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
 283int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
 284
 285struct percpu_counter tcp_orphan_count;
 286EXPORT_SYMBOL_GPL(tcp_orphan_count);
 287
 288int sysctl_tcp_wmem[3] __read_mostly;
 289int sysctl_tcp_rmem[3] __read_mostly;
 290
 291EXPORT_SYMBOL(sysctl_tcp_rmem);
 292EXPORT_SYMBOL(sysctl_tcp_wmem);
 293
 294atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
 295EXPORT_SYMBOL(tcp_memory_allocated);
 296
 297/*
 298 * Current number of TCP sockets.
 299 */
 300struct percpu_counter tcp_sockets_allocated;
 301EXPORT_SYMBOL(tcp_sockets_allocated);
 302
 303/*
 304 * TCP splice context
 305 */
 306struct tcp_splice_state {
 307        struct pipe_inode_info *pipe;
 308        size_t len;
 309        unsigned int flags;
 310};
 311
 312/*
 313 * Pressure flag: try to collapse.
 314 * Technical note: it is used by multiple contexts non atomically.
 315 * All the __sk_mem_schedule() is of this nature: accounting
 316 * is strict, actions are advisory and have some latency.
 317 */
 318int tcp_memory_pressure __read_mostly;
 319EXPORT_SYMBOL(tcp_memory_pressure);
 320
 321void tcp_enter_memory_pressure(struct sock *sk)
 322{
 323        if (!tcp_memory_pressure) {
 324                NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
 325                tcp_memory_pressure = 1;
 326        }
 327}
 328EXPORT_SYMBOL(tcp_enter_memory_pressure);
 329
 330/* Convert seconds to retransmits based on initial and max timeout */
 331static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
 332{
 333        u8 res = 0;
 334
 335        if (seconds > 0) {
 336                int period = timeout;
 337
 338                res = 1;
 339                while (seconds > period && res < 255) {
 340                        res++;
 341                        timeout <<= 1;
 342                        if (timeout > rto_max)
 343                                timeout = rto_max;
 344                        period += timeout;
 345                }
 346        }
 347        return res;
 348}
 349
 350/* Convert retransmits to seconds based on initial and max timeout */
 351static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
 352{
 353        int period = 0;
 354
 355        if (retrans > 0) {
 356                period = timeout;
 357                while (--retrans) {
 358                        timeout <<= 1;
 359                        if (timeout > rto_max)
 360                                timeout = rto_max;
 361                        period += timeout;
 362                }
 363        }
 364        return period;
 365}
 366
 367/* Address-family independent initialization for a tcp_sock.
 368 *
 369 * NOTE: A lot of things set to zero explicitly by call to
 370 *       sk_alloc() so need not be done here.
 371 */
 372void tcp_init_sock(struct sock *sk)
 373{
 374        struct inet_connection_sock *icsk = inet_csk(sk);
 375        struct tcp_sock *tp = tcp_sk(sk);
 376
 377        skb_queue_head_init(&tp->out_of_order_queue);
 378        tcp_init_xmit_timers(sk);
 379        tcp_prequeue_init(tp);
 380        INIT_LIST_HEAD(&tp->tsq_node);
 381
 382        icsk->icsk_rto = TCP_TIMEOUT_INIT;
 383        tp->mdev = TCP_TIMEOUT_INIT;
 384
 385        /* So many TCP implementations out there (incorrectly) count the
 386         * initial SYN frame in their delayed-ACK and congestion control
 387         * algorithms that we must have the following bandaid to talk
 388         * efficiently to them.  -DaveM
 389         */
 390        tp->snd_cwnd = TCP_INIT_CWND;
 391
 392        /* See draft-stevens-tcpca-spec-01 for discussion of the
 393         * initialization of these values.
 394         */
 395        tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 396        tp->snd_cwnd_clamp = ~0;
 397        tp->mss_cache = TCP_MSS_DEFAULT;
 398
 399        tp->reordering = sysctl_tcp_reordering;
 400        tcp_enable_early_retrans(tp);
 401        icsk->icsk_ca_ops = &tcp_init_congestion_ops;
 402
 403        sk->sk_state = TCP_CLOSE;
 404
 405        sk->sk_write_space = sk_stream_write_space;
 406        sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
 407
 408        icsk->icsk_sync_mss = tcp_sync_mss;
 409
 410        /* TCP Cookie Transactions */
 411        if (sysctl_tcp_cookie_size > 0) {
 412                /* Default, cookies without s_data_payload. */
 413                tp->cookie_values =
 414                        kzalloc(sizeof(*tp->cookie_values),
 415                                sk->sk_allocation);
 416                if (tp->cookie_values != NULL)
 417                        kref_init(&tp->cookie_values->kref);
 418        }
 419        /* Presumed zeroed, in order of appearance:
 420         *      cookie_in_always, cookie_out_never,
 421         *      s_data_constant, s_data_in, s_data_out
 422         */
 423        sk->sk_sndbuf = sysctl_tcp_wmem[1];
 424        sk->sk_rcvbuf = sysctl_tcp_rmem[1];
 425
 426        local_bh_disable();
 427        sock_update_memcg(sk);
 428        sk_sockets_allocated_inc(sk);
 429        local_bh_enable();
 430}
 431EXPORT_SYMBOL(tcp_init_sock);
 432
 433/*
 434 *      Wait for a TCP event.
 435 *
 436 *      Note that we don't need to lock the socket, as the upper poll layers
 437 *      take care of normal races (between the test and the event) and we don't
 438 *      go look at any of the socket buffers directly.
 439 */
 440unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 441{
 442        unsigned int mask;
 443        struct sock *sk = sock->sk;
 444        const struct tcp_sock *tp = tcp_sk(sk);
 445
 446        sock_poll_wait(file, sk_sleep(sk), wait);
 447        if (sk->sk_state == TCP_LISTEN)
 448                return inet_csk_listen_poll(sk);
 449
 450        /* Socket is not locked. We are protected from async events
 451         * by poll logic and correct handling of state changes
 452         * made by other threads is impossible in any case.
 453         */
 454
 455        mask = 0;
 456
 457        /*
 458         * POLLHUP is certainly not done right. But poll() doesn't
 459         * have a notion of HUP in just one direction, and for a
 460         * socket the read side is more interesting.
 461         *
 462         * Some poll() documentation says that POLLHUP is incompatible
 463         * with the POLLOUT/POLLWR flags, so somebody should check this
 464         * all. But careful, it tends to be safer to return too many
 465         * bits than too few, and you can easily break real applications
 466         * if you don't tell them that something has hung up!
 467         *
 468         * Check-me.
 469         *
 470         * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
 471         * our fs/select.c). It means that after we received EOF,
 472         * poll always returns immediately, making impossible poll() on write()
 473         * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
 474         * if and only if shutdown has been made in both directions.
 475         * Actually, it is interesting to look how Solaris and DUX
 476         * solve this dilemma. I would prefer, if POLLHUP were maskable,
 477         * then we could set it on SND_SHUTDOWN. BTW examples given
 478         * in Stevens' books assume exactly this behaviour, it explains
 479         * why POLLHUP is incompatible with POLLOUT.    --ANK
 480         *
 481         * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 482         * blocking on fresh not-connected or disconnected socket. --ANK
 483         */
 484        if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
 485                mask |= POLLHUP;
 486        if (sk->sk_shutdown & RCV_SHUTDOWN)
 487                mask |= POLLIN | POLLRDNORM | POLLRDHUP;
 488
 489        /* Connected? */
 490        if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
 491                int target = sock_rcvlowat(sk, 0, INT_MAX);
 492
 493                if (tp->urg_seq == tp->copied_seq &&
 494                    !sock_flag(sk, SOCK_URGINLINE) &&
 495                    tp->urg_data)
 496                        target++;
 497
 498                /* Potential race condition. If read of tp below will
 499                 * escape above sk->sk_state, we can be illegally awaken
 500                 * in SYN_* states. */
 501                if (tp->rcv_nxt - tp->copied_seq >= target)
 502                        mask |= POLLIN | POLLRDNORM;
 503
 504                if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 505                        if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
 506                                mask |= POLLOUT | POLLWRNORM;
 507                        } else {  /* send SIGIO later */
 508                                set_bit(SOCK_ASYNC_NOSPACE,
 509                                        &sk->sk_socket->flags);
 510                                set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 511
 512                                /* Race breaker. If space is freed after
 513                                 * wspace test but before the flags are set,
 514                                 * IO signal will be lost.
 515                                 */
 516                                if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
 517                                        mask |= POLLOUT | POLLWRNORM;
 518                        }
 519                } else
 520                        mask |= POLLOUT | POLLWRNORM;
 521
 522                if (tp->urg_data & TCP_URG_VALID)
 523                        mask |= POLLPRI;
 524        }
 525        /* This barrier is coupled with smp_wmb() in tcp_reset() */
 526        smp_rmb();
 527        if (sk->sk_err)
 528                mask |= POLLERR;
 529
 530        return mask;
 531}
 532EXPORT_SYMBOL(tcp_poll);
 533
 534int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 535{
 536        struct tcp_sock *tp = tcp_sk(sk);
 537        int answ;
 538
 539        switch (cmd) {
 540        case SIOCINQ:
 541                if (sk->sk_state == TCP_LISTEN)
 542                        return -EINVAL;
 543
 544                lock_sock(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                release_sock(sk);
 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
 624static inline void tcp_push(struct sock *sk, int flags, int mss_now,
 625                            int nonagle)
 626{
 627        if (tcp_send_head(sk)) {
 628                struct tcp_sock *tp = tcp_sk(sk);
 629
 630                if (!(flags & MSG_MORE) || forced_push(tp))
 631                        tcp_mark_push(tp, tcp_write_queue_tail(sk));
 632
 633                tcp_mark_urg(tp, flags);
 634                __tcp_push_pending_frames(sk, mss_now,
 635                                          (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
 636        }
 637}
 638
 639static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 640                                unsigned int offset, size_t len)
 641{
 642        struct tcp_splice_state *tss = rd_desc->arg.data;
 643        int ret;
 644
 645        ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
 646                              tss->flags);
 647        if (ret > 0)
 648                rd_desc->count -= ret;
 649        return ret;
 650}
 651
 652static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 653{
 654        /* Store TCP splice context information in read_descriptor_t. */
 655        read_descriptor_t rd_desc = {
 656                .arg.data = tss,
 657                .count    = tss->len,
 658        };
 659
 660        return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 661}
 662
 663/**
 664 *  tcp_splice_read - splice data from TCP socket to a pipe
 665 * @sock:       socket to splice from
 666 * @ppos:       position (not valid)
 667 * @pipe:       pipe to splice to
 668 * @len:        number of bytes to splice
 669 * @flags:      splice modifier flags
 670 *
 671 * Description:
 672 *    Will read pages from given socket and fill them into a pipe.
 673 *
 674 **/
 675ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 676                        struct pipe_inode_info *pipe, size_t len,
 677                        unsigned int flags)
 678{
 679        struct sock *sk = sock->sk;
 680        struct tcp_splice_state tss = {
 681                .pipe = pipe,
 682                .len = len,
 683                .flags = flags,
 684        };
 685        long timeo;
 686        ssize_t spliced;
 687        int ret;
 688
 689        sock_rps_record_flow(sk);
 690        /*
 691         * We can't seek on a socket input
 692         */
 693        if (unlikely(*ppos))
 694                return -ESPIPE;
 695
 696        ret = spliced = 0;
 697
 698        lock_sock(sk);
 699
 700        timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 701        while (tss.len) {
 702                ret = __tcp_splice_read(sk, &tss);
 703                if (ret < 0)
 704                        break;
 705                else if (!ret) {
 706                        if (spliced)
 707                                break;
 708                        if (sock_flag(sk, SOCK_DONE))
 709                                break;
 710                        if (sk->sk_err) {
 711                                ret = sock_error(sk);
 712                                break;
 713                        }
 714                        if (sk->sk_shutdown & RCV_SHUTDOWN)
 715                                break;
 716                        if (sk->sk_state == TCP_CLOSE) {
 717                                /*
 718                                 * This occurs when user tries to read
 719                                 * from never connected socket.
 720                                 */
 721                                if (!sock_flag(sk, SOCK_DONE))
 722                                        ret = -ENOTCONN;
 723                                break;
 724                        }
 725                        if (!timeo) {
 726                                ret = -EAGAIN;
 727                                break;
 728                        }
 729                        sk_wait_data(sk, &timeo);
 730                        if (signal_pending(current)) {
 731                                ret = sock_intr_errno(timeo);
 732                                break;
 733                        }
 734                        continue;
 735                }
 736                tss.len -= ret;
 737                spliced += ret;
 738
 739                if (!timeo)
 740                        break;
 741                release_sock(sk);
 742                lock_sock(sk);
 743
 744                if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 745                    (sk->sk_shutdown & RCV_SHUTDOWN) ||
 746                    signal_pending(current))
 747                        break;
 748        }
 749
 750        release_sock(sk);
 751
 752        if (spliced)
 753                return spliced;
 754
 755        return ret;
 756}
 757EXPORT_SYMBOL(tcp_splice_read);
 758
 759struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
 760{
 761        struct sk_buff *skb;
 762
 763        /* The TCP header must be at least 32-bit aligned.  */
 764        size = ALIGN(size, 4);
 765
 766        skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 767        if (skb) {
 768                if (sk_wmem_schedule(sk, skb->truesize)) {
 769                        skb_reserve(skb, sk->sk_prot->max_header);
 770                        /*
 771                         * Make sure that we have exactly size bytes
 772                         * available to the caller, no more, no less.
