net/ipv4/tcp.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the  BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              Implementation of the Transmission Control Protocol(TCP).
   7  *
   8  * Authors:     Ross Biro
   9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
  11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
  12  *              Florian La Roche, <flla@stud.uni-sb.de>
  13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
  15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
  16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
  17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  18  *              Jorge Cwik, <jorge@laser.satlink.net>
  19  *
  20  * Fixes:
  21  *              Alan Cox        :       Numerous verify_area() calls
  22  *              Alan Cox        :       Set the ACK bit on a reset
  23  *              Alan Cox        :       Stopped it crashing if it closed while
  24  *                                      sk->inuse=1 and was trying to connect
  25  *                                      (tcp_err()).
  26  *              Alan Cox        :       All icmp error handling was broken
  27  *                                      pointers passed where wrong and the
  28  *                                      socket was looked up backwards. Nobody
  29  *                                      tested any icmp error code obviously.
  30  *              Alan Cox        :       tcp_err() now handled properly. It
  31  *                                      wakes people on errors. poll
  32  *                                      behaves and the icmp error race
  33  *                                      has gone by moving it into sock.c
  34  *              Alan Cox        :       tcp_send_reset() fixed to work for
  35  *                                      everything not just packets for
  36  *                                      unknown sockets.
  37  *              Alan Cox        :       tcp option processing.
  38  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
  39  *                                      syn rule wrong]
  40  *              Herp Rosmanith  :       More reset fixes
  41  *              Alan Cox        :       No longer acks invalid rst frames.
  42  *                                      Acking any kind of RST is right out.
  43  *              Alan Cox        :       Sets an ignore me flag on an rst
  44  *                                      receive otherwise odd bits of prattle
  45  *                                      escape still
  46  *              Alan Cox        :       Fixed another acking RST frame bug.
  47  *                                      Should stop LAN workplace lockups.
  48  *              Alan Cox        :       Some tidyups using the new skb list
  49  *                                      facilities
  50  *              Alan Cox        :       sk->keepopen now seems to work
  51  *              Alan Cox        :       Pulls options out correctly on accepts
  52  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
  53  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
  54  *                                      bit to skb ops.
  55  *              Alan Cox        :       Tidied tcp_data to avoid a potential
  56  *                                      nasty.
  57  *              Alan Cox        :       Added some better commenting, as the
  58  *                                      tcp is hard to follow
  59  *              Alan Cox        :       Removed incorrect check for 20 * psh
  60  *      Michael O'Reilly        :       ack < copied bug fix.
  61  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
  62  *              Alan Cox        :       FIN with no memory -> CRASH
  63  *              Alan Cox        :       Added socket option proto entries.
  64  *                                      Also added awareness of them to accept.
  65  *              Alan Cox        :       Added TCP options (SOL_TCP)
  66  *              Alan Cox        :       Switched wakeup calls to callbacks,
  67  *                                      so the kernel can layer network
  68  *                                      sockets.
  69  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
  70  *              Alan Cox        :       Handle FIN (more) properly (we hope).
  71  *              Alan Cox        :       RST frames sent on unsynchronised
  72  *                                      state ack error.
  73  *              Alan Cox        :       Put in missing check for SYN bit.
  74  *              Alan Cox        :       Added tcp_select_window() aka NET2E
  75  *                                      window non shrink trick.
  76  *              Alan Cox        :       Added a couple of small NET2E timer
  77  *                                      fixes
  78  *              Charles Hedrick :       TCP fixes
  79  *              Toomas Tamm     :       TCP window fixes
  80  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
  81  *              Charles Hedrick :       Rewrote most of it to actually work
  82  *              Linus           :       Rewrote tcp_read() and URG handling
  83  *                                      completely
  84  *              Gerhard Koerting:       Fixed some missing timer handling
  85  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
  86  *              Gerhard Koerting:       PC/TCP workarounds
  87  *              Adam Caldwell   :       Assorted timer/timing errors
  88  *              Matthew Dillon  :       Fixed another RST bug
  89  *              Alan Cox        :       Move to kernel side addressing changes.
  90  *              Alan Cox        :       Beginning work on TCP fastpathing
  91  *                                      (not yet usable)
  92  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
  93  *              Alan Cox        :       TCP fast path debugging
  94  *              Alan Cox        :       Window clamping
  95  *              Michael Riepe   :       Bug in tcp_check()
  96  *              Matt Dillon     :       More TCP improvements and RST bug fixes
  97  *              Matt Dillon     :       Yet more small nasties remove from the
  98  *                                      TCP code (Be very nice to this man if
  99  *                                      tcp finally works 100%) 8)
 100  *              Alan Cox        :       BSD accept semantics.
 101  *              Alan Cox        :       Reset on closedown bug.
 102  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
 103  *              Michael Pall    :       Handle poll() after URG properly in
 104  *                                      all cases.
 105  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
 106  *                                      (multi URG PUSH broke rlogin).
 107  *              Michael Pall    :       Fix the multi URG PUSH problem in
 108  *                                      tcp_readable(), poll() after URG
 109  *                                      works now.
 110  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
 111  *                                      BSD api.
 112  *              Alan Cox        :       Changed the semantics of sk->socket to
 113  *                                      fix a race and a signal problem with
 114  *                                      accept() and async I/O.
 115  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
 116  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
 117  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
 118  *                                      clients/servers which listen in on
 119  *                                      fixed ports.
 120  *              Alan Cox        :       Cleaned the above up and shrank it to
 121  *                                      a sensible code size.
 122  *              Alan Cox        :       Self connect lockup fix.
 123  *              Alan Cox        :       No connect to multicast.
 124  *              Ross Biro       :       Close unaccepted children on master
 125  *                                      socket close.
 126  *              Alan Cox        :       Reset tracing code.
 127  *              Alan Cox        :       Spurious resets on shutdown.
 128  *              Alan Cox        :       Giant 15 minute/60 second timer error
 129  *              Alan Cox        :       Small whoops in polling before an
 130  *                                      accept.
 131  *              Alan Cox        :       Kept the state trace facility since
 132  *                                      it's handy for debugging.
 133  *              Alan Cox        :       More reset handler fixes.
 134  *              Alan Cox        :       Started rewriting the code based on
 135  *                                      the RFC's for other useful protocol
 136  *                                      references see: Comer, KA9Q NOS, and
 137  *                                      for a reference on the difference
 138  *                                      between specifications and how BSD
 139  *                                      works see the 4.4lite source.
 140  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
 141  *                                      close.
 142  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
 143  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
 144  *              Alan Cox        :       Reimplemented timers as per the RFC
 145  *                                      and using multiple timers for sanity.
 146  *              Alan Cox        :       Small bug fixes, and a lot of new
 147  *                                      comments.
 148  *              Alan Cox        :       Fixed dual reader crash by locking
 149  *                                      the buffers (much like datagram.c)
 150  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
 151  *                                      now gets fed up of retrying without
 152  *                                      (even a no space) answer.
 153  *              Alan Cox        :       Extracted closing code better
 154  *              Alan Cox        :       Fixed the closing state machine to
 155  *                                      resemble the RFC.
 156  *              Alan Cox        :       More 'per spec' fixes.
 157  *              Jorge Cwik      :       Even faster checksumming.
 158  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
 159  *                                      only frames. At least one pc tcp stack
 160  *                                      generates them.
 161  *              Alan Cox        :       Cache last socket.
 162  *              Alan Cox        :       Per route irtt.
 163  *              Matt Day        :       poll()->select() match BSD precisely on error
 164  *              Alan Cox        :       New buffers
 165  *              Marc Tamsky     :       Various sk->prot->retransmits and
 166  *                                      sk->retransmits misupdating fixed.
 167  *                                      Fixed tcp_write_timeout: stuck close,
 168  *                                      and TCP syn retries gets used now.
 169  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
 170  *                                      ack if state is TCP_CLOSED.
 171  *              Alan Cox        :       Look up device on a retransmit - routes may
 172  *                                      change. Doesn't yet cope with MSS shrink right
 173  *                                      but it's a start!
 174  *              Marc Tamsky     :       Closing in closing fixes.
 175  *              Mike Shaver     :       RFC1122 verifications.
 176  *              Alan Cox        :       rcv_saddr errors.
 177  *              Alan Cox        :       Block double connect().
 178  *              Alan Cox        :       Small hooks for enSKIP.
 179  *              Alexey Kuznetsov:       Path MTU discovery.
 180  *              Alan Cox        :       Support soft errors.
 181  *              Alan Cox        :       Fix MTU discovery pathological case
 182  *                                      when the remote claims no mtu!
 183  *              Marc Tamsky     :       TCP_CLOSE fix.
 184  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
 185  *                                      window but wrong (fixes NT lpd problems)
 186  *              Pedro Roque     :       Better TCP window handling, delayed ack.
 187  *              Joerg Reuter    :       No modification of locked buffers in
 188  *                                      tcp_do_retransmit()
 189  *              Eric Schenk     :       Changed receiver side silly window
 190  *                                      avoidance algorithm to BSD style
 191  *                                      algorithm. This doubles throughput
 192  *                                      against machines running Solaris,
 193  *                                      and seems to result in general
 194  *                                      improvement.
 195  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
 196  *      Willy Konynenberg       :       Transparent proxying support.
 197  *      Mike McLagan            :       Routing by source
 198  *              Keith Owens     :       Do proper merging with partial SKB's in
 199  *                                      tcp_do_sendmsg to avoid burstiness.
 200  *              Eric Schenk     :       Fix fast close down bug with
 201  *                                      shutdown() followed by close().
 202  *              Andi Kleen      :       Make poll agree with SIGIO
 203  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
 204  *                                      lingertime == 0 (RFC 793 ABORT Call)
 205  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
 206  *                                      csum_and_copy_from_user() if possible.
 207  *
 208  *              This program is free software; you can redistribute it and/or
 209  *              modify it under the terms of the GNU General Public License
 210  *              as published by the Free Software Foundation; either version
 211  *              2 of the License, or(at your option) any later version.
 212  *
 213  * Description of States:
 214  *
 215  *      TCP_SYN_SENT            sent a connection request, waiting for ack
 216  *
 217  *      TCP_SYN_RECV            received a connection request, sent ack,
 218  *                              waiting for final ack in three-way handshake.
 219  *
 220  *      TCP_ESTABLISHED         connection established
 221  *
 222  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
 223  *                              transmission of remaining buffered data
 224  *
 225  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
 226  *                              to shutdown
 227  *
 228  *      TCP_CLOSING             both sides have shutdown but we still have
 229  *                              data we have to finish sending
 230  *
 231  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
 232  *                              closed, can only be entered from FIN_WAIT2
 233  *                              or CLOSING.  Required because the other end
 234  *                              may not have gotten our last ACK causing it
 235  *                              to retransmit the data packet (which we ignore)
 236  *
 237  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
 238  *                              us to finish writing our data and to shutdown
 239  *                              (we have to close() to move on to LAST_ACK)
 240  *
 241  *      TCP_LAST_ACK            out side has shutdown after remote has
 242  *                              shutdown.  There may still be data in our
 243  *                              buffer that we have to finish sending
 244  *
 245  *      TCP_CLOSE               socket is finished
 246  */
 247
 248 #define pr_fmt(fmt) "TCP: " fmt
 249
 250 #include <crypto/hash.h>
 251 #include <linux/kernel.h>
 252 #include <linux/module.h>
 253 #include <linux/types.h>
 254 #include <linux/fcntl.h>
 255 #include <linux/poll.h>
 256 #include <linux/inet_diag.h>
 257 #include <linux/init.h>
 258 #include <linux/fs.h>
 259 #include <linux/skbuff.h>
 260 #include <linux/scatterlist.h>
 261 #include <linux/splice.h>
 262 #include <linux/net.h>
 263 #include <linux/socket.h>
 264 #include <linux/random.h>
 265 #include <linux/bootmem.h>
 266 #include <linux/highmem.h>
 267 #include <linux/swap.h>
 268 #include <linux/cache.h>
 269 #include <linux/err.h>
 270 #include <linux/time.h>
 271 #include <linux/slab.h>
 272 #include <linux/errqueue.h>
 273
 274 #include <net/icmp.h>
 275 #include <net/inet_common.h>
 276 #include <net/tcp.h>
 277 #include <net/xfrm.h>
 278 #include <net/ip.h>
 279 #include <net/sock.h>
 280
 281 #include <linux/uaccess.h>
 282 #include <asm/ioctls.h>
 283 #include <net/busy_poll.h>
 284
 285 int sysctl_tcp_min_tso_segs __read_mostly = 2;
 286
 287 int sysctl_tcp_autocorking __read_mostly = 1;
 288
 289 struct percpu_counter tcp_orphan_count;
 290 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 291
 292 long sysctl_tcp_mem[3] __read_mostly;
 293 int sysctl_tcp_wmem[3] __read_mostly;
 294 int sysctl_tcp_rmem[3] __read_mostly;
 295
 296 EXPORT_SYMBOL(sysctl_tcp_mem);
 297 EXPORT_SYMBOL(sysctl_tcp_rmem);
 298 EXPORT_SYMBOL(sysctl_tcp_wmem);
 299
 300 atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
 301 EXPORT_SYMBOL(tcp_memory_allocated);
 302
 303 /*
 304  * Current number of TCP sockets.
 305  */
 306 struct percpu_counter tcp_sockets_allocated;
 307 EXPORT_SYMBOL(tcp_sockets_allocated);
 308
 309 /*
 310  * TCP splice context
 311  */
 312 struct tcp_splice_state {
 313         struct pipe_inode_info *pipe;
 314         size_t len;
 315         unsigned int flags;
 316 };
 317
 318 /*
 319  * Pressure flag: try to collapse.
 320  * Technical note: it is used by multiple contexts non atomically.
 321  * All the __sk_mem_schedule() is of this nature: accounting
 322  * is strict, actions are advisory and have some latency.
