linux/net/ipv4/tcp.c
<<
>>
Prefs
   1/*
   2 * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3 *              operating system.  INET is implemented using the  BSD Socket
   4 *              interface as the means of communication with the user level.
   5 *
   6 *              Implementation of the Transmission Control Protocol(TCP).
   7 *
   8 * Authors:     Ross Biro
   9 *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10 *              Mark Evans, <evansmp@uhura.aston.ac.uk>
  11 *              Corey Minyard <wf-rch!minyard@relay.EU.net>
  12 *              Florian La Roche, <flla@stud.uni-sb.de>
  13 *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  14 *              Linus Torvalds, <torvalds@cs.helsinki.fi>
  15 *              Alan Cox, <gw4pts@gw4pts.ampr.org>
  16 *              Matthew Dillon, <dillon@apollo.west.oic.com>
  17 *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  18 *              Jorge Cwik, <jorge@laser.satlink.net>
  19 *
  20 * Fixes:
  21 *              Alan Cox        :       Numerous verify_area() calls
  22 *              Alan Cox        :       Set the ACK bit on a reset
  23 *              Alan Cox        :       Stopped it crashing if it closed while
  24 *                                      sk->inuse=1 and was trying to connect
  25 *                                      (tcp_err()).
  26 *              Alan Cox        :       All icmp error handling was broken
  27 *                                      pointers passed where wrong and the
  28 *                                      socket was looked up backwards. Nobody
  29 *                                      tested any icmp error code obviously.
  30 *              Alan Cox        :       tcp_err() now handled properly. It
  31 *                                      wakes people on errors. poll
  32 *                                      behaves and the icmp error race
  33 *                                      has gone by moving it into sock.c
  34 *              Alan Cox        :       tcp_send_reset() fixed to work for
  35 *                                      everything not just packets for
  36 *                                      unknown sockets.
  37 *              Alan Cox        :       tcp option processing.
  38 *              Alan Cox        :       Reset tweaked (still not 100%) [Had
  39 *                                      syn rule wrong]
  40 *              Herp Rosmanith  :       More reset fixes
  41 *              Alan Cox        :       No longer acks invalid rst frames.
  42 *                                      Acking any kind of RST is right out.
  43 *              Alan Cox        :       Sets an ignore me flag on an rst
  44 *                                      receive otherwise odd bits of prattle
  45 *                                      escape still
  46 *              Alan Cox        :       Fixed another acking RST frame bug.
  47 *                                      Should stop LAN workplace lockups.
  48 *              Alan Cox        :       Some tidyups using the new skb list
  49 *                                      facilities
  50 *              Alan Cox        :       sk->keepopen now seems to work
  51 *              Alan Cox        :       Pulls options out correctly on accepts
  52 *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
  53 *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
  54 *                                      bit to skb ops.
  55 *              Alan Cox        :       Tidied tcp_data to avoid a potential
  56 *                                      nasty.
  57 *              Alan Cox        :       Added some better commenting, as the
  58 *                                      tcp is hard to follow
  59 *              Alan Cox        :       Removed incorrect check for 20 * psh
  60 *      Michael O'Reilly        :       ack < copied bug fix.
  61 *      Johannes Stille         :       Misc tcp fixes (not all in yet).
  62 *              Alan Cox        :       FIN with no memory -> CRASH
  63 *              Alan Cox        :       Added socket option proto entries.
  64 *                                      Also added awareness of them to accept.
  65 *              Alan Cox        :       Added TCP options (SOL_TCP)
  66 *              Alan Cox        :       Switched wakeup calls to callbacks,
  67 *                                      so the kernel can layer network
  68 *                                      sockets.
  69 *              Alan Cox        :       Use ip_tos/ip_ttl settings.
  70 *              Alan Cox        :       Handle FIN (more) properly (we hope).
  71 *              Alan Cox        :       RST frames sent on unsynchronised
  72 *                                      state ack error.
  73 *              Alan Cox        :       Put in missing check for SYN bit.
  74 *              Alan Cox        :       Added tcp_select_window() aka NET2E
  75 *                                      window non shrink trick.
  76 *              Alan Cox        :       Added a couple of small NET2E timer
  77 *                                      fixes
  78 *              Charles Hedrick :       TCP fixes
  79 *              Toomas Tamm     :       TCP window fixes
  80 *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
  81 *              Charles Hedrick :       Rewrote most of it to actually work
  82 *              Linus           :       Rewrote tcp_read() and URG handling
  83 *                                      completely
  84 *              Gerhard Koerting:       Fixed some missing timer handling
  85 *              Matthew Dillon  :       Reworked TCP machine states as per RFC
  86 *              Gerhard Koerting:       PC/TCP workarounds
  87 *              Adam Caldwell   :       Assorted timer/timing errors
  88 *              Matthew Dillon  :       Fixed another RST bug
  89 *              Alan Cox        :       Move to kernel side addressing changes.
  90 *              Alan Cox        :       Beginning work on TCP fastpathing
  91 *                                      (not yet usable)
  92 *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
  93 *              Alan Cox        :       TCP fast path debugging
  94 *              Alan Cox        :       Window clamping
  95 *              Michael Riepe   :       Bug in tcp_check()
  96 *              Matt Dillon     :       More TCP improvements and RST bug fixes
  97 *              Matt Dillon     :       Yet more small nasties remove from the
  98 *                                      TCP code (Be very nice to this man if
  99 *                                      tcp finally works 100%) 8)
 100 *              Alan Cox        :       BSD accept semantics.
 101 *              Alan Cox        :       Reset on closedown bug.
 102 *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
 103 *              Michael Pall    :       Handle poll() after URG properly in
 104 *                                      all cases.
 105 *              Michael Pall    :       Undo the last fix in tcp_read_urg()
 106 *                                      (multi URG PUSH broke rlogin).
 107 *              Michael Pall    :       Fix the multi URG PUSH problem in
 108 *                                      tcp_readable(), poll() after URG
 109 *                                      works now.
 110 *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
 111 *                                      BSD api.
 112 *              Alan Cox        :       Changed the semantics of sk->socket to
 113 *                                      fix a race and a signal problem with
 114 *                                      accept() and async I/O.
 115 *              Alan Cox        :       Relaxed the rules on tcp_sendto().
 116 *              Yury Shevchuk   :       Really fixed accept() blocking problem.
 117 *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
 118 *                                      clients/servers which listen in on
 119 *                                      fixed ports.
 120 *              Alan Cox        :       Cleaned the above up and shrank it to
 121 *                                      a sensible code size.
 122 *              Alan Cox        :       Self connect lockup fix.
 123 *              Alan Cox        :       No connect to multicast.
 124 *              Ross Biro       :       Close unaccepted children on master
 125 *                                      socket close.
 126 *              Alan Cox        :       Reset tracing code.
 127 *              Alan Cox        :       Spurious resets on shutdown.
 128 *              Alan Cox        :       Giant 15 minute/60 second timer error
 129 *              Alan Cox        :       Small whoops in polling before an
 130 *                                      accept.
 131 *              Alan Cox        :       Kept the state trace facility since
 132 *                                      it's handy for debugging.
 133 *              Alan Cox        :       More reset handler fixes.
 134 *              Alan Cox        :       Started rewriting the code based on
 135 *                                      the RFC's for other useful protocol
 136 *                                      references see: Comer, KA9Q NOS, and
 137 *                                      for a reference on the difference
 138 *                                      between specifications and how BSD
 139 *                                      works see the 4.4lite source.
 140 *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
 141 *                                      close.
 142 *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
 143 *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
 144 *              Alan Cox        :       Reimplemented timers as per the RFC
 145 *                                      and using multiple timers for sanity.
 146 *              Alan Cox        :       Small bug fixes, and a lot of new
 147 *                                      comments.
 148 *              Alan Cox        :       Fixed dual reader crash by locking
 149 *                                      the buffers (much like datagram.c)
 150 *              Alan Cox        :       Fixed stuck sockets in probe. A probe
 151 *                                      now gets fed up of retrying without
 152 *                                      (even a no space) answer.
 153 *              Alan Cox        :       Extracted closing code better
 154 *              Alan Cox        :       Fixed the closing state machine to
 155 *                                      resemble the RFC.
 156 *              Alan Cox        :       More 'per spec' fixes.
 157 *              Jorge Cwik      :       Even faster checksumming.
 158 *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
 159 *                                      only frames. At least one pc tcp stack
 160 *                                      generates them.
 161 *              Alan Cox        :       Cache last socket.
 162 *              Alan Cox        :       Per route irtt.
 163 *              Matt Day        :       poll()->select() match BSD precisely on error
 164 *              Alan Cox        :       New buffers
 165 *              Marc Tamsky     :       Various sk->prot->retransmits and
 166 *                                      sk->retransmits misupdating fixed.
 167 *                                      Fixed tcp_write_timeout: stuck close,
 168 *                                      and TCP syn retries gets used now.
 169 *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
 170 *                                      ack if state is TCP_CLOSED.
 171 *              Alan Cox        :       Look up device on a retransmit - routes may
 172 *                                      change. Doesn't yet cope with MSS shrink right
 173 *                                      but it's a start!
 174 *              Marc Tamsky     :       Closing in closing fixes.
 175 *              Mike Shaver     :       RFC1122 verifications.
 176 *              Alan Cox        :       rcv_saddr errors.
 177 *              Alan Cox        :       Block double connect().
 178 *              Alan Cox        :       Small hooks for enSKIP.
 179 *              Alexey Kuznetsov:       Path MTU discovery.
 180 *              Alan Cox        :       Support soft errors.
 181 *              Alan Cox        :       Fix MTU discovery pathological case
 182 *                                      when the remote claims no mtu!
 183 *              Marc Tamsky     :       TCP_CLOSE fix.
 184 *              Colin (G3TNE)   :       Send a reset on syn ack replies in
 185 *                                      window but wrong (fixes NT lpd problems)
 186 *              Pedro Roque     :       Better TCP window handling, delayed ack.
 187 *              Joerg Reuter    :       No modification of locked buffers in
 188 *                                      tcp_do_retransmit()
 189 *              Eric Schenk     :       Changed receiver side silly window
 190 *                                      avoidance algorithm to BSD style
 191 *                                      algorithm. This doubles throughput
 192 *                                      against machines running Solaris,
 193 *                                      and seems to result in general
 194 *                                      improvement.
 195 *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
 196 *      Willy Konynenberg       :       Transparent proxying support.
 197 *      Mike McLagan            :       Routing by source
 198 *              Keith Owens     :       Do proper merging with partial SKB's in
 199 *                                      tcp_do_sendmsg to avoid burstiness.
 200 *              Eric Schenk     :       Fix fast close down bug with
 201 *                                      shutdown() followed by close().
 202 *              Andi Kleen      :       Make poll agree with SIGIO
 203 *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
 204 *                                      lingertime == 0 (RFC 793 ABORT Call)
 205 *      Hirokazu Takahashi      :       Use copy_from_user() instead of
 206 *                                      csum_and_copy_from_user() if possible.
 207 *
 208 *              This program is free software; you can redistribute it and/or
 209 *              modify it under the terms of the GNU General Public License
 210 *              as published by the Free Software Foundation; either version
 211 *              2 of the License, or(at your option) any later version.
 212 *
 213 * Description of States:
 214 *
 215 *      TCP_SYN_SENT            sent a connection request, waiting for ack
 216 *
 217 *      TCP_SYN_RECV            received a connection request, sent ack,
 218 *                              waiting for final ack in three-way handshake.
 219 *
 220 *      TCP_ESTABLISHED         connection established
 221 *
 222 *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
 223 *                              transmission of remaining buffered data
 224 *
 225 *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
 226 *                              to shutdown
 227 *
 228 *      TCP_CLOSING             both sides have shutdown but we still have
 229 *                              data we have to finish sending
 230 *
 231 *      TCP_TIME_WAIT           timeout to catch resent junk before entering
 232 *                              closed, can only be entered from FIN_WAIT2
 233 *                              or CLOSING.  Required because the other end
 234 *                              may not have gotten our last ACK causing it
 235 *                              to retransmit the data packet (which we ignore)
 236 *
 237 *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
 238 *                              us to finish writing our data and to shutdown
 239 *                              (we have to close() to move on to LAST_ACK)
 240 *
 241 *      TCP_LAST_ACK            out side has shutdown after remote has
 242 *                              shutdown.  There may still be data in our
 243 *                              buffer that we have to finish sending
 244 *
 245 *      TCP_CLOSE               socket is finished
 246 */
 247
 248#define pr_fmt(fmt) "TCP: " fmt
 249
 250#include <linux/kernel.h>
 251#include <linux/module.h>
 252#include <linux/types.h>
 253#include <linux/fcntl.h>
 254#include <linux/poll.h>
 255#include <linux/init.h>
 256#include <linux/fs.h>
 257#include <linux/skbuff.h>
 258#include <linux/scatterlist.h>
 259#include <linux/splice.h>
 260#include <linux/net.h>
 261#include <linux/socket.h>
 262#include <linux/random.h>
 263#include <linux/bootmem.h>
 264#include <linux/highmem.h>
 265#include <linux/swap.h>
 266#include <linux/cache.h>
 267#include <linux/err.h>
 268#include <linux/crypto.h>
 269#include <linux/time.h>
 270#include <linux/slab.h>
 271
 272#include <net/icmp.h>
 273#include <net/inet_common.h>
 274#include <net/tcp.h>
 275#include <net/xfrm.h>
 276#include <net/ip.h>
 277#include <net/netdma.h>
 278#include <net/sock.h>
 279
 280#include <asm/uaccess.h>
 281#include <asm/ioctls.h>
 282#include <net/busy_poll.h>
 283
 284int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
 285
 286int sysctl_tcp_min_tso_segs __read_mostly = 2;
 287
 288int sysctl_tcp_autocorking __read_mostly = 1;
 289
 290struct percpu_counter tcp_orphan_count;
 291EXPORT_SYMBOL_GPL(tcp_orphan_count);
 292
 293long sysctl_tcp_mem[3] __read_mostly;
 294int sysctl_tcp_wmem[3] __read_mostly;
 295int sysctl_tcp_rmem[3] __read_mostly;
 296
 297EXPORT_SYMBOL(sysctl_tcp_mem);
 298EXPORT_SYMBOL(sysctl_tcp_rmem);
 299EXPORT_SYMBOL(sysctl_tcp_wmem);
 300
 301atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
 302EXPORT_SYMBOL(tcp_memory_allocated);
 303
 304/*
 305 * Current number of TCP sockets.
