net/ipv4/tcp.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the  BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              Implementation of the Transmission Control Protocol(TCP).
   7  *
   8  * Authors:     Ross Biro
   9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
  11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
  12  *              Florian La Roche, <flla@stud.uni-sb.de>
  13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
  15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
  16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
  17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  18  *              Jorge Cwik, <jorge@laser.satlink.net>
  19  *
  20  * Fixes:
  21  *              Alan Cox        :       Numerous verify_area() calls
  22  *              Alan Cox        :       Set the ACK bit on a reset
  23  *              Alan Cox        :       Stopped it crashing if it closed while
  24  *                                      sk->inuse=1 and was trying to connect
  25  *                                      (tcp_err()).
  26  *              Alan Cox        :       All icmp error handling was broken
  27  *                                      pointers passed where wrong and the
  28  *                                      socket was looked up backwards. Nobody
  29  *                                      tested any icmp error code obviously.
  30  *              Alan Cox        :       tcp_err() now handled properly. It
  31  *                                      wakes people on errors. poll
  32  *                                      behaves and the icmp error race
  33  *                                      has gone by moving it into sock.c
  34  *              Alan Cox        :       tcp_send_reset() fixed to work for
  35  *                                      everything not just packets for
  36  *                                      unknown sockets.
  37  *              Alan Cox        :       tcp option processing.
  38  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
  39  *                                      syn rule wrong]
  40  *              Herp Rosmanith  :       More reset fixes
  41  *              Alan Cox        :       No longer acks invalid rst frames.
  42  *                                      Acking any kind of RST is right out.
  43  *              Alan Cox        :       Sets an ignore me flag on an rst
  44  *                                      receive otherwise odd bits of prattle
  45  *                                      escape still
  46  *              Alan Cox        :       Fixed another acking RST frame bug.
  47  *                                      Should stop LAN workplace lockups.
  48  *              Alan Cox        :       Some tidyups using the new skb list
  49  *                                      facilities
  50  *              Alan Cox        :       sk->keepopen now seems to work
  51  *              Alan Cox        :       Pulls options out correctly on accepts
  52  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
  53  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
  54  *                                      bit to skb ops.
  55  *              Alan Cox        :       Tidied tcp_data to avoid a potential
  56  *                                      nasty.
  57  *              Alan Cox        :       Added some better commenting, as the
  58  *                                      tcp is hard to follow
  59  *              Alan Cox        :       Removed incorrect check for 20 * psh
  60  *      Michael O'Reilly        :       ack < copied bug fix.
  61  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
  62  *              Alan Cox        :       FIN with no memory -> CRASH
  63  *              Alan Cox        :       Added socket option proto entries.
  64  *                                      Also added awareness of them to accept.
  65  *              Alan Cox        :       Added TCP options (SOL_TCP)
  66  *              Alan Cox        :       Switched wakeup calls to callbacks,
  67  *                                      so the kernel can layer network
  68  *                                      sockets.
  69  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
  70  *              Alan Cox        :       Handle FIN (more) properly (we hope).
  71  *              Alan Cox        :       RST frames sent on unsynchronised
  72  *                                      state ack error.
  73  *              Alan Cox        :       Put in missing check for SYN bit.
  74  *              Alan Cox        :       Added tcp_select_window() aka NET2E
  75  *                                      window non shrink trick.
  76  *              Alan Cox        :       Added a couple of small NET2E timer
  77  *                                      fixes
  78  *              Charles Hedrick :       TCP fixes
  79  *              Toomas Tamm     :       TCP window fixes
  80  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
  81  *              Charles Hedrick :       Rewrote most of it to actually work
  82  *              Linus           :       Rewrote tcp_read() and URG handling
  83  *                                      completely
  84  *              Gerhard Koerting:       Fixed some missing timer handling
  85  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
  86  *              Gerhard Koerting:       PC/TCP workarounds
  87  *              Adam Caldwell   :       Assorted timer/timing errors
  88  *              Matthew Dillon  :       Fixed another RST bug
  89  *              Alan Cox        :       Move to kernel side addressing changes.
  90  *              Alan Cox        :       Beginning work on TCP fastpathing
  91  *                                      (not yet usable)
  92  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
  93  *              Alan Cox        :       TCP fast path debugging
  94  *              Alan Cox        :       Window clamping
  95  *              Michael Riepe   :       Bug in tcp_check()
  96  *              Matt Dillon     :       More TCP improvements and RST bug fixes
  97  *              Matt Dillon     :       Yet more small nasties remove from the
  98  *                                      TCP code (Be very nice to this man if
  99  *                                      tcp finally works 100%) 8)
 100  *              Alan Cox        :       BSD accept semantics.
 101  *              Alan Cox        :       Reset on closedown bug.
 102  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
 103  *              Michael Pall    :       Handle poll() after URG properly in
 104  *                                      all cases.
 105  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
 106  *                                      (multi URG PUSH broke rlogin).
 107  *              Michael Pall    :       Fix the multi URG PUSH problem in
 108  *                                      tcp_readable(), poll() after URG
 109  *                                      works now.
 110  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
 111  *                                      BSD api.
 112  *              Alan Cox        :       Changed the semantics of sk->socket to
 113  *                                      fix a race and a signal problem with
 114  *                                      accept() and async I/O.
 115  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
 116  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
 117  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
 118  *                                      clients/servers which listen in on
 119  *                                      fixed ports.
 120  *              Alan Cox        :       Cleaned the above up and shrank it to
 121  *                                      a sensible code size.
 122  *              Alan Cox        :       Self connect lockup fix.
 123  *              Alan Cox        :       No connect to multicast.
 124  *              Ross Biro       :       Close unaccepted children on master
 125  *                                      socket close.
 126  *              Alan Cox        :       Reset tracing code.
 127  *              Alan Cox        :       Spurious resets on shutdown.
 128  *              Alan Cox        :       Giant 15 minute/60 second timer error
 129  *              Alan Cox        :       Small whoops in polling before an
 130  *                                      accept.
 131  *              Alan Cox        :       Kept the state trace facility since
 132  *                                      it's handy for debugging.
 133  *              Alan Cox        :       More reset handler fixes.
 134  *              Alan Cox        :       Started rewriting the code based on
 135  *                                      the RFC's for other useful protocol
 136  *                                      references see: Comer, KA9Q NOS, and
 137  *                                      for a reference on the difference
 138  *                                      between specifications and how BSD
 139  *                                      works see the 4.4lite source.
 140  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
 141  *                                      close.
 142  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
 143  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
 144  *              Alan Cox        :       Reimplemented timers as per the RFC
 145  *                                      and using multiple timers for sanity.
 146  *              Alan Cox        :       Small bug fixes, and a lot of new
 147  *                                      comments.
 148  *              Alan Cox        :       Fixed dual reader crash by locking
 149  *                                      the buffers (much like datagram.c)
 150  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
 151  *                                      now gets fed up of retrying without
 152  *                                      (even a no space) answer.
 153  *              Alan Cox        :       Extracted closing code better
 154  *              Alan Cox        :       Fixed the closing state machine to
 155  *                                      resemble the RFC.
 156  *              Alan Cox        :       More 'per spec' fixes.
 157  *              Jorge Cwik      :       Even faster checksumming.
 158  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
 159  *                                      only frames. At least one pc tcp stack
 160  *                                      generates them.
 161  *              Alan Cox        :       Cache last socket.
 162  *              Alan Cox        :       Per route irtt.
 163  *              Matt Day        :       poll()->select() match BSD precisely on error
 164  *              Alan Cox        :       New buffers
 165  *              Marc Tamsky     :       Various sk->prot->retransmits and
 166  *                                      sk->retransmits misupdating fixed.
 167  *                                      Fixed tcp_write_timeout: stuck close,
 168  *                                      and TCP syn retries gets used now.
 169  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
 170  *                                      ack if state is TCP_CLOSED.
 171  *              Alan Cox        :       Look up device on a retransmit - routes may
 172  *                                      change. Doesn't yet cope with MSS shrink right
 173  *                                      but it's a start!
 174  *              Marc Tamsky     :       Closing in closing fixes.
 175  *              Mike Shaver     :       RFC1122 verifications.
 176  *              Alan Cox        :       rcv_saddr errors.
 177  *              Alan Cox        :       Block double connect().
 178  *              Alan Cox        :       Small hooks for enSKIP.
 179  *              Alexey Kuznetsov:       Path MTU discovery.
 180  *              Alan Cox        :       Support soft errors.
 181  *              Alan Cox        :       Fix MTU discovery pathological case
 182  *                                      when the remote claims no mtu!
 183  *              Marc Tamsky     :       TCP_CLOSE fix.
 184  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
 185  *                                      window but wrong (fixes NT lpd problems)
 186  *              Pedro Roque     :       Better TCP window handling, delayed ack.
 187  *              Joerg Reuter    :       No modification of locked buffers in
 188  *                                      tcp_do_retransmit()
 189  *              Eric Schenk     :       Changed receiver side silly window
 190  *                                      avoidance algorithm to BSD style
 191  *                                      algorithm. This doubles throughput
 192  *                                      against machines running Solaris,
 193  *                                      and seems to result in general
 194  *                                      improvement.
 195  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
 196  *      Willy Konynenberg       :       Transparent proxying support.
 197  *      Mike McLagan            :       Routing by source
 198  *              Keith Owens     :       Do proper merging with partial SKB's in
 199  *                                      tcp_do_sendmsg to avoid burstiness.
 200  *              Eric Schenk     :       Fix fast close down bug with
 201  *                                      shutdown() followed by close().
 202  *              Andi Kleen      :       Make poll agree with SIGIO
 203  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
 204  *                                      lingertime == 0 (RFC 793 ABORT Call)
 205  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
 206  *                                      csum_and_copy_from_user() if possible.
 207  *
 208  *              This program is free software; you can redistribute it and/or
 209  *              modify it under the terms of the GNU General Public License
 210  *              as published by the Free Software Foundation; either version
 211  *              2 of the License, or(at your option) any later version.
 212  *
 213  * Description of States:
 214  *
 215  *      TCP_SYN_SENT            sent a connection request, waiting for ack
 216  *
 217  *      TCP_SYN_RECV            received a connection request, sent ack,
 218  *                              waiting for final ack in three-way handshake.
 219  *
 220  *      TCP_ESTABLISHED         connection established
 221  *
 222  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
 223  *                              transmission of remaining buffered data
 224  *
 225  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
 226  *                              to shutdown
 227  *
 228  *      TCP_CLOSING             both sides have shutdown but we still have
 229  *                              data we have to finish sending
 230  *
 231  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
 232  *                              closed, can only be entered from FIN_WAIT2
 233  *                              or CLOSING.  Required because the other end
 234  *                              may not have gotten our last ACK causing it
 235  *                              to retransmit the data packet (which we ignore)
 236  *
 237  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
 238  *                              us to finish writing our data and to shutdown
 239  *                              (we have to close() to move on to LAST_ACK)
 240  *
 241  *      TCP_LAST_ACK            out side has shutdown after remote has
 242  *                              shutdown.  There may still be data in our
 243  *                              buffer that we have to finish sending
 244  *
 245  *      TCP_CLOSE               socket is finished
 246  */
 247
 248 #define pr_fmt(fmt) "TCP: " fmt
 249
 250 #include <crypto/hash.h>
 251 #include <linux/kernel.h>
 252 #include <linux/module.h>
 253 #include <linux/types.h>
 254 #include <linux/fcntl.h>
 255 #include <linux/poll.h>
 256 #include <linux/inet_diag.h>
 257 #include <linux/init.h>
 258 #include <linux/fs.h>
 259 #include <linux/skbuff.h>
 260 #include <linux/scatterlist.h>
 261 #include <linux/splice.h>
 262 #include <linux/net.h>
 263 #include <linux/socket.h>
 264 #include <linux/random.h>
 265 #include <linux/bootmem.h>
 266 #include <linux/highmem.h>
 267 #include <linux/swap.h>
 268 #include <linux/cache.h>
 269 #include <linux/err.h>
 270 #include <linux/time.h>
 271 #include <linux/slab.h>
 272 #include <linux/errqueue.h>
 273 #include <linux/static_key.h>
 274
 275 #include <net/icmp.h>
 276 #include <net/inet_common.h>
 277 #include <net/tcp.h>
 278 #include <net/xfrm.h>
 279 #include <net/ip.h>
 280 #include <net/sock.h>
 281
 282 #include <linux/uaccess.h>
 283 #include <asm/ioctls.h>
 284 #include <net/busy_poll.h>
 285
 286 struct percpu_counter tcp_orphan_count;
 287 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 288
 289 long sysctl_tcp_mem[3] __read_mostly;
 290 EXPORT_SYMBOL(sysctl_tcp_mem);
 291
 292 atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
 293 EXPORT_SYMBOL(tcp_memory_allocated);
 294
 295 #if IS_ENABLED(CONFIG_SMC)
 296 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
 297 EXPORT_SYMBOL(tcp_have_smc);
 298 #endif
 299
 300 /*
 301  * Current number of TCP sockets.
 302  */
 303 struct percpu_counter tcp_sockets_allocated;
 304 EXPORT_SYMBOL(tcp_sockets_allocated);
 305
 306 /*
 307  * TCP splice context
 308  */
 309 struct tcp_splice_state {
 310         struct pipe_inode_info *pipe;
 311         size_t len;
 312         unsigned int flags;
 313 };
 314
 315 /*
 316  * Pressure flag: try to collapse.
 317  * Technical note: it is used by multiple contexts non atomically.
 318  * All the __sk_mem_schedule() is of this nature: accounting
 319  * is strict, actions are advisory and have some latency.
 320  */
 321 unsigned long tcp_memory_pressure __read_mostly;
 322 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
 323
 324 void tcp_enter_memory_pressure(struct sock *sk)
 325 {
 326         unsigned long val;
 327
 328         if (tcp_memory_pressure)
 329                 return;
 330         val = jiffies;
 331
 332         if (!val)
 333                 val--;
 334         if (!cmpxchg(&tcp_memory_pressure, 0, val))
 335                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
 336 }
 337 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
 338
 339 void tcp_leave_memory_pressure(struct sock *sk)
 340 {
 341         unsigned long val;
 342
 343         if (!tcp_memory_pressure)
 344                 return;
 345         val = xchg(&tcp_memory_pressure, 0);
 346         if (val)
 347                 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
 348                               jiffies_to_msecs(jiffies - val));
 349 }
 350 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
 351
 352 /* Convert seconds to retransmits based on initial and max timeout */
 353 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
 354 {
 355         u8 res = 0;
 356
 357         if (seconds > 0) {
 358                 int period = timeout;
 359
 360                 res = 1;
 361                 while (seconds > period && res < 255) {
 362                         res++;
 363                         timeout <<= 1;
 364                         if (timeout > rto_max)
 365                                 timeout = rto_max;
 366                         period += timeout;
 367                 }
 368         }
 369         return res;
 370 }
 371
 372 /* Convert retransmits to seconds based on initial and max timeout */
 373 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
 374 {
 375         int period = 0;
 376
 377         if (retrans > 0) {
 378                 period = timeout;
 379                 while (--retrans) {
 380                         timeout <<= 1;
 381                         if (timeout > rto_max)
 382                                 timeout = rto_max;
 383                         period += timeout;
 384                 }
 385         }
 386         return period;
 387 }
 388
 389 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
 390 {
 391         u32 rate = READ_ONCE(tp->rate_delivered);
 392         u32 intv = READ_ONCE(tp->rate_interval_us);
 393         u64 rate64 = 0;
 394
 395         if (rate && intv) {
 396                 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
 397                 do_div(rate64, intv);
 398         }
 399         return rate64;
 400 }
 401
 402 /* Address-family independent initialization for a tcp_sock.
 403  *
 404  * NOTE: A lot of things set to zero explicitly by call to
 405  *       sk_alloc() so need not be done here.