 323  */
 324 unsigned long tcp_memory_pressure __read_mostly;
 325 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
 326
 327 void tcp_enter_memory_pressure(struct sock *sk)
 328 {
 329         unsigned long val;
 330
 331         if (READ_ONCE(tcp_memory_pressure))
 332                 return;
 333         val = jiffies;
 334
 335         if (!val)
 336                 val--;
 337         if (!cmpxchg(&tcp_memory_pressure, 0, val))
 338                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
 339 }
 340 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
 341
 342 void tcp_leave_memory_pressure(struct sock *sk)
 343 {
 344         unsigned long val;
 345
 346         if (!READ_ONCE(tcp_memory_pressure))
 347                 return;
 348         val = xchg(&tcp_memory_pressure, 0);
 349         if (val)
 350                 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
 351                               jiffies_to_msecs(jiffies - val));
 352 }
 353 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
 354
 355 /* Convert seconds to retransmits based on initial and max timeout */
 356 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
 357 {
 358         u8 res = 0;
 359
 360         if (seconds > 0) {
 361                 int period = timeout;
 362
 363                 res = 1;
 364                 while (seconds > period && res < 255) {
 365                         res++;
 366                         timeout <<= 1;
 367                         if (timeout > rto_max)
 368                                 timeout = rto_max;
 369                         period += timeout;
 370                 }
 371         }
 372         return res;
 373 }
 374
 375 /* Convert retransmits to seconds based on initial and max timeout */
 376 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
 377 {
 378         int period = 0;
 379
 380         if (retrans > 0) {
 381                 period = timeout;
 382                 while (--retrans) {
 383                         timeout <<= 1;
 384                         if (timeout > rto_max)
 385                                 timeout = rto_max;
 386                         period += timeout;
 387                 }
 388         }
 389         return period;
 390 }
 391
 392 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
 393 {
 394         u32 rate = READ_ONCE(tp->rate_delivered);
 395         u32 intv = READ_ONCE(tp->rate_interval_us);
 396         u64 rate64 = 0;
 397
 398         if (rate && intv) {
 399                 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
 400                 do_div(rate64, intv);
 401         }
 402         return rate64;
 403 }
 404
 405 /* Address-family independent initialization for a tcp_sock.
 406  *
 407  * NOTE: A lot of things set to zero explicitly by call to
 408  *       sk_alloc() so need not be done here.
 409  */
 410 void tcp_init_sock(struct sock *sk)
 411 {
 412         struct inet_connection_sock *icsk = inet_csk(sk);
 413         struct tcp_sock *tp = tcp_sk(sk);
 414
 415         tp->out_of_order_queue = RB_ROOT;
 416         tcp_init_xmit_timers(sk);
 417         INIT_LIST_HEAD(&tp->tsq_node);
 418
 419         icsk->icsk_rto = TCP_TIMEOUT_INIT;
 420         tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
 421         minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
 422
 423         /* So many TCP implementations out there (incorrectly) count the
 424          * initial SYN frame in their delayed-ACK and congestion control
 425          * algorithms that we must have the following bandaid to talk
 426          * efficiently to them.  -DaveM
 427          */
 428         tp->snd_cwnd = TCP_INIT_CWND;
 429
 430         /* There's a bubble in the pipe until at least the first ACK. */
 431         tp->app_limited = ~0U;
 432
 433         /* See draft-stevens-tcpca-spec-01 for discussion of the
 434          * initialization of these values.
 435          */
 436         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 437         tp->snd_cwnd_clamp = ~0;
 438         tp->mss_cache = TCP_MSS_DEFAULT;
 439
 440         tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
 441         tcp_assign_congestion_control(sk);
 442
 443         tp->tsoffset = 0;
 444
 445         sk->sk_state = TCP_CLOSE;
 446
 447         sk->sk_write_space = sk_stream_write_space;
 448         sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
 449
 450         icsk->icsk_sync_mss = tcp_sync_mss;
 451
 452         sk->sk_sndbuf = sysctl_tcp_wmem[1];
 453         sk->sk_rcvbuf = sysctl_tcp_rmem[1];
 454
 455         sk_sockets_allocated_inc(sk);
 456 }
 457 EXPORT_SYMBOL(tcp_init_sock);
 458
 459 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags, struct sk_buff *skb)
 460 {
 461         if (tsflags && skb) {
 462                 struct skb_shared_info *shinfo = skb_shinfo(skb);
 463                 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 464
 465                 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
 466                 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
 467                         tcb->txstamp_ack = 1;
 468                 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
 469                         shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
 470         }
 471 }
 472
 473 /*
 474  *      Wait for a TCP event.
 475  *
 476  *      Note that we don't need to lock the socket, as the upper poll layers
 477  *      take care of normal races (between the test and the event) and we don't
 478  *      go look at any of the socket buffers directly.
 479  */
 480 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 481 {
 482         unsigned int mask;
 483         struct sock *sk = sock->sk;
 484         const struct tcp_sock *tp = tcp_sk(sk);
 485         int state;
 486
 487         sock_rps_record_flow(sk);
 488
 489         sock_poll_wait(file, sk_sleep(sk), wait);
 490
 491         state = sk_state_load(sk);
 492         if (state == TCP_LISTEN)
 493                 return inet_csk_listen_poll(sk);
 494
 495         /* Socket is not locked. We are protected from async events
 496          * by poll logic and correct handling of state changes
 497          * made by other threads is impossible in any case.
 498          */
 499
 500         mask = 0;
 501
 502         /*
 503          * POLLHUP is certainly not done right. But poll() doesn't
 504          * have a notion of HUP in just one direction, and for a
 505          * socket the read side is more interesting.
 506          *
 507          * Some poll() documentation says that POLLHUP is incompatible
 508          * with the POLLOUT/POLLWR flags, so somebody should check this
 509          * all. But careful, it tends to be safer to return too many
 510          * bits than too few, and you can easily break real applications
 511          * if you don't tell them that something has hung up!
 512          *
 513          * Check-me.
 514          *
 515          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
 516          * our fs/select.c). It means that after we received EOF,
 517          * poll always returns immediately, making impossible poll() on write()
 518          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
 519          * if and only if shutdown has been made in both directions.
 520          * Actually, it is interesting to look how Solaris and DUX
 521          * solve this dilemma. I would prefer, if POLLHUP were maskable,
 522          * then we could set it on SND_SHUTDOWN. BTW examples given
 523          * in Stevens' books assume exactly this behaviour, it explains
 524          * why POLLHUP is incompatible with POLLOUT.    --ANK
 525          *
 526          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 527          * blocking on fresh not-connected or disconnected socket. --ANK
 528          */
 529         if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
 530                 mask |= POLLHUP;
 531         if (sk->sk_shutdown & RCV_SHUTDOWN)
 532                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
 533
 534         /* Connected or passive Fast Open socket? */
 535         if (state != TCP_SYN_SENT &&
 536             (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
 537                 int target = sock_rcvlowat(sk, 0, INT_MAX);
 538
 539                 if (tp->urg_seq == tp->copied_seq &&
 540                     !sock_flag(sk, SOCK_URGINLINE) &&
 541                     tp->urg_data)
 542                         target++;
 543
 544                 if (tp->rcv_nxt - tp->copied_seq >= target)
 545                         mask |= POLLIN | POLLRDNORM;
 546
 547                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 548                         if (sk_stream_is_writeable(sk)) {
 549                                 mask |= POLLOUT | POLLWRNORM;
 550                         } else {  /* send SIGIO later */
 551                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 552                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 553
 554                                 /* Race breaker. If space is freed after
 555                                  * wspace test but before the flags are set,
 556                                  * IO signal will be lost. Memory barrier
 557                                  * pairs with the input side.
 558                                  */
 559                                 smp_mb__after_atomic();
 560                                 if (sk_stream_is_writeable(sk))
 561                                         mask |= POLLOUT | POLLWRNORM;
 562                         }
 563                 } else
 564                         mask |= POLLOUT | POLLWRNORM;
 565
 566                 if (tp->urg_data & TCP_URG_VALID)
 567                         mask |= POLLPRI;
 568         } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
 569                 /* Active TCP fastopen socket with defer_connect
 570                  * Return POLLOUT so application can call write()
 571                  * in order for kernel to generate SYN+data
 572                  */
 573                 mask |= POLLOUT | POLLWRNORM;
 574         }
 575         /* This barrier is coupled with smp_wmb() in tcp_reset() */
 576         smp_rmb();
 577         if (sk->sk_err || !skb_queue_empty_lockless(&sk->sk_error_queue))
 578                 mask |= POLLERR;
 579
 580         return mask;
 581 }
 582 EXPORT_SYMBOL(tcp_poll);
 583
 584 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 585 {
 586         struct tcp_sock *tp = tcp_sk(sk);
 587         int answ;
 588         bool slow;
 589
 590         switch (cmd) {
 591         case SIOCINQ:
 592                 if (sk->sk_state == TCP_LISTEN)
 593                         return -EINVAL;
 594
 595                 slow = lock_sock_fast(sk);
 596                 answ = tcp_inq(sk);
 597                 unlock_sock_fast(sk, slow);
 598                 break;
 599         case SIOCATMARK:
 600                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 601                 break;
 602         case SIOCOUTQ:
 603                 if (sk->sk_state == TCP_LISTEN)
 604                         return -EINVAL;
 605
 606                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 607                         answ = 0;
 608                 else
 609                         answ = tp->write_seq - tp->snd_una;
 610                 break;
 611         case SIOCOUTQNSD:
 612                 if (sk->sk_state == TCP_LISTEN)
 613                         return -EINVAL;
 614
 615                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 616                         answ = 0;
 617                 else
 618                         answ = tp->write_seq - tp->snd_nxt;
 619                 break;
 620         default:
 621                 return -ENOIOCTLCMD;
 622         }
 623
 624         return put_user(answ, (int __user *)arg);
 625 }
 626 EXPORT_SYMBOL(tcp_ioctl);
 627
 628 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 629 {
 630         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 631         tp->pushed_seq = tp->write_seq;
 632 }
 633
 634 static inline bool forced_push(const struct tcp_sock *tp)
 635 {
 636         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 637 }
 638
 639 static void skb_entail(struct sock *sk, struct sk_buff *skb)
 640 {
 641         struct tcp_sock *tp = tcp_sk(sk);
 642         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 643
 644         skb->csum    = 0;
 645         tcb->seq     = tcb->end_seq = tp->write_seq;
 646         tcb->tcp_flags = TCPHDR_ACK;
 647         tcb->sacked  = 0;
 648         __skb_header_release(skb);
 649         tcp_add_write_queue_tail(sk, skb);
 650         sk->sk_wmem_queued += skb->truesize;
 651         sk_mem_charge(sk, skb->truesize);
 652         if (tp->nonagle & TCP_NAGLE_PUSH)
 653                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 654
 655         tcp_slow_start_after_idle_check(sk);
 656 }
 657
 658 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
 659 {
 660         if (flags & MSG_OOB)
 661                 tp->snd_up = tp->write_seq;
 662 }
 663
 664 /* If a not yet filled skb is pushed, do not send it if
 665  * we have data packets in Qdisc or NIC queues :
 666  * Because TX completion will happen shortly, it gives a chance
 667  * to coalesce future sendmsg() payload into this skb, without
 668  * need for a timer, and with no latency trade off.
 669  * As packets containing data payload have a bigger truesize
 670  * than pure acks (dataless) packets, the last checks prevent
 671  * autocorking if we only have an ACK in Qdisc/NIC queues,
 672  * or if TX completion was delayed after we processed ACK packet.
 673  */
 674 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
 675                                 int size_goal)
 676 {
 677         return skb->len < size_goal &&
 678                sysctl_tcp_autocorking &&
 679                skb != tcp_write_queue_head(sk) &&
 680                refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
 681 }
 682
 683 static void tcp_push(struct sock *sk, int flags, int mss_now,
 684                      int nonagle, int size_goal)
 685 {
 686         struct tcp_sock *tp = tcp_sk(sk);
 687         struct sk_buff *skb;
 688
 689         if (!tcp_send_head(sk))
 690                 return;
 691
 692         skb = tcp_write_queue_tail(sk);
 693         if (!(flags & MSG_MORE) || forced_push(tp))
 694                 tcp_mark_push(tp, skb);
 695
 696         tcp_mark_urg(tp, flags);
 697
 698         if (tcp_should_autocork(sk, skb, size_goal)) {
 699
 700                 /* avoid atomic op if TSQ_THROTTLED bit is already set */
 701                 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
 702                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
 703                         set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
 704                 }
 705                 /* It is possible TX completion already happened
 706                  * before we set TSQ_THROTTLED.
 707                  */
 708                 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
 709                         return;
 710         }
 711
 712         if (flags & MSG_MORE)
 713                 nonagle = TCP_NAGLE_CORK;
 714
 715         __tcp_push_pending_frames(sk, mss_now, nonagle);
 716 }
 717
 718 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 719                                 unsigned int offset, size_t len)
 720 {
 721         struct tcp_splice_state *tss = rd_desc->arg.data;
 722         int ret;
 723
 724         ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
 725                               min(rd_desc->count, len), tss->flags);
 726         if (ret > 0)
 727                 rd_desc->count -= ret;
 728         return ret;
 729 }
 730
 731 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 732 {
 733         /* Store TCP splice context information in read_descriptor_t. */
 734         read_descriptor_t rd_desc = {
 735                 .arg.data = tss,
 736                 .count    = tss->len,
 737         };
 738
 739         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 740 }
 741
 742 /**
 743  *  tcp_splice_read - splice data from TCP socket to a pipe
 744  * @sock:       socket to splice from
 745  * @ppos:       position (not valid)
 746  * @pipe:       pipe to splice to
 747  * @len:        number of bytes to splice
 748  * @flags:      splice modifier flags
 749  *
 750  * Description:
 751  *    Will read pages from given socket and fill them into a pipe.
 752  *
 753  **/
 754 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 755                         struct pipe_inode_info *pipe, size_t len,
 756                         unsigned int flags)
 757 {
 758         struct sock *sk = sock->sk;
 759         struct tcp_splice_state tss = {
 760                 .pipe = pipe,
 761                 .len = len,
 762                 .flags = flags,
 763         };
 764         long timeo;
 765         ssize_t spliced;
 766         int ret;
 767
 768         sock_rps_record_flow(sk);
 769         /*
 770          * We can't seek on a socket input
 771          */
 772         if (unlikely(*ppos))
 773                 return -ESPIPE;
 774
 775         ret = spliced = 0;
 776
 777         lock_sock(sk);
 778
 779         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 780         while (tss.len) {
 781                 ret = __tcp_splice_read(sk, &tss);
 782                 if (ret < 0)
 783                         break;
 784                 else if (!ret) {
 785                         if (spliced)
 786                                 break;
 787                         if (sock_flag(sk, SOCK_DONE))
 788                                 break;
 789                         if (sk->sk_err) {
 790                                 ret = sock_error(sk);
 791                                 break;
 792                         }
 793                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 794                                 break;
 795                         if (sk->sk_state == TCP_CLOSE) {
 796                                 /*
 797                                  * This occurs when user tries to read
 798                                  * from never connected socket.
 799                                  */
 800                                 if (!sock_flag(sk, SOCK_DONE))
 801                                         ret = -ENOTCONN;
 802                                 break;
 803                         }
 804                         if (!timeo) {
 805                                 ret = -EAGAIN;
 806                                 break;
 807                         }
 808                         /* if __tcp_splice_read() got nothing while we have
 809                          * an skb in receive queue, we do not want to loop.
 810                          * This might happen with URG data.