 306 */
 307struct percpu_counter tcp_sockets_allocated;
 308EXPORT_SYMBOL(tcp_sockets_allocated);
 309
 310/*
 311 * TCP splice context
 312 */
 313struct tcp_splice_state {
 314        struct pipe_inode_info *pipe;
 315        size_t len;
 316        unsigned int flags;
 317};
 318
 319/*
 320 * Pressure flag: try to collapse.
 321 * Technical note: it is used by multiple contexts non atomically.
 322 * All the __sk_mem_schedule() is of this nature: accounting
 323 * is strict, actions are advisory and have some latency.
 324 */
 325int tcp_memory_pressure __read_mostly;
 326EXPORT_SYMBOL(tcp_memory_pressure);
 327
 328void tcp_enter_memory_pressure(struct sock *sk)
 329{
 330        if (!tcp_memory_pressure) {
 331                NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
 332                tcp_memory_pressure = 1;
 333        }
 334}
 335EXPORT_SYMBOL(tcp_enter_memory_pressure);
 336
 337/* Convert seconds to retransmits based on initial and max timeout */
 338static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
 339{
 340        u8 res = 0;
 341
 342        if (seconds > 0) {
 343                int period = timeout;
 344
 345                res = 1;
 346                while (seconds > period && res < 255) {
 347                        res++;
 348                        timeout <<= 1;
 349                        if (timeout > rto_max)
 350                                timeout = rto_max;
 351                        period += timeout;
 352                }
 353        }
 354        return res;
 355}
 356
 357/* Convert retransmits to seconds based on initial and max timeout */
 358static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
 359{
 360        int period = 0;
 361
 362        if (retrans > 0) {
 363                period = timeout;
 364                while (--retrans) {
 365                        timeout <<= 1;
 366                        if (timeout > rto_max)
 367                                timeout = rto_max;
 368                        period += timeout;
 369                }
 370        }
 371        return period;
 372}
 373
 374/* Address-family independent initialization for a tcp_sock.
 375 *
 376 * NOTE: A lot of things set to zero explicitly by call to
 377 *       sk_alloc() so need not be done here.
 378 */
 379void tcp_init_sock(struct sock *sk)
 380{
 381        struct inet_connection_sock *icsk = inet_csk(sk);
 382        struct tcp_sock *tp = tcp_sk(sk);
 383
 384        __skb_queue_head_init(&tp->out_of_order_queue);
 385        tcp_init_xmit_timers(sk);
 386        tcp_prequeue_init(tp);
 387        INIT_LIST_HEAD(&tp->tsq_node);
 388
 389        icsk->icsk_rto = TCP_TIMEOUT_INIT;
 390        tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
 391
 392        /* So many TCP implementations out there (incorrectly) count the
 393         * initial SYN frame in their delayed-ACK and congestion control
 394         * algorithms that we must have the following bandaid to talk
 395         * efficiently to them.  -DaveM
 396         */
 397        tp->snd_cwnd = TCP_INIT_CWND;
 398
 399        /* See draft-stevens-tcpca-spec-01 for discussion of the
 400         * initialization of these values.
 401         */
 402        tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 403        tp->snd_cwnd_clamp = ~0;
 404        tp->mss_cache = TCP_MSS_DEFAULT;
 405
 406        tp->reordering = sysctl_tcp_reordering;
 407        tcp_enable_early_retrans(tp);
 408        icsk->icsk_ca_ops = &tcp_init_congestion_ops;
 409
 410        tp->tsoffset = 0;
 411
 412        sk->sk_state = TCP_CLOSE;
 413
 414        sk->sk_write_space = sk_stream_write_space;
 415        sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
 416
 417        icsk->icsk_sync_mss = tcp_sync_mss;
 418
 419        sk->sk_sndbuf = sysctl_tcp_wmem[1];
 420        sk->sk_rcvbuf = sysctl_tcp_rmem[1];
 421
 422        local_bh_disable();
 423        sock_update_memcg(sk);
 424        sk_sockets_allocated_inc(sk);
 425        local_bh_enable();
 426}
 427EXPORT_SYMBOL(tcp_init_sock);
 428
 429/*
 430 *      Wait for a TCP event.
 431 *
 432 *      Note that we don't need to lock the socket, as the upper poll layers
 433 *      take care of normal races (between the test and the event) and we don't
 434 *      go look at any of the socket buffers directly.
 435 */
 436unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 437{
 438        unsigned int mask;
 439        struct sock *sk = sock->sk;
 440        const struct tcp_sock *tp = tcp_sk(sk);
 441
 442        sock_rps_record_flow(sk);
 443
 444        sock_poll_wait(file, sk_sleep(sk), wait);
 445        if (sk->sk_state == TCP_LISTEN)
 446                return inet_csk_listen_poll(sk);
 447
 448        /* Socket is not locked. We are protected from async events
 449         * by poll logic and correct handling of state changes
 450         * made by other threads is impossible in any case.
 451         */
 452
 453        mask = 0;
 454
 455        /*
 456         * POLLHUP is certainly not done right. But poll() doesn't
 457         * have a notion of HUP in just one direction, and for a
 458         * socket the read side is more interesting.
 459         *
 460         * Some poll() documentation says that POLLHUP is incompatible
 461         * with the POLLOUT/POLLWR flags, so somebody should check this
 462         * all. But careful, it tends to be safer to return too many
 463         * bits than too few, and you can easily break real applications
 464         * if you don't tell them that something has hung up!
 465         *
 466         * Check-me.
 467         *
 468         * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
 469         * our fs/select.c). It means that after we received EOF,
 470         * poll always returns immediately, making impossible poll() on write()
 471         * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
 472         * if and only if shutdown has been made in both directions.
 473         * Actually, it is interesting to look how Solaris and DUX
 474         * solve this dilemma. I would prefer, if POLLHUP were maskable,
 475         * then we could set it on SND_SHUTDOWN. BTW examples given
 476         * in Stevens' books assume exactly this behaviour, it explains
 477         * why POLLHUP is incompatible with POLLOUT.    --ANK
 478         *
 479         * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 480         * blocking on fresh not-connected or disconnected socket. --ANK
 481         */
 482        if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
 483                mask |= POLLHUP;
 484        if (sk->sk_shutdown & RCV_SHUTDOWN)
 485                mask |= POLLIN | POLLRDNORM | POLLRDHUP;
 486
 487        /* Connected or passive Fast Open socket? */
 488        if (sk->sk_state != TCP_SYN_SENT &&
 489            (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk != NULL)) {
 490                int target = sock_rcvlowat(sk, 0, INT_MAX);
 491
 492                if (tp->urg_seq == tp->copied_seq &&
 493                    !sock_flag(sk, SOCK_URGINLINE) &&
 494                    tp->urg_data)
 495                        target++;
 496
 497                /* Potential race condition. If read of tp below will
 498                 * escape above sk->sk_state, we can be illegally awaken
 499                 * in SYN_* states. */
 500                if (tp->rcv_nxt - tp->copied_seq >= target)
 501                        mask |= POLLIN | POLLRDNORM;
 502
 503                if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 504                        if (sk_stream_is_writeable(sk)) {
 505                                mask |= POLLOUT | POLLWRNORM;
 506                        } else {  /* send SIGIO later */
 507                                set_bit(SOCK_ASYNC_NOSPACE,
 508                                        &sk->sk_socket->flags);
 509                                set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 510
 511                                /* Race breaker. If space is freed after
 512                                 * wspace test but before the flags are set,
 513                                 * IO signal will be lost.
 514                                 */
 515                                if (sk_stream_is_writeable(sk))
 516                                        mask |= POLLOUT | POLLWRNORM;
 517                        }
 518                } else
 519                        mask |= POLLOUT | POLLWRNORM;
 520
 521                if (tp->urg_data & TCP_URG_VALID)
 522                        mask |= POLLPRI;
 523        }
 524        /* This barrier is coupled with smp_wmb() in tcp_reset() */
 525        smp_rmb();
 526        if (sk->sk_err)
 527                mask |= POLLERR;
 528
 529        return mask;
 530}
 531EXPORT_SYMBOL(tcp_poll);
 532
 533int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 534{
 535        struct tcp_sock *tp = tcp_sk(sk);
 536        int answ;
 537        bool slow;
 538
 539        switch (cmd) {
 540        case SIOCINQ:
 541                if (sk->sk_state == TCP_LISTEN)
 542                        return -EINVAL;
 543
 544                slow = lock_sock_fast(sk);
 545                if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 546                        answ = 0;
 547                else if (sock_flag(sk, SOCK_URGINLINE) ||
 548                         !tp->urg_data ||
 549                         before(tp->urg_seq, tp->copied_seq) ||
 550                         !before(tp->urg_seq, tp->rcv_nxt)) {
 551
 552                        answ = tp->rcv_nxt - tp->copied_seq;
 553
 554                        /* Subtract 1, if FIN was received */
 555                        if (answ && sock_flag(sk, SOCK_DONE))
 556                                answ--;
 557                } else
 558                        answ = tp->urg_seq - tp->copied_seq;
 559                unlock_sock_fast(sk, slow);
 560                break;
 561        case SIOCATMARK:
 562                answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 563                break;
 564        case SIOCOUTQ:
 565                if (sk->sk_state == TCP_LISTEN)
 566                        return -EINVAL;
 567
 568                if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 569                        answ = 0;
 570                else
 571                        answ = tp->write_seq - tp->snd_una;
 572                break;
 573        case SIOCOUTQNSD:
 574                if (sk->sk_state == TCP_LISTEN)
 575                        return -EINVAL;
 576
 577                if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 578                        answ = 0;
 579                else
 580                        answ = tp->write_seq - tp->snd_nxt;
 581                break;
 582        default:
 583                return -ENOIOCTLCMD;
 584        }
 585
 586        return put_user(answ, (int __user *)arg);
 587}
 588EXPORT_SYMBOL(tcp_ioctl);
 589
 590static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 591{
 592        TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 593        tp->pushed_seq = tp->write_seq;
 594}
 595
 596static inline bool forced_push(const struct tcp_sock *tp)
 597{
 598        return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 599}
 600
 601static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
 602{
 603        struct tcp_sock *tp = tcp_sk(sk);
 604        struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 605
 606        skb->csum    = 0;
 607        tcb->seq     = tcb->end_seq = tp->write_seq;
 608        tcb->tcp_flags = TCPHDR_ACK;
 609        tcb->sacked  = 0;
 610        skb_header_release(skb);
 611        tcp_add_write_queue_tail(sk, skb);
 612        sk->sk_wmem_queued += skb->truesize;
 613        sk_mem_charge(sk, skb->truesize);
 614        if (tp->nonagle & TCP_NAGLE_PUSH)
 615                tp->nonagle &= ~TCP_NAGLE_PUSH;
 616}
 617
 618static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
 619{
 620        if (flags & MSG_OOB)
 621                tp->snd_up = tp->write_seq;
 622}
 623
 624/* If a not yet filled skb is pushed, do not send it if
 625 * we have data packets in Qdisc or NIC queues :
 626 * Because TX completion will happen shortly, it gives a chance
 627 * to coalesce future sendmsg() payload into this skb, without
 628 * need for a timer, and with no latency trade off.
 629 * As packets containing data payload have a bigger truesize
 630 * than pure acks (dataless) packets, the last checks prevent
 631 * autocorking if we only have an ACK in Qdisc/NIC queues,
 632 * or if TX completion was delayed after we processed ACK packet.
 633 */
 634static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
 635                                int size_goal)
 636{
 637        return skb->len < size_goal &&
 638               sysctl_tcp_autocorking &&
 639               skb != tcp_write_queue_head(sk) &&
 640               atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
 641}
 642
 643static void tcp_push(struct sock *sk, int flags, int mss_now,
 644                     int nonagle, int size_goal)
 645{
 646        struct tcp_sock *tp = tcp_sk(sk);
 647        struct sk_buff *skb;
 648
 649        if (!tcp_send_head(sk))
 650                return;
 651
 652        skb = tcp_write_queue_tail(sk);
 653        if (!(flags & MSG_MORE) || forced_push(tp))
 654                tcp_mark_push(tp, skb);
 655
 656        tcp_mark_urg(tp, flags);
 657
 658        if (tcp_should_autocork(sk, skb, size_goal)) {
 659
 660                /* avoid atomic op if TSQ_THROTTLED bit is already set */
 661                if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
 662                        NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
 663                        set_bit(TSQ_THROTTLED, &tp->tsq_flags);
 664                }
 665                /* It is possible TX completion already happened
 666                 * before we set TSQ_THROTTLED.