 406  */
 407 void tcp_init_sock(struct sock *sk)
 408 {
 409         struct inet_connection_sock *icsk = inet_csk(sk);
 410         struct tcp_sock *tp = tcp_sk(sk);
 411
 412         tp->out_of_order_queue = RB_ROOT;
 413         sk->tcp_rtx_queue = RB_ROOT;
 414         tcp_init_xmit_timers(sk);
 415         INIT_LIST_HEAD(&tp->tsq_node);
 416         INIT_LIST_HEAD(&tp->tsorted_sent_queue);
 417
 418         icsk->icsk_rto = TCP_TIMEOUT_INIT;
 419         tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
 420         minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
 421
 422         /* So many TCP implementations out there (incorrectly) count the
 423          * initial SYN frame in their delayed-ACK and congestion control
 424          * algorithms that we must have the following bandaid to talk
 425          * efficiently to them.  -DaveM
 426          */
 427         tp->snd_cwnd = TCP_INIT_CWND;
 428
 429         /* There's a bubble in the pipe until at least the first ACK. */
 430         tp->app_limited = ~0U;
 431
 432         /* See draft-stevens-tcpca-spec-01 for discussion of the
 433          * initialization of these values.
 434          */
 435         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 436         tp->snd_cwnd_clamp = ~0;
 437         tp->mss_cache = TCP_MSS_DEFAULT;
 438
 439         tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
 440         tcp_assign_congestion_control(sk);
 441
 442         tp->tsoffset = 0;
 443         tp->rack.reo_wnd_steps = 1;
 444
 445         sk->sk_state = TCP_CLOSE;
 446
 447         sk->sk_write_space = sk_stream_write_space;
 448         sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
 449
 450         icsk->icsk_sync_mss = tcp_sync_mss;
 451
 452         sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
 453         sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
 454
 455         sk_sockets_allocated_inc(sk);
 456         sk->sk_route_forced_caps = NETIF_F_GSO;
 457 }
 458 EXPORT_SYMBOL(tcp_init_sock);
 459
 460 void tcp_init_transfer(struct sock *sk, int bpf_op)
 461 {
 462         struct inet_connection_sock *icsk = inet_csk(sk);
 463
 464         tcp_mtup_init(sk);
 465         icsk->icsk_af_ops->rebuild_header(sk);
 466         tcp_init_metrics(sk);
 467         tcp_call_bpf(sk, bpf_op, 0, NULL);
 468         tcp_init_congestion_control(sk);
 469         tcp_init_buffer_space(sk);
 470 }
 471
 472 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
 473 {
 474         struct sk_buff *skb = tcp_write_queue_tail(sk);
 475
 476         if (tsflags && skb) {
 477                 struct skb_shared_info *shinfo = skb_shinfo(skb);
 478                 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 479
 480                 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
 481                 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
 482                         tcb->txstamp_ack = 1;
 483                 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
 484                         shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
 485         }
 486 }
 487
 488 static inline bool tcp_stream_is_readable(const struct tcp_sock *tp,
 489                                           int target, struct sock *sk)
 490 {
 491         return (tp->rcv_nxt - tp->copied_seq >= target) ||
 492                 (sk->sk_prot->stream_memory_read ?
 493                 sk->sk_prot->stream_memory_read(sk) : false);
 494 }
 495
 496 /*
 497  * Socket is not locked. We are protected from async events by poll logic and
 498  * correct handling of state changes made by other threads is impossible in
 499  * any case.
 500  */
 501 __poll_t tcp_poll_mask(struct socket *sock, __poll_t events)
 502 {
 503         struct sock *sk = sock->sk;
 504         const struct tcp_sock *tp = tcp_sk(sk);
 505         __poll_t mask = 0;
 506         int state;
 507
 508         state = inet_sk_state_load(sk);
 509         if (state == TCP_LISTEN)
 510                 return inet_csk_listen_poll(sk);
 511
 512         /*
 513          * EPOLLHUP is certainly not done right. But poll() doesn't
 514          * have a notion of HUP in just one direction, and for a
 515          * socket the read side is more interesting.
 516          *
 517          * Some poll() documentation says that EPOLLHUP is incompatible
 518          * with the EPOLLOUT/POLLWR flags, so somebody should check this
 519          * all. But careful, it tends to be safer to return too many
 520          * bits than too few, and you can easily break real applications
 521          * if you don't tell them that something has hung up!
 522          *
 523          * Check-me.
 524          *
 525          * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
 526          * our fs/select.c). It means that after we received EOF,
 527          * poll always returns immediately, making impossible poll() on write()
 528          * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
 529          * if and only if shutdown has been made in both directions.
 530          * Actually, it is interesting to look how Solaris and DUX
 531          * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
 532          * then we could set it on SND_SHUTDOWN. BTW examples given
 533          * in Stevens' books assume exactly this behaviour, it explains
 534          * why EPOLLHUP is incompatible with EPOLLOUT.  --ANK
 535          *
 536          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 537          * blocking on fresh not-connected or disconnected socket. --ANK
 538          */
 539         if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
 540                 mask |= EPOLLHUP;
 541         if (sk->sk_shutdown & RCV_SHUTDOWN)
 542                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
 543
 544         /* Connected or passive Fast Open socket? */
 545         if (state != TCP_SYN_SENT &&
 546             (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
 547                 int target = sock_rcvlowat(sk, 0, INT_MAX);
 548
 549                 if (tp->urg_seq == tp->copied_seq &&
 550                     !sock_flag(sk, SOCK_URGINLINE) &&
 551                     tp->urg_data)
 552                         target++;
 553
 554                 if (tcp_stream_is_readable(tp, target, sk))
 555                         mask |= EPOLLIN | EPOLLRDNORM;
 556
 557                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 558                         if (sk_stream_is_writeable(sk)) {
 559                                 mask |= EPOLLOUT | EPOLLWRNORM;
 560                         } else {  /* send SIGIO later */
 561                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 562                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 563
 564                                 /* Race breaker. If space is freed after
 565                                  * wspace test but before the flags are set,
 566                                  * IO signal will be lost. Memory barrier
 567                                  * pairs with the input side.
 568                                  */
 569                                 smp_mb__after_atomic();
 570                                 if (sk_stream_is_writeable(sk))
 571                                         mask |= EPOLLOUT | EPOLLWRNORM;
 572                         }
 573                 } else
 574                         mask |= EPOLLOUT | EPOLLWRNORM;
 575
 576                 if (tp->urg_data & TCP_URG_VALID)
 577                         mask |= EPOLLPRI;
 578         } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
 579                 /* Active TCP fastopen socket with defer_connect
 580                  * Return EPOLLOUT so application can call write()
 581                  * in order for kernel to generate SYN+data
 582                  */
 583                 mask |= EPOLLOUT | EPOLLWRNORM;
 584         }
 585         /* This barrier is coupled with smp_wmb() in tcp_reset() */
 586         smp_rmb();
 587         if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
 588                 mask |= EPOLLERR;
 589
 590         return mask;
 591 }
 592 EXPORT_SYMBOL(tcp_poll_mask);
 593
 594 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 595 {
 596         struct tcp_sock *tp = tcp_sk(sk);
 597         int answ;
 598         bool slow;
 599
 600         switch (cmd) {
 601         case SIOCINQ:
 602                 if (sk->sk_state == TCP_LISTEN)
 603                         return -EINVAL;
 604
 605                 slow = lock_sock_fast(sk);
 606                 answ = tcp_inq(sk);
 607                 unlock_sock_fast(sk, slow);
 608                 break;
 609         case SIOCATMARK:
 610                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 611                 break;
 612         case SIOCOUTQ:
 613                 if (sk->sk_state == TCP_LISTEN)
 614                         return -EINVAL;
 615
 616                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 617                         answ = 0;
 618                 else
 619                         answ = tp->write_seq - tp->snd_una;
 620                 break;
 621         case SIOCOUTQNSD:
 622                 if (sk->sk_state == TCP_LISTEN)
 623                         return -EINVAL;
 624
 625                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 626                         answ = 0;
 627                 else
 628                         answ = tp->write_seq - tp->snd_nxt;
 629                 break;
 630         default:
 631                 return -ENOIOCTLCMD;
 632         }
 633
 634         return put_user(answ, (int __user *)arg);
 635 }
 636 EXPORT_SYMBOL(tcp_ioctl);
 637
 638 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 639 {
 640         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 641         tp->pushed_seq = tp->write_seq;
 642 }
 643
 644 static inline bool forced_push(const struct tcp_sock *tp)
 645 {
 646         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 647 }
 648
 649 static void skb_entail(struct sock *sk, struct sk_buff *skb)
 650 {
 651         struct tcp_sock *tp = tcp_sk(sk);
 652         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 653
 654         skb->csum    = 0;
 655         tcb->seq     = tcb->end_seq = tp->write_seq;
 656         tcb->tcp_flags = TCPHDR_ACK;
 657         tcb->sacked  = 0;
 658         __skb_header_release(skb);
 659         tcp_add_write_queue_tail(sk, skb);
 660         sk->sk_wmem_queued += skb->truesize;
 661         sk_mem_charge(sk, skb->truesize);
 662         if (tp->nonagle & TCP_NAGLE_PUSH)
 663                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 664
 665         tcp_slow_start_after_idle_check(sk);
 666 }
 667
 668 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
 669 {
 670         if (flags & MSG_OOB)
 671                 tp->snd_up = tp->write_seq;
 672 }
 673
 674 /* If a not yet filled skb is pushed, do not send it if
 675  * we have data packets in Qdisc or NIC queues :
 676  * Because TX completion will happen shortly, it gives a chance
 677  * to coalesce future sendmsg() payload into this skb, without
 678  * need for a timer, and with no latency trade off.
 679  * As packets containing data payload have a bigger truesize
 680  * than pure acks (dataless) packets, the last checks prevent
 681  * autocorking if we only have an ACK in Qdisc/NIC queues,
 682  * or if TX completion was delayed after we processed ACK packet.
 683  */
 684 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
 685                                 int size_goal)
 686 {
 687         return skb->len < size_goal &&
 688                sock_net(sk)->ipv4.sysctl_tcp_autocorking &&
 689                !tcp_rtx_queue_empty(sk) &&
 690                refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
 691 }
 692
 693 static void tcp_push(struct sock *sk, int flags, int mss_now,
 694                      int nonagle, int size_goal)
 695 {
 696         struct tcp_sock *tp = tcp_sk(sk);
 697         struct sk_buff *skb;
 698
 699         skb = tcp_write_queue_tail(sk);
 700         if (!skb)
 701                 return;
 702         if (!(flags & MSG_MORE) || forced_push(tp))
 703                 tcp_mark_push(tp, skb);
 704
 705         tcp_mark_urg(tp, flags);
 706
 707         if (tcp_should_autocork(sk, skb, size_goal)) {
 708
 709                 /* avoid atomic op if TSQ_THROTTLED bit is already set */
 710                 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
 711                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
 712                         set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
 713                 }
 714                 /* It is possible TX completion already happened
 715                  * before we set TSQ_THROTTLED.
 716                  */
 717                 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
 718                         return;
 719         }
 720
 721         if (flags & MSG_MORE)
 722                 nonagle = TCP_NAGLE_CORK;
 723
 724         __tcp_push_pending_frames(sk, mss_now, nonagle);
 725 }
 726
 727 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 728                                 unsigned int offset, size_t len)
 729 {
 730         struct tcp_splice_state *tss = rd_desc->arg.data;
 731         int ret;
 732
 733         ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
 734                               min(rd_desc->count, len), tss->flags);
 735         if (ret > 0)
 736                 rd_desc->count -= ret;
 737         return ret;
 738 }
 739
 740 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 741 {
 742         /* Store TCP splice context information in read_descriptor_t. */
 743         read_descriptor_t rd_desc = {
 744                 .arg.data = tss,
 745                 .count    = tss->len,
 746         };
 747
 748         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 749 }
 750
 751 /**
 752  *  tcp_splice_read - splice data from TCP socket to a pipe
 753  * @sock:       socket to splice from
 754  * @ppos:       position (not valid)
 755  * @pipe:       pipe to splice to
 756  * @len:        number of bytes to splice
 757  * @flags:      splice modifier flags
 758  *
 759  * Description:
 760  *    Will read pages from given socket and fill them into a pipe.
 761  *
 762  **/
 763 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 764                         struct pipe_inode_info *pipe, size_t len,
 765                         unsigned int flags)
 766 {
 767         struct sock *sk = sock->sk;
 768         struct tcp_splice_state tss = {
 769                 .pipe = pipe,
 770                 .len = len,
 771                 .flags = flags,
 772         };
 773         long timeo;
 774         ssize_t spliced;
 775         int ret;
 776
 777         sock_rps_record_flow(sk);
 778         /*
 779          * We can't seek on a socket input
 780          */
 781         if (unlikely(*ppos))
 782                 return -ESPIPE;
 783
 784         ret = spliced = 0;
 785
 786         lock_sock(sk);
 787
 788         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 789         while (tss.len) {
 790                 ret = __tcp_splice_read(sk, &tss);
 791                 if (ret < 0)
 792                         break;
 793                 else if (!ret) {
 794                         if (spliced)
 795                                 break;
 796                         if (sock_flag(sk, SOCK_DONE))
 797                                 break;
 798                         if (sk->sk_err) {
 799                                 ret = sock_error(sk);
 800                                 break;
 801                         }
 802                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 803                                 break;
 804                         if (sk->sk_state == TCP_CLOSE) {
 805                                 /*
 806                                  * This occurs when user tries to read
 807                                  * from never connected socket.
 808                                  */
 809                                 if (!sock_flag(sk, SOCK_DONE))
 810                                         ret = -ENOTCONN;
 811                                 break;
 812                         }
 813                         if (!timeo) {
 814                                 ret = -EAGAIN;
 815                                 break;
 816                         }
 817                         /* if __tcp_splice_read() got nothing while we have
 818                          * an skb in receive queue, we do not want to loop.
 819                          * This might happen with URG data.
 820                          */
 821                         if (!skb_queue_empty(&sk->sk_receive_queue))
 822                                 break;
 823                         sk_wait_data(sk, &timeo, NULL);
 824                         if (signal_pending(current)) {
 825                                 ret = sock_intr_errno(timeo);
 826                                 break;
 827                         }
 828                         continue;
 829                 }
 830                 tss.len -= ret;
 831                 spliced += ret;
 832
 833                 if (!timeo)
 834                         break;
 835                 release_sock(sk);
 836                 lock_sock(sk);
 837
 838                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 839                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 840                     signal_pending(current))
 841                         break;
 842         }
 843
 844         release_sock(sk);
 845
 846         if (spliced)
 847                 return spliced;
 848
 849         return ret;
 850 }
 851 EXPORT_SYMBOL(tcp_splice_read);
 852
 853 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
 854                                     bool force_schedule)
 855 {
 856         struct sk_buff *skb;
 857
 858         /* The TCP header must be at least 32-bit aligned.  */
 859         size = ALIGN(size, 4);
 860
 861         if (unlikely(tcp_under_memory_pressure(sk)))
 862                 sk_mem_reclaim_partial(sk);
 863
 864         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 865         if (likely(skb)) {
 866                 bool mem_scheduled;
 867
 868                 if (force_schedule) {
 869                         mem_scheduled = true;
 870                         sk_forced_mem_schedule(sk, skb->truesize);
 871                 } else {
 872                         mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
 873                 }
 874                 if (likely(mem_scheduled)) {
 875                         skb_reserve(skb, sk->sk_prot->max_header);
 876                         /*
 877                          * Make sure that we have exactly size bytes
 878                          * available to the caller, no more, no less.
