_options_est(struct tcp_sock *tp,
2263                 struct tcp_repair_opt __user *optbuf, unsigned int len)
2264 {
2265         struct tcp_repair_opt opt;
2266
2267         while (len >= sizeof(opt)) {
2268                 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2269                         return -EFAULT;
2270
2271                 optbuf++;
2272                 len -= sizeof(opt);
2273
2274                 switch (opt.opt_code) {
2275                 case TCPOPT_MSS:
2276                         tp->rx_opt.mss_clamp = opt.opt_val;
2277                         break;
2278                 case TCPOPT_WINDOW:
2279                         {
2280                                 u16 snd_wscale = opt.opt_val & 0xFFFF;
2281                                 u16 rcv_wscale = opt.opt_val >> 16;
2282
2283                                 if (snd_wscale > 14 || rcv_wscale > 14)
2284                                         return -EFBIG;
2285
2286                                 tp->rx_opt.snd_wscale = snd_wscale;
2287                                 tp->rx_opt.rcv_wscale = rcv_wscale;
2288                                 tp->rx_opt.wscale_ok = 1;
2289                         }
2290                         break;
2291                 case TCPOPT_SACK_PERM:
2292                         if (opt.opt_val != 0)
2293                                 return -EINVAL;
2294
2295                         tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2296                         if (sysctl_tcp_fack)
2297                                 tcp_enable_fack(tp);
2298                         break;
2299                 case TCPOPT_TIMESTAMP:
2300                         if (opt.opt_val != 0)
2301                                 return -EINVAL;
2302
2303                         tp->rx_opt.tstamp_ok = 1;
2304                         break;
2305                 }
2306         }
2307
2308         return 0;
2309 }
2310
2311 /*
2312  *      Socket option code for TCP.
2313  */
2314 static int do_tcp_setsockopt(struct sock *sk, int level,
2315                 int optname, char __user *optval, unsigned int optlen)
2316 {
2317         struct tcp_sock *tp = tcp_sk(sk);
2318         struct inet_connection_sock *icsk = inet_csk(sk);
2319         int val;
2320         int err = 0;
2321
2322         /* These are data/string values, all the others are ints */
2323         switch (optname) {
2324         case TCP_CONGESTION: {
2325                 char name[TCP_CA_NAME_MAX];
2326
2327                 if (optlen < 1)
2328                         return -EINVAL;
2329
2330                 val = strncpy_from_user(name, optval,
2331                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2332                 if (val < 0)
2333                         return -EFAULT;
2334                 name[val] = 0;
2335
2336                 lock_sock(sk);
2337                 err = tcp_set_congestion_control(sk, name);
2338                 release_sock(sk);
2339                 return err;
2340         }
2341         default:
2342                 /* fallthru */
2343                 break;
2344         }
2345
2346         if (optlen < sizeof(int))
2347                 return -EINVAL;
2348
2349         if (get_user(val, (int __user *)optval))
2350                 return -EFAULT;
2351
2352         lock_sock(sk);
2353
2354         switch (optname) {
2355         case TCP_MAXSEG:
2356                 /* Values greater than interface MTU won't take effect. However
2357                  * at the point when this call is done we typically don't yet
2358                  * know which interface is going to be used */
2359                 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2360                         err = -EINVAL;
2361                         break;
2362                 }
2363                 tp->rx_opt.user_mss = val;
2364                 break;
2365
2366         case TCP_NODELAY:
2367                 if (val) {
2368                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2369                          * this option on corked socket is remembered, but
2370                          * it is not activated until cork is cleared.
2371                          *
2372                          * However, when TCP_NODELAY is set we make
2373                          * an explicit push, which overrides even TCP_CORK
2374                          * for currently queued segments.