 773                         */
 774                        skb->avail_size = size;
 775                        return skb;
 776                }
 777                __kfree_skb(skb);
 778        } else {
 779                sk->sk_prot->enter_memory_pressure(sk);
 780                sk_stream_moderate_sndbuf(sk);
 781        }
 782        return NULL;
 783}
 784
 785static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 786                                       int large_allowed)
 787{
 788        struct tcp_sock *tp = tcp_sk(sk);
 789        u32 xmit_size_goal, old_size_goal;
 790
 791        xmit_size_goal = mss_now;
 792
 793        if (large_allowed && sk_can_gso(sk)) {
 794                xmit_size_goal = ((sk->sk_gso_max_size - 1) -
 795                                  inet_csk(sk)->icsk_af_ops->net_header_len -
 796                                  inet_csk(sk)->icsk_ext_hdr_len -
 797                                  tp->tcp_header_len);
 798
 799                /* TSQ : try to have two TSO segments in flight */
 800                xmit_size_goal = min_t(u32, xmit_size_goal,
 801                                       sysctl_tcp_limit_output_bytes >> 1);
 802
 803                xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
 804
 805                /* We try hard to avoid divides here */
 806                old_size_goal = tp->xmit_size_goal_segs * mss_now;
 807
 808                if (likely(old_size_goal <= xmit_size_goal &&
 809                           old_size_goal + mss_now > xmit_size_goal)) {
 810                        xmit_size_goal = old_size_goal;
 811                } else {
 812                        tp->xmit_size_goal_segs =
 813                                min_t(u16, xmit_size_goal / mss_now,
 814                                      sk->sk_gso_max_segs);
 815                        xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
 816                }
 817        }
 818
 819        return max(xmit_size_goal, mss_now);
 820}
 821
 822static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 823{
 824        int mss_now;
 825
 826        mss_now = tcp_current_mss(sk);
 827        *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 828
 829        return mss_now;
 830}
 831
 832static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
 833                         size_t psize, int flags)
 834{
 835        struct tcp_sock *tp = tcp_sk(sk);
 836        int mss_now, size_goal;
 837        int err;
 838        ssize_t copied;
 839        long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 840
 841        /* Wait for a connection to finish. */
 842        if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
 843                if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 844                        goto out_err;
 845
 846        clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 847
 848        mss_now = tcp_send_mss(sk, &size_goal, flags);
 849        copied = 0;
 850
 851        err = -EPIPE;
 852        if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 853                goto out_err;
 854
 855        while (psize > 0) {
 856                struct sk_buff *skb = tcp_write_queue_tail(sk);
 857                struct page *page = pages[poffset / PAGE_SIZE];
 858                int copy, i;
 859                int offset = poffset % PAGE_SIZE;
 860                int size = min_t(size_t, psize, PAGE_SIZE - offset);
 861                bool can_coalesce;
 862
 863                if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
 864new_segment:
 865                        if (!sk_stream_memory_free(sk))
 866                                goto wait_for_sndbuf;
 867
 868                        skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
 869                        if (!skb)
 870                                goto wait_for_memory;
 871
 872                        skb_entail(sk, skb);
 873                        copy = size_goal;
 874                }
 875
 876                if (copy > size)
 877                        copy = size;
 878
 879                i = skb_shinfo(skb)->nr_frags;
 880                can_coalesce = skb_can_coalesce(skb, i, page, offset);
 881                if (!can_coalesce && i >= MAX_SKB_FRAGS) {
 882                        tcp_mark_push(tp, skb);
 883                        goto new_segment;
 884                }
 885                if (!sk_wmem_schedule(sk, copy))
 886                        goto wait_for_memory;
 887
 888                if (can_coalesce) {
 889                        skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
 890                } else {
 891                        get_page(page);
 892                        skb_fill_page_desc(skb, i, page, offset, copy);
 893                }
 894
 895                skb->len += copy;
 896                skb->data_len += copy;
 897                skb->truesize += copy;
 898                sk->sk_wmem_queued += copy;
 899                sk_mem_charge(sk, copy);
 900                skb->ip_summed = CHECKSUM_PARTIAL;
 901                tp->write_seq += copy;
 902                TCP_SKB_CB(skb)->end_seq += copy;
 903                skb_shinfo(skb)->gso_segs = 0;
 904
 905                if (!copied)
 906                        TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
 907
 908                copied += copy;
 909                poffset += copy;
 910                if (!(psize -= copy))
 911                        goto out;
 912
 913                if (skb->len < size_goal || (flags & MSG_OOB))
 914                        continue;
 915
 916                if (forced_push(tp)) {
 917                        tcp_mark_push(tp, skb);
 918                        __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
 919                } else if (skb == tcp_send_head(sk))
 920                        tcp_push_one(sk, mss_now);
 921                continue;
 922
 923wait_for_sndbuf:
 924                set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 925wait_for_memory:
 926                tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
 927
 928                if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 929                        goto do_error;
 930
 931                mss_now = tcp_send_mss(sk, &size_goal, flags);
 932        }
 933
 934out:
 935        if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
 936                tcp_push(sk, flags, mss_now, tp->nonagle);
 937        return copied;
 938
 939do_error:
 940        if (copied)
 941                goto out;
 942out_err:
 943        return sk_stream_error(sk, flags, err);
 944}
 945
 946int tcp_sendpage(struct sock *sk, struct page *page, int offset,
 947                 size_t size, int flags)
 948{
 949        ssize_t res;
 950
 951        if (!(sk->sk_route_caps & NETIF_F_SG) ||
 952            !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
 953                return sock_no_sendpage(sk->sk_socket, page, offset, size,
 954                                        flags);
 955
 956        lock_sock(sk);
 957        res = do_tcp_sendpages(sk, &page, offset, size, flags);
 958        release_sock(sk);
 959        return res;
 960}
 961EXPORT_SYMBOL(tcp_sendpage);
 962
 963static inline int select_size(const struct sock *sk, bool sg)
 964{
 965        const struct tcp_sock *tp = tcp_sk(sk);
 966        int tmp = tp->mss_cache;
 967
 968        if (sg) {
 969                if (sk_can_gso(sk)) {
 970                        /* Small frames wont use a full page:
 971                         * Payload will immediately follow tcp header.
 972                         */
 973                        tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
 974                } else {
 975                        int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
 976
 977                        if (tmp >= pgbreak &&
 978                            tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
 979                                tmp = pgbreak;
 980                }
 981        }
 982
 983        return tmp;
 984}
 985
 986void tcp_free_fastopen_req(struct tcp_sock *tp)
 987{
 988        if (tp->fastopen_req != NULL) {
 989                kfree(tp->fastopen_req);
 990                tp->fastopen_req = NULL;
 991        }
 992}
 993
 994static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *size)
 995{
 996        struct tcp_sock *tp = tcp_sk(sk);
 997        int err, flags;
 998
 999        if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1000                return -EOPNOTSUPP;
1001        if (tp->fastopen_req != NULL)
1002                return -EALREADY; /* Another Fast Open is in progress */
1003
1004        tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1005                                   sk->sk_allocation);
1006        if (unlikely(tp->fastopen_req == NULL))
1007                return -ENOBUFS;
1008        tp->fastopen_req->data = msg;
1009
1010        flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1011        err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1012                                    msg->msg_namelen, flags);
1013        *size = tp->fastopen_req->copied;
1014        tcp_free_fastopen_req(tp);
1015        return err;
1016}
1017
1018int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1019                size_t size)
1020{
1021        struct iovec *iov;
1022        struct tcp_sock *tp = tcp_sk(sk);
1023        struct sk_buff *skb;
1024        int iovlen, flags, err, copied = 0;
1025        int mss_now = 0, size_goal, copied_syn = 0, offset = 0;
1026        bool sg;
1027        long timeo;
1028
1029        lock_sock(sk);
1030
1031        flags = msg->msg_flags;
1032        if (flags & MSG_FASTOPEN) {
1033                err = tcp_sendmsg_fastopen(sk, msg, &copied_syn);
1034                if (err == -EINPROGRESS && copied_syn > 0)
1035                        goto out;
1036                else if (err)
1037                        goto out_err;
1038                offset = copied_syn;
1039        }
1040
1041        timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1042
1043        /* Wait for a connection to finish. */
1044        if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1045                if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1046                        goto do_error;
1047
1048        if (unlikely(tp->repair)) {
1049                if (tp->repair_queue == TCP_RECV_QUEUE) {
1050                        copied = tcp_send_rcvq(sk, msg, size);
1051                        goto out;
1052                }
1053
1054                err = -EINVAL;
1055                if (tp->repair_queue == TCP_NO_QUEUE)
1056                        goto out_err;
1057
1058                /* 'common' sending to sendq */
1059        }
1060
1061        /* This should be in poll */
1062        clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1063
1064        mss_now = tcp_send_mss(sk, &size_goal, flags);
1065
1066        /* Ok commence sending. */
1067        iovlen = msg->msg_iovlen;
1068        iov = msg->msg_iov;
1069        copied = 0;
1070
1071        err = -EPIPE;
1072        if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1073                goto out_err;
1074
1075        sg = !!(sk->sk_route_caps & NETIF_F_SG);
1076
1077        while (--iovlen >= 0) {
1078                size_t seglen = iov->iov_len;
1079                unsigned char __user *from = iov->iov_base;
1080
1081                iov++;
1082                if (unlikely(offset > 0)) {  /* Skip bytes copied in SYN */
1083                        if (offset >= seglen) {
1084                                offset -= seglen;
1085                                continue;
1086                        }
1087                        seglen -= offset;
1088                        from += offset;
1089                        offset = 0;
1090                }
1091
1092                while (seglen > 0) {
1093                        int copy = 0;
1094                        int max = size_goal;
1095
1096                        skb = tcp_write_queue_tail(sk);
1097                        if (tcp_send_head(sk)) {
1098                                if (skb->ip_summed == CHECKSUM_NONE)
1099                                        max = mss_now;
1100                                copy = max - skb->len;
1101                        }
1102
1103                        if (copy <= 0) {
1104new_segment:
1105                                /* Allocate new segment. If the interface is SG,
1106                                 * allocate skb fitting to single page.
1107                                 */
1108                                if (!sk_stream_memory_free(sk))
1109                                        goto wait_for_sndbuf;
1110
1111                                skb = sk_stream_alloc_skb(sk,
1112                                                          select_size(sk, sg),
1113                                                          sk->sk_allocation);
1114                                if (!skb)
1115                                        goto wait_for_memory;
1116
1117                                /*
1118                                 * Check whether we can use HW checksum.
1119                                 */
1120                                if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1121                                        skb->ip_summed = CHECKSUM_PARTIAL;
1122
1123                                skb_entail(sk, skb);
1124                                copy = size_goal;
1125                                max = size_goal;
1126                        }
1127
1128                        /* Try to append data to the end of skb. */
1129                        if (copy > seglen)
1130                                copy = seglen;
1131
1132                        /* Where to copy to? */
1133                        if (skb_availroom(skb) > 0) {
1134                                /* We have some space in skb head. Superb! */
1135                                copy = min_t(int, copy, skb_availroom(skb));
1136                                err = skb_add_data_nocache(sk, skb, from, copy);
1137                                if (err)
1138                                        goto do_fault;
1139                        } else {
1140                                bool merge = false;
1141                                int i = skb_shinfo(skb)->nr_frags;
1142                                struct page *page = sk->sk_sndmsg_page;
1143                                int off;
1144
1145                                if (page && page_count(page) == 1)
1146                                        sk->sk_sndmsg_off = 0;
1147
1148                                off = sk->sk_sndmsg_off;
1149
1150                                if (skb_can_coalesce(skb, i, page, off) &&
1151                                    off != PAGE_SIZE) {
1152                                        /* We can extend the last page
1153                                         * fragment. */
1154                                        merge = true;
1155                                } else if (i == MAX_SKB_FRAGS || !sg) {
1156                                        /* Need to add new fragment and cannot
1157                                         * do this because interface is non-SG,
1158                                         * or because all the page slots are
1159                                         * busy. */
1160                                        tcp_mark_push(tp, skb);
1161                                        goto new_segment;
1162                                } else if (page) {
1163                                        if (off == PAGE_SIZE) {
1164                                                put_page(page);
1165                                                sk->sk_sndmsg_page = page = NULL;
1166                                                off = 0;
1167                                        }
1168                                } else
1169                                        off = 0;
1170
1171                                if (copy > PAGE_SIZE - off)
1172                                        copy = PAGE_SIZE - off;
1173
1174                                if (!sk_wmem_schedule(sk, copy))
1175                                        goto wait_for_memory;
1176
1177                                if (!page) {
1178                                        /* Allocate new cache page. */
1179                                        if (!(page = sk_stream_alloc_page(sk)))
1180                                                goto wait_for_memory;
1181                                }
1182
1183                                /* Time to copy data. We are close to
1184                                 * the end! */
1185                                err = skb_copy_to_page_nocache(sk, from, skb,
1186                                                               page, off, copy);
1187                                if (err) {
1188                                        /* If this page was new, give it to the
1189                                         * socket so it does not get leaked.