 879                          */
 880                         skb->reserved_tailroom = skb->end - skb->tail - size;
 881                         INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
 882                         return skb;
 883                 }
 884                 __kfree_skb(skb);
 885         } else {
 886                 sk->sk_prot->enter_memory_pressure(sk);
 887                 sk_stream_moderate_sndbuf(sk);
 888         }
 889         return NULL;
 890 }
 891
 892 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 893                                        int large_allowed)
 894 {
 895         struct tcp_sock *tp = tcp_sk(sk);
 896         u32 new_size_goal, size_goal;
 897
 898         if (!large_allowed)
 899                 return mss_now;
 900
 901         /* Note : tcp_tso_autosize() will eventually split this later */
 902         new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
 903         new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
 904
 905         /* We try hard to avoid divides here */
 906         size_goal = tp->gso_segs * mss_now;
 907         if (unlikely(new_size_goal < size_goal ||
 908                      new_size_goal >= size_goal + mss_now)) {
 909                 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
 910                                      sk->sk_gso_max_segs);
 911                 size_goal = tp->gso_segs * mss_now;
 912         }
 913
 914         return max(size_goal, mss_now);
 915 }
 916
 917 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 918 {
 919         int mss_now;
 920
 921         mss_now = tcp_current_mss(sk);
 922         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 923
 924         return mss_now;
 925 }
 926
 927 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
 928                          size_t size, int flags)
 929 {
 930         struct tcp_sock *tp = tcp_sk(sk);
 931         int mss_now, size_goal;
 932         int err;
 933         ssize_t copied;
 934         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 935
 936         /* Wait for a connection to finish. One exception is TCP Fast Open
 937          * (passive side) where data is allowed to be sent before a connection
 938          * is fully established.
 939          */
 940         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
 941             !tcp_passive_fastopen(sk)) {
 942                 err = sk_stream_wait_connect(sk, &timeo);
 943                 if (err != 0)
 944                         goto out_err;
 945         }
 946
 947         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 948
 949         mss_now = tcp_send_mss(sk, &size_goal, flags);
 950         copied = 0;
 951
 952         err = -EPIPE;
 953         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 954                 goto out_err;
 955
 956         while (size > 0) {
 957                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 958                 int copy, i;
 959                 bool can_coalesce;
 960
 961                 if (!skb || (copy = size_goal - skb->len) <= 0 ||
 962                     !tcp_skb_can_collapse_to(skb)) {
 963 new_segment:
 964                         if (!sk_stream_memory_free(sk))
 965                                 goto wait_for_sndbuf;
 966
 967                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
 968                                         tcp_rtx_and_write_queues_empty(sk));
 969                         if (!skb)
 970                                 goto wait_for_memory;
 971
 972                         skb_entail(sk, skb);
 973                         copy = size_goal;
 974                 }
 975
 976                 if (copy > size)
 977                         copy = size;
 978
 979                 i = skb_shinfo(skb)->nr_frags;
 980                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 981                 if (!can_coalesce && i >= sysctl_max_skb_frags) {
 982                         tcp_mark_push(tp, skb);
 983                         goto new_segment;
 984                 }
 985                 if (!sk_wmem_schedule(sk, copy))
 986                         goto wait_for_memory;
 987
 988                 if (can_coalesce) {
 989                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
 990                 } else {
 991                         get_page(page);
 992                         skb_fill_page_desc(skb, i, page, offset, copy);
 993                 }
 994
 995                 if (!(flags & MSG_NO_SHARED_FRAGS))
 996                         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
 997
 998                 skb->len += copy;
 999                 skb->data_len += copy;
1000                 skb->truesize += copy;
1001                 sk->sk_wmem_queued += copy;
1002                 sk_mem_charge(sk, copy);
1003                 skb->ip_summed = CHECKSUM_PARTIAL;
1004                 tp->write_seq += copy;
1005                 TCP_SKB_CB(skb)->end_seq += copy;
1006                 tcp_skb_pcount_set(skb, 0);
1007
1008                 if (!copied)
1009                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1010
1011                 copied += copy;
1012                 offset += copy;
1013                 size -= copy;
1014                 if (!size)
1015                         goto out;
1016
1017                 if (skb->len < size_goal || (flags & MSG_OOB))
1018                         continue;
1019
1020                 if (forced_push(tp)) {
1021                         tcp_mark_push(tp, skb);
1022                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1023                 } else if (skb == tcp_send_head(sk))
1024                         tcp_push_one(sk, mss_now);
1025                 continue;
1026
1027 wait_for_sndbuf:
1028                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1029 wait_for_memory:
1030                 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1031                          TCP_NAGLE_PUSH, size_goal);
1032
1033                 err = sk_stream_wait_memory(sk, &timeo);
1034                 if (err != 0)
1035                         goto do_error;
1036
1037                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1038         }
1039
1040 out:
1041         if (copied) {
1042                 tcp_tx_timestamp(sk, sk->sk_tsflags);
1043                 if (!(flags & MSG_SENDPAGE_NOTLAST))
1044                         tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1045         }
1046         return copied;
1047
1048 do_error:
1049         if (copied)
1050                 goto out;
1051 out_err:
1052         /* make sure we wake any epoll edge trigger waiter */
1053         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1054                      err == -EAGAIN)) {
1055                 sk->sk_write_space(sk);
1056                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1057         }
1058         return sk_stream_error(sk, flags, err);
1059 }
1060 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1061
1062 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1063                         size_t size, int flags)
1064 {
1065         if (!(sk->sk_route_caps & NETIF_F_SG))
1066                 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1067
1068         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1069
1070         return do_tcp_sendpages(sk, page, offset, size, flags);
1071 }
1072 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1073
1074 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1075                  size_t size, int flags)
1076 {
1077         int ret;
1078
1079         lock_sock(sk);
1080         ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1081         release_sock(sk);
1082
1083         return ret;
1084 }
1085 EXPORT_SYMBOL(tcp_sendpage);
1086
1087 /* Do not bother using a page frag for very small frames.
1088  * But use this heuristic only for the first skb in write queue.
1089  *
1090  * Having no payload in skb->head allows better SACK shifting
1091  * in tcp_shift_skb_data(), reducing sack/rack overhead, because
1092  * write queue has less skbs.
1093  * Each skb can hold up to MAX_SKB_FRAGS * 32Kbytes, or ~0.5 MB.
1094  * This also speeds up tso_fragment(), since it wont fallback
1095  * to tcp_fragment().
1096  */
1097 static int linear_payload_sz(bool first_skb)
1098 {
1099         if (first_skb)
1100                 return SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1101         return 0;
1102 }
1103
1104 static int select_size(bool first_skb, bool zc)
1105 {
1106         if (zc)
1107                 return 0;
1108         return linear_payload_sz(first_skb);
1109 }
1110
1111 void tcp_free_fastopen_req(struct tcp_sock *tp)
1112 {
1113         if (tp->fastopen_req) {
1114                 kfree(tp->fastopen_req);
1115                 tp->fastopen_req = NULL;
1116         }
1117 }
1118
1119 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1120                                 int *copied, size_t size)
1121 {
1122         struct tcp_sock *tp = tcp_sk(sk);
1123         struct inet_sock *inet = inet_sk(sk);
1124         struct sockaddr *uaddr = msg->msg_name;
1125         int err, flags;
1126
1127         if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1128             (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1129              uaddr->sa_family == AF_UNSPEC))
1130                 return -EOPNOTSUPP;
1131         if (tp->fastopen_req)
1132                 return -EALREADY; /* Another Fast Open is in progress */
1133
1134         tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1135                                    sk->sk_allocation);
1136         if (unlikely(!tp->fastopen_req))
1137                 return -ENOBUFS;
1138         tp->fastopen_req->data = msg;
1139         tp->fastopen_req->size = size;
1140
1141         if (inet->defer_connect) {
1142                 err = tcp_connect(sk);
1143                 /* Same failure procedure as in tcp_v4/6_connect */
1144                 if (err) {
1145                         tcp_set_state(sk, TCP_CLOSE);
1146                         inet->inet_dport = 0;
1147                         sk->sk_route_caps = 0;
1148                 }
1149         }
1150         flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1151         err = __inet_stream_connect(sk->sk_socket, uaddr,
1152                                     msg->msg_namelen, flags, 1);
1153         /* fastopen_req could already be freed in __inet_stream_connect
1154          * if the connection times out or gets rst
1155          */
1156         if (tp->fastopen_req) {
1157                 *copied = tp->fastopen_req->copied;
1158                 tcp_free_fastopen_req(tp);
1159                 inet->defer_connect = 0;
1160         }
1161         return err;
1162 }
1163
1164 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1165 {
1166         struct tcp_sock *tp = tcp_sk(sk);
1167         struct ubuf_info *uarg = NULL;
1168         struct sk_buff *skb;
1169         struct sockcm_cookie sockc;
1170         int flags, err, copied = 0;
1171         int mss_now = 0, size_goal, copied_syn = 0;
1172         bool process_backlog = false;
1173         bool zc = false;
1174         long timeo;
1175
1176         flags = msg->msg_flags;
1177
1178         if (flags & MSG_ZEROCOPY && size) {
1179                 if (sk->sk_state != TCP_ESTABLISHED) {
1180                         err = -EINVAL;
1181                         goto out_err;
1182                 }
1183
1184                 skb = tcp_write_queue_tail(sk);
1185                 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
1186                 if (!uarg) {
1187                         err = -ENOBUFS;
1188                         goto out_err;
1189                 }
1190
1191                 zc = sk->sk_route_caps & NETIF_F_SG;
1192                 if (!zc)
1193                         uarg->zerocopy = 0;
1194         }
1195
1196         if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1197             !tp->repair) {
1198                 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1199                 if (err == -EINPROGRESS && copied_syn > 0)
1200                         goto out;
1201                 else if (err)
1202                         goto out_err;
1203         }
1204
1205         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1206
1207         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1208
1209         /* Wait for a connection to finish. One exception is TCP Fast Open
1210          * (passive side) where data is allowed to be sent before a connection
1211          * is fully established.
1212          */
1213         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1214             !tcp_passive_fastopen(sk)) {
1215                 err = sk_stream_wait_connect(sk, &timeo);
1216                 if (err != 0)
1217                         goto do_error;
1218         }
1219
1220         if (unlikely(tp->repair)) {
1221                 if (tp->repair_queue == TCP_RECV_QUEUE) {
1222                         copied = tcp_send_rcvq(sk, msg, size);
1223                         goto out_nopush;
1224                 }
1225
1226                 err = -EINVAL;
1227                 if (tp->repair_queue == TCP_NO_QUEUE)
1228                         goto out_err;
1229
1230                 /* 'common' sending to sendq */
1231         }
1232
1233         sockc.tsflags = sk->sk_tsflags;
1234         if (msg->msg_controllen) {
1235                 err = sock_cmsg_send(sk, msg, &sockc);
1236                 if (unlikely(err)) {
1237                         err = -EINVAL;
1238                         goto out_err;
1239                 }
1240         }
1241
1242         /* This should be in poll */
1243         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1244
1245         /* Ok commence sending. */
1246         copied = 0;
1247
1248 restart:
1249         mss_now = tcp_send_mss(sk, &size_goal, flags);
1250
1251         err = -EPIPE;
1252         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1253                 goto do_error;
1254
1255         while (msg_data_left(msg)) {
1256                 int copy = 0;
1257
1258                 skb = tcp_write_queue_tail(sk);
1259                 if (skb)
1260                         copy = size_goal - skb->len;
1261
1262                 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1263                         bool first_skb;
1264                         int linear;
1265
1266 new_segment:
1267                         /* Allocate new segment. If the interface is SG,
1268                          * allocate skb fitting to single page.
1269                          */
1270                         if (!sk_stream_memory_free(sk))
1271                                 goto wait_for_sndbuf;
1272
1273                         if (process_backlog && sk_flush_backlog(sk)) {
1274                                 process_backlog = false;
1275                                 goto restart;
1276                         }
1277                         first_skb = tcp_rtx_and_write_queues_empty(sk);
1278                         linear = select_size(first_skb, zc);
1279                         skb = sk_stream_alloc_skb(sk, linear, sk->sk_allocation,
1280                                                   first_skb);
1281                         if (!skb)
1282                                 goto wait_for_memory;
1283
1284                         process_backlog = true;
1285                         skb->ip_summed = CHECKSUM_PARTIAL;
1286
1287                         skb_entail(sk, skb);
1288                         copy = size_goal;
1289
1290                         /* All packets are restored as if they have
1291                          * already been sent. skb_mstamp isn't set to
1292                          * avoid wrong rtt estimation.
1293                          */
1294                         if (tp->repair)
1295                                 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1296                 }
1297
1298                 /* Try to append data to the end of skb. */
1299                 if (copy > msg_data_left(msg))
1300                         copy = msg_data_left(msg);
1301
1302                 /* Where to copy to? */
1303                 if (skb_availroom(skb) > 0 && !zc) {
1304                         /* We have some space in skb head. Superb! */
1305                         copy = min_t(int, copy, skb_availroom(skb));
1306                         err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1307                         if (err)
1308                                 goto do_fault;
1309                 } else if (!zc) {
1310                         bool merge = true;
1311                         int i = skb_shinfo(skb)->nr_frags;
1312                         struct page_frag *pfrag = sk_page_frag(sk);
1313
1314                         if (!sk_page_frag_refill(sk, pfrag))
1315                                 goto wait_for_memory;
1316
1317                         if (!skb_can_coalesce(skb, i, pfrag->page,
1318                                               pfrag->offset)) {
1319                                 if (i >= sysctl_max_skb_frags) {
1320                                         tcp_mark_push(tp, skb);
1321                                         goto new_segment;
1322                                 }
1323                                 merge = false;
1324                         }
1325
1326                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1327
1328                         if (!sk_wmem_schedule(sk, copy))
1329                                 goto wait_for_memory;
1330
1331                         err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1332                                                        pfrag->page,
1333                                                        pfrag->offset,
1334                                                        copy);
1335                         if (err)
1336                                 goto do_error;
1337
1338                         /* Update the skb. */
1339                         if (merge) {
1340                                 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1341                         } else {
1342                                 skb_fill_page_desc(skb, i, pfrag->page,
1343                                                    pfrag->offset, copy);
1344                                 page_ref_inc(pfrag->page);
1345                         }
1346                         pfrag->offset += copy;
1347                 } else {
1348                         err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1349                         if (err == -EMSGSIZE || err == -EEXIST) {
1350                                 tcp_mark_push(tp, skb);
1351                                 goto new_segment;
1352                         }
1353                         if (err < 0)
1354                                 goto do_error;
1355                         copy = err;
1356                 }
1357
1358                 if (!copied)
1359                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1360
1361                 tp->write_seq += copy;
1362                 TCP_SKB_CB(skb)->end_seq += copy;
1363                 tcp_skb_pcount_set(skb, 0);
1364
1365                 copied += copy;
1366                 if (!msg_data_left(msg)) {
1367                         if (unlikely(flags & MSG_EOR))
1368                                 TCP_SKB_CB(skb)->eor = 1;
1369                         goto out;
1370                 }
1371
1372                 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1373                         continue;
1374
1375                 if (forced_push(tp)) {
1376                         tcp_mark_push(tp, skb);
1377                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1378                 } else if (skb == tcp_send_head(sk))
1379                         tcp_push_one(sk, mss_now);
1380                 continue;
1381
1382 wait_for_sndbuf:
1383                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1384 wait_for_memory:
1385                 if (copied)
1386                         tcp_push(sk, flags & ~MSG_MORE, mss_now,
1387                                  TCP_NAGLE_PUSH, size_goal);
1388
1389                 err = sk_stream_wait_memory(sk, &timeo);
1390                 if (err != 0)
1391                         goto do_error;
1392
1393                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1394         }
1395
1396 out:
1397         if (copied) {
1398                 tcp_tx_timestamp(sk, sockc.tsflags);
1399                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1400         }
1401 out_nopush:
1402         sock_zerocopy_put(uarg);
1403         return copied + copied_syn;
1404
1405 do_fault:
1406         if (!skb->len) {
1407                 tcp_unlink_write_queue(skb, sk);
1408                 /* It is the one place in all of TCP, except connection
1409                  * reset, where we can be unlinking the send_head.