2375                          */
2376                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2377                         tcp_push_pending_frames(sk);
2378                 } else {
2379                         tp->nonagle &= ~TCP_NAGLE_OFF;
2380                 }
2381                 break;
2382
2383         case TCP_THIN_LINEAR_TIMEOUTS:
2384                 if (val < 0 || val > 1)
2385                         err = -EINVAL;
2386                 else
2387                         tp->thin_lto = val;
2388                 break;
2389
2390         case TCP_THIN_DUPACK:
2391                 if (val < 0 || val > 1)
2392                         err = -EINVAL;
2393                 else {
2394                         tp->thin_dupack = val;
2395                         if (tp->thin_dupack)
2396                                 tcp_disable_early_retrans(tp);
2397                 }
2398                 break;
2399
2400         case TCP_REPAIR:
2401                 if (!tcp_can_repair_sock(sk))
2402                         err = -EPERM;
2403                 else if (val == 1) {
2404                         tp->repair = 1;
2405                         sk->sk_reuse = SK_FORCE_REUSE;
2406                         tp->repair_queue = TCP_NO_QUEUE;
2407                 } else if (val == 0) {
2408                         tp->repair = 0;
2409                         sk->sk_reuse = SK_NO_REUSE;
2410                         tcp_send_window_probe(sk);
2411                 } else
2412                         err = -EINVAL;
2413
2414                 break;
2415
2416         case TCP_REPAIR_QUEUE:
2417                 if (!tp->repair)
2418                         err = -EPERM;
2419                 else if (val < TCP_QUEUES_NR)
2420                         tp->repair_queue = val;
2421                 else
2422                         err = -EINVAL;
2423                 break;
2424
2425         case TCP_QUEUE_SEQ:
2426                 if (sk->sk_state != TCP_CLOSE)
2427                         err = -EPERM;
2428                 else if (tp->repair_queue == TCP_SEND_QUEUE)
2429                         tp->write_seq = val;
2430                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2431                         tp->rcv_nxt = val;
2432                 else
2433                         err = -EINVAL;
2434                 break;
2435
2436         case TCP_REPAIR_OPTIONS:
2437                 if (!tp->repair)
2438                         err = -EINVAL;
2439                 else if (sk->sk_state == TCP_ESTABLISHED)
2440                         err = tcp_repair_options_est(tp,
2441                                         (struct tcp_repair_opt __user *)optval,
2442                                         optlen);
2443                 else
2444                         err = -EPERM;
2445                 break;
2446
2447         case TCP_CORK:
2448                 /* When set indicates to always queue non-full frames.
2449                  * Later the user clears this option and we transmit
2450                  * any pending partial frames in the queue.  This is
2451                  * meant to be used alongside sendfile() to get properly
2452                  * filled frames when the user (for example) must write
2453                  * out headers with a write() call first and then use
2454                  * sendfile to send out the data parts.
2455                  *
2456                  * TCP_CORK can be set together with TCP_NODELAY and it is
2457                  * stronger than TCP_NODELAY.
2458                  */
2459                 if (val) {
2460                         tp->nonagle |= TCP_NAGLE_CORK;
2461                 } else {
2462                         tp->nonagle &= ~TCP_NAGLE_CORK;
2463                         if (tp->nonagle&TCP_NAGLE_OFF)
2464                                 tp->nonagle |= TCP_NAGLE_PUSH;
2465                         tcp_push_pending_frames(sk);
2466                 }
2467                 break;
2468
2469         case TCP_KEEPIDLE:
2470                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2471                         err = -EINVAL;
2472                 else {
2473                         tp->keepalive_time = val * HZ;
2474                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2475                             !((1 << sk->sk_state) &
2476                               (TCPF_CLOSE | TCPF_LISTEN))) {
2477                                 u32 elapsed = keepalive_time_elapsed(tp);
2478                                 if (tp->keepalive_time > elapsed)
2479                                         elapsed = tp->keepalive_time - elapsed;
2480                                 else
2481                                         elapsed = 0;
2482                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2483                         }
2484                 }
