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_us = jiffies_to_usecs(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         tcp_assign_congestion_control(sk);
 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 static void tcp_tx_timestamp(struct sock *sk, struct sk_buff *skb)
 429 {
 430         if (sk->sk_tsflags) {
 431                 struct skb_shared_info *shinfo = skb_shinfo(skb);
 432
 433                 sock_tx_timestamp(sk, &shinfo->tx_flags);
 434                 if (shinfo->tx_flags & SKBTX_ANY_TSTAMP)
 435                         shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
 436         }
 437 }
 438
 439 /*
 440  *      Wait for a TCP event.
 441  *
 442  *      Note that we don't need to lock the socket, as the upper poll layers
 443  *      take care of normal races (between the test and the event) and we don't
 444  *      go look at any of the socket buffers directly.
 445  */
 446 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 447 {
 448         unsigned int mask;
 449         struct sock *sk = sock->sk;
 450         const struct tcp_sock *tp = tcp_sk(sk);
 451
 452         sock_rps_record_flow(sk);
 453
 454         sock_poll_wait(file, sk_sleep(sk), wait);
 455         if (sk->sk_state == TCP_LISTEN)
 456                 return inet_csk_listen_poll(sk);
 457
 458         /* Socket is not locked. We are protected from async events
 459          * by poll logic and correct handling of state changes
 460          * made by other threads is impossible in any case.
 461          */
 462
 463         mask = 0;
 464
 465         /*
 466          * POLLHUP is certainly not done right. But poll() doesn't
 467          * have a notion of HUP in just one direction, and for a
 468          * socket the read side is more interesting.
 469          *
 470          * Some poll() documentation says that POLLHUP is incompatible
 471          * with the POLLOUT/POLLWR flags, so somebody should check this
 472          * all. But careful, it tends to be safer to return too many
 473          * bits than too few, and you can easily break real applications
 474          * if you don't tell them that something has hung up!
 475          *
 476          * Check-me.
 477          *
 478          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
 479          * our fs/select.c). It means that after we received EOF,
 480          * poll always returns immediately, making impossible poll() on write()
 481          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
 482          * if and only if shutdown has been made in both directions.
 483          * Actually, it is interesting to look how Solaris and DUX
 484          * solve this dilemma. I would prefer, if POLLHUP were maskable,
 485          * then we could set it on SND_SHUTDOWN. BTW examples given
 486          * in Stevens' books assume exactly this behaviour, it explains
 487          * why POLLHUP is incompatible with POLLOUT.    --ANK
 488          *
 489          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 490          * blocking on fresh not-connected or disconnected socket. --ANK
 491          */
 492         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
 493                 mask |= POLLHUP;
 494         if (sk->sk_shutdown & RCV_SHUTDOWN)
 495                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
 496
 497         /* Connected or passive Fast Open socket? */
 498         if (sk->sk_state != TCP_SYN_SENT &&
 499             (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk)) {
 500                 int target = sock_rcvlowat(sk, 0, INT_MAX);
 501
 502                 if (tp->urg_seq == tp->copied_seq &&
 503                     !sock_flag(sk, SOCK_URGINLINE) &&
 504                     tp->urg_data)
 505                         target++;
 506
 507                 /* Potential race condition. If read of tp below will
 508                  * escape above sk->sk_state, we can be illegally awaken
 509                  * in SYN_* states. */
 510                 if (tp->rcv_nxt - tp->copied_seq >= target)
 511                         mask |= POLLIN | POLLRDNORM;
 512
 513                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 514                         if (sk_stream_is_writeable(sk)) {
 515                                 mask |= POLLOUT | POLLWRNORM;
 516                         } else {  /* send SIGIO later */
 517                                 set_bit(SOCK_ASYNC_NOSPACE,
 518                                         &sk->sk_socket->flags);
 519                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 520
 521                                 /* Race breaker. If space is freed after
 522                                  * wspace test but before the flags are set,
 523                                  * IO signal will be lost.
 524                                  */
 525                                 if (sk_stream_is_writeable(sk))
 526                                         mask |= POLLOUT | POLLWRNORM;
 527                         }
 528                 } else
 529                         mask |= POLLOUT | POLLWRNORM;
 530
 531                 if (tp->urg_data & TCP_URG_VALID)
 532                         mask |= POLLPRI;
 533         }
 534         /* This barrier is coupled with smp_wmb() in tcp_reset() */
 535         smp_rmb();
 536         if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
 537                 mask |= POLLERR;
 538
 539         return mask;
 540 }
 541 EXPORT_SYMBOL(tcp_poll);
 542
 543 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 544 {
 545         struct tcp_sock *tp = tcp_sk(sk);
 546         int answ;
 547         bool slow;
 548
 549         switch (cmd) {
 550         case SIOCINQ:
 551                 if (sk->sk_state == TCP_LISTEN)
 552                         return -EINVAL;
 553
 554                 slow = lock_sock_fast(sk);
 555                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 556                         answ = 0;
 557                 else if (sock_flag(sk, SOCK_URGINLINE) ||
 558                          !tp->urg_data ||
 559                          before(tp->urg_seq, tp->copied_seq) ||
 560                          !before(tp->urg_seq, tp->rcv_nxt)) {
 561
 562                         answ = tp->rcv_nxt - tp->copied_seq;
 563
 564                         /* Subtract 1, if FIN was received */
 565                         if (answ && sock_flag(sk, SOCK_DONE))
 566                                 answ--;
 567                 } else
 568                         answ = tp->urg_seq - tp->copied_seq;
 569                 unlock_sock_fast(sk, slow);
 570                 break;
 571         case SIOCATMARK:
 572                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 573                 break;
 574         case SIOCOUTQ:
 575                 if (sk->sk_state == TCP_LISTEN)
 576                         return -EINVAL;
 577
 578                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 579                         answ = 0;
 580                 else
 581                         answ = tp->write_seq - tp->snd_una;
 582                 break;
 583         case SIOCOUTQNSD:
 584                 if (sk->sk_state == TCP_LISTEN)
 585                         return -EINVAL;
 586
 587                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 588                         answ = 0;
 589                 else
 590                         answ = tp->write_seq - tp->snd_nxt;
 591                 break;
 592         default:
 593                 return -ENOIOCTLCMD;
 594         }
 595
 596         return put_user(answ, (int __user *)arg);
 597 }
 598 EXPORT_SYMBOL(tcp_ioctl);
 599
 600 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 601 {
 602         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 603         tp->pushed_seq = tp->write_seq;
 604 }
 605
 606 static inline bool forced_push(const struct tcp_sock *tp)
 607 {
 608         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 609 }
 610
 611 static void skb_entail(struct sock *sk, struct sk_buff *skb)
 612 {
 613         struct tcp_sock *tp = tcp_sk(sk);
 614         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 615
 616         skb->csum    = 0;
 617         tcb->seq     = tcb->end_seq = tp->write_seq;
 618         tcb->tcp_flags = TCPHDR_ACK;
 619         tcb->sacked  = 0;
 620         __skb_header_release(skb);
 621         tcp_add_write_queue_tail(sk, skb);
 622         sk->sk_wmem_queued += skb->truesize;
 623         sk_mem_charge(sk, skb->truesize);
 624         if (tp->nonagle & TCP_NAGLE_PUSH)
 625                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 626 }
 627
 628 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
 629 {
 630         if (flags & MSG_OOB)
 631                 tp->snd_up = tp->write_seq;
 632 }
 633
 634 /* If a not yet filled skb is pushed, do not send it if
 635  * we have data packets in Qdisc or NIC queues :
 636  * Because TX completion will happen shortly, it gives a chance
 637  * to coalesce future sendmsg() payload into this skb, without
 638  * need for a timer, and with no latency trade off.
 639  * As packets containing data payload have a bigger truesize
 640  * than pure acks (dataless) packets, the last checks prevent
 641  * autocorking if we only have an ACK in Qdisc/NIC queues,
 642  * or if TX completion was delayed after we processed ACK packet.
 643  */
 644 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
 645                                 int size_goal)
 646 {
 647         return skb->len < size_goal &&
 648                sysctl_tcp_autocorking &&
 649                skb != tcp_write_queue_head(sk) &&
 650                atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
 651 }
 652
 653 static void tcp_push(struct sock *sk, int flags, int mss_now,
 654                      int nonagle, int size_goal)
 655 {
 656         struct tcp_sock *tp = tcp_sk(sk);
 657         struct sk_buff *skb;
 658
 659         if (!tcp_send_head(sk))
 660                 return;
 661
 662         skb = tcp_write_queue_tail(sk);
 663         if (!(flags & MSG_MORE) || forced_push(tp))
 664                 tcp_mark_push(tp, skb);
 665
 666         tcp_mark_urg(tp, flags);
 667
 668         if (tcp_should_autocork(sk, skb, size_goal)) {
 669
 670                 /* avoid atomic op if TSQ_THROTTLED bit is already set */
 671                 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
 672                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
 673                         set_bit(TSQ_THROTTLED, &tp->tsq_flags);
 674                 }
 675                 /* It is possible TX completion already happened
 676                  * before we set TSQ_THROTTLED.
 677                  */
 678                 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
 679                         return;
 680         }
 681
 682         if (flags & MSG_MORE)
 683                 nonagle = TCP_NAGLE_CORK;
 684
 685         __tcp_push_pending_frames(sk, mss_now, nonagle);
 686 }
 687
 688 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 689                                 unsigned int offset, size_t len)
 690 {
 691         struct tcp_splice_state *tss = rd_desc->arg.data;
 692         int ret;
 693
 694         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
 695                               tss->flags);
 696         if (ret > 0)
 697                 rd_desc->count -= ret;
 698         return ret;
 699 }
 700
 701 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 702 {
 703         /* Store TCP splice context information in read_descriptor_t. */
 704         read_descriptor_t rd_desc = {
 705                 .arg.data = tss,
 706                 .count    = tss->len,
 707         };
 708
 709         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 710 }
 711
 712 /**
 713  *  tcp_splice_read - splice data from TCP socket to a pipe
 714  * @sock:       socket to splice from
 715  * @ppos:       position (not valid)
 716  * @pipe:       pipe to splice to
 717  * @len:        number of bytes to splice
 718  * @flags:      splice modifier flags
 719  *
 720  * Description:
 721  *    Will read pages from given socket and fill them into a pipe.