 811                          */
 812                         if (!skb_queue_empty(&sk->sk_receive_queue))
 813                                 break;
 814                         sk_wait_data(sk, &timeo, NULL);
 815                         if (signal_pending(current)) {
 816                                 ret = sock_intr_errno(timeo);
 817                                 break;
 818                         }
 819                         continue;
 820                 }
 821                 tss.len -= ret;
 822                 spliced += ret;
 823
 824                 if (!timeo)
 825                         break;
 826                 release_sock(sk);
 827                 lock_sock(sk);
 828
 829                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 830                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 831                     signal_pending(current))
 832                         break;
 833         }
 834
 835         release_sock(sk);
 836
 837         if (spliced)
 838                 return spliced;
 839
 840         return ret;
 841 }
 842 EXPORT_SYMBOL(tcp_splice_read);
 843
 844 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
 845                                     bool force_schedule)
 846 {
 847         struct sk_buff *skb;
 848
 849         /* The TCP header must be at least 32-bit aligned.  */
 850         size = ALIGN(size, 4);
 851
 852         if (unlikely(tcp_under_memory_pressure(sk)))
 853                 sk_mem_reclaim_partial(sk);
 854
 855         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 856         if (likely(skb)) {
 857                 bool mem_scheduled;
 858
 859                 if (force_schedule) {
 860                         mem_scheduled = true;
 861                         sk_forced_mem_schedule(sk, skb->truesize);
 862                 } else {
 863                         mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
 864                 }
 865                 if (likely(mem_scheduled)) {
 866                         skb_reserve(skb, sk->sk_prot->max_header);
 867                         /*
 868                          * Make sure that we have exactly size bytes
 869                          * available to the caller, no more, no less.
 870                          */
 871                         skb->reserved_tailroom = skb->end - skb->tail - size;
 872                         return skb;
 873                 }
 874                 __kfree_skb(skb);
 875         } else {
 876                 sk->sk_prot->enter_memory_pressure(sk);
 877                 sk_stream_moderate_sndbuf(sk);
 878         }
 879         return NULL;
 880 }
 881
 882 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 883                                        int large_allowed)
 884 {
 885         struct tcp_sock *tp = tcp_sk(sk);
 886         u32 new_size_goal, size_goal;
 887
 888         if (!large_allowed || !sk_can_gso(sk))
 889                 return mss_now;
 890
 891         /* Note : tcp_tso_autosize() will eventually split this later */
 892         new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
 893         new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
 894
 895         /* We try hard to avoid divides here */
 896         size_goal = tp->gso_segs * mss_now;
 897         if (unlikely(new_size_goal < size_goal ||
 898                      new_size_goal >= size_goal + mss_now)) {
 899                 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
 900                                      sk->sk_gso_max_segs);
 901                 size_goal = tp->gso_segs * mss_now;
 902         }
 903
 904         return max(size_goal, mss_now);
 905 }
 906
 907 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 908 {
 909         int mss_now;
 910
 911         mss_now = tcp_current_mss(sk);
 912         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 913
 914         return mss_now;
 915 }
 916
 917 /* In some cases, both sendpage() and sendmsg() could have added
 918  * an skb to the write queue, but failed adding payload on it.
 919  * We need to remove it to consume less memory, but more
 920  * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
 921  * users.
 922  */
 923 static void tcp_remove_empty_skb(struct sock *sk, struct sk_buff *skb)
 924 {
 925         if (skb && !skb->len &&
 926             TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
 927                 tcp_unlink_write_queue(skb, sk);
 928                 tcp_check_send_head(sk, skb);
 929                 sk_wmem_free_skb(sk, skb);
 930         }
 931 }
 932
 933 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
 934                          size_t size, int flags)
 935 {
 936         struct tcp_sock *tp = tcp_sk(sk);
 937         int mss_now, size_goal;
 938         int err;
 939         ssize_t copied;
 940         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 941
 942         /* Wait for a connection to finish. One exception is TCP Fast Open
 943          * (passive side) where data is allowed to be sent before a connection
 944          * is fully established.
 945          */
 946         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
 947             !tcp_passive_fastopen(sk)) {
 948                 err = sk_stream_wait_connect(sk, &timeo);
 949                 if (err != 0)
 950                         goto out_err;
 951         }
 952
 953         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 954
 955         mss_now = tcp_send_mss(sk, &size_goal, flags);
 956         copied = 0;
 957
 958         err = -EPIPE;
 959         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 960                 goto out_err;
 961
 962         while (size > 0) {
 963                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 964                 int copy, i;
 965                 bool can_coalesce;
 966
 967                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0 ||
 968                     !tcp_skb_can_collapse_to(skb)) {
 969 new_segment:
 970                         if (!sk_stream_memory_free(sk))
 971                                 goto wait_for_sndbuf;
 972
 973                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
 974                                                   skb_queue_empty(&sk->sk_write_queue));
 975                         if (!skb)
 976                                 goto wait_for_memory;
 977
 978                         skb_entail(sk, skb);
 979                         copy = size_goal;
 980                 }
 981
 982                 if (copy > size)
 983                         copy = size;
 984
 985                 i = skb_shinfo(skb)->nr_frags;
 986                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 987                 if (!can_coalesce && i >= sysctl_max_skb_frags) {
 988                         tcp_mark_push(tp, skb);
 989                         goto new_segment;
 990                 }
 991                 if (!sk_wmem_schedule(sk, copy))
 992                         goto wait_for_memory;
 993
 994                 if (can_coalesce) {
 995                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
 996                 } else {
 997                         get_page(page);
 998                         skb_fill_page_desc(skb, i, page, offset, copy);
 999                 }
1000                 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1001
1002                 skb->len += copy;
1003                 skb->data_len += copy;
1004                 skb->truesize += copy;
1005                 sk->sk_wmem_queued += copy;
1006                 sk_mem_charge(sk, copy);
1007                 skb->ip_summed = CHECKSUM_PARTIAL;
1008                 tp->write_seq += copy;
1009                 TCP_SKB_CB(skb)->end_seq += copy;
1010                 tcp_skb_pcount_set(skb, 0);
1011
1012                 if (!copied)
1013                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1014
1015                 copied += copy;
1016                 offset += copy;
1017                 size -= copy;
1018                 if (!size)
1019                         goto out;
1020
1021                 if (skb->len < size_goal || (flags & MSG_OOB))
1022                         continue;
1023
1024                 if (forced_push(tp)) {
1025                         tcp_mark_push(tp, skb);
1026                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1027                 } else if (skb == tcp_send_head(sk))
1028                         tcp_push_one(sk, mss_now);
1029                 continue;
1030
1031 wait_for_sndbuf:
1032                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1033 wait_for_memory:
1034                 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1035                          TCP_NAGLE_PUSH, size_goal);
1036
1037                 err = sk_stream_wait_memory(sk, &timeo);
1038                 if (err != 0)
1039                         goto do_error;
1040
1041                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1042         }
1043
1044 out:
1045         if (copied) {
1046                 tcp_tx_timestamp(sk, sk->sk_tsflags, tcp_write_queue_tail(sk));
1047                 if (!(flags & MSG_SENDPAGE_NOTLAST))
1048                         tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1049         }
1050         return copied;
1051
1052 do_error:
1053         tcp_remove_empty_skb(sk, tcp_write_queue_tail(sk));
1054         if (copied)
1055                 goto out;
1056 out_err:
1057         /* make sure we wake any epoll edge trigger waiter */
1058         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1059                      err == -EAGAIN)) {
1060                 sk->sk_write_space(sk);
1061                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1062         }
1063         return sk_stream_error(sk, flags, err);
1064 }
1065 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1066
1067 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1068                         size_t size, int flags)
1069 {
1070         if (!(sk->sk_route_caps & NETIF_F_SG) ||
1071             !sk_check_csum_caps(sk))
1072                 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1073
1074         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1075
1076         return do_tcp_sendpages(sk, page, offset, size, flags);
1077 }
1078 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1079
1080 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1081                  size_t size, int flags)
1082 {
1083         int ret;
1084
1085         lock_sock(sk);
1086         ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1087         release_sock(sk);
1088
1089         return ret;
1090 }
1091 EXPORT_SYMBOL(tcp_sendpage);
1092
1093 /* Do not bother using a page frag for very small frames.
1094  * But use this heuristic only for the first skb in write queue.
1095  *
1096  * Having no payload in skb->head allows better SACK shifting
1097  * in tcp_shift_skb_data(), reducing sack/rack overhead, because
1098  * write queue has less skbs.
1099  * Each skb can hold up to MAX_SKB_FRAGS * 32Kbytes, or ~0.5 MB.
1100  * This also speeds up tso_fragment(), since it wont fallback
1101  * to tcp_fragment().
1102  */
1103 static int linear_payload_sz(bool first_skb)
1104 {
1105         if (first_skb)
1106                 return SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1107         return 0;
1108 }
1109
1110 static int select_size(const struct sock *sk, bool sg, bool first_skb)
1111 {
1112         const struct tcp_sock *tp = tcp_sk(sk);
1113         int tmp = tp->mss_cache;
1114
1115         if (sg) {
1116                 if (sk_can_gso(sk)) {
1117                         tmp = linear_payload_sz(first_skb);
1118                 } else {
1119                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1120
1121                         if (tmp >= pgbreak &&
1122                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1123                                 tmp = pgbreak;
1124                 }
1125         }
1126
1127         return tmp;
1128 }
1129
1130 void tcp_free_fastopen_req(struct tcp_sock *tp)
1131 {
1132         if (tp->fastopen_req) {
1133                 kfree(tp->fastopen_req);
1134                 tp->fastopen_req = NULL;
1135         }
1136 }
1137
1138 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1139                                 int *copied, size_t size)
1140 {
1141         struct tcp_sock *tp = tcp_sk(sk);
1142         struct inet_sock *inet = inet_sk(sk);
1143         struct sockaddr *uaddr = msg->msg_name;
1144         int err, flags;
1145
1146         if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1147             (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1148              uaddr->sa_family == AF_UNSPEC))
1149                 return -EOPNOTSUPP;
1150         if (tp->fastopen_req)
1151                 return -EALREADY; /* Another Fast Open is in progress */
1152
1153         tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1154                                    sk->sk_allocation);
1155         if (unlikely(!tp->fastopen_req))
1156                 return -ENOBUFS;
1157         tp->fastopen_req->data = msg;
1158         tp->fastopen_req->size = size;
1159
1160         if (inet->defer_connect) {
1161                 err = tcp_connect(sk);
1162                 /* Same failure procedure as in tcp_v4/6_connect */
1163                 if (err) {
1164                         tcp_set_state(sk, TCP_CLOSE);
1165                         inet->inet_dport = 0;
1166                         sk->sk_route_caps = 0;
1167                 }
1168         }
1169         flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1170         err = __inet_stream_connect(sk->sk_socket, uaddr,
1171                                     msg->msg_namelen, flags, 1);
1172         /* fastopen_req could already be freed in __inet_stream_connect
1173          * if the connection times out or gets rst
1174          */
1175         if (tp->fastopen_req) {
1176                 *copied = tp->fastopen_req->copied;
1177                 tcp_free_fastopen_req(tp);
1178                 inet->defer_connect = 0;
1179         }
1180         return err;
1181 }
1182
1183 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1184 {
1185         struct tcp_sock *tp = tcp_sk(sk);
1186         struct ubuf_info *uarg = NULL;
1187         struct sk_buff *skb;
1188         struct sockcm_cookie sockc;
1189         int flags, err, copied = 0;
1190         int mss_now = 0, size_goal, copied_syn = 0;
1191         bool process_backlog = false;
1192         bool sg;
1193         long timeo;
1194
1195         flags = msg->msg_flags;
1196
1197         if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
1198                 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1199                         err = -EINVAL;
1200                         goto out_err;
1201                 }
1202
1203                 skb = tcp_send_head(sk) ? tcp_write_queue_tail(sk) : NULL;
1204                 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
1205                 if (!uarg) {
1206                         err = -ENOBUFS;
1207                         goto out_err;
1208                 }
1209
1210                 if (!(sk_check_csum_caps(sk) && sk->sk_route_caps & NETIF_F_SG))
1211                         uarg->zerocopy = 0;
1212         }
1213
1214         if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1215             !tp->repair) {
1216                 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1217                 if (err == -EINPROGRESS && copied_syn > 0)
1218                         goto out;
1219                 else if (err)
1220                         goto out_err;
1221         }
1222
1223         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1224
1225         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1226
1227         /* Wait for a connection to finish. One exception is TCP Fast Open
1228          * (passive side) where data is allowed to be sent before a connection
1229          * is fully established.
1230          */
1231         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1232             !tcp_passive_fastopen(sk)) {
1233                 err = sk_stream_wait_connect(sk, &timeo);
1234                 if (err != 0)
1235                         goto do_error;
1236         }
1237
1238         if (unlikely(tp->repair)) {
1239                 if (tp->repair_queue == TCP_RECV_QUEUE) {
1240                         copied = tcp_send_rcvq(sk, msg, size);
1241                         goto out_nopush;
1242                 }
1243
1244                 err = -EINVAL;
1245                 if (tp->repair_queue == TCP_NO_QUEUE)
1246                         goto out_err;
1247
1248                 /* 'common' sending to sendq */
1249         }
1250
1251         sockc.tsflags = sk->sk_tsflags;
1252         if (msg->msg_controllen) {
1253                 err = sock_cmsg_send(sk, msg, &sockc);
1254                 if (unlikely(err)) {
1255                         err = -EINVAL;
1256                         goto out_err;
1257                 }
1258         }
1259
1260         /* This should be in poll */
1261         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1262
1263         /* Ok commence sending. */
1264         copied = 0;
1265
1266 restart:
1267         mss_now = tcp_send_mss(sk, &size_goal, flags);
1268
1269         err = -EPIPE;
1270         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1271                 goto do_error;
1272
1273         sg = !!(sk->sk_route_caps & NETIF_F_SG);
1274
1275         while (msg_data_left(msg)) {
1276                 int copy = 0;
1277                 int max = size_goal;
1278
1279                 skb = tcp_write_queue_tail(sk);
1280                 if (tcp_send_head(sk)) {
1281                         if (skb->ip_summed == CHECKSUM_NONE)
1282                                 max = mss_now;
1283                         copy = max - skb->len;
1284                 }
1285
1286                 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1287                         bool first_skb;
1288
1289 new_segment:
1290                         /* Allocate new segment. If the interface is SG,
1291                          * allocate skb fitting to single page.
1292                          */
1293                         if (!sk_stream_memory_free(sk))
1294                                 goto wait_for_sndbuf;
1295
1296                         if (process_backlog && sk_flush_backlog(sk)) {
1297                                 process_backlog = false;
1298                                 goto restart;
1299                         }
1300                         first_skb = skb_queue_empty(&sk->sk_write_queue);
1301                         skb = sk_stream_alloc_skb(sk,
1302                                                   select_size(sk, sg, first_skb),
1303                                                   sk->sk_allocation,
1304                                                   first_skb);
1305                         if (!skb)
1306                                 goto wait_for_memory;
1307
1308                         process_backlog = true;
1309                         /*
1310                          * Check whether we can use HW checksum.