1190                                         */
1191                                        if (!sk->sk_sndmsg_page) {
1192                                                sk->sk_sndmsg_page = page;
1193                                                sk->sk_sndmsg_off = 0;
1194                                        }
1195                                        goto do_error;
1196                                }
1197
1198                                /* Update the skb. */
1199                                if (merge) {
1200                                        skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1201                                } else {
1202                                        skb_fill_page_desc(skb, i, page, off, copy);
1203                                        if (sk->sk_sndmsg_page) {
1204                                                get_page(page);
1205                                        } else if (off + copy < PAGE_SIZE) {
1206                                                get_page(page);
1207                                                sk->sk_sndmsg_page = page;
1208                                        }
1209                                }
1210
1211                                sk->sk_sndmsg_off = off + copy;
1212                        }
1213
1214                        if (!copied)
1215                                TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1216
1217                        tp->write_seq += copy;
1218                        TCP_SKB_CB(skb)->end_seq += copy;
1219                        skb_shinfo(skb)->gso_segs = 0;
1220
1221                        from += copy;
1222                        copied += copy;
1223                        if ((seglen -= copy) == 0 && iovlen == 0)
1224                                goto out;
1225
1226                        if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1227                                continue;
1228
1229                        if (forced_push(tp)) {
1230                                tcp_mark_push(tp, skb);
1231                                __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1232                        } else if (skb == tcp_send_head(sk))
1233                                tcp_push_one(sk, mss_now);
1234                        continue;
1235
1236wait_for_sndbuf:
1237                        set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1238wait_for_memory:
1239                        if (copied)
1240                                tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1241
1242                        if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1243                                goto do_error;
1244
1245                        mss_now = tcp_send_mss(sk, &size_goal, flags);
1246                }
1247        }
1248
1249out:
1250        if (copied)
1251                tcp_push(sk, flags, mss_now, tp->nonagle);
1252        release_sock(sk);
1253        return copied + copied_syn;
1254
1255do_fault:
1256        if (!skb->len) {
1257                tcp_unlink_write_queue(skb, sk);
1258                /* It is the one place in all of TCP, except connection
1259                 * reset, where we can be unlinking the send_head.
1260                 */
1261                tcp_check_send_head(sk, skb);
1262                sk_wmem_free_skb(sk, skb);
1263        }
1264
1265do_error:
1266        if (copied + copied_syn)
1267                goto out;
1268out_err:
1269        err = sk_stream_error(sk, flags, err);
1270        release_sock(sk);
1271        return err;
1272}
1273EXPORT_SYMBOL(tcp_sendmsg);
1274
1275/*
1276 *      Handle reading urgent data. BSD has very simple semantics for
1277 *      this, no blocking and very strange errors 8)
1278 */
1279
1280static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1281{
1282        struct tcp_sock *tp = tcp_sk(sk);
1283
1284        /* No URG data to read. */
1285        if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1286            tp->urg_data == TCP_URG_READ)
1287                return -EINVAL; /* Yes this is right ! */
1288
1289        if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1290                return -ENOTCONN;
1291
1292        if (tp->urg_data & TCP_URG_VALID) {
1293                int err = 0;
1294                char c = tp->urg_data;
1295
1296                if (!(flags & MSG_PEEK))
1297                        tp->urg_data = TCP_URG_READ;
1298
1299                /* Read urgent data. */
1300                msg->msg_flags |= MSG_OOB;
1301
1302                if (len > 0) {
1303                        if (!(flags & MSG_TRUNC))
1304                                err = memcpy_toiovec(msg->msg_iov, &c, 1);
1305                        len = 1;
1306                } else
1307                        msg->msg_flags |= MSG_TRUNC;
1308
1309                return err ? -EFAULT : len;
1310        }
1311
1312        if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1313                return 0;
1314
1315        /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1316         * the available implementations agree in this case:
1317         * this call should never block, independent of the
1318         * blocking state of the socket.
1319         * Mike <pall@rz.uni-karlsruhe.de>
1320         */
1321        return -EAGAIN;
1322}
1323
1324static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1325{
1326        struct sk_buff *skb;
1327        int copied = 0, err = 0;
1328
1329        /* XXX -- need to support SO_PEEK_OFF */
1330
1331        skb_queue_walk(&sk->sk_write_queue, skb) {
1332                err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1333                if (err)
1334                        break;
1335
1336                copied += skb->len;
1337        }
1338
1339        return err ?: copied;
1340}
1341
1342/* Clean up the receive buffer for full frames taken by the user,
1343 * then send an ACK if necessary.  COPIED is the number of bytes
1344 * tcp_recvmsg has given to the user so far, it speeds up the
1345 * calculation of whether or not we must ACK for the sake of
1346 * a window update.
1347 */
1348void tcp_cleanup_rbuf(struct sock *sk, int copied)
1349{
1350        struct tcp_sock *tp = tcp_sk(sk);
1351        bool time_to_ack = false;
1352
1353        struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1354
1355        WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1356             "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1357             tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1358
1359        if (inet_csk_ack_scheduled(sk)) {
1360                const struct inet_connection_sock *icsk = inet_csk(sk);
1361                   /* Delayed ACKs frequently hit locked sockets during bulk
1362                    * receive. */
1363                if (icsk->icsk_ack.blocked ||
1364                    /* Once-per-two-segments ACK was not sent by tcp_input.c */
1365                    tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1366                    /*
1367                     * If this read emptied read buffer, we send ACK, if
1368                     * connection is not bidirectional, user drained
1369                     * receive buffer and there was a small segment
1370                     * in queue.
1371                     */
1372                    (copied > 0 &&
1373                     ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1374                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1375                       !icsk->icsk_ack.pingpong)) &&
1376                      !atomic_read(&sk->sk_rmem_alloc)))
1377                        time_to_ack = true;
1378        }
1379
1380        /* We send an ACK if we can now advertise a non-zero window
1381         * which has been raised "significantly".
1382         *
1383         * Even if window raised up to infinity, do not send window open ACK
1384         * in states, where we will not receive more. It is useless.
1385         */
1386        if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1387                __u32 rcv_window_now = tcp_receive_window(tp);
1388
1389                /* Optimize, __tcp_select_window() is not cheap. */
1390                if (2*rcv_window_now <= tp->window_clamp) {
1391                        __u32 new_window = __tcp_select_window(sk);
1392
1393                        /* Send ACK now, if this read freed lots of space
1394                         * in our buffer. Certainly, new_window is new window.
1395                         * We can advertise it now, if it is not less than current one.
1396                         * "Lots" means "at least twice" here.
1397                         */
1398                        if (new_window && new_window >= 2 * rcv_window_now)
1399                                time_to_ack = true;
1400                }
1401        }
1402        if (time_to_ack)
1403                tcp_send_ack(sk);
1404}
1405
1406static void tcp_prequeue_process(struct sock *sk)
1407{
1408        struct sk_buff *skb;
1409        struct tcp_sock *tp = tcp_sk(sk);
1410
1411        NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1412
1413        /* RX process wants to run with disabled BHs, though it is not
1414         * necessary */
1415        local_bh_disable();
1416        while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1417                sk_backlog_rcv(sk, skb);
1418        local_bh_enable();
1419
1420        /* Clear memory counter. */
1421        tp->ucopy.memory = 0;
1422}
1423
1424#ifdef CONFIG_NET_DMA
1425static void tcp_service_net_dma(struct sock *sk, bool wait)
1426{
1427        dma_cookie_t done, used;
1428        dma_cookie_t last_issued;
1429        struct tcp_sock *tp = tcp_sk(sk);
1430
1431        if (!tp->ucopy.dma_chan)
1432                return;
1433
1434        last_issued = tp->ucopy.dma_cookie;
1435        dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1436
1437        do {
1438                if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1439                                              last_issued, &done,
1440                                              &used) == DMA_SUCCESS) {
1441                        /* Safe to free early-copied skbs now */
1442                        __skb_queue_purge(&sk->sk_async_wait_queue);
1443                        break;
1444                } else {
1445                        struct sk_buff *skb;
1446                        while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1447                               (dma_async_is_complete(skb->dma_cookie, done,
1448                                                      used) == DMA_SUCCESS)) {
1449                                __skb_dequeue(&sk->sk_async_wait_queue);
1450                                kfree_skb(skb);
1451                        }
1452                }
1453        } while (wait);
1454}
1455#endif
1456
1457static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1458{
1459        struct sk_buff *skb;
1460        u32 offset;
1461
1462        skb_queue_walk(&sk->sk_receive_queue, skb) {
1463                offset = seq - TCP_SKB_CB(skb)->seq;
1464                if (tcp_hdr(skb)->syn)
1465                        offset--;
1466                if (offset < skb->len || tcp_hdr(skb)->fin) {
1467                        *off = offset;
1468                        return skb;
1469                }
1470        }
1471        return NULL;
1472}
1473
1474/*
1475 * This routine provides an alternative to tcp_recvmsg() for routines
1476 * that would like to handle copying from skbuffs directly in 'sendfile'
1477 * fashion.
1478 * Note:
1479 *      - It is assumed that the socket was locked by the caller.
1480 *      - The routine does not block.
1481 *      - At present, there is no support for reading OOB data
1482 *        or for 'peeking' the socket using this routine
1483 *        (although both would be easy to implement).
1484 */
1485int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1486                  sk_read_actor_t recv_actor)
1487{
1488        struct sk_buff *skb;
1489        struct tcp_sock *tp = tcp_sk(sk);
1490        u32 seq = tp->copied_seq;
1491        u32 offset;
1492        int copied = 0;
1493
1494        if (sk->sk_state == TCP_LISTEN)
1495                return -ENOTCONN;
1496        while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1497                if (offset < skb->len) {
1498                        int used;
1499                        size_t len;
1500
1501                        len = skb->len - offset;
1502                        /* Stop reading if we hit a patch of urgent data */
1503                        if (tp->urg_data) {
1504                                u32 urg_offset = tp->urg_seq - seq;
1505                                if (urg_offset < len)
1506                                        len = urg_offset;
1507                                if (!len)
1508                                        break;
1509                        }
1510                        used = recv_actor(desc, skb, offset, len);
1511                        if (used < 0) {
1512                                if (!copied)
1513                                        copied = used;
1514                                break;
1515                        } else if (used <= len) {
1516                                seq += used;
1517                                copied += used;
1518                                offset += used;
1519                        }
1520                        /*
1521                         * If recv_actor drops the lock (e.g. TCP splice
1522                         * receive) the skb pointer might be invalid when
1523                         * getting here: tcp_collapse might have deleted it
1524                         * while aggregating skbs from the socket queue.
1525                         */
1526                        skb = tcp_recv_skb(sk, seq-1, &offset);
1527                        if (!skb || (offset+1 != skb->len))
1528                                break;
1529                }
1530                if (tcp_hdr(skb)->fin) {
1531                        sk_eat_skb(sk, skb, false);
1532                        ++seq;
1533                        break;
1534                }
1535                sk_eat_skb(sk, skb, false);
1536                if (!desc->count)
1537                        break;
1538                tp->copied_seq = seq;
1539        }
1540        tp->copied_seq = seq;
1541
1542        tcp_rcv_space_adjust(sk);
1543
1544        /* Clean up data we have read: This will do ACK frames. */
1545        if (copied > 0)
1546                tcp_cleanup_rbuf(sk, copied);
1547        return copied;
1548}
1549EXPORT_SYMBOL(tcp_read_sock);
1550
1551/*
1552 *      This routine copies from a sock struct into the user buffer.
1553 *
1554 *      Technical note: in 2.3 we work on _locked_ socket, so that
1555 *      tricks with *seq access order and skb->users are not required.
1556 *      Probably, code can be easily improved even more.
1557 */
1558
1559int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1560                size_t len, int nonblock, int flags, int *addr_len)
1561{
1562        struct tcp_sock *tp = tcp_sk(sk);
1563        int copied = 0;
1564        u32 peek_seq;
1565        u32 *seq;
1566        unsigned long used;
1567        int err;
1568        int target;             /* Read at least this many bytes */
1569        long timeo;
1570        struct task_struct *user_recv = NULL;
1571        bool copied_early = false;
1572        struct sk_buff *skb;
1573        u32 urg_hole = 0;
1574
1575        lock_sock(sk);
1576
1577        err = -ENOTCONN;
1578        if (sk->sk_state == TCP_LISTEN)
1579                goto out;
1580
1581        timeo = sock_rcvtimeo(sk, nonblock);
1582
1583        /* Urgent data needs to be handled specially. */
1584        if (flags & MSG_OOB)
1585                goto recv_urg;
1586
1587        if (unlikely(tp->repair)) {
1588                err = -EPERM;
1589                if (!(flags & MSG_PEEK))
1590                        goto out;
1591
1592                if (tp->repair_queue == TCP_SEND_QUEUE)
1593                        goto recv_sndq;
1594
1595                err = -EINVAL;
1596                if (tp->repair_queue == TCP_NO_QUEUE)
1597                        goto out;
1598
1599                /* 'common' recv queue MSG_PEEK-ing */
1600        }
1601
1602        seq = &tp->copied_seq;
1603        if (flags & MSG_PEEK) {
1604                peek_seq = tp->copied_seq;
1605                seq = &peek_seq;
1606        }
1607
1608        target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1609
1610#ifdef CONFIG_NET_DMA
1611        tp->ucopy.dma_chan = NULL;
1612        preempt_disable();
1613        skb = skb_peek_tail(&sk->sk_receive_queue);
1614        {
1615                int available = 0;
1616
1617                if (skb)
1618                        available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1619                if ((available < target) &&
1620                    (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1621                    !sysctl_tcp_low_latency &&
1622                    net_dma_find_channel()) {
1623                        preempt_enable_no_resched();
1624                        tp->ucopy.pinned_list =
1625                                        dma_pin_iovec_pages(msg->msg_iov, len);
1626                } else {
1627                        preempt_enable_no_resched();
1628                }
1629        }
1630#endif
1631
1632        do {
1633                u32 offset;
1634
1635                /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1636                if (tp->urg_data && tp->urg_seq == *seq) {
1637                        if (copied)
1638                                break;
1639                        if (signal_pending(current)) {
1640                                copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1641                                break;
1642                        }
1643                }
1644
1645                /* Next get a buffer. */
1646
1647                skb_queue_walk(&sk->sk_receive_queue, skb) {
1648                        /* Now that we have two receive queues this
1649                         * shouldn't happen.