 722  *
 723  **/
 724 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 725                         struct pipe_inode_info *pipe, size_t len,
 726                         unsigned int flags)
 727 {
 728         struct sock *sk = sock->sk;
 729         struct tcp_splice_state tss = {
 730                 .pipe = pipe,
 731                 .len = len,
 732                 .flags = flags,
 733         };
 734         long timeo;
 735         ssize_t spliced;
 736         int ret;
 737
 738         sock_rps_record_flow(sk);
 739         /*
 740          * We can't seek on a socket input
 741          */
 742         if (unlikely(*ppos))
 743                 return -ESPIPE;
 744
 745         ret = spliced = 0;
 746
 747         lock_sock(sk);
 748
 749         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 750         while (tss.len) {
 751                 ret = __tcp_splice_read(sk, &tss);
 752                 if (ret < 0)
 753                         break;
 754                 else if (!ret) {
 755                         if (spliced)
 756                                 break;
 757                         if (sock_flag(sk, SOCK_DONE))
 758                                 break;
 759                         if (sk->sk_err) {
 760                                 ret = sock_error(sk);
 761                                 break;
 762                         }
 763                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 764                                 break;
 765                         if (sk->sk_state == TCP_CLOSE) {
 766                                 /*
 767                                  * This occurs when user tries to read
 768                                  * from never connected socket.
 769                                  */
 770                                 if (!sock_flag(sk, SOCK_DONE))
 771                                         ret = -ENOTCONN;
 772                                 break;
 773                         }
 774                         if (!timeo) {
 775                                 ret = -EAGAIN;
 776                                 break;
 777                         }
 778                         sk_wait_data(sk, &timeo);
 779                         if (signal_pending(current)) {
 780                                 ret = sock_intr_errno(timeo);
 781                                 break;
 782                         }
 783                         continue;
 784                 }
 785                 tss.len -= ret;
 786                 spliced += ret;
 787
 788                 if (!timeo)
 789                         break;
 790                 release_sock(sk);
 791                 lock_sock(sk);
 792
 793                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 794                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 795                     signal_pending(current))
 796                         break;
 797         }
 798
 799         release_sock(sk);
 800
 801         if (spliced)
 802                 return spliced;
 803
 804         return ret;
 805 }
 806 EXPORT_SYMBOL(tcp_splice_read);
 807
 808 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
 809 {
 810         struct sk_buff *skb;
 811
 812         /* The TCP header must be at least 32-bit aligned.  */
 813         size = ALIGN(size, 4);
 814
 815         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 816         if (skb) {
 817                 if (sk_wmem_schedule(sk, skb->truesize)) {
 818                         skb_reserve(skb, sk->sk_prot->max_header);
 819                         /*
 820                          * Make sure that we have exactly size bytes
 821                          * available to the caller, no more, no less.
 822                          */
 823                         skb->reserved_tailroom = skb->end - skb->tail - size;
 824                         return skb;
 825                 }
 826                 __kfree_skb(skb);
 827         } else {
 828                 sk->sk_prot->enter_memory_pressure(sk);
 829                 sk_stream_moderate_sndbuf(sk);
 830         }
 831         return NULL;
 832 }
 833
 834 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 835                                        int large_allowed)
 836 {
 837         struct tcp_sock *tp = tcp_sk(sk);
 838         u32 new_size_goal, size_goal;
 839
 840         if (!large_allowed || !sk_can_gso(sk))
 841                 return mss_now;
 842
 843         /* Note : tcp_tso_autosize() will eventually split this later */
 844         new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
 845         new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
 846
 847         /* We try hard to avoid divides here */
 848         size_goal = tp->gso_segs * mss_now;
 849         if (unlikely(new_size_goal < size_goal ||
 850                      new_size_goal >= size_goal + mss_now)) {
 851                 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
 852                                      sk->sk_gso_max_segs);
 853                 size_goal = tp->gso_segs * mss_now;
 854         }
 855
 856         return max(size_goal, mss_now);
 857 }
 858
 859 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 860 {
 861         int mss_now;
 862
 863         mss_now = tcp_current_mss(sk);
 864         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 865
 866         return mss_now;
 867 }
 868
 869 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
 870                                 size_t size, int flags)
 871 {
 872         struct tcp_sock *tp = tcp_sk(sk);
 873         int mss_now, size_goal;
 874         int err;
 875         ssize_t copied;
 876         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 877
 878         /* Wait for a connection to finish. One exception is TCP Fast Open
 879          * (passive side) where data is allowed to be sent before a connection
 880          * is fully established.
 881          */
 882         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
 883             !tcp_passive_fastopen(sk)) {
 884                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 885                         goto out_err;
 886         }
 887
 888         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 889
 890         mss_now = tcp_send_mss(sk, &size_goal, flags);
 891         copied = 0;
 892
 893         err = -EPIPE;
 894         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 895                 goto out_err;
 896
 897         while (size > 0) {
 898                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 899                 int copy, i;
 900                 bool can_coalesce;
 901
 902                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
 903 new_segment:
 904                         if (!sk_stream_memory_free(sk))
 905                                 goto wait_for_sndbuf;
 906
 907                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
 908                         if (!skb)
 909                                 goto wait_for_memory;
 910
 911                         skb_entail(sk, skb);
 912                         copy = size_goal;
 913                 }
 914
 915                 if (copy > size)
 916                         copy = size;
 917
 918                 i = skb_shinfo(skb)->nr_frags;
 919                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 920                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
 921                         tcp_mark_push(tp, skb);
 922                         goto new_segment;
 923                 }
 924                 if (!sk_wmem_schedule(sk, copy))
 925                         goto wait_for_memory;
 926
 927                 if (can_coalesce) {
 928                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
 929                 } else {
 930                         get_page(page);
 931                         skb_fill_page_desc(skb, i, page, offset, copy);
 932                 }
 933                 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
 934
 935                 skb->len += copy;
 936                 skb->data_len += copy;
 937                 skb->truesize += copy;
 938                 sk->sk_wmem_queued += copy;
 939                 sk_mem_charge(sk, copy);
 940                 skb->ip_summed = CHECKSUM_PARTIAL;
 941                 tp->write_seq += copy;
 942                 TCP_SKB_CB(skb)->end_seq += copy;
 943                 tcp_skb_pcount_set(skb, 0);
 944
 945                 if (!copied)
 946                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
 947
 948                 copied += copy;
 949                 offset += copy;
 950                 if (!(size -= copy)) {
 951                         tcp_tx_timestamp(sk, skb);
 952                         goto out;
 953                 }
 954
 955                 if (skb->len < size_goal || (flags & MSG_OOB))
 956                         continue;
 957
 958                 if (forced_push(tp)) {
 959                         tcp_mark_push(tp, skb);
 960                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
 961                 } else if (skb == tcp_send_head(sk))
 962                         tcp_push_one(sk, mss_now);
 963                 continue;
 964
 965 wait_for_sndbuf:
 966                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 967 wait_for_memory:
 968                 tcp_push(sk, flags & ~MSG_MORE, mss_now,
 969                          TCP_NAGLE_PUSH, size_goal);
 970
 971                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 972                         goto do_error;
 973
 974                 mss_now = tcp_send_mss(sk, &size_goal, flags);
 975         }
 976
 977 out:
 978         if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
 979                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
 980         return copied;
 981
 982 do_error:
 983         if (copied)
 984                 goto out;
 985 out_err:
 986         return sk_stream_error(sk, flags, err);
 987 }
 988
 989 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
 990                  size_t size, int flags)
 991 {
 992         ssize_t res;
 993
 994         if (!(sk->sk_route_caps & NETIF_F_SG) ||
 995             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
 996                 return sock_no_sendpage(sk->sk_socket, page, offset, size,
 997                                         flags);
 998
 999         lock_sock(sk);
1000         res = do_tcp_sendpages(sk, page, offset, size, flags);
1001         release_sock(sk);
1002         return res;
1003 }
1004 EXPORT_SYMBOL(tcp_sendpage);
1005
1006 static inline int select_size(const struct sock *sk, bool sg)
1007 {
1008         const struct tcp_sock *tp = tcp_sk(sk);
1009         int tmp = tp->mss_cache;
1010
1011         if (sg) {
1012                 if (sk_can_gso(sk)) {
1013                         /* Small frames wont use a full page:
1014                          * Payload will immediately follow tcp header.
1015                          */
1016                         tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1017                 } else {
1018                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1019
1020                         if (tmp >= pgbreak &&
1021                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1022                                 tmp = pgbreak;
1023                 }
1024         }
1025
1026         return tmp;
1027 }
1028
1029 void tcp_free_fastopen_req(struct tcp_sock *tp)
1030 {
1031         if (tp->fastopen_req) {
1032                 kfree(tp->fastopen_req);
1033                 tp->fastopen_req = NULL;
1034         }
1035 }
1036
1037 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1038                                 int *copied, size_t size)
1039 {
1040         struct tcp_sock *tp = tcp_sk(sk);
1041         int err, flags;
1042
1043         if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1044                 return -EOPNOTSUPP;
1045         if (tp->fastopen_req)
1046                 return -EALREADY; /* Another Fast Open is in progress */
1047
1048         tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1049                                    sk->sk_allocation);
1050         if (unlikely(!tp->fastopen_req))
1051                 return -ENOBUFS;
1052         tp->fastopen_req->data = msg;
1053         tp->fastopen_req->size = size;
1054
1055         flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1056         err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1057                                     msg->msg_namelen, flags);
1058         *copied = tp->fastopen_req->copied;
1059         tcp_free_fastopen_req(tp);
1060         return err;
1061 }
1062
1063 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1064 {
1065         struct tcp_sock *tp = tcp_sk(sk);
1066         struct sk_buff *skb;
1067         int flags, err, copied = 0;
1068         int mss_now = 0, size_goal, copied_syn = 0;
1069         bool sg;
1070         long timeo;
1071
1072         lock_sock(sk);
1073
1074         flags = msg->msg_flags;
1075         if (flags & MSG_FASTOPEN) {
1076                 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1077                 if (err == -EINPROGRESS && copied_syn > 0)
1078                         goto out;
1079                 else if (err)
1080                         goto out_err;
1081         }
1082
1083         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1084
1085         /* Wait for a connection to finish. One exception is TCP Fast Open
1086          * (passive side) where data is allowed to be sent before a connection
1087          * is fully established.
