net/ipv4/tcp.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the  BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              Implementation of the Transmission Control Protocol(TCP).
   7  *
   8  * Authors:     Ross Biro
   9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
  11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
  12  *              Florian La Roche, <flla@stud.uni-sb.de>
  13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
  15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
  16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
  17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  18  *              Jorge Cwik, <jorge@laser.satlink.net>
  19  *
  20  * Fixes:
  21  *              Alan Cox        :       Numerous verify_area() calls
  22  *              Alan Cox        :       Set the ACK bit on a reset
  23  *              Alan Cox        :       Stopped it crashing if it closed while
  24  *                                      sk->inuse=1 and was trying to connect
  25  *                                      (tcp_err()).
  26  *              Alan Cox        :       All icmp error handling was broken
  27  *                                      pointers passed where wrong and the
  28  *                                      socket was looked up backwards. Nobody
  29  *                                      tested any icmp error code obviously.
  30  *              Alan Cox        :       tcp_err() now handled properly. It
  31  *                                      wakes people on errors. poll
  32  *                                      behaves and the icmp error race
  33  *                                      has gone by moving it into sock.c
  34  *              Alan Cox        :       tcp_send_reset() fixed to work for
  35  *                                      everything not just packets for
  36  *                                      unknown sockets.
  37  *              Alan Cox        :       tcp option processing.
  38  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
  39  *                                      syn rule wrong]
  40  *              Herp Rosmanith  :       More reset fixes
  41  *              Alan Cox        :       No longer acks invalid rst frames.
  42  *                                      Acking any kind of RST is right out.
  43  *              Alan Cox        :       Sets an ignore me flag on an rst
  44  *                                      receive otherwise odd bits of prattle
  45  *                                      escape still
  46  *              Alan Cox        :       Fixed another acking RST frame bug.
  47  *                                      Should stop LAN workplace lockups.
  48  *              Alan Cox        :       Some tidyups using the new skb list
  49  *                                      facilities
  50  *              Alan Cox        :       sk->keepopen now seems to work
  51  *              Alan Cox        :       Pulls options out correctly on accepts
  52  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
  53  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
  54  *                                      bit to skb ops.
  55  *              Alan Cox        :       Tidied tcp_data to avoid a potential
  56  *                                      nasty.
  57  *              Alan Cox        :       Added some better commenting, as the
  58  *                                      tcp is hard to follow
  59  *              Alan Cox        :       Removed incorrect check for 20 * psh
  60  *      Michael O'Reilly        :       ack < copied bug fix.
  61  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
  62  *              Alan Cox        :       FIN with no memory -> CRASH
  63  *              Alan Cox        :       Added socket option proto entries.
  64  *                                      Also added awareness of them to accept.
  65  *              Alan Cox        :       Added TCP options (SOL_TCP)
  66  *              Alan Cox        :       Switched wakeup calls to callbacks,
  67  *                                      so the kernel can layer network
  68  *                                      sockets.
  69  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
  70  *              Alan Cox        :       Handle FIN (more) properly (we hope).
  71  *              Alan Cox        :       RST frames sent on unsynchronised
  72  *                                      state ack error.
  73  *              Alan Cox        :       Put in missing check for SYN bit.
  74  *              Alan Cox        :       Added tcp_select_window() aka NET2E
  75  *                                      window non shrink trick.
  76  *              Alan Cox        :       Added a couple of small NET2E timer
  77  *                                      fixes
  78  *              Charles Hedrick :       TCP fixes
  79  *              Toomas Tamm     :       TCP window fixes
  80  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
  81  *              Charles Hedrick :       Rewrote most of it to actually work
  82  *              Linus           :       Rewrote tcp_read() and URG handling
  83  *                                      completely
  84  *              Gerhard Koerting:       Fixed some missing timer handling
  85  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
  86  *              Gerhard Koerting:       PC/TCP workarounds
  87  *              Adam Caldwell   :       Assorted timer/timing errors
  88  *              Matthew Dillon  :       Fixed another RST bug
  89  *              Alan Cox        :       Move to kernel side addressing changes.
  90  *              Alan Cox        :       Beginning work on TCP fastpathing
  91  *                                      (not yet usable)
  92  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
  93  *              Alan Cox        :       TCP fast path debugging
  94  *              Alan Cox        :       Window clamping
  95  *              Michael Riepe   :       Bug in tcp_check()
  96  *              Matt Dillon     :       More TCP improvements and RST bug fixes
  97  *              Matt Dillon     :       Yet more small nasties remove from the
  98  *                                      TCP code (Be very nice to this man if
  99  *                                      tcp finally works 100%) 8)
 100  *              Alan Cox        :       BSD accept semantics.
 101  *              Alan Cox        :       Reset on closedown bug.
 102  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
 103  *              Michael Pall    :       Handle poll() after URG properly in
 104  *                                      all cases.
 105  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
 106  *                                      (multi URG PUSH broke rlogin).
 107  *              Michael Pall    :       Fix the multi URG PUSH problem in
 108  *                                      tcp_readable(), poll() after URG
 109  *                                      works now.
 110  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
 111  *                                      BSD api.
 112  *              Alan Cox        :       Changed the semantics of sk->socket to
 113  *                                      fix a race and a signal problem with
 114  *                                      accept() and async I/O.
 115  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
 116  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
 117  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
 118  *                                      clients/servers which listen in on
 119  *                                      fixed ports.
 120  *              Alan Cox        :       Cleaned the above up and shrank it to
 121  *                                      a sensible code size.
 122  *              Alan Cox        :       Self connect lockup fix.
 123  *              Alan Cox        :       No connect to multicast.
 124  *              Ross Biro       :       Close unaccepted children on master
 125  *                                      socket close.
 126  *              Alan Cox        :       Reset tracing code.
 127  *              Alan Cox        :       Spurious resets on shutdown.
 128  *              Alan Cox        :       Giant 15 minute/60 second timer error
 129  *              Alan Cox        :       Small whoops in polling before an
 130  *                                      accept.
 131  *              Alan Cox        :       Kept the state trace facility since
 132  *                                      it's handy for debugging.
 133  *              Alan Cox        :       More reset handler fixes.
 134  *              Alan Cox        :       Started rewriting the code based on
 135  *                                      the RFC's for other useful protocol
 136  *                                      references see: Comer, KA9Q NOS, and
 137  *                                      for a reference on the difference
 138  *                                      between specifications and how BSD
 139  *                                      works see the 4.4lite source.
 140  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
 141  *                                      close.
 142  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
 143  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
 144  *              Alan Cox        :       Reimplemented timers as per the RFC
 145  *                                      and using multiple timers for sanity.
 146  *              Alan Cox        :       Small bug fixes, and a lot of new
 147  *                                      comments.
 148  *              Alan Cox        :       Fixed dual reader crash by locking
 149  *                                      the buffers (much like datagram.c)
 150  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
 151  *                                      now gets fed up of retrying without
 152  *                                      (even a no space) answer.
 153  *              Alan Cox        :       Extracted closing code better
 154  *              Alan Cox        :       Fixed the closing state machine to
 155  *                                      resemble the RFC.
 156  *              Alan Cox        :       More 'per spec' fixes.
 157  *              Jorge Cwik      :       Even faster checksumming.
 158  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
 159  *                                      only frames. At least one pc tcp stack
 160  *                                      generates them.
 161  *              Alan Cox        :       Cache last socket.
 162  *              Alan Cox        :       Per route irtt.
 163  *              Matt Day        :       poll()->select() match BSD precisely on error
 164  *              Alan Cox        :       New buffers
 165  *              Marc Tamsky     :       Various sk->prot->retransmits and
 166  *                                      sk->retransmits misupdating fixed.
 167  *                                      Fixed tcp_write_timeout: stuck close,
 168  *                                      and TCP syn retries gets used now.
 169  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
 170  *                                      ack if state is TCP_CLOSED.
 171  *              Alan Cox        :       Look up device on a retransmit - routes may
 172  *                                      change. Doesn't yet cope with MSS shrink right
 173  *                                      but it's a start!
 174  *              Marc Tamsky     :       Closing in closing fixes.
 175  *              Mike Shaver     :       RFC1122 verifications.
 176  *              Alan Cox        :       rcv_saddr errors.
 177  *              Alan Cox        :       Block double connect().
 178  *              Alan Cox        :       Small hooks for enSKIP.
 179  *              Alexey Kuznetsov:       Path MTU discovery.
 180  *              Alan Cox        :       Support soft errors.
 181  *              Alan Cox        :       Fix MTU discovery pathological case
 182  *                                      when the remote claims no mtu!
 183  *              Marc Tamsky     :       TCP_CLOSE fix.
 184  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
 185  *                                      window but wrong (fixes NT lpd problems)
 186  *              Pedro Roque     :       Better TCP window handling, delayed ack.
 187  *              Joerg Reuter    :       No modification of locked buffers in
 188  *                                      tcp_do_retransmit()
 189  *              Eric Schenk     :       Changed receiver side silly window
 190  *                                      avoidance algorithm to BSD style
 191  *                                      algorithm. This doubles throughput
 192  *                                      against machines running Solaris,
 193  *                                      and seems to result in general
 194  *                                      improvement.
 195  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
 196  *      Willy Konynenberg       :       Transparent proxying support.
 197  *      Mike McLagan            :       Routing by source
 198  *              Keith Owens     :       Do proper merging with partial SKB's in
 199  *                                      tcp_do_sendmsg to avoid burstiness.
 200  *              Eric Schenk     :       Fix fast close down bug with
 201  *                                      shutdown() followed by close().
 202  *              Andi Kleen      :       Make poll agree with SIGIO
 203  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
 204  *                                      lingertime == 0 (RFC 793 ABORT Call)
 205  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
 206  *                                      csum_and_copy_from_user() if possible.
 207  *
 208  *              This program is free software; you can redistribute it and/or
 209  *              modify it under the terms of the GNU General Public License
 210  *              as published by the Free Software Foundation; either version
 211  *              2 of the License, or(at your option) any later version.
 212  *
 213  * Description of States:
 214  *
 215  *      TCP_SYN_SENT            sent a connection request, waiting for ack
 216  *
 217  *      TCP_SYN_RECV            received a connection request, sent ack,
 218  *                              waiting for final ack in three-way handshake.
 219  *
 220  *      TCP_ESTABLISHED         connection established
 221  *
 222  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
 223  *                              transmission of remaining buffered data
 224  *
 225  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
 226  *                              to shutdown
 227  *
 228  *      TCP_CLOSING             both sides have shutdown but we still have
 229  *                              data we have to finish sending
 230  *
 231  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
 232  *                              closed, can only be entered from FIN_WAIT2
 233  *                              or CLOSING.  Required because the other end
 234  *                              may not have gotten our last ACK causing it
 235  *                              to retransmit the data packet (which we ignore)
 236  *
 237  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
 238  *                              us to finish writing our data and to shutdown
 239  *                              (we have to close() to move on to LAST_ACK)
 240  *
 241  *      TCP_LAST_ACK            out side has shutdown after remote has
 242  *                              shutdown.  There may still be data in our
 243  *                              buffer that we have to finish sending
 244  *
 245  *      TCP_CLOSE               socket is finished
 246  */
 247
 248 #define pr_fmt(fmt) "TCP: " fmt
 249
 250 #include <crypto/hash.h>
 251 #include <linux/kernel.h>
 252 #include <linux/module.h>
 253 #include <linux/types.h>
 254 #include <linux/fcntl.h>
 255 #include <linux/poll.h>
 256 #include <linux/inet_diag.h>
 257 #include <linux/init.h>
 258 #include <linux/fs.h>
 259 #include <linux/skbuff.h>
 260 #include <linux/scatterlist.h>
 261 #include <linux/splice.h>
 262 #include <linux/net.h>
 263 #include <linux/socket.h>
 264 #include <linux/random.h>
 265 #include <linux/bootmem.h>
 266 #include <linux/highmem.h>
 267 #include <linux/swap.h>
 268 #include <linux/cache.h>
 269 #include <linux/err.h>
 270 #include <linux/time.h>
 271 #include <linux/slab.h>
 272 #include <linux/errqueue.h>
 273 #include <linux/static_key.h>
 274
 275 #include <net/icmp.h>
 276 #include <net/inet_common.h>
 277 #include <net/tcp.h>
 278 #include <net/xfrm.h>
 279 #include <net/ip.h>
 280 #include <net/sock.h>
 281
 282 #include <linux/uaccess.h>
 283 #include <asm/ioctls.h>
 284 #include <net/busy_poll.h>
 285
 286 struct percpu_counter tcp_orphan_count;
 287 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 288
 289 long sysctl_tcp_mem[3] __read_mostly;
 290 EXPORT_SYMBOL(sysctl_tcp_mem);
 291
 292 atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
 293 EXPORT_SYMBOL(tcp_memory_allocated);
 294
 295 #if IS_ENABLED(CONFIG_SMC)
 296 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
 297 EXPORT_SYMBOL(tcp_have_smc);
 298 #endif
 299
 300 /*
 301  * Current number of TCP sockets.
 302  */
 303 struct percpu_counter tcp_sockets_allocated;
 304 EXPORT_SYMBOL(tcp_sockets_allocated);
 305
 306 /*
 307  * TCP splice context
 308  */
 309 struct tcp_splice_state {
 310         struct pipe_inode_info *pipe;
 311         size_t len;
 312         unsigned int flags;
 313 };
 314
 315 /*
 316  * Pressure flag: try to collapse.
 317  * Technical note: it is used by multiple contexts non atomically.
 318  * All the __sk_mem_schedule() is of this nature: accounting
 319  * is strict, actions are advisory and have some latency.
 320  */
 321 unsigned long tcp_memory_pressure __read_mostly;
 322 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
 323
 324 void tcp_enter_memory_pressure(struct sock *sk)
 325 {
 326         unsigned long val;
 327
 328         if (tcp_memory_pressure)
 329                 return;
 330         val = jiffies;
 331
 332         if (!val)
 333                 val--;
 334         if (!cmpxchg(&tcp_memory_pressure, 0, val))
 335                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
 336 }
 337 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
 338
 339 void tcp_leave_memory_pressure(struct sock *sk)
 340 {
 341         unsigned long val;
 342
 343         if (!tcp_memory_pressure)
 344                 return;
 345         val = xchg(&tcp_memory_pressure, 0);
 346         if (val)
 347                 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
 348                               jiffies_to_msecs(jiffies - val));
 349 }
 350 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
 351
 352 /* Convert seconds to retransmits based on initial and max timeout */
 353 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
 354 {
 355         u8 res = 0;
 356
 357         if (seconds > 0) {
 358                 int period = timeout;
 359
 360                 res = 1;
 361                 while (seconds > period && res < 255) {
 362                         res++;
 363                         timeout <<= 1;
 364                         if (timeout > rto_max)
 365                                 timeout = rto_max;
 366                         period += timeout;
 367                 }
 368         }
 369         return res;
 370 }
 371
 372 /* Convert retransmits to seconds based on initial and max timeout */
 373 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
 374 {
 375         int period = 0;
 376
 377         if (retrans > 0) {
 378                 period = timeout;
 379                 while (--retrans) {
 380                         timeout <<= 1;
 381                         if (timeout > rto_max)
 382                                 timeout = rto_max;
 383                         period += timeout;
 384                 }
 385         }
 386         return period;
 387 }
 388
 389 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
 390 {
 391         u32 rate = READ_ONCE(tp->rate_delivered);
 392         u32 intv = READ_ONCE(tp->rate_interval_us);
 393         u64 rate64 = 0;
 394
 395         if (rate && intv) {
 396                 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
 397                 do_div(rate64, intv);
 398         }
 399         return rate64;
 400 }
 401
 402 /* Address-family independent initialization for a tcp_sock.