 667                 */
 668                if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
 669                        return;
 670        }
 671
 672        if (flags & MSG_MORE)
 673                nonagle = TCP_NAGLE_CORK;
 674
 675        __tcp_push_pending_frames(sk, mss_now, nonagle);
 676}
 677
 678static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 679                                unsigned int offset, size_t len)
 680{
 681        struct tcp_splice_state *tss = rd_desc->arg.data;
 682        int ret;
 683
 684        ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
 685                              tss->flags);
 686        if (ret > 0)
 687                rd_desc->count -= ret;
 688        return ret;
 689}
 690
 691static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 692{
 693        /* Store TCP splice context information in read_descriptor_t. */
 694        read_descriptor_t rd_desc = {
 695                .arg.data = tss,
 696                .count    = tss->len,
 697        };
 698
 699        return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 700}
 701
 702/**
 703 *  tcp_splice_read - splice data from TCP socket to a pipe
 704 * @sock:       socket to splice from
 705 * @ppos:       position (not valid)
 706 * @pipe:       pipe to splice to
 707 * @len:        number of bytes to splice
 708 * @flags:      splice modifier flags
 709 *
 710 * Description:
 711 *    Will read pages from given socket and fill them into a pipe.
 712 *
 713 **/
 714ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 715                        struct pipe_inode_info *pipe, size_t len,
 716                        unsigned int flags)
 717{
 718        struct sock *sk = sock->sk;
 719        struct tcp_splice_state tss = {
 720                .pipe = pipe,
 721                .len = len,
 722                .flags = flags,
 723        };
 724        long timeo;
 725        ssize_t spliced;
 726        int ret;
 727
 728        sock_rps_record_flow(sk);
 729        /*
 730         * We can't seek on a socket input
 731         */
 732        if (unlikely(*ppos))
 733                return -ESPIPE;
 734
 735        ret = spliced = 0;
 736
 737        lock_sock(sk);
 738
 739        timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 740        while (tss.len) {
 741                ret = __tcp_splice_read(sk, &tss);
 742                if (ret < 0)
 743                        break;
 744                else if (!ret) {
 745                        if (spliced)
 746                                break;
 747                        if (sock_flag(sk, SOCK_DONE))
 748                                break;
 749                        if (sk->sk_err) {
 750                                ret = sock_error(sk);
 751                                break;
 752                        }
 753                        if (sk->sk_shutdown & RCV_SHUTDOWN)
 754                                break;
 755                        if (sk->sk_state == TCP_CLOSE) {
 756                                /*
 757                                 * This occurs when user tries to read
 758                                 * from never connected socket.
 759                                 */
 760                                if (!sock_flag(sk, SOCK_DONE))
 761                                        ret = -ENOTCONN;
 762                                break;
 763                        }
 764                        if (!timeo) {
 765                                ret = -EAGAIN;
 766                                break;
 767                        }
 768                        sk_wait_data(sk, &timeo);
 769                        if (signal_pending(current)) {
 770                                ret = sock_intr_errno(timeo);
 771                                break;
 772                        }
 773                        continue;
 774                }
 775                tss.len -= ret;
 776                spliced += ret;
 777
 778                if (!timeo)
 779                        break;
 780                release_sock(sk);
 781                lock_sock(sk);
 782
 783                if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 784                    (sk->sk_shutdown & RCV_SHUTDOWN) ||
 785                    signal_pending(current))
 786                        break;
 787        }
 788
 789        release_sock(sk);
 790
 791        if (spliced)
 792                return spliced;
 793
 794        return ret;
 795}
 796EXPORT_SYMBOL(tcp_splice_read);
 797
 798struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
 799{
 800        struct sk_buff *skb;
 801
 802        /* The TCP header must be at least 32-bit aligned.  */
 803        size = ALIGN(size, 4);
 804
 805        skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 806        if (skb) {
 807                if (sk_wmem_schedule(sk, skb->truesize)) {
 808                        skb_reserve(skb, sk->sk_prot->max_header);
 809                        /*
 810                         * Make sure that we have exactly size bytes
 811                         * available to the caller, no more, no less.
 812                         */
 813                        skb->reserved_tailroom = skb->end - skb->tail - size;
 814                        return skb;
 815                }
 816                __kfree_skb(skb);
 817        } else {
 818                sk->sk_prot->enter_memory_pressure(sk);
 819                sk_stream_moderate_sndbuf(sk);
 820        }
 821        return NULL;
 822}
 823
 824static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 825                                       int large_allowed)
 826{
 827        struct tcp_sock *tp = tcp_sk(sk);
 828        u32 xmit_size_goal, old_size_goal;
 829
 830        xmit_size_goal = mss_now;
 831
 832        if (large_allowed && sk_can_gso(sk)) {
 833                u32 gso_size, hlen;
 834
 835                /* Maybe we should/could use sk->sk_prot->max_header here ? */
 836                hlen = inet_csk(sk)->icsk_af_ops->net_header_len +
 837                       inet_csk(sk)->icsk_ext_hdr_len +
 838                       tp->tcp_header_len;
 839
 840                /* Goal is to send at least one packet per ms,
 841                 * not one big TSO packet every 100 ms.
 842                 * This preserves ACK clocking and is consistent
 843                 * with tcp_tso_should_defer() heuristic.
 844                 */
 845                gso_size = sk->sk_pacing_rate / (2 * MSEC_PER_SEC);
 846                gso_size = max_t(u32, gso_size,
 847                                 sysctl_tcp_min_tso_segs * mss_now);
 848
 849                xmit_size_goal = min_t(u32, gso_size,
 850                                       sk->sk_gso_max_size - 1 - hlen);
 851
 852                xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
 853
 854                /* We try hard to avoid divides here */
 855                old_size_goal = tp->xmit_size_goal_segs * mss_now;
 856
 857                if (likely(old_size_goal <= xmit_size_goal &&
 858                           old_size_goal + mss_now > xmit_size_goal)) {
 859                        xmit_size_goal = old_size_goal;
 860                } else {
 861                        tp->xmit_size_goal_segs =
 862                                min_t(u16, xmit_size_goal / mss_now,
 863                                      sk->sk_gso_max_segs);
 864                        xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
 865                }
 866        }
 867
 868        return max(xmit_size_goal, mss_now);
 869}
 870
 871static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 872{
 873        int mss_now;
 874
 875        mss_now = tcp_current_mss(sk);
 876        *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 877
 878        return mss_now;
 879}
 880
 881static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
 882                                size_t size, int flags)
 883{
 884        struct tcp_sock *tp = tcp_sk(sk);
 885        int mss_now, size_goal;
 886        int err;
 887        ssize_t copied;
 888        long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 889
 890        /* Wait for a connection to finish. One exception is TCP Fast Open
 891         * (passive side) where data is allowed to be sent before a connection
 892         * is fully established.
 893         */
 894        if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
 895            !tcp_passive_fastopen(sk)) {
 896                if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 897                        goto out_err;
 898        }
 899
 900        clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 901
 902        mss_now = tcp_send_mss(sk, &size_goal, flags);
 903        copied = 0;
 904
 905        err = -EPIPE;
 906        if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 907                goto out_err;
 908
 909        while (size > 0) {
 910                struct sk_buff *skb = tcp_write_queue_tail(sk);
 911                int copy, i;
 912                bool can_coalesce;
 913
 914                if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
 915new_segment:
 916                        if (!sk_stream_memory_free(sk))
 917                                goto wait_for_sndbuf;
 918
 919                        skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
 920                        if (!skb)
 921                                goto wait_for_memory;
 922
 923                        skb_entail(sk, skb);
 924                        copy = size_goal;
 925                }
 926
 927                if (copy > size)
 928                        copy = size;
 929
 930                i = skb_shinfo(skb)->nr_frags;
 931                can_coalesce = skb_can_coalesce(skb, i, page, offset);
 932                if (!can_coalesce && i >= MAX_SKB_FRAGS) {
 933                        tcp_mark_push(tp, skb);
 934                        goto new_segment;
 935                }
 936                if (!sk_wmem_schedule(sk, copy))
 937                        goto wait_for_memory;
 938
 939                if (can_coalesce) {
 940                        skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
 941                } else {
 942                        get_page(page);
 943                        skb_fill_page_desc(skb, i, page, offset, copy);
 944                }
 945                skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
 946
 947                skb->len += copy;
 948                skb->data_len += copy;
 949                skb->truesize += copy;
 950                sk->sk_wmem_queued += copy;
 951                sk_mem_charge(sk, copy);
 952                skb->ip_summed = CHECKSUM_PARTIAL;
 953                tp->write_seq += copy;
 954                TCP_SKB_CB(skb)->end_seq += copy;
 955                skb_shinfo(skb)->gso_segs = 0;
 956
 957                if (!copied)
 958                        TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
 959
 960                copied += copy;
 961                offset += copy;
 962                if (!(size -= copy))
 963                        goto out;
 964
 965                if (skb->len < size_goal || (flags & MSG_OOB))
 966                        continue;
 967
 968                if (forced_push(tp)) {
 969                        tcp_mark_push(tp, skb);
 970                        __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
 971                } else if (skb == tcp_send_head(sk))
 972                        tcp_push_one(sk, mss_now);
 973                continue;
 974
 975wait_for_sndbuf:
 976                set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 977wait_for_memory:
 978                tcp_push(sk, flags & ~MSG_MORE, mss_now,
 979                         TCP_NAGLE_PUSH, size_goal);
 980
 981                if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 982                        goto do_error;
 983
 984                mss_now = tcp_send_mss(sk, &size_goal, flags);
 985        }
 986
 987out:
 988        if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
 989                tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
 990        return copied;
 991
 992do_error:
 993        if (copied)
 994                goto out;
 995out_err:
 996        return sk_stream_error(sk, flags, err);
 997}
 998
 999int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1000                 size_t size, int flags)
1001{
1002        ssize_t res;
1003
1004        if (!(sk->sk_route_caps & NETIF_F_SG) ||
1005            !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
1006                return sock_no_sendpage(sk->sk_socket, page, offset, size,
1007                                        flags);
1008
1009        lock_sock(sk);
1010        res = do_tcp_sendpages(sk, page, offset, size, flags);
1011        release_sock(sk);
1012        return res;
1013}
1014EXPORT_SYMBOL(tcp_sendpage);
1015
1016static inline int select_size(const struct sock *sk, bool sg)
1017{
1018        const struct tcp_sock *tp = tcp_sk(sk);
1019        int tmp = tp->mss_cache;
1020
1021        if (sg) {
1022                if (sk_can_gso(sk)) {
1023                        /* Small frames wont use a full page:
1024                         * Payload will immediately follow tcp header.
1025                         */
1026                        tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1027                } else {
1028                        int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1029
1030                        if (tmp >= pgbreak &&
1031                            tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1032                                tmp = pgbreak;
1033                }
1034        }
1035
1036        return tmp;
1037}
1038
1039void tcp_free_fastopen_req(struct tcp_sock *tp)
1040{
1041        if (tp->fastopen_req != NULL) {
1042                kfree(tp->fastopen_req);
1043                tp->fastopen_req = NULL;
1044        }
1045}
1046
1047static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1048                                int *copied, size_t size)
1049{
1050        struct tcp_sock *tp = tcp_sk(sk);
1051        int err, flags;
1052
1053        if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1054                return -EOPNOTSUPP;
1055        if (tp->fastopen_req != NULL)
1056                return -EALREADY; /* Another Fast Open is in progress */
1057
1058        tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1059                                   sk->sk_allocation);
1060        if (unlikely(tp->fastopen_req == NULL))
1061                return -ENOBUFS;
1062        tp->fastopen_req->data = msg;
1063        tp->fastopen_req->size = size;
1064
1065        flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1066        err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1067                                    msg->msg_namelen, flags);
1068        *copied = tp->fastopen_req->copied;
1069        tcp_free_fastopen_req(tp);
1070        return err;
1071}
1072
1073int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1074                size_t size)
1075{
1076        struct iovec *iov;
1077        struct tcp_sock *tp = tcp_sk(sk);
1078        struct sk_buff *skb;
1079        int iovlen, flags, err, copied = 0;
1080        int mss_now = 0, size_goal, copied_syn = 0, offset = 0;
1081        bool sg;
1082        long timeo;
1083
1084        lock_sock(sk);
1085
1086        flags = msg->msg_flags;
1087        if (flags & MSG_FASTOPEN) {
1088                err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1089                if (err == -EINPROGRESS && copied_syn > 0)
1090                        goto out;
1091                else if (err)
1092                        goto out_err;
1093                offset = copied_syn;
1094        }
1095
1096        timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1097
1098        /* Wait for a connection to finish. One exception is TCP Fast Open
1099         * (passive side) where data is allowed to be sent before a connection
1100         * is fully established.
1101         */
1102        if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1103            !tcp_passive_fastopen(sk)) {
1104                if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1105                        goto do_error;
1106        }
1107
1108        if (unlikely(tp->repair)) {
1109                if (tp->repair_queue == TCP_RECV_QUEUE) {
1110                        copied = tcp_send_rcvq(sk, msg, size);
1111                        goto out_nopush;
1112                }
1113
1114                err = -EINVAL;
1115                if (tp->repair_queue == TCP_NO_QUEUE)
1116                        goto out_err;
1117
1118                /* 'common' sending to sendq */
1119        }
1120
1121        /* This should be in poll */
1122        clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1123
1124        mss_now = tcp_send_mss(sk, &size_goal, flags);
1125
1126        /* Ok commence sending. */
1127        iovlen = msg->msg_iovlen;
1128        iov = msg->msg_iov;
1129        copied = 0;
1130
1131        err = -EPIPE;
1132        if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1133                goto out_err;
1134
1135        sg = !!(sk->sk_route_caps & NETIF_F_SG);
1136
1137        while (--iovlen >= 0) {
1138                size_t seglen = iov->iov_len;
1139                unsigned char __user *from = iov->iov_base;
1140
1141                iov++;
1142                if (unlikely(offset > 0)) {  /* Skip bytes copied in SYN */
1143                        if (offset >= seglen) {
1144                                offset -= seglen;
1145                                continue;
1146                        }
1147                        seglen -= offset;
1148                        from += offset;
1149                        offset = 0;
1150                }
1151
1152                while (seglen > 0) {
1153                        int copy = 0;
1154                        int max = size_goal;
1155
1156                        skb = tcp_write_queue_tail(sk);
1157                        if (tcp_send_head(sk)) {
1158                                if (skb->ip_summed == CHECKSUM_NONE)
1159                                        max = mss_now;
1160                                copy = max - skb->len;
1161                        }
1162
1163                        if (copy <= 0) {
1164new_segment:
1165                                /* Allocate new segment. If the interface is SG,
1166                                 * allocate skb fitting to single page.