1088          */
1089         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1090             !tcp_passive_fastopen(sk)) {
1091                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1092                         goto do_error;
1093         }
1094
1095         if (unlikely(tp->repair)) {
1096                 if (tp->repair_queue == TCP_RECV_QUEUE) {
1097                         copied = tcp_send_rcvq(sk, msg, size);
1098                         goto out_nopush;
1099                 }
1100
1101                 err = -EINVAL;
1102                 if (tp->repair_queue == TCP_NO_QUEUE)
1103                         goto out_err;
1104
1105                 /* 'common' sending to sendq */
1106         }
1107
1108         /* This should be in poll */
1109         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1110
1111         mss_now = tcp_send_mss(sk, &size_goal, flags);
1112
1113         /* Ok commence sending. */
1114         copied = 0;
1115
1116         err = -EPIPE;
1117         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1118                 goto out_err;
1119
1120         sg = !!(sk->sk_route_caps & NETIF_F_SG);
1121
1122         while (msg_data_left(msg)) {
1123                 int copy = 0;
1124                 int max = size_goal;
1125
1126                 skb = tcp_write_queue_tail(sk);
1127                 if (tcp_send_head(sk)) {
1128                         if (skb->ip_summed == CHECKSUM_NONE)
1129                                 max = mss_now;
1130                         copy = max - skb->len;
1131                 }
1132
1133                 if (copy <= 0) {
1134 new_segment:
1135                         /* Allocate new segment. If the interface is SG,
1136                          * allocate skb fitting to single page.
1137                          */
1138                         if (!sk_stream_memory_free(sk))
1139                                 goto wait_for_sndbuf;
1140
1141                         skb = sk_stream_alloc_skb(sk,
1142                                                   select_size(sk, sg),
1143                                                   sk->sk_allocation);
1144                         if (!skb)
1145                                 goto wait_for_memory;
1146
1147                         /*
1148                          * Check whether we can use HW checksum.
1149                          */
1150                         if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1151                                 skb->ip_summed = CHECKSUM_PARTIAL;
1152
1153                         skb_entail(sk, skb);
1154                         copy = size_goal;
1155                         max = size_goal;
1156
1157                         /* All packets are restored as if they have
1158                          * already been sent. skb_mstamp isn't set to
1159                          * avoid wrong rtt estimation.
1160                          */
1161                         if (tp->repair)
1162                                 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1163                 }
1164
1165                 /* Try to append data to the end of skb. */
1166                 if (copy > msg_data_left(msg))
1167                         copy = msg_data_left(msg);
1168
1169                 /* Where to copy to? */
1170                 if (skb_availroom(skb) > 0) {
1171                         /* We have some space in skb head. Superb! */
1172                         copy = min_t(int, copy, skb_availroom(skb));
1173                         err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1174                         if (err)
1175                                 goto do_fault;
1176                 } else {
1177                         bool merge = true;
1178                         int i = skb_shinfo(skb)->nr_frags;
1179                         struct page_frag *pfrag = sk_page_frag(sk);
1180
1181                         if (!sk_page_frag_refill(sk, pfrag))
1182                                 goto wait_for_memory;
1183
1184                         if (!skb_can_coalesce(skb, i, pfrag->page,
1185                                               pfrag->offset)) {
1186                                 if (i == MAX_SKB_FRAGS || !sg) {
1187                                         tcp_mark_push(tp, skb);
1188                                         goto new_segment;
1189                                 }
1190                                 merge = false;
1191                         }
1192
1193                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1194
1195                         if (!sk_wmem_schedule(sk, copy))
1196                                 goto wait_for_memory;
1197
1198                         err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1199                                                        pfrag->page,
1200                                                        pfrag->offset,
1201                                                        copy);
1202                         if (err)
1203                                 goto do_error;
1204
1205                         /* Update the skb. */
1206                         if (merge) {
1207                                 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1208                         } else {
1209                                 skb_fill_page_desc(skb, i, pfrag->page,
1210                                                    pfrag->offset, copy);
1211                                 get_page(pfrag->page);
1212                         }
1213                         pfrag->offset += copy;
1214                 }
1215
1216                 if (!copied)
1217                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1218
1219                 tp->write_seq += copy;
1220                 TCP_SKB_CB(skb)->end_seq += copy;
1221                 tcp_skb_pcount_set(skb, 0);
1222
1223                 copied += copy;
1224                 if (!msg_data_left(msg)) {
1225                         tcp_tx_timestamp(sk, skb);
1226                         goto out;
1227                 }
1228
1229                 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1230                         continue;
1231
1232                 if (forced_push(tp)) {
1233                         tcp_mark_push(tp, skb);
1234                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1235                 } else if (skb == tcp_send_head(sk))
1236                         tcp_push_one(sk, mss_now);
1237                 continue;
1238
1239 wait_for_sndbuf:
1240                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1241 wait_for_memory:
1242                 if (copied)
1243                         tcp_push(sk, flags & ~MSG_MORE, mss_now,
1244                                  TCP_NAGLE_PUSH, size_goal);
1245
1246                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1247                         goto do_error;
1248
1249                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1250         }
1251
1252 out:
1253         if (copied)
1254                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1255 out_nopush:
1256         release_sock(sk);
1257         return copied + copied_syn;
1258
1259 do_fault:
1260         if (!skb->len) {
1261                 tcp_unlink_write_queue(skb, sk);
1262                 /* It is the one place in all of TCP, except connection
1263                  * reset, where we can be unlinking the send_head.
1264                  */
1265                 tcp_check_send_head(sk, skb);
1266                 sk_wmem_free_skb(sk, skb);
1267         }
1268
1269 do_error:
1270         if (copied + copied_syn)
1271                 goto out;
1272 out_err:
1273         err = sk_stream_error(sk, flags, err);
1274         release_sock(sk);
1275         return err;
1276 }
1277 EXPORT_SYMBOL(tcp_sendmsg);
1278
1279 /*
1280  *      Handle reading urgent data. BSD has very simple semantics for
1281  *      this, no blocking and very strange errors 8)
1282  */
1283
1284 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1285 {
1286         struct tcp_sock *tp = tcp_sk(sk);
1287
1288         /* No URG data to read. */
1289         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1290             tp->urg_data == TCP_URG_READ)
1291                 return -EINVAL; /* Yes this is right ! */
1292
1293         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1294                 return -ENOTCONN;
1295
1296         if (tp->urg_data & TCP_URG_VALID) {
1297                 int err = 0;
1298                 char c = tp->urg_data;
1299
1300                 if (!(flags & MSG_PEEK))
1301                         tp->urg_data = TCP_URG_READ;
1302
1303                 /* Read urgent data. */
1304                 msg->msg_flags |= MSG_OOB;
1305
1306                 if (len > 0) {
1307                         if (!(flags & MSG_TRUNC))
1308                                 err = memcpy_to_msg(msg, &c, 1);
1309                         len = 1;
1310                 } else
1311                         msg->msg_flags |= MSG_TRUNC;
1312
1313                 return err ? -EFAULT : len;
1314         }
1315
1316         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1317                 return 0;
1318
1319         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1320          * the available implementations agree in this case:
1321          * this call should never block, independent of the
1322          * blocking state of the socket.
1323          * Mike <pall@rz.uni-karlsruhe.de>
1324          */
1325         return -EAGAIN;
1326 }
1327
1328 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1329 {
1330         struct sk_buff *skb;
1331         int copied = 0, err = 0;
1332
1333         /* XXX -- need to support SO_PEEK_OFF */
1334
1335         skb_queue_walk(&sk->sk_write_queue, skb) {
1336                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1337                 if (err)
1338                         break;
1339
1340                 copied += skb->len;
1341         }
1342
1343         return err ?: copied;
1344 }
1345
1346 /* Clean up the receive buffer for full frames taken by the user,
1347  * then send an ACK if necessary.  COPIED is the number of bytes
1348  * tcp_recvmsg has given to the user so far, it speeds up the
1349  * calculation of whether or not we must ACK for the sake of
1350  * a window update.
1351  */
1352 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1353 {
1354         struct tcp_sock *tp = tcp_sk(sk);
1355         bool time_to_ack = false;
1356
1357         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1358
1359         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1360              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1361              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1362
1363         if (inet_csk_ack_scheduled(sk)) {
1364                 const struct inet_connection_sock *icsk = inet_csk(sk);
1365                    /* Delayed ACKs frequently hit locked sockets during bulk
1366                     * receive. */
1367                 if (icsk->icsk_ack.blocked ||
1368                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1369                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1370                     /*
1371                      * If this read emptied read buffer, we send ACK, if
1372                      * connection is not bidirectional, user drained
1373                      * receive buffer and there was a small segment
1374                      * in queue.
1375                      */
1376                     (copied > 0 &&
1377                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1378                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1379                        !icsk->icsk_ack.pingpong)) &&
1380                       !atomic_read(&sk->sk_rmem_alloc)))
1381                         time_to_ack = true;
1382         }
1383
1384         /* We send an ACK if we can now advertise a non-zero window
1385          * which has been raised "significantly".
1386          *
1387          * Even if window raised up to infinity, do not send window open ACK
1388          * in states, where we will not receive more. It is useless.
1389          */
1390         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1391                 __u32 rcv_window_now = tcp_receive_window(tp);
1392
1393                 /* Optimize, __tcp_select_window() is not cheap. */
1394                 if (2*rcv_window_now <= tp->window_clamp) {
1395                         __u32 new_window = __tcp_select_window(sk);
1396
1397                         /* Send ACK now, if this read freed lots of space
1398                          * in our buffer. Certainly, new_window is new window.
1399                          * We can advertise it now, if it is not less than current one.
1400                          * "Lots" means "at least twice" here.
1401                          */
1402                         if (new_window && new_window >= 2 * rcv_window_now)
1403                                 time_to_ack = true;
1404                 }
1405         }
1406         if (time_to_ack)
1407                 tcp_send_ack(sk);
1408 }
1409
1410 static void tcp_prequeue_process(struct sock *sk)
1411 {
1412         struct sk_buff *skb;
1413         struct tcp_sock *tp = tcp_sk(sk);
1414
1415         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1416
1417         /* RX process wants to run with disabled BHs, though it is not
1418          * necessary */
1419         local_bh_disable();
1420         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1421                 sk_backlog_rcv(sk, skb);
1422         local_bh_enable();
1423
1424         /* Clear memory counter. */
1425         tp->ucopy.memory = 0;
1426 }
1427
1428 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1429 {
1430         struct sk_buff *skb;
1431         u32 offset;
1432
1433         while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1434                 offset = seq - TCP_SKB_CB(skb)->seq;
1435                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1436                         offset--;
1437                 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1438                         *off = offset;
1439                         return skb;
1440                 }
1441                 /* This looks weird, but this can happen if TCP collapsing
1442                  * splitted a fat GRO packet, while we released socket lock
1443                  * in skb_splice_bits()
1444                  */
1445                 sk_eat_skb(sk, skb);
1446         }
1447         return NULL;
1448 }
1449
1450 /*
1451  * This routine provides an alternative to tcp_recvmsg() for routines
1452  * that would like to handle copying from skbuffs directly in 'sendfile'
1453  * fashion.
1454  * Note:
1455  *      - It is assumed that the socket was locked by the caller.
1456  *      - The routine does not block.