1410                  */
1411                 tcp_check_send_head(sk, skb);
1412                 sk_wmem_free_skb(sk, skb);
1413         }
1414
1415 do_error:
1416         if (copied + copied_syn)
1417                 goto out;
1418 out_err:
1419         sock_zerocopy_put_abort(uarg);
1420         err = sk_stream_error(sk, flags, err);
1421         /* make sure we wake any epoll edge trigger waiter */
1422         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1423                      err == -EAGAIN)) {
1424                 sk->sk_write_space(sk);
1425                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1426         }
1427         return err;
1428 }
1429 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1430
1431 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1432 {
1433         int ret;
1434
1435         lock_sock(sk);
1436         ret = tcp_sendmsg_locked(sk, msg, size);
1437         release_sock(sk);
1438
1439         return ret;
1440 }
1441 EXPORT_SYMBOL(tcp_sendmsg);
1442
1443 /*
1444  *      Handle reading urgent data. BSD has very simple semantics for
1445  *      this, no blocking and very strange errors 8)
1446  */
1447
1448 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1449 {
1450         struct tcp_sock *tp = tcp_sk(sk);
1451
1452         /* No URG data to read. */
1453         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1454             tp->urg_data == TCP_URG_READ)
1455                 return -EINVAL; /* Yes this is right ! */
1456
1457         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1458                 return -ENOTCONN;
1459
1460         if (tp->urg_data & TCP_URG_VALID) {
1461                 int err = 0;
1462                 char c = tp->urg_data;
1463
1464                 if (!(flags & MSG_PEEK))
1465                         tp->urg_data = TCP_URG_READ;
1466
1467                 /* Read urgent data. */
1468                 msg->msg_flags |= MSG_OOB;
1469
1470                 if (len > 0) {
1471                         if (!(flags & MSG_TRUNC))
1472                                 err = memcpy_to_msg(msg, &c, 1);
1473                         len = 1;
1474                 } else
1475                         msg->msg_flags |= MSG_TRUNC;
1476
1477                 return err ? -EFAULT : len;
1478         }
1479
1480         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1481                 return 0;
1482
1483         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1484          * the available implementations agree in this case:
1485          * this call should never block, independent of the
1486          * blocking state of the socket.
1487          * Mike <pall@rz.uni-karlsruhe.de>
1488          */
1489         return -EAGAIN;
1490 }
1491
1492 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1493 {
1494         struct sk_buff *skb;
1495         int copied = 0, err = 0;
1496
1497         /* XXX -- need to support SO_PEEK_OFF */
1498
1499         skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1500                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1501                 if (err)
1502                         return err;
1503                 copied += skb->len;
1504         }
1505
1506         skb_queue_walk(&sk->sk_write_queue, skb) {
1507                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1508                 if (err)
1509                         break;
1510
1511                 copied += skb->len;
1512         }
1513
1514         return err ?: copied;
1515 }
1516
1517 /* Clean up the receive buffer for full frames taken by the user,
1518  * then send an ACK if necessary.  COPIED is the number of bytes
1519  * tcp_recvmsg has given to the user so far, it speeds up the
1520  * calculation of whether or not we must ACK for the sake of
1521  * a window update.
1522  */
1523 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1524 {
1525         struct tcp_sock *tp = tcp_sk(sk);
1526         bool time_to_ack = false;
1527
1528         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1529
1530         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1531              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1532              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1533
1534         if (inet_csk_ack_scheduled(sk)) {
1535                 const struct inet_connection_sock *icsk = inet_csk(sk);
1536                    /* Delayed ACKs frequently hit locked sockets during bulk
1537                     * receive. */
1538                 if (icsk->icsk_ack.blocked ||
1539                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1540                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1541                     /*
1542                      * If this read emptied read buffer, we send ACK, if
1543                      * connection is not bidirectional, user drained
1544                      * receive buffer and there was a small segment
1545                      * in queue.
1546                      */
1547                     (copied > 0 &&
1548                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1549                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1550                        !icsk->icsk_ack.pingpong)) &&
1551                       !atomic_read(&sk->sk_rmem_alloc)))
1552                         time_to_ack = true;
1553         }
1554
1555         /* We send an ACK if we can now advertise a non-zero window
1556          * which has been raised "significantly".
1557          *
1558          * Even if window raised up to infinity, do not send window open ACK
1559          * in states, where we will not receive more. It is useless.
1560          */
1561         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1562                 __u32 rcv_window_now = tcp_receive_window(tp);
1563
1564                 /* Optimize, __tcp_select_window() is not cheap. */
1565                 if (2*rcv_window_now <= tp->window_clamp) {
1566                         __u32 new_window = __tcp_select_window(sk);
1567
1568                         /* Send ACK now, if this read freed lots of space
1569                          * in our buffer. Certainly, new_window is new window.
1570                          * We can advertise it now, if it is not less than current one.
1571                          * "Lots" means "at least twice" here.
1572                          */
1573                         if (new_window && new_window >= 2 * rcv_window_now)
1574                                 time_to_ack = true;
1575                 }
1576         }
1577         if (time_to_ack)
1578                 tcp_send_ack(sk);
1579 }
1580
1581 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1582 {
1583         struct sk_buff *skb;
1584         u32 offset;
1585
1586         while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1587                 offset = seq - TCP_SKB_CB(skb)->seq;
1588                 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1589                         pr_err_once("%s: found a SYN, please report !\n", __func__);
1590                         offset--;
1591                 }
1592                 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1593                         *off = offset;
1594                         return skb;
1595                 }
1596                 /* This looks weird, but this can happen if TCP collapsing
1597                  * splitted a fat GRO packet, while we released socket lock
1598                  * in skb_splice_bits()
1599                  */
1600                 sk_eat_skb(sk, skb);
1601         }
1602         return NULL;
1603 }
1604
1605 /*
1606  * This routine provides an alternative to tcp_recvmsg() for routines
1607  * that would like to handle copying from skbuffs directly in 'sendfile'
1608  * fashion.
1609  * Note:
1610  *      - It is assumed that the socket was locked by the caller.
1611  *      - The routine does not block.
1612  *      - At present, there is no support for reading OOB data
1613  *        or for 'peeking' the socket using this routine
1614  *        (although both would be easy to implement).
1615  */
1616 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1617                   sk_read_actor_t recv_actor)
1618 {
1619         struct sk_buff *skb;
1620         struct tcp_sock *tp = tcp_sk(sk);
1621         u32 seq = tp->copied_seq;
1622         u32 offset;
1623         int copied = 0;
1624
1625         if (sk->sk_state == TCP_LISTEN)
1626                 return -ENOTCONN;
1627         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1628                 if (offset < skb->len) {
1629                         int used;
1630                         size_t len;
1631
1632                         len = skb->len - offset;
1633                         /* Stop reading if we hit a patch of urgent data */
1634                         if (tp->urg_data) {
1635                                 u32 urg_offset = tp->urg_seq - seq;
1636                                 if (urg_offset < len)
1637                                         len = urg_offset;
1638                                 if (!len)
1639                                         break;
1640                         }
1641                         used = recv_actor(desc, skb, offset, len);
1642                         if (used <= 0) {
1643                                 if (!copied)
1644                                         copied = used;
1645                                 break;
1646                         } else if (used <= len) {
1647                                 seq += used;
1648                                 copied += used;
1649                                 offset += used;
1650                         }
1651                         /* If recv_actor drops the lock (e.g. TCP splice
1652                          * receive) the skb pointer might be invalid when
1653                          * getting here: tcp_collapse might have deleted it
1654                          * while aggregating skbs from the socket queue.
1655                          */
1656                         skb = tcp_recv_skb(sk, seq - 1, &offset);
1657                         if (!skb)
1658                                 break;
1659                         /* TCP coalescing might have appended data to the skb.
1660                          * Try to splice more frags
1661                          */
1662                         if (offset + 1 != skb->len)
1663                                 continue;
1664                 }
1665                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1666                         sk_eat_skb(sk, skb);
1667                         ++seq;
1668                         break;
1669                 }
1670                 sk_eat_skb(sk, skb);
1671                 if (!desc->count)
1672                         break;
1673                 tp->copied_seq = seq;
1674         }
1675         tp->copied_seq = seq;
1676
1677         tcp_rcv_space_adjust(sk);
1678
1679         /* Clean up data we have read: This will do ACK frames. */
1680         if (copied > 0) {
1681                 tcp_recv_skb(sk, seq, &offset);
1682                 tcp_cleanup_rbuf(sk, copied);
1683         }
1684         return copied;
1685 }
1686 EXPORT_SYMBOL(tcp_read_sock);
1687
1688 int tcp_peek_len(struct socket *sock)
1689 {
1690         return tcp_inq(sock->sk);
1691 }
1692 EXPORT_SYMBOL(tcp_peek_len);
1693
1694 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1695 int tcp_set_rcvlowat(struct sock *sk, int val)
1696 {
1697         int cap;
1698
1699         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1700                 cap = sk->sk_rcvbuf >> 1;
1701         else
1702                 cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1;
1703         val = min(val, cap);
1704         sk->sk_rcvlowat = val ? : 1;
1705
1706         /* Check if we need to signal EPOLLIN right now */
1707         tcp_data_ready(sk);
1708
1709         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1710                 return 0;
1711
1712         val <<= 1;
1713         if (val > sk->sk_rcvbuf) {
1714                 sk->sk_rcvbuf = val;
1715                 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1716         }
1717         return 0;
1718 }
1719 EXPORT_SYMBOL(tcp_set_rcvlowat);
1720
1721 #ifdef CONFIG_MMU
1722 static const struct vm_operations_struct tcp_vm_ops = {
1723 };
1724
1725 int tcp_mmap(struct file *file, struct socket *sock,
1726              struct vm_area_struct *vma)
1727 {
1728         if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1729                 return -EPERM;
1730         vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1731
1732         /* Instruct vm_insert_page() to not down_read(mmap_sem) */
1733         vma->vm_flags |= VM_MIXEDMAP;
1734
1735         vma->vm_ops = &tcp_vm_ops;
1736         return 0;
1737 }
1738 EXPORT_SYMBOL(tcp_mmap);
1739
1740 static int tcp_zerocopy_receive(struct sock *sk,
1741                                 struct tcp_zerocopy_receive *zc)
1742 {
1743         unsigned long address = (unsigned long)zc->address;
1744         const skb_frag_t *frags = NULL;
1745         u32 length = 0, seq, offset;
1746         struct vm_area_struct *vma;
1747         struct sk_buff *skb = NULL;
1748         struct tcp_sock *tp;
1749         int ret;
1750
1751         if (address & (PAGE_SIZE - 1) || address != zc->address)
1752                 return -EINVAL;
1753
1754         if (sk->sk_state == TCP_LISTEN)
1755                 return -ENOTCONN;
1756
1757         sock_rps_record_flow(sk);
1758
1759         down_read(&current->mm->mmap_sem);
1760
1761         ret = -EINVAL;
1762         vma = find_vma(current->mm, address);
1763         if (!vma || vma->vm_start > address || vma->vm_ops != &tcp_vm_ops)
1764                 goto out;
1765         zc->length = min_t(unsigned long, zc->length, vma->vm_end - address);
1766
1767         tp = tcp_sk(sk);
1768         seq = tp->copied_seq;
1769         zc->length = min_t(u32, zc->length, tcp_inq(sk));
1770         zc->length &= ~(PAGE_SIZE - 1);
1771
1772         zap_page_range(vma, address, zc->length);
1773
1774         zc->recv_skip_hint = 0;
1775         ret = 0;
1776         while (length + PAGE_SIZE <= zc->length) {
1777                 if (zc->recv_skip_hint < PAGE_SIZE) {
1778                         if (skb) {
1779                                 skb = skb->next;
1780                                 offset = seq - TCP_SKB_CB(skb)->seq;
1781                         } else {
1782                                 skb = tcp_recv_skb(sk, seq, &offset);
1783                         }
1784
1785                         zc->recv_skip_hint = skb->len - offset;
1786                         offset -= skb_headlen(skb);
1787                         if ((int)offset < 0 || skb_has_frag_list(skb))
1788                                 break;
1789                         frags = skb_shinfo(skb)->frags;
1790                         while (offset) {
1791                                 if (frags->size > offset)
1792                                         goto out;
1793                                 offset -= frags->size;
1794                                 frags++;
1795                         }
1796                 }
1797                 if (frags->size != PAGE_SIZE || frags->page_offset)
1798                         break;
1799                 ret = vm_insert_page(vma, address + length,
1800                                      skb_frag_page(frags));
1801                 if (ret)
1802                         break;
1803                 length += PAGE_SIZE;
1804                 seq += PAGE_SIZE;
1805                 zc->recv_skip_hint -= PAGE_SIZE;
1806                 frags++;
1807         }
1808 out:
1809         up_read(&current->mm->mmap_sem);
1810         if (length) {
1811                 tp->copied_seq = seq;
1812                 tcp_rcv_space_adjust(sk);
1813
1814                 /* Clean up data we have read: This will do ACK frames. */
1815                 tcp_recv_skb(sk, seq, &offset);
1816                 tcp_cleanup_rbuf(sk, length);
1817                 ret = 0;
1818                 if (length == zc->length)
1819                         zc->recv_skip_hint = 0;
1820         } else {
1821                 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
1822                         ret = -EIO;
1823         }
1824         zc->length = length;
1825         return ret;
1826 }
1827 #endif
1828
1829 static void tcp_update_recv_tstamps(struct sk_buff *skb,
1830                                     struct scm_timestamping *tss)
1831 {
1832         if (skb->tstamp)
1833                 tss->ts[0] = ktime_to_timespec(skb->tstamp);
1834         else
1835                 tss->ts[0] = (struct timespec) {0};
1836
1837         if (skb_hwtstamps(skb)->hwtstamp)
1838                 tss->ts[2] = ktime_to_timespec(skb_hwtstamps(skb)->hwtstamp);
1839         else
1840                 tss->ts[2] = (struct timespec) {0};
1841 }
1842
1843 /* Similar to __sock_recv_timestamp, but does not require an skb */
1844 static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
1845                                struct scm_timestamping *tss)
1846 {
1847         struct timeval tv;
1848         bool has_timestamping = false;
1849
1850         if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
1851                 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
1852                         if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
1853                                 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPNS,
1854                                          sizeof(tss->ts[0]), &tss->ts[0]);
1855                         } else {
1856                                 tv.tv_sec = tss->ts[0].tv_sec;
1857                                 tv.tv_usec = tss->ts[0].tv_nsec / 1000;
1858
1859                                 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMP,
1860                                          sizeof(tv), &tv);
1861                         }
1862                 }
1863
1864                 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
1865                         has_timestamping = true;
1866                 else
1867                         tss->ts[0] = (struct timespec) {0};
1868         }
1869
1870         if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
1871                 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
1872                         has_timestamping = true;
1873                 else
1874                         tss->ts[2] = (struct timespec) {0};
1875         }
1876
1877         if (has_timestamping) {
1878                 tss->ts[1] = (struct timespec) {0};
1879                 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPING,
1880                          sizeof(*tss), tss);
1881         }
1882 }
1883
1884 static int tcp_inq_hint(struct sock *sk)
1885 {
1886         const struct tcp_sock *tp = tcp_sk(sk);
1887         u32 copied_seq = READ_ONCE(tp->copied_seq);
1888         u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
1889         int inq;
1890
1891         inq = rcv_nxt - copied_seq;
1892         if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
1893                 lock_sock(sk);
1894                 inq = tp->rcv_nxt - tp->copied_seq;
1895                 release_sock(sk);
1896         }
1897         return inq;
1898 }
1899
1900 /*
1901  *      This routine copies from a sock struct into the user buffer.
1902  *
1903  *      Technical note: in 2.3 we work on _locked_ socket, so that
1904  *      tricks with *seq access order and skb->users are not required.