 403  *
 404  * NOTE: A lot of things set to zero explicitly by call to
 405  *       sk_alloc() so need not be done here.
 406  */
 407 void tcp_init_sock(struct sock *sk)
 408 {
 409         struct inet_connection_sock *icsk = inet_csk(sk);
 410         struct tcp_sock *tp = tcp_sk(sk);
 411
 412         tp->out_of_order_queue = RB_ROOT;
 413         sk->tcp_rtx_queue = RB_ROOT;
 414         tcp_init_xmit_timers(sk);
 415         INIT_LIST_HEAD(&tp->tsq_node);
 416         INIT_LIST_HEAD(&tp->tsorted_sent_queue);
 417
 418         icsk->icsk_rto = TCP_TIMEOUT_INIT;
 419         tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
 420         minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
 421
 422         /* So many TCP implementations out there (incorrectly) count the
 423          * initial SYN frame in their delayed-ACK and congestion control
 424          * algorithms that we must have the following bandaid to talk
 425          * efficiently to them.  -DaveM
 426          */
 427         tp->snd_cwnd = TCP_INIT_CWND;
 428
 429         /* There's a bubble in the pipe until at least the first ACK. */
 430         tp->app_limited = ~0U;
 431
 432         /* See draft-stevens-tcpca-spec-01 for discussion of the
 433          * initialization of these values.
 434          */
 435         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 436         tp->snd_cwnd_clamp = ~0;
 437         tp->mss_cache = TCP_MSS_DEFAULT;
 438
 439         tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
 440         tcp_assign_congestion_control(sk);
 441
 442         tp->tsoffset = 0;
 443         tp->rack.reo_wnd_steps = 1;
 444
 445         sk->sk_state = TCP_CLOSE;
 446
 447         sk->sk_write_space = sk_stream_write_space;
 448         sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
 449
 450         icsk->icsk_sync_mss = tcp_sync_mss;
 451
 452         sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
 453         sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
 454
 455         sk_sockets_allocated_inc(sk);
 456 }
 457 EXPORT_SYMBOL(tcp_init_sock);
 458
 459 void tcp_init_transfer(struct sock *sk, int bpf_op)
 460 {
 461         struct inet_connection_sock *icsk = inet_csk(sk);
 462
 463         tcp_mtup_init(sk);
 464         icsk->icsk_af_ops->rebuild_header(sk);
 465         tcp_init_metrics(sk);
 466         tcp_call_bpf(sk, bpf_op);
 467         tcp_init_congestion_control(sk);
 468         tcp_init_buffer_space(sk);
 469 }
 470
 471 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
 472 {
 473         struct sk_buff *skb = tcp_write_queue_tail(sk);
 474
 475         if (tsflags && skb) {
 476                 struct skb_shared_info *shinfo = skb_shinfo(skb);
 477                 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 478
 479                 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
 480                 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
 481                         tcb->txstamp_ack = 1;
 482                 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
 483                         shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
 484         }
 485 }
 486
 487 /*
 488  *      Wait for a TCP event.
 489  *
 490  *      Note that we don't need to lock the socket, as the upper poll layers
 491  *      take care of normal races (between the test and the event) and we don't
 492  *      go look at any of the socket buffers directly.
 493  */
 494 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 495 {
 496         unsigned int mask;
 497         struct sock *sk = sock->sk;
 498         const struct tcp_sock *tp = tcp_sk(sk);
 499         int state;
 500
 501         sock_rps_record_flow(sk);
 502
 503         sock_poll_wait(file, sk_sleep(sk), wait);
 504
 505         state = inet_sk_state_load(sk);
 506         if (state == TCP_LISTEN)
 507                 return inet_csk_listen_poll(sk);
 508
 509         /* Socket is not locked. We are protected from async events
 510          * by poll logic and correct handling of state changes
 511          * made by other threads is impossible in any case.
 512          */
 513
 514         mask = 0;
 515
 516         /*
 517          * POLLHUP is certainly not done right. But poll() doesn't
 518          * have a notion of HUP in just one direction, and for a
 519          * socket the read side is more interesting.
 520          *
 521          * Some poll() documentation says that POLLHUP is incompatible
 522          * with the POLLOUT/POLLWR flags, so somebody should check this
 523          * all. But careful, it tends to be safer to return too many
 524          * bits than too few, and you can easily break real applications
 525          * if you don't tell them that something has hung up!
 526          *
 527          * Check-me.
 528          *
 529          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
 530          * our fs/select.c). It means that after we received EOF,
 531          * poll always returns immediately, making impossible poll() on write()
 532          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
 533          * if and only if shutdown has been made in both directions.
 534          * Actually, it is interesting to look how Solaris and DUX
 535          * solve this dilemma. I would prefer, if POLLHUP were maskable,
 536          * then we could set it on SND_SHUTDOWN. BTW examples given
 537          * in Stevens' books assume exactly this behaviour, it explains
 538          * why POLLHUP is incompatible with POLLOUT.    --ANK
 539          *
 540          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 541          * blocking on fresh not-connected or disconnected socket. --ANK
 542          */
 543         if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
 544                 mask |= POLLHUP;
 545         if (sk->sk_shutdown & RCV_SHUTDOWN)
 546                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
 547
 548         /* Connected or passive Fast Open socket? */
 549         if (state != TCP_SYN_SENT &&
 550             (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
 551                 int target = sock_rcvlowat(sk, 0, INT_MAX);
 552
 553                 if (tp->urg_seq == tp->copied_seq &&
 554                     !sock_flag(sk, SOCK_URGINLINE) &&
 555                     tp->urg_data)
 556                         target++;
 557
 558                 if (tp->rcv_nxt - tp->copied_seq >= target)
 559                         mask |= POLLIN | POLLRDNORM;
 560
 561                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 562                         if (sk_stream_is_writeable(sk)) {
 563                                 mask |= POLLOUT | POLLWRNORM;
 564                         } else {  /* send SIGIO later */
 565                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 566                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 567
 568                                 /* Race breaker. If space is freed after
 569                                  * wspace test but before the flags are set,
 570                                  * IO signal will be lost. Memory barrier
 571                                  * pairs with the input side.
 572                                  */
 573                                 smp_mb__after_atomic();
 574                                 if (sk_stream_is_writeable(sk))
 575                                         mask |= POLLOUT | POLLWRNORM;
 576                         }
 577                 } else
 578                         mask |= POLLOUT | POLLWRNORM;
 579
 580                 if (tp->urg_data & TCP_URG_VALID)
 581                         mask |= POLLPRI;
 582         } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
 583                 /* Active TCP fastopen socket with defer_connect
 584                  * Return POLLOUT so application can call write()
 585                  * in order for kernel to generate SYN+data
 586                  */
 587                 mask |= POLLOUT | POLLWRNORM;
 588         }
 589         /* This barrier is coupled with smp_wmb() in tcp_reset() */
 590         smp_rmb();
 591         if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
 592                 mask |= POLLERR;
 593
 594         return mask;
 595 }
 596 EXPORT_SYMBOL(tcp_poll);
 597
 598 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 599 {
 600         struct tcp_sock *tp = tcp_sk(sk);
 601         int answ;
 602         bool slow;
 603
 604         switch (cmd) {
 605         case SIOCINQ:
 606                 if (sk->sk_state == TCP_LISTEN)
 607                         return -EINVAL;
 608
 609                 slow = lock_sock_fast(sk);
 610                 answ = tcp_inq(sk);
 611                 unlock_sock_fast(sk, slow);
 612                 break;
 613         case SIOCATMARK:
 614                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 615                 break;
 616         case SIOCOUTQ:
 617                 if (sk->sk_state == TCP_LISTEN)
 618                         return -EINVAL;
 619
 620                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 621                         answ = 0;
 622                 else
 623                         answ = tp->write_seq - tp->snd_una;
 624                 break;
 625         case SIOCOUTQNSD:
 626                 if (sk->sk_state == TCP_LISTEN)
 627                         return -EINVAL;
 628
 629                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 630                         answ = 0;
 631                 else
 632                         answ = tp->write_seq - tp->snd_nxt;
 633                 break;
 634         default:
 635                 return -ENOIOCTLCMD;
 636         }
 637
 638         return put_user(answ, (int __user *)arg);
 639 }
 640 EXPORT_SYMBOL(tcp_ioctl);
 641
 642 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 643 {
 644         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 645         tp->pushed_seq = tp->write_seq;
 646 }
 647
 648 static inline bool forced_push(const struct tcp_sock *tp)
 649 {
 650         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 651 }
 652
 653 static void skb_entail(struct sock *sk, struct sk_buff *skb)
 654 {
 655         struct tcp_sock *tp = tcp_sk(sk);
 656         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 657
 658         skb->csum    = 0;
 659         tcb->seq     = tcb->end_seq = tp->write_seq;
 660         tcb->tcp_flags = TCPHDR_ACK;
 661         tcb->sacked  = 0;
 662         __skb_header_release(skb);
 663         tcp_add_write_queue_tail(sk, skb);
 664         sk->sk_wmem_queued += skb->truesize;
 665         sk_mem_charge(sk, skb->truesize);
 666         if (tp->nonagle & TCP_NAGLE_PUSH)
 667                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 668
 669         tcp_slow_start_after_idle_check(sk);
 670 }
 671
 672 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
 673 {
 674         if (flags & MSG_OOB)
 675                 tp->snd_up = tp->write_seq;
 676 }
 677
 678 /* If a not yet filled skb is pushed, do not send it if
 679  * we have data packets in Qdisc or NIC queues :
 680  * Because TX completion will happen shortly, it gives a chance
 681  * to coalesce future sendmsg() payload into this skb, without
 682  * need for a timer, and with no latency trade off.
 683  * As packets containing data payload have a bigger truesize
 684  * than pure acks (dataless) packets, the last checks prevent
 685  * autocorking if we only have an ACK in Qdisc/NIC queues,
 686  * or if TX completion was delayed after we processed ACK packet.
 687  */
 688 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
 689                                 int size_goal)
 690 {
 691         return skb->len < size_goal &&
 692                sock_net(sk)->ipv4.sysctl_tcp_autocorking &&
 693                skb != tcp_write_queue_head(sk) &&
 694                refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
 695 }
 696
 697 static void tcp_push(struct sock *sk, int flags, int mss_now,
 698                      int nonagle, int size_goal)
 699 {
 700         struct tcp_sock *tp = tcp_sk(sk);
 701         struct sk_buff *skb;
 702
 703         skb = tcp_write_queue_tail(sk);
 704         if (!skb)
 705                 return;
 706         if (!(flags & MSG_MORE) || forced_push(tp))
 707                 tcp_mark_push(tp, skb);
 708
 709         tcp_mark_urg(tp, flags);
 710
 711         if (tcp_should_autocork(sk, skb, size_goal)) {
 712
 713                 /* avoid atomic op if TSQ_THROTTLED bit is already set */
 714                 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
 715                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
 716                         set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
 717                 }
 718                 /* It is possible TX completion already happened
 719                  * before we set TSQ_THROTTLED.
 720                  */
 721                 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
 722                         return;
 723         }
 724
 725         if (flags & MSG_MORE)
 726                 nonagle = TCP_NAGLE_CORK;
 727
 728         __tcp_push_pending_frames(sk, mss_now, nonagle);
 729 }
 730
 731 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 732                                 unsigned int offset, size_t len)
 733 {
 734         struct tcp_splice_state *tss = rd_desc->arg.data;
 735         int ret;
 736
 737         ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
 738                               min(rd_desc->count, len), tss->flags);
 739         if (ret > 0)
 740                 rd_desc->count -= ret;
 741         return ret;
 742 }
 743
 744 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 745 {
 746         /* Store TCP splice context information in read_descriptor_t. */
 747         read_descriptor_t rd_desc = {
 748                 .arg.data = tss,
 749                 .count    = tss->len,
 750         };
 751
 752         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 753 }
 754
 755 /**
 756  *  tcp_splice_read - splice data from TCP socket to a pipe
 757  * @sock:       socket to splice from
 758  * @ppos:       position (not valid)
 759  * @pipe:       pipe to splice to
 760  * @len:        number of bytes to splice
 761  * @flags:      splice modifier flags
 762  *
 763  * Description:
 764  *    Will read pages from given socket and fill them into a pipe.
 765  *
 766  **/
 767 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 768                         struct pipe_inode_info *pipe, size_t len,
 769                         unsigned int flags)
 770 {
 771         struct sock *sk = sock->sk;
 772         struct tcp_splice_state tss = {
 773                 .pipe = pipe,
 774                 .len = len,
 775                 .flags = flags,
 776         };
 777         long timeo;
 778         ssize_t spliced;
 779         int ret;
 780
 781         sock_rps_record_flow(sk);
 782         /*
 783          * We can't seek on a socket input
 784          */
 785         if (unlikely(*ppos))
 786                 return -ESPIPE;
 787
 788         ret = spliced = 0;
 789
 790         lock_sock(sk);
 791
 792         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 793         while (tss.len) {
 794                 ret = __tcp_splice_read(sk, &tss);
 795                 if (ret < 0)
 796                         break;
 797                 else if (!ret) {
 798                         if (spliced)
 799                                 break;
 800                         if (sock_flag(sk, SOCK_DONE))
 801                                 break;
 802                         if (sk->sk_err) {
 803                                 ret = sock_error(sk);
 804                                 break;
 805                         }
 806                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 807                                 break;
 808                         if (sk->sk_state == TCP_CLOSE) {
 809                                 /*
 810                                  * This occurs when user tries to read
 811                                  * from never connected socket.
 812                                  */
 813                                 if (!sock_flag(sk, SOCK_DONE))
 814                                         ret = -ENOTCONN;
 815                                 break;
 816                         }
 817                         if (!timeo) {
 818                                 ret = -EAGAIN;
 819                                 break;
 820                         }
 821                         /* if __tcp_splice_read() got nothing while we have
 822                          * an skb in receive queue, we do not want to loop.
 823                          * This might happen with URG data.