2485                 break;
2486         case TCP_KEEPINTVL:
2487                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2488                         err = -EINVAL;
2489                 else
2490                         tp->keepalive_intvl = val * HZ;
2491                 break;
2492         case TCP_KEEPCNT:
2493                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2494                         err = -EINVAL;
2495                 else
2496                         tp->keepalive_probes = val;
2497                 break;
2498         case TCP_SYNCNT:
2499                 if (val < 1 || val > MAX_TCP_SYNCNT)
2500                         err = -EINVAL;
2501                 else
2502                         icsk->icsk_syn_retries = val;
2503                 break;
2504
2505         case TCP_LINGER2:
2506                 if (val < 0)
2507                         tp->linger2 = -1;
2508                 else if (val > sysctl_tcp_fin_timeout / HZ)
2509                         tp->linger2 = 0;
2510                 else
2511                         tp->linger2 = val * HZ;
2512                 break;
2513
2514         case TCP_DEFER_ACCEPT:
2515                 /* Translate value in seconds to number of retransmits */
2516                 icsk->icsk_accept_queue.rskq_defer_accept =
2517                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2518                                         TCP_RTO_MAX / HZ);
2519                 break;
2520
2521         case TCP_WINDOW_CLAMP:
2522                 if (!val) {
2523                         if (sk->sk_state != TCP_CLOSE) {
2524                                 err = -EINVAL;
2525                                 break;
2526                         }
2527                         tp->window_clamp = 0;
2528                 } else
2529                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2530                                                 SOCK_MIN_RCVBUF / 2 : val;
2531                 break;
2532
2533         case TCP_QUICKACK:
2534                 if (!val) {
2535                         icsk->icsk_ack.pingpong = 1;
2536                 } else {
2537                         icsk->icsk_ack.pingpong = 0;
2538                         if ((1 << sk->sk_state) &
2539                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2540                             inet_csk_ack_scheduled(sk)) {
2541                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2542                                 tcp_cleanup_rbuf(sk, 1);
2543                                 if (!(val & 1))
2544                                         icsk->icsk_ack.pingpong = 1;
2545                         }
2546                 }
2547                 break;
2548
2549 #ifdef CONFIG_TCP_MD5SIG
2550         case TCP_MD5SIG:
2551                 /* Read the IP->Key mappings from userspace */
2552                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2553                 break;
2554 #endif
2555         case TCP_USER_TIMEOUT:
2556                 /* Cap the max timeout in ms TCP will retry/retrans
2557                  * before giving up and aborting (ETIMEDOUT) a connection.
2558                  */
2559                 if (val < 0)
2560                         err = -EINVAL;
2561                 else
2562                         icsk->icsk_user_timeout = msecs_to_jiffies(val);
2563                 break;
2564
2565         case TCP_FASTOPEN:
2566                 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2567                     TCPF_LISTEN)))
2568                         err = fastopen_init_queue(sk, val);
2569                 else
2570                         err = -EINVAL;
2571                 break;
2572         case TCP_TIMESTAMP:
2573                 if (!tp->repair)
2574                         err = -EPERM;
2575                 else
2576                         tp->tsoffset = val - tcp_time_stamp;
2577                 break;
2578         case TCP_NOTSENT_LOWAT:
2579                 tp->notsent_lowat = val;
2580                 sk->sk_write_space(sk);
2581                 break;
2582         default:
2583                 err = -ENOPROTOOPT;
2584                 break;
2585         }
2586
2587         release_sock(sk);
2588         return err;
2589 }
2590
2591 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2592                    unsigned int optlen)
2593 {
2594         const struct inet_connection_sock *icsk = inet_csk(sk);
2595
2596         if (level != SOL_TCP)
2597                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2598                                                      optval, optlen);
2599         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2600 }
2601 EXPORT_SYMBOL(tcp_setsockopt);
2602
2603 #ifdef CONFIG_COMPAT