1457  *      - At present, there is no support for reading OOB data
1458  *        or for 'peeking' the socket using this routine
1459  *        (although both would be easy to implement).
1460  */
1461 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1462                   sk_read_actor_t recv_actor)
1463 {
1464         struct sk_buff *skb;
1465         struct tcp_sock *tp = tcp_sk(sk);
1466         u32 seq = tp->copied_seq;
1467         u32 offset;
1468         int copied = 0;
1469
1470         if (sk->sk_state == TCP_LISTEN)
1471                 return -ENOTCONN;
1472         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1473                 if (offset < skb->len) {
1474                         int used;
1475                         size_t len;
1476
1477                         len = skb->len - offset;
1478                         /* Stop reading if we hit a patch of urgent data */
1479                         if (tp->urg_data) {
1480                                 u32 urg_offset = tp->urg_seq - seq;
1481                                 if (urg_offset < len)
1482                                         len = urg_offset;
1483                                 if (!len)
1484                                         break;
1485                         }
1486                         used = recv_actor(desc, skb, offset, len);
1487                         if (used <= 0) {
1488                                 if (!copied)
1489                                         copied = used;
1490                                 break;
1491                         } else if (used <= len) {
1492                                 seq += used;
1493                                 copied += used;
1494                                 offset += used;
1495                         }
1496                         /* If recv_actor drops the lock (e.g. TCP splice
1497                          * receive) the skb pointer might be invalid when
1498                          * getting here: tcp_collapse might have deleted it
1499                          * while aggregating skbs from the socket queue.
1500                          */
1501                         skb = tcp_recv_skb(sk, seq - 1, &offset);
1502                         if (!skb)
1503                                 break;
1504                         /* TCP coalescing might have appended data to the skb.
1505                          * Try to splice more frags
1506                          */
1507                         if (offset + 1 != skb->len)
1508                                 continue;
1509                 }
1510                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1511                         sk_eat_skb(sk, skb);
1512                         ++seq;
1513                         break;
1514                 }
1515                 sk_eat_skb(sk, skb);
1516                 if (!desc->count)
1517                         break;
1518                 tp->copied_seq = seq;
1519         }
1520         tp->copied_seq = seq;
1521
1522         tcp_rcv_space_adjust(sk);
1523
1524         /* Clean up data we have read: This will do ACK frames. */
1525         if (copied > 0) {
1526                 tcp_recv_skb(sk, seq, &offset);
1527                 tcp_cleanup_rbuf(sk, copied);
1528         }
1529         return copied;
1530 }
1531 EXPORT_SYMBOL(tcp_read_sock);
1532
1533 /*
1534  *      This routine copies from a sock struct into the user buffer.
1535  *
1536  *      Technical note: in 2.3 we work on _locked_ socket, so that
1537  *      tricks with *seq access order and skb->users are not required.
1538  *      Probably, code can be easily improved even more.
1539  */
1540
1541 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1542                 int flags, int *addr_len)
1543 {
1544         struct tcp_sock *tp = tcp_sk(sk);
1545         int copied = 0;
1546         u32 peek_seq;
1547         u32 *seq;
1548         unsigned long used;
1549         int err;
1550         int target;             /* Read at least this many bytes */
1551         long timeo;
1552         struct task_struct *user_recv = NULL;
1553         struct sk_buff *skb;
1554         u32 urg_hole = 0;
1555
1556         if (unlikely(flags & MSG_ERRQUEUE))
1557                 return inet_recv_error(sk, msg, len, addr_len);
1558
1559         if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1560             (sk->sk_state == TCP_ESTABLISHED))
1561                 sk_busy_loop(sk, nonblock);
1562
1563         lock_sock(sk);
1564
1565         err = -ENOTCONN;
1566         if (sk->sk_state == TCP_LISTEN)
1567                 goto out;
1568
1569         timeo = sock_rcvtimeo(sk, nonblock);
1570
1571         /* Urgent data needs to be handled specially. */
1572         if (flags & MSG_OOB)
1573                 goto recv_urg;
1574
1575         if (unlikely(tp->repair)) {
1576                 err = -EPERM;
1577                 if (!(flags & MSG_PEEK))
1578                         goto out;
1579
1580                 if (tp->repair_queue == TCP_SEND_QUEUE)
1581                         goto recv_sndq;
1582
1583                 err = -EINVAL;
1584                 if (tp->repair_queue == TCP_NO_QUEUE)
1585                         goto out;
1586
1587                 /* 'common' recv queue MSG_PEEK-ing */
1588         }
1589
1590         seq = &tp->copied_seq;
1591         if (flags & MSG_PEEK) {
1592                 peek_seq = tp->copied_seq;
1593                 seq = &peek_seq;
1594         }
1595
1596         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1597
1598         do {
1599                 u32 offset;
1600
1601                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1602                 if (tp->urg_data && tp->urg_seq == *seq) {
1603                         if (copied)
1604                                 break;
1605                         if (signal_pending(current)) {
1606                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1607                                 break;
1608                         }
1609                 }
1610
1611                 /* Next get a buffer. */
1612
1613                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1614                         /* Now that we have two receive queues this
1615                          * shouldn't happen.
1616                          */
1617                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1618                                  "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1619                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1620                                  flags))
1621                                 break;
1622
1623                         offset = *seq - TCP_SKB_CB(skb)->seq;
1624                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1625                                 offset--;
1626                         if (offset < skb->len)
1627                                 goto found_ok_skb;
1628                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1629                                 goto found_fin_ok;
1630                         WARN(!(flags & MSG_PEEK),
1631                              "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1632                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1633                 }
1634
1635                 /* Well, if we have backlog, try to process it now yet. */
1636
1637                 if (copied >= target && !sk->sk_backlog.tail)
1638                         break;
1639
1640                 if (copied) {
1641                         if (sk->sk_err ||
1642                             sk->sk_state == TCP_CLOSE ||
1643                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1644                             !timeo ||
1645                             signal_pending(current))
1646                                 break;
1647                 } else {
1648                         if (sock_flag(sk, SOCK_DONE))
1649                                 break;
1650
1651                         if (sk->sk_err) {
1652                                 copied = sock_error(sk);
1653                                 break;
1654                         }
1655
1656                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1657                                 break;
1658
1659                         if (sk->sk_state == TCP_CLOSE) {
1660                                 if (!sock_flag(sk, SOCK_DONE)) {
1661                                         /* This occurs when user tries to read
1662                                          * from never connected socket.
1663                                          */
1664                                         copied = -ENOTCONN;
1665                                         break;
1666                                 }
1667                                 break;
1668                         }
1669
1670                         if (!timeo) {
1671                                 copied = -EAGAIN;
1672                                 break;
1673                         }
1674
1675                         if (signal_pending(current)) {
1676                                 copied = sock_intr_errno(timeo);
1677                                 break;
1678                         }
1679                 }
1680
1681                 tcp_cleanup_rbuf(sk, copied);
1682
1683                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1684                         /* Install new reader */
1685                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1686                                 user_recv = current;
1687                                 tp->ucopy.task = user_recv;
1688                                 tp->ucopy.msg = msg;
1689                         }
1690
1691                         tp->ucopy.len = len;
1692
1693                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1694                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1695
1696                         /* Ugly... If prequeue is not empty, we have to
1697                          * process it before releasing socket, otherwise
1698                          * order will be broken at second iteration.
1699                          * More elegant solution is required!!!
1700                          *
1701                          * Look: we have the following (pseudo)queues:
1702                          *
1703                          * 1. packets in flight
1704                          * 2. backlog
1705                          * 3. prequeue
1706                          * 4. receive_queue
1707                          *
1708                          * Each queue can be processed only if the next ones
1709                          * are empty. At this point we have empty receive_queue.
1710                          * But prequeue _can_ be not empty after 2nd iteration,
1711                          * when we jumped to start of loop because backlog
1712                          * processing added something to receive_queue.
1713                          * We cannot release_sock(), because backlog contains
1714                          * packets arrived _after_ prequeued ones.
1715                          *
1716                          * Shortly, algorithm is clear --- to process all
1717                          * the queues in order. We could make it more directly,
1718                          * requeueing packets from backlog to prequeue, if
1719                          * is not empty. It is more elegant, but eats cycles,
1720                          * unfortunately.