 824                          */
 825                         if (!skb_queue_empty(&sk->sk_receive_queue))
 826                                 break;
 827                         sk_wait_data(sk, &timeo, NULL);
 828                         if (signal_pending(current)) {
 829                                 ret = sock_intr_errno(timeo);
 830                                 break;
 831                         }
 832                         continue;
 833                 }
 834                 tss.len -= ret;
 835                 spliced += ret;
 836
 837                 if (!timeo)
 838                         break;
 839                 release_sock(sk);
 840                 lock_sock(sk);
 841
 842                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 843                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 844                     signal_pending(current))
 845                         break;
 846         }
 847
 848         release_sock(sk);
 849
 850         if (spliced)
 851                 return spliced;
 852
 853         return ret;
 854 }
 855 EXPORT_SYMBOL(tcp_splice_read);
 856
 857 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
 858                                     bool force_schedule)
 859 {
 860         struct sk_buff *skb;
 861
 862         /* The TCP header must be at least 32-bit aligned.  */
 863         size = ALIGN(size, 4);
 864
 865         if (unlikely(tcp_under_memory_pressure(sk)))
 866                 sk_mem_reclaim_partial(sk);
 867
 868         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 869         if (likely(skb)) {
 870                 bool mem_scheduled;
 871
 872                 if (force_schedule) {
 873                         mem_scheduled = true;
 874                         sk_forced_mem_schedule(sk, skb->truesize);
 875                 } else {
 876                         mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
 877                 }
 878                 if (likely(mem_scheduled)) {
 879                         skb_reserve(skb, sk->sk_prot->max_header);
 880                         /*
 881                          * Make sure that we have exactly size bytes
 882                          * available to the caller, no more, no less.
 883                          */
 884                         skb->reserved_tailroom = skb->end - skb->tail - size;
 885                         INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
 886                         return skb;
 887                 }
 888                 __kfree_skb(skb);
 889         } else {
 890                 sk->sk_prot->enter_memory_pressure(sk);
 891                 sk_stream_moderate_sndbuf(sk);
 892         }
 893         return NULL;
 894 }
 895
 896 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 897                                        int large_allowed)
 898 {
 899         struct tcp_sock *tp = tcp_sk(sk);
 900         u32 new_size_goal, size_goal;
 901
 902         if (!large_allowed || !sk_can_gso(sk))
 903                 return mss_now;
 904
 905         /* Note : tcp_tso_autosize() will eventually split this later */
 906         new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
 907         new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
 908
 909         /* We try hard to avoid divides here */
 910         size_goal = tp->gso_segs * mss_now;
 911         if (unlikely(new_size_goal < size_goal ||
 912                      new_size_goal >= size_goal + mss_now)) {
 913                 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
 914                                      sk->sk_gso_max_segs);
 915                 size_goal = tp->gso_segs * mss_now;
 916         }
 917
 918         return max(size_goal, mss_now);
 919 }
 920
 921 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 922 {
 923         int mss_now;
 924
 925         mss_now = tcp_current_mss(sk);
 926         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 927
 928         return mss_now;
 929 }
 930
 931 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
 932                          size_t size, int flags)
 933 {
 934         struct tcp_sock *tp = tcp_sk(sk);
 935         int mss_now, size_goal;
 936         int err;
 937         ssize_t copied;
 938         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 939
 940         /* Wait for a connection to finish. One exception is TCP Fast Open
 941          * (passive side) where data is allowed to be sent before a connection
 942          * is fully established.
 943          */
 944         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
 945             !tcp_passive_fastopen(sk)) {
 946                 err = sk_stream_wait_connect(sk, &timeo);
 947                 if (err != 0)
 948                         goto out_err;
 949         }
 950
 951         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 952
 953         mss_now = tcp_send_mss(sk, &size_goal, flags);
 954         copied = 0;
 955
 956         err = -EPIPE;
 957         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 958                 goto out_err;
 959
 960         while (size > 0) {
 961                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 962                 int copy, i;
 963                 bool can_coalesce;
 964
 965                 if (!skb || (copy = size_goal - skb->len) <= 0 ||
 966                     !tcp_skb_can_collapse_to(skb)) {
 967 new_segment:
 968                         if (!sk_stream_memory_free(sk))
 969                                 goto wait_for_sndbuf;
 970
 971                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
 972                                         tcp_rtx_and_write_queues_empty(sk));
 973                         if (!skb)
 974                                 goto wait_for_memory;
 975
 976                         skb_entail(sk, skb);
 977                         copy = size_goal;
 978                 }
 979
 980                 if (copy > size)
 981                         copy = size;
 982
 983                 i = skb_shinfo(skb)->nr_frags;
 984                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 985                 if (!can_coalesce && i >= sysctl_max_skb_frags) {
 986                         tcp_mark_push(tp, skb);
 987                         goto new_segment;
 988                 }
 989                 if (!sk_wmem_schedule(sk, copy))
 990                         goto wait_for_memory;
 991
 992                 if (can_coalesce) {
 993                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
 994                 } else {
 995                         get_page(page);
 996                         skb_fill_page_desc(skb, i, page, offset, copy);
 997                 }
 998                 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
 999
1000                 skb->len += copy;
1001                 skb->data_len += copy;
1002                 skb->truesize += copy;
1003                 sk->sk_wmem_queued += copy;
1004                 sk_mem_charge(sk, copy);
1005                 skb->ip_summed = CHECKSUM_PARTIAL;
1006                 tp->write_seq += copy;
1007                 TCP_SKB_CB(skb)->end_seq += copy;
1008                 tcp_skb_pcount_set(skb, 0);
1009
1010                 if (!copied)
1011                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1012
1013                 copied += copy;
1014                 offset += copy;
1015                 size -= copy;
1016                 if (!size)
1017                         goto out;
1018
1019                 if (skb->len < size_goal || (flags & MSG_OOB))
1020                         continue;
1021
1022                 if (forced_push(tp)) {
1023                         tcp_mark_push(tp, skb);
1024                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1025                 } else if (skb == tcp_send_head(sk))
1026                         tcp_push_one(sk, mss_now);
1027                 continue;
1028
1029 wait_for_sndbuf:
1030                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1031 wait_for_memory:
1032                 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1033                          TCP_NAGLE_PUSH, size_goal);
1034
1035                 err = sk_stream_wait_memory(sk, &timeo);
1036                 if (err != 0)
1037                         goto do_error;
1038
1039                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1040         }
1041
1042 out:
1043         if (copied) {
1044                 tcp_tx_timestamp(sk, sk->sk_tsflags);
1045                 if (!(flags & MSG_SENDPAGE_NOTLAST))
1046                         tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1047         }
1048         return copied;
1049
1050 do_error:
1051         if (copied)
1052                 goto out;
1053 out_err:
1054         /* make sure we wake any epoll edge trigger waiter */
1055         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1056                      err == -EAGAIN)) {
1057                 sk->sk_write_space(sk);
1058                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1059         }
1060         return sk_stream_error(sk, flags, err);
1061 }
1062 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1063
1064 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1065                         size_t size, int flags)
1066 {
1067         if (!(sk->sk_route_caps & NETIF_F_SG) ||
1068             !sk_check_csum_caps(sk))
1069                 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1070
1071         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1072
1073         return do_tcp_sendpages(sk, page, offset, size, flags);
1074 }
1075 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1076
1077 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1078                  size_t size, int flags)
1079 {
1080         int ret;
1081
1082         lock_sock(sk);
1083         ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1084         release_sock(sk);
1085
1086         return ret;
1087 }
1088 EXPORT_SYMBOL(tcp_sendpage);
1089
1090 /* Do not bother using a page frag for very small frames.
1091  * But use this heuristic only for the first skb in write queue.
1092  *
1093  * Having no payload in skb->head allows better SACK shifting
1094  * in tcp_shift_skb_data(), reducing sack/rack overhead, because
1095  * write queue has less skbs.
1096  * Each skb can hold up to MAX_SKB_FRAGS * 32Kbytes, or ~0.5 MB.
1097  * This also speeds up tso_fragment(), since it wont fallback
1098  * to tcp_fragment().
1099  */
1100 static int linear_payload_sz(bool first_skb)
1101 {
1102         if (first_skb)
1103                 return SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1104         return 0;
1105 }
1106
1107 static int select_size(const struct sock *sk, bool sg, bool first_skb)
1108 {
1109         const struct tcp_sock *tp = tcp_sk(sk);
1110         int tmp = tp->mss_cache;
1111
1112         if (sg) {
1113                 if (sk_can_gso(sk)) {
1114                         tmp = linear_payload_sz(first_skb);
1115                 } else {
1116                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1117
1118                         if (tmp >= pgbreak &&
1119                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1120                                 tmp = pgbreak;
1121                 }
1122         }
1123
1124         return tmp;
1125 }
1126
1127 void tcp_free_fastopen_req(struct tcp_sock *tp)
1128 {
1129         if (tp->fastopen_req) {
1130                 kfree(tp->fastopen_req);
1131                 tp->fastopen_req = NULL;
1132         }
1133 }
1134
1135 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1136                                 int *copied, size_t size)
1137 {
1138         struct tcp_sock *tp = tcp_sk(sk);
1139         struct inet_sock *inet = inet_sk(sk);
1140         struct sockaddr *uaddr = msg->msg_name;
1141         int err, flags;
1142
1143         if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1144             (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1145              uaddr->sa_family == AF_UNSPEC))
1146                 return -EOPNOTSUPP;
1147         if (tp->fastopen_req)
1148                 return -EALREADY; /* Another Fast Open is in progress */
1149
1150         tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1151                                    sk->sk_allocation);
1152         if (unlikely(!tp->fastopen_req))
1153                 return -ENOBUFS;
1154         tp->fastopen_req->data = msg;
1155         tp->fastopen_req->size = size;
1156
1157         if (inet->defer_connect) {
1158                 err = tcp_connect(sk);
1159                 /* Same failure procedure as in tcp_v4/6_connect */
1160                 if (err) {
1161                         tcp_set_state(sk, TCP_CLOSE);
1162                         inet->inet_dport = 0;
1163                         sk->sk_route_caps = 0;
1164                 }
1165         }
1166         flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1167         err = __inet_stream_connect(sk->sk_socket, uaddr,
1168                                     msg->msg_namelen, flags, 1);
1169         /* fastopen_req could already be freed in __inet_stream_connect
1170          * if the connection times out or gets rst
1171          */
1172         if (tp->fastopen_req) {
1173                 *copied = tp->fastopen_req->copied;
1174                 tcp_free_fastopen_req(tp);
1175                 inet->defer_connect = 0;
1176         }
1177         return err;
1178 }
1179
1180 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1181 {
1182         struct tcp_sock *tp = tcp_sk(sk);
1183         struct ubuf_info *uarg = NULL;
1184         struct sk_buff *skb;
1185         struct sockcm_cookie sockc;
1186         int flags, err, copied = 0;
1187         int mss_now = 0, size_goal, copied_syn = 0;
1188         bool process_backlog = false;
1189         bool sg;
1190         long timeo;
1191
1192         flags = msg->msg_flags;
1193
1194         if (flags & MSG_ZEROCOPY && size) {
1195                 if (sk->sk_state != TCP_ESTABLISHED) {
1196                         err = -EINVAL;
1197                         goto out_err;
1198                 }
1199
1200                 skb = tcp_write_queue_tail(sk);
1201                 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
1202                 if (!uarg) {
1203                         err = -ENOBUFS;
1204                         goto out_err;
1205                 }
1206
1207                 if (!(sk_check_csum_caps(sk) && sk->sk_route_caps & NETIF_F_SG))
1208                         uarg->zerocopy = 0;
1209         }
1210
1211         if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect)) {
1212                 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1213                 if (err == -EINPROGRESS && copied_syn > 0)
1214                         goto out;
1215                 else if (err)
1216                         goto out_err;
1217         }
1218
1219         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1220
1221         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1222
1223         /* Wait for a connection to finish. One exception is TCP Fast Open
1224          * (passive side) where data is allowed to be sent before a connection
1225          * is fully established.
1226          */
1227         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1228             !tcp_passive_fastopen(sk)) {
1229                 err = sk_stream_wait_connect(sk, &timeo);
1230                 if (err != 0)
1231                         goto do_error;
1232         }
1233
1234         if (unlikely(tp->repair)) {
1235                 if (tp->repair_queue == TCP_RECV_QUEUE) {
1236                         copied = tcp_send_rcvq(sk, msg, size);
1237                         goto out_nopush;
1238                 }
1239
1240                 err = -EINVAL;
1241                 if (tp->repair_queue == TCP_NO_QUEUE)
1242                         goto out_err;
1243
1244                 /* 'common' sending to sendq */
1245         }
1246
1247         sockc.tsflags = sk->sk_tsflags;
1248         if (msg->msg_controllen) {
1249                 err = sock_cmsg_send(sk, msg, &sockc);
1250                 if (unlikely(err)) {
1251                         err = -EINVAL;
1252                         goto out_err;
1253                 }
1254         }
1255
1256         /* This should be in poll */
1257         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1258
1259         /* Ok commence sending. */
1260         copied = 0;
1261
1262 restart:
1263         mss_now = tcp_send_mss(sk, &size_goal, flags);
1264
1265         err = -EPIPE;
1266         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1267                 goto do_error;
1268
1269         sg = !!(sk->sk_route_caps & NETIF_F_SG);
1270
1271         while (msg_data_left(msg)) {
1272                 int copy = 0;
1273                 int max = size_goal;
1274
1275                 skb = tcp_write_queue_tail(sk);
1276                 if (skb) {
1277                         if (skb->ip_summed == CHECKSUM_NONE)
1278                                 max = mss_now;
1279                         copy = max - skb->len;
1280                 }
1281
1282                 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1283                         bool first_skb;
1284
1285 new_segment:
1286                         /* Allocate new segment. If the interface is SG,
1287                          * allocate skb fitting to single page.
1288                          */
1289                         if (!sk_stream_memory_free(sk))
1290                                 goto wait_for_sndbuf;
1291
1292                         if (process_backlog && sk_flush_backlog(sk)) {
1293                                 process_backlog = false;
1294                                 goto restart;
1295                         }
1296                         first_skb = tcp_rtx_and_write_queues_empty(sk);
1297                         skb = sk_stream_alloc_skb(sk,
1298                                                   select_size(sk, sg, first_skb),
1299                                                   sk->sk_allocation,
1300                                                   first_skb);
1301                         if (!skb)
1302                                 goto wait_for_memory;
1303
1304                         process_backlog = true;
1305                         /*
1306                          * Check whether we can use HW checksum.
1307                          */
1308                         if (sk_check_csum_caps(sk))
1309                                 skb->ip_summed = CHECKSUM_PARTIAL;
1310
1311                         skb_entail(sk, skb);
1312                         copy = size_goal;
1313                         max = size_goal;
1314
1315                         /* All packets are restored as if they have
1316                          * already been sent. skb_mstamp isn't set to
1317                          * avoid wrong rtt estimation.
