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, <[email protected]>
  10  *              Mark Evans, <[email protected]>
  11  *              Corey Minyard <[email protected]>
  12  *              Florian La Roche, <[email protected]>
  13  *              Charles Hedrick, <[email protected]>
  14  *              Linus Torvalds, <[email protected]>
  15  *              Alan Cox, <[email protected]>
  16  *              Matthew Dillon, <[email protected]>
  17  *              Arnt Gulbrandsen, <[email protected]>
  18  *              Jorge Cwik, <[email protected]>
  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. Memory barrier
 524                                  * pairs with the input side.
 525                                  */
 526                                 smp_mb__after_atomic();
 527                                 if (sk_stream_is_writeable(sk))
 528                                         mask |= POLLOUT | POLLWRNORM;
 529                         }
 530                 } else
 531                         mask |= POLLOUT | POLLWRNORM;
 532
 533                 if (tp->urg_data & TCP_URG_VALID)
 534                         mask |= POLLPRI;
 535         }
 536         /* This barrier is coupled with smp_wmb() in tcp_reset() */
 537         smp_rmb();
 538         if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
 539                 mask |= POLLERR;
 540
 541         return mask;
 542 }
 543 EXPORT_SYMBOL(tcp_poll);
 544
 545 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 546 {
 547         struct tcp_sock *tp = tcp_sk(sk);
 548         int answ;
 549         bool slow;
 550
 551         switch (cmd) {
 552         case SIOCINQ:
 553                 if (sk->sk_state == TCP_LISTEN)
 554                         return -EINVAL;
 555
 556                 slow = lock_sock_fast(sk);
 557                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 558                         answ = 0;
 559                 else if (sock_flag(sk, SOCK_URGINLINE) ||
 560                          !tp->urg_data ||
 561                          before(tp->urg_seq, tp->copied_seq) ||
 562                          !before(tp->urg_seq, tp->rcv_nxt)) {
 563
 564                         answ = tp->rcv_nxt - tp->copied_seq;
 565
 566                         /* Subtract 1, if FIN was received */
 567                         if (answ && sock_flag(sk, SOCK_DONE))
 568                                 answ--;
 569                 } else
 570                         answ = tp->urg_seq - tp->copied_seq;
 571                 unlock_sock_fast(sk, slow);
 572                 break;
 573         case SIOCATMARK:
 574                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 575                 break;
 576         case SIOCOUTQ:
 577                 if (sk->sk_state == TCP_LISTEN)
 578                         return -EINVAL;
 579
 580                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 581                         answ = 0;
 582                 else
 583                         answ = tp->write_seq - tp->snd_una;
 584                 break;
 585         case SIOCOUTQNSD:
 586                 if (sk->sk_state == TCP_LISTEN)
 587                         return -EINVAL;
 588
 589                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 590                         answ = 0;
 591                 else
 592                         answ = tp->write_seq - tp->snd_nxt;
 593                 break;
 594         default:
 595                 return -ENOIOCTLCMD;
 596         }
 597
 598         return put_user(answ, (int __user *)arg);
 599 }
 600 EXPORT_SYMBOL(tcp_ioctl);
 601
 602 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 603 {
 604         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 605         tp->pushed_seq = tp->write_seq;
 606 }
 607
 608 static inline bool forced_push(const struct tcp_sock *tp)
 609 {
 610         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 611 }
 612
 613 static void skb_entail(struct sock *sk, struct sk_buff *skb)
 614 {
 615         struct tcp_sock *tp = tcp_sk(sk);
 616         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 617
 618         skb->csum    = 0;
 619         tcb->seq     = tcb->end_seq = tp->write_seq;
 620         tcb->tcp_flags = TCPHDR_ACK;
 621         tcb->sacked  = 0;
 622         __skb_header_release(skb);
 623         tcp_add_write_queue_tail(sk, skb);
 624         sk->sk_wmem_queued += skb->truesize;
 625         sk_mem_charge(sk, skb->truesize);
 626         if (tp->nonagle & TCP_NAGLE_PUSH)
 627                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 628 }
 629
 630 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
 631 {
 632         if (flags & MSG_OOB)
 633                 tp->snd_up = tp->write_seq;
 634 }
 635
 636 /* If a not yet filled skb is pushed, do not send it if
 637  * we have data packets in Qdisc or NIC queues :
 638  * Because TX completion will happen shortly, it gives a chance
 639  * to coalesce future sendmsg() payload into this skb, without
 640  * need for a timer, and with no latency trade off.
 641  * As packets containing data payload have a bigger truesize
 642  * than pure acks (dataless) packets, the last checks prevent
 643  * autocorking if we only have an ACK in Qdisc/NIC queues,
 644  * or if TX completion was delayed after we processed ACK packet.
 645  */
 646 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
 647                                 int size_goal)
 648 {
 649         return skb->len < size_goal &&
 650                sysctl_tcp_autocorking &&
 651                skb != tcp_write_queue_head(sk) &&
 652                atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
 653 }
 654
 655 static void tcp_push(struct sock *sk, int flags, int mss_now,
 656                      int nonagle, int size_goal)
 657 {
 658         struct tcp_sock *tp = tcp_sk(sk);
 659         struct sk_buff *skb;
 660
 661         if (!tcp_send_head(sk))
 662                 return;
 663
 664         skb = tcp_write_queue_tail(sk);
 665         if (!(flags & MSG_MORE) || forced_push(tp))
 666                 tcp_mark_push(tp, skb);
 667
 668         tcp_mark_urg(tp, flags);
 669
 670         if (tcp_should_autocork(sk, skb, size_goal)) {
 671
 672                 /* avoid atomic op if TSQ_THROTTLED bit is already set */
 673                 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
 674                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
 675                         set_bit(TSQ_THROTTLED, &tp->tsq_flags);
 676                 }
 677                 /* It is possible TX completion already happened
 678                  * before we set TSQ_THROTTLED.
 679                  */
 680                 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
 681                         return;
 682         }
 683
 684         if (flags & MSG_MORE)
 685                 nonagle = TCP_NAGLE_CORK;
 686
 687         __tcp_push_pending_frames(sk, mss_now, nonagle);
 688 }
 689
 690 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 691                                 unsigned int offset, size_t len)
 692 {
 693         struct tcp_splice_state *tss = rd_desc->arg.data;
 694         int ret;
 695
 696         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
 697                               tss->flags);
 698         if (ret > 0)
 699                 rd_desc->count -= ret;
 700         return ret;
 701 }
 702
 703 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 704 {
 705         /* Store TCP splice context information in read_descriptor_t. */
 706         read_descriptor_t rd_desc = {
 707                 .arg.data = tss,
 708                 .count    = tss->len,
 709         };
 710
 711         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 712 }
 713
 714 /**
 715  *  tcp_splice_read - splice data from TCP socket to a pipe
 716  * @sock:       socket to splice from
 717  * @ppos:       position (not valid)
 718  * @pipe:       pipe to splice to
 719  * @len:        number of bytes to splice
 720  * @flags:      splice modifier flags
 721  *
 722  * Description:
 723  *    Will read pages from given socket and fill them into a pipe.
 724  *
 725  **/
 726 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 727                         struct pipe_inode_info *pipe, size_t len,
 728                         unsigned int flags)
 729 {
 730         struct sock *sk = sock->sk;
 731         struct tcp_splice_state tss = {
 732                 .pipe = pipe,
 733                 .len = len,
 734                 .flags = flags,
 735         };
 736         long timeo;
 737         ssize_t spliced;
 738         int ret;
 739
 740         sock_rps_record_flow(sk);
 741         /*
 742          * We can't seek on a socket input
 743          */
 744         if (unlikely(*ppos))
 745                 return -ESPIPE;
 746
 747         ret = spliced = 0;
 748
 749         lock_sock(sk);
 750
 751         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 752         while (tss.len) {
 753                 ret = __tcp_splice_read(sk, &tss);
 754                 if (ret < 0)
 755                         break;
 756                 else if (!ret) {
 757                         if (spliced)
 758                                 break;
 759                         if (sock_flag(sk, SOCK_DONE))
 760                                 break;
 761                         if (sk->sk_err) {
 762                                 ret = sock_error(sk);
 763                                 break;
 764                         }
 765                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 766                                 break;
 767                         if (sk->sk_state == TCP_CLOSE) {
 768                                 /*
 769                                  * This occurs when user tries to read
 770                                  * from never connected socket.
 771                                  */
 772                                 if (!sock_flag(sk, SOCK_DONE))
 773                                         ret = -ENOTCONN;
 774                                 break;
 775                         }
 776                         if (!timeo) {
 777                                 ret = -EAGAIN;
 778                                 break;
 779                         }
 780                         sk_wait_data(sk, &timeo);
 781                         if (signal_pending(current)) {
 782                                 ret = sock_intr_errno(timeo);
 783                                 break;
 784                         }
 785                         continue;
 786                 }
 787                 tss.len -= ret;
 788                 spliced += ret;
 789
 790                 if (!timeo)
 791                         break;
 792                 release_sock(sk);
 793                 lock_sock(sk);
 794
 795                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 796                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 797                     signal_pending(current))
 798                         break;
 799         }
 800
 801         release_sock(sk);
 802
 803         if (spliced)
 804                 return spliced;
 805
 806         return ret;
 807 }
 808 EXPORT_SYMBOL(tcp_splice_read);
 809
 810 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
 811 {
 812         struct sk_buff *skb;
 813
 814         /* The TCP header must be at least 32-bit aligned.  */
 815         size = ALIGN(size, 4);
 816
 817         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 818         if (skb) {
 819                 if (sk_wmem_schedule(sk, skb->truesize)) {
 820                         skb_reserve(skb, sk->sk_prot->max_header);
 821                         /*
 822                          * Make sure that we have exactly size bytes
 823                          * available to the caller, no more, no less.