1721                          */
1722                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1723                                 goto do_prequeue;
1724
1725                         /* __ Set realtime policy in scheduler __ */
1726                 }
1727
1728                 if (copied >= target) {
1729                         /* Do not sleep, just process backlog. */
1730                         release_sock(sk);
1731                         lock_sock(sk);
1732                 } else
1733                         sk_wait_data(sk, &timeo);
1734
1735                 if (user_recv) {
1736                         int chunk;
1737
1738                         /* __ Restore normal policy in scheduler __ */
1739
1740                         if ((chunk = len - tp->ucopy.len) != 0) {
1741                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1742                                 len -= chunk;
1743                                 copied += chunk;
1744                         }
1745
1746                         if (tp->rcv_nxt == tp->copied_seq &&
1747                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1748 do_prequeue:
1749                                 tcp_prequeue_process(sk);
1750
1751                                 if ((chunk = len - tp->ucopy.len) != 0) {
1752                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1753                                         len -= chunk;
1754                                         copied += chunk;
1755                                 }
1756                         }
1757                 }
1758                 if ((flags & MSG_PEEK) &&
1759                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1760                         net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1761                                             current->comm,
1762                                             task_pid_nr(current));
1763                         peek_seq = tp->copied_seq;
1764                 }
1765                 continue;
1766
1767         found_ok_skb:
1768                 /* Ok so how much can we use? */
1769                 used = skb->len - offset;
1770                 if (len < used)
1771                         used = len;
1772
1773                 /* Do we have urgent data here? */
1774                 if (tp->urg_data) {
1775                         u32 urg_offset = tp->urg_seq - *seq;
1776                         if (urg_offset < used) {
1777                                 if (!urg_offset) {
1778                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1779                                                 ++*seq;
1780                                                 urg_hole++;
1781                                                 offset++;
1782                                                 used--;
1783                                                 if (!used)
1784                                                         goto skip_copy;
1785                                         }
1786                                 } else
1787                                         used = urg_offset;
1788                         }
1789                 }
1790
1791                 if (!(flags & MSG_TRUNC)) {
1792                         err = skb_copy_datagram_msg(skb, offset, msg, used);
1793                         if (err) {
1794                                 /* Exception. Bailout! */
1795                                 if (!copied)
1796                                         copied = -EFAULT;
1797                                 break;
1798                         }
1799                 }
1800
1801                 *seq += used;
1802                 copied += used;
1803                 len -= used;
1804
1805                 tcp_rcv_space_adjust(sk);
1806
1807 skip_copy:
1808                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1809                         tp->urg_data = 0;
1810                         tcp_fast_path_check(sk);
1811                 }
1812                 if (used + offset < skb->len)
1813                         continue;
1814
1815                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1816                         goto found_fin_ok;
1817                 if (!(flags & MSG_PEEK))
1818                         sk_eat_skb(sk, skb);
1819                 continue;
1820
1821         found_fin_ok:
1822                 /* Process the FIN. */
1823                 ++*seq;
1824                 if (!(flags & MSG_PEEK))
1825                         sk_eat_skb(sk, skb);
1826                 break;
1827         } while (len > 0);
1828
1829         if (user_recv) {
1830                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1831                         int chunk;
1832
1833                         tp->ucopy.len = copied > 0 ? len : 0;
1834
1835                         tcp_prequeue_process(sk);
1836
1837                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1838                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1839                                 len -= chunk;
1840                                 copied += chunk;
1841                         }
1842                 }
1843
1844                 tp->ucopy.task = NULL;
1845                 tp->ucopy.len = 0;
1846         }
1847
1848         /* According to UNIX98, msg_name/msg_namelen are ignored
1849          * on connected socket. I was just happy when found this 8) --ANK
1850          */
1851
1852         /* Clean up data we have read: This will do ACK frames. */
1853         tcp_cleanup_rbuf(sk, copied);
1854
1855         release_sock(sk);
1856         return copied;
1857
1858 out:
1859         release_sock(sk);
1860         return err;
1861
1862 recv_urg:
1863         err = tcp_recv_urg(sk, msg, len, flags);
1864         goto out;
1865
1866 recv_sndq:
1867         err = tcp_peek_sndq(sk, msg, len);
1868         goto out;
1869 }
1870 EXPORT_SYMBOL(tcp_recvmsg);
1871
1872 void tcp_set_state(struct sock *sk, int state)
1873 {
1874         int oldstate = sk->sk_state;
1875
1876         switch (state) {
1877         case TCP_ESTABLISHED:
1878                 if (oldstate != TCP_ESTABLISHED)
1879                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1880                 break;
1881
1882         case TCP_CLOSE:
1883                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1884                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1885
1886                 sk->sk_prot->unhash(sk);
1887                 if (inet_csk(sk)->icsk_bind_hash &&
1888                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1889                         inet_put_port(sk);
1890                 /* fall through */
1891         default:
1892                 if (oldstate == TCP_ESTABLISHED)
1893                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1894         }
1895
1896         /* Change state AFTER socket is unhashed to avoid closed
1897          * socket sitting in hash tables.
1898          */
1899         sk->sk_state = state;
1900
1901 #ifdef STATE_TRACE
1902         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1903 #endif
1904 }
1905 EXPORT_SYMBOL_GPL(tcp_set_state);
1906
1907 /*
1908  *      State processing on a close. This implements the state shift for
1909  *      sending our FIN frame. Note that we only send a FIN for some
1910  *      states. A shutdown() may have already sent the FIN, or we may be
1911  *      closed.
1912  */
1913
1914 static const unsigned char new_state[16] = {
1915   /* current state:        new state:      action:      */
1916   [0 /* (Invalid) */]   = TCP_CLOSE,
1917   [TCP_ESTABLISHED]     = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1918   [TCP_SYN_SENT]        = TCP_CLOSE,
1919   [TCP_SYN_RECV]        = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1920   [TCP_FIN_WAIT1]       = TCP_FIN_WAIT1,
1921   [TCP_FIN_WAIT2]       = TCP_FIN_WAIT2,
1922   [TCP_TIME_WAIT]       = TCP_CLOSE,
1923   [TCP_CLOSE]           = TCP_CLOSE,
1924   [TCP_CLOSE_WAIT]      = TCP_LAST_ACK  | TCP_ACTION_FIN,
1925   [TCP_LAST_ACK]        = TCP_LAST_ACK,
1926   [TCP_LISTEN]          = TCP_CLOSE,
1927   [TCP_CLOSING]         = TCP_CLOSING,
1928   [TCP_NEW_SYN_RECV]    = TCP_CLOSE,    /* should not happen ! */
1929 };
1930
1931 static int tcp_close_state(struct sock *sk)
1932 {
1933         int next = (int)new_state[sk->sk_state];
1934         int ns = next & TCP_STATE_MASK;
1935
1936         tcp_set_state(sk, ns);
1937
1938         return next & TCP_ACTION_FIN;
1939 }
1940
1941 /*
1942  *      Shutdown the sending side of a connection. Much like close except
1943  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1944  */
1945
1946 void tcp_shutdown(struct sock *sk, int how)
1947 {
1948         /*      We need to grab some memory, and put together a FIN,
1949          *      and then put it into the queue to be sent.
1950          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1951          */
1952         if (!(how & SEND_SHUTDOWN))
1953                 return;
1954
1955         /* If we've already sent a FIN, or it's a closed state, skip this. */
1956         if ((1 << sk->sk_state) &
1957             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1958              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1959                 /* Clear out any half completed packets.  FIN if needed. */
1960                 if (tcp_close_state(sk))
1961                         tcp_send_fin(sk);
1962         }
1963 }
1964 EXPORT_SYMBOL(tcp_shutdown);
1965
1966 bool tcp_check_oom(struct sock *sk, int shift)
1967 {
1968         bool too_many_orphans, out_of_socket_memory;
1969
1970         too_many_orphans = tcp_too_many_orphans(sk, shift);
1971         out_of_socket_memory = tcp_out_of_memory(sk);
1972
1973         if (too_many_orphans)
1974                 net_info_ratelimited("too many orphaned sockets\n");
1975         if (out_of_socket_memory)
1976                 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
1977         return too_many_orphans || out_of_socket_memory;
1978 }
1979
1980 void tcp_close(struct sock *sk, long timeout)
1981 {
1982         struct sk_buff *skb;
1983         int data_was_unread = 0;
1984         int state;
1985
1986         lock_sock(sk);
1987         sk->sk_shutdown = SHUTDOWN_MASK;
1988
1989         if (sk->sk_state == TCP_LISTEN) {
1990                 tcp_set_state(sk, TCP_CLOSE);
1991
1992                 /* Special case. */
1993                 inet_csk_listen_stop(sk);
1994
1995                 goto adjudge_to_death;
1996         }
1997
1998         /*  We need to flush the recv. buffs.  We do this only on the
1999          *  descriptor close, not protocol-sourced closes, because the
2000          *  reader process may not have drained the data yet!
2001          */
2002         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2003                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2004
2005                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2006                         len--;
2007                 data_was_unread += len;
2008                 __kfree_skb(skb);
2009         }
2010
2011         sk_mem_reclaim(sk);
2012
2013         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2014         if (sk->sk_state == TCP_CLOSE)
2015                 goto adjudge_to_death;
2016
2017         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2018          * data was lost. To witness the awful effects of the old behavior of
2019          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2020          * GET in an FTP client, suspend the process, wait for the client to
2021          * advertise a zero window, then kill -9 the FTP client, wheee...
2022          * Note: timeout is always zero in such a case.
2023          */
2024         if (unlikely(tcp_sk(sk)->repair)) {
2025                 sk->sk_prot->disconnect(sk, 0);
2026         } else if (data_was_unread) {
2027                 /* Unread data was tossed, zap the connection. */
2028                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2029                 tcp_set_state(sk, TCP_CLOSE);
2030                 tcp_send_active_reset(sk, sk->sk_allocation);
2031         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2032                 /* Check zero linger _after_ checking for unread data. */
2033                 sk->sk_prot->disconnect(sk, 0);
2034                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2035         } else if (tcp_close_state(sk)) {
2036                 /* We FIN if the application ate all the data before
2037                  * zapping the connection.
2038                  */
2039
2040                 /* RED-PEN. Formally speaking, we have broken TCP state
2041                  * machine. State transitions:
2042                  *
2043                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2044                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2045                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2046                  *
2047                  * are legal only when FIN has been sent (i.e. in window),
2048                  * rather than queued out of window. Purists blame.
2049                  *
2050                  * F.e. "RFC state" is ESTABLISHED,
2051                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2052                  *
2053                  * The visible declinations are that sometimes
2054                  * we enter time-wait state, when it is not required really
2055                  * (harmless), do not send active resets, when they are
2056                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2057                  * they look as CLOSING or LAST_ACK for Linux)
2058                  * Probably, I missed some more holelets.
2059                  *                                              --ANK
2060                  * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2061                  * in a single packet! (May consider it later but will
2062                  * probably need API support or TCP_CORK SYN-ACK until
2063                  * data is written and socket is closed.)
2064                  */
2065                 tcp_send_fin(sk);
2066         }
2067
2068         sk_stream_wait_close(sk, timeout);
2069
2070 adjudge_to_death:
2071         state = sk->sk_state;
2072         sock_hold(sk);
2073         sock_orphan(sk);
2074
2075         /* It is the last release_sock in its life. It will remove backlog. */
2076         release_sock(sk);
2077
2078
2079         /* Now socket is owned by kernel and we acquire BH lock
2080            to finish close. No need to check for user refs.
2081          */
2082         local_bh_disable();
2083         bh_lock_sock(sk);
2084         WARN_ON(sock_owned_by_user(sk));
2085
2086         percpu_counter_inc(sk->sk_prot->orphan_count);
2087
2088         /* Have we already been destroyed by a softirq or backlog? */
2089         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2090                 goto out;
2091
2092         /*      This is a (useful) BSD violating of the RFC. There is a
2093          *      problem with TCP as specified in that the other end could
2094          *      keep a socket open forever with no application left this end.
2095          *      We use a 1 minute timeout (about the same as BSD) then kill
2096          *      our end. If they send after that then tough - BUT: long enough
2097          *      that we won't make the old 4*rto = almost no time - whoops
2098          *      reset mistake.