2604 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2605                           char __user *optval, unsigned int optlen)
2606 {
2607         if (level != SOL_TCP)
2608                 return inet_csk_compat_setsockopt(sk, level, optname,
2609                                                   optval, optlen);
2610         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2611 }
2612 EXPORT_SYMBOL(compat_tcp_setsockopt);
2613 #endif
2614
2615 /* Return information about state of tcp endpoint in API format. */
2616 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2617 {
2618         const struct tcp_sock *tp = tcp_sk(sk);
2619         const struct inet_connection_sock *icsk = inet_csk(sk);
2620         u32 now = tcp_time_stamp;
2621
2622         memset(info, 0, sizeof(*info));
2623
2624         info->tcpi_state = sk->sk_state;
2625         info->tcpi_ca_state = icsk->icsk_ca_state;
2626         info->tcpi_retransmits = icsk->icsk_retransmits;
2627         info->tcpi_probes = icsk->icsk_probes_out;
2628         info->tcpi_backoff = icsk->icsk_backoff;
2629
2630         if (tp->rx_opt.tstamp_ok)
2631                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2632         if (tcp_is_sack(tp))
2633                 info->tcpi_options |= TCPI_OPT_SACK;
2634         if (tp->rx_opt.wscale_ok) {
2635                 info->tcpi_options |= TCPI_OPT_WSCALE;
2636                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2637                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2638         }
2639
2640         if (tp->ecn_flags & TCP_ECN_OK)
2641                 info->tcpi_options |= TCPI_OPT_ECN;
2642         if (tp->ecn_flags & TCP_ECN_SEEN)
2643                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2644         if (tp->syn_data_acked)
2645                 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2646
2647         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2648         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2649         info->tcpi_snd_mss = tp->mss_cache;
2650         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2651
2652         if (sk->sk_state == TCP_LISTEN) {
2653                 info->tcpi_unacked = sk->sk_ack_backlog;
2654                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2655         } else {
2656                 info->tcpi_unacked = tp->packets_out;
2657                 info->tcpi_sacked = tp->sacked_out;
2658         }
2659         info->tcpi_lost = tp->lost_out;
2660         info->tcpi_retrans = tp->retrans_out;
2661         info->tcpi_fackets = tp->fackets_out;
2662
2663         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2664         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2665         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2666
2667         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2668         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2669         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2670         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2671         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2672         info->tcpi_snd_cwnd = tp->snd_cwnd;
2673         info->tcpi_advmss = tp->advmss;
2674         info->tcpi_reordering = tp->reordering;
2675
2676         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2677         info->tcpi_rcv_space = tp->rcvq_space.space;
2678
2679         info->tcpi_total_retrans = tp->total_retrans;
2680 }
2681 EXPORT_SYMBOL_GPL(tcp_get_info);
2682
2683 static int do_tcp_getsockopt(struct sock *sk, int level,
2684                 int optname, char __user *optval, int __user *optlen)
2685 {
2686         struct inet_connection_sock *icsk = inet_csk(sk);
2687         struct tcp_sock *tp = tcp_sk(sk);
2688         int val, len;
2689
2690         if (get_user(len, optlen))
2691                 return -EFAULT;
2692
2693         len = min_t(unsigned int, len, sizeof(int));
2694
2695         if (len < 0)
2696                 return -EINVAL;
2697
2698         switch (optname) {
2699         case TCP_MAXSEG:
2700                 val = tp->mss_cache;
2701                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2702                         val = tp->rx_opt.user_mss;
2703                 if (tp->repair)
2704                         val = tp->rx_opt.mss_clamp;
2705                 break;
2706         case TCP_NODELAY:
2707                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2708                 break;
2709         case TCP_CORK:
2710                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2711                 break;
2712         case TCP_KEEPIDLE:
2713                 val = keepalive_time_when(tp) / HZ;
2714                 break;
2715         case TCP_KEEPINTVL:
2716                 val = keepalive_intvl_when(tp) / HZ;
2717                 break;
2718         case TCP_KEEPCNT:
