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/memblock.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  *      Wait for a TCP event.
 498  *
 499  *      Note that we don't need to lock the socket, as the upper poll layers
 500  *      take care of normal races (between the test and the event) and we don't
 501  *      go look at any of the socket buffers directly.
 502  */
 503 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 504 {
 505         __poll_t mask;
 506         struct sock *sk = sock->sk;
 507         const struct tcp_sock *tp = tcp_sk(sk);
 508         int state;
 509
 510         sock_poll_wait(file, sock, wait);
 511
 512         state = inet_sk_state_load(sk);
 513         if (state == TCP_LISTEN)
 514                 return inet_csk_listen_poll(sk);
 515
 516         /* Socket is not locked. We are protected from async events
 517          * by poll logic and correct handling of state changes
 518          * made by other threads is impossible in any case.
 519          */
 520
 521         mask = 0;
 522
 523         /*
 524          * EPOLLHUP is certainly not done right. But poll() doesn't
 525          * have a notion of HUP in just one direction, and for a
 526          * socket the read side is more interesting.
 527          *
 528          * Some poll() documentation says that EPOLLHUP is incompatible
 529          * with the EPOLLOUT/POLLWR flags, so somebody should check this
 530          * all. But careful, it tends to be safer to return too many
 531          * bits than too few, and you can easily break real applications
 532          * if you don't tell them that something has hung up!
 533          *
 534          * Check-me.
 535          *
 536          * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
 537          * our fs/select.c). It means that after we received EOF,
 538          * poll always returns immediately, making impossible poll() on write()
 539          * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
 540          * if and only if shutdown has been made in both directions.
 541          * Actually, it is interesting to look how Solaris and DUX
 542          * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
 543          * then we could set it on SND_SHUTDOWN. BTW examples given
 544          * in Stevens' books assume exactly this behaviour, it explains
 545          * why EPOLLHUP is incompatible with EPOLLOUT.  --ANK
 546          *
 547          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 548          * blocking on fresh not-connected or disconnected socket. --ANK
 549          */
 550         if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
 551                 mask |= EPOLLHUP;
 552         if (sk->sk_shutdown & RCV_SHUTDOWN)
 553                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
 554
 555         /* Connected or passive Fast Open socket? */
 556         if (state != TCP_SYN_SENT &&
 557             (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
 558                 int target = sock_rcvlowat(sk, 0, INT_MAX);
 559
 560                 if (tp->urg_seq == tp->copied_seq &&
 561                     !sock_flag(sk, SOCK_URGINLINE) &&
 562                     tp->urg_data)
 563                         target++;
 564
 565                 if (tcp_stream_is_readable(tp, target, sk))
 566                         mask |= EPOLLIN | EPOLLRDNORM;
 567
 568                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 569                         if (sk_stream_is_writeable(sk)) {
 570                                 mask |= EPOLLOUT | EPOLLWRNORM;
 571                         } else {  /* send SIGIO later */
 572                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 573                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 574
 575                                 /* Race breaker. If space is freed after
 576                                  * wspace test but before the flags are set,
 577                                  * IO signal will be lost. Memory barrier
 578                                  * pairs with the input side.