1318                          */
1319                         if (tp->repair)
1320                                 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1321                 }
1322
1323                 /* Try to append data to the end of skb. */
1324                 if (copy > msg_data_left(msg))
1325                         copy = msg_data_left(msg);
1326
1327                 /* Where to copy to? */
1328                 if (skb_availroom(skb) > 0) {
1329                         /* We have some space in skb head. Superb! */
1330                         copy = min_t(int, copy, skb_availroom(skb));
1331                         err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1332                         if (err)
1333                                 goto do_fault;
1334                 } else if (!uarg || !uarg->zerocopy) {
1335                         bool merge = true;
1336                         int i = skb_shinfo(skb)->nr_frags;
1337                         struct page_frag *pfrag = sk_page_frag(sk);
1338
1339                         if (!sk_page_frag_refill(sk, pfrag))
1340                                 goto wait_for_memory;
1341
1342                         if (!skb_can_coalesce(skb, i, pfrag->page,
1343                                               pfrag->offset)) {
1344                                 if (i >= sysctl_max_skb_frags || !sg) {
1345                                         tcp_mark_push(tp, skb);
1346                                         goto new_segment;
1347                                 }
1348                                 merge = false;
1349                         }
1350
1351                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1352
1353                         if (!sk_wmem_schedule(sk, copy))
1354                                 goto wait_for_memory;
1355
1356                         err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1357                                                        pfrag->page,
1358                                                        pfrag->offset,
1359                                                        copy);
1360                         if (err)
1361                                 goto do_error;
1362
1363                         /* Update the skb. */
1364                         if (merge) {
1365                                 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1366                         } else {
1367                                 skb_fill_page_desc(skb, i, pfrag->page,
1368                                                    pfrag->offset, copy);
1369                                 page_ref_inc(pfrag->page);
1370                         }
1371                         pfrag->offset += copy;
1372                 } else {
1373                         err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1374                         if (err == -EMSGSIZE || err == -EEXIST) {
1375                                 tcp_mark_push(tp, skb);
1376                                 goto new_segment;
1377                         }
1378                         if (err < 0)
1379                                 goto do_error;
1380                         copy = err;
1381                 }
1382
1383                 if (!copied)
1384                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1385
1386                 tp->write_seq += copy;
1387                 TCP_SKB_CB(skb)->end_seq += copy;
1388                 tcp_skb_pcount_set(skb, 0);
1389
1390                 copied += copy;
1391                 if (!msg_data_left(msg)) {
1392                         if (unlikely(flags & MSG_EOR))
1393                                 TCP_SKB_CB(skb)->eor = 1;
1394                         goto out;
1395                 }
1396
1397                 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1398                         continue;
1399
1400                 if (forced_push(tp)) {
1401                         tcp_mark_push(tp, skb);
1402                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1403                 } else if (skb == tcp_send_head(sk))
1404                         tcp_push_one(sk, mss_now);
1405                 continue;
1406
1407 wait_for_sndbuf:
1408                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1409 wait_for_memory:
1410                 if (copied)
1411                         tcp_push(sk, flags & ~MSG_MORE, mss_now,
1412                                  TCP_NAGLE_PUSH, size_goal);
1413
1414                 err = sk_stream_wait_memory(sk, &timeo);
1415                 if (err != 0)
1416                         goto do_error;
1417
1418                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1419         }
1420
1421 out:
1422         if (copied) {
1423                 tcp_tx_timestamp(sk, sockc.tsflags);
1424                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1425         }
1426 out_nopush:
1427         sock_zerocopy_put(uarg);
1428         return copied + copied_syn;
1429
1430 do_fault:
1431         if (!skb->len) {
1432                 tcp_unlink_write_queue(skb, sk);
1433                 /* It is the one place in all of TCP, except connection
1434                  * reset, where we can be unlinking the send_head.
1435                  */
1436                 tcp_check_send_head(sk, skb);
1437                 sk_wmem_free_skb(sk, skb);
1438         }
1439
1440 do_error:
1441         if (copied + copied_syn)
1442                 goto out;
1443 out_err:
1444         sock_zerocopy_put_abort(uarg);
1445         err = sk_stream_error(sk, flags, err);
1446         /* make sure we wake any epoll edge trigger waiter */
1447         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1448                      err == -EAGAIN)) {
1449                 sk->sk_write_space(sk);
1450                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1451         }
1452         return err;
1453 }
1454 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1455
1456 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1457 {
1458         int ret;
1459
1460         lock_sock(sk);
1461         ret = tcp_sendmsg_locked(sk, msg, size);
1462         release_sock(sk);
1463
1464         return ret;
1465 }
1466 EXPORT_SYMBOL(tcp_sendmsg);
1467
1468 /*
1469  *      Handle reading urgent data. BSD has very simple semantics for
1470  *      this, no blocking and very strange errors 8)
1471  */
1472
1473 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1474 {
1475         struct tcp_sock *tp = tcp_sk(sk);
1476
1477         /* No URG data to read. */
1478         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1479             tp->urg_data == TCP_URG_READ)
1480                 return -EINVAL; /* Yes this is right ! */
1481
1482         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1483                 return -ENOTCONN;
1484
1485         if (tp->urg_data & TCP_URG_VALID) {
1486                 int err = 0;
1487                 char c = tp->urg_data;
1488
1489                 if (!(flags & MSG_PEEK))
1490                         tp->urg_data = TCP_URG_READ;
1491
1492                 /* Read urgent data. */
1493                 msg->msg_flags |= MSG_OOB;
1494
1495                 if (len > 0) {
1496                         if (!(flags & MSG_TRUNC))
1497                                 err = memcpy_to_msg(msg, &c, 1);
1498                         len = 1;
1499                 } else
1500                         msg->msg_flags |= MSG_TRUNC;
1501
1502                 return err ? -EFAULT : len;
1503         }
1504
1505         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1506                 return 0;
1507
1508         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1509          * the available implementations agree in this case:
1510          * this call should never block, independent of the
1511          * blocking state of the socket.
1512          * Mike <pall@rz.uni-karlsruhe.de>
1513          */
1514         return -EAGAIN;
1515 }
1516
1517 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1518 {
1519         struct sk_buff *skb;
1520         int copied = 0, err = 0;
1521
1522         /* XXX -- need to support SO_PEEK_OFF */
1523
1524         skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1525                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1526                 if (err)
1527                         return err;
1528                 copied += skb->len;
1529         }
1530
1531         skb_queue_walk(&sk->sk_write_queue, skb) {
1532                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1533                 if (err)
1534                         break;
1535
1536                 copied += skb->len;
1537         }
1538
1539         return err ?: copied;
1540 }
1541
1542 /* Clean up the receive buffer for full frames taken by the user,
1543  * then send an ACK if necessary.  COPIED is the number of bytes
1544  * tcp_recvmsg has given to the user so far, it speeds up the
1545  * calculation of whether or not we must ACK for the sake of
1546  * a window update.
1547  */
1548 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1549 {
1550         struct tcp_sock *tp = tcp_sk(sk);
1551         bool time_to_ack = false;
1552
1553         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1554
1555         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1556              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1557              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1558
1559         if (inet_csk_ack_scheduled(sk)) {
1560                 const struct inet_connection_sock *icsk = inet_csk(sk);
1561                    /* Delayed ACKs frequently hit locked sockets during bulk
1562                     * receive. */
1563                 if (icsk->icsk_ack.blocked ||
1564                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1565                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1566                     /*
1567                      * If this read emptied read buffer, we send ACK, if
1568                      * connection is not bidirectional, user drained
1569                      * receive buffer and there was a small segment
1570                      * in queue.
1571                      */
1572                     (copied > 0 &&
1573                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1574                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1575                        !icsk->icsk_ack.pingpong)) &&
1576                       !atomic_read(&sk->sk_rmem_alloc)))
1577                         time_to_ack = true;
1578         }
1579
1580         /* We send an ACK if we can now advertise a non-zero window
1581          * which has been raised "significantly".
1582          *
1583          * Even if window raised up to infinity, do not send window open ACK
1584          * in states, where we will not receive more. It is useless.
1585          */
1586         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1587                 __u32 rcv_window_now = tcp_receive_window(tp);
1588
1589                 /* Optimize, __tcp_select_window() is not cheap. */
1590                 if (2*rcv_window_now <= tp->window_clamp) {
1591                         __u32 new_window = __tcp_select_window(sk);
1592
1593                         /* Send ACK now, if this read freed lots of space
1594                          * in our buffer. Certainly, new_window is new window.
1595                          * We can advertise it now, if it is not less than current one.
1596                          * "Lots" means "at least twice" here.
1597                          */
1598                         if (new_window && new_window >= 2 * rcv_window_now)
1599                                 time_to_ack = true;
1600                 }
1601         }
1602         if (time_to_ack)
1603                 tcp_send_ack(sk);
1604 }
1605
1606 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1607 {
1608         struct sk_buff *skb;
1609         u32 offset;
1610
1611         while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1612                 offset = seq - TCP_SKB_CB(skb)->seq;
1613                 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1614                         pr_err_once("%s: found a SYN, please report !\n", __func__);
1615                         offset--;
1616                 }
1617                 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1618                         *off = offset;
1619                         return skb;
1620                 }
1621                 /* This looks weird, but this can happen if TCP collapsing
1622                  * splitted a fat GRO packet, while we released socket lock
1623                  * in skb_splice_bits()
1624                  */
1625                 sk_eat_skb(sk, skb);
1626         }
1627         return NULL;
1628 }
1629
1630 /*
1631  * This routine provides an alternative to tcp_recvmsg() for routines
1632  * that would like to handle copying from skbuffs directly in 'sendfile'
1633  * fashion.
1634  * Note:
1635  *      - It is assumed that the socket was locked by the caller.
1636  *      - The routine does not block.
1637  *      - At present, there is no support for reading OOB data
1638  *        or for 'peeking' the socket using this routine
1639  *        (although both would be easy to implement).
1640  */
1641 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1642                   sk_read_actor_t recv_actor)
1643 {
1644         struct sk_buff *skb;
1645         struct tcp_sock *tp = tcp_sk(sk);
1646         u32 seq = tp->copied_seq;
1647         u32 offset;
1648         int copied = 0;
1649
1650         if (sk->sk_state == TCP_LISTEN)
1651                 return -ENOTCONN;
1652         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1653                 if (offset < skb->len) {
1654                         int used;
1655                         size_t len;
1656
1657                         len = skb->len - offset;
1658                         /* Stop reading if we hit a patch of urgent data */
1659                         if (tp->urg_data) {
1660                                 u32 urg_offset = tp->urg_seq - seq;
1661                                 if (urg_offset < len)
1662                                         len = urg_offset;
1663                                 if (!len)
1664                                         break;
1665                         }
1666                         used = recv_actor(desc, skb, offset, len);
1667                         if (used <= 0) {
1668                                 if (!copied)
1669                                         copied = used;
1670                                 break;
1671                         } else if (used <= len) {
1672                                 seq += used;
1673                                 copied += used;
1674                                 offset += used;
1675                         }
1676                         /* If recv_actor drops the lock (e.g. TCP splice
1677                          * receive) the skb pointer might be invalid when
1678                          * getting here: tcp_collapse might have deleted it
1679                          * while aggregating skbs from the socket queue.
1680                          */
1681                         skb = tcp_recv_skb(sk, seq - 1, &offset);
1682                         if (!skb)
1683                                 break;
1684                         /* TCP coalescing might have appended data to the skb.
1685                          * Try to splice more frags
1686                          */
1687                         if (offset + 1 != skb->len)
1688                                 continue;
1689                 }
1690                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1691                         sk_eat_skb(sk, skb);
1692                         ++seq;
1693                         break;
1694                 }
1695                 sk_eat_skb(sk, skb);
1696                 if (!desc->count)
1697                         break;
1698                 tp->copied_seq = seq;
1699         }
1700         tp->copied_seq = seq;
1701
1702         tcp_rcv_space_adjust(sk);
1703
1704         /* Clean up data we have read: This will do ACK frames. */
1705         if (copied > 0) {
1706                 tcp_recv_skb(sk, seq, &offset);
1707                 tcp_cleanup_rbuf(sk, copied);
1708         }
1709         return copied;
1710 }
1711 EXPORT_SYMBOL(tcp_read_sock);
1712
1713 int tcp_peek_len(struct socket *sock)
1714 {
1715         return tcp_inq(sock->sk);
1716 }
1717 EXPORT_SYMBOL(tcp_peek_len);
1718
1719 static void tcp_update_recv_tstamps(struct sk_buff *skb,
1720                                     struct scm_timestamping *tss)
1721 {
1722         if (skb->tstamp)
1723                 tss->ts[0] = ktime_to_timespec(skb->tstamp);
1724         else
1725                 tss->ts[0] = (struct timespec) {0};
1726
1727         if (skb_hwtstamps(skb)->hwtstamp)
1728                 tss->ts[2] = ktime_to_timespec(skb_hwtstamps(skb)->hwtstamp);
1729         else
1730                 tss->ts[2] = (struct timespec) {0};
1731 }
1732
1733 /* Similar to __sock_recv_timestamp, but does not require an skb */
1734 void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
1735                         struct scm_timestamping *tss)
1736 {
1737         struct timeval tv;
1738         bool has_timestamping = false;
1739
1740         if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
1741                 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
1742                         if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
1743                                 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPNS,
1744                                          sizeof(tss->ts[0]), &tss->ts[0]);
1745                         } else {
1746                                 tv.tv_sec = tss->ts[0].tv_sec;
1747                                 tv.tv_usec = tss->ts[0].tv_nsec / 1000;
1748
1749                                 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMP,
1750                                          sizeof(tv), &tv);
1751                         }
1752                 }
1753
1754                 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
1755                         has_timestamping = true;
1756                 else
1757                         tss->ts[0] = (struct timespec) {0};
1758         }
1759
1760         if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
1761                 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
1762                         has_timestamping = true;
1763                 else
1764                         tss->ts[2] = (struct timespec) {0};
1765         }
1766
1767         if (has_timestamping) {
1768                 tss->ts[1] = (struct timespec) {0};
1769                 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPING,
1770                          sizeof(*tss), tss);
1771         }
1772 }
1773
1774 /*
1775  *      This routine copies from a sock struct into the user buffer.
1776  *
1777  *      Technical note: in 2.3 we work on _locked_ socket, so that
1778  *      tricks with *seq access order and skb->users are not required.
1779  *      Probably, code can be easily improved even more.