2719                 val = keepalive_probes(tp);
2720                 break;
2721         case TCP_SYNCNT:
2722                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2723                 break;
2724         case TCP_LINGER2:
2725                 val = tp->linger2;
2726                 if (val >= 0)
2727                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2728                 break;
2729         case TCP_DEFER_ACCEPT:
2730                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2731                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2732                 break;
2733         case TCP_WINDOW_CLAMP:
2734                 val = tp->window_clamp;
2735                 break;
2736         case TCP_INFO: {
2737                 struct tcp_info info;
2738
2739                 if (get_user(len, optlen))
2740                         return -EFAULT;
2741
2742                 tcp_get_info(sk, &info);
2743
2744                 len = min_t(unsigned int, len, sizeof(info));
2745                 if (put_user(len, optlen))
2746                         return -EFAULT;
2747                 if (copy_to_user(optval, &info, len))
2748                         return -EFAULT;
2749                 return 0;
2750         }
2751         case TCP_QUICKACK:
2752                 val = !icsk->icsk_ack.pingpong;
2753                 break;
2754
2755         case TCP_CONGESTION:
2756                 if (get_user(len, optlen))
2757                         return -EFAULT;
2758                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2759                 if (put_user(len, optlen))
2760                         return -EFAULT;
2761                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2762                         return -EFAULT;
2763                 return 0;
2764
2765         case TCP_THIN_LINEAR_TIMEOUTS:
2766                 val = tp->thin_lto;
2767                 break;
2768         case TCP_THIN_DUPACK:
2769                 val = tp->thin_dupack;
2770                 break;
2771
2772         case TCP_REPAIR:
2773                 val = tp->repair;
2774                 break;
2775
2776         case TCP_REPAIR_QUEUE:
2777                 if (tp->repair)
2778                         val = tp->repair_queue;
2779                 else
2780                         return -EINVAL;
2781                 break;
2782
2783         case TCP_QUEUE_SEQ:
2784                 if (tp->repair_queue == TCP_SEND_QUEUE)
2785                         val = tp->write_seq;
2786                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2787                         val = tp->rcv_nxt;
2788                 else
2789                         return -EINVAL;
2790                 break;
2791
2792         case TCP_USER_TIMEOUT:
2793                 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2794                 break;
2795         case TCP_TIMESTAMP:
2796                 val = tcp_time_stamp + tp->tsoffset;
2797                 break;
2798         case TCP_NOTSENT_LOWAT:
2799                 val = tp->notsent_lowat;
2800                 break;
2801         default:
2802                 return -ENOPROTOOPT;
2803         }
2804
2805         if (put_user(len, optlen))
2806                 return -EFAULT;
2807         if (copy_to_user(optval, &val, len))
2808                 return -EFAULT;
2809         return 0;
2810 }
2811
2812 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2813                    int __user *optlen)
2814 {
2815         struct inet_connection_sock *icsk = inet_csk(sk);
2816
2817         if (level != SOL_TCP)
2818                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2819                                                      optval, optlen);
2820         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2821 }
2822 EXPORT_SYMBOL(tcp_getsockopt);
2823
2824 #ifdef CONFIG_COMPAT
2825 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2826                           char __user *optval, int __user *optlen)
2827 {
2828         if (level != SOL_TCP)
2829                 return inet_csk_compat_getsockopt(sk, level, optname,
2830                                                   optval, optlen);
2831         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2832 }
2833 EXPORT_SYMBOL(compat_tcp_getsockopt);
2834 #endif
2835
2836 #ifdef CONFIG_TCP_MD5SIG
2837 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool __read_mostly;
2838 static DEFINE_MUTEX(tcp_md5sig_mutex);
2839
2840 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
2841 {
2842         int cpu;
2843
2844         for_each_possible_cpu(cpu) {
2845                 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
2846
2847                 if (p->md5_desc.tfm)
2848                         crypto_free_hash(p->md5_desc.tfm);
2849         }
2850         free_percpu(pool);