 824                          */
 825                         skb->reserved_tailroom = skb->end - skb->tail - size;
 826                         return skb;
 827                 }
 828                 __kfree_skb(skb);
 829         } else {
 830                 sk->sk_prot->enter_memory_pressure(sk);
 831                 sk_stream_moderate_sndbuf(sk);
 832         }
 833         return NULL;
 834 }
 835
 836 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 837                                        int large_allowed)
 838 {
 839         struct tcp_sock *tp = tcp_sk(sk);
 840         u32 new_size_goal, size_goal;
 841
 842         if (!large_allowed || !sk_can_gso(sk))
 843                 return mss_now;
 844
 845         /* Note : tcp_tso_autosize() will eventually split this later */
 846         new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
 847         new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
 848
 849         /* We try hard to avoid divides here */
 850         size_goal = tp->gso_segs * mss_now;
 851         if (unlikely(new_size_goal < size_goal ||
 852                      new_size_goal >= size_goal + mss_now)) {
 853                 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
 854                                      sk->sk_gso_max_segs);
 855                 size_goal = tp->gso_segs * mss_now;
 856         }
 857
 858         return max(size_goal, mss_now);
 859 }
 860
 861 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 862 {
 863         int mss_now;
 864
 865         mss_now = tcp_current_mss(sk);
 866         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 867
 868         return mss_now;
 869 }
 870
 871 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
 872                                 size_t size, int flags)
 873 {
 874         struct tcp_sock *tp = tcp_sk(sk);
 875         int mss_now, size_goal;
 876         int err;
 877         ssize_t copied;
 878         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 879
 880         /* Wait for a connection to finish. One exception is TCP Fast Open
 881          * (passive side) where data is allowed to be sent before a connection
 882          * is fully established.
 883          */
 884         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
 885             !tcp_passive_fastopen(sk)) {
 886                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 887                         goto out_err;
 888         }
 889
 890         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 891
 892         mss_now = tcp_send_mss(sk, &size_goal, flags);
 893         copied = 0;
 894
 895         err = -EPIPE;
 896         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 897                 goto out_err;
 898
 899         while (size > 0) {
 900                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 901                 int copy, i;
 902                 bool can_coalesce;
 903
 904                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
 905 new_segment:
 906                         if (!sk_stream_memory_free(sk))
 907                                 goto wait_for_sndbuf;
 908
 909                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
 910                         if (!skb)
 911                                 goto wait_for_memory;
 912
 913                         skb_entail(sk, skb);
 914                         copy = size_goal;
 915                 }
 916
 917                 if (copy > size)
 918                         copy = size;
 919
 920                 i = skb_shinfo(skb)->nr_frags;
 921                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 922                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
 923                         tcp_mark_push(tp, skb);
 924                         goto new_segment;
 925                 }
 926                 if (!sk_wmem_schedule(sk, copy))
 927                         goto wait_for_memory;
 928
 929                 if (can_coalesce) {
 930                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
 931                 } else {
 932                         get_page(page);
 933                         skb_fill_page_desc(skb, i, page, offset, copy);
 934                 }
 935                 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
 936
 937                 skb->len += copy;
 938                 skb->data_len += copy;
 939                 skb->truesize += copy;
 940                 sk->sk_wmem_queued += copy;
 941                 sk_mem_charge(sk, copy);
 942                 skb->ip_summed = CHECKSUM_PARTIAL;
 943                 tp->write_seq += copy;
 944                 TCP_SKB_CB(skb)->end_seq += copy;
 945                 tcp_skb_pcount_set(skb, 0);
 946
 947                 if (!copied)
 948                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
 949
 950                 copied += copy;
 951                 offset += copy;
 952                 if (!(size -= copy)) {
 953                         tcp_tx_timestamp(sk, skb);
 954                         goto out;
 955                 }
 956
 957                 if (skb->len < size_goal || (flags & MSG_OOB))
 958                         continue;
 959
 960                 if (forced_push(tp)) {
 961                         tcp_mark_push(tp, skb);
 962                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
 963                 } else if (skb == tcp_send_head(sk))
 964                         tcp_push_one(sk, mss_now);
 965                 continue;
 966
 967 wait_for_sndbuf:
 968                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 969 wait_for_memory:
 970                 tcp_push(sk, flags & ~MSG_MORE, mss_now,
 971                          TCP_NAGLE_PUSH, size_goal);
 972
 973                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 974                         goto do_error;
 975
 976                 mss_now = tcp_send_mss(sk, &size_goal, flags);
 977         }
 978
 979 out:
 980         if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
 981                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
 982         return copied;
 983
 984 do_error:
 985         if (copied)
 986                 goto out;
 987 out_err:
 988         return sk_stream_error(sk, flags, err);
 989 }
 990
 991 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
 992                  size_t size, int flags)
 993 {
 994         ssize_t res;
 995
 996         if (!(sk->sk_route_caps & NETIF_F_SG) ||
 997             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
 998                 return sock_no_sendpage(sk->sk_socket, page, offset, size,
 999                                         flags);
1000
1001         lock_sock(sk);
1002         res = do_tcp_sendpages(sk, page, offset, size, flags);
1003         release_sock(sk);
1004         return res;
1005 }
1006 EXPORT_SYMBOL(tcp_sendpage);
1007
1008 static inline int select_size(const struct sock *sk, bool sg)
1009 {
1010         const struct tcp_sock *tp = tcp_sk(sk);
1011         int tmp = tp->mss_cache;
1012
1013         if (sg) {
1014                 if (sk_can_gso(sk)) {
1015                         /* Small frames wont use a full page:
1016                          * Payload will immediately follow tcp header.
1017                          */
1018                         tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1019                 } else {
1020                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1021
1022                         if (tmp >= pgbreak &&
1023                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1024                                 tmp = pgbreak;
1025                 }
1026         }
1027
1028         return tmp;
1029 }
1030
1031 void tcp_free_fastopen_req(struct tcp_sock *tp)
1032 {
1033         if (tp->fastopen_req) {
1034                 kfree(tp->fastopen_req);
1035                 tp->fastopen_req = NULL;
1036         }
1037 }
1038
1039 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1040                                 int *copied, size_t size)
1041 {
1042         struct tcp_sock *tp = tcp_sk(sk);
1043         int err, flags;
1044
1045         if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1046                 return -EOPNOTSUPP;
1047         if (tp->fastopen_req)
1048                 return -EALREADY; /* Another Fast Open is in progress */
1049
1050         tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1051                                    sk->sk_allocation);
1052         if (unlikely(!tp->fastopen_req))
1053                 return -ENOBUFS;
1054         tp->fastopen_req->data = msg;
1055         tp->fastopen_req->size = size;
1056
1057         flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1058         err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1059                                     msg->msg_namelen, flags);
1060         *copied = tp->fastopen_req->copied;
1061         tcp_free_fastopen_req(tp);
1062         return err;
1063 }
1064
1065 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1066 {
1067         struct tcp_sock *tp = tcp_sk(sk);
1068         struct sk_buff *skb;
1069         int flags, err, copied = 0;
1070         int mss_now = 0, size_goal, copied_syn = 0;
1071         bool sg;
1072         long timeo;
1073
1074         lock_sock(sk);
1075
1076         flags = msg->msg_flags;
1077         if (flags & MSG_FASTOPEN) {
1078                 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1079                 if (err == -EINPROGRESS && copied_syn > 0)
1080                         goto out;
1081                 else if (err)
1082                         goto out_err;
1083         }
1084
1085         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1086
1087         /* Wait for a connection to finish. One exception is TCP Fast Open
1088          * (passive side) where data is allowed to be sent before a connection
1089          * is fully established.
1090          */
1091         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1092             !tcp_passive_fastopen(sk)) {
1093                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1094                         goto do_error;
1095         }
1096
1097         if (unlikely(tp->repair)) {
1098                 if (tp->repair_queue == TCP_RECV_QUEUE) {
1099                         copied = tcp_send_rcvq(sk, msg, size);
1100                         goto out_nopush;
1101                 }
1102
1103                 err = -EINVAL;
1104                 if (tp->repair_queue == TCP_NO_QUEUE)
1105                         goto out_err;
1106
1107                 /* 'common' sending to sendq */
1108         }
1109
1110         /* This should be in poll */
1111         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1112
1113         mss_now = tcp_send_mss(sk, &size_goal, flags);
1114
1115         /* Ok commence sending. */
1116         copied = 0;
1117
1118         err = -EPIPE;
1119         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1120                 goto out_err;
1121
1122         sg = !!(sk->sk_route_caps & NETIF_F_SG);
1123
1124         while (msg_data_left(msg)) {
1125                 int copy = 0;
1126                 int max = size_goal;
1127
1128                 skb = tcp_write_queue_tail(sk);
1129                 if (tcp_send_head(sk)) {
1130                         if (skb->ip_summed == CHECKSUM_NONE)
1131                                 max = mss_now;
1132                         copy = max - skb->len;
1133                 }
1134
1135                 if (copy <= 0) {
1136 new_segment:
1137                         /* Allocate new segment. If the interface is SG,
1138                          * allocate skb fitting to single page.
1139                          */
1140                         if (!sk_stream_memory_free(sk))
1141                                 goto wait_for_sndbuf;
1142
1143                         skb = sk_stream_alloc_skb(sk,
1144                                                   select_size(sk, sg),
1145                                                   sk->sk_allocation);
1146                         if (!skb)
1147                                 goto wait_for_memory;
1148
1149                         /*
1150                          * Check whether we can use HW checksum.
1151                          */
1152                         if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1153                                 skb->ip_summed = CHECKSUM_PARTIAL;
1154
1155                         skb_entail(sk, skb);
1156                         copy = size_goal;
1157                         max = size_goal;
1158
1159                         /* All packets are restored as if they have
1160                          * already been sent. skb_mstamp isn't set to
1161                          * avoid wrong rtt estimation.