1167                                 */
1168                                if (!sk_stream_memory_free(sk))
1169                                        goto wait_for_sndbuf;
1170
1171                                skb = sk_stream_alloc_skb(sk,
1172                                                          select_size(sk, sg),
1173                                                          sk->sk_allocation);
1174                                if (!skb)
1175                                        goto wait_for_memory;
1176
1177                                /*
1178                                 * All packets are restored as if they have
1179                                 * already been sent.
1180                                 */
1181                                if (tp->repair)
1182                                        TCP_SKB_CB(skb)->when = tcp_time_stamp;
1183
1184                                /*
1185                                 * Check whether we can use HW checksum.
1186                                 */
1187                                if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1188                                        skb->ip_summed = CHECKSUM_PARTIAL;
1189
1190                                skb_entail(sk, skb);
1191                                copy = size_goal;
1192                                max = size_goal;
1193                        }
1194
1195                        /* Try to append data to the end of skb. */
1196                        if (copy > seglen)
1197                                copy = seglen;
1198
1199                        /* Where to copy to? */
1200                        if (skb_availroom(skb) > 0) {
1201                                /* We have some space in skb head. Superb! */
1202                                copy = min_t(int, copy, skb_availroom(skb));
1203                                err = skb_add_data_nocache(sk, skb, from, copy);
1204                                if (err)
1205                                        goto do_fault;
1206                        } else {
1207                                bool merge = true;
1208                                int i = skb_shinfo(skb)->nr_frags;
1209                                struct page_frag *pfrag = sk_page_frag(sk);
1210
1211                                if (!sk_page_frag_refill(sk, pfrag))
1212                                        goto wait_for_memory;
1213
1214                                if (!skb_can_coalesce(skb, i, pfrag->page,
1215                                                      pfrag->offset)) {
1216                                        if (i == MAX_SKB_FRAGS || !sg) {
1217                                                tcp_mark_push(tp, skb);
1218                                                goto new_segment;
1219                                        }
1220                                        merge = false;
1221                                }
1222
1223                                copy = min_t(int, copy, pfrag->size - pfrag->offset);
1224
1225                                if (!sk_wmem_schedule(sk, copy))
1226                                        goto wait_for_memory;
1227
1228                                err = skb_copy_to_page_nocache(sk, from, skb,
1229                                                               pfrag->page,
1230                                                               pfrag->offset,
1231                                                               copy);
1232                                if (err)
1233                                        goto do_error;
1234
1235                                /* Update the skb. */
1236                                if (merge) {
1237                                        skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1238                                } else {
1239                                        skb_fill_page_desc(skb, i, pfrag->page,
1240                                                           pfrag->offset, copy);
1241                                        get_page(pfrag->page);
1242                                }
1243                                pfrag->offset += copy;
1244                        }
1245
1246                        if (!copied)
1247                                TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1248
1249                        tp->write_seq += copy;
1250                        TCP_SKB_CB(skb)->end_seq += copy;
1251                        skb_shinfo(skb)->gso_segs = 0;
1252
1253                        from += copy;
1254                        copied += copy;
1255                        if ((seglen -= copy) == 0 && iovlen == 0)
1256                                goto out;
1257
1258                        if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1259                                continue;
1260
1261                        if (forced_push(tp)) {
1262                                tcp_mark_push(tp, skb);
1263                                __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1264                        } else if (skb == tcp_send_head(sk))
1265                                tcp_push_one(sk, mss_now);
1266                        continue;
1267
1268wait_for_sndbuf:
1269                        set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1270wait_for_memory:
1271                        if (copied)
1272                                tcp_push(sk, flags & ~MSG_MORE, mss_now,
1273                                         TCP_NAGLE_PUSH, size_goal);
1274
1275                        if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1276                                goto do_error;
1277
1278                        mss_now = tcp_send_mss(sk, &size_goal, flags);
1279                }
1280        }
1281
1282out:
1283        if (copied)
1284                tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1285out_nopush:
1286        release_sock(sk);
1287        return copied + copied_syn;
1288
1289do_fault:
1290        if (!skb->len) {
1291                tcp_unlink_write_queue(skb, sk);
1292                /* It is the one place in all of TCP, except connection
1293                 * reset, where we can be unlinking the send_head.
1294                 */
1295                tcp_check_send_head(sk, skb);
1296                sk_wmem_free_skb(sk, skb);
1297        }
1298
1299do_error:
1300        if (copied + copied_syn)
1301                goto out;
1302out_err:
1303        err = sk_stream_error(sk, flags, err);
1304        release_sock(sk);
1305        return err;
1306}
1307EXPORT_SYMBOL(tcp_sendmsg);
1308
1309/*
1310 *      Handle reading urgent data. BSD has very simple semantics for
1311 *      this, no blocking and very strange errors 8)
1312 */
1313
1314static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1315{
1316        struct tcp_sock *tp = tcp_sk(sk);
1317
1318        /* No URG data to read. */
1319        if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1320            tp->urg_data == TCP_URG_READ)
1321                return -EINVAL; /* Yes this is right ! */
1322
1323        if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1324                return -ENOTCONN;
1325
1326        if (tp->urg_data & TCP_URG_VALID) {
1327                int err = 0;
1328                char c = tp->urg_data;
1329
1330                if (!(flags & MSG_PEEK))
1331                        tp->urg_data = TCP_URG_READ;
1332
1333                /* Read urgent data. */
1334                msg->msg_flags |= MSG_OOB;
1335
1336                if (len > 0) {
1337                        if (!(flags & MSG_TRUNC))
1338                                err = memcpy_toiovec(msg->msg_iov, &c, 1);
1339                        len = 1;
1340                } else
1341                        msg->msg_flags |= MSG_TRUNC;
1342
1343                return err ? -EFAULT : len;
1344        }
1345
1346        if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1347                return 0;
1348
1349        /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1350         * the available implementations agree in this case:
1351         * this call should never block, independent of the
1352         * blocking state of the socket.
1353         * Mike <pall@rz.uni-karlsruhe.de>
1354         */
1355        return -EAGAIN;
1356}
1357
1358static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1359{
1360        struct sk_buff *skb;
1361        int copied = 0, err = 0;
1362
1363        /* XXX -- need to support SO_PEEK_OFF */
1364
1365        skb_queue_walk(&sk->sk_write_queue, skb) {
1366                err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1367                if (err)
1368                        break;
1369
1370                copied += skb->len;
1371        }
1372
1373        return err ?: copied;
1374}
1375
1376/* Clean up the receive buffer for full frames taken by the user,
1377 * then send an ACK if necessary.  COPIED is the number of bytes
1378 * tcp_recvmsg has given to the user so far, it speeds up the
1379 * calculation of whether or not we must ACK for the sake of
1380 * a window update.
1381 */
1382void tcp_cleanup_rbuf(struct sock *sk, int copied)
1383{
1384        struct tcp_sock *tp = tcp_sk(sk);
1385        bool time_to_ack = false;
1386
1387        struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1388
1389        WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1390             "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1391             tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1392
1393        if (inet_csk_ack_scheduled(sk)) {
1394                const struct inet_connection_sock *icsk = inet_csk(sk);
1395                   /* Delayed ACKs frequently hit locked sockets during bulk
1396                    * receive. */
1397                if (icsk->icsk_ack.blocked ||
1398                    /* Once-per-two-segments ACK was not sent by tcp_input.c */
1399                    tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1400                    /*
1401                     * If this read emptied read buffer, we send ACK, if
1402                     * connection is not bidirectional, user drained
1403                     * receive buffer and there was a small segment
1404                     * in queue.
1405                     */
1406                    (copied > 0 &&
1407                     ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1408                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1409                       !icsk->icsk_ack.pingpong)) &&
1410                      !atomic_read(&sk->sk_rmem_alloc)))
1411                        time_to_ack = true;
1412        }
1413
1414        /* We send an ACK if we can now advertise a non-zero window
1415         * which has been raised "significantly".
1416         *
1417         * Even if window raised up to infinity, do not send window open ACK
1418         * in states, where we will not receive more. It is useless.
1419         */
1420        if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1421                __u32 rcv_window_now = tcp_receive_window(tp);
1422
1423                /* Optimize, __tcp_select_window() is not cheap. */
1424                if (2*rcv_window_now <= tp->window_clamp) {
1425                        __u32 new_window = __tcp_select_window(sk);
1426
1427                        /* Send ACK now, if this read freed lots of space
1428                         * in our buffer. Certainly, new_window is new window.
1429                         * We can advertise it now, if it is not less than current one.
1430                         * "Lots" means "at least twice" here.
1431                         */
1432                        if (new_window && new_window >= 2 * rcv_window_now)
1433                                time_to_ack = true;
1434                }
1435        }
1436        if (time_to_ack)
1437                tcp_send_ack(sk);
1438}
1439
1440static void tcp_prequeue_process(struct sock *sk)
1441{
1442        struct sk_buff *skb;
1443        struct tcp_sock *tp = tcp_sk(sk);
1444
1445        NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1446
1447        /* RX process wants to run with disabled BHs, though it is not
1448         * necessary */
1449        local_bh_disable();
1450        while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1451                sk_backlog_rcv(sk, skb);
1452        local_bh_enable();
1453
1454        /* Clear memory counter. */
1455        tp->ucopy.memory = 0;
1456}
1457
1458#ifdef CONFIG_NET_DMA
1459static void tcp_service_net_dma(struct sock *sk, bool wait)
1460{
1461        dma_cookie_t done, used;
1462        dma_cookie_t last_issued;
1463        struct tcp_sock *tp = tcp_sk(sk);
1464
1465        if (!tp->ucopy.dma_chan)
1466                return;
1467
1468        last_issued = tp->ucopy.dma_cookie;
1469        dma_async_issue_pending(tp->ucopy.dma_chan);
1470
1471        do {
1472                if (dma_async_is_tx_complete(tp->ucopy.dma_chan,
1473                                              last_issued, &done,
1474                                              &used) == DMA_COMPLETE) {
1475                        /* Safe to free early-copied skbs now */
1476                        __skb_queue_purge(&sk->sk_async_wait_queue);
1477                        break;
1478                } else {
1479                        struct sk_buff *skb;
1480                        while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1481                               (dma_async_is_complete(skb->dma_cookie, done,
1482                                                      used) == DMA_COMPLETE)) {
1483                                __skb_dequeue(&sk->sk_async_wait_queue);
1484                                kfree_skb(skb);
1485                        }
1486                }
1487        } while (wait);
1488}
1489#endif
1490
1491static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1492{
1493        struct sk_buff *skb;
1494        u32 offset;
1495
1496        while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1497                offset = seq - TCP_SKB_CB(skb)->seq;
1498                if (tcp_hdr(skb)->syn)
1499                        offset--;
1500                if (offset < skb->len || tcp_hdr(skb)->fin) {
1501                        *off = offset;
1502                        return skb;
1503                }
1504                /* This looks weird, but this can happen if TCP collapsing
1505                 * splitted a fat GRO packet, while we released socket lock
1506                 * in skb_splice_bits()
1507                 */
1508                sk_eat_skb(sk, skb, false);
1509        }
1510        return NULL;
1511}
1512
1513/*
1514 * This routine provides an alternative to tcp_recvmsg() for routines
1515 * that would like to handle copying from skbuffs directly in 'sendfile'
1516 * fashion.
1517 * Note:
1518 *      - It is assumed that the socket was locked by the caller.
1519 *      - The routine does not block.
1520 *      - At present, there is no support for reading OOB data
1521 *        or for 'peeking' the socket using this routine
1522 *        (although both would be easy to implement).
1523 */
1524int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1525                  sk_read_actor_t recv_actor)
1526{
1527        struct sk_buff *skb;
1528        struct tcp_sock *tp = tcp_sk(sk);
1529        u32 seq = tp->copied_seq;
1530        u32 offset;
1531        int copied = 0;
1532
1533        if (sk->sk_state == TCP_LISTEN)
1534                return -ENOTCONN;
1535        while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1536                if (offset < skb->len) {
1537                        int used;
1538                        size_t len;
1539
1540                        len = skb->len - offset;
1541                        /* Stop reading if we hit a patch of urgent data */
1542                        if (tp->urg_data) {
1543                                u32 urg_offset = tp->urg_seq - seq;
1544                                if (urg_offset < len)
1545                                        len = urg_offset;
1546                                if (!len)
1547                                        break;
1548                        }
1549                        used = recv_actor(desc, skb, offset, len);
1550                        if (used <= 0) {
1551                                if (!copied)
1552                                        copied = used;
1553                                break;
1554                        } else if (used <= len) {
1555                                seq += used;
1556                                copied += used;
1557                                offset += used;
1558                        }
1559                        /* If recv_actor drops the lock (e.g. TCP splice
1560                         * receive) the skb pointer might be invalid when
1561                         * getting here: tcp_collapse might have deleted it
1562                         * while aggregating skbs from the socket queue.
1563                         */
1564                        skb = tcp_recv_skb(sk, seq - 1, &offset);
1565                        if (!skb)
1566                                break;
1567                        /* TCP coalescing might have appended data to the skb.