2851 }
2852
2853 static void __tcp_alloc_md5sig_pool(void)
2854 {
2855         int cpu;
2856         struct tcp_md5sig_pool __percpu *pool;
2857
2858         pool = alloc_percpu(struct tcp_md5sig_pool);
2859         if (!pool)
2860                 return;
2861
2862         for_each_possible_cpu(cpu) {
2863                 struct crypto_hash *hash;
2864
2865                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2866                 if (IS_ERR_OR_NULL(hash))
2867                         goto out_free;
2868
2869                 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
2870         }
2871         /* before setting tcp_md5sig_pool, we must commit all writes
2872          * to memory. See ACCESS_ONCE() in tcp_get_md5sig_pool()
2873          */
2874         smp_wmb();
2875         tcp_md5sig_pool = pool;
2876         return;
2877 out_free:
2878         __tcp_free_md5sig_pool(pool);
2879 }
2880
2881 bool tcp_alloc_md5sig_pool(void)
2882 {
2883         if (unlikely(!tcp_md5sig_pool)) {
2884                 mutex_lock(&tcp_md5sig_mutex);
2885
2886                 if (!tcp_md5sig_pool)
2887                         __tcp_alloc_md5sig_pool();
2888
2889                 mutex_unlock(&tcp_md5sig_mutex);
2890         }
2891         return tcp_md5sig_pool != NULL;
2892 }
2893 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2894
2895
2896 /**
2897  *      tcp_get_md5sig_pool - get md5sig_pool for this user
2898  *
2899  *      We use percpu structure, so if we succeed, we exit with preemption
2900  *      and BH disabled, to make sure another thread or softirq handling
2901  *      wont try to get same context.
2902  */
2903 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2904 {
2905         struct tcp_md5sig_pool __percpu *p;
2906
2907         local_bh_disable();
2908         p = ACCESS_ONCE(tcp_md5sig_pool);
2909         if (p)
2910                 return __this_cpu_ptr(p);
2911
2912         local_bh_enable();
2913         return NULL;
2914 }
2915 EXPORT_SYMBOL(tcp_get_md5sig_pool);
2916
2917 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2918                         const struct tcphdr *th)
2919 {
2920         struct scatterlist sg;
2921         struct tcphdr hdr;
2922         int err;
2923
2924         /* We are not allowed to change tcphdr, make a local copy */
2925         memcpy(&hdr, th, sizeof(hdr));
2926         hdr.check = 0;
2927
2928         /* options aren't included in the hash */
2929         sg_init_one(&sg, &hdr, sizeof(hdr));
2930         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
2931         return err;
2932 }
2933 EXPORT_SYMBOL(tcp_md5_hash_header);
2934
2935 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2936                           const struct sk_buff *skb, unsigned int header_len)
2937 {
2938         struct scatterlist sg;
2939         const struct tcphdr *tp = tcp_hdr(skb);
2940         struct hash_desc *desc = &hp->md5_desc;
2941         unsigned int i;
2942         const unsigned int head_data_len = skb_headlen(skb) > header_len ?
2943                                            skb_headlen(skb) - header_len : 0;
2944         const struct skb_shared_info *shi = skb_shinfo(skb);
2945         struct sk_buff *frag_iter;
2946
2947         sg_init_table(&sg, 1);
2948
2949         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
2950         if (crypto_hash_update(desc, &sg, head_data_len))
2951                 return 1;
2952
2953         for (i = 0; i < shi->nr_frags; ++i) {
2954                 const struct skb_frag_struct *f = &shi->frags[i];
2955                 unsigned int offset = f->page_offset;
2956                 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
2957
2958                 sg_set_page(&sg, page, skb_frag_size(f),
2959                             offset_in_page(offset));
2960                 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
2961                         return 1;
2962         }
2963
2964         skb_walk_frags(skb, frag_iter)
2965                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
2966                         return 1;
2967
2968         return 0;
2969 }
2970 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
2971
2972 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
2973 {
2974         struct scatterlist sg;
2975
2976         sg_init_one(&sg, key->key, key->keylen);
2977         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
2978 }
2979 EXPORT_SYMBOL(tcp_md5_hash_key);
2980
2981 #endif
2982
2983 void tcp_done(struct sock *sk)
2984 {
2985         struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2986
2987         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2988                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
2989
2990         tcp_set_state(sk, TCP_CLOSE);
2991         tcp_clear_xmit_timers(sk);
2992         if (req != NULL)