1905  *      Probably, code can be easily improved even more.
1906  */
1907
1908 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1909                 int flags, int *addr_len)
1910 {
1911         struct tcp_sock *tp = tcp_sk(sk);
1912         int copied = 0;
1913         u32 peek_seq;
1914         u32 *seq;
1915         unsigned long used;
1916         int err, inq;
1917         int target;             /* Read at least this many bytes */
1918         long timeo;
1919         struct sk_buff *skb, *last;
1920         u32 urg_hole = 0;
1921         struct scm_timestamping tss;
1922         bool has_tss = false;
1923         bool has_cmsg;
1924
1925         if (unlikely(flags & MSG_ERRQUEUE))
1926                 return inet_recv_error(sk, msg, len, addr_len);
1927
1928         if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1929             (sk->sk_state == TCP_ESTABLISHED))
1930                 sk_busy_loop(sk, nonblock);
1931
1932         lock_sock(sk);
1933
1934         err = -ENOTCONN;
1935         if (sk->sk_state == TCP_LISTEN)
1936                 goto out;
1937
1938         has_cmsg = tp->recvmsg_inq;
1939         timeo = sock_rcvtimeo(sk, nonblock);
1940
1941         /* Urgent data needs to be handled specially. */
1942         if (flags & MSG_OOB)
1943                 goto recv_urg;
1944
1945         if (unlikely(tp->repair)) {
1946                 err = -EPERM;
1947                 if (!(flags & MSG_PEEK))
1948                         goto out;
1949
1950                 if (tp->repair_queue == TCP_SEND_QUEUE)
1951                         goto recv_sndq;
1952
1953                 err = -EINVAL;
1954                 if (tp->repair_queue == TCP_NO_QUEUE)
1955                         goto out;
1956
1957                 /* 'common' recv queue MSG_PEEK-ing */
1958         }
1959
1960         seq = &tp->copied_seq;
1961         if (flags & MSG_PEEK) {
1962                 peek_seq = tp->copied_seq;
1963                 seq = &peek_seq;
1964         }
1965
1966         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1967
1968         do {
1969                 u32 offset;
1970
1971                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1972                 if (tp->urg_data && tp->urg_seq == *seq) {
1973                         if (copied)
1974                                 break;
1975                         if (signal_pending(current)) {
1976                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1977                                 break;
1978                         }
1979                 }
1980
1981                 /* Next get a buffer. */
1982
1983                 last = skb_peek_tail(&sk->sk_receive_queue);
1984                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1985                         last = skb;
1986                         /* Now that we have two receive queues this
1987                          * shouldn't happen.
1988                          */
1989                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1990                                  "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1991                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1992                                  flags))
1993                                 break;
1994
1995                         offset = *seq - TCP_SKB_CB(skb)->seq;
1996                         if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1997                                 pr_err_once("%s: found a SYN, please report !\n", __func__);
1998                                 offset--;
1999                         }
2000                         if (offset < skb->len)
2001                                 goto found_ok_skb;
2002                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2003                                 goto found_fin_ok;
2004                         WARN(!(flags & MSG_PEEK),
2005                              "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
2006                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2007                 }
2008
2009                 /* Well, if we have backlog, try to process it now yet. */
2010
2011                 if (copied >= target && !sk->sk_backlog.tail)
2012                         break;
2013
2014                 if (copied) {
2015                         if (sk->sk_err ||
2016                             sk->sk_state == TCP_CLOSE ||
2017                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
2018                             !timeo ||
2019                             signal_pending(current))
2020                                 break;
2021                 } else {
2022                         if (sock_flag(sk, SOCK_DONE))
2023                                 break;
2024
2025                         if (sk->sk_err) {
2026                                 copied = sock_error(sk);
2027                                 break;
2028                         }
2029
2030                         if (sk->sk_shutdown & RCV_SHUTDOWN)
2031                                 break;
2032
2033                         if (sk->sk_state == TCP_CLOSE) {
2034                                 if (!sock_flag(sk, SOCK_DONE)) {
2035                                         /* This occurs when user tries to read
2036                                          * from never connected socket.
2037                                          */
2038                                         copied = -ENOTCONN;
2039                                         break;
2040                                 }
2041                                 break;
2042                         }
2043
2044                         if (!timeo) {
2045                                 copied = -EAGAIN;
2046                                 break;
2047                         }
2048
2049                         if (signal_pending(current)) {
2050                                 copied = sock_intr_errno(timeo);
2051                                 break;
2052                         }
2053                 }
2054
2055                 tcp_cleanup_rbuf(sk, copied);
2056
2057                 if (copied >= target) {
2058                         /* Do not sleep, just process backlog. */
2059                         release_sock(sk);
2060                         lock_sock(sk);
2061                 } else {
2062                         sk_wait_data(sk, &timeo, last);
2063                 }
2064
2065                 if ((flags & MSG_PEEK) &&
2066                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
2067                         net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2068                                             current->comm,
2069                                             task_pid_nr(current));
2070                         peek_seq = tp->copied_seq;
2071                 }
2072                 continue;
2073
2074         found_ok_skb:
2075                 /* Ok so how much can we use? */
2076                 used = skb->len - offset;
2077                 if (len < used)
2078                         used = len;
2079
2080                 /* Do we have urgent data here? */
2081                 if (tp->urg_data) {
2082                         u32 urg_offset = tp->urg_seq - *seq;
2083                         if (urg_offset < used) {
2084                                 if (!urg_offset) {
2085                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
2086                                                 ++*seq;
2087                                                 urg_hole++;
2088                                                 offset++;
2089                                                 used--;
2090                                                 if (!used)
2091                                                         goto skip_copy;
2092                                         }
2093                                 } else
2094                                         used = urg_offset;
2095                         }
2096                 }
2097
2098                 if (!(flags & MSG_TRUNC)) {
2099                         err = skb_copy_datagram_msg(skb, offset, msg, used);
2100                         if (err) {
2101                                 /* Exception. Bailout! */
2102                                 if (!copied)
2103                                         copied = -EFAULT;
2104                                 break;
2105                         }
2106                 }
2107
2108                 *seq += used;
2109                 copied += used;
2110                 len -= used;
2111
2112                 tcp_rcv_space_adjust(sk);
2113
2114 skip_copy:
2115                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
2116                         tp->urg_data = 0;
2117                         tcp_fast_path_check(sk);
2118                 }
2119                 if (used + offset < skb->len)
2120                         continue;
2121
2122                 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2123                         tcp_update_recv_tstamps(skb, &tss);
2124                         has_tss = true;
2125                         has_cmsg = true;
2126                 }
2127                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2128                         goto found_fin_ok;
2129                 if (!(flags & MSG_PEEK))
2130                         sk_eat_skb(sk, skb);
2131                 continue;
2132
2133         found_fin_ok:
2134                 /* Process the FIN. */
2135                 ++*seq;
2136                 if (!(flags & MSG_PEEK))
2137                         sk_eat_skb(sk, skb);
2138                 break;
2139         } while (len > 0);
2140
2141         /* According to UNIX98, msg_name/msg_namelen are ignored
2142          * on connected socket. I was just happy when found this 8) --ANK
2143          */
2144
2145         /* Clean up data we have read: This will do ACK frames. */
2146         tcp_cleanup_rbuf(sk, copied);
2147
2148         release_sock(sk);
2149
2150         if (has_cmsg) {
2151                 if (has_tss)
2152                         tcp_recv_timestamp(msg, sk, &tss);
2153                 if (tp->recvmsg_inq) {
2154                         inq = tcp_inq_hint(sk);
2155                         put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2156                 }
2157         }
2158
2159         return copied;
2160
2161 out:
2162         release_sock(sk);
2163         return err;
2164
2165 recv_urg:
2166         err = tcp_recv_urg(sk, msg, len, flags);
2167         goto out;
2168
2169 recv_sndq:
2170         err = tcp_peek_sndq(sk, msg, len);
2171         goto out;
2172 }
2173 EXPORT_SYMBOL(tcp_recvmsg);
2174
2175 void tcp_set_state(struct sock *sk, int state)
2176 {
2177         int oldstate = sk->sk_state;
2178
2179         /* We defined a new enum for TCP states that are exported in BPF
2180          * so as not force the internal TCP states to be frozen. The
2181          * following checks will detect if an internal state value ever
2182          * differs from the BPF value. If this ever happens, then we will
2183          * need to remap the internal value to the BPF value before calling
2184          * tcp_call_bpf_2arg.
2185          */
2186         BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2187         BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2188         BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2189         BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2190         BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2191         BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2192         BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2193         BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2194         BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2195         BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2196         BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2197         BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2198         BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2199
2200         if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2201                 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2202
2203         switch (state) {
2204         case TCP_ESTABLISHED:
2205                 if (oldstate != TCP_ESTABLISHED)
2206                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2207                 break;
2208
2209         case TCP_CLOSE:
2210                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2211                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2212
2213                 sk->sk_prot->unhash(sk);
2214                 if (inet_csk(sk)->icsk_bind_hash &&
2215                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2216                         inet_put_port(sk);
2217                 /* fall through */
2218         default:
2219                 if (oldstate == TCP_ESTABLISHED)
2220                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2221         }
2222
2223         /* Change state AFTER socket is unhashed to avoid closed
2224          * socket sitting in hash tables.
2225          */
2226         inet_sk_state_store(sk, state);
2227
2228 #ifdef STATE_TRACE
2229         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
2230 #endif
2231 }
2232 EXPORT_SYMBOL_GPL(tcp_set_state);
2233
2234 /*
2235  *      State processing on a close. This implements the state shift for
2236  *      sending our FIN frame. Note that we only send a FIN for some
2237  *      states. A shutdown() may have already sent the FIN, or we may be
2238  *      closed.
2239  */
2240
2241 static const unsigned char new_state[16] = {
2242   /* current state:        new state:      action:      */
2243   [0 /* (Invalid) */]   = TCP_CLOSE,
2244   [TCP_ESTABLISHED]     = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2245   [TCP_SYN_SENT]        = TCP_CLOSE,
2246   [TCP_SYN_RECV]        = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2247   [TCP_FIN_WAIT1]       = TCP_FIN_WAIT1,
2248   [TCP_FIN_WAIT2]       = TCP_FIN_WAIT2,
2249   [TCP_TIME_WAIT]       = TCP_CLOSE,
2250   [TCP_CLOSE]           = TCP_CLOSE,
2251   [TCP_CLOSE_WAIT]      = TCP_LAST_ACK  | TCP_ACTION_FIN,
2252   [TCP_LAST_ACK]        = TCP_LAST_ACK,
2253   [TCP_LISTEN]          = TCP_CLOSE,
2254   [TCP_CLOSING]         = TCP_CLOSING,
2255   [TCP_NEW_SYN_RECV]    = TCP_CLOSE,    /* should not happen ! */
2256 };
2257
2258 static int tcp_close_state(struct sock *sk)
2259 {
2260         int next = (int)new_state[sk->sk_state];
2261         int ns = next & TCP_STATE_MASK;
2262
2263         tcp_set_state(sk, ns);
2264
2265         return next & TCP_ACTION_FIN;
2266 }
2267
2268 /*
2269  *      Shutdown the sending side of a connection. Much like close except
2270  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2271  */
2272
2273 void tcp_shutdown(struct sock *sk, int how)
2274 {
2275         /*      We need to grab some memory, and put together a FIN,
2276          *      and then put it into the queue to be sent.
2277          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2278          */
2279         if (!(how & SEND_SHUTDOWN))
2280                 return;
2281
2282         /* If we've already sent a FIN, or it's a closed state, skip this. */
2283         if ((1 << sk->sk_state) &
2284             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2285              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2286                 /* Clear out any half completed packets.  FIN if needed. */
2287                 if (tcp_close_state(sk))
2288                         tcp_send_fin(sk);
2289         }
2290 }
2291 EXPORT_SYMBOL(tcp_shutdown);
2292
2293 bool tcp_check_oom(struct sock *sk, int shift)
2294 {
2295         bool too_many_orphans, out_of_socket_memory;
2296
2297         too_many_orphans = tcp_too_many_orphans(sk, shift);
2298         out_of_socket_memory = tcp_out_of_memory(sk);
2299
2300         if (too_many_orphans)
2301                 net_info_ratelimited("too many orphaned sockets\n");
2302         if (out_of_socket_memory)
2303                 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2304         return too_many_orphans || out_of_socket_memory;
2305 }
2306
2307 void tcp_close(struct sock *sk, long timeout)
2308 {
2309         struct sk_buff *skb;
2310         int data_was_unread = 0;
2311         int state;
2312
2313         lock_sock(sk);
2314         sk->sk_shutdown = SHUTDOWN_MASK;
2315
2316         if (sk->sk_state == TCP_LISTEN) {
2317                 tcp_set_state(sk, TCP_CLOSE);
2318
2319                 /* Special case. */
2320                 inet_csk_listen_stop(sk);
2321
2322                 goto adjudge_to_death;
2323         }
2324
2325         /*  We need to flush the recv. buffs.  We do this only on the
2326          *  descriptor close, not protocol-sourced closes, because the
2327          *  reader process may not have drained the data yet!
2328          */
2329         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2330                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2331
2332                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2333                         len--;
2334                 data_was_unread += len;
2335                 __kfree_skb(skb);
2336         }
2337
2338         sk_mem_reclaim(sk);
2339
2340         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2341         if (sk->sk_state == TCP_CLOSE)
2342                 goto adjudge_to_death;
2343
2344         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2345          * data was lost. To witness the awful effects of the old behavior of
2346          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2347          * GET in an FTP client, suspend the process, wait for the client to
2348          * advertise a zero window, then kill -9 the FTP client, wheee...
2349          * Note: timeout is always zero in such a case.
2350          */
2351         if (unlikely(tcp_sk(sk)->repair)) {
2352                 sk->sk_prot->disconnect(sk, 0);
2353         } else if (data_was_unread) {
2354                 /* Unread data was tossed, zap the connection. */
2355                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2356                 tcp_set_state(sk, TCP_CLOSE);
2357                 tcp_send_active_reset(sk, sk->sk_allocation);
2358         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2359                 /* Check zero linger _after_ checking for unread data. */
2360                 sk->sk_prot->disconnect(sk, 0);
2361                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2362         } else if (tcp_close_state(sk)) {
2363                 /* We FIN if the application ate all the data before
2364                  * zapping the connection.
2365                  */
2366
2367                 /* RED-PEN. Formally speaking, we have broken TCP state
2368                  * machine. State transitions:
2369                  *
2370                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2371                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2372                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2373                  *
2374                  * are legal only when FIN has been sent (i.e. in window),
2375                  * rather than queued out of window. Purists blame.
2376                  *
2377                  * F.e. "RFC state" is ESTABLISHED,
2378                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2379                  *
2380                  * The visible declinations are that sometimes
2381                  * we enter time-wait state, when it is not required really
2382                  * (harmless), do not send active resets, when they are
2383                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2384                  * they look as CLOSING or LAST_ACK for Linux)
2385                  * Probably, I missed some more holelets.
2386                  *                                              --ANK
2387                  * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2388                  * in a single packet! (May consider it later but will
2389                  * probably need API support or TCP_CORK SYN-ACK until
2390                  * data is written and socket is closed.)
2391                  */
2392                 tcp_send_fin(sk);
2393         }
2394
2395         sk_stream_wait_close(sk, timeout);
2396
2397 adjudge_to_death:
2398         state = sk->sk_state;
2399         sock_hold(sk);
2400         sock_orphan(sk);
2401
2402         /* It is the last release_sock in its life. It will remove backlog. */
2403         release_sock(sk);
2404
2405
2406         /* Now socket is owned by kernel and we acquire BH lock
2407          *  to finish close. No need to check for user refs.