1650                         */
1651                        if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1652                                 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1653                                 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1654                                 flags))
1655                                break;
1656
1657                        offset = *seq - TCP_SKB_CB(skb)->seq;
1658                        if (tcp_hdr(skb)->syn)
1659                                offset--;
1660                        if (offset < skb->len)
1661                                goto found_ok_skb;
1662                        if (tcp_hdr(skb)->fin)
1663                                goto found_fin_ok;
1664                        WARN(!(flags & MSG_PEEK),
1665                             "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1666                             *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1667                }
1668
1669                /* Well, if we have backlog, try to process it now yet. */
1670
1671                if (copied >= target && !sk->sk_backlog.tail)
1672                        break;
1673
1674                if (copied) {
1675                        if (sk->sk_err ||
1676                            sk->sk_state == TCP_CLOSE ||
1677                            (sk->sk_shutdown & RCV_SHUTDOWN) ||
1678                            !timeo ||
1679                            signal_pending(current))
1680                                break;
1681                } else {
1682                        if (sock_flag(sk, SOCK_DONE))
1683                                break;
1684
1685                        if (sk->sk_err) {
1686                                copied = sock_error(sk);
1687                                break;
1688                        }
1689
1690                        if (sk->sk_shutdown & RCV_SHUTDOWN)
1691                                break;
1692
1693                        if (sk->sk_state == TCP_CLOSE) {
1694                                if (!sock_flag(sk, SOCK_DONE)) {
1695                                        /* This occurs when user tries to read
1696                                         * from never connected socket.
1697                                         */
1698                                        copied = -ENOTCONN;
1699                                        break;
1700                                }
1701                                break;
1702                        }
1703
1704                        if (!timeo) {
1705                                copied = -EAGAIN;
1706                                break;
1707                        }
1708
1709                        if (signal_pending(current)) {
1710                                copied = sock_intr_errno(timeo);
1711                                break;
1712                        }
1713                }
1714
1715                tcp_cleanup_rbuf(sk, copied);
1716
1717                if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1718                        /* Install new reader */
1719                        if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1720                                user_recv = current;
1721                                tp->ucopy.task = user_recv;
1722                                tp->ucopy.iov = msg->msg_iov;
1723                        }
1724
1725                        tp->ucopy.len = len;
1726
1727                        WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1728                                !(flags & (MSG_PEEK | MSG_TRUNC)));
1729
1730                        /* Ugly... If prequeue is not empty, we have to
1731                         * process it before releasing socket, otherwise
1732                         * order will be broken at second iteration.
1733                         * More elegant solution is required!!!
1734                         *
1735                         * Look: we have the following (pseudo)queues:
1736                         *
1737                         * 1. packets in flight
1738                         * 2. backlog
1739                         * 3. prequeue
1740                         * 4. receive_queue
1741                         *
1742                         * Each queue can be processed only if the next ones
1743                         * are empty. At this point we have empty receive_queue.
1744                         * But prequeue _can_ be not empty after 2nd iteration,
1745                         * when we jumped to start of loop because backlog
1746                         * processing added something to receive_queue.
1747                         * We cannot release_sock(), because backlog contains
1748                         * packets arrived _after_ prequeued ones.
1749                         *
1750                         * Shortly, algorithm is clear --- to process all
1751                         * the queues in order. We could make it more directly,
1752                         * requeueing packets from backlog to prequeue, if
1753                         * is not empty. It is more elegant, but eats cycles,
1754                         * unfortunately.
1755                         */
1756                        if (!skb_queue_empty(&tp->ucopy.prequeue))
1757                                goto do_prequeue;
1758
1759                        /* __ Set realtime policy in scheduler __ */
1760                }
1761
1762#ifdef CONFIG_NET_DMA
1763                if (tp->ucopy.dma_chan) {
1764                        if (tp->rcv_wnd == 0 &&
1765                            !skb_queue_empty(&sk->sk_async_wait_queue)) {
1766                                tcp_service_net_dma(sk, true);
1767                                tcp_cleanup_rbuf(sk, copied);
1768                        } else
1769                                dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1770                }
1771#endif
1772                if (copied >= target) {
1773                        /* Do not sleep, just process backlog. */
1774                        release_sock(sk);
1775                        lock_sock(sk);
1776                } else
1777                        sk_wait_data(sk, &timeo);
1778
1779#ifdef CONFIG_NET_DMA
1780                tcp_service_net_dma(sk, false);  /* Don't block */
1781                tp->ucopy.wakeup = 0;
1782#endif
1783
1784                if (user_recv) {
1785                        int chunk;
1786
1787                        /* __ Restore normal policy in scheduler __ */
1788
1789                        if ((chunk = len - tp->ucopy.len) != 0) {
1790                                NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1791                                len -= chunk;
1792                                copied += chunk;
1793                        }
1794
1795                        if (tp->rcv_nxt == tp->copied_seq &&
1796                            !skb_queue_empty(&tp->ucopy.prequeue)) {
1797do_prequeue:
1798                                tcp_prequeue_process(sk);
1799
1800                                if ((chunk = len - tp->ucopy.len) != 0) {
1801                                        NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1802                                        len -= chunk;
1803                                        copied += chunk;
1804                                }
1805                        }
1806                }
1807                if ((flags & MSG_PEEK) &&
1808                    (peek_seq - copied - urg_hole != tp->copied_seq)) {
1809                        net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1810                                            current->comm,
1811                                            task_pid_nr(current));
1812                        peek_seq = tp->copied_seq;
1813                }
1814                continue;
1815
1816        found_ok_skb:
1817                /* Ok so how much can we use? */
1818                used = skb->len - offset;
1819                if (len < used)
1820                        used = len;
1821
1822                /* Do we have urgent data here? */
1823                if (tp->urg_data) {
1824                        u32 urg_offset = tp->urg_seq - *seq;
1825                        if (urg_offset < used) {
1826                                if (!urg_offset) {
1827                                        if (!sock_flag(sk, SOCK_URGINLINE)) {
1828                                                ++*seq;
1829                                                urg_hole++;
1830                                                offset++;
1831                                                used--;
1832                                                if (!used)
1833                                                        goto skip_copy;
1834                                        }
1835                                } else
1836                                        used = urg_offset;
1837                        }
1838                }
1839
1840                if (!(flags & MSG_TRUNC)) {
1841#ifdef CONFIG_NET_DMA
1842                        if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1843                                tp->ucopy.dma_chan = net_dma_find_channel();
1844
1845                        if (tp->ucopy.dma_chan) {
1846                                tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1847                                        tp->ucopy.dma_chan, skb, offset,
1848                                        msg->msg_iov, used,
1849                                        tp->ucopy.pinned_list);
1850
1851                                if (tp->ucopy.dma_cookie < 0) {
1852
1853                                        pr_alert("%s: dma_cookie < 0\n",
1854                                                 __func__);
1855
1856                                        /* Exception. Bailout! */
1857                                        if (!copied)
1858                                                copied = -EFAULT;
1859                                        break;
1860                                }
1861
1862                                dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1863
1864                                if ((offset + used) == skb->len)
1865                                        copied_early = true;
1866
1867                        } else
1868#endif
1869                        {
1870                                err = skb_copy_datagram_iovec(skb, offset,
1871                                                msg->msg_iov, used);
1872                                if (err) {
1873                                        /* Exception. Bailout! */
1874                                        if (!copied)
1875                                                copied = -EFAULT;
1876                                        break;
1877                                }
1878                        }
1879                }
1880
1881                *seq += used;
1882                copied += used;
1883                len -= used;
1884
1885                tcp_rcv_space_adjust(sk);
1886
1887skip_copy:
1888                if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1889                        tp->urg_data = 0;
1890                        tcp_fast_path_check(sk);
1891                }
1892                if (used + offset < skb->len)
1893                        continue;
1894
1895                if (tcp_hdr(skb)->fin)
1896                        goto found_fin_ok;
1897                if (!(flags & MSG_PEEK)) {
1898                        sk_eat_skb(sk, skb, copied_early);
1899                        copied_early = false;
1900                }
1901                continue;
1902
1903        found_fin_ok:
1904                /* Process the FIN. */
1905                ++*seq;
1906                if (!(flags & MSG_PEEK)) {
1907                        sk_eat_skb(sk, skb, copied_early);
1908                        copied_early = false;
1909                }
1910                break;
1911        } while (len > 0);
1912
1913        if (user_recv) {
1914                if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1915                        int chunk;
1916
1917                        tp->ucopy.len = copied > 0 ? len : 0;
1918
1919                        tcp_prequeue_process(sk);
1920
1921                        if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1922                                NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1923                                len -= chunk;
1924                                copied += chunk;
1925                        }
1926                }
1927
1928                tp->ucopy.task = NULL;
1929                tp->ucopy.len = 0;
1930        }
1931
1932#ifdef CONFIG_NET_DMA
1933        tcp_service_net_dma(sk, true);  /* Wait for queue to drain */
1934        tp->ucopy.dma_chan = NULL;
1935
1936        if (tp->ucopy.pinned_list) {
1937                dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1938                tp->ucopy.pinned_list = NULL;
1939        }
1940#endif
1941
1942        /* According to UNIX98, msg_name/msg_namelen are ignored
1943         * on connected socket. I was just happy when found this 8) --ANK
1944         */
1945
1946        /* Clean up data we have read: This will do ACK frames. */
1947        tcp_cleanup_rbuf(sk, copied);
1948
1949        release_sock(sk);
1950        return copied;
1951
1952out:
1953        release_sock(sk);
1954        return err;
1955
1956recv_urg:
1957        err = tcp_recv_urg(sk, msg, len, flags);
1958        goto out;
1959
1960recv_sndq:
1961        err = tcp_peek_sndq(sk, msg, len);
1962        goto out;
1963}
1964EXPORT_SYMBOL(tcp_recvmsg);
1965
1966void tcp_set_state(struct sock *sk, int state)
1967{
1968        int oldstate = sk->sk_state;
1969
1970        switch (state) {
1971        case TCP_ESTABLISHED:
1972                if (oldstate != TCP_ESTABLISHED)
1973                        TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1974                break;
1975
1976        case TCP_CLOSE:
1977                if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1978                        TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1979
1980                sk->sk_prot->unhash(sk);
1981                if (inet_csk(sk)->icsk_bind_hash &&
1982                    !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1983                        inet_put_port(sk);
1984                /* fall through */
1985        default:
1986                if (oldstate == TCP_ESTABLISHED)
1987                        TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1988        }
1989
1990        /* Change state AFTER socket is unhashed to avoid closed
1991         * socket sitting in hash tables.
1992         */
1993        sk->sk_state = state;
1994
1995#ifdef STATE_TRACE
1996        SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1997#endif
1998}
1999EXPORT_SYMBOL_GPL(tcp_set_state);
2000
2001/*
2002 *      State processing on a close. This implements the state shift for
2003 *      sending our FIN frame. Note that we only send a FIN for some
2004 *      states. A shutdown() may have already sent the FIN, or we may be
2005 *      closed.
2006 */
2007
2008static const unsigned char new_state[16] = {
2009  /* current state:        new state:      action:      */
2010  /* (Invalid)          */ TCP_CLOSE,
2011  /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2012  /* TCP_SYN_SENT       */ TCP_CLOSE,
2013  /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2014  /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
2015  /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
2016  /* TCP_TIME_WAIT      */ TCP_CLOSE,
2017  /* TCP_CLOSE          */ TCP_CLOSE,
2018  /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
2019  /* TCP_LAST_ACK       */ TCP_LAST_ACK,
2020  /* TCP_LISTEN         */ TCP_CLOSE,
2021  /* TCP_CLOSING        */ TCP_CLOSING,
2022};
2023
2024static int tcp_close_state(struct sock *sk)
2025{
2026        int next = (int)new_state[sk->sk_state];
2027        int ns = next & TCP_STATE_MASK;
2028
2029        tcp_set_state(sk, ns);
2030
2031        return next & TCP_ACTION_FIN;
2032}
2033
2034/*
2035 *      Shutdown the sending side of a connection. Much like close except
2036 *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2037 */
2038
2039void tcp_shutdown(struct sock *sk, int how)
2040{
2041        /*      We need to grab some memory, and put together a FIN,
2042         *      and then put it into the queue to be sent.