2993                 reqsk_fastopen_remove(sk, req, false);
2994
2995         sk->sk_shutdown = SHUTDOWN_MASK;
2996
2997         if (!sock_flag(sk, SOCK_DEAD))
2998                 sk->sk_state_change(sk);
2999         else
3000                 inet_csk_destroy_sock(sk);
3001 }
3002 EXPORT_SYMBOL_GPL(tcp_done);
3003
3004 extern struct tcp_congestion_ops tcp_reno;
3005
3006 static __initdata unsigned long thash_entries;
3007 static int __init set_thash_entries(char *str)
3008 {
3009         ssize_t ret;
3010
3011         if (!str)
3012                 return 0;
3013
3014         ret = kstrtoul(str, 0, &thash_entries);
3015         if (ret)
3016                 return 0;
3017
3018         return 1;
3019 }
3020 __setup("thash_entries=", set_thash_entries);
3021
3022 static void tcp_init_mem(void)
3023 {
3024         unsigned long limit = nr_free_buffer_pages() / 8;
3025         limit = max(limit, 128UL);
3026         sysctl_tcp_mem[0] = limit / 4 * 3;
3027         sysctl_tcp_mem[1] = limit;
3028         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3029 }
3030
3031 void __init tcp_init(void)
3032 {
3033         struct sk_buff *skb = NULL;
3034         unsigned long limit;
3035         int max_rshare, max_wshare, cnt;
3036         unsigned int i;
3037
3038         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3039
3040         percpu_counter_init(&tcp_sockets_allocated, 0);
3041         percpu_counter_init(&tcp_orphan_count, 0);
3042         tcp_hashinfo.bind_bucket_cachep =
3043                 kmem_cache_create("tcp_bind_bucket",
3044                                   sizeof(struct inet_bind_bucket), 0,
3045                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3046
3047         /* Size and allocate the main established and bind bucket
3048          * hash tables.
3049          *
3050          * The methodology is similar to that of the buffer cache.
3051          */
3052         tcp_hashinfo.ehash =
3053                 alloc_large_system_hash("TCP established",
3054                                         sizeof(struct inet_ehash_bucket),
3055                                         thash_entries,
3056                                         17, /* one slot per 128 KB of memory */
3057                                         0,
3058                                         NULL,
3059                                         &tcp_hashinfo.ehash_mask,
3060                                         0,
3061                                         thash_entries ? 0 : 512 * 1024);
3062         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3063                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3064
3065         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3066                 panic("TCP: failed to alloc ehash_locks");
3067         tcp_hashinfo.bhash =
3068                 alloc_large_system_hash("TCP bind",
3069                                         sizeof(struct inet_bind_hashbucket),
3070                                         tcp_hashinfo.ehash_mask + 1,
3071                                         17, /* one slot per 128 KB of memory */
3072                                         0,
3073                                         &tcp_hashinfo.bhash_size,
3074                                         NULL,
3075                                         0,
3076                                         64 * 1024);
3077         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3078         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3079                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3080                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3081         }
3082
3083
3084         cnt = tcp_hashinfo.ehash_mask + 1;
3085
3086         tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3087         sysctl_tcp_max_orphans = cnt / 2;
3088         sysctl_max_syn_backlog = max(128, cnt / 256);
3089
3090         tcp_init_mem();
3091         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3092         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3093         max_wshare = min(4UL*1024*1024, limit);
3094         max_rshare = min(6UL*1024*1024, limit);
3095
3096         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3097         sysctl_tcp_wmem[1] = 16*1024;
3098         sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3099
3100         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3101         sysctl_tcp_rmem[1] = 87380;
3102         sysctl_tcp_rmem[2] = max(87380, max_rshare);
3103
3104         pr_info("Hash tables configured (established %u bind %u)\n",
3105                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3106
3107         tcp_metrics_init();
3108
3109         tcp_register_congestion_control(&tcp_reno);
3110
3111         tcp_tasklet_init();
3112 }