2099          *
2100          *      Nope, it was not mistake. It is really desired behaviour
2101          *      f.e. on http servers, when such sockets are useless, but
2102          *      consume significant resources. Let's do it with special
2103          *      linger2 option.                                 --ANK
2104          */
2105
2106         if (sk->sk_state == TCP_FIN_WAIT2) {
2107                 struct tcp_sock *tp = tcp_sk(sk);
2108                 if (tp->linger2 < 0) {
2109                         tcp_set_state(sk, TCP_CLOSE);
2110                         tcp_send_active_reset(sk, GFP_ATOMIC);
2111                         NET_INC_STATS_BH(sock_net(sk),
2112                                         LINUX_MIB_TCPABORTONLINGER);
2113                 } else {
2114                         const int tmo = tcp_fin_time(sk);
2115
2116                         if (tmo > TCP_TIMEWAIT_LEN) {
2117                                 inet_csk_reset_keepalive_timer(sk,
2118                                                 tmo - TCP_TIMEWAIT_LEN);
2119                         } else {
2120                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2121                                 goto out;
2122                         }
2123                 }
2124         }
2125         if (sk->sk_state != TCP_CLOSE) {
2126                 sk_mem_reclaim(sk);
2127                 if (tcp_check_oom(sk, 0)) {
2128                         tcp_set_state(sk, TCP_CLOSE);
2129                         tcp_send_active_reset(sk, GFP_ATOMIC);
2130                         NET_INC_STATS_BH(sock_net(sk),
2131                                         LINUX_MIB_TCPABORTONMEMORY);
2132                 }
2133         }
2134
2135         if (sk->sk_state == TCP_CLOSE) {
2136                 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2137                 /* We could get here with a non-NULL req if the socket is
2138                  * aborted (e.g., closed with unread data) before 3WHS
2139                  * finishes.
2140                  */
2141                 if (req)
2142                         reqsk_fastopen_remove(sk, req, false);
2143                 inet_csk_destroy_sock(sk);
2144         }
2145         /* Otherwise, socket is reprieved until protocol close. */
2146
2147 out:
2148         bh_unlock_sock(sk);
2149         local_bh_enable();
2150         sock_put(sk);
2151 }
2152 EXPORT_SYMBOL(tcp_close);
2153
2154 /* These states need RST on ABORT according to RFC793 */
2155
2156 static inline bool tcp_need_reset(int state)
2157 {
2158         return (1 << state) &
2159                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2160                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2161 }
2162
2163 int tcp_disconnect(struct sock *sk, int flags)
2164 {
2165         struct inet_sock *inet = inet_sk(sk);
2166         struct inet_connection_sock *icsk = inet_csk(sk);
2167         struct tcp_sock *tp = tcp_sk(sk);
2168         int err = 0;
2169         int old_state = sk->sk_state;
2170
2171         if (old_state != TCP_CLOSE)
2172                 tcp_set_state(sk, TCP_CLOSE);
2173
2174         /* ABORT function of RFC793 */
2175         if (old_state == TCP_LISTEN) {
2176                 inet_csk_listen_stop(sk);
2177         } else if (unlikely(tp->repair)) {
2178                 sk->sk_err = ECONNABORTED;
2179         } else if (tcp_need_reset(old_state) ||
2180                    (tp->snd_nxt != tp->write_seq &&
2181                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2182                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2183                  * states
2184                  */
2185                 tcp_send_active_reset(sk, gfp_any());
2186                 sk->sk_err = ECONNRESET;
2187         } else if (old_state == TCP_SYN_SENT)
2188                 sk->sk_err = ECONNRESET;
2189
2190         tcp_clear_xmit_timers(sk);
2191         __skb_queue_purge(&sk->sk_receive_queue);
2192         tcp_write_queue_purge(sk);
2193         __skb_queue_purge(&tp->out_of_order_queue);
2194
2195         inet->inet_dport = 0;
2196
2197         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2198                 inet_reset_saddr(sk);
2199
2200         sk->sk_shutdown = 0;
2201         sock_reset_flag(sk, SOCK_DONE);
2202         tp->srtt_us = 0;
2203         if ((tp->write_seq += tp->max_window + 2) == 0)
2204                 tp->write_seq = 1;
2205         icsk->icsk_backoff = 0;
2206         tp->snd_cwnd = 2;
2207         icsk->icsk_probes_out = 0;
2208         tp->packets_out = 0;
2209         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2210         tp->snd_cwnd_cnt = 0;
2211         tp->window_clamp = 0;
2212         tcp_set_ca_state(sk, TCP_CA_Open);
2213         tcp_clear_retrans(tp);
2214         inet_csk_delack_init(sk);
2215         tcp_init_send_head(sk);
2216         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2217         __sk_dst_reset(sk);
2218
2219         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2220
2221         sk->sk_error_report(sk);
2222         return err;
2223 }
2224 EXPORT_SYMBOL(tcp_disconnect);
2225
2226 void tcp_sock_destruct(struct sock *sk)
2227 {
2228         inet_sock_destruct(sk);
2229
2230         kfree(inet_csk(sk)->icsk_accept_queue.fastopenq);
2231 }
2232
2233 static inline bool tcp_can_repair_sock(const struct sock *sk)
2234 {
2235         return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2236                 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2237 }
2238
2239 static int tcp_repair_options_est(struct tcp_sock *tp,
2240                 struct tcp_repair_opt __user *optbuf, unsigned int len)
2241 {
2242         struct tcp_repair_opt opt;
2243
2244         while (len >= sizeof(opt)) {
2245                 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2246                         return -EFAULT;
2247
2248                 optbuf++;
2249                 len -= sizeof(opt);
2250
2251                 switch (opt.opt_code) {
2252                 case TCPOPT_MSS:
2253                         tp->rx_opt.mss_clamp = opt.opt_val;
2254                         break;
2255                 case TCPOPT_WINDOW:
2256                         {
2257                                 u16 snd_wscale = opt.opt_val & 0xFFFF;
2258                                 u16 rcv_wscale = opt.opt_val >> 16;
2259
2260                                 if (snd_wscale > 14 || rcv_wscale > 14)
2261                                         return -EFBIG;
2262
2263                                 tp->rx_opt.snd_wscale = snd_wscale;
2264                                 tp->rx_opt.rcv_wscale = rcv_wscale;
2265                                 tp->rx_opt.wscale_ok = 1;
2266                         }
2267                         break;
2268                 case TCPOPT_SACK_PERM:
2269                         if (opt.opt_val != 0)
2270                                 return -EINVAL;
2271
2272                         tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2273                         if (sysctl_tcp_fack)
2274                                 tcp_enable_fack(tp);
2275                         break;
2276                 case TCPOPT_TIMESTAMP:
2277                         if (opt.opt_val != 0)
2278                                 return -EINVAL;
2279
2280                         tp->rx_opt.tstamp_ok = 1;
2281                         break;
2282                 }
2283         }
2284
2285         return 0;
2286 }
2287
2288 /*
2289  *      Socket option code for TCP.
2290  */
2291 static int do_tcp_setsockopt(struct sock *sk, int level,
2292                 int optname, char __user *optval, unsigned int optlen)
2293 {
2294         struct tcp_sock *tp = tcp_sk(sk);
2295         struct inet_connection_sock *icsk = inet_csk(sk);
2296         int val;
2297         int err = 0;
2298
2299         /* These are data/string values, all the others are ints */
2300         switch (optname) {
2301         case TCP_CONGESTION: {
2302                 char name[TCP_CA_NAME_MAX];
2303
2304                 if (optlen < 1)
2305                         return -EINVAL;
2306
2307                 val = strncpy_from_user(name, optval,
2308                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2309                 if (val < 0)
2310                         return -EFAULT;
2311                 name[val] = 0;
2312
2313                 lock_sock(sk);
2314                 err = tcp_set_congestion_control(sk, name);
2315                 release_sock(sk);
2316                 return err;
2317         }
2318         default:
2319                 /* fallthru */
2320                 break;
2321         }
2322
2323         if (optlen < sizeof(int))
2324                 return -EINVAL;
2325
2326         if (get_user(val, (int __user *)optval))
2327                 return -EFAULT;
2328
2329         lock_sock(sk);
2330
2331         switch (optname) {
2332         case TCP_MAXSEG:
2333                 /* Values greater than interface MTU won't take effect. However
2334                  * at the point when this call is done we typically don't yet
2335                  * know which interface is going to be used */
2336                 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2337                         err = -EINVAL;
2338                         break;
2339                 }
2340                 tp->rx_opt.user_mss = val;
2341                 break;
2342
2343         case TCP_NODELAY:
2344                 if (val) {
2345                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2346                          * this option on corked socket is remembered, but
2347                          * it is not activated until cork is cleared.
2348                          *
2349                          * However, when TCP_NODELAY is set we make
2350                          * an explicit push, which overrides even TCP_CORK
2351                          * for currently queued segments.
2352                          */
2353                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2354                         tcp_push_pending_frames(sk);
2355                 } else {
2356                         tp->nonagle &= ~TCP_NAGLE_OFF;
2357                 }
2358                 break;
2359
2360         case TCP_THIN_LINEAR_TIMEOUTS:
2361                 if (val < 0 || val > 1)
2362                         err = -EINVAL;
2363                 else
2364                         tp->thin_lto = val;
2365                 break;
2366
2367         case TCP_THIN_DUPACK:
2368                 if (val < 0 || val > 1)
2369                         err = -EINVAL;
2370                 else {
2371                         tp->thin_dupack = val;
2372                         if (tp->thin_dupack)
2373                                 tcp_disable_early_retrans(tp);
2374                 }
2375                 break;
2376
2377         case TCP_REPAIR:
2378                 if (!tcp_can_repair_sock(sk))
2379                         err = -EPERM;
2380                 else if (val == 1) {
2381                         tp->repair = 1;
2382                         sk->sk_reuse = SK_FORCE_REUSE;
2383                         tp->repair_queue = TCP_NO_QUEUE;
2384                 } else if (val == 0) {
2385                         tp->repair = 0;
2386                         sk->sk_reuse = SK_NO_REUSE;
2387                         tcp_send_window_probe(sk);
2388                 } else
2389                         err = -EINVAL;
2390
2391                 break;
2392
2393         case TCP_REPAIR_QUEUE:
2394                 if (!tp->repair)
2395                         err = -EPERM;
2396                 else if (val < TCP_QUEUES_NR)
2397                         tp->repair_queue = val;
2398                 else
2399                         err = -EINVAL;
2400                 break;
2401
2402         case TCP_QUEUE_SEQ:
2403                 if (sk->sk_state != TCP_CLOSE)
2404                         err = -EPERM;
2405                 else if (tp->repair_queue == TCP_SEND_QUEUE)
2406                         tp->write_seq = val;
2407                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2408                         tp->rcv_nxt = val;
2409                 else
2410                         err = -EINVAL;
2411                 break;
2412
2413         case TCP_REPAIR_OPTIONS:
2414                 if (!tp->repair)
2415                         err = -EINVAL;
2416                 else if (sk->sk_state == TCP_ESTABLISHED)
2417                         err = tcp_repair_options_est(tp,
2418                                         (struct tcp_repair_opt __user *)optval,
2419                                         optlen);
2420                 else
2421                         err = -EPERM;
2422                 break;
2423
2424         case TCP_CORK:
2425                 /* When set indicates to always queue non-full frames.