2408          */
2409         local_bh_disable();
2410         bh_lock_sock(sk);
2411         WARN_ON(sock_owned_by_user(sk));
2412
2413         percpu_counter_inc(sk->sk_prot->orphan_count);
2414
2415         /* Have we already been destroyed by a softirq or backlog? */
2416         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2417                 goto out;
2418
2419         /*      This is a (useful) BSD violating of the RFC. There is a
2420          *      problem with TCP as specified in that the other end could
2421          *      keep a socket open forever with no application left this end.
2422          *      We use a 1 minute timeout (about the same as BSD) then kill
2423          *      our end. If they send after that then tough - BUT: long enough
2424          *      that we won't make the old 4*rto = almost no time - whoops
2425          *      reset mistake.
2426          *
2427          *      Nope, it was not mistake. It is really desired behaviour
2428          *      f.e. on http servers, when such sockets are useless, but
2429          *      consume significant resources. Let's do it with special
2430          *      linger2 option.                                 --ANK
2431          */
2432
2433         if (sk->sk_state == TCP_FIN_WAIT2) {
2434                 struct tcp_sock *tp = tcp_sk(sk);
2435                 if (tp->linger2 < 0) {
2436                         tcp_set_state(sk, TCP_CLOSE);
2437                         tcp_send_active_reset(sk, GFP_ATOMIC);
2438                         __NET_INC_STATS(sock_net(sk),
2439                                         LINUX_MIB_TCPABORTONLINGER);
2440                 } else {
2441                         const int tmo = tcp_fin_time(sk);
2442
2443                         if (tmo > TCP_TIMEWAIT_LEN) {
2444                                 inet_csk_reset_keepalive_timer(sk,
2445                                                 tmo - TCP_TIMEWAIT_LEN);
2446                         } else {
2447                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2448                                 goto out;
2449                         }
2450                 }
2451         }
2452         if (sk->sk_state != TCP_CLOSE) {
2453                 sk_mem_reclaim(sk);
2454                 if (tcp_check_oom(sk, 0)) {
2455                         tcp_set_state(sk, TCP_CLOSE);
2456                         tcp_send_active_reset(sk, GFP_ATOMIC);
2457                         __NET_INC_STATS(sock_net(sk),
2458                                         LINUX_MIB_TCPABORTONMEMORY);
2459                 } else if (!check_net(sock_net(sk))) {
2460                         /* Not possible to send reset; just close */
2461                         tcp_set_state(sk, TCP_CLOSE);
2462                 }
2463         }
2464
2465         if (sk->sk_state == TCP_CLOSE) {
2466                 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2467                 /* We could get here with a non-NULL req if the socket is
2468                  * aborted (e.g., closed with unread data) before 3WHS
2469                  * finishes.
2470                  */
2471                 if (req)
2472                         reqsk_fastopen_remove(sk, req, false);
2473                 inet_csk_destroy_sock(sk);
2474         }
2475         /* Otherwise, socket is reprieved until protocol close. */
2476
2477 out:
2478         bh_unlock_sock(sk);
2479         local_bh_enable();
2480         sock_put(sk);
2481 }
2482 EXPORT_SYMBOL(tcp_close);
2483
2484 /* These states need RST on ABORT according to RFC793 */
2485
2486 static inline bool tcp_need_reset(int state)
2487 {
2488         return (1 << state) &
2489                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2490                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2491 }
2492
2493 static void tcp_rtx_queue_purge(struct sock *sk)
2494 {
2495         struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2496
2497         while (p) {
2498                 struct sk_buff *skb = rb_to_skb(p);
2499
2500                 p = rb_next(p);
2501                 /* Since we are deleting whole queue, no need to
2502                  * list_del(&skb->tcp_tsorted_anchor)
2503                  */
2504                 tcp_rtx_queue_unlink(skb, sk);
2505                 sk_wmem_free_skb(sk, skb);
2506         }
2507 }
2508
2509 void tcp_write_queue_purge(struct sock *sk)
2510 {
2511         struct sk_buff *skb;
2512
2513         tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2514         while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2515                 tcp_skb_tsorted_anchor_cleanup(skb);
2516                 sk_wmem_free_skb(sk, skb);
2517         }
2518         tcp_rtx_queue_purge(sk);
2519         INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2520         sk_mem_reclaim(sk);
2521         tcp_clear_all_retrans_hints(tcp_sk(sk));
2522         tcp_sk(sk)->packets_out = 0;
2523 }
2524
2525 int tcp_disconnect(struct sock *sk, int flags)
2526 {
2527         struct inet_sock *inet = inet_sk(sk);
2528         struct inet_connection_sock *icsk = inet_csk(sk);
2529         struct tcp_sock *tp = tcp_sk(sk);
2530         int err = 0;
2531         int old_state = sk->sk_state;
2532
2533         if (old_state != TCP_CLOSE)
2534                 tcp_set_state(sk, TCP_CLOSE);
2535
2536         /* ABORT function of RFC793 */
2537         if (old_state == TCP_LISTEN) {
2538                 inet_csk_listen_stop(sk);
2539         } else if (unlikely(tp->repair)) {
2540                 sk->sk_err = ECONNABORTED;
2541         } else if (tcp_need_reset(old_state) ||
2542                    (tp->snd_nxt != tp->write_seq &&
2543                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2544                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2545                  * states
2546                  */
2547                 tcp_send_active_reset(sk, gfp_any());
2548                 sk->sk_err = ECONNRESET;
2549         } else if (old_state == TCP_SYN_SENT)
2550                 sk->sk_err = ECONNRESET;
2551
2552         tcp_clear_xmit_timers(sk);
2553         __skb_queue_purge(&sk->sk_receive_queue);
2554         tcp_write_queue_purge(sk);
2555         tcp_fastopen_active_disable_ofo_check(sk);
2556         skb_rbtree_purge(&tp->out_of_order_queue);
2557
2558         inet->inet_dport = 0;
2559
2560         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2561                 inet_reset_saddr(sk);
2562
2563         sk->sk_shutdown = 0;
2564         sock_reset_flag(sk, SOCK_DONE);
2565         tp->srtt_us = 0;
2566         tp->write_seq += tp->max_window + 2;
2567         if (tp->write_seq == 0)
2568                 tp->write_seq = 1;
2569         icsk->icsk_backoff = 0;
2570         tp->snd_cwnd = 2;
2571         icsk->icsk_probes_out = 0;
2572         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2573         tp->snd_cwnd_cnt = 0;
2574         tp->window_clamp = 0;
2575         tp->delivered_ce = 0;
2576         tcp_set_ca_state(sk, TCP_CA_Open);
2577         tp->is_sack_reneg = 0;
2578         tcp_clear_retrans(tp);
2579         inet_csk_delack_init(sk);
2580         /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2581          * issue in __tcp_select_window()
2582          */
2583         icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2584         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2585         __sk_dst_reset(sk);
2586         dst_release(sk->sk_rx_dst);
2587         sk->sk_rx_dst = NULL;
2588         tcp_saved_syn_free(tp);
2589         tp->compressed_ack = 0;
2590
2591         /* Clean up fastopen related fields */
2592         tcp_free_fastopen_req(tp);
2593         inet->defer_connect = 0;
2594
2595         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2596
2597         if (sk->sk_frag.page) {
2598                 put_page(sk->sk_frag.page);
2599                 sk->sk_frag.page = NULL;
2600                 sk->sk_frag.offset = 0;
2601         }
2602
2603         sk->sk_error_report(sk);
2604         return err;
2605 }
2606 EXPORT_SYMBOL(tcp_disconnect);
2607
2608 static inline bool tcp_can_repair_sock(const struct sock *sk)
2609 {
2610         return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2611                 (sk->sk_state != TCP_LISTEN);
2612 }
2613
2614 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2615 {
2616         struct tcp_repair_window opt;
2617
2618         if (!tp->repair)
2619                 return -EPERM;
2620
2621         if (len != sizeof(opt))
2622                 return -EINVAL;
2623
2624         if (copy_from_user(&opt, optbuf, sizeof(opt)))
2625                 return -EFAULT;
2626
2627         if (opt.max_window < opt.snd_wnd)
2628                 return -EINVAL;
2629
2630         if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2631                 return -EINVAL;
2632
2633         if (after(opt.rcv_wup, tp->rcv_nxt))
2634                 return -EINVAL;
2635
2636         tp->snd_wl1     = opt.snd_wl1;
2637         tp->snd_wnd     = opt.snd_wnd;
2638         tp->max_window  = opt.max_window;
2639
2640         tp->rcv_wnd     = opt.rcv_wnd;
2641         tp->rcv_wup     = opt.rcv_wup;
2642
2643         return 0;
2644 }
2645
2646 static int tcp_repair_options_est(struct sock *sk,
2647                 struct tcp_repair_opt __user *optbuf, unsigned int len)
2648 {
2649         struct tcp_sock *tp = tcp_sk(sk);
2650         struct tcp_repair_opt opt;
2651
2652         while (len >= sizeof(opt)) {
2653                 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2654                         return -EFAULT;
2655
2656                 optbuf++;
2657                 len -= sizeof(opt);
2658
2659                 switch (opt.opt_code) {
2660                 case TCPOPT_MSS:
2661                         tp->rx_opt.mss_clamp = opt.opt_val;
2662                         tcp_mtup_init(sk);
2663                         break;
2664                 case TCPOPT_WINDOW:
2665                         {
2666                                 u16 snd_wscale = opt.opt_val & 0xFFFF;
2667                                 u16 rcv_wscale = opt.opt_val >> 16;
2668
2669                                 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
2670                                         return -EFBIG;
2671
2672                                 tp->rx_opt.snd_wscale = snd_wscale;
2673                                 tp->rx_opt.rcv_wscale = rcv_wscale;
2674                                 tp->rx_opt.wscale_ok = 1;
2675                         }
2676                         break;
2677                 case TCPOPT_SACK_PERM:
2678                         if (opt.opt_val != 0)
2679                                 return -EINVAL;
2680
2681                         tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2682                         break;
2683                 case TCPOPT_TIMESTAMP:
2684                         if (opt.opt_val != 0)
2685                                 return -EINVAL;
2686
2687                         tp->rx_opt.tstamp_ok = 1;
2688                         break;
2689                 }
2690         }
2691
2692         return 0;
2693 }
2694
2695 /*
2696  *      Socket option code for TCP.
2697  */
2698 static int do_tcp_setsockopt(struct sock *sk, int level,
2699                 int optname, char __user *optval, unsigned int optlen)
2700 {
2701         struct tcp_sock *tp = tcp_sk(sk);
2702         struct inet_connection_sock *icsk = inet_csk(sk);
2703         struct net *net = sock_net(sk);
2704         int val;
2705         int err = 0;
2706
2707         /* These are data/string values, all the others are ints */
2708         switch (optname) {
2709         case TCP_CONGESTION: {
2710                 char name[TCP_CA_NAME_MAX];
2711
2712                 if (optlen < 1)
2713                         return -EINVAL;
2714
2715                 val = strncpy_from_user(name, optval,
2716                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2717                 if (val < 0)
2718                         return -EFAULT;
2719                 name[val] = 0;
2720
2721                 lock_sock(sk);
2722                 err = tcp_set_congestion_control(sk, name, true, true);
2723                 release_sock(sk);
2724                 return err;
2725         }
2726         case TCP_ULP: {
2727                 char name[TCP_ULP_NAME_MAX];
2728
2729                 if (optlen < 1)
2730                         return -EINVAL;
2731
2732                 val = strncpy_from_user(name, optval,
2733                                         min_t(long, TCP_ULP_NAME_MAX - 1,
2734                                               optlen));
2735                 if (val < 0)
2736                         return -EFAULT;
2737                 name[val] = 0;
2738
2739                 lock_sock(sk);
2740                 err = tcp_set_ulp(sk, name);
2741                 release_sock(sk);
2742                 return err;
2743         }
2744         case TCP_FASTOPEN_KEY: {
2745                 __u8 key[TCP_FASTOPEN_KEY_LENGTH];
2746
2747                 if (optlen != sizeof(key))
2748                         return -EINVAL;
2749
2750                 if (copy_from_user(key, optval, optlen))
2751                         return -EFAULT;
2752
2753                 return tcp_fastopen_reset_cipher(net, sk, key, sizeof(key));
2754         }
2755         default:
2756                 /* fallthru */
2757                 break;
2758         }
2759
2760         if (optlen < sizeof(int))
2761                 return -EINVAL;
2762
2763         if (get_user(val, (int __user *)optval))
2764                 return -EFAULT;
2765
2766         lock_sock(sk);
2767
2768         switch (optname) {
2769         case TCP_MAXSEG:
2770                 /* Values greater than interface MTU won't take effect. However
2771                  * at the point when this call is done we typically don't yet
2772                  * know which interface is going to be used
2773                  */
2774                 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
2775                         err = -EINVAL;
2776                         break;
2777                 }
2778                 tp->rx_opt.user_mss = val;
2779                 break;
2780
2781         case TCP_NODELAY:
2782                 if (val) {
2783                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2784                          * this option on corked socket is remembered, but
2785                          * it is not activated until cork is cleared.
2786                          *
2787                          * However, when TCP_NODELAY is set we make
2788                          * an explicit push, which overrides even TCP_CORK
2789                          * for currently queued segments.
2790                          */
2791                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2792                         tcp_push_pending_frames(sk);
2793                 } else {
2794                         tp->nonagle &= ~TCP_NAGLE_OFF;
2795                 }
2796                 break;
2797
2798         case TCP_THIN_LINEAR_TIMEOUTS:
2799                 if (val < 0 || val > 1)
2800                         err = -EINVAL;
2801                 else
2802                         tp->thin_lto = val;
2803                 break;
2804
2805         case TCP_THIN_DUPACK:
2806                 if (val < 0 || val > 1)
2807                         err = -EINVAL;
2808                 break;
2809
2810         case TCP_REPAIR:
2811                 if (!tcp_can_repair_sock(sk))
2812                         err = -EPERM;
2813                 else if (val == 1) {
2814                         tp->repair = 1;
2815                         sk->sk_reuse = SK_FORCE_REUSE;
2816                         tp->repair_queue = TCP_NO_QUEUE;
2817                 } else if (val == 0) {
2818                         tp->repair = 0;
2819                         sk->sk_reuse = SK_NO_REUSE;
2820                         tcp_send_window_probe(sk);
2821                 } else
2822                         err = -EINVAL;
2823
2824                 break;
2825
2826         case TCP_REPAIR_QUEUE:
2827                 if (!tp->repair)
2828                         err = -EPERM;
2829                 else if ((unsigned int)val < TCP_QUEUES_NR)
2830                         tp->repair_queue = val;
2831                 else
2832                         err = -EINVAL;
2833                 break;
2834
2835         case TCP_QUEUE_SEQ:
2836                 if (sk->sk_state != TCP_CLOSE)
2837                         err = -EPERM;
2838                 else if (tp->repair_queue == TCP_SEND_QUEUE)
2839                         tp->write_seq = val;
2840                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2841                         tp->rcv_nxt = val;
2842                 else
2843                         err = -EINVAL;
2844                 break;
2845
2846         case TCP_REPAIR_OPTIONS:
2847                 if (!tp->repair)
2848                         err = -EINVAL;
2849                 else if (sk->sk_state == TCP_ESTABLISHED)
2850                         err = tcp_repair_options_est(sk,
2851                                         (struct tcp_repair_opt __user *)optval,
2852                                         optlen);
2853                 else
2854                         err = -EPERM;
2855                 break;
2856
2857         case TCP_CORK:
2858                 /* When set indicates to always queue non-full frames.
2859                  * Later the user clears this option and we transmit
2860                  * any pending partial frames in the queue.  This is
2861                  * meant to be used alongside sendfile() to get properly
2862                  * filled frames when the user (for example) must write
2863                  * out headers with a write() call first and then use
2864                  * sendfile to send out the data parts.
2865                  *
2866                  * TCP_CORK can be set together with TCP_NODELAY and it is
2867                  * stronger than TCP_NODELAY.