1568                         * Try to splice more frags
1569                         */
1570                        if (offset + 1 != skb->len)
1571                                continue;
1572                }
1573                if (tcp_hdr(skb)->fin) {
1574                        sk_eat_skb(sk, skb, false);
1575                        ++seq;
1576                        break;
1577                }
1578                sk_eat_skb(sk, skb, false);
1579                if (!desc->count)
1580                        break;
1581                tp->copied_seq = seq;
1582        }
1583        tp->copied_seq = seq;
1584
1585        tcp_rcv_space_adjust(sk);
1586
1587        /* Clean up data we have read: This will do ACK frames. */
1588        if (copied > 0) {
1589                tcp_recv_skb(sk, seq, &offset);
1590                tcp_cleanup_rbuf(sk, copied);
1591        }
1592        return copied;
1593}
1594EXPORT_SYMBOL(tcp_read_sock);
1595
1596/*
1597 *      This routine copies from a sock struct into the user buffer.
1598 *
1599 *      Technical note: in 2.3 we work on _locked_ socket, so that
1600 *      tricks with *seq access order and skb->users are not required.
1601 *      Probably, code can be easily improved even more.
1602 */
1603
1604int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1605                size_t len, int nonblock, int flags, int *addr_len)
1606{
1607        struct tcp_sock *tp = tcp_sk(sk);
1608        int copied = 0;
1609        u32 peek_seq;
1610        u32 *seq;
1611        unsigned long used;
1612        int err;
1613        int target;             /* Read at least this many bytes */
1614        long timeo;
1615        struct task_struct *user_recv = NULL;
1616        bool copied_early = false;
1617        struct sk_buff *skb;
1618        u32 urg_hole = 0;
1619
1620        if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1621            (sk->sk_state == TCP_ESTABLISHED))
1622                sk_busy_loop(sk, nonblock);
1623
1624        lock_sock(sk);
1625
1626        err = -ENOTCONN;
1627        if (sk->sk_state == TCP_LISTEN)
1628                goto out;
1629
1630        timeo = sock_rcvtimeo(sk, nonblock);
1631
1632        /* Urgent data needs to be handled specially. */
1633        if (flags & MSG_OOB)
1634                goto recv_urg;
1635
1636        if (unlikely(tp->repair)) {
1637                err = -EPERM;
1638                if (!(flags & MSG_PEEK))
1639                        goto out;
1640
1641                if (tp->repair_queue == TCP_SEND_QUEUE)
1642                        goto recv_sndq;
1643
1644                err = -EINVAL;
1645                if (tp->repair_queue == TCP_NO_QUEUE)
1646                        goto out;
1647
1648                /* 'common' recv queue MSG_PEEK-ing */
1649        }
1650
1651        seq = &tp->copied_seq;
1652        if (flags & MSG_PEEK) {
1653                peek_seq = tp->copied_seq;
1654                seq = &peek_seq;
1655        }
1656
1657        target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1658
1659#ifdef CONFIG_NET_DMA
1660        tp->ucopy.dma_chan = NULL;
1661        preempt_disable();
1662        skb = skb_peek_tail(&sk->sk_receive_queue);
1663        {
1664                int available = 0;
1665
1666                if (skb)
1667                        available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1668                if ((available < target) &&
1669                    (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1670                    !sysctl_tcp_low_latency &&
1671                    net_dma_find_channel()) {
1672                        preempt_enable();
1673                        tp->ucopy.pinned_list =
1674                                        dma_pin_iovec_pages(msg->msg_iov, len);
1675                } else {
1676                        preempt_enable();
1677                }
1678        }
1679#endif
1680
1681        do {
1682                u32 offset;
1683
1684                /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1685                if (tp->urg_data && tp->urg_seq == *seq) {
1686                        if (copied)
1687                                break;
1688                        if (signal_pending(current)) {
1689                                copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1690                                break;
1691                        }
1692                }
1693
1694                /* Next get a buffer. */
1695
1696                skb_queue_walk(&sk->sk_receive_queue, skb) {
1697                        /* Now that we have two receive queues this
1698                         * shouldn't happen.
1699                         */
1700                        if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1701                                 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1702                                 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1703                                 flags))
1704                                break;
1705
1706                        offset = *seq - TCP_SKB_CB(skb)->seq;
1707                        if (tcp_hdr(skb)->syn)
1708                                offset--;
1709                        if (offset < skb->len)
1710                                goto found_ok_skb;
1711                        if (tcp_hdr(skb)->fin)
1712                                goto found_fin_ok;
1713                        WARN(!(flags & MSG_PEEK),
1714                             "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1715                             *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1716                }
1717
1718                /* Well, if we have backlog, try to process it now yet. */
1719
1720                if (copied >= target && !sk->sk_backlog.tail)
1721                        break;
1722
1723                if (copied) {
1724                        if (sk->sk_err ||
1725                            sk->sk_state == TCP_CLOSE ||
1726                            (sk->sk_shutdown & RCV_SHUTDOWN) ||
1727                            !timeo ||
1728                            signal_pending(current))
1729                                break;
1730                } else {
1731                        if (sock_flag(sk, SOCK_DONE))
1732                                break;
1733
1734                        if (sk->sk_err) {
1735                                copied = sock_error(sk);
1736                                break;
1737                        }
1738
1739                        if (sk->sk_shutdown & RCV_SHUTDOWN)
1740                                break;
1741
1742                        if (sk->sk_state == TCP_CLOSE) {
1743                                if (!sock_flag(sk, SOCK_DONE)) {
1744                                        /* This occurs when user tries to read
1745                                         * from never connected socket.
1746                                         */
1747                                        copied = -ENOTCONN;
1748                                        break;
1749                                }
1750                                break;
1751                        }
1752
1753                        if (!timeo) {
1754                                copied = -EAGAIN;
1755                                break;
1756                        }
1757
1758                        if (signal_pending(current)) {
1759                                copied = sock_intr_errno(timeo);
1760                                break;
1761                        }
1762                }
1763
1764                tcp_cleanup_rbuf(sk, copied);
1765
1766                if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1767                        /* Install new reader */
1768                        if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1769                                user_recv = current;
1770                                tp->ucopy.task = user_recv;
1771                                tp->ucopy.iov = msg->msg_iov;
1772                        }
1773
1774                        tp->ucopy.len = len;
1775
1776                        WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1777                                !(flags & (MSG_PEEK | MSG_TRUNC)));
1778
1779                        /* Ugly... If prequeue is not empty, we have to
1780                         * process it before releasing socket, otherwise
1781                         * order will be broken at second iteration.
1782                         * More elegant solution is required!!!
1783                         *
1784                         * Look: we have the following (pseudo)queues:
1785                         *
1786                         * 1. packets in flight
1787                         * 2. backlog
1788                         * 3. prequeue
1789                         * 4. receive_queue
1790                         *
1791                         * Each queue can be processed only if the next ones
1792                         * are empty. At this point we have empty receive_queue.
1793                         * But prequeue _can_ be not empty after 2nd iteration,
1794                         * when we jumped to start of loop because backlog
1795                         * processing added something to receive_queue.
1796                         * We cannot release_sock(), because backlog contains
1797                         * packets arrived _after_ prequeued ones.
1798                         *
1799                         * Shortly, algorithm is clear --- to process all
1800                         * the queues in order. We could make it more directly,
1801                         * requeueing packets from backlog to prequeue, if
1802                         * is not empty. It is more elegant, but eats cycles,
1803                         * unfortunately.
1804                         */
1805                        if (!skb_queue_empty(&tp->ucopy.prequeue))
1806                                goto do_prequeue;
1807
1808                        /* __ Set realtime policy in scheduler __ */
1809                }
1810
1811#ifdef CONFIG_NET_DMA
1812                if (tp->ucopy.dma_chan) {
1813                        if (tp->rcv_wnd == 0 &&
1814                            !skb_queue_empty(&sk->sk_async_wait_queue)) {
1815                                tcp_service_net_dma(sk, true);
1816                                tcp_cleanup_rbuf(sk, copied);
1817                        } else
1818                                dma_async_issue_pending(tp->ucopy.dma_chan);
1819                }
1820#endif
1821                if (copied >= target) {
1822                        /* Do not sleep, just process backlog. */
1823                        release_sock(sk);
1824                        lock_sock(sk);
1825                } else
1826                        sk_wait_data(sk, &timeo);
1827
1828#ifdef CONFIG_NET_DMA
1829                tcp_service_net_dma(sk, false);  /* Don't block */
1830                tp->ucopy.wakeup = 0;
1831#endif
1832
1833                if (user_recv) {
1834                        int chunk;
1835
1836                        /* __ Restore normal policy in scheduler __ */
1837
1838                        if ((chunk = len - tp->ucopy.len) != 0) {
1839                                NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1840                                len -= chunk;
1841                                copied += chunk;
1842                        }
1843
1844                        if (tp->rcv_nxt == tp->copied_seq &&
1845                            !skb_queue_empty(&tp->ucopy.prequeue)) {
1846do_prequeue:
1847                                tcp_prequeue_process(sk);
1848
1849                                if ((chunk = len - tp->ucopy.len) != 0) {
1850                                        NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1851                                        len -= chunk;
1852                                        copied += chunk;
1853                                }
1854                        }
1855                }
1856                if ((flags & MSG_PEEK) &&
1857                    (peek_seq - copied - urg_hole != tp->copied_seq)) {
1858                        net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1859                                            current->comm,
1860                                            task_pid_nr(current));
1861                        peek_seq = tp->copied_seq;
1862                }
1863                continue;
1864
1865        found_ok_skb:
1866                /* Ok so how much can we use? */
1867                used = skb->len - offset;
1868                if (len < used)
1869                        used = len;
1870
1871                /* Do we have urgent data here? */
1872                if (tp->urg_data) {
1873                        u32 urg_offset = tp->urg_seq - *seq;
1874                        if (urg_offset < used) {
1875                                if (!urg_offset) {
1876                                        if (!sock_flag(sk, SOCK_URGINLINE)) {
1877                                                ++*seq;
1878                                                urg_hole++;
1879                                                offset++;
1880                                                used--;
1881                                                if (!used)
1882                                                        goto skip_copy;
1883                                        }
1884                                } else
1885                                        used = urg_offset;
1886                        }
1887                }
1888
1889                if (!(flags & MSG_TRUNC)) {
1890#ifdef CONFIG_NET_DMA
1891                        if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1892                                tp->ucopy.dma_chan = net_dma_find_channel();
1893
1894                        if (tp->ucopy.dma_chan) {
1895                                tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1896                                        tp->ucopy.dma_chan, skb, offset,
1897                                        msg->msg_iov, used,
1898                                        tp->ucopy.pinned_list);
1899
1900                                if (tp->ucopy.dma_cookie < 0) {
1901
1902                                        pr_alert("%s: dma_cookie < 0\n",
1903                                                 __func__);
1904
1905                                        /* Exception. Bailout! */
1906                                        if (!copied)
1907                                                copied = -EFAULT;
1908                                        break;
1909                                }
1910
1911                                dma_async_issue_pending(tp->ucopy.dma_chan);
1912
1913                                if ((offset + used) == skb->len)
1914                                        copied_early = true;
1915
1916                        } else
1917#endif
1918                        {
1919                                err = skb_copy_datagram_iovec(skb, offset,
1920                                                msg->msg_iov, used);
1921                                if (err) {
1922                                        /* Exception. Bailout! */
1923                                        if (!copied)
1924                                                copied = -EFAULT;
1925                                        break;
1926                                }
1927                        }
1928                }
1929
1930                *seq += used;
1931                copied += used;
1932                len -= used;
1933
1934                tcp_rcv_space_adjust(sk);
1935
1936skip_copy:
1937                if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1938                        tp->urg_data = 0;
1939                        tcp_fast_path_check(sk);
1940                }
1941                if (used + offset < skb->len)
1942                        continue;
1943
1944                if (tcp_hdr(skb)->fin)
1945                        goto found_fin_ok;
1946                if (!(flags & MSG_PEEK)) {
1947                        sk_eat_skb(sk, skb, copied_early);
1948                        copied_early = false;
1949                }
1950                continue;
1951
1952        found_fin_ok:
1953                /* Process the FIN. */
1954                ++*seq;
1955                if (!(flags & MSG_PEEK)) {
1956                        sk_eat_skb(sk, skb, copied_early);
1957                        copied_early = false;
1958                }
1959                break;
1960        } while (len > 0);
1961
1962        if (user_recv) {
1963                if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1964                        int chunk;
1965
1966                        tp->ucopy.len = copied > 0 ? len : 0;
1967
1968                        tcp_prequeue_process(sk);
1969
1970                        if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1971                                NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1972                                len -= chunk;
1973                                copied += chunk;
1974                        }
1975                }
1976
1977                tp->ucopy.task = NULL;
1978                tp->ucopy.len = 0;
1979        }
1980
1981#ifdef CONFIG_NET_DMA
1982        tcp_service_net_dma(sk, true);  /* Wait for queue to drain */
1983        tp->ucopy.dma_chan = NULL;
1984
1985        if (tp->ucopy.pinned_list) {
1986                dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1987                tp->ucopy.pinned_list = NULL;
1988        }
1989#endif
1990
1991        /* According to UNIX98, msg_name/msg_namelen are ignored
1992         * on connected socket. I was just happy when found this 8) --ANK
1993         */
1994
1995        /* Clean up data we have read: This will do ACK frames. */
1996        tcp_cleanup_rbuf(sk, copied);
1997
1998        release_sock(sk);
1999        return copied;
2000
2001out:
2002        release_sock(sk);
2003        return err;
2004
2005recv_urg:
2006        err = tcp_recv_urg(sk, msg, len, flags);
2007        goto out;
2008
2009recv_sndq:
2010        err = tcp_peek_sndq(sk, msg, len);
2011        goto out;
2012}
2013EXPORT_SYMBOL(tcp_recvmsg);
2014
2015void tcp_set_state(struct sock *sk, int state)
2016{
2017        int oldstate = sk->sk_state;
2018
2019        switch (state) {
2020        case TCP_ESTABLISHED:
2021                if (oldstate != TCP_ESTABLISHED)
2022                        TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2023                break;
2024
2025        case TCP_CLOSE:
2026                if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2027                        TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2028
2029                sk->sk_prot->unhash(sk);
2030                if (inet_csk(sk)->icsk_bind_hash &&
2031                    !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2032                        inet_put_port(sk);
2033                /* fall through */
2034        default:
2035                if (oldstate == TCP_ESTABLISHED)
2036                        TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2037        }
2038
2039        /* Change state AFTER socket is unhashed to avoid closed
2040         * socket sitting in hash tables.