1311                          */
1312                         if (sk_check_csum_caps(sk))
1313                                 skb->ip_summed = CHECKSUM_PARTIAL;
1314
1315                         skb_entail(sk, skb);
1316                         copy = size_goal;
1317                         max = size_goal;
1318
1319                         /* All packets are restored as if they have
1320                          * already been sent. skb_mstamp isn't set to
1321                          * avoid wrong rtt estimation.
1322                          */
1323                         if (tp->repair)
1324                                 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1325                 }
1326
1327                 /* Try to append data to the end of skb. */
1328                 if (copy > msg_data_left(msg))
1329                         copy = msg_data_left(msg);
1330
1331                 /* Where to copy to? */
1332                 if (skb_availroom(skb) > 0) {
1333                         /* We have some space in skb head. Superb! */
1334                         copy = min_t(int, copy, skb_availroom(skb));
1335                         err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1336                         if (err)
1337                                 goto do_fault;
1338                 } else if (!uarg || !uarg->zerocopy) {
1339                         bool merge = true;
1340                         int i = skb_shinfo(skb)->nr_frags;
1341                         struct page_frag *pfrag = sk_page_frag(sk);
1342
1343                         if (!sk_page_frag_refill(sk, pfrag))
1344                                 goto wait_for_memory;
1345
1346                         if (!skb_can_coalesce(skb, i, pfrag->page,
1347                                               pfrag->offset)) {
1348                                 if (i >= sysctl_max_skb_frags || !sg) {
1349                                         tcp_mark_push(tp, skb);
1350                                         goto new_segment;
1351                                 }
1352                                 merge = false;
1353                         }
1354
1355                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1356
1357                         if (!sk_wmem_schedule(sk, copy))
1358                                 goto wait_for_memory;
1359
1360                         err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1361                                                        pfrag->page,
1362                                                        pfrag->offset,
1363                                                        copy);
1364                         if (err)
1365                                 goto do_error;
1366
1367                         /* Update the skb. */
1368                         if (merge) {
1369                                 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1370                         } else {
1371                                 skb_fill_page_desc(skb, i, pfrag->page,
1372                                                    pfrag->offset, copy);
1373                                 page_ref_inc(pfrag->page);
1374                         }
1375                         pfrag->offset += copy;
1376                 } else {
1377                         err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1378                         if (err == -EMSGSIZE || err == -EEXIST)
1379                                 goto new_segment;
1380                         if (err < 0)
1381                                 goto do_error;
1382                         copy = err;
1383                 }
1384
1385                 if (!copied)
1386                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1387
1388                 tp->write_seq += copy;
1389                 TCP_SKB_CB(skb)->end_seq += copy;
1390                 tcp_skb_pcount_set(skb, 0);
1391
1392                 copied += copy;
1393                 if (!msg_data_left(msg)) {
1394                         if (unlikely(flags & MSG_EOR))
1395                                 TCP_SKB_CB(skb)->eor = 1;
1396                         goto out;
1397                 }
1398
1399                 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1400                         continue;
1401
1402                 if (forced_push(tp)) {
1403                         tcp_mark_push(tp, skb);
1404                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1405                 } else if (skb == tcp_send_head(sk))
1406                         tcp_push_one(sk, mss_now);
1407                 continue;
1408
1409 wait_for_sndbuf:
1410                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1411 wait_for_memory:
1412                 if (copied)
1413                         tcp_push(sk, flags & ~MSG_MORE, mss_now,
1414                                  TCP_NAGLE_PUSH, size_goal);
1415
1416                 err = sk_stream_wait_memory(sk, &timeo);
1417                 if (err != 0)
1418                         goto do_error;
1419
1420                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1421         }
1422
1423 out:
1424         if (copied) {
1425                 tcp_tx_timestamp(sk, sockc.tsflags, tcp_write_queue_tail(sk));
1426                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1427         }
1428 out_nopush:
1429         sock_zerocopy_put(uarg);
1430         return copied + copied_syn;
1431
1432 do_error:
1433         skb = tcp_write_queue_tail(sk);
1434 do_fault:
1435         tcp_remove_empty_skb(sk, skb);
1436
1437         if (copied + copied_syn)
1438                 goto out;
1439 out_err:
1440         sock_zerocopy_put_abort(uarg);
1441         err = sk_stream_error(sk, flags, err);
1442         /* make sure we wake any epoll edge trigger waiter */
1443         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1444                      err == -EAGAIN)) {
1445                 sk->sk_write_space(sk);
1446                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1447         }
1448         return err;
1449 }
1450 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1451
1452 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1453 {
1454         int ret;
1455
1456         lock_sock(sk);
1457         ret = tcp_sendmsg_locked(sk, msg, size);
1458         release_sock(sk);
1459
1460         return ret;
1461 }
1462 EXPORT_SYMBOL(tcp_sendmsg);
1463
1464 /*
1465  *      Handle reading urgent data. BSD has very simple semantics for
1466  *      this, no blocking and very strange errors 8)
1467  */
1468
1469 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1470 {
1471         struct tcp_sock *tp = tcp_sk(sk);
1472
1473         /* No URG data to read. */
1474         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1475             tp->urg_data == TCP_URG_READ)
1476                 return -EINVAL; /* Yes this is right ! */
1477
1478         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1479                 return -ENOTCONN;
1480
1481         if (tp->urg_data & TCP_URG_VALID) {
1482                 int err = 0;
1483                 char c = tp->urg_data;
1484
1485                 if (!(flags & MSG_PEEK))
1486                         tp->urg_data = TCP_URG_READ;
1487
1488                 /* Read urgent data. */
1489                 msg->msg_flags |= MSG_OOB;
1490
1491                 if (len > 0) {
1492                         if (!(flags & MSG_TRUNC))
1493                                 err = memcpy_to_msg(msg, &c, 1);
1494                         len = 1;
1495                 } else
1496                         msg->msg_flags |= MSG_TRUNC;
1497
1498                 return err ? -EFAULT : len;
1499         }
1500
1501         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1502                 return 0;
1503
1504         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1505          * the available implementations agree in this case:
1506          * this call should never block, independent of the
1507          * blocking state of the socket.
1508          * Mike <pall@rz.uni-karlsruhe.de>
1509          */
1510         return -EAGAIN;
1511 }
1512
1513 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1514 {
1515         struct sk_buff *skb;
1516         int copied = 0, err = 0;
1517
1518         /* XXX -- need to support SO_PEEK_OFF */
1519
1520         skb_queue_walk(&sk->sk_write_queue, skb) {
1521                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1522                 if (err)
1523                         break;
1524
1525                 copied += skb->len;
1526         }
1527
1528         return err ?: copied;
1529 }
1530
1531 /* Clean up the receive buffer for full frames taken by the user,
1532  * then send an ACK if necessary.  COPIED is the number of bytes
1533  * tcp_recvmsg has given to the user so far, it speeds up the
1534  * calculation of whether or not we must ACK for the sake of
1535  * a window update.
1536  */
1537 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1538 {
1539         struct tcp_sock *tp = tcp_sk(sk);
1540         bool time_to_ack = false;
1541
1542         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1543
1544         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1545              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1546              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1547
1548         if (inet_csk_ack_scheduled(sk)) {
1549                 const struct inet_connection_sock *icsk = inet_csk(sk);
1550                    /* Delayed ACKs frequently hit locked sockets during bulk
1551                     * receive. */
1552                 if (icsk->icsk_ack.blocked ||
1553                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1554                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1555                     /*
1556                      * If this read emptied read buffer, we send ACK, if
1557                      * connection is not bidirectional, user drained
1558                      * receive buffer and there was a small segment
1559                      * in queue.
1560                      */
1561                     (copied > 0 &&
1562                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1563                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1564                        !icsk->icsk_ack.pingpong)) &&
1565                       !atomic_read(&sk->sk_rmem_alloc)))
1566                         time_to_ack = true;
1567         }
1568
1569         /* We send an ACK if we can now advertise a non-zero window
1570          * which has been raised "significantly".
1571          *
1572          * Even if window raised up to infinity, do not send window open ACK
1573          * in states, where we will not receive more. It is useless.
1574          */
1575         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1576                 __u32 rcv_window_now = tcp_receive_window(tp);
1577
1578                 /* Optimize, __tcp_select_window() is not cheap. */
1579                 if (2*rcv_window_now <= tp->window_clamp) {
1580                         __u32 new_window = __tcp_select_window(sk);
1581
1582                         /* Send ACK now, if this read freed lots of space
1583                          * in our buffer. Certainly, new_window is new window.
1584                          * We can advertise it now, if it is not less than current one.
1585                          * "Lots" means "at least twice" here.
1586                          */
1587                         if (new_window && new_window >= 2 * rcv_window_now)
1588                                 time_to_ack = true;
1589                 }
1590         }
1591         if (time_to_ack)
1592                 tcp_send_ack(sk);
1593 }
1594
1595 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1596 {
1597         struct sk_buff *skb;
1598         u32 offset;
1599
1600         while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1601                 offset = seq - TCP_SKB_CB(skb)->seq;
1602                 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1603                         pr_err_once("%s: found a SYN, please report !\n", __func__);
1604                         offset--;
1605                 }
1606                 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1607                         *off = offset;
1608                         return skb;
1609                 }
1610                 /* This looks weird, but this can happen if TCP collapsing
1611                  * splitted a fat GRO packet, while we released socket lock
1612                  * in skb_splice_bits()
1613                  */
1614                 sk_eat_skb(sk, skb);
1615         }
1616         return NULL;
1617 }
1618
1619 /*
1620  * This routine provides an alternative to tcp_recvmsg() for routines
1621  * that would like to handle copying from skbuffs directly in 'sendfile'
1622  * fashion.
1623  * Note:
1624  *      - It is assumed that the socket was locked by the caller.
1625  *      - The routine does not block.
1626  *      - At present, there is no support for reading OOB data
1627  *        or for 'peeking' the socket using this routine
1628  *        (although both would be easy to implement).
1629  */
1630 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1631                   sk_read_actor_t recv_actor)
1632 {
1633         struct sk_buff *skb;
1634         struct tcp_sock *tp = tcp_sk(sk);
1635         u32 seq = tp->copied_seq;
1636         u32 offset;
1637         int copied = 0;
1638
1639         if (sk->sk_state == TCP_LISTEN)
1640                 return -ENOTCONN;
1641         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1642                 if (offset < skb->len) {
1643                         int used;
1644                         size_t len;
1645
1646                         len = skb->len - offset;
1647                         /* Stop reading if we hit a patch of urgent data */
1648                         if (tp->urg_data) {
1649                                 u32 urg_offset = tp->urg_seq - seq;
1650                                 if (urg_offset < len)
1651                                         len = urg_offset;
1652                                 if (!len)
1653                                         break;
1654                         }
1655                         used = recv_actor(desc, skb, offset, len);
1656                         if (used <= 0) {
1657                                 if (!copied)
1658                                         copied = used;
1659                                 break;
1660                         }
1661                         if (WARN_ON_ONCE(used > len))
1662                                 used = len;
1663                         seq += used;
1664                         copied += used;
1665                         offset += used;
1666
1667                         /* If recv_actor drops the lock (e.g. TCP splice
1668                          * receive) the skb pointer might be invalid when
1669                          * getting here: tcp_collapse might have deleted it
1670                          * while aggregating skbs from the socket queue.
1671                          */
1672                         skb = tcp_recv_skb(sk, seq - 1, &offset);
1673                         if (!skb)
1674                                 break;
1675                         /* TCP coalescing might have appended data to the skb.
1676                          * Try to splice more frags
1677                          */
1678                         if (offset + 1 != skb->len)
1679                                 continue;
1680                 }
1681                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1682                         sk_eat_skb(sk, skb);
1683                         ++seq;
1684                         break;
1685                 }
1686                 sk_eat_skb(sk, skb);
1687                 if (!desc->count)
1688                         break;
1689                 tp->copied_seq = seq;
1690         }
1691         tp->copied_seq = seq;
1692
1693         tcp_rcv_space_adjust(sk);
1694
1695         /* Clean up data we have read: This will do ACK frames. */
1696         if (copied > 0) {
1697                 tcp_recv_skb(sk, seq, &offset);
1698                 tcp_cleanup_rbuf(sk, copied);
1699         }
1700         return copied;
1701 }
1702 EXPORT_SYMBOL(tcp_read_sock);
1703
1704 int tcp_peek_len(struct socket *sock)
1705 {
1706         return tcp_inq(sock->sk);
1707 }
1708 EXPORT_SYMBOL(tcp_peek_len);
1709
1710 static void tcp_update_recv_tstamps(struct sk_buff *skb,
1711                                     struct scm_timestamping *tss)
1712 {
1713         if (skb->tstamp)
1714                 tss->ts[0] = ktime_to_timespec(skb->tstamp);
1715         else
1716                 tss->ts[0] = (struct timespec) {0};
1717
1718         if (skb_hwtstamps(skb)->hwtstamp)
1719                 tss->ts[2] = ktime_to_timespec(skb_hwtstamps(skb)->hwtstamp);
1720         else
1721                 tss->ts[2] = (struct timespec) {0};
1722 }
1723
1724 /* Similar to __sock_recv_timestamp, but does not require an skb */
1725 void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
1726                         struct scm_timestamping *tss)
1727 {
1728         struct timeval tv;
1729         bool has_timestamping = false;
1730
1731         if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
1732                 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
1733                         if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
1734                                 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPNS,
1735                                          sizeof(tss->ts[0]), &tss->ts[0]);
1736                         } else {
1737                                 tv.tv_sec = tss->ts[0].tv_sec;
1738                                 tv.tv_usec = tss->ts[0].tv_nsec / 1000;
1739
1740                                 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMP,
1741                                          sizeof(tv), &tv);
1742                         }
1743                 }
1744
1745                 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
1746                         has_timestamping = true;
1747                 else
1748                         tss->ts[0] = (struct timespec) {0};
1749         }
1750
1751         if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
1752                 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
1753                         has_timestamping = true;
1754                 else
1755                         tss->ts[2] = (struct timespec) {0};
1756         }
1757
1758         if (has_timestamping) {
1759                 tss->ts[1] = (struct timespec) {0};
1760                 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPING,
1761                          sizeof(*tss), tss);
1762         }
1763 }
1764
1765 /*
1766  *      This routine copies from a sock struct into the user buffer.
1767  *
1768  *      Technical note: in 2.3 we work on _locked_ socket, so that
1769  *      tricks with *seq access order and skb->users are not required.
1770  *      Probably, code can be easily improved even more.