1162                          */
1163                         if (tp->repair)
1164                                 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1165                 }
1166
1167                 /* Try to append data to the end of skb. */
1168                 if (copy > msg_data_left(msg))
1169                         copy = msg_data_left(msg);
1170
1171                 /* Where to copy to? */
1172                 if (skb_availroom(skb) > 0) {
1173                         /* We have some space in skb head. Superb! */
1174                         copy = min_t(int, copy, skb_availroom(skb));
1175                         err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1176                         if (err)
1177                                 goto do_fault;
1178                 } else {
1179                         bool merge = true;
1180                         int i = skb_shinfo(skb)->nr_frags;
1181                         struct page_frag *pfrag = sk_page_frag(sk);
1182
1183                         if (!sk_page_frag_refill(sk, pfrag))
1184                                 goto wait_for_memory;
1185
1186                         if (!skb_can_coalesce(skb, i, pfrag->page,
1187                                               pfrag->offset)) {
1188                                 if (i == MAX_SKB_FRAGS || !sg) {
1189                                         tcp_mark_push(tp, skb);
1190                                         goto new_segment;
1191                                 }
1192                                 merge = false;
1193                         }
1194
1195                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1196
1197                         if (!sk_wmem_schedule(sk, copy))
1198                                 goto wait_for_memory;
1199
1200                         err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1201                                                        pfrag->page,
1202                                                        pfrag->offset,
1203                                                        copy);
1204                         if (err)
1205                                 goto do_error;
1206
1207                         /* Update the skb. */
1208                         if (merge) {
1209                                 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1210                         } else {
1211                                 skb_fill_page_desc(skb, i, pfrag->page,
1212                                                    pfrag->offset, copy);
1213                                 get_page(pfrag->page);
1214                         }
1215                         pfrag->offset += copy;
1216                 }
1217
1218                 if (!copied)
1219                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1220
1221                 tp->write_seq += copy;
1222                 TCP_SKB_CB(skb)->end_seq += copy;
1223                 tcp_skb_pcount_set(skb, 0);
1224
1225                 copied += copy;
1226                 if (!msg_data_left(msg)) {
1227                         tcp_tx_timestamp(sk, skb);
1228                         goto out;
1229                 }
1230
1231                 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1232                         continue;
1233
1234                 if (forced_push(tp)) {
1235                         tcp_mark_push(tp, skb);
1236                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1237                 } else if (skb == tcp_send_head(sk))
1238                         tcp_push_one(sk, mss_now);
1239                 continue;
1240
1241 wait_for_sndbuf:
1242                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1243 wait_for_memory:
1244                 if (copied)
1245                         tcp_push(sk, flags & ~MSG_MORE, mss_now,
1246                                  TCP_NAGLE_PUSH, size_goal);
1247
1248                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1249                         goto do_error;
1250
1251                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1252         }
1253
1254 out:
1255         if (copied)
1256                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1257 out_nopush:
1258         release_sock(sk);
1259         return copied + copied_syn;
1260
1261 do_fault:
1262         if (!skb->len) {
1263                 tcp_unlink_write_queue(skb, sk);
1264                 /* It is the one place in all of TCP, except connection
1265                  * reset, where we can be unlinking the send_head.
1266                  */
1267                 tcp_check_send_head(sk, skb);
1268                 sk_wmem_free_skb(sk, skb);
1269         }
1270
1271 do_error:
1272         if (copied + copied_syn)
1273                 goto out;
1274 out_err:
1275         err = sk_stream_error(sk, flags, err);
1276         release_sock(sk);
1277         return err;
1278 }
1279 EXPORT_SYMBOL(tcp_sendmsg);
1280
1281 /*
1282  *      Handle reading urgent data. BSD has very simple semantics for
1283  *      this, no blocking and very strange errors 8)
1284  */
1285
1286 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1287 {
1288         struct tcp_sock *tp = tcp_sk(sk);
1289
1290         /* No URG data to read. */
1291         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1292             tp->urg_data == TCP_URG_READ)
1293                 return -EINVAL; /* Yes this is right ! */
1294
1295         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1296                 return -ENOTCONN;
1297
1298         if (tp->urg_data & TCP_URG_VALID) {
1299                 int err = 0;
1300                 char c = tp->urg_data;
1301
1302                 if (!(flags & MSG_PEEK))
1303                         tp->urg_data = TCP_URG_READ;
1304
1305                 /* Read urgent data. */
1306                 msg->msg_flags |= MSG_OOB;
1307
1308                 if (len > 0) {
1309                         if (!(flags & MSG_TRUNC))
1310                                 err = memcpy_to_msg(msg, &c, 1);
1311                         len = 1;
1312                 } else
1313                         msg->msg_flags |= MSG_TRUNC;
1314
1315                 return err ? -EFAULT : len;
1316         }
1317
1318         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1319                 return 0;
1320
1321         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1322          * the available implementations agree in this case:
1323          * this call should never block, independent of the
1324          * blocking state of the socket.
1325          * Mike <[email protected]>
1326          */
1327         return -EAGAIN;
1328 }
1329
1330 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1331 {
1332         struct sk_buff *skb;
1333         int copied = 0, err = 0;
1334
1335         /* XXX -- need to support SO_PEEK_OFF */
1336
1337         skb_queue_walk(&sk->sk_write_queue, skb) {
1338                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1339                 if (err)
1340                         break;
1341
1342                 copied += skb->len;
1343         }
1344
1345         return err ?: copied;
1346 }
1347
1348 /* Clean up the receive buffer for full frames taken by the user,
1349  * then send an ACK if necessary.  COPIED is the number of bytes
1350  * tcp_recvmsg has given to the user so far, it speeds up the
1351  * calculation of whether or not we must ACK for the sake of
1352  * a window update.
1353  */
1354 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1355 {
1356         struct tcp_sock *tp = tcp_sk(sk);
1357         bool time_to_ack = false;
1358
1359         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1360
1361         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1362              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1363              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1364
1365         if (inet_csk_ack_scheduled(sk)) {
1366                 const struct inet_connection_sock *icsk = inet_csk(sk);
1367                    /* Delayed ACKs frequently hit locked sockets during bulk
1368                     * receive. */
1369                 if (icsk->icsk_ack.blocked ||
1370                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1371                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1372                     /*
1373                      * If this read emptied read buffer, we send ACK, if
1374                      * connection is not bidirectional, user drained
1375                      * receive buffer and there was a small segment
1376                      * in queue.
1377                      */
1378                     (copied > 0 &&
1379                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1380                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1381                        !icsk->icsk_ack.pingpong)) &&
1382                       !atomic_read(&sk->sk_rmem_alloc)))
1383                         time_to_ack = true;
1384         }
1385
1386         /* We send an ACK if we can now advertise a non-zero window
1387          * which has been raised "significantly".
1388          *
1389          * Even if window raised up to infinity, do not send window open ACK
1390          * in states, where we will not receive more. It is useless.
1391          */
1392         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1393                 __u32 rcv_window_now = tcp_receive_window(tp);
1394
1395                 /* Optimize, __tcp_select_window() is not cheap. */
1396                 if (2*rcv_window_now <= tp->window_clamp) {
1397                         __u32 new_window = __tcp_select_window(sk);
1398
1399                         /* Send ACK now, if this read freed lots of space
1400                          * in our buffer. Certainly, new_window is new window.
1401                          * We can advertise it now, if it is not less than current one.
1402                          * "Lots" means "at least twice" here.
1403                          */
1404                         if (new_window && new_window >= 2 * rcv_window_now)
1405                                 time_to_ack = true;
1406                 }
1407         }
1408         if (time_to_ack)
1409                 tcp_send_ack(sk);
1410 }
1411
1412 static void tcp_prequeue_process(struct sock *sk)
1413 {
1414         struct sk_buff *skb;
1415         struct tcp_sock *tp = tcp_sk(sk);
1416
1417         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1418
1419         /* RX process wants to run with disabled BHs, though it is not
1420          * necessary */
1421         local_bh_disable();
1422         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1423                 sk_backlog_rcv(sk, skb);
1424         local_bh_enable();
1425
1426         /* Clear memory counter. */
1427         tp->ucopy.memory = 0;
1428 }
1429
1430 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1431 {
1432         struct sk_buff *skb;
1433         u32 offset;
1434
1435         while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1436                 offset = seq - TCP_SKB_CB(skb)->seq;
1437                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1438                         offset--;
1439                 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1440                         *off = offset;
1441                         return skb;
1442                 }
1443                 /* This looks weird, but this can happen if TCP collapsing
1444                  * splitted a fat GRO packet, while we released socket lock
1445                  * in skb_splice_bits()
1446                  */
1447                 sk_eat_skb(sk, skb);
1448         }
1449         return NULL;
1450 }
1451
1452 /*
1453  * This routine provides an alternative to tcp_recvmsg() for routines
1454  * that would like to handle copying from skbuffs directly in 'sendfile'
1455  * fashion.
1456  * Note:
1457  *      - It is assumed that the socket was locked by the caller.
1458  *      - The routine does not block.
1459  *      - At present, there is no support for reading OOB data
1460  *        or for 'peeking' the socket using this routine
1461  *        (although both would be easy to implement).
1462  */
1463 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1464                   sk_read_actor_t recv_actor)
1465 {
1466         struct sk_buff *skb;
1467         struct tcp_sock *tp = tcp_sk(sk);
1468         u32 seq = tp->copied_seq;
1469         u32 offset;
1470         int copied = 0;
1471
1472         if (sk->sk_state == TCP_LISTEN)
1473                 return -ENOTCONN;
1474         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1475                 if (offset < skb->len) {
1476                         int used;
1477                         size_t len;
1478
1479                         len = skb->len - offset;
1480                         /* Stop reading if we hit a patch of urgent data */
1481                         if (tp->urg_data) {
1482                                 u32 urg_offset = tp->urg_seq - seq;
1483                                 if (urg_offset < len)
1484                                         len = urg_offset;
1485                                 if (!len)
1486                                         break;
1487                         }
1488                         used = recv_actor(desc, skb, offset, len);
1489                         if (used <= 0) {
1490                                 if (!copied)
1491                                         copied = used;
1492                                 break;
1493                         } else if (used <= len) {
1494                                 seq += used;
1495                                 copied += used;
1496                                 offset += used;
1497                         }
1498                         /* If recv_actor drops the lock (e.g. TCP splice
1499                          * receive) the skb pointer might be invalid when
1500                          * getting here: tcp_collapse might have deleted it
1501                          * while aggregating skbs from the socket queue.
1502                          */
1503                         skb = tcp_recv_skb(sk, seq - 1, &offset);
1504                         if (!skb)
1505                                 break;
1506                         /* TCP coalescing might have appended data to the skb.