2043         *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2044         */
2045        if (!(how & SEND_SHUTDOWN))
2046                return;
2047
2048        /* If we've already sent a FIN, or it's a closed state, skip this. */
2049        if ((1 << sk->sk_state) &
2050            (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2051             TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2052                /* Clear out any half completed packets.  FIN if needed. */
2053                if (tcp_close_state(sk))
2054                        tcp_send_fin(sk);
2055        }
2056}
2057EXPORT_SYMBOL(tcp_shutdown);
2058
2059bool tcp_check_oom(struct sock *sk, int shift)
2060{
2061        bool too_many_orphans, out_of_socket_memory;
2062
2063        too_many_orphans = tcp_too_many_orphans(sk, shift);
2064        out_of_socket_memory = tcp_out_of_memory(sk);
2065
2066        if (too_many_orphans)
2067                net_info_ratelimited("too many orphaned sockets\n");
2068        if (out_of_socket_memory)
2069                net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2070        return too_many_orphans || out_of_socket_memory;
2071}
2072
2073void tcp_close(struct sock *sk, long timeout)
2074{
2075        struct sk_buff *skb;
2076        int data_was_unread = 0;
2077        int state;
2078
2079        lock_sock(sk);
2080        sk->sk_shutdown = SHUTDOWN_MASK;
2081
2082        if (sk->sk_state == TCP_LISTEN) {
2083                tcp_set_state(sk, TCP_CLOSE);
2084
2085                /* Special case. */
2086                inet_csk_listen_stop(sk);
2087
2088                goto adjudge_to_death;
2089        }
2090
2091        /*  We need to flush the recv. buffs.  We do this only on the
2092         *  descriptor close, not protocol-sourced closes, because the
2093         *  reader process may not have drained the data yet!
2094         */
2095        while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2096                u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2097                          tcp_hdr(skb)->fin;
2098                data_was_unread += len;
2099                __kfree_skb(skb);
2100        }
2101
2102        sk_mem_reclaim(sk);
2103
2104        /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2105        if (sk->sk_state == TCP_CLOSE)
2106                goto adjudge_to_death;
2107
2108        /* As outlined in RFC 2525, section 2.17, we send a RST here because
2109         * data was lost. To witness the awful effects of the old behavior of
2110         * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2111         * GET in an FTP client, suspend the process, wait for the client to
2112         * advertise a zero window, then kill -9 the FTP client, wheee...
2113         * Note: timeout is always zero in such a case.
2114         */
2115        if (unlikely(tcp_sk(sk)->repair)) {
2116                sk->sk_prot->disconnect(sk, 0);
2117        } else if (data_was_unread) {
2118                /* Unread data was tossed, zap the connection. */
2119                NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2120                tcp_set_state(sk, TCP_CLOSE);
2121                tcp_send_active_reset(sk, sk->sk_allocation);
2122        } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2123                /* Check zero linger _after_ checking for unread data. */
2124                sk->sk_prot->disconnect(sk, 0);
2125                NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2126        } else if (tcp_close_state(sk)) {
2127                /* We FIN if the application ate all the data before
2128                 * zapping the connection.
2129                 */
2130
2131                /* RED-PEN. Formally speaking, we have broken TCP state
2132                 * machine. State transitions:
2133                 *
2134                 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2135                 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2136                 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2137                 *
2138                 * are legal only when FIN has been sent (i.e. in window),
2139                 * rather than queued out of window. Purists blame.
2140                 *
2141                 * F.e. "RFC state" is ESTABLISHED,
2142                 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2143                 *
2144                 * The visible declinations are that sometimes
2145                 * we enter time-wait state, when it is not required really
2146                 * (harmless), do not send active resets, when they are
2147                 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2148                 * they look as CLOSING or LAST_ACK for Linux)
2149                 * Probably, I missed some more holelets.
2150                 *                                              --ANK
2151                 */
2152                tcp_send_fin(sk);
2153        }
2154
2155        sk_stream_wait_close(sk, timeout);
2156
2157adjudge_to_death:
2158        state = sk->sk_state;
2159        sock_hold(sk);
2160        sock_orphan(sk);
2161
2162        /* It is the last release_sock in its life. It will remove backlog. */
2163        release_sock(sk);
2164
2165
2166        /* Now socket is owned by kernel and we acquire BH lock
2167           to finish close. No need to check for user refs.
2168         */
2169        local_bh_disable();
2170        bh_lock_sock(sk);
2171        WARN_ON(sock_owned_by_user(sk));
2172
2173        percpu_counter_inc(sk->sk_prot->orphan_count);
2174
2175        /* Have we already been destroyed by a softirq or backlog? */
2176        if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2177                goto out;
2178
2179        /*      This is a (useful) BSD violating of the RFC. There is a
2180         *      problem with TCP as specified in that the other end could
2181         *      keep a socket open forever with no application left this end.
2182         *      We use a 3 minute timeout (about the same as BSD) then kill
2183         *      our end. If they send after that then tough - BUT: long enough
2184         *      that we won't make the old 4*rto = almost no time - whoops
2185         *      reset mistake.
2186         *
2187         *      Nope, it was not mistake. It is really desired behaviour
2188         *      f.e. on http servers, when such sockets are useless, but
2189         *      consume significant resources. Let's do it with special
2190         *      linger2 option.                                 --ANK
2191         */
2192
2193        if (sk->sk_state == TCP_FIN_WAIT2) {
2194                struct tcp_sock *tp = tcp_sk(sk);
2195                if (tp->linger2 < 0) {
2196                        tcp_set_state(sk, TCP_CLOSE);
2197                        tcp_send_active_reset(sk, GFP_ATOMIC);
2198                        NET_INC_STATS_BH(sock_net(sk),
2199                                        LINUX_MIB_TCPABORTONLINGER);
2200                } else {
2201                        const int tmo = tcp_fin_time(sk);
2202
2203                        if (tmo > TCP_TIMEWAIT_LEN) {
2204                                inet_csk_reset_keepalive_timer(sk,
2205                                                tmo - TCP_TIMEWAIT_LEN);
2206                        } else {
2207                                tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2208                                goto out;
2209                        }
2210                }
2211        }
2212        if (sk->sk_state != TCP_CLOSE) {
2213                sk_mem_reclaim(sk);
2214                if (tcp_check_oom(sk, 0)) {
2215                        tcp_set_state(sk, TCP_CLOSE);
2216                        tcp_send_active_reset(sk, GFP_ATOMIC);
2217                        NET_INC_STATS_BH(sock_net(sk),
2218                                        LINUX_MIB_TCPABORTONMEMORY);
2219                }
2220        }
2221
2222        if (sk->sk_state == TCP_CLOSE)
2223                inet_csk_destroy_sock(sk);
2224        /* Otherwise, socket is reprieved until protocol close. */
2225
2226out:
2227        bh_unlock_sock(sk);
2228        local_bh_enable();
2229        sock_put(sk);
2230}
2231EXPORT_SYMBOL(tcp_close);
2232
2233/* These states need RST on ABORT according to RFC793 */
2234
2235static inline bool tcp_need_reset(int state)
2236{
2237        return (1 << state) &
2238               (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2239                TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2240}
2241
2242int tcp_disconnect(struct sock *sk, int flags)
2243{
2244        struct inet_sock *inet = inet_sk(sk);
2245        struct inet_connection_sock *icsk = inet_csk(sk);
2246        struct tcp_sock *tp = tcp_sk(sk);
2247        int err = 0;
2248        int old_state = sk->sk_state;
2249
2250        if (old_state != TCP_CLOSE)
2251                tcp_set_state(sk, TCP_CLOSE);
2252
2253        /* ABORT function of RFC793 */
2254        if (old_state == TCP_LISTEN) {
2255                inet_csk_listen_stop(sk);
2256        } else if (unlikely(tp->repair)) {
2257                sk->sk_err = ECONNABORTED;
2258        } else if (tcp_need_reset(old_state) ||
2259                   (tp->snd_nxt != tp->write_seq &&
2260                    (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2261                /* The last check adjusts for discrepancy of Linux wrt. RFC
2262                 * states
2263                 */
2264                tcp_send_active_reset(sk, gfp_any());
2265                sk->sk_err = ECONNRESET;
2266        } else if (old_state == TCP_SYN_SENT)
2267                sk->sk_err = ECONNRESET;
2268
2269        tcp_clear_xmit_timers(sk);
2270        __skb_queue_purge(&sk->sk_receive_queue);
2271        tcp_write_queue_purge(sk);
2272        __skb_queue_purge(&tp->out_of_order_queue);
2273#ifdef CONFIG_NET_DMA
2274        __skb_queue_purge(&sk->sk_async_wait_queue);
2275#endif
2276
2277        inet->inet_dport = 0;
2278
2279        if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2280                inet_reset_saddr(sk);
2281
2282        sk->sk_shutdown = 0;
2283        sock_reset_flag(sk, SOCK_DONE);
2284        tp->srtt = 0;
2285        if ((tp->write_seq += tp->max_window + 2) == 0)
2286                tp->write_seq = 1;
2287        icsk->icsk_backoff = 0;
2288        tp->snd_cwnd = 2;
2289        icsk->icsk_probes_out = 0;
2290        tp->packets_out = 0;
2291        tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2292        tp->snd_cwnd_cnt = 0;
2293        tp->bytes_acked = 0;
2294        tp->window_clamp = 0;
2295        tcp_set_ca_state(sk, TCP_CA_Open);
2296        tcp_clear_retrans(tp);
2297        inet_csk_delack_init(sk);
2298        tcp_init_send_head(sk);
2299        memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2300        __sk_dst_reset(sk);
2301
2302        WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2303
2304        sk->sk_error_report(sk);
2305        return err;
2306}
2307EXPORT_SYMBOL(tcp_disconnect);
2308
2309static inline bool tcp_can_repair_sock(const struct sock *sk)
2310{
2311        return capable(CAP_NET_ADMIN) &&
2312                ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2313}
2314
2315static int tcp_repair_options_est(struct tcp_sock *tp,
2316                struct tcp_repair_opt __user *optbuf, unsigned int len)
2317{
2318        struct tcp_repair_opt opt;
2319
2320        while (len >= sizeof(opt)) {
2321                if (copy_from_user(&opt, optbuf, sizeof(opt)))
2322                        return -EFAULT;
2323
2324                optbuf++;
2325                len -= sizeof(opt);
2326
2327                switch (opt.opt_code) {
2328                case TCPOPT_MSS:
2329                        tp->rx_opt.mss_clamp = opt.opt_val;
2330                        break;
2331                case TCPOPT_WINDOW:
2332                        {
2333                                u16 snd_wscale = opt.opt_val & 0xFFFF;
2334                                u16 rcv_wscale = opt.opt_val >> 16;
2335
2336                                if (snd_wscale > 14 || rcv_wscale > 14)
2337                                        return -EFBIG;
2338
2339                                tp->rx_opt.snd_wscale = snd_wscale;
2340                                tp->rx_opt.rcv_wscale = rcv_wscale;
2341                                tp->rx_opt.wscale_ok = 1;
2342                        }
2343                        break;
2344                case TCPOPT_SACK_PERM:
2345                        if (opt.opt_val != 0)
2346                                return -EINVAL;
2347
2348                        tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2349                        if (sysctl_tcp_fack)
2350                                tcp_enable_fack(tp);
2351                        break;
2352                case TCPOPT_TIMESTAMP:
2353                        if (opt.opt_val != 0)
2354                                return -EINVAL;
2355
2356                        tp->rx_opt.tstamp_ok = 1;
2357                        break;
2358                }
2359        }
2360
2361        return 0;
2362}
2363
2364/*
2365 *      Socket option code for TCP.
2366 */
2367static int do_tcp_setsockopt(struct sock *sk, int level,
2368                int optname, char __user *optval, unsigned int optlen)
2369{
2370        struct tcp_sock *tp = tcp_sk(sk);
2371        struct inet_connection_sock *icsk = inet_csk(sk);
2372        int val;
2373        int err = 0;
2374
2375        /* These are data/string values, all the others are ints */
2376        switch (optname) {
2377        case TCP_CONGESTION: {
2378                char name[TCP_CA_NAME_MAX];
2379
2380                if (optlen < 1)
2381                        return -EINVAL;
2382
2383                val = strncpy_from_user(name, optval,
2384                                        min_t(long, TCP_CA_NAME_MAX-1, optlen));
2385                if (val < 0)
2386                        return -EFAULT;
2387                name[val] = 0;
2388
2389                lock_sock(sk);
2390                err = tcp_set_congestion_control(sk, name);
2391                release_sock(sk);
2392                return err;
2393        }
2394        case TCP_COOKIE_TRANSACTIONS: {
2395                struct tcp_cookie_transactions ctd;
2396                struct tcp_cookie_values *cvp = NULL;
2397
2398                if (sizeof(ctd) > optlen)
2399                        return -EINVAL;
2400                if (copy_from_user(&ctd, optval, sizeof(ctd)))
2401                        return -EFAULT;
2402
2403                if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2404                    ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2405                        return -EINVAL;
2406
2407                if (ctd.tcpct_cookie_desired == 0) {
2408                        /* default to global value */
2409                } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2410                           ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2411                           ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2412                        return -EINVAL;
2413                }
2414
2415                if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2416                        /* Supercedes all other values */
2417                        lock_sock(sk);
2418                        if (tp->cookie_values != NULL) {
2419                                kref_put(&tp->cookie_values->kref,
2420                                         tcp_cookie_values_release);
2421                                tp->cookie_values = NULL;
2422                        }
2423                        tp->rx_opt.cookie_in_always = 0; /* false */
2424                        tp->rx_opt.cookie_out_never = 1; /* true */
2425                        release_sock(sk);
2426                        return err;
2427                }
2428
2429                /* Allocate ancillary memory before locking.