2041         */
2042        sk->sk_state = state;
2043
2044#ifdef STATE_TRACE
2045        SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
2046#endif
2047}
2048EXPORT_SYMBOL_GPL(tcp_set_state);
2049
2050/*
2051 *      State processing on a close. This implements the state shift for
2052 *      sending our FIN frame. Note that we only send a FIN for some
2053 *      states. A shutdown() may have already sent the FIN, or we may be
2054 *      closed.
2055 */
2056
2057static const unsigned char new_state[16] = {
2058  /* current state:        new state:      action:      */
2059  /* (Invalid)          */ TCP_CLOSE,
2060  /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2061  /* TCP_SYN_SENT       */ TCP_CLOSE,
2062  /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2063  /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
2064  /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
2065  /* TCP_TIME_WAIT      */ TCP_CLOSE,
2066  /* TCP_CLOSE          */ TCP_CLOSE,
2067  /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
2068  /* TCP_LAST_ACK       */ TCP_LAST_ACK,
2069  /* TCP_LISTEN         */ TCP_CLOSE,
2070  /* TCP_CLOSING        */ TCP_CLOSING,
2071};
2072
2073static int tcp_close_state(struct sock *sk)
2074{
2075        int next = (int)new_state[sk->sk_state];
2076        int ns = next & TCP_STATE_MASK;
2077
2078        tcp_set_state(sk, ns);
2079
2080        return next & TCP_ACTION_FIN;
2081}
2082
2083/*
2084 *      Shutdown the sending side of a connection. Much like close except
2085 *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2086 */
2087
2088void tcp_shutdown(struct sock *sk, int how)
2089{
2090        /*      We need to grab some memory, and put together a FIN,
2091         *      and then put it into the queue to be sent.
2092         *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2093         */
2094        if (!(how & SEND_SHUTDOWN))
2095                return;
2096
2097        /* If we've already sent a FIN, or it's a closed state, skip this. */
2098        if ((1 << sk->sk_state) &
2099            (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2100             TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2101                /* Clear out any half completed packets.  FIN if needed. */
2102                if (tcp_close_state(sk))
2103                        tcp_send_fin(sk);
2104        }
2105}
2106EXPORT_SYMBOL(tcp_shutdown);
2107
2108bool tcp_check_oom(struct sock *sk, int shift)
2109{
2110        bool too_many_orphans, out_of_socket_memory;
2111
2112        too_many_orphans = tcp_too_many_orphans(sk, shift);
2113        out_of_socket_memory = tcp_out_of_memory(sk);
2114
2115        if (too_many_orphans)
2116                net_info_ratelimited("too many orphaned sockets\n");
2117        if (out_of_socket_memory)
2118                net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2119        return too_many_orphans || out_of_socket_memory;
2120}
2121
2122void tcp_close(struct sock *sk, long timeout)
2123{
2124        struct sk_buff *skb;
2125        int data_was_unread = 0;
2126        int state;
2127
2128        lock_sock(sk);
2129        sk->sk_shutdown = SHUTDOWN_MASK;
2130
2131        if (sk->sk_state == TCP_LISTEN) {
2132                tcp_set_state(sk, TCP_CLOSE);
2133
2134                /* Special case. */
2135                inet_csk_listen_stop(sk);
2136
2137                goto adjudge_to_death;
2138        }
2139
2140        /*  We need to flush the recv. buffs.  We do this only on the
2141         *  descriptor close, not protocol-sourced closes, because the
2142         *  reader process may not have drained the data yet!
2143         */
2144        while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2145                u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2146                          tcp_hdr(skb)->fin;
2147                data_was_unread += len;
2148                __kfree_skb(skb);
2149        }
2150
2151        sk_mem_reclaim(sk);
2152
2153        /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2154        if (sk->sk_state == TCP_CLOSE)
2155                goto adjudge_to_death;
2156
2157        /* As outlined in RFC 2525, section 2.17, we send a RST here because
2158         * data was lost. To witness the awful effects of the old behavior of
2159         * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2160         * GET in an FTP client, suspend the process, wait for the client to
2161         * advertise a zero window, then kill -9 the FTP client, wheee...
2162         * Note: timeout is always zero in such a case.
2163         */
2164        if (unlikely(tcp_sk(sk)->repair)) {
2165                sk->sk_prot->disconnect(sk, 0);
2166        } else if (data_was_unread) {
2167                /* Unread data was tossed, zap the connection. */
2168                NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2169                tcp_set_state(sk, TCP_CLOSE);
2170                tcp_send_active_reset(sk, sk->sk_allocation);
2171        } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2172                /* Check zero linger _after_ checking for unread data. */
2173                sk->sk_prot->disconnect(sk, 0);
2174                NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2175        } else if (tcp_close_state(sk)) {
2176                /* We FIN if the application ate all the data before
2177                 * zapping the connection.
2178                 */
2179
2180                /* RED-PEN. Formally speaking, we have broken TCP state
2181                 * machine. State transitions:
2182                 *
2183                 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2184                 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2185                 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2186                 *
2187                 * are legal only when FIN has been sent (i.e. in window),
2188                 * rather than queued out of window. Purists blame.
2189                 *
2190                 * F.e. "RFC state" is ESTABLISHED,
2191                 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2192                 *
2193                 * The visible declinations are that sometimes
2194                 * we enter time-wait state, when it is not required really
2195                 * (harmless), do not send active resets, when they are
2196                 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2197                 * they look as CLOSING or LAST_ACK for Linux)
2198                 * Probably, I missed some more holelets.
2199                 *                                              --ANK
2200                 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2201                 * in a single packet! (May consider it later but will
2202                 * probably need API support or TCP_CORK SYN-ACK until
2203                 * data is written and socket is closed.)
2204                 */
2205                tcp_send_fin(sk);
2206        }
2207
2208        sk_stream_wait_close(sk, timeout);
2209
2210adjudge_to_death:
2211        state = sk->sk_state;
2212        sock_hold(sk);
2213        sock_orphan(sk);
2214
2215        /* It is the last release_sock in its life. It will remove backlog. */
2216        release_sock(sk);
2217
2218
2219        /* Now socket is owned by kernel and we acquire BH lock
2220           to finish close. No need to check for user refs.
2221         */
2222        local_bh_disable();
2223        bh_lock_sock(sk);
2224        WARN_ON(sock_owned_by_user(sk));
2225
2226        percpu_counter_inc(sk->sk_prot->orphan_count);
2227
2228        /* Have we already been destroyed by a softirq or backlog? */
2229        if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2230                goto out;
2231
2232        /*      This is a (useful) BSD violating of the RFC. There is a
2233         *      problem with TCP as specified in that the other end could
2234         *      keep a socket open forever with no application left this end.
2235         *      We use a 1 minute timeout (about the same as BSD) then kill
2236         *      our end. If they send after that then tough - BUT: long enough
2237         *      that we won't make the old 4*rto = almost no time - whoops
2238         *      reset mistake.
2239         *
2240         *      Nope, it was not mistake. It is really desired behaviour
2241         *      f.e. on http servers, when such sockets are useless, but
2242         *      consume significant resources. Let's do it with special
2243         *      linger2 option.                                 --ANK
2244         */
2245
2246        if (sk->sk_state == TCP_FIN_WAIT2) {
2247                struct tcp_sock *tp = tcp_sk(sk);
2248                if (tp->linger2 < 0) {
2249                        tcp_set_state(sk, TCP_CLOSE);
2250                        tcp_send_active_reset(sk, GFP_ATOMIC);
2251                        NET_INC_STATS_BH(sock_net(sk),
2252                                        LINUX_MIB_TCPABORTONLINGER);
2253                } else {
2254                        const int tmo = tcp_fin_time(sk);
2255
2256                        if (tmo > TCP_TIMEWAIT_LEN) {
2257                                inet_csk_reset_keepalive_timer(sk,
2258                                                tmo - TCP_TIMEWAIT_LEN);
2259                        } else {
2260                                tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2261                                goto out;
2262                        }
2263                }
2264        }
2265        if (sk->sk_state != TCP_CLOSE) {
2266                sk_mem_reclaim(sk);
2267                if (tcp_check_oom(sk, 0)) {
2268                        tcp_set_state(sk, TCP_CLOSE);
2269                        tcp_send_active_reset(sk, GFP_ATOMIC);
2270                        NET_INC_STATS_BH(sock_net(sk),
2271                                        LINUX_MIB_TCPABORTONMEMORY);
2272                }
2273        }
2274
2275        if (sk->sk_state == TCP_CLOSE) {
2276                struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2277                /* We could get here with a non-NULL req if the socket is
2278                 * aborted (e.g., closed with unread data) before 3WHS
2279                 * finishes.
2280                 */
2281                if (req != NULL)
2282                        reqsk_fastopen_remove(sk, req, false);
2283                inet_csk_destroy_sock(sk);
2284        }
2285        /* Otherwise, socket is reprieved until protocol close. */
2286
2287out:
2288        bh_unlock_sock(sk);
2289        local_bh_enable();
2290        sock_put(sk);
2291}
2292EXPORT_SYMBOL(tcp_close);
2293
2294/* These states need RST on ABORT according to RFC793 */
2295
2296static inline bool tcp_need_reset(int state)
2297{
2298        return (1 << state) &
2299               (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2300                TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2301}
2302
2303int tcp_disconnect(struct sock *sk, int flags)
2304{
2305        struct inet_sock *inet = inet_sk(sk);
2306        struct inet_connection_sock *icsk = inet_csk(sk);
2307        struct tcp_sock *tp = tcp_sk(sk);
2308        int err = 0;
2309        int old_state = sk->sk_state;
2310
2311        if (old_state != TCP_CLOSE)
2312                tcp_set_state(sk, TCP_CLOSE);
2313
2314        /* ABORT function of RFC793 */
2315        if (old_state == TCP_LISTEN) {
2316                inet_csk_listen_stop(sk);
2317        } else if (unlikely(tp->repair)) {
2318                sk->sk_err = ECONNABORTED;
2319        } else if (tcp_need_reset(old_state) ||
2320                   (tp->snd_nxt != tp->write_seq &&
2321                    (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2322                /* The last check adjusts for discrepancy of Linux wrt. RFC
2323                 * states
2324                 */
2325                tcp_send_active_reset(sk, gfp_any());
2326                sk->sk_err = ECONNRESET;
2327        } else if (old_state == TCP_SYN_SENT)
2328                sk->sk_err = ECONNRESET;
2329
2330        tcp_clear_xmit_timers(sk);
2331        __skb_queue_purge(&sk->sk_receive_queue);
2332        tcp_write_queue_purge(sk);
2333        __skb_queue_purge(&tp->out_of_order_queue);
2334#ifdef CONFIG_NET_DMA
2335        __skb_queue_purge(&sk->sk_async_wait_queue);
2336#endif
2337
2338        inet->inet_dport = 0;
2339
2340        if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2341                inet_reset_saddr(sk);
2342
2343        sk->sk_shutdown = 0;
2344        sock_reset_flag(sk, SOCK_DONE);
2345        tp->srtt_us = 0;
2346        if ((tp->write_seq += tp->max_window + 2) == 0)
2347                tp->write_seq = 1;
2348        icsk->icsk_backoff = 0;
2349        tp->snd_cwnd = 2;
2350        icsk->icsk_probes_out = 0;
2351        tp->packets_out = 0;
2352        tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2353        tp->snd_cwnd_cnt = 0;
2354        tp->window_clamp = 0;
2355        tcp_set_ca_state(sk, TCP_CA_Open);
2356        tcp_clear_retrans(tp);
2357        inet_csk_delack_init(sk);
2358        tcp_init_send_head(sk);
2359        memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2360        __sk_dst_reset(sk);
2361
2362        WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2363
2364        sk->sk_error_report(sk);
2365        return err;
2366}
2367EXPORT_SYMBOL(tcp_disconnect);
2368
2369void tcp_sock_destruct(struct sock *sk)
2370{
2371        inet_sock_destruct(sk);
2372
2373        kfree(inet_csk(sk)->icsk_accept_queue.fastopenq);
2374}
2375
2376static inline bool tcp_can_repair_sock(const struct sock *sk)
2377{
2378        return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2379                ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2380}
2381
2382static int tcp_repair_options_est(struct tcp_sock *tp,
2383                struct tcp_repair_opt __user *optbuf, unsigned int len)
2384{
2385        struct tcp_repair_opt opt;
2386
2387        while (len >= sizeof(opt)) {
2388                if (copy_from_user(&opt, optbuf, sizeof(opt)))
2389                        return -EFAULT;
2390
2391                optbuf++;
2392                len -= sizeof(opt);
2393
2394                switch (opt.opt_code) {
2395                case TCPOPT_MSS:
2396                        tp->rx_opt.mss_clamp = opt.opt_val;
2397                        break;
2398                case TCPOPT_WINDOW:
2399                        {
2400                                u16 snd_wscale = opt.opt_val & 0xFFFF;
2401                                u16 rcv_wscale = opt.opt_val >> 16;
2402
2403                                if (snd_wscale > 14 || rcv_wscale > 14)
2404                                        return -EFBIG;
2405
2406                                tp->rx_opt.snd_wscale = snd_wscale;
2407                                tp->rx_opt.rcv_wscale = rcv_wscale;
2408                                tp->rx_opt.wscale_ok = 1;
2409                        }
2410                        break;
2411                case TCPOPT_SACK_PERM:
2412                        if (opt.opt_val != 0)
2413                                return -EINVAL;
2414
2415                        tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2416                        if (sysctl_tcp_fack)
2417                                tcp_enable_fack(tp);
2418                        break;
2419                case TCPOPT_TIMESTAMP:
2420                        if (opt.opt_val != 0)
2421                                return -EINVAL;
2422
2423                        tp->rx_opt.tstamp_ok = 1;
2424                        break;
2425                }
2426        }
2427
2428        return 0;
2429}
2430
2431/*
2432 *      Socket option code for TCP.