2868                  */
2869                 if (val) {
2870                         tp->nonagle |= TCP_NAGLE_CORK;
2871                 } else {
2872                         tp->nonagle &= ~TCP_NAGLE_CORK;
2873                         if (tp->nonagle&TCP_NAGLE_OFF)
2874                                 tp->nonagle |= TCP_NAGLE_PUSH;
2875                         tcp_push_pending_frames(sk);
2876                 }
2877                 break;
2878
2879         case TCP_KEEPIDLE:
2880                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2881                         err = -EINVAL;
2882                 else {
2883                         tp->keepalive_time = val * HZ;
2884                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2885                             !((1 << sk->sk_state) &
2886                               (TCPF_CLOSE | TCPF_LISTEN))) {
2887                                 u32 elapsed = keepalive_time_elapsed(tp);
2888                                 if (tp->keepalive_time > elapsed)
2889                                         elapsed = tp->keepalive_time - elapsed;
2890                                 else
2891                                         elapsed = 0;
2892                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2893                         }
2894                 }
2895                 break;
2896         case TCP_KEEPINTVL:
2897                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2898                         err = -EINVAL;
2899                 else
2900                         tp->keepalive_intvl = val * HZ;
2901                 break;
2902         case TCP_KEEPCNT:
2903                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2904                         err = -EINVAL;
2905                 else
2906                         tp->keepalive_probes = val;
2907                 break;
2908         case TCP_SYNCNT:
2909                 if (val < 1 || val > MAX_TCP_SYNCNT)
2910                         err = -EINVAL;
2911                 else
2912                         icsk->icsk_syn_retries = val;
2913                 break;
2914
2915         case TCP_SAVE_SYN:
2916                 if (val < 0 || val > 1)
2917                         err = -EINVAL;
2918                 else
2919                         tp->save_syn = val;
2920                 break;
2921
2922         case TCP_LINGER2:
2923                 if (val < 0)
2924                         tp->linger2 = -1;
2925                 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
2926                         tp->linger2 = 0;
2927                 else
2928                         tp->linger2 = val * HZ;
2929                 break;
2930
2931         case TCP_DEFER_ACCEPT:
2932                 /* Translate value in seconds to number of retransmits */
2933                 icsk->icsk_accept_queue.rskq_defer_accept =
2934                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2935                                         TCP_RTO_MAX / HZ);
2936                 break;
2937
2938         case TCP_WINDOW_CLAMP:
2939                 if (!val) {
2940                         if (sk->sk_state != TCP_CLOSE) {
2941                                 err = -EINVAL;
2942                                 break;
2943                         }
2944                         tp->window_clamp = 0;
2945                 } else
2946                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2947                                                 SOCK_MIN_RCVBUF / 2 : val;
2948                 break;
2949
2950         case TCP_QUICKACK:
2951                 if (!val) {
2952                         icsk->icsk_ack.pingpong = 1;
2953                 } else {
2954                         icsk->icsk_ack.pingpong = 0;
2955                         if ((1 << sk->sk_state) &
2956                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2957                             inet_csk_ack_scheduled(sk)) {
2958                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2959                                 tcp_cleanup_rbuf(sk, 1);
2960                                 if (!(val & 1))
2961                                         icsk->icsk_ack.pingpong = 1;
2962                         }
2963                 }
2964                 break;
2965
2966 #ifdef CONFIG_TCP_MD5SIG
2967         case TCP_MD5SIG:
2968         case TCP_MD5SIG_EXT:
2969                 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
2970                         err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
2971                 else
2972                         err = -EINVAL;
2973                 break;
2974 #endif
2975         case TCP_USER_TIMEOUT:
2976                 /* Cap the max time in ms TCP will retry or probe the window
2977                  * before giving up and aborting (ETIMEDOUT) a connection.
2978                  */
2979                 if (val < 0)
2980                         err = -EINVAL;
2981                 else
2982                         icsk->icsk_user_timeout = msecs_to_jiffies(val);
2983                 break;
2984
2985         case TCP_FASTOPEN:
2986                 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2987                     TCPF_LISTEN))) {
2988                         tcp_fastopen_init_key_once(net);
2989
2990                         fastopen_queue_tune(sk, val);
2991                 } else {
2992                         err = -EINVAL;
2993                 }
2994                 break;
2995         case TCP_FASTOPEN_CONNECT:
2996                 if (val > 1 || val < 0) {
2997                         err = -EINVAL;
2998                 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
2999                         if (sk->sk_state == TCP_CLOSE)
3000                                 tp->fastopen_connect = val;
3001                         else
3002                                 err = -EINVAL;
3003                 } else {
3004                         err = -EOPNOTSUPP;
3005                 }
3006                 break;
3007         case TCP_FASTOPEN_NO_COOKIE:
3008                 if (val > 1 || val < 0)
3009                         err = -EINVAL;
3010                 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3011                         err = -EINVAL;
3012                 else
3013                         tp->fastopen_no_cookie = val;
3014                 break;
3015         case TCP_TIMESTAMP:
3016                 if (!tp->repair)
3017                         err = -EPERM;
3018                 else
3019                         tp->tsoffset = val - tcp_time_stamp_raw();
3020                 break;
3021         case TCP_REPAIR_WINDOW:
3022                 err = tcp_repair_set_window(tp, optval, optlen);
3023                 break;
3024         case TCP_NOTSENT_LOWAT:
3025                 tp->notsent_lowat = val;
3026                 sk->sk_write_space(sk);
3027                 break;
3028         case TCP_INQ:
3029                 if (val > 1 || val < 0)
3030                         err = -EINVAL;
3031                 else
3032                         tp->recvmsg_inq = val;
3033                 break;
3034         default:
3035                 err = -ENOPROTOOPT;
3036                 break;
3037         }
3038
3039         release_sock(sk);
3040         return err;
3041 }
3042
3043 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
3044                    unsigned int optlen)
3045 {
3046         const struct inet_connection_sock *icsk = inet_csk(sk);
3047
3048         if (level != SOL_TCP)
3049                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3050                                                      optval, optlen);
3051         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3052 }
3053 EXPORT_SYMBOL(tcp_setsockopt);
3054
3055 #ifdef CONFIG_COMPAT
3056 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
3057                           char __user *optval, unsigned int optlen)
3058 {
3059         if (level != SOL_TCP)
3060                 return inet_csk_compat_setsockopt(sk, level, optname,
3061                                                   optval, optlen);
3062         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3063 }
3064 EXPORT_SYMBOL(compat_tcp_setsockopt);
3065 #endif
3066
3067 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3068                                       struct tcp_info *info)
3069 {
3070         u64 stats[__TCP_CHRONO_MAX], total = 0;
3071         enum tcp_chrono i;
3072
3073         for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3074                 stats[i] = tp->chrono_stat[i - 1];
3075                 if (i == tp->chrono_type)
3076                         stats[i] += tcp_jiffies32 - tp->chrono_start;
3077                 stats[i] *= USEC_PER_SEC / HZ;
3078                 total += stats[i];
3079         }
3080
3081         info->tcpi_busy_time = total;
3082         info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3083         info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3084 }
3085
3086 /* Return information about state of tcp endpoint in API format. */
3087 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3088 {
3089         const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3090         const struct inet_connection_sock *icsk = inet_csk(sk);
3091         u32 now;
3092         u64 rate64;
3093         bool slow;
3094         u32 rate;
3095
3096         memset(info, 0, sizeof(*info));
3097         if (sk->sk_type != SOCK_STREAM)
3098                 return;
3099
3100         info->tcpi_state = inet_sk_state_load(sk);
3101
3102         /* Report meaningful fields for all TCP states, including listeners */
3103         rate = READ_ONCE(sk->sk_pacing_rate);
3104         rate64 = rate != ~0U ? rate : ~0ULL;
3105         info->tcpi_pacing_rate = rate64;
3106
3107         rate = READ_ONCE(sk->sk_max_pacing_rate);
3108         rate64 = rate != ~0U ? rate : ~0ULL;
3109         info->tcpi_max_pacing_rate = rate64;
3110
3111         info->tcpi_reordering = tp->reordering;
3112         info->tcpi_snd_cwnd = tp->snd_cwnd;
3113
3114         if (info->tcpi_state == TCP_LISTEN) {
3115                 /* listeners aliased fields :
3116                  * tcpi_unacked -> Number of children ready for accept()
3117                  * tcpi_sacked  -> max backlog
3118                  */
3119                 info->tcpi_unacked = sk->sk_ack_backlog;
3120                 info->tcpi_sacked = sk->sk_max_ack_backlog;
3121                 return;
3122         }
3123
3124         slow = lock_sock_fast(sk);
3125
3126         info->tcpi_ca_state = icsk->icsk_ca_state;
3127         info->tcpi_retransmits = icsk->icsk_retransmits;
3128         info->tcpi_probes = icsk->icsk_probes_out;
3129         info->tcpi_backoff = icsk->icsk_backoff;
3130
3131         if (tp->rx_opt.tstamp_ok)
3132                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3133         if (tcp_is_sack(tp))
3134                 info->tcpi_options |= TCPI_OPT_SACK;
3135         if (tp->rx_opt.wscale_ok) {
3136                 info->tcpi_options |= TCPI_OPT_WSCALE;
3137                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3138                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3139         }
3140
3141         if (tp->ecn_flags & TCP_ECN_OK)
3142                 info->tcpi_options |= TCPI_OPT_ECN;
3143         if (tp->ecn_flags & TCP_ECN_SEEN)
3144                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3145         if (tp->syn_data_acked)
3146                 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3147
3148         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3149         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3150         info->tcpi_snd_mss = tp->mss_cache;
3151         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3152
3153         info->tcpi_unacked = tp->packets_out;
3154         info->tcpi_sacked = tp->sacked_out;
3155
3156         info->tcpi_lost = tp->lost_out;
3157         info->tcpi_retrans = tp->retrans_out;
3158
3159         now = tcp_jiffies32;
3160         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3161         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3162         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3163
3164         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3165         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3166         info->tcpi_rtt = tp->srtt_us >> 3;
3167         info->tcpi_rttvar = tp->mdev_us >> 2;
3168         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3169         info->tcpi_advmss = tp->advmss;
3170
3171         info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3172         info->tcpi_rcv_space = tp->rcvq_space.space;
3173
3174         info->tcpi_total_retrans = tp->total_retrans;
3175
3176         info->tcpi_bytes_acked = tp->bytes_acked;
3177         info->tcpi_bytes_received = tp->bytes_received;
3178         info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3179         tcp_get_info_chrono_stats(tp, info);
3180
3181         info->tcpi_segs_out = tp->segs_out;
3182         info->tcpi_segs_in = tp->segs_in;
3183
3184         info->tcpi_min_rtt = tcp_min_rtt(tp);
3185         info->tcpi_data_segs_in = tp->data_segs_in;
3186         info->tcpi_data_segs_out = tp->data_segs_out;
3187
3188         info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3189         rate64 = tcp_compute_delivery_rate(tp);
3190         if (rate64)
3191                 info->tcpi_delivery_rate = rate64;
3192         info->tcpi_delivered = tp->delivered;
3193         info->tcpi_delivered_ce = tp->delivered_ce;
3194         unlock_sock_fast(sk, slow);
3195 }
3196 EXPORT_SYMBOL_GPL(tcp_get_info);
3197
3198 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk)
3199 {
3200         const struct tcp_sock *tp = tcp_sk(sk);
3201         struct sk_buff *stats;
3202         struct tcp_info info;
3203         u64 rate64;
3204         u32 rate;
3205
3206         stats = alloc_skb(7 * nla_total_size_64bit(sizeof(u64)) +
3207                           7 * nla_total_size(sizeof(u32)) +
3208                           3 * nla_total_size(sizeof(u8)), GFP_ATOMIC);
3209         if (!stats)
3210                 return NULL;
3211
3212         tcp_get_info_chrono_stats(tp, &info);
3213         nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3214                           info.tcpi_busy_time, TCP_NLA_PAD);
3215         nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3216                           info.tcpi_rwnd_limited, TCP_NLA_PAD);
3217         nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3218                           info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3219         nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3220                           tp->data_segs_out, TCP_NLA_PAD);
3221         nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3222                           tp->total_retrans, TCP_NLA_PAD);
3223
3224         rate = READ_ONCE(sk->sk_pacing_rate);
3225         rate64 = rate != ~0U ? rate : ~0ULL;
3226         nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3227
3228         rate64 = tcp_compute_delivery_rate(tp);
3229         nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3230
3231         nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3232         nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3233         nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3234
3235         nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3236         nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3237         nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3238         nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3239         nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3240
3241         nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3242         nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3243
3244         return stats;
3245 }
3246
3247 static int do_tcp_getsockopt(struct sock *sk, int level,
3248                 int optname, char __user *optval, int __user *optlen)
3249 {
3250         struct inet_connection_sock *icsk = inet_csk(sk);
3251         struct tcp_sock *tp = tcp_sk(sk);
3252         struct net *net = sock_net(sk);
3253         int val, len;
3254
3255         if (get_user(len, optlen))
3256                 return -EFAULT;
3257
3258         len = min_t(unsigned int, len, sizeof(int));
3259
3260         if (len < 0)
3261                 return -EINVAL;
3262
3263         switch (optname) {
3264         case TCP_MAXSEG:
3265                 val = tp->mss_cache;
3266                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3267                         val = tp->rx_opt.user_mss;
3268                 if (tp->repair)
3269                         val = tp->rx_opt.mss_clamp;
3270                 break;
3271         case TCP_NODELAY:
3272                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3273                 break;
3274         case TCP_CORK:
3275                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3276                 break;
3277         case TCP_KEEPIDLE:
3278                 val = keepalive_time_when(tp) / HZ;
3279                 break;
3280         case TCP_KEEPINTVL:
3281                 val = keepalive_intvl_when(tp) / HZ;
3282                 break;
3283         case TCP_KEEPCNT:
3284                 val = keepalive_probes(tp);
3285                 break;
3286         case TCP_SYNCNT:
3287                 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3288                 break;
3289         case TCP_LINGER2:
3290                 val = tp->linger2;
3291                 if (val >= 0)
3292                         val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3293                 break;
3294         case TCP_DEFER_ACCEPT:
3295                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3296                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3297                 break;
3298         case TCP_WINDOW_CLAMP:
3299                 val = tp->window_clamp;
3300                 break;
3301         case TCP_INFO: {
3302                 struct tcp_info info;
3303
3304                 if (get_user(len, optlen))
3305                         return -EFAULT;
3306
3307                 tcp_get_info(sk, &info);
3308
3309                 len = min_t(unsigned int, len, sizeof(info));
3310                 if (put_user(len, optlen))
3311                         return -EFAULT;
3312                 if (copy_to_user(optval, &info, len))
3313                         return -EFAULT;
3314                 return 0;
3315         }
3316         case TCP_CC_INFO: {
3317                 const struct tcp_congestion_ops *ca_ops;
3318                 union tcp_cc_info info;
3319                 size_t sz = 0;
3320                 int attr;
3321
3322                 if (get_user(len, optlen))
3323                         return -EFAULT;
3324
3325                 ca_ops = icsk->icsk_ca_ops;
3326                 if (ca_ops && ca_ops->get_info)
3327                         sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3328
3329                 len = min_t(unsigned int, len, sz);
3330                 if (put_user(len, optlen))
3331                         return -EFAULT;
3332                 if (copy_to_user(optval, &info, len))
3333                         return -EFAULT;
3334                 return 0;
3335         }
3336         case TCP_QUICKACK:
3337                 val = !icsk->icsk_ack.pingpong;
3338                 break;
3339
3340         case TCP_CONGESTION:
3341                 if (get_user(len, optlen))
3342                         return -EFAULT;
3343                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
3344                 if (put_user(len, optlen))
3345                         return -EFAULT;