2430                 */
2431                if (ctd.tcpct_used > 0 ||
2432                    (tp->cookie_values == NULL &&
2433                     (sysctl_tcp_cookie_size > 0 ||
2434                      ctd.tcpct_cookie_desired > 0 ||
2435                      ctd.tcpct_s_data_desired > 0))) {
2436                        cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2437                                      GFP_KERNEL);
2438                        if (cvp == NULL)
2439                                return -ENOMEM;
2440
2441                        kref_init(&cvp->kref);
2442                }
2443                lock_sock(sk);
2444                tp->rx_opt.cookie_in_always =
2445                        (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2446                tp->rx_opt.cookie_out_never = 0; /* false */
2447
2448                if (tp->cookie_values != NULL) {
2449                        if (cvp != NULL) {
2450                                /* Changed values are recorded by a changed
2451                                 * pointer, ensuring the cookie will differ,
2452                                 * without separately hashing each value later.
2453                                 */
2454                                kref_put(&tp->cookie_values->kref,
2455                                         tcp_cookie_values_release);
2456                        } else {
2457                                cvp = tp->cookie_values;
2458                        }
2459                }
2460
2461                if (cvp != NULL) {
2462                        cvp->cookie_desired = ctd.tcpct_cookie_desired;
2463
2464                        if (ctd.tcpct_used > 0) {
2465                                memcpy(cvp->s_data_payload, ctd.tcpct_value,
2466                                       ctd.tcpct_used);
2467                                cvp->s_data_desired = ctd.tcpct_used;
2468                                cvp->s_data_constant = 1; /* true */
2469                        } else {
2470                                /* No constant payload data. */
2471                                cvp->s_data_desired = ctd.tcpct_s_data_desired;
2472                                cvp->s_data_constant = 0; /* false */
2473                        }
2474
2475                        tp->cookie_values = cvp;
2476                }
2477                release_sock(sk);
2478                return err;
2479        }
2480        default:
2481                /* fallthru */
2482                break;
2483        }
2484
2485        if (optlen < sizeof(int))
2486                return -EINVAL;
2487
2488        if (get_user(val, (int __user *)optval))
2489                return -EFAULT;
2490
2491        lock_sock(sk);
2492
2493        switch (optname) {
2494        case TCP_MAXSEG:
2495                /* Values greater than interface MTU won't take effect. However
2496                 * at the point when this call is done we typically don't yet
2497                 * know which interface is going to be used */
2498                if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2499                        err = -EINVAL;
2500                        break;
2501                }
2502                tp->rx_opt.user_mss = val;
2503                break;
2504
2505        case TCP_NODELAY:
2506                if (val) {
2507                        /* TCP_NODELAY is weaker than TCP_CORK, so that
2508                         * this option on corked socket is remembered, but
2509                         * it is not activated until cork is cleared.
2510                         *
2511                         * However, when TCP_NODELAY is set we make
2512                         * an explicit push, which overrides even TCP_CORK
2513                         * for currently queued segments.
2514                         */
2515                        tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2516                        tcp_push_pending_frames(sk);
2517                } else {
2518                        tp->nonagle &= ~TCP_NAGLE_OFF;
2519                }
2520                break;
2521
2522        case TCP_THIN_LINEAR_TIMEOUTS:
2523                if (val < 0 || val > 1)
2524                        err = -EINVAL;
2525                else
2526                        tp->thin_lto = val;
2527                break;
2528
2529        case TCP_THIN_DUPACK:
2530                if (val < 0 || val > 1)
2531                        err = -EINVAL;
2532                else
2533                        tp->thin_dupack = val;
2534                        if (tp->thin_dupack)
2535                                tcp_disable_early_retrans(tp);
2536                break;
2537
2538        case TCP_REPAIR:
2539                if (!tcp_can_repair_sock(sk))
2540                        err = -EPERM;
2541                else if (val == 1) {
2542                        tp->repair = 1;
2543                        sk->sk_reuse = SK_FORCE_REUSE;
2544                        tp->repair_queue = TCP_NO_QUEUE;
2545                } else if (val == 0) {
2546                        tp->repair = 0;
2547                        sk->sk_reuse = SK_NO_REUSE;
2548                        tcp_send_window_probe(sk);
2549                } else
2550                        err = -EINVAL;
2551
2552                break;
2553
2554        case TCP_REPAIR_QUEUE:
2555                if (!tp->repair)
2556                        err = -EPERM;
2557                else if (val < TCP_QUEUES_NR)
2558                        tp->repair_queue = val;
2559                else
2560                        err = -EINVAL;
2561                break;
2562
2563        case TCP_QUEUE_SEQ:
2564                if (sk->sk_state != TCP_CLOSE)
2565                        err = -EPERM;
2566                else if (tp->repair_queue == TCP_SEND_QUEUE)
2567                        tp->write_seq = val;
2568                else if (tp->repair_queue == TCP_RECV_QUEUE)
2569                        tp->rcv_nxt = val;
2570                else
2571                        err = -EINVAL;
2572                break;
2573
2574        case TCP_REPAIR_OPTIONS:
2575                if (!tp->repair)
2576                        err = -EINVAL;
2577                else if (sk->sk_state == TCP_ESTABLISHED)
2578                        err = tcp_repair_options_est(tp,
2579                                        (struct tcp_repair_opt __user *)optval,
2580                                        optlen);
2581                else
2582                        err = -EPERM;
2583                break;
2584
2585        case TCP_CORK:
2586                /* When set indicates to always queue non-full frames.
2587                 * Later the user clears this option and we transmit
2588                 * any pending partial frames in the queue.  This is
2589                 * meant to be used alongside sendfile() to get properly
2590                 * filled frames when the user (for example) must write
2591                 * out headers with a write() call first and then use
2592                 * sendfile to send out the data parts.
2593                 *
2594                 * TCP_CORK can be set together with TCP_NODELAY and it is
2595                 * stronger than TCP_NODELAY.
2596                 */
2597                if (val) {
2598                        tp->nonagle |= TCP_NAGLE_CORK;
2599                } else {
2600                        tp->nonagle &= ~TCP_NAGLE_CORK;
2601                        if (tp->nonagle&TCP_NAGLE_OFF)
2602                                tp->nonagle |= TCP_NAGLE_PUSH;
2603                        tcp_push_pending_frames(sk);
2604                }
2605                break;
2606
2607        case TCP_KEEPIDLE:
2608                if (val < 1 || val > MAX_TCP_KEEPIDLE)
2609                        err = -EINVAL;
2610                else {
2611                        tp->keepalive_time = val * HZ;
2612                        if (sock_flag(sk, SOCK_KEEPOPEN) &&
2613                            !((1 << sk->sk_state) &
2614                              (TCPF_CLOSE | TCPF_LISTEN))) {
2615                                u32 elapsed = keepalive_time_elapsed(tp);
2616                                if (tp->keepalive_time > elapsed)
2617                                        elapsed = tp->keepalive_time - elapsed;
2618                                else
2619                                        elapsed = 0;
2620                                inet_csk_reset_keepalive_timer(sk, elapsed);
2621                        }
2622                }
2623                break;
2624        case TCP_KEEPINTVL:
2625                if (val < 1 || val > MAX_TCP_KEEPINTVL)
2626                        err = -EINVAL;
2627                else
2628                        tp->keepalive_intvl = val * HZ;
2629                break;
2630        case TCP_KEEPCNT:
2631                if (val < 1 || val > MAX_TCP_KEEPCNT)
2632                        err = -EINVAL;
2633                else
2634                        tp->keepalive_probes = val;
2635                break;
2636        case TCP_SYNCNT:
2637                if (val < 1 || val > MAX_TCP_SYNCNT)
2638                        err = -EINVAL;
2639                else
2640                        icsk->icsk_syn_retries = val;
2641                break;
2642
2643        case TCP_LINGER2:
2644                if (val < 0)
2645                        tp->linger2 = -1;
2646                else if (val > sysctl_tcp_fin_timeout / HZ)
2647                        tp->linger2 = 0;
2648                else
2649                        tp->linger2 = val * HZ;
2650                break;
2651
2652        case TCP_DEFER_ACCEPT:
2653                /* Translate value in seconds to number of retransmits */
2654                icsk->icsk_accept_queue.rskq_defer_accept =
2655                        secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2656                                        TCP_RTO_MAX / HZ);
2657                break;
2658
2659        case TCP_WINDOW_CLAMP:
2660                if (!val) {
2661                        if (sk->sk_state != TCP_CLOSE) {
2662                                err = -EINVAL;
2663                                break;
2664                        }
2665                        tp->window_clamp = 0;
2666                } else
2667                        tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2668                                                SOCK_MIN_RCVBUF / 2 : val;
2669                break;
2670
2671        case TCP_QUICKACK:
2672                if (!val) {
2673                        icsk->icsk_ack.pingpong = 1;
2674                } else {
2675                        icsk->icsk_ack.pingpong = 0;
2676                        if ((1 << sk->sk_state) &
2677                            (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2678                            inet_csk_ack_scheduled(sk)) {
2679                                icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2680                                tcp_cleanup_rbuf(sk, 1);
2681                                if (!(val & 1))
2682                                        icsk->icsk_ack.pingpong = 1;
2683                        }
2684                }
2685                break;
2686
2687#ifdef CONFIG_TCP_MD5SIG
2688        case TCP_MD5SIG:
2689                /* Read the IP->Key mappings from userspace */
2690                err = tp->af_specific->md5_parse(sk, optval, optlen);
2691                break;
2692#endif
2693        case TCP_USER_TIMEOUT:
2694                /* Cap the max timeout in ms TCP will retry/retrans
2695                 * before giving up and aborting (ETIMEDOUT) a connection.
2696                 */
2697                if (val < 0)
2698                        err = -EINVAL;
2699                else
2700                        icsk->icsk_user_timeout = msecs_to_jiffies(val);
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
2765        info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2766        info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2767        info->tcpi_snd_mss = tp->mss_cache;
2768        info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2769
2770        if (sk->sk_state == TCP_LISTEN) {
2771                info->tcpi_unacked = sk->sk_ack_backlog;
2772                info->tcpi_sacked = sk->sk_max_ack_backlog;
2773        } else {
2774                info->tcpi_unacked = tp->packets_out;
2775                info->tcpi_sacked = tp->sacked_out;
2776        }
2777        info->tcpi_lost = tp->lost_out;
2778        info->tcpi_retrans = tp->retrans_out;
2779        info->tcpi_fackets = tp->fackets_out;
2780
2781        info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2782        info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2783        info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2784
2785        info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2786        info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2787        info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2788        info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2789        info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2790        info->tcpi_snd_cwnd = tp->snd_cwnd;
2791        info->tcpi_advmss = tp->advmss;
2792        info->tcpi_reordering = tp->reordering;
2793
2794        info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2795        info->tcpi_rcv_space = tp->rcvq_space.space;
2796
2797        info->tcpi_total_retrans = tp->total_retrans;
2798}
2799EXPORT_SYMBOL_GPL(tcp_get_info);
2800
2801static int do_tcp_getsockopt(struct sock *sk, int level,
2802                int optname, char __user *optval, int __user *optlen)
2803{
2804        struct inet_connection_sock *icsk = inet_csk(sk);
2805        struct tcp_sock *tp = tcp_sk(sk);
2806        int val, len;
2807
2808        if (get_user(len, optlen))
2809                return -EFAULT;
2810
2811        len = min_t(unsigned int, len, sizeof(int));
2812
2813        if (len < 0)
2814                return -EINVAL;
2815
2816        switch (optname) {
2817        case TCP_MAXSEG:
2818                val = tp->mss_cache;
2819                if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2820                        val = tp->rx_opt.user_mss;
2821                if (tp->repair)
2822                        val = tp->rx_opt.mss_clamp;
2823                break;
2824        case TCP_NODELAY:
2825                val = !!(tp->nonagle&TCP_NAGLE_OFF);
2826                break;
2827        case TCP_CORK:
2828                val = !!(tp->nonagle&TCP_NAGLE_CORK);
2829                break;
2830        case TCP_KEEPIDLE:
2831                val = keepalive_time_when(tp) / HZ;
2832                break;
2833        case TCP_KEEPINTVL:
2834                val = keepalive_intvl_when(tp) / HZ;
2835                break;
2836        case TCP_KEEPCNT:
2837                val = keepalive_probes(tp);
2838                break;
2839        case TCP_SYNCNT:
2840                val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2841                break;
2842        case TCP_LINGER2:
2843                val = tp->linger2;
2844                if (val >= 0)
2845                        val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2846                break;
2847        case TCP_DEFER_ACCEPT:
2848                val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2849                                      TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2850                break;
2851        case TCP_WINDOW_CLAMP:
2852                val = tp->window_clamp;
2853                break;
2854        case TCP_INFO: {
2855                struct tcp_info info;
2856
2857                if (get_user(len, optlen))
2858                        return -EFAULT;
2859
2860                tcp_get_info(sk, &info);
2861
2862                len = min_t(unsigned int, len, sizeof(info));
2863                if (put_user(len, optlen))
2864                        return -EFAULT;
2865                if (copy_to_user(optval, &info, len))
2866                        return -EFAULT;
2867                return 0;
2868        }
2869        case TCP_QUICKACK:
2870                val = !icsk->icsk_ack.pingpong;
2871                break;
2872
2873        case TCP_CONGESTION:
2874                if (get_user(len, optlen))
2875                        return -EFAULT;
2876                len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2877                if (put_user(len, optlen))
2878                        return -EFAULT;
2879                if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2880                        return -EFAULT;
2881                return 0;
2882
2883        case TCP_COOKIE_TRANSACTIONS: {
2884                struct tcp_cookie_transactions ctd;
2885                struct tcp_cookie_values *cvp = tp->cookie_values;
2886
2887                if (get_user(len, optlen))
2888                        return -EFAULT;
2889                if (len < sizeof(ctd))
2890                        return -EINVAL;
2891
2892                memset(&ctd, 0, sizeof(ctd));
2893                ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2894                                   TCP_COOKIE_IN_ALWAYS : 0)
2895                                | (tp->rx_opt.cookie_out_never ?