2433 */
2434static int do_tcp_setsockopt(struct sock *sk, int level,
2435                int optname, char __user *optval, unsigned int optlen)
2436{
2437        struct tcp_sock *tp = tcp_sk(sk);
2438        struct inet_connection_sock *icsk = inet_csk(sk);
2439        int val;
2440        int err = 0;
2441
2442        /* These are data/string values, all the others are ints */
2443        switch (optname) {
2444        case TCP_CONGESTION: {
2445                char name[TCP_CA_NAME_MAX];
2446
2447                if (optlen < 1)
2448                        return -EINVAL;
2449
2450                val = strncpy_from_user(name, optval,
2451                                        min_t(long, TCP_CA_NAME_MAX-1, optlen));
2452                if (val < 0)
2453                        return -EFAULT;
2454                name[val] = 0;
2455
2456                lock_sock(sk);
2457                err = tcp_set_congestion_control(sk, name);
2458                release_sock(sk);
2459                return err;
2460        }
2461        default:
2462                /* fallthru */
2463                break;
2464        }
2465
2466        if (optlen < sizeof(int))
2467                return -EINVAL;
2468
2469        if (get_user(val, (int __user *)optval))
2470                return -EFAULT;
2471
2472        lock_sock(sk);
2473
2474        switch (optname) {
2475        case TCP_MAXSEG:
2476                /* Values greater than interface MTU won't take effect. However
2477                 * at the point when this call is done we typically don't yet
2478                 * know which interface is going to be used */
2479                if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2480                        err = -EINVAL;
2481                        break;
2482                }
2483                tp->rx_opt.user_mss = val;
2484                break;
2485
2486        case TCP_NODELAY:
2487                if (val) {
2488                        /* TCP_NODELAY is weaker than TCP_CORK, so that
2489                         * this option on corked socket is remembered, but
2490                         * it is not activated until cork is cleared.
2491                         *
2492                         * However, when TCP_NODELAY is set we make
2493                         * an explicit push, which overrides even TCP_CORK
2494                         * for currently queued segments.
2495                         */
2496                        tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2497                        tcp_push_pending_frames(sk);
2498                } else {
2499                        tp->nonagle &= ~TCP_NAGLE_OFF;
2500                }
2501                break;
2502
2503        case TCP_THIN_LINEAR_TIMEOUTS:
2504                if (val < 0 || val > 1)
2505                        err = -EINVAL;
2506                else
2507                        tp->thin_lto = val;
2508                break;
2509
2510        case TCP_THIN_DUPACK:
2511                if (val < 0 || val > 1)
2512                        err = -EINVAL;
2513                else {
2514                        tp->thin_dupack = val;
2515                        if (tp->thin_dupack)
2516                                tcp_disable_early_retrans(tp);
2517                }
2518                break;
2519
2520        case TCP_REPAIR:
2521                if (!tcp_can_repair_sock(sk))
2522                        err = -EPERM;
2523                else if (val == 1) {
2524                        tp->repair = 1;
2525                        sk->sk_reuse = SK_FORCE_REUSE;
2526                        tp->repair_queue = TCP_NO_QUEUE;
2527                } else if (val == 0) {
2528                        tp->repair = 0;
2529                        sk->sk_reuse = SK_NO_REUSE;
2530                        tcp_send_window_probe(sk);
2531                } else
2532                        err = -EINVAL;
2533
2534                break;
2535
2536        case TCP_REPAIR_QUEUE:
2537                if (!tp->repair)
2538                        err = -EPERM;
2539                else if (val < TCP_QUEUES_NR)
2540                        tp->repair_queue = val;
2541                else
2542                        err = -EINVAL;
2543                break;
2544
2545        case TCP_QUEUE_SEQ:
2546                if (sk->sk_state != TCP_CLOSE)
2547                        err = -EPERM;
2548                else if (tp->repair_queue == TCP_SEND_QUEUE)
2549                        tp->write_seq = val;
2550                else if (tp->repair_queue == TCP_RECV_QUEUE)
2551                        tp->rcv_nxt = val;
2552                else
2553                        err = -EINVAL;
2554                break;
2555
2556        case TCP_REPAIR_OPTIONS:
2557                if (!tp->repair)
2558                        err = -EINVAL;
2559                else if (sk->sk_state == TCP_ESTABLISHED)
2560                        err = tcp_repair_options_est(tp,
2561                                        (struct tcp_repair_opt __user *)optval,
2562                                        optlen);
2563                else
2564                        err = -EPERM;
2565                break;
2566
2567        case TCP_CORK:
2568                /* When set indicates to always queue non-full frames.
2569                 * Later the user clears this option and we transmit
2570                 * any pending partial frames in the queue.  This is
2571                 * meant to be used alongside sendfile() to get properly
2572                 * filled frames when the user (for example) must write
2573                 * out headers with a write() call first and then use
2574                 * sendfile to send out the data parts.
2575                 *
2576                 * TCP_CORK can be set together with TCP_NODELAY and it is
2577                 * stronger than TCP_NODELAY.
2578                 */
2579                if (val) {
2580                        tp->nonagle |= TCP_NAGLE_CORK;
2581                } else {
2582                        tp->nonagle &= ~TCP_NAGLE_CORK;
2583                        if (tp->nonagle&TCP_NAGLE_OFF)
2584                                tp->nonagle |= TCP_NAGLE_PUSH;
2585                        tcp_push_pending_frames(sk);
2586                }
2587                break;
2588
2589        case TCP_KEEPIDLE:
2590                if (val < 1 || val > MAX_TCP_KEEPIDLE)
2591                        err = -EINVAL;
2592                else {
2593                        tp->keepalive_time = val * HZ;
2594                        if (sock_flag(sk, SOCK_KEEPOPEN) &&
2595                            !((1 << sk->sk_state) &
2596                              (TCPF_CLOSE | TCPF_LISTEN))) {
2597                                u32 elapsed = keepalive_time_elapsed(tp);
2598                                if (tp->keepalive_time > elapsed)
2599                                        elapsed = tp->keepalive_time - elapsed;
2600                                else
2601                                        elapsed = 0;
2602                                inet_csk_reset_keepalive_timer(sk, elapsed);
2603                        }
2604                }
2605                break;
2606        case TCP_KEEPINTVL:
2607                if (val < 1 || val > MAX_TCP_KEEPINTVL)
2608                        err = -EINVAL;
2609                else
2610                        tp->keepalive_intvl = val * HZ;
2611                break;
2612        case TCP_KEEPCNT:
2613                if (val < 1 || val > MAX_TCP_KEEPCNT)
2614                        err = -EINVAL;
2615                else
2616                        tp->keepalive_probes = val;
2617                break;
2618        case TCP_SYNCNT:
2619                if (val < 1 || val > MAX_TCP_SYNCNT)
2620                        err = -EINVAL;
2621                else
2622                        icsk->icsk_syn_retries = val;
2623                break;
2624
2625        case TCP_LINGER2:
2626                if (val < 0)
2627                        tp->linger2 = -1;
2628                else if (val > sysctl_tcp_fin_timeout / HZ)
2629                        tp->linger2 = 0;
2630                else
2631                        tp->linger2 = val * HZ;
2632                break;
2633
2634        case TCP_DEFER_ACCEPT:
2635                /* Translate value in seconds to number of retransmits */
2636                icsk->icsk_accept_queue.rskq_defer_accept =
2637                        secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2638                                        TCP_RTO_MAX / HZ);
2639                break;
2640
2641        case TCP_WINDOW_CLAMP:
2642                if (!val) {
2643                        if (sk->sk_state != TCP_CLOSE) {
2644                                err = -EINVAL;
2645                                break;
2646                        }
2647                        tp->window_clamp = 0;
2648                } else
2649                        tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2650                                                SOCK_MIN_RCVBUF / 2 : val;
2651                break;
2652
2653        case TCP_QUICKACK:
2654                if (!val) {
2655                        icsk->icsk_ack.pingpong = 1;
2656                } else {
2657                        icsk->icsk_ack.pingpong = 0;
2658                        if ((1 << sk->sk_state) &
2659                            (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2660                            inet_csk_ack_scheduled(sk)) {
2661                                icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2662                                tcp_cleanup_rbuf(sk, 1);
2663                                if (!(val & 1))
2664                                        icsk->icsk_ack.pingpong = 1;
2665                        }
2666                }
2667                break;
2668
2669#ifdef CONFIG_TCP_MD5SIG
2670        case TCP_MD5SIG:
2671                /* Read the IP->Key mappings from userspace */
2672                err = tp->af_specific->md5_parse(sk, optval, optlen);
2673                break;
2674#endif
2675        case TCP_USER_TIMEOUT:
2676                /* Cap the max timeout in ms TCP will retry/retrans
2677                 * before giving up and aborting (ETIMEDOUT) a connection.
2678                 */
2679                if (val < 0)
2680                        err = -EINVAL;
2681                else
2682                        icsk->icsk_user_timeout = msecs_to_jiffies(val);
2683                break;
2684
2685        case TCP_FASTOPEN:
2686                if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2687                    TCPF_LISTEN)))
2688                        err = fastopen_init_queue(sk, val);
2689                else
2690                        err = -EINVAL;
2691                break;
2692        case TCP_TIMESTAMP:
2693                if (!tp->repair)
2694                        err = -EPERM;
2695                else
2696                        tp->tsoffset = val - tcp_time_stamp;
2697                break;
2698        case TCP_NOTSENT_LOWAT:
2699                tp->notsent_lowat = val;
2700                sk->sk_write_space(sk);
2701                break;
2702        default:
2703                err = -ENOPROTOOPT;
2704                break;
2705        }
2706
2707        release_sock(sk);
2708        return err;
2709}
2710
2711int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2712                   unsigned int optlen)
2713{
2714        const struct inet_connection_sock *icsk = inet_csk(sk);
2715
2716        if (level != SOL_TCP)
2717                return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2718                                                     optval, optlen);
2719        return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2720}
2721EXPORT_SYMBOL(tcp_setsockopt);
2722
2723#ifdef CONFIG_COMPAT
2724int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2725                          char __user *optval, unsigned int optlen)
2726{
2727        if (level != SOL_TCP)
2728                return inet_csk_compat_setsockopt(sk, level, optname,
2729                                                  optval, optlen);
2730        return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2731}
2732EXPORT_SYMBOL(compat_tcp_setsockopt);
2733#endif
2734
2735/* Return information about state of tcp endpoint in API format. */
2736void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2737{
2738        const struct tcp_sock *tp = tcp_sk(sk);
2739        const struct inet_connection_sock *icsk = inet_csk(sk);
2740        u32 now = tcp_time_stamp;
2741
2742        memset(info, 0, sizeof(*info));
2743
2744        info->tcpi_state = sk->sk_state;
2745        info->tcpi_ca_state = icsk->icsk_ca_state;
2746        info->tcpi_retransmits = icsk->icsk_retransmits;
2747        info->tcpi_probes = icsk->icsk_probes_out;
2748        info->tcpi_backoff = icsk->icsk_backoff;
2749
2750        if (tp->rx_opt.tstamp_ok)
2751                info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2752        if (tcp_is_sack(tp))
2753                info->tcpi_options |= TCPI_OPT_SACK;
2754        if (tp->rx_opt.wscale_ok) {
2755                info->tcpi_options |= TCPI_OPT_WSCALE;
2756                info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2757                info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2758        }
2759
2760        if (tp->ecn_flags & TCP_ECN_OK)
2761                info->tcpi_options |= TCPI_OPT_ECN;
2762        if (tp->ecn_flags & TCP_ECN_SEEN)
2763                info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2764        if (tp->syn_data_acked)
2765                info->tcpi_options |= TCPI_OPT_SYN_DATA;
2766
2767        info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2768        info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2769        info->tcpi_snd_mss = tp->mss_cache;
2770        info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2771
2772        if (sk->sk_state == TCP_LISTEN) {
2773                info->tcpi_unacked = sk->sk_ack_backlog;
2774                info->tcpi_sacked = sk->sk_max_ack_backlog;
2775        } else {
2776                info->tcpi_unacked = tp->packets_out;
2777                info->tcpi_sacked = tp->sacked_out;
2778        }
2779        info->tcpi_lost = tp->lost_out;
2780        info->tcpi_retrans = tp->retrans_out;
2781        info->tcpi_fackets = tp->fackets_out;
2782
2783        info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2784        info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2785        info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2786
2787        info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2788        info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2789        info->tcpi_rtt = tp->srtt_us >> 3;
2790        info->tcpi_rttvar = tp->mdev_us >> 2;
2791        info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2792        info->tcpi_snd_cwnd = tp->snd_cwnd;
2793        info->tcpi_advmss = tp->advmss;
2794        info->tcpi_reordering = tp->reordering;
2795
2796        info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2797        info->tcpi_rcv_space = tp->rcvq_space.space;
2798
2799        info->tcpi_total_retrans = tp->total_retrans;
2800
2801        info->tcpi_pacing_rate = sk->sk_pacing_rate != ~0U ?
2802                                        sk->sk_pacing_rate : ~0ULL;
2803        info->tcpi_max_pacing_rate = sk->sk_max_pacing_rate != ~0U ?