1507                          * Try to splice more frags
1508                          */
1509                         if (offset + 1 != skb->len)
1510                                 continue;
1511                 }
1512                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1513                         sk_eat_skb(sk, skb);
1514                         ++seq;
1515                         break;
1516                 }
1517                 sk_eat_skb(sk, skb);
1518                 if (!desc->count)
1519                         break;
1520                 tp->copied_seq = seq;
1521         }
1522         tp->copied_seq = seq;
1523
1524         tcp_rcv_space_adjust(sk);
1525
1526         /* Clean up data we have read: This will do ACK frames. */
1527         if (copied > 0) {
1528                 tcp_recv_skb(sk, seq, &offset);
1529                 tcp_cleanup_rbuf(sk, copied);
1530         }
1531         return copied;
1532 }
1533 EXPORT_SYMBOL(tcp_read_sock);
1534
1535 /*
1536  *      This routine copies from a sock struct into the user buffer.
1537  *
1538  *      Technical note: in 2.3 we work on _locked_ socket, so that
1539  *      tricks with *seq access order and skb->users are not required.
1540  *      Probably, code can be easily improved even more.
1541  */
1542
1543 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1544                 int flags, int *addr_len)
1545 {
1546         struct tcp_sock *tp = tcp_sk(sk);
1547         int copied = 0;
1548         u32 peek_seq;
1549         u32 *seq;
1550         unsigned long used;
1551         int err;
1552         int target;             /* Read at least this many bytes */
1553         long timeo;
1554         struct task_struct *user_recv = NULL;
1555         struct sk_buff *skb;
1556         u32 urg_hole = 0;
1557
1558         if (unlikely(flags & MSG_ERRQUEUE))
1559                 return inet_recv_error(sk, msg, len, addr_len);
1560
1561         if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1562             (sk->sk_state == TCP_ESTABLISHED))
1563                 sk_busy_loop(sk, nonblock);
1564
1565         lock_sock(sk);
1566
1567         err = -ENOTCONN;
1568         if (sk->sk_state == TCP_LISTEN)
1569                 goto out;
1570
1571         timeo = sock_rcvtimeo(sk, nonblock);
1572
1573         /* Urgent data needs to be handled specially. */
1574         if (flags & MSG_OOB)
1575                 goto recv_urg;
1576
1577         if (unlikely(tp->repair)) {
1578                 err = -EPERM;
1579                 if (!(flags & MSG_PEEK))
1580                         goto out;
1581
1582                 if (tp->repair_queue == TCP_SEND_QUEUE)
1583                         goto recv_sndq;
1584
1585                 err = -EINVAL;
1586                 if (tp->repair_queue == TCP_NO_QUEUE)
1587                         goto out;
1588
1589                 /* 'common' recv queue MSG_PEEK-ing */
1590         }
1591
1592         seq = &tp->copied_seq;
1593         if (flags & MSG_PEEK) {
1594                 peek_seq = tp->copied_seq;
1595                 seq = &peek_seq;
1596         }
1597
1598         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1599
1600         do {
1601                 u32 offset;
1602
1603                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1604                 if (tp->urg_data && tp->urg_seq == *seq) {
1605                         if (copied)
1606                                 break;
1607                         if (signal_pending(current)) {
1608                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1609                                 break;
1610                         }
1611                 }
1612
1613                 /* Next get a buffer. */
1614
1615                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1616                         /* Now that we have two receive queues this
1617                          * shouldn't happen.
1618                          */
1619                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1620                                  "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1621                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1622                                  flags))
1623                                 break;
1624
1625                         offset = *seq - TCP_SKB_CB(skb)->seq;
1626                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1627                                 offset--;
1628                         if (offset < skb->len)
1629                                 goto found_ok_skb;
1630                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1631                                 goto found_fin_ok;
1632                         WARN(!(flags & MSG_PEEK),
1633                              "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1634                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1635                 }
1636
1637                 /* Well, if we have backlog, try to process it now yet. */
1638
1639                 if (copied >= target && !sk->sk_backlog.tail)
1640                         break;
1641
1642                 if (copied) {
1643                         if (sk->sk_err ||
1644                             sk->sk_state == TCP_CLOSE ||
1645                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1646                             !timeo ||
1647                             signal_pending(current))
1648                                 break;
1649                 } else {
1650                         if (sock_flag(sk, SOCK_DONE))
1651                                 break;
1652
1653                         if (sk->sk_err) {
1654                                 copied = sock_error(sk);
1655                                 break;
1656                         }
1657
1658                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1659                                 break;
1660
1661                         if (sk->sk_state == TCP_CLOSE) {
1662                                 if (!sock_flag(sk, SOCK_DONE)) {
1663                                         /* This occurs when user tries to read
1664                                          * from never connected socket.
1665                                          */
1666                                         copied = -ENOTCONN;
1667                                         break;
1668                                 }
1669                                 break;
1670                         }
1671
1672                         if (!timeo) {
1673                                 copied = -EAGAIN;
1674                                 break;
1675                         }
1676
1677                         if (signal_pending(current)) {
1678                                 copied = sock_intr_errno(timeo);
1679                                 break;
1680                         }
1681                 }
1682
1683                 tcp_cleanup_rbuf(sk, copied);
1684
1685                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1686                         /* Install new reader */
1687                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1688                                 user_recv = current;
1689                                 tp->ucopy.task = user_recv;
1690                                 tp->ucopy.msg = msg;
1691                         }
1692
1693                         tp->ucopy.len = len;
1694
1695                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1696                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1697
1698                         /* Ugly... If prequeue is not empty, we have to
1699                          * process it before releasing socket, otherwise
1700                          * order will be broken at second iteration.
1701                          * More elegant solution is required!!!
1702                          *
1703                          * Look: we have the following (pseudo)queues:
1704                          *
1705                          * 1. packets in flight
1706                          * 2. backlog
1707                          * 3. prequeue
1708                          * 4. receive_queue
1709                          *
1710                          * Each queue can be processed only if the next ones
1711                          * are empty. At this point we have empty receive_queue.
1712                          * But prequeue _can_ be not empty after 2nd iteration,
1713                          * when we jumped to start of loop because backlog
1714                          * processing added something to receive_queue.
1715                          * We cannot release_sock(), because backlog contains
1716                          * packets arrived _after_ prequeued ones.
1717                          *
1718                          * Shortly, algorithm is clear --- to process all
1719                          * the queues in order. We could make it more directly,
1720                          * requeueing packets from backlog to prequeue, if
1721                          * is not empty. It is more elegant, but eats cycles,
1722                          * unfortunately.
1723                          */
1724                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1725                                 goto do_prequeue;
1726
1727                         /* __ Set realtime policy in scheduler __ */
1728                 }
1729
1730                 if (copied >= target) {
1731                         /* Do not sleep, just process backlog. */
1732                         release_sock(sk);
1733                         lock_sock(sk);
1734                 } else
1735                         sk_wait_data(sk, &timeo);
1736
1737                 if (user_recv) {
1738                         int chunk;
1739
1740                         /* __ Restore normal policy in scheduler __ */
1741
1742                         if ((chunk = len - tp->ucopy.len) != 0) {
1743                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1744                                 len -= chunk;
1745                                 copied += chunk;
1746                         }
1747
1748                         if (tp->rcv_nxt == tp->copied_seq &&
1749                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1750 do_prequeue:
1751                                 tcp_prequeue_process(sk);
1752
1753                                 if ((chunk = len - tp->ucopy.len) != 0) {
1754                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1755                                         len -= chunk;
1756                                         copied += chunk;
1757                                 }
1758                         }
1759                 }
1760                 if ((flags & MSG_PEEK) &&
1761                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1762                         net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1763                                             current->comm,
1764                                             task_pid_nr(current));
1765                         peek_seq = tp->copied_seq;
1766                 }
1767                 continue;
1768
1769         found_ok_skb:
1770                 /* Ok so how much can we use? */
1771                 used = skb->len - offset;
1772                 if (len < used)
1773                         used = len;
1774
1775                 /* Do we have urgent data here? */
1776                 if (tp->urg_data) {
1777                         u32 urg_offset = tp->urg_seq - *seq;
1778                         if (urg_offset < used) {
1779                                 if (!urg_offset) {
1780                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1781                                                 ++*seq;
1782                                                 urg_hole++;
1783                                                 offset++;
1784                                                 used--;
1785                                                 if (!used)
1786                                                         goto skip_copy;
1787                                         }
1788                                 } else
1789                                         used = urg_offset;
1790                         }
1791                 }
1792
1793                 if (!(flags & MSG_TRUNC)) {
1794                         err = skb_copy_datagram_msg(skb, offset, msg, used);
1795                         if (err) {
1796                                 /* Exception. Bailout! */
1797                                 if (!copied)
1798                                         copied = -EFAULT;
1799                                 break;
1800                         }
1801                 }
1802
1803                 *seq += used;
1804                 copied += used;
1805                 len -= used;
1806
1807                 tcp_rcv_space_adjust(sk);
1808
1809 skip_copy:
1810                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1811                         tp->urg_data = 0;
1812                         tcp_fast_path_check(sk);
1813                 }
1814                 if (used + offset < skb->len)
1815                         continue;
1816
1817                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1818                         goto found_fin_ok;
1819                 if (!(flags & MSG_PEEK))
1820                         sk_eat_skb(sk, skb);
1821                 continue;
1822
1823         found_fin_ok:
1824                 /* Process the FIN. */
1825                 ++*seq;
1826                 if (!(flags & MSG_PEEK))
1827                         sk_eat_skb(sk, skb);
1828                 break;
1829         } while (len > 0);
1830
1831         if (user_recv) {
1832                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1833                         int chunk;
1834
1835                         tp->ucopy.len = copied > 0 ? len : 0;
1836
1837                         tcp_prequeue_process(sk);
1838
1839                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1840                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1841                                 len -= chunk;
1842                                 copied += chunk;
1843                         }
1844                 }
1845
1846                 tp->ucopy.task = NULL;
1847                 tp->ucopy.len = 0;
1848         }
1849
1850         /* According to UNIX98, msg_name/msg_namelen are ignored
1851          * on connected socket. I was just happy when found this 8) --ANK
1852          */
1853
1854         /* Clean up data we have read: This will do ACK frames. */
1855         tcp_cleanup_rbuf(sk, copied);
1856
1857         release_sock(sk);
1858         return copied;
1859
1860 out:
1861         release_sock(sk);
1862         return err;
1863
1864 recv_urg:
1865         err = tcp_recv_urg(sk, msg, len, flags);
1866         goto out;
1867
1868 recv_sndq:
1869         err = tcp_peek_sndq(sk, msg, len);
1870         goto out;
1871 }
1872 EXPORT_SYMBOL(tcp_recvmsg);
1873
1874 void tcp_set_state(struct sock *sk, int state)
1875 {
1876         int oldstate = sk->sk_state;
1877
1878         switch (state) {
1879         case TCP_ESTABLISHED:
1880                 if (oldstate != TCP_ESTABLISHED)
1881                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1882                 break;
1883
1884         case TCP_CLOSE:
1885                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1886                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1887
1888                 sk->sk_prot->unhash(sk);
1889                 if (inet_csk(sk)->icsk_bind_hash &&
1890                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1891                         inet_put_port(sk);
1892                 /* fall through */
1893         default:
1894                 if (oldstate == TCP_ESTABLISHED)
1895                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1896         }
1897
1898         /* Change state AFTER socket is unhashed to avoid closed
1899          * socket sitting in hash tables.