1780  */
1781
1782 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1783                 int flags, int *addr_len)
1784 {
1785         struct tcp_sock *tp = tcp_sk(sk);
1786         int copied = 0;
1787         u32 peek_seq;
1788         u32 *seq;
1789         unsigned long used;
1790         int err;
1791         int target;             /* Read at least this many bytes */
1792         long timeo;
1793         struct sk_buff *skb, *last;
1794         u32 urg_hole = 0;
1795         struct scm_timestamping tss;
1796         bool has_tss = false;
1797
1798         if (unlikely(flags & MSG_ERRQUEUE))
1799                 return inet_recv_error(sk, msg, len, addr_len);
1800
1801         if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1802             (sk->sk_state == TCP_ESTABLISHED))
1803                 sk_busy_loop(sk, nonblock);
1804
1805         lock_sock(sk);
1806
1807         err = -ENOTCONN;
1808         if (sk->sk_state == TCP_LISTEN)
1809                 goto out;
1810
1811         timeo = sock_rcvtimeo(sk, nonblock);
1812
1813         /* Urgent data needs to be handled specially. */
1814         if (flags & MSG_OOB)
1815                 goto recv_urg;
1816
1817         if (unlikely(tp->repair)) {
1818                 err = -EPERM;
1819                 if (!(flags & MSG_PEEK))
1820                         goto out;
1821
1822                 if (tp->repair_queue == TCP_SEND_QUEUE)
1823                         goto recv_sndq;
1824
1825                 err = -EINVAL;
1826                 if (tp->repair_queue == TCP_NO_QUEUE)
1827                         goto out;
1828
1829                 /* 'common' recv queue MSG_PEEK-ing */
1830         }
1831
1832         seq = &tp->copied_seq;
1833         if (flags & MSG_PEEK) {
1834                 peek_seq = tp->copied_seq;
1835                 seq = &peek_seq;
1836         }
1837
1838         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1839
1840         do {
1841                 u32 offset;
1842
1843                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1844                 if (tp->urg_data && tp->urg_seq == *seq) {
1845                         if (copied)
1846                                 break;
1847                         if (signal_pending(current)) {
1848                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1849                                 break;
1850                         }
1851                 }
1852
1853                 /* Next get a buffer. */
1854
1855                 last = skb_peek_tail(&sk->sk_receive_queue);
1856                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1857                         last = skb;
1858                         /* Now that we have two receive queues this
1859                          * shouldn't happen.
1860                          */
1861                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1862                                  "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1863                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1864                                  flags))
1865                                 break;
1866
1867                         offset = *seq - TCP_SKB_CB(skb)->seq;
1868                         if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1869                                 pr_err_once("%s: found a SYN, please report !\n", __func__);
1870                                 offset--;
1871                         }
1872                         if (offset < skb->len)
1873                                 goto found_ok_skb;
1874                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1875                                 goto found_fin_ok;
1876                         WARN(!(flags & MSG_PEEK),
1877                              "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1878                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1879                 }
1880
1881                 /* Well, if we have backlog, try to process it now yet. */
1882
1883                 if (copied >= target && !sk->sk_backlog.tail)
1884                         break;
1885
1886                 if (copied) {
1887                         if (sk->sk_err ||
1888                             sk->sk_state == TCP_CLOSE ||
1889                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1890                             !timeo ||
1891                             signal_pending(current))
1892                                 break;
1893                 } else {
1894                         if (sock_flag(sk, SOCK_DONE))
1895                                 break;
1896
1897                         if (sk->sk_err) {
1898                                 copied = sock_error(sk);
1899                                 break;
1900                         }
1901
1902                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1903                                 break;
1904
1905                         if (sk->sk_state == TCP_CLOSE) {
1906                                 if (!sock_flag(sk, SOCK_DONE)) {
1907                                         /* This occurs when user tries to read
1908                                          * from never connected socket.
1909                                          */
1910                                         copied = -ENOTCONN;
1911                                         break;
1912                                 }
1913                                 break;
1914                         }
1915
1916                         if (!timeo) {
1917                                 copied = -EAGAIN;
1918                                 break;
1919                         }
1920
1921                         if (signal_pending(current)) {
1922                                 copied = sock_intr_errno(timeo);
1923                                 break;
1924                         }
1925                 }
1926
1927                 tcp_cleanup_rbuf(sk, copied);
1928
1929                 if (copied >= target) {
1930                         /* Do not sleep, just process backlog. */
1931                         release_sock(sk);
1932                         lock_sock(sk);
1933                 } else {
1934                         sk_wait_data(sk, &timeo, last);
1935                 }
1936
1937                 if ((flags & MSG_PEEK) &&
1938                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1939                         net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1940                                             current->comm,
1941                                             task_pid_nr(current));
1942                         peek_seq = tp->copied_seq;
1943                 }
1944                 continue;
1945
1946         found_ok_skb:
1947                 /* Ok so how much can we use? */
1948                 used = skb->len - offset;
1949                 if (len < used)
1950                         used = len;
1951
1952                 /* Do we have urgent data here? */
1953                 if (tp->urg_data) {
1954                         u32 urg_offset = tp->urg_seq - *seq;
1955                         if (urg_offset < used) {
1956                                 if (!urg_offset) {
1957                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1958                                                 ++*seq;
1959                                                 urg_hole++;
1960                                                 offset++;
1961                                                 used--;
1962                                                 if (!used)
1963                                                         goto skip_copy;
1964                                         }
1965                                 } else
1966                                         used = urg_offset;
1967                         }
1968                 }
1969
1970                 if (!(flags & MSG_TRUNC)) {
1971                         err = skb_copy_datagram_msg(skb, offset, msg, used);
1972                         if (err) {
1973                                 /* Exception. Bailout! */
1974                                 if (!copied)
1975                                         copied = -EFAULT;
1976                                 break;
1977                         }
1978                 }
1979
1980                 *seq += used;
1981                 copied += used;
1982                 len -= used;
1983
1984                 tcp_rcv_space_adjust(sk);
1985
1986 skip_copy:
1987                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1988                         tp->urg_data = 0;
1989                         tcp_fast_path_check(sk);
1990                 }
1991                 if (used + offset < skb->len)
1992                         continue;
1993
1994                 if (TCP_SKB_CB(skb)->has_rxtstamp) {
1995                         tcp_update_recv_tstamps(skb, &tss);
1996                         has_tss = true;
1997                 }
1998                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1999                         goto found_fin_ok;
2000                 if (!(flags & MSG_PEEK))
2001                         sk_eat_skb(sk, skb);
2002                 continue;
2003
2004         found_fin_ok:
2005                 /* Process the FIN. */
2006                 ++*seq;
2007                 if (!(flags & MSG_PEEK))
2008                         sk_eat_skb(sk, skb);
2009                 break;
2010         } while (len > 0);
2011
2012         /* According to UNIX98, msg_name/msg_namelen are ignored
2013          * on connected socket. I was just happy when found this 8) --ANK
2014          */
2015
2016         if (has_tss)
2017                 tcp_recv_timestamp(msg, sk, &tss);
2018
2019         /* Clean up data we have read: This will do ACK frames. */
2020         tcp_cleanup_rbuf(sk, copied);
2021
2022         release_sock(sk);
2023         return copied;
2024
2025 out:
2026         release_sock(sk);
2027         return err;
2028
2029 recv_urg:
2030         err = tcp_recv_urg(sk, msg, len, flags);
2031         goto out;
2032
2033 recv_sndq:
2034         err = tcp_peek_sndq(sk, msg, len);
2035         goto out;
2036 }
2037 EXPORT_SYMBOL(tcp_recvmsg);
2038
2039 void tcp_set_state(struct sock *sk, int state)
2040 {
2041         int oldstate = sk->sk_state;
2042
2043         switch (state) {
2044         case TCP_ESTABLISHED:
2045                 if (oldstate != TCP_ESTABLISHED)
2046                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2047                 break;
2048
2049         case TCP_CLOSE:
2050                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2051                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2052
2053                 sk->sk_prot->unhash(sk);
2054                 if (inet_csk(sk)->icsk_bind_hash &&
2055                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2056                         inet_put_port(sk);
2057                 /* fall through */
2058         default:
2059                 if (oldstate == TCP_ESTABLISHED)
2060                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2061         }
2062
2063         /* Change state AFTER socket is unhashed to avoid closed
2064          * socket sitting in hash tables.
2065          */
2066         inet_sk_state_store(sk, state);
2067
2068 #ifdef STATE_TRACE
2069         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
2070 #endif
2071 }
2072 EXPORT_SYMBOL_GPL(tcp_set_state);
2073
2074 /*
2075  *      State processing on a close. This implements the state shift for
2076  *      sending our FIN frame. Note that we only send a FIN for some
2077  *      states. A shutdown() may have already sent the FIN, or we may be
2078  *      closed.
2079  */
2080
2081 static const unsigned char new_state[16] = {
2082   /* current state:        new state:      action:      */
2083   [0 /* (Invalid) */]   = TCP_CLOSE,
2084   [TCP_ESTABLISHED]     = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2085   [TCP_SYN_SENT]        = TCP_CLOSE,
2086   [TCP_SYN_RECV]        = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2087   [TCP_FIN_WAIT1]       = TCP_FIN_WAIT1,
2088   [TCP_FIN_WAIT2]       = TCP_FIN_WAIT2,
2089   [TCP_TIME_WAIT]       = TCP_CLOSE,
2090   [TCP_CLOSE]           = TCP_CLOSE,
2091   [TCP_CLOSE_WAIT]      = TCP_LAST_ACK  | TCP_ACTION_FIN,
2092   [TCP_LAST_ACK]        = TCP_LAST_ACK,
2093   [TCP_LISTEN]          = TCP_CLOSE,
2094   [TCP_CLOSING]         = TCP_CLOSING,
2095   [TCP_NEW_SYN_RECV]    = TCP_CLOSE,    /* should not happen ! */
2096 };
2097
2098 static int tcp_close_state(struct sock *sk)
2099 {
2100         int next = (int)new_state[sk->sk_state];
2101         int ns = next & TCP_STATE_MASK;
2102
2103         tcp_set_state(sk, ns);
2104
2105         return next & TCP_ACTION_FIN;
2106 }
2107
2108 /*
2109  *      Shutdown the sending side of a connection. Much like close except
2110  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2111  */
2112
2113 void tcp_shutdown(struct sock *sk, int how)
2114 {
2115         /*      We need to grab some memory, and put together a FIN,
2116          *      and then put it into the queue to be sent.
2117          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2118          */
2119         if (!(how & SEND_SHUTDOWN))
2120                 return;
2121
2122         /* If we've already sent a FIN, or it's a closed state, skip this. */
2123         if ((1 << sk->sk_state) &
2124             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2125              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2126                 /* Clear out any half completed packets.  FIN if needed. */
2127                 if (tcp_close_state(sk))
2128                         tcp_send_fin(sk);
2129         }
2130 }
2131 EXPORT_SYMBOL(tcp_shutdown);
2132
2133 bool tcp_check_oom(struct sock *sk, int shift)
2134 {
2135         bool too_many_orphans, out_of_socket_memory;
2136
2137         too_many_orphans = tcp_too_many_orphans(sk, shift);
2138         out_of_socket_memory = tcp_out_of_memory(sk);
2139
2140         if (too_many_orphans)
2141                 net_info_ratelimited("too many orphaned sockets\n");
2142         if (out_of_socket_memory)
2143                 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2144         return too_many_orphans || out_of_socket_memory;
2145 }
2146
2147 void tcp_close(struct sock *sk, long timeout)
2148 {
2149         struct sk_buff *skb;
2150         int data_was_unread = 0;
2151         int state;
2152
2153         lock_sock(sk);
2154         sk->sk_shutdown = SHUTDOWN_MASK;
2155
2156         if (sk->sk_state == TCP_LISTEN) {
2157                 tcp_set_state(sk, TCP_CLOSE);
2158
2159                 /* Special case. */
2160                 inet_csk_listen_stop(sk);
2161
2162                 goto adjudge_to_death;
2163         }
2164
2165         /*  We need to flush the recv. buffs.  We do this only on the
2166          *  descriptor close, not protocol-sourced closes, because the
2167          *  reader process may not have drained the data yet!
2168          */
2169         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2170                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2171
2172                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2173                         len--;
2174                 data_was_unread += len;
2175                 __kfree_skb(skb);
2176         }
2177
2178         sk_mem_reclaim(sk);
2179
2180         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2181         if (sk->sk_state == TCP_CLOSE)
2182                 goto adjudge_to_death;
2183
2184         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2185          * data was lost. To witness the awful effects of the old behavior of
2186          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2187          * GET in an FTP client, suspend the process, wait for the client to
2188          * advertise a zero window, then kill -9 the FTP client, wheee...
2189          * Note: timeout is always zero in such a case.
2190          */
2191         if (unlikely(tcp_sk(sk)->repair)) {
2192                 sk->sk_prot->disconnect(sk, 0);
2193         } else if (data_was_unread) {
2194                 /* Unread data was tossed, zap the connection. */
2195                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2196                 tcp_set_state(sk, TCP_CLOSE);
2197                 tcp_send_active_reset(sk, sk->sk_allocation);
2198         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2199                 /* Check zero linger _after_ checking for unread data. */
2200                 sk->sk_prot->disconnect(sk, 0);
2201                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2202         } else if (tcp_close_state(sk)) {
2203                 /* We FIN if the application ate all the data before
2204                  * zapping the connection.
2205                  */
2206
2207                 /* RED-PEN. Formally speaking, we have broken TCP state
2208                  * machine. State transitions:
2209                  *
2210                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2211                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2212                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2213                  *
2214                  * are legal only when FIN has been sent (i.e. in window),
2215                  * rather than queued out of window. Purists blame.
2216                  *
2217                  * F.e. "RFC state" is ESTABLISHED,
2218                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2219                  *
2220                  * The visible declinations are that sometimes
2221                  * we enter time-wait state, when it is not required really
2222                  * (harmless), do not send active resets, when they are
2223                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2224                  * they look as CLOSING or LAST_ACK for Linux)
2225                  * Probably, I missed some more holelets.
2226                  *                                              --ANK
2227                  * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2228                  * in a single packet! (May consider it later but will
2229                  * probably need API support or TCP_CORK SYN-ACK until
2230                  * data is written and socket is closed.)
2231                  */
2232                 tcp_send_fin(sk);
2233         }
2234
2235         sk_stream_wait_close(sk, timeout);
2236
2237 adjudge_to_death:
2238         state = sk->sk_state;
2239         sock_hold(sk);
2240         sock_orphan(sk);
2241
2242         /* It is the last release_sock in its life. It will remove backlog. */
2243         release_sock(sk);
2244
2245
2246         /* Now socket is owned by kernel and we acquire BH lock
2247          *  to finish close. No need to check for user refs.
2248          */
2249         local_bh_disable();
2250         bh_lock_sock(sk);
2251         WARN_ON(sock_owned_by_user(sk));
2252
2253         percpu_counter_inc(sk->sk_prot->orphan_count);
2254
2255         /* Have we already been destroyed by a softirq or backlog? */
2256         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2257                 goto out;
2258
2259         /*      This is a (useful) BSD violating of the RFC. There is a
2260          *      problem with TCP as specified in that the other end could
2261          *      keep a socket open forever with no application left this end.
2262          *      We use a 1 minute timeout (about the same as BSD) then kill
2263          *      our end. If they send after that then tough - BUT: long enough
2264          *      that we won't make the old 4*rto = almost no time - whoops
2265          *      reset mistake.