2426                  * Later the user clears this option and we transmit
2427                  * any pending partial frames in the queue.  This is
2428                  * meant to be used alongside sendfile() to get properly
2429                  * filled frames when the user (for example) must write
2430                  * out headers with a write() call first and then use
2431                  * sendfile to send out the data parts.
2432                  *
2433                  * TCP_CORK can be set together with TCP_NODELAY and it is
2434                  * stronger than TCP_NODELAY.
2435                  */
2436                 if (val) {
2437                         tp->nonagle |= TCP_NAGLE_CORK;
2438                 } else {
2439                         tp->nonagle &= ~TCP_NAGLE_CORK;
2440                         if (tp->nonagle&TCP_NAGLE_OFF)
2441                                 tp->nonagle |= TCP_NAGLE_PUSH;
2442                         tcp_push_pending_frames(sk);
2443                 }
2444                 break;
2445
2446         case TCP_KEEPIDLE:
2447                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2448                         err = -EINVAL;
2449                 else {
2450                         tp->keepalive_time = val * HZ;
2451                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2452                             !((1 << sk->sk_state) &
2453                               (TCPF_CLOSE | TCPF_LISTEN))) {
2454                                 u32 elapsed = keepalive_time_elapsed(tp);
2455                                 if (tp->keepalive_time > elapsed)
2456                                         elapsed = tp->keepalive_time - elapsed;
2457                                 else
2458                                         elapsed = 0;
2459                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2460                         }
2461                 }
2462                 break;
2463         case TCP_KEEPINTVL:
2464                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2465                         err = -EINVAL;
2466                 else
2467                         tp->keepalive_intvl = val * HZ;
2468                 break;
2469         case TCP_KEEPCNT:
2470                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2471                         err = -EINVAL;
2472                 else
2473                         tp->keepalive_probes = val;
2474                 break;
2475         case TCP_SYNCNT:
2476                 if (val < 1 || val > MAX_TCP_SYNCNT)
2477                         err = -EINVAL;
2478                 else
2479                         icsk->icsk_syn_retries = val;
2480                 break;
2481
2482         case TCP_LINGER2:
2483                 if (val < 0)
2484                         tp->linger2 = -1;
2485                 else if (val > sysctl_tcp_fin_timeout / HZ)
2486                         tp->linger2 = 0;
2487                 else
2488                         tp->linger2 = val * HZ;
2489                 break;
2490
2491         case TCP_DEFER_ACCEPT:
2492                 /* Translate value in seconds to number of retransmits */
2493                 icsk->icsk_accept_queue.rskq_defer_accept =
2494                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2495                                         TCP_RTO_MAX / HZ);
2496                 break;
2497
2498         case TCP_WINDOW_CLAMP:
2499                 if (!val) {
2500                         if (sk->sk_state != TCP_CLOSE) {
2501                                 err = -EINVAL;
2502                                 break;
2503                         }
2504                         tp->window_clamp = 0;
2505                 } else
2506                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2507                                                 SOCK_MIN_RCVBUF / 2 : val;
2508                 break;
2509
2510         case TCP_QUICKACK:
2511                 if (!val) {
2512                         icsk->icsk_ack.pingpong = 1;
2513                 } else {
2514                         icsk->icsk_ack.pingpong = 0;
2515                         if ((1 << sk->sk_state) &
2516                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2517                             inet_csk_ack_scheduled(sk)) {
2518                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2519                                 tcp_cleanup_rbuf(sk, 1);
2520                                 if (!(val & 1))
2521                                         icsk->icsk_ack.pingpong = 1;
2522                         }
2523                 }
2524                 break;
2525
2526 #ifdef CONFIG_TCP_MD5SIG
2527         case TCP_MD5SIG:
2528                 /* Read the IP->Key mappings from userspace */
2529                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2530                 break;
2531 #endif
2532         case TCP_USER_TIMEOUT:
2533                 /* Cap the max time in ms TCP will retry or probe the window
2534                  * before giving up and aborting (ETIMEDOUT) a connection.
2535                  */
2536                 if (val < 0)
2537                         err = -EINVAL;
2538                 else
2539                         icsk->icsk_user_timeout = msecs_to_jiffies(val);
2540                 break;
2541
2542         case TCP_FASTOPEN:
2543                 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2544                     TCPF_LISTEN)))
2545                         err = fastopen_init_queue(sk, val);
2546                 else
2547                         err = -EINVAL;
2548                 break;
2549         case TCP_TIMESTAMP:
2550                 if (!tp->repair)
2551                         err = -EPERM;
2552                 else
2553                         tp->tsoffset = val - tcp_time_stamp;
2554                 break;
2555         case TCP_NOTSENT_LOWAT:
2556                 tp->notsent_lowat = val;
2557                 sk->sk_write_space(sk);
2558                 break;
2559         default:
2560                 err = -ENOPROTOOPT;
2561                 break;
2562         }
2563
2564         release_sock(sk);
2565         return err;
2566 }
2567
2568 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2569                    unsigned int optlen)
2570 {
2571         const struct inet_connection_sock *icsk = inet_csk(sk);
2572
2573         if (level != SOL_TCP)
2574                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2575                                                      optval, optlen);
2576         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2577 }
2578 EXPORT_SYMBOL(tcp_setsockopt);
2579
2580 #ifdef CONFIG_COMPAT
2581 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2582                           char __user *optval, unsigned int optlen)
2583 {
2584         if (level != SOL_TCP)
2585                 return inet_csk_compat_setsockopt(sk, level, optname,
2586                                                   optval, optlen);
2587         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2588 }
2589 EXPORT_SYMBOL(compat_tcp_setsockopt);
2590 #endif
2591
2592 /* Return information about state of tcp endpoint in API format. */
2593 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2594 {
2595         const struct tcp_sock *tp = tcp_sk(sk);
2596         const struct inet_connection_sock *icsk = inet_csk(sk);
2597         u32 now = tcp_time_stamp;
2598
2599         memset(info, 0, sizeof(*info));
2600
2601         info->tcpi_state = sk->sk_state;
2602         info->tcpi_ca_state = icsk->icsk_ca_state;
2603         info->tcpi_retransmits = icsk->icsk_retransmits;
2604         info->tcpi_probes = icsk->icsk_probes_out;
2605         info->tcpi_backoff = icsk->icsk_backoff;
2606
2607         if (tp->rx_opt.tstamp_ok)
2608                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2609         if (tcp_is_sack(tp))
2610                 info->tcpi_options |= TCPI_OPT_SACK;
2611         if (tp->rx_opt.wscale_ok) {
2612                 info->tcpi_options |= TCPI_OPT_WSCALE;
2613                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2614                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2615         }
2616
2617         if (tp->ecn_flags & TCP_ECN_OK)
2618                 info->tcpi_options |= TCPI_OPT_ECN;
2619         if (tp->ecn_flags & TCP_ECN_SEEN)
2620                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2621         if (tp->syn_data_acked)
2622                 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2623
2624         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2625         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2626         info->tcpi_snd_mss = tp->mss_cache;
2627         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2628
2629         if (sk->sk_state == TCP_LISTEN) {
2630                 info->tcpi_unacked = sk->sk_ack_backlog;
2631                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2632         } else {
2633                 info->tcpi_unacked = tp->packets_out;
2634                 info->tcpi_sacked = tp->sacked_out;
2635         }
2636         info->tcpi_lost = tp->lost_out;
2637         info->tcpi_retrans = tp->retrans_out;
2638         info->tcpi_fackets = tp->fackets_out;
2639
2640         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2641         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2642         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2643
2644         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2645         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2646         info->tcpi_rtt = tp->srtt_us >> 3;
2647         info->tcpi_rttvar = tp->mdev_us >> 2;
2648         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2649         info->tcpi_snd_cwnd = tp->snd_cwnd;
2650         info->tcpi_advmss = tp->advmss;
2651         info->tcpi_reordering = tp->reordering;
2652
2653         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2654         info->tcpi_rcv_space = tp->rcvq_space.space;
2655
2656         info->tcpi_total_retrans = tp->total_retrans;
2657
2658         info->tcpi_pacing_rate = sk->sk_pacing_rate != ~0U ?
2659                                         sk->sk_pacing_rate : ~0ULL;
2660         info->tcpi_max_pacing_rate = sk->sk_max_pacing_rate != ~0U ?