1900          */
1901         sk->sk_state = state;
1902
1903 #ifdef STATE_TRACE
1904         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1905 #endif
1906 }
1907 EXPORT_SYMBOL_GPL(tcp_set_state);
1908
1909 /*
1910  *      State processing on a close. This implements the state shift for
1911  *      sending our FIN frame. Note that we only send a FIN for some
1912  *      states. A shutdown() may have already sent the FIN, or we may be
1913  *      closed.
1914  */
1915
1916 static const unsigned char new_state[16] = {
1917   /* current state:        new state:      action:      */
1918   [0 /* (Invalid) */]   = TCP_CLOSE,
1919   [TCP_ESTABLISHED]     = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1920   [TCP_SYN_SENT]        = TCP_CLOSE,
1921   [TCP_SYN_RECV]        = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1922   [TCP_FIN_WAIT1]       = TCP_FIN_WAIT1,
1923   [TCP_FIN_WAIT2]       = TCP_FIN_WAIT2,
1924   [TCP_TIME_WAIT]       = TCP_CLOSE,
1925   [TCP_CLOSE]           = TCP_CLOSE,
1926   [TCP_CLOSE_WAIT]      = TCP_LAST_ACK  | TCP_ACTION_FIN,
1927   [TCP_LAST_ACK]        = TCP_LAST_ACK,
1928   [TCP_LISTEN]          = TCP_CLOSE,
1929   [TCP_CLOSING]         = TCP_CLOSING,
1930   [TCP_NEW_SYN_RECV]    = TCP_CLOSE,    /* should not happen ! */
1931 };
1932
1933 static int tcp_close_state(struct sock *sk)
1934 {
1935         int next = (int)new_state[sk->sk_state];
1936         int ns = next & TCP_STATE_MASK;
1937
1938         tcp_set_state(sk, ns);
1939
1940         return next & TCP_ACTION_FIN;
1941 }
1942
1943 /*
1944  *      Shutdown the sending side of a connection. Much like close except
1945  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1946  */
1947
1948 void tcp_shutdown(struct sock *sk, int how)
1949 {
1950         /*      We need to grab some memory, and put together a FIN,
1951          *      and then put it into the queue to be sent.
1952          *              Tim MacKenzie([email protected]) 4 Dec '92.
1953          */
1954         if (!(how & SEND_SHUTDOWN))
1955                 return;
1956
1957         /* If we've already sent a FIN, or it's a closed state, skip this. */
1958         if ((1 << sk->sk_state) &
1959             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1960              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1961                 /* Clear out any half completed packets.  FIN if needed. */
1962                 if (tcp_close_state(sk))
1963                         tcp_send_fin(sk);
1964         }
1965 }
1966 EXPORT_SYMBOL(tcp_shutdown);
1967
1968 bool tcp_check_oom(struct sock *sk, int shift)
1969 {
1970         bool too_many_orphans, out_of_socket_memory;
1971
1972         too_many_orphans = tcp_too_many_orphans(sk, shift);
1973         out_of_socket_memory = tcp_out_of_memory(sk);
1974
1975         if (too_many_orphans)
1976                 net_info_ratelimited("too many orphaned sockets\n");
1977         if (out_of_socket_memory)
1978                 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
1979         return too_many_orphans || out_of_socket_memory;
1980 }
1981
1982 void tcp_close(struct sock *sk, long timeout)
1983 {
1984         struct sk_buff *skb;
1985         int data_was_unread = 0;
1986         int state;
1987
1988         lock_sock(sk);
1989         sk->sk_shutdown = SHUTDOWN_MASK;
1990
1991         if (sk->sk_state == TCP_LISTEN) {
1992                 tcp_set_state(sk, TCP_CLOSE);
1993
1994                 /* Special case. */
1995                 inet_csk_listen_stop(sk);
1996
1997                 goto adjudge_to_death;
1998         }
1999
2000         /*  We need to flush the recv. buffs.  We do this only on the
2001          *  descriptor close, not protocol-sourced closes, because the
2002          *  reader process may not have drained the data yet!
2003          */
2004         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2005                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2006
2007                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2008                         len--;
2009                 data_was_unread += len;
2010                 __kfree_skb(skb);
2011         }
2012
2013         sk_mem_reclaim(sk);
2014
2015         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2016         if (sk->sk_state == TCP_CLOSE)
2017                 goto adjudge_to_death;
2018
2019         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2020          * data was lost. To witness the awful effects of the old behavior of
2021          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2022          * GET in an FTP client, suspend the process, wait for the client to
2023          * advertise a zero window, then kill -9 the FTP client, wheee...
2024          * Note: timeout is always zero in such a case.
2025          */
2026         if (unlikely(tcp_sk(sk)->repair)) {
2027                 sk->sk_prot->disconnect(sk, 0);
2028         } else if (data_was_unread) {
2029                 /* Unread data was tossed, zap the connection. */
2030                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2031                 tcp_set_state(sk, TCP_CLOSE);
2032                 tcp_send_active_reset(sk, sk->sk_allocation);
2033         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2034                 /* Check zero linger _after_ checking for unread data. */
2035                 sk->sk_prot->disconnect(sk, 0);
2036                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2037         } else if (tcp_close_state(sk)) {
2038                 /* We FIN if the application ate all the data before
2039                  * zapping the connection.
2040                  */
2041
2042                 /* RED-PEN. Formally speaking, we have broken TCP state
2043                  * machine. State transitions:
2044                  *
2045                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2046                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2047                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2048                  *
2049                  * are legal only when FIN has been sent (i.e. in window),
2050                  * rather than queued out of window. Purists blame.
2051                  *
2052                  * F.e. "RFC state" is ESTABLISHED,
2053                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2054                  *
2055                  * The visible declinations are that sometimes
2056                  * we enter time-wait state, when it is not required really
2057                  * (harmless), do not send active resets, when they are
2058                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2059                  * they look as CLOSING or LAST_ACK for Linux)
2060                  * Probably, I missed some more holelets.
2061                  *                                              --ANK
2062                  * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2063                  * in a single packet! (May consider it later but will
2064                  * probably need API support or TCP_CORK SYN-ACK until
2065                  * data is written and socket is closed.)
2066                  */
2067                 tcp_send_fin(sk);
2068         }
2069
2070         sk_stream_wait_close(sk, timeout);
2071
2072 adjudge_to_death:
2073         state = sk->sk_state;
2074         sock_hold(sk);
2075         sock_orphan(sk);
2076
2077         /* It is the last release_sock in its life. It will remove backlog. */
2078         release_sock(sk);
2079
2080
2081         /* Now socket is owned by kernel and we acquire BH lock
2082            to finish close. No need to check for user refs.
2083          */
2084         local_bh_disable();
2085         bh_lock_sock(sk);
2086         WARN_ON(sock_owned_by_user(sk));
2087
2088         percpu_counter_inc(sk->sk_prot->orphan_count);
2089
2090         /* Have we already been destroyed by a softirq or backlog? */
2091         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2092                 goto out;
2093
2094         /*      This is a (useful) BSD violating of the RFC. There is a
2095          *      problem with TCP as specified in that the other end could
2096          *      keep a socket open forever with no application left this end.
2097          *      We use a 1 minute timeout (about the same as BSD) then kill
2098          *      our end. If they send after that then tough - BUT: long enough
2099          *      that we won't make the old 4*rto = almost no time - whoops
2100          *      reset mistake.
2101          *
2102          *      Nope, it was not mistake. It is really desired behaviour
2103          *      f.e. on http servers, when such sockets are useless, but
2104          *      consume significant resources. Let's do it with special
2105          *      linger2 option.                                 --ANK
2106          */
2107
2108         if (sk->sk_state == TCP_FIN_WAIT2) {
2109                 struct tcp_sock *tp = tcp_sk(sk);
2110                 if (tp->linger2 < 0) {
2111                         tcp_set_state(sk, TCP_CLOSE);
2112                         tcp_send_active_reset(sk, GFP_ATOMIC);
2113                         NET_INC_STATS_BH(sock_net(sk),
2114                                         LINUX_MIB_TCPABORTONLINGER);
2115                 } else {
2116                         const int tmo = tcp_fin_time(sk);
2117
2118                         if (tmo > TCP_TIMEWAIT_LEN) {
2119                                 inet_csk_reset_keepalive_timer(sk,
2120                                                 tmo - TCP_TIMEWAIT_LEN);
2121                         } else {
2122                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2123                                 goto out;
2124                         }
2125                 }
2126         }
2127         if (sk->sk_state != TCP_CLOSE) {
2128                 sk_mem_reclaim(sk);
2129                 if (tcp_check_oom(sk, 0)) {
2130                         tcp_set_state(sk, TCP_CLOSE);
2131                         tcp_send_active_reset(sk, GFP_ATOMIC);
2132                         NET_INC_STATS_BH(sock_net(sk),
2133                                         LINUX_MIB_TCPABORTONMEMORY);
2134                 }
2135         }
2136
2137         if (sk->sk_state == TCP_CLOSE) {
2138                 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2139                 /* We could get here with a non-NULL req if the socket is
2140                  * aborted (e.g., closed with unread data) before 3WHS
2141                  * finishes.