 579                                  */
 580                                 smp_mb__after_atomic();
 581                                 if (sk_stream_is_writeable(sk))
 582                                         mask |= EPOLLOUT | EPOLLWRNORM;
 583                         }
 584                 } else
 585                         mask |= EPOLLOUT | EPOLLWRNORM;
 586
 587                 if (tp->urg_data & TCP_URG_VALID)
 588                         mask |= EPOLLPRI;
 589         } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
 590                 /* Active TCP fastopen socket with defer_connect
 591                  * Return EPOLLOUT so application can call write()
 592                  * in order for kernel to generate SYN+data
 593                  */
 594                 mask |= EPOLLOUT | EPOLLWRNORM;
 595         }
 596         /* This barrier is coupled with smp_wmb() in tcp_reset() */
 597         smp_rmb();
 598         if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
 599                 mask |= EPOLLERR;
 600
 601         return mask;
 602 }
 603 EXPORT_SYMBOL(tcp_poll);
 604
 605 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 606 {
 607         struct tcp_sock *tp = tcp_sk(sk);
 608         int answ;
 609         bool slow;
 610
 611         switch (cmd) {
 612         case SIOCINQ:
 613                 if (sk->sk_state == TCP_LISTEN)
 614                         return -EINVAL;
 615
 616                 slow = lock_sock_fast(sk);
 617                 answ = tcp_inq(sk);
 618                 unlock_sock_fast(sk, slow);
 619                 break;
 620         case SIOCATMARK:
 621                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 622                 break;
 623         case SIOCOUTQ:
 624                 if (sk->sk_state == TCP_LISTEN)
 625                         return -EINVAL;
 626
 627                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 628                         answ = 0;
 629                 else
 630                         answ = tp->write_seq - tp->snd_una;
 631                 break;
 632         case SIOCOUTQNSD:
 633                 if (sk->sk_state == TCP_LISTEN)
 634                         return -EINVAL;
 635
 636                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 637                         answ = 0;
 638                 else
 639                         answ = tp->write_seq - tp->snd_nxt;
 640                 break;
 641         default:
 642                 return -ENOIOCTLCMD;
 643         }
 644
 645         return put_user(answ, (int __user *)arg);
 646 }
 647 EXPORT_SYMBOL(tcp_ioctl);
 648
 649 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 650 {
 651         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 652         tp->pushed_seq = tp->write_seq;
 653 }
 654
 655 static inline bool forced_push(const struct tcp_sock *tp)
 656 {
 657         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 658 }
 659
 660 static void skb_entail(struct sock *sk, struct sk_buff *skb)
 661 {
 662         struct tcp_sock *tp = tcp_sk(sk);
 663         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 664
 665         skb->csum    = 0;
 666         tcb->seq     = tcb->end_seq = tp->write_seq;
 667         tcb->tcp_flags = TCPHDR_ACK;
 668         tcb->sacked  = 0;
 669         __skb_header_release(skb);
 670         tcp_add_write_queue_tail(sk, skb);
 671         sk->sk_wmem_queued += skb->truesize;
 672         sk_mem_charge(sk, skb->truesize);
 673         if (tp->nonagle & TCP_NAGLE_PUSH)
 674                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 675
 676         tcp_slow_start_after_idle_check(sk);
 677 }
 678
 679 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
 680 {
 681         if (flags & MSG_OOB)
 682                 tp->snd_up = tp->write_seq;
 683 }
 684
 685 /* If a not yet filled skb is pushed, do not send it if
 686  * we have data packets in Qdisc or NIC queues :
 687  * Because TX completion will happen shortly, it gives a chance
 688  * to coalesce future sendmsg() payload into this skb, without
 689  * need for a timer, and with no latency trade off.
 690  * As packets containing data payload have a bigger truesize
 691  * than pure acks (dataless) packets, the last checks prevent
 692  * autocorking if we only have an ACK in Qdisc/NIC queues,
 693  * or if TX completion was delayed after we processed ACK packet.
 694  */
 695 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
 696                                 int size_goal)
 697 {
 698         return skb->len < size_goal &&
 699                sock_net(sk)->ipv4.sysctl_tcp_autocorking &&
 700                !tcp_rtx_queue_empty(sk) &&
 701                refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
 702 }
 703
 704 static void tcp_push(struct sock *sk, int flags, int mss_now,
 705                      int nonagle, int size_goal)
 706 {
 707         struct tcp_sock *tp = tcp_sk(sk);
 708         struct sk_buff *skb;
 709
 710         skb = tcp_write_queue_tail(sk);
 711         if (!skb)
 712                 return;
 713         if (!(flags & MSG_MORE) || forced_push(tp))
 714                 tcp_mark_push(tp, skb);
 715
 716         tcp_mark_urg(tp, flags);
 717
 718         if (tcp_should_autocork(sk, skb, size_goal)) {
 719
 720                 /* avoid atomic op if TSQ_THROTTLED bit is already set */
 721                 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
 722                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
 723                         set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
 724                 }
 725                 /* It is possible TX completion already happened
 726                  * before we set TSQ_THROTTLED.