1771  */
1772
1773 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1774                 int flags, int *addr_len)
1775 {
1776         struct tcp_sock *tp = tcp_sk(sk);
1777         int copied = 0;
1778         u32 peek_seq;
1779         u32 *seq;
1780         unsigned long used;
1781         int err;
1782         int target;             /* Read at least this many bytes */
1783         long timeo;
1784         struct sk_buff *skb, *last;
1785         u32 urg_hole = 0;
1786         struct scm_timestamping tss;
1787         bool has_tss = false;
1788
1789         if (unlikely(flags & MSG_ERRQUEUE))
1790                 return inet_recv_error(sk, msg, len, addr_len);
1791
1792         if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue) &&
1793             (sk->sk_state == TCP_ESTABLISHED))
1794                 sk_busy_loop(sk, nonblock);
1795
1796         lock_sock(sk);
1797
1798         err = -ENOTCONN;
1799         if (sk->sk_state == TCP_LISTEN)
1800                 goto out;
1801
1802         timeo = sock_rcvtimeo(sk, nonblock);
1803
1804         /* Urgent data needs to be handled specially. */
1805         if (flags & MSG_OOB)
1806                 goto recv_urg;
1807
1808         if (unlikely(tp->repair)) {
1809                 err = -EPERM;
1810                 if (!(flags & MSG_PEEK))
1811                         goto out;
1812
1813                 if (tp->repair_queue == TCP_SEND_QUEUE)
1814                         goto recv_sndq;
1815
1816                 err = -EINVAL;
1817                 if (tp->repair_queue == TCP_NO_QUEUE)
1818                         goto out;
1819
1820                 /* 'common' recv queue MSG_PEEK-ing */
1821         }
1822
1823         seq = &tp->copied_seq;
1824         if (flags & MSG_PEEK) {
1825                 peek_seq = tp->copied_seq;
1826                 seq = &peek_seq;
1827         }
1828
1829         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1830
1831         do {
1832                 u32 offset;
1833
1834                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1835                 if (tp->urg_data && tp->urg_seq == *seq) {
1836                         if (copied)
1837                                 break;
1838                         if (signal_pending(current)) {
1839                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1840                                 break;
1841                         }
1842                 }
1843
1844                 /* Next get a buffer. */
1845
1846                 last = skb_peek_tail(&sk->sk_receive_queue);
1847                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1848                         last = skb;
1849                         /* Now that we have two receive queues this
1850                          * shouldn't happen.
1851                          */
1852                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1853                                  "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
1854                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1855                                  flags))
1856                                 break;
1857
1858                         offset = *seq - TCP_SKB_CB(skb)->seq;
1859                         if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1860                                 pr_err_once("%s: found a SYN, please report !\n", __func__);
1861                                 offset--;
1862                         }
1863                         if (offset < skb->len)
1864                                 goto found_ok_skb;
1865                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1866                                 goto found_fin_ok;
1867                         WARN(!(flags & MSG_PEEK),
1868                              "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
1869                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1870                 }
1871
1872                 /* Well, if we have backlog, try to process it now yet. */
1873
1874                 if (copied >= target && !sk->sk_backlog.tail)
1875                         break;
1876
1877                 if (copied) {
1878                         if (sk->sk_err ||
1879                             sk->sk_state == TCP_CLOSE ||
1880                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1881                             !timeo ||
1882                             signal_pending(current))
1883                                 break;
1884                 } else {
1885                         if (sock_flag(sk, SOCK_DONE))
1886                                 break;
1887
1888                         if (sk->sk_err) {
1889                                 copied = sock_error(sk);
1890                                 break;
1891                         }
1892
1893                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1894                                 break;
1895
1896                         if (sk->sk_state == TCP_CLOSE) {
1897                                 if (!sock_flag(sk, SOCK_DONE)) {
1898                                         /* This occurs when user tries to read
1899                                          * from never connected socket.
1900                                          */
1901                                         copied = -ENOTCONN;
1902                                         break;
1903                                 }
1904                                 break;
1905                         }
1906
1907                         if (!timeo) {
1908                                 copied = -EAGAIN;
1909                                 break;
1910                         }
1911
1912                         if (signal_pending(current)) {
1913                                 copied = sock_intr_errno(timeo);
1914                                 break;
1915                         }
1916                 }
1917
1918                 tcp_cleanup_rbuf(sk, copied);
1919
1920                 if (copied >= target) {
1921                         /* Do not sleep, just process backlog. */
1922                         release_sock(sk);
1923                         lock_sock(sk);
1924                 } else {
1925                         sk_wait_data(sk, &timeo, last);
1926                 }
1927
1928                 if ((flags & MSG_PEEK) &&
1929                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1930                         net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1931                                             current->comm,
1932                                             task_pid_nr(current));
1933                         peek_seq = tp->copied_seq;
1934                 }
1935                 continue;
1936
1937         found_ok_skb:
1938                 /* Ok so how much can we use? */
1939                 used = skb->len - offset;
1940                 if (len < used)
1941                         used = len;
1942
1943                 /* Do we have urgent data here? */
1944                 if (tp->urg_data) {
1945                         u32 urg_offset = tp->urg_seq - *seq;
1946                         if (urg_offset < used) {
1947                                 if (!urg_offset) {
1948                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1949                                                 ++*seq;
1950                                                 urg_hole++;
1951                                                 offset++;
1952                                                 used--;
1953                                                 if (!used)
1954                                                         goto skip_copy;
1955                                         }
1956                                 } else
1957                                         used = urg_offset;
1958                         }
1959                 }
1960
1961                 if (!(flags & MSG_TRUNC)) {
1962                         err = skb_copy_datagram_msg(skb, offset, msg, used);
1963                         if (err) {
1964                                 /* Exception. Bailout! */
1965                                 if (!copied)
1966                                         copied = -EFAULT;
1967                                 break;
1968                         }
1969                 }
1970
1971                 *seq += used;
1972                 copied += used;
1973                 len -= used;
1974
1975                 tcp_rcv_space_adjust(sk);
1976
1977 skip_copy:
1978                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1979                         tp->urg_data = 0;
1980                         tcp_fast_path_check(sk);
1981                 }
1982
1983                 if (TCP_SKB_CB(skb)->has_rxtstamp) {
1984                         tcp_update_recv_tstamps(skb, &tss);
1985                         has_tss = true;
1986                 }
1987
1988                 if (used + offset < skb->len)
1989                         continue;
1990
1991                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1992                         goto found_fin_ok;
1993                 if (!(flags & MSG_PEEK))
1994                         sk_eat_skb(sk, skb);
1995                 continue;
1996
1997         found_fin_ok:
1998                 /* Process the FIN. */
1999                 ++*seq;
2000                 if (!(flags & MSG_PEEK))
2001                         sk_eat_skb(sk, skb);
2002                 break;
2003         } while (len > 0);
2004
2005         /* According to UNIX98, msg_name/msg_namelen are ignored
2006          * on connected socket. I was just happy when found this 8) --ANK
2007          */
2008
2009         if (has_tss)
2010                 tcp_recv_timestamp(msg, sk, &tss);
2011
2012         /* Clean up data we have read: This will do ACK frames. */
2013         tcp_cleanup_rbuf(sk, copied);
2014
2015         release_sock(sk);
2016         return copied;
2017
2018 out:
2019         release_sock(sk);
2020         return err;
2021
2022 recv_urg:
2023         err = tcp_recv_urg(sk, msg, len, flags);
2024         goto out;
2025
2026 recv_sndq:
2027         err = tcp_peek_sndq(sk, msg, len);
2028         goto out;
2029 }
2030 EXPORT_SYMBOL(tcp_recvmsg);
2031
2032 void tcp_set_state(struct sock *sk, int state)
2033 {
2034         int oldstate = sk->sk_state;
2035
2036         switch (state) {
2037         case TCP_ESTABLISHED:
2038                 if (oldstate != TCP_ESTABLISHED)
2039                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2040                 break;
2041
2042         case TCP_CLOSE:
2043                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2044                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2045
2046                 sk->sk_prot->unhash(sk);
2047                 if (inet_csk(sk)->icsk_bind_hash &&
2048                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2049                         inet_put_port(sk);
2050                 /* fall through */
2051         default:
2052                 if (oldstate == TCP_ESTABLISHED)
2053                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2054         }
2055
2056         /* Change state AFTER socket is unhashed to avoid closed
2057          * socket sitting in hash tables.
2058          */
2059         sk_state_store(sk, state);
2060
2061 #ifdef STATE_TRACE
2062         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
2063 #endif
2064 }
2065 EXPORT_SYMBOL_GPL(tcp_set_state);
2066
2067 /*
2068  *      State processing on a close. This implements the state shift for
2069  *      sending our FIN frame. Note that we only send a FIN for some
2070  *      states. A shutdown() may have already sent the FIN, or we may be
2071  *      closed.
2072  */
2073
2074 static const unsigned char new_state[16] = {
2075   /* current state:        new state:      action:      */
2076   [0 /* (Invalid) */]   = TCP_CLOSE,
2077   [TCP_ESTABLISHED]     = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2078   [TCP_SYN_SENT]        = TCP_CLOSE,
2079   [TCP_SYN_RECV]        = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2080   [TCP_FIN_WAIT1]       = TCP_FIN_WAIT1,
2081   [TCP_FIN_WAIT2]       = TCP_FIN_WAIT2,
2082   [TCP_TIME_WAIT]       = TCP_CLOSE,
2083   [TCP_CLOSE]           = TCP_CLOSE,
2084   [TCP_CLOSE_WAIT]      = TCP_LAST_ACK  | TCP_ACTION_FIN,
2085   [TCP_LAST_ACK]        = TCP_LAST_ACK,
2086   [TCP_LISTEN]          = TCP_CLOSE,
2087   [TCP_CLOSING]         = TCP_CLOSING,
2088   [TCP_NEW_SYN_RECV]    = TCP_CLOSE,    /* should not happen ! */
2089 };
2090
2091 static int tcp_close_state(struct sock *sk)
2092 {
2093         int next = (int)new_state[sk->sk_state];
2094         int ns = next & TCP_STATE_MASK;
2095
2096         tcp_set_state(sk, ns);
2097
2098         return next & TCP_ACTION_FIN;
2099 }
2100
2101 /*
2102  *      Shutdown the sending side of a connection. Much like close except
2103  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2104  */
2105
2106 void tcp_shutdown(struct sock *sk, int how)
2107 {
2108         /*      We need to grab some memory, and put together a FIN,
2109          *      and then put it into the queue to be sent.
2110          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2111          */
2112         if (!(how & SEND_SHUTDOWN))
2113                 return;
2114
2115         /* If we've already sent a FIN, or it's a closed state, skip this. */
2116         if ((1 << sk->sk_state) &
2117             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2118              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2119                 /* Clear out any half completed packets.  FIN if needed. */
2120                 if (tcp_close_state(sk))
2121                         tcp_send_fin(sk);
2122         }
2123 }
2124 EXPORT_SYMBOL(tcp_shutdown);
2125
2126 bool tcp_check_oom(struct sock *sk, int shift)
2127 {
2128         bool too_many_orphans, out_of_socket_memory;
2129
2130         too_many_orphans = tcp_too_many_orphans(sk, shift);
2131         out_of_socket_memory = tcp_out_of_memory(sk);
2132
2133         if (too_many_orphans)
2134                 net_info_ratelimited("too many orphaned sockets\n");
2135         if (out_of_socket_memory)
2136                 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2137         return too_many_orphans || out_of_socket_memory;
2138 }
2139
2140 void tcp_close(struct sock *sk, long timeout)
2141 {
2142         struct sk_buff *skb;
2143         int data_was_unread = 0;
2144         int state;
2145
2146         lock_sock(sk);
2147         sk->sk_shutdown = SHUTDOWN_MASK;
2148
2149         if (sk->sk_state == TCP_LISTEN) {
2150                 tcp_set_state(sk, TCP_CLOSE);
2151
2152                 /* Special case. */
2153                 inet_csk_listen_stop(sk);
2154
2155                 goto adjudge_to_death;
2156         }
2157
2158         /*  We need to flush the recv. buffs.  We do this only on the
2159          *  descriptor close, not protocol-sourced closes, because the
2160          *  reader process may not have drained the data yet!
2161          */
2162         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2163                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2164
2165                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2166                         len--;
2167                 data_was_unread += len;
2168                 __kfree_skb(skb);
2169         }
2170
2171         sk_mem_reclaim(sk);
2172
2173         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2174         if (sk->sk_state == TCP_CLOSE)
2175                 goto adjudge_to_death;
2176
2177         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2178          * data was lost. To witness the awful effects of the old behavior of
2179          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2180          * GET in an FTP client, suspend the process, wait for the client to
2181          * advertise a zero window, then kill -9 the FTP client, wheee...
2182          * Note: timeout is always zero in such a case.
2183          */
2184         if (unlikely(tcp_sk(sk)->repair)) {
2185                 sk->sk_prot->disconnect(sk, 0);
2186         } else if (data_was_unread) {
2187                 /* Unread data was tossed, zap the connection. */
2188                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2189                 tcp_set_state(sk, TCP_CLOSE);
2190                 tcp_send_active_reset(sk, sk->sk_allocation);
2191         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2192                 /* Check zero linger _after_ checking for unread data. */
2193                 sk->sk_prot->disconnect(sk, 0);
2194                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2195         } else if (tcp_close_state(sk)) {
2196                 /* We FIN if the application ate all the data before
2197                  * zapping the connection.
2198                  */
2199
2200                 /* RED-PEN. Formally speaking, we have broken TCP state
2201                  * machine. State transitions:
2202                  *
2203                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2204                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2205                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2206                  *
2207                  * are legal only when FIN has been sent (i.e. in window),
2208                  * rather than queued out of window. Purists blame.
2209                  *
2210                  * F.e. "RFC state" is ESTABLISHED,
2211                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2212                  *
2213                  * The visible declinations are that sometimes
2214                  * we enter time-wait state, when it is not required really
2215                  * (harmless), do not send active resets, when they are
2216                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2217                  * they look as CLOSING or LAST_ACK for Linux)
2218                  * Probably, I missed some more holelets.
2219                  *                                              --ANK
2220                  * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2221                  * in a single packet! (May consider it later but will
2222                  * probably need API support or TCP_CORK SYN-ACK until
2223                  * data is written and socket is closed.)
2224                  */
2225                 tcp_send_fin(sk);
2226         }
2227
2228         sk_stream_wait_close(sk, timeout);
2229
2230 adjudge_to_death:
2231         state = sk->sk_state;
2232         sock_hold(sk);
2233         sock_orphan(sk);
2234
2235         local_bh_disable();
2236         bh_lock_sock(sk);
2237         /* remove backlog if any, without releasing ownership. */
2238         __release_sock(sk);
2239
2240         percpu_counter_inc(sk->sk_prot->orphan_count);
2241
2242         /* Have we already been destroyed by a softirq or backlog? */
2243         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2244                 goto out;
2245
2246         /*      This is a (useful) BSD violating of the RFC. There is a
2247          *      problem with TCP as specified in that the other end could
2248          *      keep a socket open forever with no application left this end.