2142                  */
2143                 if (req)
2144                         reqsk_fastopen_remove(sk, req, false);
2145                 inet_csk_destroy_sock(sk);
2146         }
2147         /* Otherwise, socket is reprieved until protocol close. */
2148
2149 out:
2150         bh_unlock_sock(sk);
2151         local_bh_enable();
2152         sock_put(sk);
2153 }
2154 EXPORT_SYMBOL(tcp_close);
2155
2156 /* These states need RST on ABORT according to RFC793 */
2157
2158 static inline bool tcp_need_reset(int state)
2159 {
2160         return (1 << state) &
2161                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2162                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2163 }
2164
2165 int tcp_disconnect(struct sock *sk, int flags)
2166 {
2167         struct inet_sock *inet = inet_sk(sk);
2168         struct inet_connection_sock *icsk = inet_csk(sk);
2169         struct tcp_sock *tp = tcp_sk(sk);
2170         int err = 0;
2171         int old_state = sk->sk_state;
2172
2173         if (old_state != TCP_CLOSE)
2174                 tcp_set_state(sk, TCP_CLOSE);
2175
2176         /* ABORT function of RFC793 */
2177         if (old_state == TCP_LISTEN) {
2178                 inet_csk_listen_stop(sk);
2179         } else if (unlikely(tp->repair)) {
2180                 sk->sk_err = ECONNABORTED;
2181         } else if (tcp_need_reset(old_state) ||
2182                    (tp->snd_nxt != tp->write_seq &&
2183                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2184                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2185                  * states
2186                  */
2187                 tcp_send_active_reset(sk, gfp_any());
2188                 sk->sk_err = ECONNRESET;
2189         } else if (old_state == TCP_SYN_SENT)
2190                 sk->sk_err = ECONNRESET;
2191
2192         tcp_clear_xmit_timers(sk);
2193         __skb_queue_purge(&sk->sk_receive_queue);
2194         tcp_write_queue_purge(sk);
2195         __skb_queue_purge(&tp->out_of_order_queue);
2196
2197         inet->inet_dport = 0;
2198
2199         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2200                 inet_reset_saddr(sk);
2201
2202         sk->sk_shutdown = 0;
2203         sock_reset_flag(sk, SOCK_DONE);
2204         tp->srtt_us = 0;
2205         if ((tp->write_seq += tp->max_window + 2) == 0)
2206                 tp->write_seq = 1;
2207         icsk->icsk_backoff = 0;
2208         tp->snd_cwnd = 2;
2209         icsk->icsk_probes_out = 0;
2210         tp->packets_out = 0;
2211         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2212         tp->snd_cwnd_cnt = 0;
2213         tp->window_clamp = 0;
2214         tcp_set_ca_state(sk, TCP_CA_Open);
2215         tcp_clear_retrans(tp);
2216         inet_csk_delack_init(sk);
2217         tcp_init_send_head(sk);
2218         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2219         __sk_dst_reset(sk);
2220
2221         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2222
2223         sk->sk_error_report(sk);
2224         return err;
2225 }
2226 EXPORT_SYMBOL(tcp_disconnect);
2227
2228 void tcp_sock_destruct(struct sock *sk)
2229 {
2230         inet_sock_destruct(sk);
2231
2232         kfree(inet_csk(sk)->icsk_accept_queue.fastopenq);
2233 }
2234
2235 static inline bool tcp_can_repair_sock(const struct sock *sk)
2236 {
2237         return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2238                 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2239 }
2240
2241 static int tcp_repair_options_est(struct tcp_sock *tp,
2242                 struct tcp_repair_opt __user *optbuf, unsigned int len)
2243 {
2244         struct tcp_repair_opt opt;
2245
2246         while (len >= sizeof(opt)) {
2247                 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2248                         return -EFAULT;
2249
2250                 optbuf++;
2251                 len -= sizeof(opt);
2252
2253                 switch (opt.opt_code) {
2254                 case TCPOPT_MSS:
2255                         tp->rx_opt.mss_clamp = opt.opt_val;
2256                         break;
2257                 case TCPOPT_WINDOW:
2258                         {
2259                                 u16 snd_wscale = opt.opt_val & 0xFFFF;
2260                                 u16 rcv_wscale = opt.opt_val >> 16;
2261
2262                                 if (snd_wscale > 14 || rcv_wscale > 14)
2263                                         return -EFBIG;
2264
2265                                 tp->rx_opt.snd_wscale = snd_wscale;
2266                                 tp->rx_opt.rcv_wscale = rcv_wscale;
2267                                 tp->rx_opt.wscale_ok = 1;
2268                         }
2269                         break;
2270                 case TCPOPT_SACK_PERM:
2271                         if (opt.opt_val != 0)
2272                                 return -EINVAL;
2273
2274                         tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2275                         if (sysctl_tcp_fack)
2276                                 tcp_enable_fack(tp);
2277                         break;
2278                 case TCPOPT_TIMESTAMP:
2279                         if (opt.opt_val != 0)
2280                                 return -EINVAL;
2281
2282                         tp->rx_opt.tstamp_ok = 1;
2283                         break;
2284                 }
2285         }
2286
2287         return 0;
2288 }
2289
2290 /*
2291  *      Socket option code for TCP.
2292  */
2293 static int do_tcp_setsockopt(struct sock *sk, int level,
2294                 int optname, char __user *optval, unsigned int optlen)
2295 {
2296         struct tcp_sock *tp = tcp_sk(sk);
2297         struct inet_connection_sock *icsk = inet_csk(sk);
2298         int val;
2299         int err = 0;
2300
2301         /* These are data/string values, all the others are ints */
2302         switch (optname) {
2303         case TCP_CONGESTION: {
2304                 char name[TCP_CA_NAME_MAX];
2305
2306                 if (optlen < 1)
2307                         return -EINVAL;
2308
2309                 val = strncpy_from_user(name, optval,
2310                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2311                 if (val < 0)
2312                         return -EFAULT;
2313                 name[val] = 0;
2314
2315                 lock_sock(sk);
2316                 err = tcp_set_congestion_control(sk, name);
2317                 release_sock(sk);
2318                 return err;
2319         }
2320         default:
2321                 /* fallthru */
2322                 break;
2323         }
2324
2325         if (optlen < sizeof(int))
2326                 return -EINVAL;
2327
2328         if (get_user(val, (int __user *)optval))
2329                 return -EFAULT;
2330
2331         lock_sock(sk);
2332
2333         switch (optname) {
2334         case TCP_MAXSEG:
2335                 /* Values greater than interface MTU won't take effect. However
2336                  * at the point when this call is done we typically don't yet
2337                  * know which interface is going to be used */
2338                 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2339                         err = -EINVAL;
2340                         break;
2341                 }
2342                 tp->rx_opt.user_mss = val;
2343                 break;
2344
2345         case TCP_NODELAY:
2346                 if (val) {
2347                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2348                          * this option on corked socket is remembered, but
2349                          * it is not activated until cork is cleared.
2350                          *
2351                          * However, when TCP_NODELAY is set we make
2352                          * an explicit push, which overrides even TCP_CORK
2353                          * for currently queued segments.
2354                          */
2355                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2356                         tcp_push_pending_frames(sk);
2357                 } else {
2358                         tp->nonagle &= ~TCP_NAGLE_OFF;
2359                 }
2360                 break;
2361
2362         case TCP_THIN_LINEAR_TIMEOUTS:
2363                 if (val < 0 || val > 1)
2364                         err = -EINVAL;
2365                 else
2366                         tp->thin_lto = val;
2367                 break;
2368
2369         case TCP_THIN_DUPACK:
2370                 if (val < 0 || val > 1)
2371                         err = -EINVAL;
2372                 else {
2373                         tp->thin_dupack = val;
2374                         if (tp->thin_dupack)
2375                                 tcp_disable_early_retrans(tp);
2376                 }
2377                 break;
2378
2379         case TCP_REPAIR:
2380                 if (!tcp_can_repair_sock(sk))
2381                         err = -EPERM;
2382                 else if (val == 1) {
2383                         tp->repair = 1;
2384                         sk->sk_reuse = SK_FORCE_REUSE;
2385                         tp->repair_queue = TCP_NO_QUEUE;
2386                 } else if (val == 0) {
2387                         tp->repair = 0;
2388                         sk->sk_reuse = SK_NO_REUSE;
2389                         tcp_send_window_probe(sk);
2390                 } else
2391                         err = -EINVAL;
2392
2393                 break;
2394
2395         case TCP_REPAIR_QUEUE:
2396                 if (!tp->repair)
2397                         err = -EPERM;
2398                 else if (val < TCP_QUEUES_NR)
2399                         tp->repair_queue = val;
2400                 else
2401                         err = -EINVAL;
2402                 break;
2403
2404         case TCP_QUEUE_SEQ:
2405                 if (sk->sk_state != TCP_CLOSE)
2406                         err = -EPERM;
2407                 else if (tp->repair_queue == TCP_SEND_QUEUE)
2408                         tp->write_seq = val;
2409                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2410                         tp->rcv_nxt = val;
2411                 else
2412                         err = -EINVAL;
2413                 break;
2414
2415         case TCP_REPAIR_OPTIONS:
2416                 if (!tp->repair)
2417                         err = -EINVAL;
2418                 else if (sk->sk_state == TCP_ESTABLISHED)
2419                         err = tcp_repair_options_est(tp,
2420                                         (struct tcp_repair_opt __user *)optval,
2421                                         optlen);
2422                 else
2423                         err = -EPERM;
2424                 break;
2425
2426         case TCP_CORK:
2427                 /* When set indicates to always queue non-full frames.
2428                  * Later the user clears this option and we transmit
2429                  * any pending partial frames in the queue.  This is
2430                  * meant to be used alongside sendfile() to get properly
2431                  * filled frames when the user (for example) must write
2432                  * out headers with a write() call first and then use
2433                  * sendfile to send out the data parts.
2434                  *
2435                  * TCP_CORK can be set together with TCP_NODELAY and it is
2436                  * stronger than TCP_NODELAY.