2804                                        sk->sk_max_pacing_rate : ~0ULL;
2805}
2806EXPORT_SYMBOL_GPL(tcp_get_info);
2807
2808static int do_tcp_getsockopt(struct sock *sk, int level,
2809                int optname, char __user *optval, int __user *optlen)
2810{
2811        struct inet_connection_sock *icsk = inet_csk(sk);
2812        struct tcp_sock *tp = tcp_sk(sk);
2813        int val, len;
2814
2815        if (get_user(len, optlen))
2816                return -EFAULT;
2817
2818        len = min_t(unsigned int, len, sizeof(int));
2819
2820        if (len < 0)
2821                return -EINVAL;
2822
2823        switch (optname) {
2824        case TCP_MAXSEG:
2825                val = tp->mss_cache;
2826                if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2827                        val = tp->rx_opt.user_mss;
2828                if (tp->repair)
2829                        val = tp->rx_opt.mss_clamp;
2830                break;
2831        case TCP_NODELAY:
2832                val = !!(tp->nonagle&TCP_NAGLE_OFF);
2833                break;
2834        case TCP_CORK:
2835                val = !!(tp->nonagle&TCP_NAGLE_CORK);
2836                break;
2837        case TCP_KEEPIDLE:
2838                val = keepalive_time_when(tp) / HZ;
2839                break;
2840        case TCP_KEEPINTVL:
2841                val = keepalive_intvl_when(tp) / HZ;
2842                break;
2843        case TCP_KEEPCNT:
2844                val = keepalive_probes(tp);
2845                break;
2846        case TCP_SYNCNT:
2847                val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2848                break;
2849        case TCP_LINGER2:
2850                val = tp->linger2;
2851                if (val >= 0)
2852                        val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2853                break;
2854        case TCP_DEFER_ACCEPT:
2855                val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2856                                      TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2857                break;
2858        case TCP_WINDOW_CLAMP:
2859                val = tp->window_clamp;
2860                break;
2861        case TCP_INFO: {
2862                struct tcp_info info;
2863
2864                if (get_user(len, optlen))
2865                        return -EFAULT;
2866
2867                tcp_get_info(sk, &info);
2868
2869                len = min_t(unsigned int, len, sizeof(info));
2870                if (put_user(len, optlen))
2871                        return -EFAULT;
2872                if (copy_to_user(optval, &info, len))
2873                        return -EFAULT;
2874                return 0;
2875        }
2876        case TCP_QUICKACK:
2877                val = !icsk->icsk_ack.pingpong;
2878                break;
2879
2880        case TCP_CONGESTION:
2881                if (get_user(len, optlen))
2882                        return -EFAULT;
2883                len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2884                if (put_user(len, optlen))
2885                        return -EFAULT;
2886                if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2887                        return -EFAULT;
2888                return 0;
2889
2890        case TCP_THIN_LINEAR_TIMEOUTS:
2891                val = tp->thin_lto;
2892                break;
2893        case TCP_THIN_DUPACK:
2894                val = tp->thin_dupack;
2895                break;
2896
2897        case TCP_REPAIR:
2898                val = tp->repair;
2899                break;
2900
2901        case TCP_REPAIR_QUEUE:
2902                if (tp->repair)
2903                        val = tp->repair_queue;
2904                else
2905                        return -EINVAL;
2906                break;
2907
2908        case TCP_QUEUE_SEQ:
2909                if (tp->repair_queue == TCP_SEND_QUEUE)
2910                        val = tp->write_seq;
2911                else if (tp->repair_queue == TCP_RECV_QUEUE)
2912                        val = tp->rcv_nxt;
2913                else
2914                        return -EINVAL;
2915                break;
2916
2917        case TCP_USER_TIMEOUT:
2918                val = jiffies_to_msecs(icsk->icsk_user_timeout);
2919                break;
2920
2921        case TCP_FASTOPEN:
2922                if (icsk->icsk_accept_queue.fastopenq != NULL)
2923                        val = icsk->icsk_accept_queue.fastopenq->max_qlen;
2924                else
2925                        val = 0;
2926                break;
2927
2928        case TCP_TIMESTAMP:
2929                val = tcp_time_stamp + tp->tsoffset;
2930                break;
2931        case TCP_NOTSENT_LOWAT:
2932                val = tp->notsent_lowat;
2933                break;
2934        default:
2935                return -ENOPROTOOPT;
2936        }
2937
2938        if (put_user(len, optlen))
2939                return -EFAULT;
2940        if (copy_to_user(optval, &val, len))
2941                return -EFAULT;
2942        return 0;
2943}
2944
2945int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2946                   int __user *optlen)
2947{
2948        struct inet_connection_sock *icsk = inet_csk(sk);
2949
2950        if (level != SOL_TCP)
2951                return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2952                                                     optval, optlen);
2953        return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2954}
2955EXPORT_SYMBOL(tcp_getsockopt);
2956
2957#ifdef CONFIG_COMPAT
2958int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2959                          char __user *optval, int __user *optlen)
2960{
2961        if (level != SOL_TCP)
2962                return inet_csk_compat_getsockopt(sk, level, optname,
2963                                                  optval, optlen);
2964        return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2965}
2966EXPORT_SYMBOL(compat_tcp_getsockopt);
2967#endif
2968
2969#ifdef CONFIG_TCP_MD5SIG
2970static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool __read_mostly;
2971static DEFINE_MUTEX(tcp_md5sig_mutex);
2972
2973static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
2974{
2975        int cpu;
2976
2977        for_each_possible_cpu(cpu) {
2978                struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
2979
2980                if (p->md5_desc.tfm)
2981                        crypto_free_hash(p->md5_desc.tfm);
2982        }
2983        free_percpu(pool);
2984}
2985
2986static void __tcp_alloc_md5sig_pool(void)
2987{
2988        int cpu;
2989        struct tcp_md5sig_pool __percpu *pool;
2990
2991        pool = alloc_percpu(struct tcp_md5sig_pool);
2992        if (!pool)
2993                return;
2994
2995        for_each_possible_cpu(cpu) {
2996                struct crypto_hash *hash;
2997
2998                hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2999                if (IS_ERR_OR_NULL(hash))
3000                        goto out_free;
3001
3002                per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
3003        }
3004        /* before setting tcp_md5sig_pool, we must commit all writes
3005         * to memory. See ACCESS_ONCE() in tcp_get_md5sig_pool()
3006         */
3007        smp_wmb();
3008        tcp_md5sig_pool = pool;
3009        return;
3010out_free:
3011        __tcp_free_md5sig_pool(pool);
3012}
3013
3014bool tcp_alloc_md5sig_pool(void)
3015{
3016        if (unlikely(!tcp_md5sig_pool)) {
3017                mutex_lock(&tcp_md5sig_mutex);
3018
3019                if (!tcp_md5sig_pool)
3020                        __tcp_alloc_md5sig_pool();
3021
3022                mutex_unlock(&tcp_md5sig_mutex);
3023        }
3024        return tcp_md5sig_pool != NULL;
3025}
3026EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3027
3028
3029/**
3030 *      tcp_get_md5sig_pool - get md5sig_pool for this user
3031 *
3032 *      We use percpu structure, so if we succeed, we exit with preemption
3033 *      and BH disabled, to make sure another thread or softirq handling
3034 *      wont try to get same context.
3035 */
3036struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3037{
3038        struct tcp_md5sig_pool __percpu *p;
3039
3040        local_bh_disable();
3041        p = ACCESS_ONCE(tcp_md5sig_pool);
3042        if (p)
3043                return __this_cpu_ptr(p);
3044
3045        local_bh_enable();
3046        return NULL;
3047}
3048EXPORT_SYMBOL(tcp_get_md5sig_pool);
3049
3050int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3051                        const struct tcphdr *th)
3052{
3053        struct scatterlist sg;
3054        struct tcphdr hdr;
3055        int err;
3056
3057        /* We are not allowed to change tcphdr, make a local copy */
3058        memcpy(&hdr, th, sizeof(hdr));
3059        hdr.check = 0;
3060
3061        /* options aren't included in the hash */
3062        sg_init_one(&sg, &hdr, sizeof(hdr));
3063        err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3064        return err;
3065}
3066EXPORT_SYMBOL(tcp_md5_hash_header);
3067
3068int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3069                          const struct sk_buff *skb, unsigned int header_len)
3070{
3071        struct scatterlist sg;
3072        const struct tcphdr *tp = tcp_hdr(skb);
3073        struct hash_desc *desc = &hp->md5_desc;
3074        unsigned int i;
3075        const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3076                                           skb_headlen(skb) - header_len : 0;
3077        const struct skb_shared_info *shi = skb_shinfo(skb);
3078        struct sk_buff *frag_iter;
3079
3080        sg_init_table(&sg, 1);
3081
3082        sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3083        if (crypto_hash_update(desc, &sg, head_data_len))
3084                return 1;
3085
3086        for (i = 0; i < shi->nr_frags; ++i) {
3087                const struct skb_frag_struct *f = &shi->frags[i];
3088                unsigned int offset = f->page_offset;
3089                struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3090
3091                sg_set_page(&sg, page, skb_frag_size(f),
3092                            offset_in_page(offset));
3093                if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3094                        return 1;
3095        }
3096
3097        skb_walk_frags(skb, frag_iter)
3098                if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3099                        return 1;
3100
3101        return 0;
3102}
3103EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3104
3105int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3106{
3107        struct scatterlist sg;
3108
3109        sg_init_one(&sg, key->key, key->keylen);
3110        return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3111}
3112EXPORT_SYMBOL(tcp_md5_hash_key);
3113
3114#endif
3115
3116void tcp_done(struct sock *sk)
3117{
3118        struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3119
3120        if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3121                TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3122
3123        tcp_set_state(sk, TCP_CLOSE);
3124        tcp_clear_xmit_timers(sk);
3125        if (req != NULL)
3126                reqsk_fastopen_remove(sk, req, false);
3127
3128        sk->sk_shutdown = SHUTDOWN_MASK;
3129
3130        if (!sock_flag(sk, SOCK_DEAD))
3131                sk->sk_state_change(sk);
3132        else
3133                inet_csk_destroy_sock(sk);
3134}
3135EXPORT_SYMBOL_GPL(tcp_done);
3136
3137extern struct tcp_congestion_ops tcp_reno;
3138
3139static __initdata unsigned long thash_entries;
3140static int __init set_thash_entries(char *str)
3141{
3142        ssize_t ret;
3143
3144        if (!str)
3145                return 0;
3146
3147        ret = kstrtoul(str, 0, &thash_entries);
3148        if (ret)
3149                return 0;
3150
3151        return 1;
3152}
3153__setup("thash_entries=", set_thash_entries);
3154
3155static void tcp_init_mem(void)
3156{
3157        unsigned long limit = nr_free_buffer_pages() / 8;
3158        limit = max(limit, 128UL);
3159        sysctl_tcp_mem[0] = limit / 4 * 3;
3160        sysctl_tcp_mem[1] = limit;
3161        sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3162}
3163
3164void __init tcp_init(void)
3165{
3166        struct sk_buff *skb = NULL;
3167        unsigned long limit;
3168        int max_rshare, max_wshare, cnt;
3169        unsigned int i;
3170
3171        BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3172
3173        percpu_counter_init(&tcp_sockets_allocated, 0);
3174        percpu_counter_init(&tcp_orphan_count, 0);
3175        tcp_hashinfo.bind_bucket_cachep =
3176                kmem_cache_create("tcp_bind_bucket",
3177                                  sizeof(struct inet_bind_bucket), 0,
3178                                  SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3179
3180        /* Size and allocate the main established and bind bucket
3181         * hash tables.
3182         *
3183         * The methodology is similar to that of the buffer cache.
3184         */
3185        tcp_hashinfo.ehash =
3186                alloc_large_system_hash("TCP established",
3187                                        sizeof(struct inet_ehash_bucket),
3188                                        thash_entries,
3189                                        17, /* one slot per 128 KB of memory */
3190                                        0,
3191                                        NULL,
3192                                        &tcp_hashinfo.ehash_mask,
3193                                        0,
3194                                        thash_entries ? 0 : 512 * 1024);
3195        for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3196                INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3197
3198        if (inet_ehash_locks_alloc(&tcp_hashinfo))
3199                panic("TCP: failed to alloc ehash_locks");
3200        tcp_hashinfo.bhash =
3201                alloc_large_system_hash("TCP bind",
3202                                        sizeof(struct inet_bind_hashbucket),
3203                                        tcp_hashinfo.ehash_mask + 1,
3204                                        17, /* one slot per 128 KB of memory */
3205                                        0,
3206                                        &tcp_hashinfo.bhash_size,
3207                                        NULL,
3208                                        0,
3209                                        64 * 1024);
3210        tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3211        for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3212                spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3213                INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3214        }
3215
3216
3217        cnt = tcp_hashinfo.ehash_mask + 1;
3218
3219        tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3220        sysctl_tcp_max_orphans = cnt / 2;
3221        sysctl_max_syn_backlog = max(128, cnt / 256);
3222
3223        tcp_init_mem();
3224        /* Set per-socket limits to no more than 1/128 the pressure threshold */
3225        limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3226        max_wshare = min(4UL*1024*1024, limit);
3227        max_rshare = min(6UL*1024*1024, limit);
3228
3229        sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3230        sysctl_tcp_wmem[1] = 16*1024;
3231        sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3232
3233        sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3234        sysctl_tcp_rmem[1] = 87380;
3235        sysctl_tcp_rmem[2] = max(87380, max_rshare);
3236
3237        pr_info("Hash tables configured (established %u bind %u)\n",
3238                tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3239
3240        tcp_metrics_init();
3241
3242        tcp_register_congestion_control(&tcp_reno);
3243
3244        tcp_tasklet_init();
3245}
3246
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.