2896                                   TCP_COOKIE_OUT_NEVER : 0);
2897
2898                if (cvp != NULL) {
2899                        ctd.tcpct_flags |= (cvp->s_data_in ?
2900                                            TCP_S_DATA_IN : 0)
2901                                         | (cvp->s_data_out ?
2902                                            TCP_S_DATA_OUT : 0);
2903
2904                        ctd.tcpct_cookie_desired = cvp->cookie_desired;
2905                        ctd.tcpct_s_data_desired = cvp->s_data_desired;
2906
2907                        memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2908                               cvp->cookie_pair_size);
2909                        ctd.tcpct_used = cvp->cookie_pair_size;
2910                }
2911
2912                if (put_user(sizeof(ctd), optlen))
2913                        return -EFAULT;
2914                if (copy_to_user(optval, &ctd, sizeof(ctd)))
2915                        return -EFAULT;
2916                return 0;
2917        }
2918        case TCP_THIN_LINEAR_TIMEOUTS:
2919                val = tp->thin_lto;
2920                break;
2921        case TCP_THIN_DUPACK:
2922                val = tp->thin_dupack;
2923                break;
2924
2925        case TCP_REPAIR:
2926                val = tp->repair;
2927                break;
2928
2929        case TCP_REPAIR_QUEUE:
2930                if (tp->repair)
2931                        val = tp->repair_queue;
2932                else
2933                        return -EINVAL;
2934                break;
2935
2936        case TCP_QUEUE_SEQ:
2937                if (tp->repair_queue == TCP_SEND_QUEUE)
2938                        val = tp->write_seq;
2939                else if (tp->repair_queue == TCP_RECV_QUEUE)
2940                        val = tp->rcv_nxt;
2941                else
2942                        return -EINVAL;
2943                break;
2944
2945        case TCP_USER_TIMEOUT:
2946                val = jiffies_to_msecs(icsk->icsk_user_timeout);
2947                break;
2948        default:
2949                return -ENOPROTOOPT;
2950        }
2951
2952        if (put_user(len, optlen))
2953                return -EFAULT;
2954        if (copy_to_user(optval, &val, len))
2955                return -EFAULT;
2956        return 0;
2957}
2958
2959int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2960                   int __user *optlen)
2961{
2962        struct inet_connection_sock *icsk = inet_csk(sk);
2963
2964        if (level != SOL_TCP)
2965                return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2966                                                     optval, optlen);
2967        return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2968}
2969EXPORT_SYMBOL(tcp_getsockopt);
2970
2971#ifdef CONFIG_COMPAT
2972int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2973                          char __user *optval, int __user *optlen)
2974{
2975        if (level != SOL_TCP)
2976                return inet_csk_compat_getsockopt(sk, level, optname,
2977                                                  optval, optlen);
2978        return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2979}
2980EXPORT_SYMBOL(compat_tcp_getsockopt);
2981#endif
2982
2983struct sk_buff *tcp_tso_segment(struct sk_buff *skb,
2984        netdev_features_t features)
2985{
2986        struct sk_buff *segs = ERR_PTR(-EINVAL);
2987        struct tcphdr *th;
2988        unsigned int thlen;
2989        unsigned int seq;
2990        __be32 delta;
2991        unsigned int oldlen;
2992        unsigned int mss;
2993
2994        if (!pskb_may_pull(skb, sizeof(*th)))
2995                goto out;
2996
2997        th = tcp_hdr(skb);
2998        thlen = th->doff * 4;
2999        if (thlen < sizeof(*th))
3000                goto out;
3001
3002        if (!pskb_may_pull(skb, thlen))
3003                goto out;
3004
3005        oldlen = (u16)~skb->len;
3006        __skb_pull(skb, thlen);
3007
3008        mss = skb_shinfo(skb)->gso_size;
3009        if (unlikely(skb->len <= mss))
3010                goto out;
3011
3012        if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
3013                /* Packet is from an untrusted source, reset gso_segs. */
3014                int type = skb_shinfo(skb)->gso_type;
3015
3016                if (unlikely(type &
3017                             ~(SKB_GSO_TCPV4 |
3018                               SKB_GSO_DODGY |
3019                               SKB_GSO_TCP_ECN |
3020                               SKB_GSO_TCPV6 |
3021                               0) ||
3022                             !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
3023                        goto out;
3024
3025                skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
3026
3027                segs = NULL;
3028                goto out;
3029        }
3030
3031        segs = skb_segment(skb, features);
3032        if (IS_ERR(segs))
3033                goto out;
3034
3035        delta = htonl(oldlen + (thlen + mss));
3036
3037        skb = segs;
3038        th = tcp_hdr(skb);
3039        seq = ntohl(th->seq);
3040
3041        do {
3042                th->fin = th->psh = 0;
3043
3044                th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
3045                                       (__force u32)delta));
3046                if (skb->ip_summed != CHECKSUM_PARTIAL)
3047                        th->check =
3048                             csum_fold(csum_partial(skb_transport_header(skb),
3049                                                    thlen, skb->csum));
3050
3051                seq += mss;
3052                skb = skb->next;
3053                th = tcp_hdr(skb);
3054
3055                th->seq = htonl(seq);
3056                th->cwr = 0;
3057        } while (skb->next);
3058
3059        delta = htonl(oldlen + (skb->tail - skb->transport_header) +
3060                      skb->data_len);
3061        th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
3062                                (__force u32)delta));
3063        if (skb->ip_summed != CHECKSUM_PARTIAL)
3064                th->check = csum_fold(csum_partial(skb_transport_header(skb),
3065                                                   thlen, skb->csum));
3066
3067out:
3068        return segs;
3069}
3070EXPORT_SYMBOL(tcp_tso_segment);
3071
3072struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
3073{
3074        struct sk_buff **pp = NULL;
3075        struct sk_buff *p;
3076        struct tcphdr *th;
3077        struct tcphdr *th2;
3078        unsigned int len;
3079        unsigned int thlen;
3080        __be32 flags;
3081        unsigned int mss = 1;
3082        unsigned int hlen;
3083        unsigned int off;
3084        int flush = 1;
3085        int i;
3086
3087        off = skb_gro_offset(skb);
3088        hlen = off + sizeof(*th);
3089        th = skb_gro_header_fast(skb, off);
3090        if (skb_gro_header_hard(skb, hlen)) {
3091                th = skb_gro_header_slow(skb, hlen, off);
3092                if (unlikely(!th))
3093                        goto out;
3094        }
3095
3096        thlen = th->doff * 4;
3097        if (thlen < sizeof(*th))
3098                goto out;
3099
3100        hlen = off + thlen;
3101        if (skb_gro_header_hard(skb, hlen)) {
3102                th = skb_gro_header_slow(skb, hlen, off);
3103                if (unlikely(!th))
3104                        goto out;
3105        }
3106
3107        skb_gro_pull(skb, thlen);
3108
3109        len = skb_gro_len(skb);
3110        flags = tcp_flag_word(th);
3111
3112        for (; (p = *head); head = &p->next) {
3113                if (!NAPI_GRO_CB(p)->same_flow)
3114                        continue;
3115
3116                th2 = tcp_hdr(p);
3117
3118                if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
3119                        NAPI_GRO_CB(p)->same_flow = 0;
3120                        continue;
3121                }
3122
3123                goto found;
3124        }
3125
3126        goto out_check_final;
3127
3128found:
3129        flush = NAPI_GRO_CB(p)->flush;
3130        flush |= (__force int)(flags & TCP_FLAG_CWR);
3131        flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
3132                  ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
3133        flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
3134        for (i = sizeof(*th); i < thlen; i += 4)
3135                flush |= *(u32 *)((u8 *)th + i) ^
3136                         *(u32 *)((u8 *)th2 + i);
3137
3138        mss = skb_shinfo(p)->gso_size;
3139
3140        flush |= (len - 1) >= mss;
3141        flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
3142
3143        if (flush || skb_gro_receive(head, skb)) {
3144                mss = 1;
3145                goto out_check_final;
3146        }
3147
3148        p = *head;
3149        th2 = tcp_hdr(p);
3150        tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
3151
3152out_check_final:
3153        flush = len < mss;
3154        flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
3155                                        TCP_FLAG_RST | TCP_FLAG_SYN |
3156                                        TCP_FLAG_FIN));
3157
3158        if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
3159                pp = head;
3160
3161out:
3162        NAPI_GRO_CB(skb)->flush |= flush;
3163
3164        return pp;
3165}
3166EXPORT_SYMBOL(tcp_gro_receive);
3167
3168int tcp_gro_complete(struct sk_buff *skb)
3169{
3170        struct tcphdr *th = tcp_hdr(skb);
3171
3172        skb->csum_start = skb_transport_header(skb) - skb->head;
3173        skb->csum_offset = offsetof(struct tcphdr, check);
3174        skb->ip_summed = CHECKSUM_PARTIAL;
3175
3176        skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
3177
3178        if (th->cwr)
3179                skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
3180
3181        return 0;
3182}
3183EXPORT_SYMBOL(tcp_gro_complete);
3184
3185#ifdef CONFIG_TCP_MD5SIG
3186static unsigned long tcp_md5sig_users;
3187static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool;
3188static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
3189
3190static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
3191{
3192        int cpu;
3193
3194        for_each_possible_cpu(cpu) {
3195                struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
3196
3197                if (p->md5_desc.tfm)
3198                        crypto_free_hash(p->md5_desc.tfm);
3199        }
3200        free_percpu(pool);
3201}
3202
3203void tcp_free_md5sig_pool(void)
3204{
3205        struct tcp_md5sig_pool __percpu *pool = NULL;
3206
3207        spin_lock_bh(&tcp_md5sig_pool_lock);
3208        if (--tcp_md5sig_users == 0) {
3209                pool = tcp_md5sig_pool;
3210                tcp_md5sig_pool = NULL;
3211        }
3212        spin_unlock_bh(&tcp_md5sig_pool_lock);
3213        if (pool)
3214                __tcp_free_md5sig_pool(pool);
3215}
3216EXPORT_SYMBOL(tcp_free_md5sig_pool);
3217
3218static struct tcp_md5sig_pool __percpu *
3219__tcp_alloc_md5sig_pool(struct sock *sk)
3220{
3221        int cpu;
3222        struct tcp_md5sig_pool __percpu *pool;
3223
3224        pool = alloc_percpu(struct tcp_md5sig_pool);
3225        if (!pool)
3226                return NULL;
3227
3228        for_each_possible_cpu(cpu) {
3229                struct crypto_hash *hash;
3230
3231                hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
3232                if (!hash || IS_ERR(hash))
3233                        goto out_free;
3234
3235                per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
3236        }
3237        return pool;
3238out_free:
3239        __tcp_free_md5sig_pool(pool);
3240        return NULL;
3241}
3242
3243struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
3244{
3245        struct tcp_md5sig_pool __percpu *pool;
3246        bool alloc = false;
3247
3248retry:
3249        spin_lock_bh(&tcp_md5sig_pool_lock);
3250        pool = tcp_md5sig_pool;
3251        if (tcp_md5sig_users++ == 0) {
3252                alloc = true;
3253                spin_unlock_bh(&tcp_md5sig_pool_lock);
3254        } else if (!pool) {
3255                tcp_md5sig_users--;
3256                spin_unlock_bh(&tcp_md5sig_pool_lock);
3257                cpu_relax();
3258                goto retry;
3259        } else
3260                spin_unlock_bh(&tcp_md5sig_pool_lock);
3261
3262        if (alloc) {
3263                /* we cannot hold spinlock here because this may sleep. */
3264                struct tcp_md5sig_pool __percpu *p;
3265
3266                p = __tcp_alloc_md5sig_pool(sk);
3267                spin_lock_bh(&tcp_md5sig_pool_lock);
3268                if (!p) {
3269                        tcp_md5sig_users--;
3270                        spin_unlock_bh(&tcp_md5sig_pool_lock);
3271                        return NULL;
3272                }
3273                pool = tcp_md5sig_pool;
3274                if (pool) {
3275                        /* oops, it has already been assigned. */
3276                        spin_unlock_bh(&tcp_md5sig_pool_lock);
3277                        __tcp_free_md5sig_pool(p);
3278                } else {
3279                        tcp_md5sig_pool = pool = p;
3280                        spin_unlock_bh(&tcp_md5sig_pool_lock);
3281                }
3282        }
3283        return pool;
3284}
3285EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3286
3287
3288/**
3289 *      tcp_get_md5sig_pool - get md5sig_pool for this user
3290 *
3291 *      We use percpu structure, so if we succeed, we exit with preemption
3292 *      and BH disabled, to make sure another thread or softirq handling
3293 *      wont try to get same context.