2266          *
2267          *      Nope, it was not mistake. It is really desired behaviour
2268          *      f.e. on http servers, when such sockets are useless, but
2269          *      consume significant resources. Let's do it with special
2270          *      linger2 option.                                 --ANK
2271          */
2272
2273         if (sk->sk_state == TCP_FIN_WAIT2) {
2274                 struct tcp_sock *tp = tcp_sk(sk);
2275                 if (tp->linger2 < 0) {
2276                         tcp_set_state(sk, TCP_CLOSE);
2277                         tcp_send_active_reset(sk, GFP_ATOMIC);
2278                         __NET_INC_STATS(sock_net(sk),
2279                                         LINUX_MIB_TCPABORTONLINGER);
2280                 } else {
2281                         const int tmo = tcp_fin_time(sk);
2282
2283                         if (tmo > TCP_TIMEWAIT_LEN) {
2284                                 inet_csk_reset_keepalive_timer(sk,
2285                                                 tmo - TCP_TIMEWAIT_LEN);
2286                         } else {
2287                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2288                                 goto out;
2289                         }
2290                 }
2291         }
2292         if (sk->sk_state != TCP_CLOSE) {
2293                 sk_mem_reclaim(sk);
2294                 if (tcp_check_oom(sk, 0)) {
2295                         tcp_set_state(sk, TCP_CLOSE);
2296                         tcp_send_active_reset(sk, GFP_ATOMIC);
2297                         __NET_INC_STATS(sock_net(sk),
2298                                         LINUX_MIB_TCPABORTONMEMORY);
2299                 }
2300         }
2301
2302         if (sk->sk_state == TCP_CLOSE) {
2303                 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2304                 /* We could get here with a non-NULL req if the socket is
2305                  * aborted (e.g., closed with unread data) before 3WHS
2306                  * finishes.
2307                  */
2308                 if (req)
2309                         reqsk_fastopen_remove(sk, req, false);
2310                 inet_csk_destroy_sock(sk);
2311         }
2312         /* Otherwise, socket is reprieved until protocol close. */
2313
2314 out:
2315         bh_unlock_sock(sk);
2316         local_bh_enable();
2317         sock_put(sk);
2318 }
2319 EXPORT_SYMBOL(tcp_close);
2320
2321 /* These states need RST on ABORT according to RFC793 */
2322
2323 static inline bool tcp_need_reset(int state)
2324 {
2325         return (1 << state) &
2326                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2327                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2328 }
2329
2330 static void tcp_rtx_queue_purge(struct sock *sk)
2331 {
2332         struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2333
2334         while (p) {
2335                 struct sk_buff *skb = rb_to_skb(p);
2336
2337                 p = rb_next(p);
2338                 /* Since we are deleting whole queue, no need to
2339                  * list_del(&skb->tcp_tsorted_anchor)
2340                  */
2341                 tcp_rtx_queue_unlink(skb, sk);
2342                 sk_wmem_free_skb(sk, skb);
2343         }
2344 }
2345
2346 void tcp_write_queue_purge(struct sock *sk)
2347 {
2348         struct sk_buff *skb;
2349
2350         tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2351         while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2352                 tcp_skb_tsorted_anchor_cleanup(skb);
2353                 sk_wmem_free_skb(sk, skb);
2354         }
2355         tcp_rtx_queue_purge(sk);
2356         INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2357         sk_mem_reclaim(sk);
2358         tcp_clear_all_retrans_hints(tcp_sk(sk));
2359 }
2360
2361 int tcp_disconnect(struct sock *sk, int flags)
2362 {
2363         struct inet_sock *inet = inet_sk(sk);
2364         struct inet_connection_sock *icsk = inet_csk(sk);
2365         struct tcp_sock *tp = tcp_sk(sk);
2366         int err = 0;
2367         int old_state = sk->sk_state;
2368
2369         if (old_state != TCP_CLOSE)
2370                 tcp_set_state(sk, TCP_CLOSE);
2371
2372         /* ABORT function of RFC793 */
2373         if (old_state == TCP_LISTEN) {
2374                 inet_csk_listen_stop(sk);
2375         } else if (unlikely(tp->repair)) {
2376                 sk->sk_err = ECONNABORTED;
2377         } else if (tcp_need_reset(old_state) ||
2378                    (tp->snd_nxt != tp->write_seq &&
2379                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2380                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2381                  * states
2382                  */
2383                 tcp_send_active_reset(sk, gfp_any());
2384                 sk->sk_err = ECONNRESET;
2385         } else if (old_state == TCP_SYN_SENT)
2386                 sk->sk_err = ECONNRESET;
2387
2388         tcp_clear_xmit_timers(sk);
2389         __skb_queue_purge(&sk->sk_receive_queue);
2390         tcp_write_queue_purge(sk);
2391         tcp_fastopen_active_disable_ofo_check(sk);
2392         skb_rbtree_purge(&tp->out_of_order_queue);
2393
2394         inet->inet_dport = 0;
2395
2396         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2397                 inet_reset_saddr(sk);
2398
2399         sk->sk_shutdown = 0;
2400         sock_reset_flag(sk, SOCK_DONE);
2401         tp->srtt_us = 0;
2402         tp->write_seq += tp->max_window + 2;
2403         if (tp->write_seq == 0)
2404                 tp->write_seq = 1;
2405         icsk->icsk_backoff = 0;
2406         tp->snd_cwnd = 2;
2407         icsk->icsk_probes_out = 0;
2408         tp->packets_out = 0;
2409         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2410         tp->snd_cwnd_cnt = 0;
2411         tp->window_clamp = 0;
2412         tcp_set_ca_state(sk, TCP_CA_Open);
2413         tp->is_sack_reneg = 0;
2414         tcp_clear_retrans(tp);
2415         inet_csk_delack_init(sk);
2416         /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2417          * issue in __tcp_select_window()
2418          */
2419         icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2420         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2421         __sk_dst_reset(sk);
2422         dst_release(sk->sk_rx_dst);
2423         sk->sk_rx_dst = NULL;
2424         tcp_saved_syn_free(tp);
2425
2426         /* Clean up fastopen related fields */
2427         tcp_free_fastopen_req(tp);
2428         inet->defer_connect = 0;
2429
2430         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2431
2432         sk->sk_error_report(sk);
2433         return err;
2434 }
2435 EXPORT_SYMBOL(tcp_disconnect);
2436
2437 static inline bool tcp_can_repair_sock(const struct sock *sk)
2438 {
2439         return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2440                 (sk->sk_state != TCP_LISTEN);
2441 }
2442
2443 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2444 {
2445         struct tcp_repair_window opt;
2446
2447         if (!tp->repair)
2448                 return -EPERM;
2449
2450         if (len != sizeof(opt))
2451                 return -EINVAL;
2452
2453         if (copy_from_user(&opt, optbuf, sizeof(opt)))
2454                 return -EFAULT;
2455
2456         if (opt.max_window < opt.snd_wnd)
2457                 return -EINVAL;
2458
2459         if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2460                 return -EINVAL;
2461
2462         if (after(opt.rcv_wup, tp->rcv_nxt))
2463                 return -EINVAL;
2464
2465         tp->snd_wl1     = opt.snd_wl1;
2466         tp->snd_wnd     = opt.snd_wnd;
2467         tp->max_window  = opt.max_window;
2468
2469         tp->rcv_wnd     = opt.rcv_wnd;
2470         tp->rcv_wup     = opt.rcv_wup;
2471
2472         return 0;
2473 }
2474
2475 static int tcp_repair_options_est(struct sock *sk,
2476                 struct tcp_repair_opt __user *optbuf, unsigned int len)
2477 {
2478         struct tcp_sock *tp = tcp_sk(sk);
2479         struct tcp_repair_opt opt;
2480
2481         while (len >= sizeof(opt)) {
2482                 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2483                         return -EFAULT;
2484
2485                 optbuf++;
2486                 len -= sizeof(opt);
2487
2488                 switch (opt.opt_code) {
2489                 case TCPOPT_MSS:
2490                         tp->rx_opt.mss_clamp = opt.opt_val;
2491                         tcp_mtup_init(sk);
2492                         break;
2493                 case TCPOPT_WINDOW:
2494                         {
2495                                 u16 snd_wscale = opt.opt_val & 0xFFFF;
2496                                 u16 rcv_wscale = opt.opt_val >> 16;
2497
2498                                 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
2499                                         return -EFBIG;
2500
2501                                 tp->rx_opt.snd_wscale = snd_wscale;
2502                                 tp->rx_opt.rcv_wscale = rcv_wscale;
2503                                 tp->rx_opt.wscale_ok = 1;
2504                         }
2505                         break;
2506                 case TCPOPT_SACK_PERM:
2507                         if (opt.opt_val != 0)
2508                                 return -EINVAL;
2509
2510                         tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2511                         break;
2512                 case TCPOPT_TIMESTAMP:
2513                         if (opt.opt_val != 0)
2514                                 return -EINVAL;
2515
2516                         tp->rx_opt.tstamp_ok = 1;
2517                         break;
2518                 }
2519         }
2520
2521         return 0;
2522 }
2523
2524 /*
2525  *      Socket option code for TCP.
2526  */
2527 static int do_tcp_setsockopt(struct sock *sk, int level,
2528                 int optname, char __user *optval, unsigned int optlen)
2529 {
2530         struct tcp_sock *tp = tcp_sk(sk);
2531         struct inet_connection_sock *icsk = inet_csk(sk);
2532         struct net *net = sock_net(sk);
2533         int val;
2534         int err = 0;
2535
2536         /* These are data/string values, all the others are ints */
2537         switch (optname) {
2538         case TCP_CONGESTION: {
2539                 char name[TCP_CA_NAME_MAX];
2540
2541                 if (optlen < 1)
2542                         return -EINVAL;
2543
2544                 val = strncpy_from_user(name, optval,
2545                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2546                 if (val < 0)
2547                         return -EFAULT;
2548                 name[val] = 0;
2549
2550                 lock_sock(sk);
2551                 err = tcp_set_congestion_control(sk, name, true, true);
2552                 release_sock(sk);
2553                 return err;
2554         }
2555         case TCP_ULP: {
2556                 char name[TCP_ULP_NAME_MAX];
2557
2558                 if (optlen < 1)
2559                         return -EINVAL;
2560
2561                 val = strncpy_from_user(name, optval,
2562                                         min_t(long, TCP_ULP_NAME_MAX - 1,
2563                                               optlen));
2564                 if (val < 0)
2565                         return -EFAULT;
2566                 name[val] = 0;
2567
2568                 lock_sock(sk);
2569                 err = tcp_set_ulp(sk, name);
2570                 release_sock(sk);
2571                 return err;
2572         }
2573         case TCP_FASTOPEN_KEY: {
2574                 __u8 key[TCP_FASTOPEN_KEY_LENGTH];
2575
2576                 if (optlen != sizeof(key))
2577                         return -EINVAL;
2578
2579                 if (copy_from_user(key, optval, optlen))
2580                         return -EFAULT;
2581
2582                 return tcp_fastopen_reset_cipher(net, sk, key, sizeof(key));
2583         }
2584         default:
2585                 /* fallthru */
2586                 break;
2587         }
2588
2589         if (optlen < sizeof(int))
2590                 return -EINVAL;
2591
2592         if (get_user(val, (int __user *)optval))
2593                 return -EFAULT;
2594
2595         lock_sock(sk);
2596
2597         switch (optname) {
2598         case TCP_MAXSEG:
2599                 /* Values greater than interface MTU won't take effect. However
2600                  * at the point when this call is done we typically don't yet
2601                  * know which interface is going to be used
2602                  */
2603                 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
2604                         err = -EINVAL;
2605                         break;
2606                 }
2607                 tp->rx_opt.user_mss = val;
2608                 break;
2609
2610         case TCP_NODELAY:
2611                 if (val) {
2612                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2613                          * this option on corked socket is remembered, but
2614                          * it is not activated until cork is cleared.
2615                          *
2616                          * However, when TCP_NODELAY is set we make
2617                          * an explicit push, which overrides even TCP_CORK
2618                          * for currently queued segments.
2619                          */
2620                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2621                         tcp_push_pending_frames(sk);
2622                 } else {
2623                         tp->nonagle &= ~TCP_NAGLE_OFF;
2624                 }
2625                 break;
2626
2627         case TCP_THIN_LINEAR_TIMEOUTS:
2628                 if (val < 0 || val > 1)
2629                         err = -EINVAL;
2630                 else
2631                         tp->thin_lto = val;
2632                 break;
2633
2634         case TCP_THIN_DUPACK:
2635                 if (val < 0 || val > 1)
2636                         err = -EINVAL;
2637                 break;
2638
2639         case TCP_REPAIR:
2640                 if (!tcp_can_repair_sock(sk))
2641                         err = -EPERM;
2642                 else if (val == 1) {
2643                         tp->repair = 1;
2644                         sk->sk_reuse = SK_FORCE_REUSE;
2645                         tp->repair_queue = TCP_NO_QUEUE;
2646                 } else if (val == 0) {
2647                         tp->repair = 0;
2648                         sk->sk_reuse = SK_NO_REUSE;
2649                         tcp_send_window_probe(sk);
2650                 } else
2651                         err = -EINVAL;
2652
2653                 break;
2654
2655         case TCP_REPAIR_QUEUE:
2656                 if (!tp->repair)
2657                         err = -EPERM;
2658                 else if (val < TCP_QUEUES_NR)
2659                         tp->repair_queue = val;
2660                 else
2661                         err = -EINVAL;
2662                 break;
2663
2664         case TCP_QUEUE_SEQ:
2665                 if (sk->sk_state != TCP_CLOSE)
2666                         err = -EPERM;
2667                 else if (tp->repair_queue == TCP_SEND_QUEUE)
2668                         tp->write_seq = val;
2669                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2670                         tp->rcv_nxt = val;
2671                 else
2672                         err = -EINVAL;
2673                 break;
2674
2675         case TCP_REPAIR_OPTIONS:
2676                 if (!tp->repair)
2677                         err = -EINVAL;
2678                 else if (sk->sk_state == TCP_ESTABLISHED)
2679                         err = tcp_repair_options_est(sk,
2680                                         (struct tcp_repair_opt __user *)optval,
2681                                         optlen);
2682                 else
2683                         err = -EPERM;
2684                 break;
2685
2686         case TCP_CORK:
2687                 /* When set indicates to always queue non-full frames.
2688                  * Later the user clears this option and we transmit
2689                  * any pending partial frames in the queue.  This is
2690                  * meant to be used alongside sendfile() to get properly
2691                  * filled frames when the user (for example) must write
2692                  * out headers with a write() call first and then use
2693                  * sendfile to send out the data parts.
2694                  *
2695                  * TCP_CORK can be set together with TCP_NODELAY and it is
2696                  * stronger than TCP_NODELAY.