3346                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3347                         return -EFAULT;
3348                 return 0;
3349
3350         case TCP_ULP:
3351                 if (get_user(len, optlen))
3352                         return -EFAULT;
3353                 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
3354                 if (!icsk->icsk_ulp_ops) {
3355                         if (put_user(0, optlen))
3356                                 return -EFAULT;
3357                         return 0;
3358                 }
3359                 if (put_user(len, optlen))
3360                         return -EFAULT;
3361                 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
3362                         return -EFAULT;
3363                 return 0;
3364
3365         case TCP_FASTOPEN_KEY: {
3366                 __u8 key[TCP_FASTOPEN_KEY_LENGTH];
3367                 struct tcp_fastopen_context *ctx;
3368
3369                 if (get_user(len, optlen))
3370                         return -EFAULT;
3371
3372                 rcu_read_lock();
3373                 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
3374                 if (ctx)
3375                         memcpy(key, ctx->key, sizeof(key));
3376                 else
3377                         len = 0;
3378                 rcu_read_unlock();
3379
3380                 len = min_t(unsigned int, len, sizeof(key));
3381                 if (put_user(len, optlen))
3382                         return -EFAULT;
3383                 if (copy_to_user(optval, key, len))
3384                         return -EFAULT;
3385                 return 0;
3386         }
3387         case TCP_THIN_LINEAR_TIMEOUTS:
3388                 val = tp->thin_lto;
3389                 break;
3390
3391         case TCP_THIN_DUPACK:
3392                 val = 0;
3393                 break;
3394
3395         case TCP_REPAIR:
3396                 val = tp->repair;
3397                 break;
3398
3399         case TCP_REPAIR_QUEUE:
3400                 if (tp->repair)
3401                         val = tp->repair_queue;
3402                 else
3403                         return -EINVAL;
3404                 break;
3405
3406         case TCP_REPAIR_WINDOW: {
3407                 struct tcp_repair_window opt;
3408
3409                 if (get_user(len, optlen))
3410                         return -EFAULT;
3411
3412                 if (len != sizeof(opt))
3413                         return -EINVAL;
3414
3415                 if (!tp->repair)
3416                         return -EPERM;
3417
3418                 opt.snd_wl1     = tp->snd_wl1;
3419                 opt.snd_wnd     = tp->snd_wnd;
3420                 opt.max_window  = tp->max_window;
3421                 opt.rcv_wnd     = tp->rcv_wnd;
3422                 opt.rcv_wup     = tp->rcv_wup;
3423
3424                 if (copy_to_user(optval, &opt, len))
3425                         return -EFAULT;
3426                 return 0;
3427         }
3428         case TCP_QUEUE_SEQ:
3429                 if (tp->repair_queue == TCP_SEND_QUEUE)
3430                         val = tp->write_seq;
3431                 else if (tp->repair_queue == TCP_RECV_QUEUE)
3432                         val = tp->rcv_nxt;
3433                 else
3434                         return -EINVAL;
3435                 break;
3436
3437         case TCP_USER_TIMEOUT:
3438                 val = jiffies_to_msecs(icsk->icsk_user_timeout);
3439                 break;
3440
3441         case TCP_FASTOPEN:
3442                 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3443                 break;
3444
3445         case TCP_FASTOPEN_CONNECT:
3446                 val = tp->fastopen_connect;
3447                 break;
3448
3449         case TCP_FASTOPEN_NO_COOKIE:
3450                 val = tp->fastopen_no_cookie;
3451                 break;
3452
3453         case TCP_TIMESTAMP:
3454                 val = tcp_time_stamp_raw() + tp->tsoffset;
3455                 break;
3456         case TCP_NOTSENT_LOWAT:
3457                 val = tp->notsent_lowat;
3458                 break;
3459         case TCP_INQ:
3460                 val = tp->recvmsg_inq;
3461                 break;
3462         case TCP_SAVE_SYN:
3463                 val = tp->save_syn;
3464                 break;
3465         case TCP_SAVED_SYN: {
3466                 if (get_user(len, optlen))
3467                         return -EFAULT;
3468
3469                 lock_sock(sk);
3470                 if (tp->saved_syn) {
3471                         if (len < tp->saved_syn[0]) {
3472                                 if (put_user(tp->saved_syn[0], optlen)) {
3473                                         release_sock(sk);
3474                                         return -EFAULT;
3475                                 }
3476                                 release_sock(sk);
3477                                 return -EINVAL;
3478                         }
3479                         len = tp->saved_syn[0];
3480                         if (put_user(len, optlen)) {
3481                                 release_sock(sk);
3482                                 return -EFAULT;
3483                         }
3484                         if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3485                                 release_sock(sk);
3486                                 return -EFAULT;
3487                         }
3488                         tcp_saved_syn_free(tp);
3489                         release_sock(sk);
3490                 } else {
3491                         release_sock(sk);
3492                         len = 0;
3493                         if (put_user(len, optlen))
3494                                 return -EFAULT;
3495                 }
3496                 return 0;
3497         }
3498 #ifdef CONFIG_MMU
3499         case TCP_ZEROCOPY_RECEIVE: {
3500                 struct tcp_zerocopy_receive zc;
3501                 int err;
3502
3503                 if (get_user(len, optlen))
3504                         return -EFAULT;
3505                 if (len != sizeof(zc))
3506                         return -EINVAL;
3507                 if (copy_from_user(&zc, optval, len))
3508                         return -EFAULT;
3509                 lock_sock(sk);
3510                 err = tcp_zerocopy_receive(sk, &zc);
3511                 release_sock(sk);
3512                 if (!err && copy_to_user(optval, &zc, len))
3513                         err = -EFAULT;
3514                 return err;
3515         }
3516 #endif
3517         default:
3518                 return -ENOPROTOOPT;
3519         }
3520
3521         if (put_user(len, optlen))
3522                 return -EFAULT;
3523         if (copy_to_user(optval, &val, len))
3524                 return -EFAULT;
3525         return 0;
3526 }
3527
3528 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3529                    int __user *optlen)
3530 {
3531         struct inet_connection_sock *icsk = inet_csk(sk);
3532
3533         if (level != SOL_TCP)
3534                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3535                                                      optval, optlen);
3536         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3537 }
3538 EXPORT_SYMBOL(tcp_getsockopt);
3539
3540 #ifdef CONFIG_COMPAT
3541 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3542                           char __user *optval, int __user *optlen)
3543 {
3544         if (level != SOL_TCP)
3545                 return inet_csk_compat_getsockopt(sk, level, optname,
3546                                                   optval, optlen);
3547         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3548 }
3549 EXPORT_SYMBOL(compat_tcp_getsockopt);
3550 #endif
3551
3552 #ifdef CONFIG_TCP_MD5SIG
3553 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3554 static DEFINE_MUTEX(tcp_md5sig_mutex);
3555 static bool tcp_md5sig_pool_populated = false;
3556
3557 static void __tcp_alloc_md5sig_pool(void)
3558 {
3559         struct crypto_ahash *hash;
3560         int cpu;
3561
3562         hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3563         if (IS_ERR(hash))
3564                 return;
3565
3566         for_each_possible_cpu(cpu) {
3567                 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3568                 struct ahash_request *req;
3569
3570                 if (!scratch) {
3571                         scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3572                                                sizeof(struct tcphdr),
3573                                                GFP_KERNEL,
3574                                                cpu_to_node(cpu));
3575                         if (!scratch)
3576                                 return;
3577                         per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3578                 }
3579                 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3580                         continue;
3581
3582                 req = ahash_request_alloc(hash, GFP_KERNEL);
3583                 if (!req)
3584                         return;
3585
3586                 ahash_request_set_callback(req, 0, NULL, NULL);
3587
3588                 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3589         }
3590         /* before setting tcp_md5sig_pool_populated, we must commit all writes
3591          * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3592          */
3593         smp_wmb();
3594         tcp_md5sig_pool_populated = true;
3595 }
3596
3597 bool tcp_alloc_md5sig_pool(void)
3598 {
3599         if (unlikely(!tcp_md5sig_pool_populated)) {
3600                 mutex_lock(&tcp_md5sig_mutex);
3601
3602                 if (!tcp_md5sig_pool_populated)
3603                         __tcp_alloc_md5sig_pool();
3604
3605                 mutex_unlock(&tcp_md5sig_mutex);
3606         }
3607         return tcp_md5sig_pool_populated;
3608 }
3609 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3610
3611
3612 /**
3613  *      tcp_get_md5sig_pool - get md5sig_pool for this user
3614  *
3615  *      We use percpu structure, so if we succeed, we exit with preemption
3616  *      and BH disabled, to make sure another thread or softirq handling
3617  *      wont try to get same context.
3618  */
3619 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3620 {
3621         local_bh_disable();
3622
3623         if (tcp_md5sig_pool_populated) {
3624                 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3625                 smp_rmb();
3626                 return this_cpu_ptr(&tcp_md5sig_pool);
3627         }
3628         local_bh_enable();
3629         return NULL;
3630 }
3631 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3632
3633 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3634                           const struct sk_buff *skb, unsigned int header_len)
3635 {
3636         struct scatterlist sg;
3637         const struct tcphdr *tp = tcp_hdr(skb);
3638         struct ahash_request *req = hp->md5_req;
3639         unsigned int i;
3640         const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3641                                            skb_headlen(skb) - header_len : 0;
3642         const struct skb_shared_info *shi = skb_shinfo(skb);
3643         struct sk_buff *frag_iter;
3644
3645         sg_init_table(&sg, 1);
3646
3647         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3648         ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3649         if (crypto_ahash_update(req))
3650                 return 1;
3651
3652         for (i = 0; i < shi->nr_frags; ++i) {
3653                 const struct skb_frag_struct *f = &shi->frags[i];
3654                 unsigned int offset = f->page_offset;
3655                 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3656
3657                 sg_set_page(&sg, page, skb_frag_size(f),
3658                             offset_in_page(offset));
3659                 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3660                 if (crypto_ahash_update(req))
3661                         return 1;
3662         }
3663
3664         skb_walk_frags(skb, frag_iter)
3665                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3666                         return 1;
3667
3668         return 0;
3669 }
3670 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3671
3672 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3673 {
3674         struct scatterlist sg;
3675
3676         sg_init_one(&sg, key->key, key->keylen);
3677         ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen);
3678         return crypto_ahash_update(hp->md5_req);
3679 }
3680 EXPORT_SYMBOL(tcp_md5_hash_key);
3681
3682 #endif
3683
3684 void tcp_done(struct sock *sk)
3685 {
3686         struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3687
3688         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3689                 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3690
3691         tcp_set_state(sk, TCP_CLOSE);
3692         tcp_clear_xmit_timers(sk);
3693         if (req)
3694                 reqsk_fastopen_remove(sk, req, false);
3695
3696         sk->sk_shutdown = SHUTDOWN_MASK;
3697
3698         if (!sock_flag(sk, SOCK_DEAD))
3699                 sk->sk_state_change(sk);
3700         else
3701                 inet_csk_destroy_sock(sk);
3702 }
3703 EXPORT_SYMBOL_GPL(tcp_done);
3704
3705 int tcp_abort(struct sock *sk, int err)
3706 {
3707         if (!sk_fullsock(sk)) {
3708                 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3709                         struct request_sock *req = inet_reqsk(sk);
3710
3711                         local_bh_disable();
3712                         inet_csk_reqsk_queue_drop_and_put(req->rsk_listener,
3713                                                           req);
3714                         local_bh_enable();
3715                         return 0;
3716                 }
3717                 return -EOPNOTSUPP;
3718         }
3719
3720         /* Don't race with userspace socket closes such as tcp_close. */
3721         lock_sock(sk);
3722
3723         if (sk->sk_state == TCP_LISTEN) {
3724                 tcp_set_state(sk, TCP_CLOSE);
3725                 inet_csk_listen_stop(sk);
3726         }
3727
3728         /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3729         local_bh_disable();
3730         bh_lock_sock(sk);
3731
3732         if (!sock_flag(sk, SOCK_DEAD)) {
3733                 sk->sk_err = err;
3734                 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3735                 smp_wmb();
3736                 sk->sk_error_report(sk);
3737                 if (tcp_need_reset(sk->sk_state))
3738                         tcp_send_active_reset(sk, GFP_ATOMIC);
3739                 tcp_done(sk);
3740         }
3741
3742         bh_unlock_sock(sk);
3743         local_bh_enable();
3744         tcp_write_queue_purge(sk);
3745         release_sock(sk);
3746         return 0;
3747 }
3748 EXPORT_SYMBOL_GPL(tcp_abort);
3749
3750 extern struct tcp_congestion_ops tcp_reno;
3751
3752 static __initdata unsigned long thash_entries;
3753 static int __init set_thash_entries(char *str)
3754 {
3755         ssize_t ret;
3756
3757         if (!str)
3758                 return 0;
3759
3760         ret = kstrtoul(str, 0, &thash_entries);
3761         if (ret)
3762                 return 0;
3763
3764         return 1;
3765 }
3766 __setup("thash_entries=", set_thash_entries);
3767
3768 static void __init tcp_init_mem(void)
3769 {
3770         unsigned long limit = nr_free_buffer_pages() / 16;
3771
3772         limit = max(limit, 128UL);
3773         sysctl_tcp_mem[0] = limit / 4 * 3;              /* 4.68 % */
3774         sysctl_tcp_mem[1] = limit;                      /* 6.25 % */
3775         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;      /* 9.37 % */
3776 }
3777
3778 void __init tcp_init(void)
3779 {
3780         int max_rshare, max_wshare, cnt;
3781         unsigned long limit;
3782         unsigned int i;
3783
3784         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3785                      FIELD_SIZEOF(struct sk_buff, cb));
3786
3787         percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3788         percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3789         inet_hashinfo_init(&tcp_hashinfo);
3790         inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
3791                             thash_entries, 21,  /* one slot per 2 MB*/
3792                             0, 64 * 1024);
3793         tcp_hashinfo.bind_bucket_cachep =
3794                 kmem_cache_create("tcp_bind_bucket",
3795                                   sizeof(struct inet_bind_bucket), 0,
3796                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3797
3798         /* Size and allocate the main established and bind bucket
3799          * hash tables.
3800          *
3801          * The methodology is similar to that of the buffer cache.
3802          */
3803         tcp_hashinfo.ehash =
3804                 alloc_large_system_hash("TCP established",
3805                                         sizeof(struct inet_ehash_bucket),
3806                                         thash_entries,
3807                                         17, /* one slot per 128 KB of memory */
3808                                         0,
3809                                         NULL,
3810                                         &tcp_hashinfo.ehash_mask,
3811                                         0,
3812                                         thash_entries ? 0 : 512 * 1024);
3813         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3814                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3815
3816         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3817                 panic("TCP: failed to alloc ehash_locks");
3818         tcp_hashinfo.bhash =
3819                 alloc_large_system_hash("TCP bind",
3820                                         sizeof(struct inet_bind_hashbucket),
3821                                         tcp_hashinfo.ehash_mask + 1,
3822                                         17, /* one slot per 128 KB of memory */
3823                                         0,
3824                                         &tcp_hashinfo.bhash_size,
3825                                         NULL,
3826                                         0,
3827                                         64 * 1024);
3828         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3829         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3830                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3831                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3832         }
3833
3834
3835         cnt = tcp_hashinfo.ehash_mask + 1;
3836         sysctl_tcp_max_orphans = cnt / 2;
3837
3838         tcp_init_mem();
3839         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3840         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3841         max_wshare = min(4UL*1024*1024, limit);
3842         max_rshare = min(6UL*1024*1024, limit);
3843
3844         init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3845         init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
3846         init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3847
3848         init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3849         init_net.ipv4.sysctl_tcp_rmem[1] = 87380;
3850         init_net.ipv4.sysctl_tcp_rmem[2] = max(87380, max_rshare);
3851
3852         pr_info("Hash tables configured (established %u bind %u)\n",
3853                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3854
3855         tcp_v4_init();
3856         tcp_metrics_init();
3857         BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3858         tcp_tasklet_init();
3859 }