2437                  */
2438                 if (val) {
2439                         tp->nonagle |= TCP_NAGLE_CORK;
2440                 } else {
2441                         tp->nonagle &= ~TCP_NAGLE_CORK;
2442                         if (tp->nonagle&TCP_NAGLE_OFF)
2443                                 tp->nonagle |= TCP_NAGLE_PUSH;
2444                         tcp_push_pending_frames(sk);
2445                 }
2446                 break;
2447
2448         case TCP_KEEPIDLE:
2449                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2450                         err = -EINVAL;
2451                 else {
2452                         tp->keepalive_time = val * HZ;
2453                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2454                             !((1 << sk->sk_state) &
2455                               (TCPF_CLOSE | TCPF_LISTEN))) {
2456                                 u32 elapsed = keepalive_time_elapsed(tp);
2457                                 if (tp->keepalive_time > elapsed)
2458                                         elapsed = tp->keepalive_time - elapsed;
2459                                 else
2460                                         elapsed = 0;
2461                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2462                         }
2463                 }
2464                 break;
2465         case TCP_KEEPINTVL:
2466                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2467                         err = -EINVAL;
2468                 else
2469                         tp->keepalive_intvl = val * HZ;
2470                 break;
2471         case TCP_KEEPCNT:
2472                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2473                         err = -EINVAL;
2474                 else
2475                         tp->keepalive_probes = val;
2476                 break;
2477         case TCP_SYNCNT:
2478                 if (val < 1 || val > MAX_TCP_SYNCNT)
2479                         err = -EINVAL;
2480                 else
2481                         icsk->icsk_syn_retries = val;
2482                 break;
2483
2484         case TCP_LINGER2:
2485                 if (val < 0)
2486                         tp->linger2 = -1;
2487                 else if (val > sysctl_tcp_fin_timeout / HZ)
2488                         tp->linger2 = 0;
2489                 else
2490                         tp->linger2 = val * HZ;
2491                 break;
2492
2493         case TCP_DEFER_ACCEPT:
2494                 /* Translate value in seconds to number of retransmits */
2495                 icsk->icsk_accept_queue.rskq_defer_accept =
2496                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2497                                         TCP_RTO_MAX / HZ);
2498                 break;
2499
2500         case TCP_WINDOW_CLAMP:
2501                 if (!val) {
2502                         if (sk->sk_state != TCP_CLOSE) {
2503                                 err = -EINVAL;
2504                                 break;
2505                         }
2506                         tp->window_clamp = 0;
2507                 } else
2508                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2509                                                 SOCK_MIN_RCVBUF / 2 : val;
2510                 break;
2511
2512         case TCP_QUICKACK:
2513                 if (!val) {
2514                         icsk->icsk_ack.pingpong = 1;
2515                 } else {
2516                         icsk->icsk_ack.pingpong = 0;
2517                         if ((1 << sk->sk_state) &
2518                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2519                             inet_csk_ack_scheduled(sk)) {
2520                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2521                                 tcp_cleanup_rbuf(sk, 1);
2522                                 if (!(val & 1))
2523                                         icsk->icsk_ack.pingpong = 1;
2524                         }
2525                 }
2526                 break;
2527
2528 #ifdef CONFIG_TCP_MD5SIG
2529         case TCP_MD5SIG:
2530                 /* Read the IP->Key mappings from userspace */
2531                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2532                 break;
2533 #endif
2534         case TCP_USER_TIMEOUT:
2535                 /* Cap the max time in ms TCP will retry or probe the window
2536                  * before giving up and aborting (ETIMEDOUT) a connection.
2537                  */
2538                 if (val < 0)
2539                         err = -EINVAL;
2540                 else
2541                         icsk->icsk_user_timeout = msecs_to_jiffies(val);
2542                 break;
2543
2544         case TCP_FASTOPEN:
2545                 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2546                     TCPF_LISTEN)))
2547                         err = fastopen_init_queue(sk, val);
2548                 else
2549                         err = -EINVAL;
2550                 break;
2551         case TCP_TIMESTAMP:
2552                 if (!tp->repair)
2553                         err = -EPERM;
2554                 else
2555                         tp->tsoffset = val - tcp_time_stamp;
2556                 break;
2557         case TCP_NOTSENT_LOWAT:
2558                 tp->notsent_lowat = val;
2559                 sk->sk_write_space(sk);
2560                 break;
2561         default:
2562                 err = -ENOPROTOOPT;
2563                 break;
2564         }
2565
2566         release_sock(sk);
2567         return err;
2568 }
2569
2570 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2571                    unsigned int optlen)
2572 {
2573         const struct inet_connection_sock *icsk = inet_csk(sk);
2574
2575         if (level != SOL_TCP)
2576                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2577                                                      optval, optlen);
2578         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2579 }
2580 EXPORT_SYMBOL(tcp_setsockopt);
2581
2582 #ifdef CONFIG_COMPAT
2583 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2584                           char __user *optval, unsigned int optlen)
2585 {
2586         if (level != SOL_TCP)
2587                 return inet_csk_compat_setsockopt(sk, level, optname,
2588                                                   optval, optlen);
2589         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2590 }
2591 EXPORT_SYMBOL(compat_tcp_setsockopt);
2592 #endif
2593
2594 /* Return information about state of tcp endpoint in API format. */
2595 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2596 {
2597         const struct tcp_sock *tp = tcp_sk(sk);
2598         const struct inet_connection_sock *icsk = inet_csk(sk);
2599         u32 now = tcp_time_stamp;
2600         u32 rate;
2601
2602         memset(info, 0, sizeof(*info));
2603
2604         info->tcpi_state = sk->sk_state;
2605         info->tcpi_ca_state = icsk->icsk_ca_state;
2606         info->tcpi_retransmits = icsk->icsk_retransmits;
2607         info->tcpi_probes = icsk->icsk_probes_out;
2608         info->tcpi_backoff = icsk->icsk_backoff;
2609
2610         if (tp->rx_opt.tstamp_ok)
2611                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2612         if (tcp_is_sack(tp))
2613                 info->tcpi_options |= TCPI_OPT_SACK;
2614         if (tp->rx_opt.wscale_ok) {
2615                 info->tcpi_options |= TCPI_OPT_WSCALE;
2616                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2617                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2618         }
2619
2620         if (tp->ecn_flags & TCP_ECN_OK)
2621                 info->tcpi_options |= TCPI_OPT_ECN;
2622         if (tp->ecn_flags & TCP_ECN_SEEN)
2623                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2624         if (tp->syn_data_acked)
2625                 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2626
2627         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2628         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2629         info->tcpi_snd_mss = tp->mss_cache;
2630         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2631
2632         if (sk->sk_state == TCP_LISTEN) {
2633                 info->tcpi_unacked = sk->sk_ack_backlog;
2634                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2635         } else {
2636                 info->tcpi_unacked = tp->packets_out;
2637                 info->tcpi_sacked = tp->sacked_out;
2638         }
2639         info->tcpi_lost = tp->lost_out;
2640         info->tcpi_retrans = tp->retrans_out;
2641         info->tcpi_fackets = tp->fackets_out;
2642
2643         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2644         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2645         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2646
2647         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2648         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2649         info->tcpi_rtt = tp->srtt_us >> 3;
2650         info->tcpi_rttvar = tp->mdev_us >> 2;
2651         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2652         info->tcpi_snd_cwnd = tp->snd_cwnd;
2653         info->tcpi_advmss = tp->advmss;
2654         info->tcpi_reordering = tp->reordering;
2655
2656         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2657         info->tcpi_rcv_space = tp->rcvq_space.space;
2658
2659         info->tcpi_total_retrans = tp->total_retrans;
2660
2661         rate = READ_ONCE(sk->sk_pacing_rate);
2662         info->tcpi_pacing_rate = rate != ~0U ? rate : ~0ULL;
2663
2664         rate = READ_ONCE(sk->sk_max_pacing_rate);
2665         info->tcpi_max_pacing_rate = rate != ~0U ? rate : ~0ULL;
2666 }
2667 EXPORT_SYMBOL_GPL(tcp_get_info);
2668
2669 static int do_tcp_getsockopt(struct sock *sk, int level,
2670                 int optname, char __user *optval, int __user *optlen)
2671 {
2672         struct inet_connection_sock *icsk = inet_csk(sk);
2673         struct tcp_sock *tp = tcp_sk(sk);
2674         int val, len;
2675
2676         if (get_user(len, optlen))
2677                 return -EFAULT;
2678
2679         len = min_t(unsigned int, len, sizeof(int));
2680
2681         if (len < 0)
2682                 return -EINVAL;
2683
2684         switch (optname) {
2685         case TCP_MAXSEG:
2686                 val = tp->mss_cache;
2687                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2688                         val = tp->rx_opt.user_mss;
2689                 if (tp->repair)
2690                         val = tp->rx_opt.mss_clamp;
2691                 break;
2692         case TCP_NODELAY:
2693                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2694                 break;
2695         case TCP_CORK:
2696                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2697                 break;
2698         case TCP_KEEPIDLE:
2699                 val = keepalive_time_when(tp) / HZ;
2700                 break;
2701         case TCP_KEEPINTVL:
2702                 val = keepalive_intvl_when(tp) / HZ;
2703                 break;
2704         case TCP_KEEPCNT:
2705                 val = keepalive_probes(tp);
2706                 break;
2707         case TCP_SYNCNT:
2708                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2709                 break;
2710         case TCP_LINGER2:
2711                 val = tp->linger2;
2712                 if (val >= 0)
2713                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2714                 break;
2715         case TCP_DEFER_ACCEPT:
2716                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2717                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2718                 break;
2719         case TCP_WINDOW_CLAMP:
2720                 val = tp->window_clamp;
2721                 break;
2722         case TCP_INFO: {
2723                 struct tcp_info info;
2724
2725                 if (get_user(len, optlen))
2726                         return -EFAULT;
2727
2728                 tcp_get_info(sk, &info);
2729
2730                 len = min_t(unsigned int, len, sizeof(info));
2731                 if (put_user(len, optlen))
2732                         return -EFAULT;
2733                 if (copy_to_user(optval, &info, len))
2734                         return -EFAULT;
2735                 return 0;
2736         }
2737         case TCP_QUICKACK:
2738                 val = !icsk->icsk_ack.pingpong;
2739                 break;
2740
2741         case TCP_CONGESTION:
2742                 if (get_user(len, optlen))
2743                         return -EFAULT;
2744                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2745                 if (put_user(len, optlen))