2249          *      We use a 1 minute timeout (about the same as BSD) then kill
2250          *      our end. If they send after that then tough - BUT: long enough
2251          *      that we won't make the old 4*rto = almost no time - whoops
2252          *      reset mistake.
2253          *
2254          *      Nope, it was not mistake. It is really desired behaviour
2255          *      f.e. on http servers, when such sockets are useless, but
2256          *      consume significant resources. Let's do it with special
2257          *      linger2 option.                                 --ANK
2258          */
2259
2260         if (sk->sk_state == TCP_FIN_WAIT2) {
2261                 struct tcp_sock *tp = tcp_sk(sk);
2262                 if (tp->linger2 < 0) {
2263                         tcp_set_state(sk, TCP_CLOSE);
2264                         tcp_send_active_reset(sk, GFP_ATOMIC);
2265                         __NET_INC_STATS(sock_net(sk),
2266                                         LINUX_MIB_TCPABORTONLINGER);
2267                 } else {
2268                         const int tmo = tcp_fin_time(sk);
2269
2270                         if (tmo > TCP_TIMEWAIT_LEN) {
2271                                 inet_csk_reset_keepalive_timer(sk,
2272                                                 tmo - TCP_TIMEWAIT_LEN);
2273                         } else {
2274                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2275                                 goto out;
2276                         }
2277                 }
2278         }
2279         if (sk->sk_state != TCP_CLOSE) {
2280                 sk_mem_reclaim(sk);
2281                 if (tcp_check_oom(sk, 0)) {
2282                         tcp_set_state(sk, TCP_CLOSE);
2283                         tcp_send_active_reset(sk, GFP_ATOMIC);
2284                         __NET_INC_STATS(sock_net(sk),
2285                                         LINUX_MIB_TCPABORTONMEMORY);
2286                 } else if (!check_net(sock_net(sk))) {
2287                         /* Not possible to send reset; just close */
2288                         tcp_set_state(sk, TCP_CLOSE);
2289                 }
2290         }
2291
2292         if (sk->sk_state == TCP_CLOSE) {
2293                 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2294                 /* We could get here with a non-NULL req if the socket is
2295                  * aborted (e.g., closed with unread data) before 3WHS
2296                  * finishes.
2297                  */
2298                 if (req)
2299                         reqsk_fastopen_remove(sk, req, false);
2300                 inet_csk_destroy_sock(sk);
2301         }
2302         /* Otherwise, socket is reprieved until protocol close. */
2303
2304 out:
2305         bh_unlock_sock(sk);
2306         local_bh_enable();
2307         release_sock(sk);
2308         sock_put(sk);
2309 }
2310 EXPORT_SYMBOL(tcp_close);
2311
2312 /* These states need RST on ABORT according to RFC793 */
2313
2314 static inline bool tcp_need_reset(int state)
2315 {
2316         return (1 << state) &
2317                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2318                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2319 }
2320
2321 int tcp_disconnect(struct sock *sk, int flags)
2322 {
2323         struct inet_sock *inet = inet_sk(sk);
2324         struct inet_connection_sock *icsk = inet_csk(sk);
2325         struct tcp_sock *tp = tcp_sk(sk);
2326         int err = 0;
2327         int old_state = sk->sk_state;
2328
2329         if (old_state != TCP_CLOSE)
2330                 tcp_set_state(sk, TCP_CLOSE);
2331
2332         /* ABORT function of RFC793 */
2333         if (old_state == TCP_LISTEN) {
2334                 inet_csk_listen_stop(sk);
2335         } else if (unlikely(tp->repair)) {
2336                 sk->sk_err = ECONNABORTED;
2337         } else if (tcp_need_reset(old_state) ||
2338                    (tp->snd_nxt != tp->write_seq &&
2339                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2340                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2341                  * states
2342                  */
2343                 tcp_send_active_reset(sk, gfp_any());
2344                 sk->sk_err = ECONNRESET;
2345         } else if (old_state == TCP_SYN_SENT)
2346                 sk->sk_err = ECONNRESET;
2347
2348         tcp_clear_xmit_timers(sk);
2349         __skb_queue_purge(&sk->sk_receive_queue);
2350         tcp_write_queue_purge(sk);
2351         tcp_fastopen_active_disable_ofo_check(sk);
2352         skb_rbtree_purge(&tp->out_of_order_queue);
2353
2354         inet->inet_dport = 0;
2355
2356         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2357                 inet_reset_saddr(sk);
2358
2359         sk->sk_shutdown = 0;
2360         sock_reset_flag(sk, SOCK_DONE);
2361         tp->srtt_us = 0;
2362         tp->write_seq += tp->max_window + 2;
2363         if (tp->write_seq == 0)
2364                 tp->write_seq = 1;
2365         tp->snd_cwnd = 2;
2366         icsk->icsk_probes_out = 0;
2367         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2368         tp->snd_cwnd_cnt = 0;
2369         tp->window_clamp = 0;
2370         tp->delivered = 0;
2371         if (icsk->icsk_ca_ops->release)
2372                 icsk->icsk_ca_ops->release(sk);
2373         memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
2374         tcp_set_ca_state(sk, TCP_CA_Open);
2375         tp->is_sack_reneg = 0;
2376         tcp_clear_retrans(tp);
2377         tp->total_retrans = 0;
2378         inet_csk_delack_init(sk);
2379         /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2380          * issue in __tcp_select_window()
2381          */
2382         icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2383         tcp_init_send_head(sk);
2384         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2385         __sk_dst_reset(sk);
2386         dst_release(sk->sk_rx_dst);
2387         sk->sk_rx_dst = NULL;
2388         tcp_saved_syn_free(tp);
2389         tp->segs_in = 0;
2390         tp->segs_out = 0;
2391         tp->bytes_acked = 0;
2392         tp->bytes_received = 0;
2393         tp->data_segs_in = 0;
2394         tp->data_segs_out = 0;
2395
2396         /* Clean up fastopen related fields */
2397         tcp_free_fastopen_req(tp);
2398         inet->defer_connect = 0;
2399
2400         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2401
2402         if (sk->sk_frag.page) {
2403                 put_page(sk->sk_frag.page);
2404                 sk->sk_frag.page = NULL;
2405                 sk->sk_frag.offset = 0;
2406         }
2407
2408         sk->sk_error_report(sk);
2409         return err;
2410 }
2411 EXPORT_SYMBOL(tcp_disconnect);
2412
2413 static inline bool tcp_can_repair_sock(const struct sock *sk)
2414 {
2415         return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2416                 (sk->sk_state != TCP_LISTEN);
2417 }
2418
2419 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2420 {
2421         struct tcp_repair_window opt;
2422
2423         if (!tp->repair)
2424                 return -EPERM;
2425
2426         if (len != sizeof(opt))
2427                 return -EINVAL;
2428
2429         if (copy_from_user(&opt, optbuf, sizeof(opt)))
2430                 return -EFAULT;
2431
2432         if (opt.max_window < opt.snd_wnd)
2433                 return -EINVAL;
2434
2435         if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2436                 return -EINVAL;
2437
2438         if (after(opt.rcv_wup, tp->rcv_nxt))
2439                 return -EINVAL;
2440
2441         tp->snd_wl1     = opt.snd_wl1;
2442         tp->snd_wnd     = opt.snd_wnd;
2443         tp->max_window  = opt.max_window;
2444
2445         tp->rcv_wnd     = opt.rcv_wnd;
2446         tp->rcv_wup     = opt.rcv_wup;
2447
2448         return 0;
2449 }
2450
2451 static int tcp_repair_options_est(struct sock *sk,
2452                 struct tcp_repair_opt __user *optbuf, unsigned int len)
2453 {
2454         struct tcp_sock *tp = tcp_sk(sk);
2455         struct tcp_repair_opt opt;
2456
2457         while (len >= sizeof(opt)) {
2458                 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2459                         return -EFAULT;
2460
2461                 optbuf++;
2462                 len -= sizeof(opt);
2463
2464                 switch (opt.opt_code) {
2465                 case TCPOPT_MSS:
2466                         tp->rx_opt.mss_clamp = opt.opt_val;
2467                         tcp_mtup_init(sk);
2468                         break;
2469                 case TCPOPT_WINDOW:
2470                         {
2471                                 u16 snd_wscale = opt.opt_val & 0xFFFF;
2472                                 u16 rcv_wscale = opt.opt_val >> 16;
2473
2474                                 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
2475                                         return -EFBIG;
2476
2477                                 tp->rx_opt.snd_wscale = snd_wscale;
2478                                 tp->rx_opt.rcv_wscale = rcv_wscale;
2479                                 tp->rx_opt.wscale_ok = 1;
2480                         }
2481                         break;
2482                 case TCPOPT_SACK_PERM:
2483                         if (opt.opt_val != 0)
2484                                 return -EINVAL;
2485
2486                         tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2487                         if (sysctl_tcp_fack)
2488                                 tcp_enable_fack(tp);
2489                         break;
2490                 case TCPOPT_TIMESTAMP:
2491                         if (opt.opt_val != 0)
2492                                 return -EINVAL;
2493
2494                         tp->rx_opt.tstamp_ok = 1;
2495                         break;
2496                 }
2497         }
2498
2499         return 0;
2500 }
2501
2502 /*
2503  *      Socket option code for TCP.
2504  */
2505 static int do_tcp_setsockopt(struct sock *sk, int level,
2506                 int optname, char __user *optval, unsigned int optlen)
2507 {
2508         struct tcp_sock *tp = tcp_sk(sk);
2509         struct inet_connection_sock *icsk = inet_csk(sk);
2510         struct net *net = sock_net(sk);
2511         int val;
2512         int err = 0;
2513
2514         /* These are data/string values, all the others are ints */
2515         switch (optname) {
2516         case TCP_CONGESTION: {
2517                 char name[TCP_CA_NAME_MAX];
2518
2519                 if (optlen < 1)
2520                         return -EINVAL;
2521
2522                 val = strncpy_from_user(name, optval,
2523                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2524                 if (val < 0)
2525                         return -EFAULT;
2526                 name[val] = 0;
2527
2528                 lock_sock(sk);
2529                 err = tcp_set_congestion_control(sk, name, true, true,
2530                                                  ns_capable(sock_net(sk)->user_ns,
2531                                                             CAP_NET_ADMIN));
2532                 release_sock(sk);
2533                 return err;
2534         }
2535         case TCP_ULP: {
2536                 char name[TCP_ULP_NAME_MAX];
2537
2538                 if (optlen < 1)
2539                         return -EINVAL;
2540
2541                 val = strncpy_from_user(name, optval,
2542                                         min_t(long, TCP_ULP_NAME_MAX - 1,
2543                                               optlen));
2544                 if (val < 0)
2545                         return -EFAULT;
2546                 name[val] = 0;
2547
2548                 lock_sock(sk);
2549                 err = tcp_set_ulp(sk, name);
2550                 release_sock(sk);
2551                 return err;
2552         }
2553         default:
2554                 /* fallthru */
2555                 break;
2556         }
2557
2558         if (optlen < sizeof(int))
2559                 return -EINVAL;
2560
2561         if (get_user(val, (int __user *)optval))
2562                 return -EFAULT;
2563
2564         lock_sock(sk);
2565
2566         switch (optname) {
2567         case TCP_MAXSEG:
2568                 /* Values greater than interface MTU won't take effect. However
2569                  * at the point when this call is done we typically don't yet
2570                  * know which interface is going to be used
2571                  */
2572                 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
2573                         err = -EINVAL;
2574                         break;
2575                 }
2576                 tp->rx_opt.user_mss = val;
2577                 break;
2578
2579         case TCP_NODELAY:
2580                 if (val) {
2581                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2582                          * this option on corked socket is remembered, but
2583                          * it is not activated until cork is cleared.
2584                          *
2585                          * However, when TCP_NODELAY is set we make
2586                          * an explicit push, which overrides even TCP_CORK
2587                          * for currently queued segments.
2588                          */
2589                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2590                         tcp_push_pending_frames(sk);
2591                 } else {
2592                         tp->nonagle &= ~TCP_NAGLE_OFF;
2593                 }
2594                 break;
2595
2596         case TCP_THIN_LINEAR_TIMEOUTS:
2597                 if (val < 0 || val > 1)
2598                         err = -EINVAL;
2599                 else
2600                         tp->thin_lto = val;
2601                 break;
2602
2603         case TCP_THIN_DUPACK:
2604                 if (val < 0 || val > 1)
2605                         err = -EINVAL;
2606                 break;
2607
2608         case TCP_REPAIR:
2609                 if (!tcp_can_repair_sock(sk))
2610                         err = -EPERM;
2611                 else if (val == 1) {
2612                         tp->repair = 1;
2613                         sk->sk_reuse = SK_FORCE_REUSE;
2614                         tp->repair_queue = TCP_NO_QUEUE;
2615                 } else if (val == 0) {
2616                         tp->repair = 0;
2617                         sk->sk_reuse = SK_NO_REUSE;
2618                         tcp_send_window_probe(sk);
2619                 } else
2620                         err = -EINVAL;
2621
2622                 break;
2623
2624         case TCP_REPAIR_QUEUE:
2625                 if (!tp->repair)
2626                         err = -EPERM;
2627                 else if ((unsigned int)val < TCP_QUEUES_NR)
2628                         tp->repair_queue = val;
2629                 else
2630                         err = -EINVAL;
2631                 break;
2632
2633         case TCP_QUEUE_SEQ:
2634                 if (sk->sk_state != TCP_CLOSE)
2635                         err = -EPERM;
2636                 else if (tp->repair_queue == TCP_SEND_QUEUE)
2637                         tp->write_seq = val;
2638                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2639                         tp->rcv_nxt = val;
2640                 else
2641                         err = -EINVAL;
2642                 break;
2643
2644         case TCP_REPAIR_OPTIONS:
2645                 if (!tp->repair)
2646                         err = -EINVAL;
2647                 else if (sk->sk_state == TCP_ESTABLISHED)
2648                         err = tcp_repair_options_est(sk,
2649                                         (struct tcp_repair_opt __user *)optval,
2650                                         optlen);
2651                 else
2652                         err = -EPERM;
2653                 break;
2654
2655         case TCP_CORK:
2656                 /* When set indicates to always queue non-full frames.
2657                  * Later the user clears this option and we transmit
2658                  * any pending partial frames in the queue.  This is
2659                  * meant to be used alongside sendfile() to get properly
2660                  * filled frames when the user (for example) must write
2661                  * out headers with a write() call first and then use
2662                  * sendfile to send out the data parts.