2697                  */
2698                 if (val) {
2699                         tp->nonagle |= TCP_NAGLE_CORK;
2700                 } else {
2701                         tp->nonagle &= ~TCP_NAGLE_CORK;
2702                         if (tp->nonagle&TCP_NAGLE_OFF)
2703                                 tp->nonagle |= TCP_NAGLE_PUSH;
2704                         tcp_push_pending_frames(sk);
2705                 }
2706                 break;
2707
2708         case TCP_KEEPIDLE:
2709                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2710                         err = -EINVAL;
2711                 else {
2712                         tp->keepalive_time = val * HZ;
2713                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2714                             !((1 << sk->sk_state) &
2715                               (TCPF_CLOSE | TCPF_LISTEN))) {
2716                                 u32 elapsed = keepalive_time_elapsed(tp);
2717                                 if (tp->keepalive_time > elapsed)
2718                                         elapsed = tp->keepalive_time - elapsed;
2719                                 else
2720                                         elapsed = 0;
2721                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2722                         }
2723                 }
2724                 break;
2725         case TCP_KEEPINTVL:
2726                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2727                         err = -EINVAL;
2728                 else
2729                         tp->keepalive_intvl = val * HZ;
2730                 break;
2731         case TCP_KEEPCNT:
2732                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2733                         err = -EINVAL;
2734                 else
2735                         tp->keepalive_probes = val;
2736                 break;
2737         case TCP_SYNCNT:
2738                 if (val < 1 || val > MAX_TCP_SYNCNT)
2739                         err = -EINVAL;
2740                 else
2741                         icsk->icsk_syn_retries = val;
2742                 break;
2743
2744         case TCP_SAVE_SYN:
2745                 if (val < 0 || val > 1)
2746                         err = -EINVAL;
2747                 else
2748                         tp->save_syn = val;
2749                 break;
2750
2751         case TCP_LINGER2:
2752                 if (val < 0)
2753                         tp->linger2 = -1;
2754                 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
2755                         tp->linger2 = 0;
2756                 else
2757                         tp->linger2 = val * HZ;
2758                 break;
2759
2760         case TCP_DEFER_ACCEPT:
2761                 /* Translate value in seconds to number of retransmits */
2762                 icsk->icsk_accept_queue.rskq_defer_accept =
2763                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2764                                         TCP_RTO_MAX / HZ);
2765                 break;
2766
2767         case TCP_WINDOW_CLAMP:
2768                 if (!val) {
2769                         if (sk->sk_state != TCP_CLOSE) {
2770                                 err = -EINVAL;
2771                                 break;
2772                         }
2773                         tp->window_clamp = 0;
2774                 } else
2775                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2776                                                 SOCK_MIN_RCVBUF / 2 : val;
2777                 break;
2778
2779         case TCP_QUICKACK:
2780                 if (!val) {
2781                         icsk->icsk_ack.pingpong = 1;
2782                 } else {
2783                         icsk->icsk_ack.pingpong = 0;
2784                         if ((1 << sk->sk_state) &
2785                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2786                             inet_csk_ack_scheduled(sk)) {
2787                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2788                                 tcp_cleanup_rbuf(sk, 1);
2789                                 if (!(val & 1))
2790                                         icsk->icsk_ack.pingpong = 1;
2791                         }
2792                 }
2793                 break;
2794
2795 #ifdef CONFIG_TCP_MD5SIG
2796         case TCP_MD5SIG:
2797         case TCP_MD5SIG_EXT:
2798                 /* Read the IP->Key mappings from userspace */
2799                 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
2800                 break;
2801 #endif
2802         case TCP_USER_TIMEOUT:
2803                 /* Cap the max time in ms TCP will retry or probe the window
2804                  * before giving up and aborting (ETIMEDOUT) a connection.
2805                  */
2806                 if (val < 0)
2807                         err = -EINVAL;
2808                 else
2809                         icsk->icsk_user_timeout = msecs_to_jiffies(val);
2810                 break;
2811
2812         case TCP_FASTOPEN:
2813                 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2814                     TCPF_LISTEN))) {
2815                         tcp_fastopen_init_key_once(net);
2816
2817                         fastopen_queue_tune(sk, val);
2818                 } else {
2819                         err = -EINVAL;
2820                 }
2821                 break;
2822         case TCP_FASTOPEN_CONNECT:
2823                 if (val > 1 || val < 0) {
2824                         err = -EINVAL;
2825                 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
2826                         if (sk->sk_state == TCP_CLOSE)
2827                                 tp->fastopen_connect = val;
2828                         else
2829                                 err = -EINVAL;
2830                 } else {
2831                         err = -EOPNOTSUPP;
2832                 }
2833                 break;
2834         case TCP_FASTOPEN_NO_COOKIE:
2835                 if (val > 1 || val < 0)
2836                         err = -EINVAL;
2837                 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2838                         err = -EINVAL;
2839                 else
2840                         tp->fastopen_no_cookie = val;
2841                 break;
2842         case TCP_TIMESTAMP:
2843                 if (!tp->repair)
2844                         err = -EPERM;
2845                 else
2846                         tp->tsoffset = val - tcp_time_stamp_raw();
2847                 break;
2848         case TCP_REPAIR_WINDOW:
2849                 err = tcp_repair_set_window(tp, optval, optlen);
2850                 break;
2851         case TCP_NOTSENT_LOWAT:
2852                 tp->notsent_lowat = val;
2853                 sk->sk_write_space(sk);
2854                 break;
2855         default:
2856                 err = -ENOPROTOOPT;
2857                 break;
2858         }
2859
2860         release_sock(sk);
2861         return err;
2862 }
2863
2864 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2865                    unsigned int optlen)
2866 {
2867         const struct inet_connection_sock *icsk = inet_csk(sk);
2868
2869         if (level != SOL_TCP)
2870                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2871                                                      optval, optlen);
2872         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2873 }
2874 EXPORT_SYMBOL(tcp_setsockopt);
2875
2876 #ifdef CONFIG_COMPAT
2877 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2878                           char __user *optval, unsigned int optlen)
2879 {
2880         if (level != SOL_TCP)
2881                 return inet_csk_compat_setsockopt(sk, level, optname,
2882                                                   optval, optlen);
2883         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2884 }
2885 EXPORT_SYMBOL(compat_tcp_setsockopt);
2886 #endif
2887
2888 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
2889                                       struct tcp_info *info)
2890 {
2891         u64 stats[__TCP_CHRONO_MAX], total = 0;
2892         enum tcp_chrono i;
2893
2894         for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
2895                 stats[i] = tp->chrono_stat[i - 1];
2896                 if (i == tp->chrono_type)
2897                         stats[i] += tcp_jiffies32 - tp->chrono_start;
2898                 stats[i] *= USEC_PER_SEC / HZ;
2899                 total += stats[i];
2900         }
2901
2902         info->tcpi_busy_time = total;
2903         info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
2904         info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
2905 }
2906
2907 /* Return information about state of tcp endpoint in API format. */
2908 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2909 {
2910         const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
2911         const struct inet_connection_sock *icsk = inet_csk(sk);
2912         u32 now;
2913         u64 rate64;
2914         bool slow;
2915         u32 rate;
2916
2917         memset(info, 0, sizeof(*info));
2918         if (sk->sk_type != SOCK_STREAM)
2919                 return;
2920
2921         info->tcpi_state = inet_sk_state_load(sk);
2922
2923         /* Report meaningful fields for all TCP states, including listeners */
2924         rate = READ_ONCE(sk->sk_pacing_rate);
2925         rate64 = rate != ~0U ? rate : ~0ULL;
2926         info->tcpi_pacing_rate = rate64;
2927
2928         rate = READ_ONCE(sk->sk_max_pacing_rate);
2929         rate64 = rate != ~0U ? rate : ~0ULL;
2930         info->tcpi_max_pacing_rate = rate64;
2931
2932         info->tcpi_reordering = tp->reordering;
2933         info->tcpi_snd_cwnd = tp->snd_cwnd;
2934
2935         if (info->tcpi_state == TCP_LISTEN) {
2936                 /* listeners aliased fields :
2937                  * tcpi_unacked -> Number of children ready for accept()
2938                  * tcpi_sacked  -> max backlog
2939                  */
2940                 info->tcpi_unacked = sk->sk_ack_backlog;
2941                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2942                 return;
2943         }
2944
2945         slow = lock_sock_fast(sk);
2946
2947         info->tcpi_ca_state = icsk->icsk_ca_state;
2948         info->tcpi_retransmits = icsk->icsk_retransmits;
2949         info->tcpi_probes = icsk->icsk_probes_out;
2950         info->tcpi_backoff = icsk->icsk_backoff;
2951
2952         if (tp->rx_opt.tstamp_ok)
2953                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2954         if (tcp_is_sack(tp))
2955                 info->tcpi_options |= TCPI_OPT_SACK;
2956         if (tp->rx_opt.wscale_ok) {
2957                 info->tcpi_options |= TCPI_OPT_WSCALE;
2958                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2959                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2960         }
2961
2962         if (tp->ecn_flags & TCP_ECN_OK)
2963                 info->tcpi_options |= TCPI_OPT_ECN;
2964         if (tp->ecn_flags & TCP_ECN_SEEN)
2965                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2966         if (tp->syn_data_acked)
2967                 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2968
2969         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2970         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2971         info->tcpi_snd_mss = tp->mss_cache;
2972         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2973
2974         info->tcpi_unacked = tp->packets_out;
2975         info->tcpi_sacked = tp->sacked_out;
2976
2977         info->tcpi_lost = tp->lost_out;
2978         info->tcpi_retrans = tp->retrans_out;
2979
2980         now = tcp_jiffies32;
2981         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2982         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2983         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2984
2985         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2986         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2987         info->tcpi_rtt = tp->srtt_us >> 3;
2988         info->tcpi_rttvar = tp->mdev_us >> 2;
2989         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2990         info->tcpi_advmss = tp->advmss;
2991
2992         info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
2993         info->tcpi_rcv_space = tp->rcvq_space.space;
2994
2995         info->tcpi_total_retrans = tp->total_retrans;
2996
2997         info->tcpi_bytes_acked = tp->bytes_acked;
2998         info->tcpi_bytes_received = tp->bytes_received;
2999         info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3000         tcp_get_info_chrono_stats(tp, info);
3001
3002         info->tcpi_segs_out = tp->segs_out;
3003         info->tcpi_segs_in = tp->segs_in;
3004
3005         info->tcpi_min_rtt = tcp_min_rtt(tp);
3006         info->tcpi_data_segs_in = tp->data_segs_in;
3007         info->tcpi_data_segs_out = tp->data_segs_out;
3008
3009         info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3010         rate64 = tcp_compute_delivery_rate(tp);
3011         if (rate64)
3012                 info->tcpi_delivery_rate = rate64;
3013         unlock_sock_fast(sk, slow);
3014 }
3015 EXPORT_SYMBOL_GPL(tcp_get_info);
3016
3017 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk)
3018 {
3019         const struct tcp_sock *tp = tcp_sk(sk);
3020         struct sk_buff *stats;
3021         struct tcp_info info;
3022         u64 rate64;
3023         u32 rate;
3024
3025         stats = alloc_skb(7 * nla_total_size_64bit(sizeof(u64)) +
3026                           3 * nla_total_size(sizeof(u32)) +
3027                           2 * nla_total_size(sizeof(u8)), GFP_ATOMIC);
3028         if (!stats)
3029                 return NULL;
3030
3031         tcp_get_info_chrono_stats(tp, &info);
3032         nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3033                           info.tcpi_busy_time, TCP_NLA_PAD);
3034         nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3035                           info.tcpi_rwnd_limited, TCP_NLA_PAD);
3036         nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3037                           info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3038         nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3039                           tp->data_segs_out, TCP_NLA_PAD);
3040         nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3041                           tp->total_retrans, TCP_NLA_PAD);
3042
3043         rate = READ_ONCE(sk->sk_pacing_rate);
3044         rate64 = rate != ~0U ? rate : ~0ULL;
3045         nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3046
3047         rate64 = tcp_compute_delivery_rate(tp);
3048         nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3049
3050         nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3051         nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3052         nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3053
3054         nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3055         nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3056         return stats;
3057 }
3058
3059 static int do_tcp_getsockopt(struct sock *sk, int level,
3060                 int optname, char __user *optval, int __user *optlen)
3061 {
3062         struct inet_connection_sock *icsk = inet_csk(sk);
3063         struct tcp_sock *tp = tcp_sk(sk);
3064         struct net *net = sock_net(sk);
3065         int val, len;
3066
3067         if (get_user(len, optlen))
3068                 return -EFAULT;
3069
3070         len = min_t(unsigned int, len, sizeof(int));
3071
3072         if (len < 0)
3073                 return -EINVAL;
3074
3075         switch (optname) {
3076         case TCP_MAXSEG:
3077                 val = tp->mss_cache;
3078                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3079                         val = tp->rx_opt.user_mss;
3080                 if (tp->repair)
3081                         val = tp->rx_opt.mss_clamp;
3082                 break;
3083         case TCP_NODELAY:
3084                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3085                 break;
3086         case TCP_CORK:
3087                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3088                 break;
3089         case TCP_KEEPIDLE:
3090                 val = keepalive_time_when(tp) / HZ;
3091                 break;
3092         case TCP_KEEPINTVL:
3093                 val = keepalive_intvl_when(tp) / HZ;
3094                 break;
3095         case TCP_KEEPCNT:
3096                 val = keepalive_probes(tp);
3097                 break;
3098         case TCP_SYNCNT:
3099                 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3100                 break;
3101         case TCP_LINGER2:
3102                 val = tp->linger2;
3103                 if (val >= 0)
3104                         val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3105                 break;
3106         case TCP_DEFER_ACCEPT:
3107                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3108                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3109                 break;
3110         case TCP_WINDOW_CLAMP:
3111                 val = tp->window_clamp;
3112                 break;
3113         case TCP_INFO: {
3114                 struct tcp_info info;
3115
3116                 if (get_user(len, optlen))
3117                         return -EFAULT;
3118
3119                 tcp_get_info(sk, &info);
3120
3121                 len = min_t(unsigned int, len, sizeof(info));
3122                 if (put_user(len, optlen))
3123                         return -EFAULT;
3124                 if (copy_to_user(optval, &info, len))
3125                         return -EFAULT;
3126                 return 0;
3127         }
3128         case TCP_CC_INFO: {
3129                 const struct tcp_congestion_ops *ca_ops;
3130                 union tcp_cc_info info;
3131                 size_t sz = 0;
3132                 int attr;
3133
3134                 if (get_user(len, optlen))
3135                         return -EFAULT;
3136
3137                 ca_ops = icsk->icsk_ca_ops;
3138                 if (ca_ops && ca_ops->get_info)
3139                         sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3140
3141                 len = min_t(unsigned int, len, sz);
3142                 if (put_user(len, optlen))
3143                         return -EFAULT;
3144                 if (copy_to_user(optval, &info, len))
3145                         return -EFAULT;
3146                 return 0;
3147         }
3148         case TCP_QUICKACK:
3149                 val = !icsk->icsk_ack.pingpong;
3150                 break;
3151
3152         case TCP_CONGESTION:
3153                 if (get_user(len, optlen))
3154                         return -EFAULT;
3155                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
3156                 if (put_user(len, optlen))