2746                         return -EFAULT;
2747                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2748                         return -EFAULT;
2749                 return 0;
2750
2751         case TCP_THIN_LINEAR_TIMEOUTS:
2752                 val = tp->thin_lto;
2753                 break;
2754         case TCP_THIN_DUPACK:
2755                 val = tp->thin_dupack;
2756                 break;
2757
2758         case TCP_REPAIR:
2759                 val = tp->repair;
2760                 break;
2761
2762         case TCP_REPAIR_QUEUE:
2763                 if (tp->repair)
2764                         val = tp->repair_queue;
2765                 else
2766                         return -EINVAL;
2767                 break;
2768
2769         case TCP_QUEUE_SEQ:
2770                 if (tp->repair_queue == TCP_SEND_QUEUE)
2771                         val = tp->write_seq;
2772                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2773                         val = tp->rcv_nxt;
2774                 else
2775                         return -EINVAL;
2776                 break;
2777
2778         case TCP_USER_TIMEOUT:
2779                 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2780                 break;
2781
2782         case TCP_FASTOPEN:
2783                 if (icsk->icsk_accept_queue.fastopenq)
2784                         val = icsk->icsk_accept_queue.fastopenq->max_qlen;
2785                 else
2786                         val = 0;
2787                 break;
2788
2789         case TCP_TIMESTAMP:
2790                 val = tcp_time_stamp + tp->tsoffset;
2791                 break;
2792         case TCP_NOTSENT_LOWAT:
2793                 val = tp->notsent_lowat;
2794                 break;
2795         default:
2796                 return -ENOPROTOOPT;
2797         }
2798
2799         if (put_user(len, optlen))
2800                 return -EFAULT;
2801         if (copy_to_user(optval, &val, len))
2802                 return -EFAULT;
2803         return 0;
2804 }
2805
2806 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2807                    int __user *optlen)
2808 {
2809         struct inet_connection_sock *icsk = inet_csk(sk);
2810
2811         if (level != SOL_TCP)
2812                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2813                                                      optval, optlen);
2814         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2815 }
2816 EXPORT_SYMBOL(tcp_getsockopt);
2817
2818 #ifdef CONFIG_COMPAT
2819 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2820                           char __user *optval, int __user *optlen)
2821 {
2822         if (level != SOL_TCP)
2823                 return inet_csk_compat_getsockopt(sk, level, optname,
2824                                                   optval, optlen);
2825         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2826 }
2827 EXPORT_SYMBOL(compat_tcp_getsockopt);
2828 #endif
2829
2830 #ifdef CONFIG_TCP_MD5SIG
2831 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
2832 static DEFINE_MUTEX(tcp_md5sig_mutex);
2833 static bool tcp_md5sig_pool_populated = false;
2834
2835 static void __tcp_alloc_md5sig_pool(void)
2836 {
2837         int cpu;
2838
2839         for_each_possible_cpu(cpu) {
2840                 if (!per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm) {
2841                         struct crypto_hash *hash;
2842
2843                         hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2844                         if (IS_ERR_OR_NULL(hash))
2845                                 return;
2846                         per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm = hash;
2847                 }
2848         }
2849         /* before setting tcp_md5sig_pool_populated, we must commit all writes
2850          * to memory. See smp_rmb() in tcp_get_md5sig_pool()
2851          */
2852         smp_wmb();
2853         tcp_md5sig_pool_populated = true;
2854 }
2855
2856 bool tcp_alloc_md5sig_pool(void)
2857 {
2858         if (unlikely(!tcp_md5sig_pool_populated)) {
2859                 mutex_lock(&tcp_md5sig_mutex);
2860
2861                 if (!tcp_md5sig_pool_populated)
2862                         __tcp_alloc_md5sig_pool();
2863
2864                 mutex_unlock(&tcp_md5sig_mutex);
2865         }
2866         return tcp_md5sig_pool_populated;
2867 }
2868 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2869
2870
2871 /**
2872  *      tcp_get_md5sig_pool - get md5sig_pool for this user
2873  *
2874  *      We use percpu structure, so if we succeed, we exit with preemption
2875  *      and BH disabled, to make sure another thread or softirq handling
2876  *      wont try to get same context.
2877  */
2878 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2879 {
2880         local_bh_disable();
2881
2882         if (tcp_md5sig_pool_populated) {
2883                 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
2884                 smp_rmb();
2885                 return this_cpu_ptr(&tcp_md5sig_pool);
2886         }
2887         local_bh_enable();
2888         return NULL;
2889 }
2890 EXPORT_SYMBOL(tcp_get_md5sig_pool);
2891
2892 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2893                         const struct tcphdr *th)
2894 {
2895         struct scatterlist sg;
2896         struct tcphdr hdr;
2897         int err;
2898
2899         /* We are not allowed to change tcphdr, make a local copy */
2900         memcpy(&hdr, th, sizeof(hdr));
2901         hdr.check = 0;
2902
2903         /* options aren't included in the hash */
2904         sg_init_one(&sg, &hdr, sizeof(hdr));
2905         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
2906         return err;
2907 }
2908 EXPORT_SYMBOL(tcp_md5_hash_header);
2909
2910 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2911                           const struct sk_buff *skb, unsigned int header_len)
2912 {
2913         struct scatterlist sg;
2914         const struct tcphdr *tp = tcp_hdr(skb);
2915         struct hash_desc *desc = &hp->md5_desc;
2916         unsigned int i;
2917         const unsigned int head_data_len = skb_headlen(skb) > header_len ?
2918                                            skb_headlen(skb) - header_len : 0;
2919         const struct skb_shared_info *shi = skb_shinfo(skb);
2920         struct sk_buff *frag_iter;
2921
2922         sg_init_table(&sg, 1);
2923
2924         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
2925         if (crypto_hash_update(desc, &sg, head_data_len))
2926                 return 1;
2927
2928         for (i = 0; i < shi->nr_frags; ++i) {
2929                 const struct skb_frag_struct *f = &shi->frags[i];
2930                 unsigned int offset = f->page_offset;
2931                 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
2932
2933                 sg_set_page(&sg, page, skb_frag_size(f),
2934                             offset_in_page(offset));
2935                 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
2936                         return 1;
2937         }
2938
2939         skb_walk_frags(skb, frag_iter)
2940                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
2941                         return 1;
2942
2943         return 0;
2944 }
2945 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
2946
2947 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
2948 {
2949         struct scatterlist sg;
2950
2951         sg_init_one(&sg, key->key, key->keylen);
2952         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
2953 }
2954 EXPORT_SYMBOL(tcp_md5_hash_key);
2955
2956 #endif
2957
2958 void tcp_done(struct sock *sk)
2959 {
2960         struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2961
2962         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2963                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
2964
2965         tcp_set_state(sk, TCP_CLOSE);
2966         tcp_clear_xmit_timers(sk);
2967         if (req)
2968                 reqsk_fastopen_remove(sk, req, false);
2969
2970         sk->sk_shutdown = SHUTDOWN_MASK;
2971
2972         if (!sock_flag(sk, SOCK_DEAD))
2973                 sk->sk_state_change(sk);
2974         else
2975                 inet_csk_destroy_sock(sk);
2976 }
2977 EXPORT_SYMBOL_GPL(tcp_done);
2978
2979 extern struct tcp_congestion_ops tcp_reno;
2980
2981 static __initdata unsigned long thash_entries;
2982 static int __init set_thash_entries(char *str)
2983 {
2984         ssize_t ret;
2985
2986         if (!str)
2987                 return 0;
2988
2989         ret = kstrtoul(str, 0, &thash_entries);
2990         if (ret)
2991                 return 0;
2992
2993         return 1;
2994 }
2995 __setup("thash_entries=", set_thash_entries);
2996
2997 static void __init tcp_init_mem(void)
2998 {
2999         unsigned long limit = nr_free_buffer_pages() / 8;
3000         limit = max(limit, 128UL);
3001         sysctl_tcp_mem[0] = limit / 4 * 3;
3002         sysctl_tcp_mem[1] = limit;
3003         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3004 }
3005
3006 void __init tcp_init(void)
3007 {
3008         unsigned long limit;
3009         int max_rshare, max_wshare, cnt;
3010         unsigned int i;
3011
3012         sock_skb_cb_check_size(sizeof(struct tcp_skb_cb));
3013
3014         percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3015         percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3016         tcp_hashinfo.bind_bucket_cachep =
3017                 kmem_cache_create("tcp_bind_bucket",
3018                                   sizeof(struct inet_bind_bucket), 0,
3019                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3020
3021         /* Size and allocate the main established and bind bucket
3022          * hash tables.
3023          *
3024          * The methodology is similar to that of the buffer cache.
3025          */
3026         tcp_hashinfo.ehash =
3027                 alloc_large_system_hash("TCP established",
3028                                         sizeof(struct inet_ehash_bucket),
3029                                         thash_entries,
3030                                         17, /* one slot per 128 KB of memory */
3031                                         0,
3032                                         NULL,
3033                                         &tcp_hashinfo.ehash_mask,
3034                                         0,
3035                                         thash_entries ? 0 : 512 * 1024);
3036         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3037                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3038
3039         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3040                 panic("TCP: failed to alloc ehash_locks");
3041         tcp_hashinfo.bhash =
3042                 alloc_large_system_hash("TCP bind",
3043                                         sizeof(struct inet_bind_hashbucket),
3044                                         tcp_hashinfo.ehash_mask + 1,
3045                                         17, /* one slot per 128 KB of memory */
3046                                         0,
3047                                         &tcp_hashinfo.bhash_size,
3048                                         NULL,
3049                                         0,
3050                                         64 * 1024);
3051         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3052         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3053                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3054                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3055         }
3056
3057
3058         cnt = tcp_hashinfo.ehash_mask + 1;
3059
3060         tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3061         sysctl_tcp_max_orphans = cnt / 2;
3062         sysctl_max_syn_backlog = max(128, cnt / 256);
3063
3064         tcp_init_mem();
3065         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3066         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3067         max_wshare = min(4UL*1024*1024, limit);
3068         max_rshare = min(6UL*1024*1024, limit);
3069
3070         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3071         sysctl_tcp_wmem[1] = 16*1024;
3072         sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3073
3074         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3075         sysctl_tcp_rmem[1] = 87380;
3076         sysctl_tcp_rmem[2] = max(87380, max_rshare);
3077
3078         pr_info("Hash tables configured (established %u bind %u)\n",
3079                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3080
3081         tcp_metrics_init();
3082         BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3083         tcp_tasklet_init();
3084 }