2663                  *
2664                  * TCP_CORK can be set together with TCP_NODELAY and it is
2665                  * stronger than TCP_NODELAY.
2666                  */
2667                 if (val) {
2668                         tp->nonagle |= TCP_NAGLE_CORK;
2669                 } else {
2670                         tp->nonagle &= ~TCP_NAGLE_CORK;
2671                         if (tp->nonagle&TCP_NAGLE_OFF)
2672                                 tp->nonagle |= TCP_NAGLE_PUSH;
2673                         tcp_push_pending_frames(sk);
2674                 }
2675                 break;
2676
2677         case TCP_KEEPIDLE:
2678                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2679                         err = -EINVAL;
2680                 else {
2681                         tp->keepalive_time = val * HZ;
2682                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2683                             !((1 << sk->sk_state) &
2684                               (TCPF_CLOSE | TCPF_LISTEN))) {
2685                                 u32 elapsed = keepalive_time_elapsed(tp);
2686                                 if (tp->keepalive_time > elapsed)
2687                                         elapsed = tp->keepalive_time - elapsed;
2688                                 else
2689                                         elapsed = 0;
2690                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2691                         }
2692                 }
2693                 break;
2694         case TCP_KEEPINTVL:
2695                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2696                         err = -EINVAL;
2697                 else
2698                         tp->keepalive_intvl = val * HZ;
2699                 break;
2700         case TCP_KEEPCNT:
2701                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2702                         err = -EINVAL;
2703                 else
2704                         tp->keepalive_probes = val;
2705                 break;
2706         case TCP_SYNCNT:
2707                 if (val < 1 || val > MAX_TCP_SYNCNT)
2708                         err = -EINVAL;
2709                 else
2710                         icsk->icsk_syn_retries = val;
2711                 break;
2712
2713         case TCP_SAVE_SYN:
2714                 if (val < 0 || val > 1)
2715                         err = -EINVAL;
2716                 else
2717                         tp->save_syn = val;
2718                 break;
2719
2720         case TCP_LINGER2:
2721                 if (val < 0)
2722                         tp->linger2 = -1;
2723                 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
2724                         tp->linger2 = 0;
2725                 else
2726                         tp->linger2 = val * HZ;
2727                 break;
2728
2729         case TCP_DEFER_ACCEPT:
2730                 /* Translate value in seconds to number of retransmits */
2731                 icsk->icsk_accept_queue.rskq_defer_accept =
2732                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2733                                         TCP_RTO_MAX / HZ);
2734                 break;
2735
2736         case TCP_WINDOW_CLAMP:
2737                 if (!val) {
2738                         if (sk->sk_state != TCP_CLOSE) {
2739                                 err = -EINVAL;
2740                                 break;
2741                         }
2742                         tp->window_clamp = 0;
2743                 } else
2744                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2745                                                 SOCK_MIN_RCVBUF / 2 : val;
2746                 break;
2747
2748         case TCP_QUICKACK:
2749                 if (!val) {
2750                         icsk->icsk_ack.pingpong = 1;
2751                 } else {
2752                         icsk->icsk_ack.pingpong = 0;
2753                         if ((1 << sk->sk_state) &
2754                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2755                             inet_csk_ack_scheduled(sk)) {
2756                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2757                                 tcp_cleanup_rbuf(sk, 1);
2758                                 if (!(val & 1))
2759                                         icsk->icsk_ack.pingpong = 1;
2760                         }
2761                 }
2762                 break;
2763
2764 #ifdef CONFIG_TCP_MD5SIG
2765         case TCP_MD5SIG:
2766         case TCP_MD5SIG_EXT:
2767                 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
2768                 break;
2769 #endif
2770         case TCP_USER_TIMEOUT:
2771                 /* Cap the max time in ms TCP will retry or probe the window
2772                  * before giving up and aborting (ETIMEDOUT) a connection.
2773                  */
2774                 if (val < 0)
2775                         err = -EINVAL;
2776                 else
2777                         icsk->icsk_user_timeout = msecs_to_jiffies(val);
2778                 break;
2779
2780         case TCP_FASTOPEN:
2781                 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2782                     TCPF_LISTEN))) {
2783                         tcp_fastopen_init_key_once(true);
2784
2785                         fastopen_queue_tune(sk, val);
2786                 } else {
2787                         err = -EINVAL;
2788                 }
2789                 break;
2790         case TCP_FASTOPEN_CONNECT:
2791                 if (val > 1 || val < 0) {
2792                         err = -EINVAL;
2793                 } else if (sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
2794                         if (sk->sk_state == TCP_CLOSE)
2795                                 tp->fastopen_connect = val;
2796                         else
2797                                 err = -EINVAL;
2798                 } else {
2799                         err = -EOPNOTSUPP;
2800                 }
2801                 break;
2802         case TCP_TIMESTAMP:
2803                 if (!tp->repair)
2804                         err = -EPERM;
2805                 else
2806                         tp->tsoffset = val - tcp_time_stamp_raw();
2807                 break;
2808         case TCP_REPAIR_WINDOW:
2809                 err = tcp_repair_set_window(tp, optval, optlen);
2810                 break;
2811         case TCP_NOTSENT_LOWAT:
2812                 tp->notsent_lowat = val;
2813                 sk->sk_write_space(sk);
2814                 break;
2815         default:
2816                 err = -ENOPROTOOPT;
2817                 break;
2818         }
2819
2820         release_sock(sk);
2821         return err;
2822 }
2823
2824 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2825                    unsigned int optlen)
2826 {
2827         const struct inet_connection_sock *icsk = inet_csk(sk);
2828
2829         if (level != SOL_TCP)
2830                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2831                                                      optval, optlen);
2832         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2833 }
2834 EXPORT_SYMBOL(tcp_setsockopt);
2835
2836 #ifdef CONFIG_COMPAT
2837 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2838                           char __user *optval, unsigned int optlen)
2839 {
2840         if (level != SOL_TCP)
2841                 return inet_csk_compat_setsockopt(sk, level, optname,
2842                                                   optval, optlen);
2843         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2844 }
2845 EXPORT_SYMBOL(compat_tcp_setsockopt);
2846 #endif
2847
2848 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
2849                                       struct tcp_info *info)
2850 {
2851         u64 stats[__TCP_CHRONO_MAX], total = 0;
2852         enum tcp_chrono i;
2853
2854         for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
2855                 stats[i] = tp->chrono_stat[i - 1];
2856                 if (i == tp->chrono_type)
2857                         stats[i] += tcp_jiffies32 - tp->chrono_start;
2858                 stats[i] *= USEC_PER_SEC / HZ;
2859                 total += stats[i];
2860         }
2861
2862         info->tcpi_busy_time = total;
2863         info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
2864         info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
2865 }
2866
2867 /* Return information about state of tcp endpoint in API format. */
2868 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2869 {
2870         const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
2871         const struct inet_connection_sock *icsk = inet_csk(sk);
2872         u32 now;
2873         u64 rate64;
2874         bool slow;
2875         u32 rate;
2876
2877         memset(info, 0, sizeof(*info));
2878         if (sk->sk_type != SOCK_STREAM)
2879                 return;
2880
2881         info->tcpi_state = sk_state_load(sk);
2882
2883         /* Report meaningful fields for all TCP states, including listeners */
2884         rate = READ_ONCE(sk->sk_pacing_rate);
2885         rate64 = rate != ~0U ? rate : ~0ULL;
2886         info->tcpi_pacing_rate = rate64;
2887
2888         rate = READ_ONCE(sk->sk_max_pacing_rate);
2889         rate64 = rate != ~0U ? rate : ~0ULL;
2890         info->tcpi_max_pacing_rate = rate64;
2891
2892         info->tcpi_reordering = tp->reordering;
2893         info->tcpi_snd_cwnd = tp->snd_cwnd;
2894
2895         if (info->tcpi_state == TCP_LISTEN) {
2896                 /* listeners aliased fields :
2897                  * tcpi_unacked -> Number of children ready for accept()
2898                  * tcpi_sacked  -> max backlog
2899                  */
2900                 info->tcpi_unacked = sk->sk_ack_backlog;
2901                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2902                 return;
2903         }
2904
2905         slow = lock_sock_fast(sk);
2906
2907         info->tcpi_ca_state = icsk->icsk_ca_state;
2908         info->tcpi_retransmits = icsk->icsk_retransmits;
2909         info->tcpi_probes = icsk->icsk_probes_out;
2910         info->tcpi_backoff = icsk->icsk_backoff;
2911
2912         if (tp->rx_opt.tstamp_ok)
2913                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2914         if (tcp_is_sack(tp))
2915                 info->tcpi_options |= TCPI_OPT_SACK;
2916         if (tp->rx_opt.wscale_ok) {
2917                 info->tcpi_options |= TCPI_OPT_WSCALE;
2918                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2919                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2920         }
2921
2922         if (tp->ecn_flags & TCP_ECN_OK)
2923                 info->tcpi_options |= TCPI_OPT_ECN;
2924         if (tp->ecn_flags & TCP_ECN_SEEN)
2925                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2926         if (tp->syn_data_acked)
2927                 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2928
2929         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2930         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2931         info->tcpi_snd_mss = tp->mss_cache;
2932         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2933
2934         info->tcpi_unacked = tp->packets_out;
2935         info->tcpi_sacked = tp->sacked_out;
2936
2937         info->tcpi_lost = tp->lost_out;
2938         info->tcpi_retrans = tp->retrans_out;
2939         info->tcpi_fackets = tp->fackets_out;
2940
2941         now = tcp_jiffies32;
2942         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2943         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2944         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2945
2946         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2947         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2948         info->tcpi_rtt = tp->srtt_us >> 3;
2949         info->tcpi_rttvar = tp->mdev_us >> 2;
2950         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2951         info->tcpi_advmss = tp->advmss;
2952
2953         info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
2954         info->tcpi_rcv_space = tp->rcvq_space.space;
2955
2956         info->tcpi_total_retrans = tp->total_retrans;
2957
2958         info->tcpi_bytes_acked = tp->bytes_acked;
2959         info->tcpi_bytes_received = tp->bytes_received;
2960         info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
2961         tcp_get_info_chrono_stats(tp, info);
2962
2963         info->tcpi_segs_out = tp->segs_out;
2964         info->tcpi_segs_in = tp->segs_in;
2965
2966         info->tcpi_min_rtt = tcp_min_rtt(tp);
2967         info->tcpi_data_segs_in = tp->data_segs_in;
2968         info->tcpi_data_segs_out = tp->data_segs_out;
2969
2970         info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
2971         rate64 = tcp_compute_delivery_rate(tp);
2972         if (rate64)
2973                 info->tcpi_delivery_rate = rate64;
2974         unlock_sock_fast(sk, slow);
2975 }
2976 EXPORT_SYMBOL_GPL(tcp_get_info);
2977
2978 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk)
2979 {
2980         const struct tcp_sock *tp = tcp_sk(sk);
2981         struct sk_buff *stats;
2982         struct tcp_info info;
2983         u64 rate64;
2984         u32 rate;
2985
2986         stats = alloc_skb(7 * nla_total_size_64bit(sizeof(u64)) +
2987                           3 * nla_total_size(sizeof(u32)) +
2988                           2 * nla_total_size(sizeof(u8)), GFP_ATOMIC);
2989         if (!stats)
2990                 return NULL;
2991
2992         tcp_get_info_chrono_stats(tp, &info);
2993         nla_put_u64_64bit(stats, TCP_NLA_BUSY,
2994                           info.tcpi_busy_time, TCP_NLA_PAD);
2995         nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
2996                           info.tcpi_rwnd_limited, TCP_NLA_PAD);
2997         nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
2998                           info.tcpi_sndbuf_limited, TCP_NLA_PAD);
2999         nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3000                           tp->data_segs_out, TCP_NLA_PAD);
3001         nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3002                           tp->total_retrans, TCP_NLA_PAD);
3003
3004         rate = READ_ONCE(sk->sk_pacing_rate);
3005         rate64 = rate != ~0U ? rate : ~0ULL;
3006         nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3007
3008         rate64 = tcp_compute_delivery_rate(tp);
3009         nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3010
3011         nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3012         nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3013         nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3014
3015         nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3016         nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3017         return stats;
3018 }
3019
3020 static int do_tcp_getsockopt(struct sock *sk, int level,
3021                 int optname, char __user *optval, int __user *optlen)
3022 {
3023         struct inet_connection_sock *icsk = inet_csk(sk);
3024         struct tcp_sock *tp = tcp_sk(sk);
3025         struct net *net = sock_net(sk);
3026         int val, len;
3027
3028         if (get_user(len, optlen))
3029                 return -EFAULT;
3030
3031         len = min_t(unsigned int, len, sizeof(int));
3032
3033         if (len < 0)
3034                 return -EINVAL;
3035
3036         switch (optname) {
3037         case TCP_MAXSEG:
3038                 val = tp->mss_cache;
3039                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3040                         val = tp->rx_opt.user_mss;
3041                 if (tp->repair)
3042                         val = tp->rx_opt.mss_clamp;
3043                 break;
3044         case TCP_NODELAY:
3045                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3046                 break;
3047         case TCP_CORK:
3048                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3049                 break;
3050         case TCP_KEEPIDLE:
3051                 val = keepalive_time_when(tp) / HZ;
3052                 break;
3053         case TCP_KEEPINTVL:
3054                 val = keepalive_intvl_when(tp) / HZ;
3055                 break;
3056         case TCP_KEEPCNT:
3057                 val = keepalive_probes(tp);
3058                 break;
3059         case TCP_SYNCNT:
3060                 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3061                 break;
3062         case TCP_LINGER2:
3063                 val = tp->linger2;
3064                 if (val >= 0)
3065                         val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3066                 break;
3067         case TCP_DEFER_ACCEPT:
3068                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3069                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3070                 break;
3071         case TCP_WINDOW_CLAMP:
3072                 val = tp->window_clamp;
3073                 break;
3074         case TCP_INFO: {
3075                 struct tcp_info info;
3076
3077                 if (get_user(len, optlen))
3078                         return -EFAULT;
3079
3080                 tcp_get_info(sk, &info);
3081
3082                 len = min_t(unsigned int, len, sizeof(info));
3083                 if (put_user(len, optlen))
3084                         return -EFAULT;
3085                 if (copy_to_user(optval, &info, len))
3086                         return -EFAULT;
3087                 return 0;
3088         }
3089         case TCP_CC_INFO: {
3090                 const struct tcp_congestion_ops *ca_ops;
3091                 union tcp_cc_info info;
3092                 size_t sz = 0;
3093                 int attr;
3094
3095                 if (get_user(len, optlen))
3096                         return -EFAULT;
3097
3098                 ca_ops = icsk->icsk_ca_ops;
3099                 if (ca_ops && ca_ops->get_info)
3100                         sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3101
3102                 len = min_t(unsigned int, len, sz);
3103                 if (put_user(len, optlen))
3104                         return -EFAULT;
3105                 if (copy_to_user(optval, &info, len))
3106                         return -EFAULT;