3157                         return -EFAULT;
3158                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3159                         return -EFAULT;
3160                 return 0;
3161
3162         case TCP_ULP:
3163                 if (get_user(len, optlen))
3164                         return -EFAULT;
3165                 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
3166                 if (!icsk->icsk_ulp_ops) {
3167                         if (put_user(0, optlen))
3168                                 return -EFAULT;
3169                         return 0;
3170                 }
3171                 if (put_user(len, optlen))
3172                         return -EFAULT;
3173                 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
3174                         return -EFAULT;
3175                 return 0;
3176
3177         case TCP_FASTOPEN_KEY: {
3178                 __u8 key[TCP_FASTOPEN_KEY_LENGTH];
3179                 struct tcp_fastopen_context *ctx;
3180
3181                 if (get_user(len, optlen))
3182                         return -EFAULT;
3183
3184                 rcu_read_lock();
3185                 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
3186                 if (ctx)
3187                         memcpy(key, ctx->key, sizeof(key));
3188                 else
3189                         len = 0;
3190                 rcu_read_unlock();
3191
3192                 len = min_t(unsigned int, len, sizeof(key));
3193                 if (put_user(len, optlen))
3194                         return -EFAULT;
3195                 if (copy_to_user(optval, key, len))
3196                         return -EFAULT;
3197                 return 0;
3198         }
3199         case TCP_THIN_LINEAR_TIMEOUTS:
3200                 val = tp->thin_lto;
3201                 break;
3202
3203         case TCP_THIN_DUPACK:
3204                 val = 0;
3205                 break;
3206
3207         case TCP_REPAIR:
3208                 val = tp->repair;
3209                 break;
3210
3211         case TCP_REPAIR_QUEUE:
3212                 if (tp->repair)
3213                         val = tp->repair_queue;
3214                 else
3215                         return -EINVAL;
3216                 break;
3217
3218         case TCP_REPAIR_WINDOW: {
3219                 struct tcp_repair_window opt;
3220
3221                 if (get_user(len, optlen))
3222                         return -EFAULT;
3223
3224                 if (len != sizeof(opt))
3225                         return -EINVAL;
3226
3227                 if (!tp->repair)
3228                         return -EPERM;
3229
3230                 opt.snd_wl1     = tp->snd_wl1;
3231                 opt.snd_wnd     = tp->snd_wnd;
3232                 opt.max_window  = tp->max_window;
3233                 opt.rcv_wnd     = tp->rcv_wnd;
3234                 opt.rcv_wup     = tp->rcv_wup;
3235
3236                 if (copy_to_user(optval, &opt, len))
3237                         return -EFAULT;
3238                 return 0;
3239         }
3240         case TCP_QUEUE_SEQ:
3241                 if (tp->repair_queue == TCP_SEND_QUEUE)
3242                         val = tp->write_seq;
3243                 else if (tp->repair_queue == TCP_RECV_QUEUE)
3244                         val = tp->rcv_nxt;
3245                 else
3246                         return -EINVAL;
3247                 break;
3248
3249         case TCP_USER_TIMEOUT:
3250                 val = jiffies_to_msecs(icsk->icsk_user_timeout);
3251                 break;
3252
3253         case TCP_FASTOPEN:
3254                 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3255                 break;
3256
3257         case TCP_FASTOPEN_CONNECT:
3258                 val = tp->fastopen_connect;
3259                 break;
3260
3261         case TCP_FASTOPEN_NO_COOKIE:
3262                 val = tp->fastopen_no_cookie;
3263                 break;
3264
3265         case TCP_TIMESTAMP:
3266                 val = tcp_time_stamp_raw() + tp->tsoffset;
3267                 break;
3268         case TCP_NOTSENT_LOWAT:
3269                 val = tp->notsent_lowat;
3270                 break;
3271         case TCP_SAVE_SYN:
3272                 val = tp->save_syn;
3273                 break;
3274         case TCP_SAVED_SYN: {
3275                 if (get_user(len, optlen))
3276                         return -EFAULT;
3277
3278                 lock_sock(sk);
3279                 if (tp->saved_syn) {
3280                         if (len < tp->saved_syn[0]) {
3281                                 if (put_user(tp->saved_syn[0], optlen)) {
3282                                         release_sock(sk);
3283                                         return -EFAULT;
3284                                 }
3285                                 release_sock(sk);
3286                                 return -EINVAL;
3287                         }
3288                         len = tp->saved_syn[0];
3289                         if (put_user(len, optlen)) {
3290                                 release_sock(sk);
3291                                 return -EFAULT;
3292                         }
3293                         if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3294                                 release_sock(sk);
3295                                 return -EFAULT;
3296                         }
3297                         tcp_saved_syn_free(tp);
3298                         release_sock(sk);
3299                 } else {
3300                         release_sock(sk);
3301                         len = 0;
3302                         if (put_user(len, optlen))
3303                                 return -EFAULT;
3304                 }
3305                 return 0;
3306         }
3307         default:
3308                 return -ENOPROTOOPT;
3309         }
3310
3311         if (put_user(len, optlen))
3312                 return -EFAULT;
3313         if (copy_to_user(optval, &val, len))
3314                 return -EFAULT;
3315         return 0;
3316 }
3317
3318 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3319                    int __user *optlen)
3320 {
3321         struct inet_connection_sock *icsk = inet_csk(sk);
3322
3323         if (level != SOL_TCP)
3324                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3325                                                      optval, optlen);
3326         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3327 }
3328 EXPORT_SYMBOL(tcp_getsockopt);
3329
3330 #ifdef CONFIG_COMPAT
3331 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3332                           char __user *optval, int __user *optlen)
3333 {
3334         if (level != SOL_TCP)
3335                 return inet_csk_compat_getsockopt(sk, level, optname,
3336                                                   optval, optlen);
3337         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3338 }
3339 EXPORT_SYMBOL(compat_tcp_getsockopt);
3340 #endif
3341
3342 #ifdef CONFIG_TCP_MD5SIG
3343 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3344 static DEFINE_MUTEX(tcp_md5sig_mutex);
3345 static bool tcp_md5sig_pool_populated = false;
3346
3347 static void __tcp_alloc_md5sig_pool(void)
3348 {
3349         struct crypto_ahash *hash;
3350         int cpu;
3351
3352         hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3353         if (IS_ERR(hash))
3354                 return;
3355
3356         for_each_possible_cpu(cpu) {
3357                 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3358                 struct ahash_request *req;
3359
3360                 if (!scratch) {
3361                         scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3362                                                sizeof(struct tcphdr),
3363                                                GFP_KERNEL,
3364                                                cpu_to_node(cpu));
3365                         if (!scratch)
3366                                 return;
3367                         per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3368                 }
3369                 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3370                         continue;
3371
3372                 req = ahash_request_alloc(hash, GFP_KERNEL);
3373                 if (!req)
3374                         return;
3375
3376                 ahash_request_set_callback(req, 0, NULL, NULL);
3377
3378                 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3379         }
3380         /* before setting tcp_md5sig_pool_populated, we must commit all writes
3381          * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3382          */
3383         smp_wmb();
3384         tcp_md5sig_pool_populated = true;
3385 }
3386
3387 bool tcp_alloc_md5sig_pool(void)
3388 {
3389         if (unlikely(!tcp_md5sig_pool_populated)) {
3390                 mutex_lock(&tcp_md5sig_mutex);
3391
3392                 if (!tcp_md5sig_pool_populated)
3393                         __tcp_alloc_md5sig_pool();
3394
3395                 mutex_unlock(&tcp_md5sig_mutex);
3396         }
3397         return tcp_md5sig_pool_populated;
3398 }
3399 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3400
3401
3402 /**
3403  *      tcp_get_md5sig_pool - get md5sig_pool for this user
3404  *
3405  *      We use percpu structure, so if we succeed, we exit with preemption
3406  *      and BH disabled, to make sure another thread or softirq handling
3407  *      wont try to get same context.
3408  */
3409 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3410 {
3411         local_bh_disable();
3412
3413         if (tcp_md5sig_pool_populated) {
3414                 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3415                 smp_rmb();
3416                 return this_cpu_ptr(&tcp_md5sig_pool);
3417         }
3418         local_bh_enable();
3419         return NULL;
3420 }
3421 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3422
3423 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3424                           const struct sk_buff *skb, unsigned int header_len)
3425 {
3426         struct scatterlist sg;
3427         const struct tcphdr *tp = tcp_hdr(skb);
3428         struct ahash_request *req = hp->md5_req;
3429         unsigned int i;
3430         const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3431                                            skb_headlen(skb) - header_len : 0;
3432         const struct skb_shared_info *shi = skb_shinfo(skb);
3433         struct sk_buff *frag_iter;
3434
3435         sg_init_table(&sg, 1);
3436
3437         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3438         ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3439         if (crypto_ahash_update(req))
3440                 return 1;
3441
3442         for (i = 0; i < shi->nr_frags; ++i) {
3443                 const struct skb_frag_struct *f = &shi->frags[i];
3444                 unsigned int offset = f->page_offset;
3445                 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3446
3447                 sg_set_page(&sg, page, skb_frag_size(f),
3448                             offset_in_page(offset));
3449                 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3450                 if (crypto_ahash_update(req))
3451                         return 1;
3452         }
3453
3454         skb_walk_frags(skb, frag_iter)
3455                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3456                         return 1;
3457
3458         return 0;
3459 }
3460 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3461
3462 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3463 {
3464         struct scatterlist sg;
3465
3466         sg_init_one(&sg, key->key, key->keylen);
3467         ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen);
3468         return crypto_ahash_update(hp->md5_req);
3469 }
3470 EXPORT_SYMBOL(tcp_md5_hash_key);
3471
3472 #endif
3473
3474 void tcp_done(struct sock *sk)
3475 {
3476         struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3477
3478         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3479                 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3480
3481         tcp_set_state(sk, TCP_CLOSE);
3482         tcp_clear_xmit_timers(sk);
3483         if (req)
3484                 reqsk_fastopen_remove(sk, req, false);
3485
3486         sk->sk_shutdown = SHUTDOWN_MASK;
3487
3488         if (!sock_flag(sk, SOCK_DEAD))
3489                 sk->sk_state_change(sk);
3490         else
3491                 inet_csk_destroy_sock(sk);
3492 }
3493 EXPORT_SYMBOL_GPL(tcp_done);
3494
3495 int tcp_abort(struct sock *sk, int err)
3496 {
3497         if (!sk_fullsock(sk)) {
3498                 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3499                         struct request_sock *req = inet_reqsk(sk);
3500
3501                         local_bh_disable();
3502                         inet_csk_reqsk_queue_drop_and_put(req->rsk_listener,
3503                                                           req);
3504                         local_bh_enable();
3505                         return 0;
3506                 }
3507                 return -EOPNOTSUPP;
3508         }
3509
3510         /* Don't race with userspace socket closes such as tcp_close. */
3511         lock_sock(sk);
3512
3513         if (sk->sk_state == TCP_LISTEN) {
3514                 tcp_set_state(sk, TCP_CLOSE);
3515                 inet_csk_listen_stop(sk);
3516         }
3517
3518         /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3519         local_bh_disable();
3520         bh_lock_sock(sk);
3521
3522         if (!sock_flag(sk, SOCK_DEAD)) {
3523                 sk->sk_err = err;
3524                 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3525                 smp_wmb();
3526                 sk->sk_error_report(sk);
3527                 if (tcp_need_reset(sk->sk_state))
3528                         tcp_send_active_reset(sk, GFP_ATOMIC);
3529                 tcp_done(sk);
3530         }
3531
3532         bh_unlock_sock(sk);
3533         local_bh_enable();
3534         release_sock(sk);
3535         return 0;
3536 }
3537 EXPORT_SYMBOL_GPL(tcp_abort);
3538
3539 extern struct tcp_congestion_ops tcp_reno;
3540
3541 static __initdata unsigned long thash_entries;
3542 static int __init set_thash_entries(char *str)
3543 {
3544         ssize_t ret;
3545
3546         if (!str)
3547                 return 0;
3548
3549         ret = kstrtoul(str, 0, &thash_entries);
3550         if (ret)
3551                 return 0;
3552
3553         return 1;
3554 }
3555 __setup("thash_entries=", set_thash_entries);
3556
3557 static void __init tcp_init_mem(void)
3558 {
3559         unsigned long limit = nr_free_buffer_pages() / 16;
3560
3561         limit = max(limit, 128UL);
3562         sysctl_tcp_mem[0] = limit / 4 * 3;              /* 4.68 % */
3563         sysctl_tcp_mem[1] = limit;                      /* 6.25 % */
3564         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;      /* 9.37 % */
3565 }
3566
3567 void __init tcp_init(void)
3568 {
3569         int max_rshare, max_wshare, cnt;
3570         unsigned long limit;
3571         unsigned int i;
3572
3573         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3574                      FIELD_SIZEOF(struct sk_buff, cb));
3575
3576         percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3577         percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3578         inet_hashinfo_init(&tcp_hashinfo);
3579         inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
3580                             thash_entries, 21,  /* one slot per 2 MB*/
3581                             0, 64 * 1024);
3582         tcp_hashinfo.bind_bucket_cachep =
3583                 kmem_cache_create("tcp_bind_bucket",
3584                                   sizeof(struct inet_bind_bucket), 0,
3585                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3586
3587         /* Size and allocate the main established and bind bucket
3588          * hash tables.
3589          *
3590          * The methodology is similar to that of the buffer cache.
3591          */
3592         tcp_hashinfo.ehash =
3593                 alloc_large_system_hash("TCP established",
3594                                         sizeof(struct inet_ehash_bucket),
3595                                         thash_entries,
3596                                         17, /* one slot per 128 KB of memory */
3597                                         0,
3598                                         NULL,
3599                                         &tcp_hashinfo.ehash_mask,
3600                                         0,
3601                                         thash_entries ? 0 : 512 * 1024);
3602         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3603                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3604
3605         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3606                 panic("TCP: failed to alloc ehash_locks");
3607         tcp_hashinfo.bhash =
3608                 alloc_large_system_hash("TCP bind",
3609                                         sizeof(struct inet_bind_hashbucket),
3610                                         tcp_hashinfo.ehash_mask + 1,
3611                                         17, /* one slot per 128 KB of memory */
3612                                         0,
3613                                         &tcp_hashinfo.bhash_size,
3614                                         NULL,
3615                                         0,
3616                                         64 * 1024);
3617         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3618         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3619                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3620                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3621         }
3622
3623
3624         cnt = tcp_hashinfo.ehash_mask + 1;
3625         sysctl_tcp_max_orphans = cnt / 2;
3626
3627         tcp_init_mem();
3628         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3629         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3630         max_wshare = min(4UL*1024*1024, limit);
3631         max_rshare = min(6UL*1024*1024, limit);
3632
3633         init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3634         init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
3635         init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3636
3637         init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3638         init_net.ipv4.sysctl_tcp_rmem[1] = 87380;
3639         init_net.ipv4.sysctl_tcp_rmem[2] = max(87380, max_rshare);
3640
3641         pr_info("Hash tables configured (established %u bind %u)\n",
3642                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3643
3644         tcp_v4_init();
3645         tcp_metrics_init();
3646         BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3647         tcp_tasklet_init();
3648 }