3294 */
3295struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3296{
3297        struct tcp_md5sig_pool __percpu *p;
3298
3299        local_bh_disable();
3300
3301        spin_lock(&tcp_md5sig_pool_lock);
3302        p = tcp_md5sig_pool;
3303        if (p)
3304                tcp_md5sig_users++;
3305        spin_unlock(&tcp_md5sig_pool_lock);
3306
3307        if (p)
3308                return this_cpu_ptr(p);
3309
3310        local_bh_enable();
3311        return NULL;
3312}
3313EXPORT_SYMBOL(tcp_get_md5sig_pool);
3314
3315void tcp_put_md5sig_pool(void)
3316{
3317        local_bh_enable();
3318        tcp_free_md5sig_pool();
3319}
3320EXPORT_SYMBOL(tcp_put_md5sig_pool);
3321
3322int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3323                        const struct tcphdr *th)
3324{
3325        struct scatterlist sg;
3326        struct tcphdr hdr;
3327        int err;
3328
3329        /* We are not allowed to change tcphdr, make a local copy */
3330        memcpy(&hdr, th, sizeof(hdr));
3331        hdr.check = 0;
3332
3333        /* options aren't included in the hash */
3334        sg_init_one(&sg, &hdr, sizeof(hdr));
3335        err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3336        return err;
3337}
3338EXPORT_SYMBOL(tcp_md5_hash_header);
3339
3340int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3341                          const struct sk_buff *skb, unsigned int header_len)
3342{
3343        struct scatterlist sg;
3344        const struct tcphdr *tp = tcp_hdr(skb);
3345        struct hash_desc *desc = &hp->md5_desc;
3346        unsigned int i;
3347        const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3348                                           skb_headlen(skb) - header_len : 0;
3349        const struct skb_shared_info *shi = skb_shinfo(skb);
3350        struct sk_buff *frag_iter;
3351
3352        sg_init_table(&sg, 1);
3353
3354        sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3355        if (crypto_hash_update(desc, &sg, head_data_len))
3356                return 1;
3357
3358        for (i = 0; i < shi->nr_frags; ++i) {
3359                const struct skb_frag_struct *f = &shi->frags[i];
3360                struct page *page = skb_frag_page(f);
3361                sg_set_page(&sg, page, skb_frag_size(f), f->page_offset);
3362                if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3363                        return 1;
3364        }
3365
3366        skb_walk_frags(skb, frag_iter)
3367                if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3368                        return 1;
3369
3370        return 0;
3371}
3372EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3373
3374int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3375{
3376        struct scatterlist sg;
3377
3378        sg_init_one(&sg, key->key, key->keylen);
3379        return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3380}
3381EXPORT_SYMBOL(tcp_md5_hash_key);
3382
3383#endif
3384
3385/* Each Responder maintains up to two secret values concurrently for
3386 * efficient secret rollover.  Each secret value has 4 states:
3387 *
3388 * Generating.  (tcp_secret_generating != tcp_secret_primary)
3389 *    Generates new Responder-Cookies, but not yet used for primary
3390 *    verification.  This is a short-term state, typically lasting only
3391 *    one round trip time (RTT).
3392 *
3393 * Primary.  (tcp_secret_generating == tcp_secret_primary)
3394 *    Used both for generation and primary verification.
3395 *
3396 * Retiring.  (tcp_secret_retiring != tcp_secret_secondary)
3397 *    Used for verification, until the first failure that can be
3398 *    verified by the newer Generating secret.  At that time, this
3399 *    cookie's state is changed to Secondary, and the Generating
3400 *    cookie's state is changed to Primary.  This is a short-term state,
3401 *    typically lasting only one round trip time (RTT).
3402 *
3403 * Secondary.  (tcp_secret_retiring == tcp_secret_secondary)
3404 *    Used for secondary verification, after primary verification
3405 *    failures.  This state lasts no more than twice the Maximum Segment
3406 *    Lifetime (2MSL).  Then, the secret is discarded.
3407 */
3408struct tcp_cookie_secret {
3409        /* The secret is divided into two parts.  The digest part is the
3410         * equivalent of previously hashing a secret and saving the state,
3411         * and serves as an initialization vector (IV).  The message part
3412         * serves as the trailing secret.
3413         */
3414        u32                             secrets[COOKIE_WORKSPACE_WORDS];
3415        unsigned long                   expires;
3416};
3417
3418#define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3419#define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3420#define TCP_SECRET_LIFE (HZ * 600)
3421
3422static struct tcp_cookie_secret tcp_secret_one;
3423static struct tcp_cookie_secret tcp_secret_two;
3424
3425/* Essentially a circular list, without dynamic allocation. */
3426static struct tcp_cookie_secret *tcp_secret_generating;
3427static struct tcp_cookie_secret *tcp_secret_primary;
3428static struct tcp_cookie_secret *tcp_secret_retiring;
3429static struct tcp_cookie_secret *tcp_secret_secondary;
3430
3431static DEFINE_SPINLOCK(tcp_secret_locker);
3432
3433/* Select a pseudo-random word in the cookie workspace.
3434 */
3435static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3436{
3437        return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3438}
3439
3440/* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3441 * Called in softirq context.
3442 * Returns: 0 for success.
3443 */
3444int tcp_cookie_generator(u32 *bakery)
3445{
3446        unsigned long jiffy = jiffies;
3447
3448        if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3449                spin_lock_bh(&tcp_secret_locker);
3450                if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3451                        /* refreshed by another */
3452                        memcpy(bakery,
3453                               &tcp_secret_generating->secrets[0],
3454                               COOKIE_WORKSPACE_WORDS);
3455                } else {
3456                        /* still needs refreshing */
3457                        get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3458
3459                        /* The first time, paranoia assumes that the
3460                         * randomization function isn't as strong.  But,
3461                         * this secret initialization is delayed until
3462                         * the last possible moment (packet arrival).
3463                         * Although that time is observable, it is
3464                         * unpredictably variable.  Mash in the most
3465                         * volatile clock bits available, and expire the
3466                         * secret extra quickly.
3467                         */
3468                        if (unlikely(tcp_secret_primary->expires ==
3469                                     tcp_secret_secondary->expires)) {
3470                                struct timespec tv;
3471
3472                                getnstimeofday(&tv);
3473                                bakery[COOKIE_DIGEST_WORDS+0] ^=
3474                                        (u32)tv.tv_nsec;
3475
3476                                tcp_secret_secondary->expires = jiffy
3477                                        + TCP_SECRET_1MSL
3478                                        + (0x0f & tcp_cookie_work(bakery, 0));
3479                        } else {
3480                                tcp_secret_secondary->expires = jiffy
3481                                        + TCP_SECRET_LIFE
3482                                        + (0xff & tcp_cookie_work(bakery, 1));
3483                                tcp_secret_primary->expires = jiffy
3484                                        + TCP_SECRET_2MSL
3485                                        + (0x1f & tcp_cookie_work(bakery, 2));
3486                        }
3487                        memcpy(&tcp_secret_secondary->secrets[0],
3488                               bakery, COOKIE_WORKSPACE_WORDS);
3489
3490                        rcu_assign_pointer(tcp_secret_generating,
3491                                           tcp_secret_secondary);
3492                        rcu_assign_pointer(tcp_secret_retiring,
3493                                           tcp_secret_primary);
3494                        /*
3495                         * Neither call_rcu() nor synchronize_rcu() needed.
3496                         * Retiring data is not freed.  It is replaced after
3497                         * further (locked) pointer updates, and a quiet time
3498                         * (minimum 1MSL, maximum LIFE - 2MSL).
3499                         */
3500                }
3501                spin_unlock_bh(&tcp_secret_locker);
3502        } else {
3503                rcu_read_lock_bh();
3504                memcpy(bakery,
3505                       &rcu_dereference(tcp_secret_generating)->secrets[0],
3506                       COOKIE_WORKSPACE_WORDS);
3507                rcu_read_unlock_bh();
3508        }
3509        return 0;
3510}
3511EXPORT_SYMBOL(tcp_cookie_generator);
3512
3513void tcp_done(struct sock *sk)
3514{
3515        if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3516                TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3517
3518        tcp_set_state(sk, TCP_CLOSE);
3519        tcp_clear_xmit_timers(sk);
3520
3521        sk->sk_shutdown = SHUTDOWN_MASK;
3522
3523        if (!sock_flag(sk, SOCK_DEAD))
3524                sk->sk_state_change(sk);
3525        else
3526                inet_csk_destroy_sock(sk);
3527}
3528EXPORT_SYMBOL_GPL(tcp_done);
3529
3530extern struct tcp_congestion_ops tcp_reno;
3531
3532static __initdata unsigned long thash_entries;
3533static int __init set_thash_entries(char *str)
3534{
3535        ssize_t ret;
3536
3537        if (!str)
3538                return 0;
3539
3540        ret = kstrtoul(str, 0, &thash_entries);
3541        if (ret)
3542                return 0;
3543
3544        return 1;
3545}
3546__setup("thash_entries=", set_thash_entries);
3547
3548void tcp_init_mem(struct net *net)
3549{
3550        unsigned long limit = nr_free_buffer_pages() / 8;
3551        limit = max(limit, 128UL);
3552        net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3;
3553        net->ipv4.sysctl_tcp_mem[1] = limit;
3554        net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2;
3555}
3556
3557void __init tcp_init(void)
3558{
3559        struct sk_buff *skb = NULL;
3560        unsigned long limit;
3561        int max_rshare, max_wshare, cnt;
3562        unsigned int i;
3563        unsigned long jiffy = jiffies;
3564
3565        BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3566
3567        percpu_counter_init(&tcp_sockets_allocated, 0);
3568        percpu_counter_init(&tcp_orphan_count, 0);
3569        tcp_hashinfo.bind_bucket_cachep =
3570                kmem_cache_create("tcp_bind_bucket",
3571                                  sizeof(struct inet_bind_bucket), 0,
3572                                  SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3573
3574        /* Size and allocate the main established and bind bucket
3575         * hash tables.
3576         *
3577         * The methodology is similar to that of the buffer cache.
3578         */
3579        tcp_hashinfo.ehash =
3580                alloc_large_system_hash("TCP established",
3581                                        sizeof(struct inet_ehash_bucket),
3582                                        thash_entries,
3583                                        (totalram_pages >= 128 * 1024) ?
3584                                        13 : 15,
3585                                        0,
3586                                        NULL,
3587                                        &tcp_hashinfo.ehash_mask,
3588                                        0,
3589                                        thash_entries ? 0 : 512 * 1024);
3590        for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3591                INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3592                INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3593        }
3594        if (inet_ehash_locks_alloc(&tcp_hashinfo))
3595                panic("TCP: failed to alloc ehash_locks");
3596        tcp_hashinfo.bhash =
3597                alloc_large_system_hash("TCP bind",
3598                                        sizeof(struct inet_bind_hashbucket),
3599                                        tcp_hashinfo.ehash_mask + 1,
3600                                        (totalram_pages >= 128 * 1024) ?
3601                                        13 : 15,
3602                                        0,
3603                                        &tcp_hashinfo.bhash_size,
3604                                        NULL,
3605                                        0,
3606                                        64 * 1024);
3607        tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3608        for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3609                spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3610                INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3611        }
3612
3613
3614        cnt = tcp_hashinfo.ehash_mask + 1;
3615
3616        tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3617        sysctl_tcp_max_orphans = cnt / 2;
3618        sysctl_max_syn_backlog = max(128, cnt / 256);
3619
3620        tcp_init_mem(&init_net);
3621        /* Set per-socket limits to no more than 1/128 the pressure threshold */
3622        limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3623        max_wshare = min(4UL*1024*1024, limit);
3624        max_rshare = min(6UL*1024*1024, limit);
3625
3626        sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3627        sysctl_tcp_wmem[1] = 16*1024;
3628        sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3629
3630        sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3631        sysctl_tcp_rmem[1] = 87380;
3632        sysctl_tcp_rmem[2] = max(87380, max_rshare);
3633
3634        pr_info("Hash tables configured (established %u bind %u)\n",
3635                tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3636
3637        tcp_metrics_init();
3638
3639        tcp_register_congestion_control(&tcp_reno);
3640
3641        memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3642        memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3643        tcp_secret_one.expires = jiffy; /* past due */
3644        tcp_secret_two.expires = jiffy; /* past due */
3645        tcp_secret_generating = &tcp_secret_one;
3646        tcp_secret_primary = &tcp_secret_one;
3647        tcp_secret_retiring = &tcp_secret_two;
3648        tcp_secret_secondary = &tcp_secret_two;
3649        tcp_tasklet_init();
3650}
3651
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.