3107                 return 0;
3108         }
3109         case TCP_QUICKACK:
3110                 val = !icsk->icsk_ack.pingpong;
3111                 break;
3112
3113         case TCP_CONGESTION:
3114                 if (get_user(len, optlen))
3115                         return -EFAULT;
3116                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
3117                 if (put_user(len, optlen))
3118                         return -EFAULT;
3119                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3120                         return -EFAULT;
3121                 return 0;
3122
3123         case TCP_ULP:
3124                 if (get_user(len, optlen))
3125                         return -EFAULT;
3126                 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
3127                 if (!icsk->icsk_ulp_ops) {
3128                         if (put_user(0, optlen))
3129                                 return -EFAULT;
3130                         return 0;
3131                 }
3132                 if (put_user(len, optlen))
3133                         return -EFAULT;
3134                 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
3135                         return -EFAULT;
3136                 return 0;
3137
3138         case TCP_THIN_LINEAR_TIMEOUTS:
3139                 val = tp->thin_lto;
3140                 break;
3141
3142         case TCP_THIN_DUPACK:
3143                 val = 0;
3144                 break;
3145
3146         case TCP_REPAIR:
3147                 val = tp->repair;
3148                 break;
3149
3150         case TCP_REPAIR_QUEUE:
3151                 if (tp->repair)
3152                         val = tp->repair_queue;
3153                 else
3154                         return -EINVAL;
3155                 break;
3156
3157         case TCP_REPAIR_WINDOW: {
3158                 struct tcp_repair_window opt;
3159
3160                 if (get_user(len, optlen))
3161                         return -EFAULT;
3162
3163                 if (len != sizeof(opt))
3164                         return -EINVAL;
3165
3166                 if (!tp->repair)
3167                         return -EPERM;
3168
3169                 opt.snd_wl1     = tp->snd_wl1;
3170                 opt.snd_wnd     = tp->snd_wnd;
3171                 opt.max_window  = tp->max_window;
3172                 opt.rcv_wnd     = tp->rcv_wnd;
3173                 opt.rcv_wup     = tp->rcv_wup;
3174
3175                 if (copy_to_user(optval, &opt, len))
3176                         return -EFAULT;
3177                 return 0;
3178         }
3179         case TCP_QUEUE_SEQ:
3180                 if (tp->repair_queue == TCP_SEND_QUEUE)
3181                         val = tp->write_seq;
3182                 else if (tp->repair_queue == TCP_RECV_QUEUE)
3183                         val = tp->rcv_nxt;
3184                 else
3185                         return -EINVAL;
3186                 break;
3187
3188         case TCP_USER_TIMEOUT:
3189                 val = jiffies_to_msecs(icsk->icsk_user_timeout);
3190                 break;
3191
3192         case TCP_FASTOPEN:
3193                 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3194                 break;
3195
3196         case TCP_FASTOPEN_CONNECT:
3197                 val = tp->fastopen_connect;
3198                 break;
3199
3200         case TCP_TIMESTAMP:
3201                 val = tcp_time_stamp_raw() + tp->tsoffset;
3202                 break;
3203         case TCP_NOTSENT_LOWAT:
3204                 val = tp->notsent_lowat;
3205                 break;
3206         case TCP_SAVE_SYN:
3207                 val = tp->save_syn;
3208                 break;
3209         case TCP_SAVED_SYN: {
3210                 if (get_user(len, optlen))
3211                         return -EFAULT;
3212
3213                 lock_sock(sk);
3214                 if (tp->saved_syn) {
3215                         if (len < tp->saved_syn[0]) {
3216                                 if (put_user(tp->saved_syn[0], optlen)) {
3217                                         release_sock(sk);
3218                                         return -EFAULT;
3219                                 }
3220                                 release_sock(sk);
3221                                 return -EINVAL;
3222                         }
3223                         len = tp->saved_syn[0];
3224                         if (put_user(len, optlen)) {
3225                                 release_sock(sk);
3226                                 return -EFAULT;
3227                         }
3228                         if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3229                                 release_sock(sk);
3230                                 return -EFAULT;
3231                         }
3232                         tcp_saved_syn_free(tp);
3233                         release_sock(sk);
3234                 } else {
3235                         release_sock(sk);
3236                         len = 0;
3237                         if (put_user(len, optlen))
3238                                 return -EFAULT;
3239                 }
3240                 return 0;
3241         }
3242         default:
3243                 return -ENOPROTOOPT;
3244         }
3245
3246         if (put_user(len, optlen))
3247                 return -EFAULT;
3248         if (copy_to_user(optval, &val, len))
3249                 return -EFAULT;
3250         return 0;
3251 }
3252
3253 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3254                    int __user *optlen)
3255 {
3256         struct inet_connection_sock *icsk = inet_csk(sk);
3257
3258         if (level != SOL_TCP)
3259                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3260                                                      optval, optlen);
3261         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3262 }
3263 EXPORT_SYMBOL(tcp_getsockopt);
3264
3265 #ifdef CONFIG_COMPAT
3266 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3267                           char __user *optval, int __user *optlen)
3268 {
3269         if (level != SOL_TCP)
3270                 return inet_csk_compat_getsockopt(sk, level, optname,
3271                                                   optval, optlen);
3272         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3273 }
3274 EXPORT_SYMBOL(compat_tcp_getsockopt);
3275 #endif
3276
3277 #ifdef CONFIG_TCP_MD5SIG
3278 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3279 static DEFINE_MUTEX(tcp_md5sig_mutex);
3280 static bool tcp_md5sig_pool_populated = false;
3281
3282 static void __tcp_alloc_md5sig_pool(void)
3283 {
3284         struct crypto_ahash *hash;
3285         int cpu;
3286
3287         hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3288         if (IS_ERR(hash))
3289                 return;
3290
3291         for_each_possible_cpu(cpu) {
3292                 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3293                 struct ahash_request *req;
3294
3295                 if (!scratch) {
3296                         scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3297                                                sizeof(struct tcphdr),
3298                                                GFP_KERNEL,
3299                                                cpu_to_node(cpu));
3300                         if (!scratch)
3301                                 return;
3302                         per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3303                 }
3304                 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3305                         continue;
3306
3307                 req = ahash_request_alloc(hash, GFP_KERNEL);
3308                 if (!req)
3309                         return;
3310
3311                 ahash_request_set_callback(req, 0, NULL, NULL);
3312
3313                 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3314         }
3315         /* before setting tcp_md5sig_pool_populated, we must commit all writes
3316          * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3317          */
3318         smp_wmb();
3319         tcp_md5sig_pool_populated = true;
3320 }
3321
3322 bool tcp_alloc_md5sig_pool(void)
3323 {
3324         if (unlikely(!tcp_md5sig_pool_populated)) {
3325                 mutex_lock(&tcp_md5sig_mutex);
3326
3327                 if (!tcp_md5sig_pool_populated)
3328                         __tcp_alloc_md5sig_pool();
3329
3330                 mutex_unlock(&tcp_md5sig_mutex);
3331         }
3332         return tcp_md5sig_pool_populated;
3333 }
3334 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3335
3336
3337 /**
3338  *      tcp_get_md5sig_pool - get md5sig_pool for this user
3339  *
3340  *      We use percpu structure, so if we succeed, we exit with preemption
3341  *      and BH disabled, to make sure another thread or softirq handling
3342  *      wont try to get same context.
3343  */
3344 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3345 {
3346         local_bh_disable();
3347
3348         if (tcp_md5sig_pool_populated) {
3349                 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3350                 smp_rmb();
3351                 return this_cpu_ptr(&tcp_md5sig_pool);
3352         }
3353         local_bh_enable();
3354         return NULL;
3355 }
3356 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3357
3358 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3359                           const struct sk_buff *skb, unsigned int header_len)
3360 {
3361         struct scatterlist sg;
3362         const struct tcphdr *tp = tcp_hdr(skb);
3363         struct ahash_request *req = hp->md5_req;
3364         unsigned int i;
3365         const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3366                                            skb_headlen(skb) - header_len : 0;
3367         const struct skb_shared_info *shi = skb_shinfo(skb);
3368         struct sk_buff *frag_iter;
3369
3370         sg_init_table(&sg, 1);
3371
3372         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3373         ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3374         if (crypto_ahash_update(req))
3375                 return 1;
3376
3377         for (i = 0; i < shi->nr_frags; ++i) {
3378                 const struct skb_frag_struct *f = &shi->frags[i];
3379                 unsigned int offset = f->page_offset;
3380                 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3381
3382                 sg_set_page(&sg, page, skb_frag_size(f),
3383                             offset_in_page(offset));
3384                 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3385                 if (crypto_ahash_update(req))
3386                         return 1;
3387         }
3388
3389         skb_walk_frags(skb, frag_iter)
3390                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3391                         return 1;
3392
3393         return 0;
3394 }
3395 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3396
3397 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3398 {
3399         u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
3400         struct scatterlist sg;
3401
3402         sg_init_one(&sg, key->key, keylen);
3403         ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
3404
3405         /* tcp_md5_do_add() might change key->key under us */
3406         return crypto_ahash_update(hp->md5_req);
3407 }
3408 EXPORT_SYMBOL(tcp_md5_hash_key);
3409
3410 #endif
3411
3412 void tcp_done(struct sock *sk)
3413 {
3414         struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3415
3416         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3417                 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3418
3419         tcp_set_state(sk, TCP_CLOSE);
3420         tcp_clear_xmit_timers(sk);
3421         if (req)
3422                 reqsk_fastopen_remove(sk, req, false);
3423
3424         sk->sk_shutdown = SHUTDOWN_MASK;
3425
3426         if (!sock_flag(sk, SOCK_DEAD))
3427                 sk->sk_state_change(sk);
3428         else
3429                 inet_csk_destroy_sock(sk);
3430 }
3431 EXPORT_SYMBOL_GPL(tcp_done);
3432
3433 int tcp_abort(struct sock *sk, int err)
3434 {
3435         if (!sk_fullsock(sk)) {
3436                 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3437                         struct request_sock *req = inet_reqsk(sk);
3438
3439                         local_bh_disable();
3440                         inet_csk_reqsk_queue_drop(req->rsk_listener, req);
3441                         local_bh_enable();
3442                         return 0;
3443                 }
3444                 return -EOPNOTSUPP;
3445         }
3446
3447         /* Don't race with userspace socket closes such as tcp_close. */
3448         lock_sock(sk);
3449
3450         if (sk->sk_state == TCP_LISTEN) {
3451                 tcp_set_state(sk, TCP_CLOSE);
3452                 inet_csk_listen_stop(sk);
3453         }
3454
3455         /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3456         local_bh_disable();
3457         bh_lock_sock(sk);
3458
3459         if (!sock_flag(sk, SOCK_DEAD)) {
3460                 sk->sk_err = err;
3461                 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3462                 smp_wmb();
3463                 sk->sk_error_report(sk);
3464                 if (tcp_need_reset(sk->sk_state))
3465                         tcp_send_active_reset(sk, GFP_ATOMIC);
3466                 tcp_done(sk);
3467         }
3468
3469         bh_unlock_sock(sk);
3470         local_bh_enable();
3471         tcp_write_queue_purge(sk);
3472         release_sock(sk);
3473         return 0;
3474 }
3475 EXPORT_SYMBOL_GPL(tcp_abort);
3476
3477 extern struct tcp_congestion_ops tcp_reno;
3478
3479 static __initdata unsigned long thash_entries;
3480 static int __init set_thash_entries(char *str)
3481 {
3482         ssize_t ret;
3483
3484         if (!str)
3485                 return 0;
3486
3487         ret = kstrtoul(str, 0, &thash_entries);
3488         if (ret)
3489                 return 0;
3490
3491         return 1;
3492 }
3493 __setup("thash_entries=", set_thash_entries);
3494
3495 static void __init tcp_init_mem(void)
3496 {
3497         unsigned long limit = nr_free_buffer_pages() / 16;
3498
3499         limit = max(limit, 128UL);
3500         sysctl_tcp_mem[0] = limit / 4 * 3;              /* 4.68 % */
3501         sysctl_tcp_mem[1] = limit;                      /* 6.25 % */
3502         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;      /* 9.37 % */
3503 }
3504
3505 void __init tcp_init(void)
3506 {
3507         int max_rshare, max_wshare, cnt;
3508         unsigned long limit;
3509         unsigned int i;
3510
3511         BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
3512         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3513                      FIELD_SIZEOF(struct sk_buff, cb));
3514
3515         percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3516         percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3517         inet_hashinfo_init(&tcp_hashinfo);
3518         tcp_hashinfo.bind_bucket_cachep =
3519                 kmem_cache_create("tcp_bind_bucket",
3520                                   sizeof(struct inet_bind_bucket), 0,
3521                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3522
3523         /* Size and allocate the main established and bind bucket
3524          * hash tables.
3525          *
3526          * The methodology is similar to that of the buffer cache.
3527          */
3528         tcp_hashinfo.ehash =
3529                 alloc_large_system_hash("TCP established",
3530                                         sizeof(struct inet_ehash_bucket),
3531                                         thash_entries,
3532                                         17, /* one slot per 128 KB of memory */
3533                                         0,
3534                                         NULL,
3535                                         &tcp_hashinfo.ehash_mask,
3536                                         0,
3537                                         thash_entries ? 0 : 512 * 1024);
3538         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3539                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3540
3541         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3542                 panic("TCP: failed to alloc ehash_locks");
3543         tcp_hashinfo.bhash =
3544                 alloc_large_system_hash("TCP bind",
3545                                         sizeof(struct inet_bind_hashbucket),
3546                                         tcp_hashinfo.ehash_mask + 1,
3547                                         17, /* one slot per 128 KB of memory */
3548                                         0,
3549                                         &tcp_hashinfo.bhash_size,
3550                                         NULL,
3551                                         0,
3552                                         64 * 1024);
3553         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3554         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3555                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3556                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3557         }
3558
3559
3560         cnt = tcp_hashinfo.ehash_mask + 1;
3561         sysctl_tcp_max_orphans = cnt / 2;
3562
3563         tcp_init_mem();
3564         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3565         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3566         max_wshare = min(4UL*1024*1024, limit);
3567         max_rshare = min(6UL*1024*1024, limit);
3568
3569         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3570         sysctl_tcp_wmem[1] = 16*1024;
3571         sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3572
3573         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3574         sysctl_tcp_rmem[1] = 87380;
3575         sysctl_tcp_rmem[2] = max(87380, max_rshare);
3576
3577         pr_info("Hash tables configured (established %u bind %u)\n",
3578                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3579
3580         tcp_v4_init();
3581         tcp_metrics_init();
3582         BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3583         tcp_tasklet_init();
3584 }