2661                                         sk->sk_max_pacing_rate : ~0ULL;
2662 }
2663 EXPORT_SYMBOL_GPL(tcp_get_info);
2664
2665 static int do_tcp_getsockopt(struct sock *sk, int level,
2666                 int optname, char __user *optval, int __user *optlen)
2667 {
2668         struct inet_connection_sock *icsk = inet_csk(sk);
2669         struct tcp_sock *tp = tcp_sk(sk);
2670         int val, len;
2671
2672         if (get_user(len, optlen))
2673                 return -EFAULT;
2674
2675         len = min_t(unsigned int, len, sizeof(int));
2676
2677         if (len < 0)
2678                 return -EINVAL;
2679
2680         switch (optname) {
2681         case TCP_MAXSEG:
2682                 val = tp->mss_cache;
2683                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2684                         val = tp->rx_opt.user_mss;
2685                 if (tp->repair)
2686                         val = tp->rx_opt.mss_clamp;
2687                 break;
2688         case TCP_NODELAY:
2689                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2690                 break;
2691         case TCP_CORK:
2692                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2693                 break;
2694         case TCP_KEEPIDLE:
2695                 val = keepalive_time_when(tp) / HZ;
2696                 break;
2697         case TCP_KEEPINTVL:
2698                 val = keepalive_intvl_when(tp) / HZ;
2699                 break;
2700         case TCP_KEEPCNT:
2701                 val = keepalive_probes(tp);
2702                 break;
2703         case TCP_SYNCNT:
2704                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2705                 break;
2706         case TCP_LINGER2:
2707                 val = tp->linger2;
2708                 if (val >= 0)
2709                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2710                 break;
2711         case TCP_DEFER_ACCEPT:
2712                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2713                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2714                 break;
2715         case TCP_WINDOW_CLAMP:
2716                 val = tp->window_clamp;
2717                 break;
2718         case TCP_INFO: {
2719                 struct tcp_info info;
2720
2721                 if (get_user(len, optlen))
2722                         return -EFAULT;
2723
2724                 tcp_get_info(sk, &info);
2725
2726                 len = min_t(unsigned int, len, sizeof(info));
2727                 if (put_user(len, optlen))
2728                         return -EFAULT;
2729                 if (copy_to_user(optval, &info, len))
2730                         return -EFAULT;
2731                 return 0;
2732         }
2733         case TCP_QUICKACK:
2734                 val = !icsk->icsk_ack.pingpong;
2735                 break;
2736
2737         case TCP_CONGESTION:
2738                 if (get_user(len, optlen))
2739                         return -EFAULT;
2740                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2741                 if (put_user(len, optlen))
2742                         return -EFAULT;
2743                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2744                         return -EFAULT;
2745                 return 0;
2746
2747         case TCP_THIN_LINEAR_TIMEOUTS:
2748                 val = tp->thin_lto;
2749                 break;
2750         case TCP_THIN_DUPACK:
2751                 val = tp->thin_dupack;
2752                 break;
2753
2754         case TCP_REPAIR:
2755                 val = tp->repair;
2756                 break;
2757
2758         case TCP_REPAIR_QUEUE:
2759                 if (tp->repair)
2760                         val = tp->repair_queue;
2761                 else
2762                         return -EINVAL;
2763                 break;
2764
2765         case TCP_QUEUE_SEQ:
2766                 if (tp->repair_queue == TCP_SEND_QUEUE)
2767                         val = tp->write_seq;
2768                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2769                         val = tp->rcv_nxt;
2770                 else
2771                         return -EINVAL;
2772                 break;
2773
2774         case TCP_USER_TIMEOUT:
2775                 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2776                 break;
2777
2778         case TCP_FASTOPEN:
2779                 if (icsk->icsk_accept_queue.fastopenq)
2780                         val = icsk->icsk_accept_queue.fastopenq->max_qlen;
2781                 else
2782                         val = 0;
2783                 break;
2784
2785         case TCP_TIMESTAMP:
2786                 val = tcp_time_stamp + tp->tsoffset;
2787                 break;
2788         case TCP_NOTSENT_LOWAT:
2789                 val = tp->notsent_lowat;
2790                 break;
2791         default:
2792                 return -ENOPROTOOPT;
2793         }
2794
2795         if (put_user(len, optlen))
2796                 return -EFAULT;
2797         if (copy_to_user(optval, &val, len))
2798                 return -EFAULT;
2799         return 0;
2800 }
2801
2802 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2803                    int __user *optlen)
2804 {
2805         struct inet_connection_sock *icsk = inet_csk(sk);
2806
2807         if (level != SOL_TCP)
2808                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2809                                                      optval, optlen);
2810         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2811 }
2812 EXPORT_SYMBOL(tcp_getsockopt);
2813
2814 #ifdef CONFIG_COMPAT
2815 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2816                           char __user *optval, int __user *optlen)
2817 {
2818         if (level != SOL_TCP)
2819                 return inet_csk_compat_getsockopt(sk, level, optname,
2820                                                   optval, optlen);
2821         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2822 }
2823 EXPORT_SYMBOL(compat_tcp_getsockopt);
2824 #endif
2825
2826 #ifdef CONFIG_TCP_MD5SIG
2827 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
2828 static DEFINE_MUTEX(tcp_md5sig_mutex);
2829 static bool tcp_md5sig_pool_populated = false;
2830
2831 static void __tcp_alloc_md5sig_pool(void)
2832 {
2833         int cpu;
2834
2835         for_each_possible_cpu(cpu) {
2836                 if (!per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm) {
2837                         struct crypto_hash *hash;
2838
2839                         hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2840                         if (IS_ERR_OR_NULL(hash))
2841                                 return;
2842                         per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm = hash;
2843                 }
2844         }
2845         /* before setting tcp_md5sig_pool_populated, we must commit all writes
2846          * to memory. See smp_rmb() in tcp_get_md5sig_pool()
2847          */
2848         smp_wmb();
2849         tcp_md5sig_pool_populated = true;
2850 }
2851
2852 bool tcp_alloc_md5sig_pool(void)
2853 {
2854         if (unlikely(!tcp_md5sig_pool_populated)) {
2855                 mutex_lock(&tcp_md5sig_mutex);
2856
2857                 if (!tcp_md5sig_pool_populated)
2858                         __tcp_alloc_md5sig_pool();
2859
2860                 mutex_unlock(&tcp_md5sig_mutex);
2861         }
2862         return tcp_md5sig_pool_populated;
2863 }
2864 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2865
2866
2867 /**
2868  *      tcp_get_md5sig_pool - get md5sig_pool for this user
2869  *
2870  *      We use percpu structure, so if we succeed, we exit with preemption
2871  *      and BH disabled, to make sure another thread or softirq handling
2872  *      wont try to get same context.
2873  */
2874 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2875 {
2876         local_bh_disable();
2877
2878         if (tcp_md5sig_pool_populated) {
2879                 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
2880                 smp_rmb();
2881                 return this_cpu_ptr(&tcp_md5sig_pool);
2882         }
2883         local_bh_enable();
2884         return NULL;
2885 }
2886 EXPORT_SYMBOL(tcp_get_md5sig_pool);
2887
2888 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2889                         const struct tcphdr *th)
2890 {
2891         struct scatterlist sg;
2892         struct tcphdr hdr;
2893         int err;
2894
2895         /* We are not allowed to change tcphdr, make a local copy */
2896         memcpy(&hdr, th, sizeof(hdr));
2897         hdr.check = 0;
2898
2899         /* options aren't included in the hash */
2900         sg_init_one(&sg, &hdr, sizeof(hdr));
2901         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
2902         return err;
2903 }
2904 EXPORT_SYMBOL(tcp_md5_hash_header);
2905
2906 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2907                           const struct sk_buff *skb, unsigned int header_len)
2908 {
2909         struct scatterlist sg;
2910         const struct tcphdr *tp = tcp_hdr(skb);
2911         struct hash_desc *desc = &hp->md5_desc;
2912         unsigned int i;
2913         const unsigned int head_data_len = skb_headlen(skb) > header_len ?
2914                                            skb_headlen(skb) - header_len : 0;
2915         const struct skb_shared_info *shi = skb_shinfo(skb);
2916         struct sk_buff *frag_iter;
2917
2918         sg_init_table(&sg, 1);
2919
2920         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
2921         if (crypto_hash_update(desc, &sg, head_data_len))
2922                 return 1;
2923
2924         for (i = 0; i < shi->nr_frags; ++i) {
2925                 const struct skb_frag_struct *f = &shi->frags[i];
2926                 unsigned int offset = f->page_offset;
2927                 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
2928
2929                 sg_set_page(&sg, page, skb_frag_size(f),
2930                             offset_in_page(offset));
2931                 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
2932                         return 1;
2933         }
2934
2935         skb_walk_frags(skb, frag_iter)
2936                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
2937                         return 1;
2938
2939         return 0;
2940 }
2941 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
2942
2943 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
2944 {
2945         struct scatterlist sg;
2946
2947         sg_init_one(&sg, key->key, key->keylen);
2948         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
2949 }
2950 EXPORT_SYMBOL(tcp_md5_hash_key);
2951
2952 #endif
2953
2954 void tcp_done(struct sock *sk)
2955 {
2956         struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2957
2958         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2959                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
2960
2961         tcp_set_state(sk, TCP_CLOSE);
2962         tcp_clear_xmit_timers(sk);
2963         if (req)
2964                 reqsk_fastopen_remove(sk, req, false);
2965
2966         sk->sk_shutdown = SHUTDOWN_MASK;
2967
2968         if (!sock_flag(sk, SOCK_DEAD))
2969                 sk->sk_state_change(sk);
2970         else
2971                 inet_csk_destroy_sock(sk);
2972 }
2973 EXPORT_SYMBOL_GPL(tcp_done);
2974
2975 extern struct tcp_congestion_ops tcp_reno;
2976
2977 static __initdata unsigned long thash_entries;
2978 static int __init set_thash_entries(char *str)
2979 {
2980         ssize_t ret;
2981
2982         if (!str)
2983                 return 0;
2984
2985         ret = kstrtoul(str, 0, &thash_entries);
2986         if (ret)
2987                 return 0;
2988
2989         return 1;
2990 }
2991 __setup("thash_entries=", set_thash_entries);
2992
2993 static void __init tcp_init_mem(void)
2994 {
2995         unsigned long limit = nr_free_buffer_pages() / 8;
2996         limit = max(limit, 128UL);
2997         sysctl_tcp_mem[0] = limit / 4 * 3;
2998         sysctl_tcp_mem[1] = limit;
2999         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3000 }
3001
3002 void __init tcp_init(void)
3003 {
3004         unsigned long limit;
3005         int max_rshare, max_wshare, cnt;
3006         unsigned int i;
3007
3008         sock_skb_cb_check_size(sizeof(struct tcp_skb_cb));
3009
3010         percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3011         percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3012         tcp_hashinfo.bind_bucket_cachep =
3013                 kmem_cache_create("tcp_bind_bucket",
3014                                   sizeof(struct inet_bind_bucket), 0,
3015                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3016
3017         /* Size and allocate the main established and bind bucket
3018          * hash tables.
3019          *
3020          * The methodology is similar to that of the buffer cache.
3021          */
3022         tcp_hashinfo.ehash =
3023                 alloc_large_system_hash("TCP established",
3024                                         sizeof(struct inet_ehash_bucket),
3025                                         thash_entries,
3026                                         17, /* one slot per 128 KB of memory */
3027                                         0,
3028                                         NULL,
3029                                         &tcp_hashinfo.ehash_mask,
3030                                         0,
3031                                         thash_entries ? 0 : 512 * 1024);
3032         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3033                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3034
3035         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3036                 panic("TCP: failed to alloc ehash_locks");
3037         tcp_hashinfo.bhash =
3038                 alloc_large_system_hash("TCP bind",
3039                                         sizeof(struct inet_bind_hashbucket),
3040                                         tcp_hashinfo.ehash_mask + 1,
3041                                         17, /* one slot per 128 KB of memory */
3042                                         0,
3043                                         &tcp_hashinfo.bhash_size,
3044                                         NULL,
3045                                         0,
3046                                         64 * 1024);
3047         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3048         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3049                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3050                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3051         }
3052
3053
3054         cnt = tcp_hashinfo.ehash_mask + 1;
3055
3056         tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3057         sysctl_tcp_max_orphans = cnt / 2;
3058         sysctl_max_syn_backlog = max(128, cnt / 256);
3059
3060         tcp_init_mem();
3061         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3062         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3063         max_wshare = min(4UL*1024*1024, limit);
3064         max_rshare = min(6UL*1024*1024, limit);
3065
3066         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3067         sysctl_tcp_wmem[1] = 16*1024;
3068         sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3069
3070         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3071         sysctl_tcp_rmem[1] = 87380;
3072         sysctl_tcp_rmem[2] = max(87380, max_rshare);
3073
3074         pr_info("Hash tables configured (established %u bind %u)\n",
3075                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3076
3077         tcp_metrics_init();
3078         BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3079         tcp_tasklet_init();
3080 }