 727                  */
 728                 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
 729                         return;
 730         }
 731
 732         if (flags & MSG_MORE)
 733                 nonagle = TCP_NAGLE_CORK;
 734
 735         __tcp_push_pending_frames(sk, mss_now, nonagle);
 736 }
 737
 738 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 739                                 unsigned int offset, size_t len)
 740 {
 741         struct tcp_splice_state *tss = rd_desc->arg.data;
 742         int ret;
 743
 744         ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
 745                               min(rd_desc->count, len), tss->flags);
 746         if (ret > 0)
 747                 rd_desc->count -= ret;
 748         return ret;
 749 }
 750
 751 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 752 {
 753         /* Store TCP splice context information in read_descriptor_t. */
 754         read_descriptor_t rd_desc = {
 755                 .arg.data = tss,
 756                 .count    = tss->len,
 757         };
 758
 759         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 760 }
 761
 762 /**
 763  *  tcp_splice_read - splice data from TCP socket to a pipe
 764  * @sock:       socket to splice from
 765  * @ppos:       position (not valid)
 766  * @pipe:       pipe to splice to
 767  * @len:        number of bytes to splice
 768  * @flags:      splice modifier flags
 769  *
 770  * Description:
 771  *    Will read pages from given socket and fill them into a pipe.
 772  *
 773  **/
 774 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 775                         struct pipe_inode_info *pipe, size_t len,
 776                         unsigned int flags)
 777 {
 778         struct sock *sk = sock->sk;
 779         struct tcp_splice_state tss = {
 780                 .pipe = pipe,
 781                 .len = len,
 782                 .flags = flags,
 783         };
 784         long timeo;
 785         ssize_t spliced;
 786         int ret;
 787
 788         sock_rps_record_flow(sk);
 789         /*
 790          * We can't seek on a socket input
 791          */
 792         if (unlikely(*ppos))
 793                 return -ESPIPE;
 794
 795         ret = spliced = 0;
 796
 797         lock_sock(sk);
 798
 799         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 800         while (tss.len) {
 801                 ret = __tcp_splice_read(sk, &tss);
 802                 if (ret < 0)
 803                         break;
 804                 else if (!ret) {
 805                         if (spliced)
 806                                 break;
 807                         if (sock_flag(sk, SOCK_DONE))
 808                                 break;
 809                         if (sk->sk_err) {
 810                                 ret = sock_error(sk);
 811                                 break;
 812                         }
 813                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 814                                 break;
 815                         if (sk->sk_state == TCP_CLOSE) {
 816                                 /*
 817                                  * This occurs when user tries to read
 818                                  * from never connected socket.
 819                                  */
 820                                 ret = -ENOTCONN;
 821                                 break;
 822                         }
 823                         if (!timeo) {
 824                                 ret = -EAGAIN;
 825                                 break;
 826                         }
 827                         /* if __tcp_splice_read() got nothing while we have
 828                          * an skb in receive queue, we do not want to loop.
 829                          * This might happen with URG data.
 830                          */
 831                         if (!skb_queue_empty(&sk->sk_receive_queue))
 832                                 break;
 833                         sk_wait_data(sk, &timeo, NULL);
 834                         if (signal_pending(current)) {
 835                                 ret = sock_intr_errno(timeo);
 836                                 break;
 837                         }
 838                         continue;
 839                 }
 840                 tss.len -= ret;
 841                 spliced += ret;
 842
 843                 if (!timeo)
 844                         break;
 845                 release_sock(sk);
 846                 lock_sock(sk);
 847
 848                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 849                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 850                     signal_pending(current))
 851                         break;
 852         }
 853
 854         release_sock(sk);
 855
 856         if (spliced)
 857                 return spliced;
 858
 859         return ret;
 860 }
 861 EXPORT_SYMBOL(tcp_splice_read);
 862
 863 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
 864                                     bool force_schedule)
 865 {
 866         struct sk_buff *skb;
 867
 868         /* The TCP header must be at least 32-bit aligned.  */
 869         size = ALIGN(size, 4);
 870
 871         if (unlikely(tcp_under_memory_pressure(sk)))
 872                 sk_mem_reclaim_partial(sk);
 873
 874         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 875         if (likely(skb)) {
 876                 bool mem_scheduled;
 877
 878                 if (force_schedule) {
 879                         mem_scheduled = true;
 880                         sk_forced_mem_schedule(sk, skb->truesize);
 881                 } else {
 882                         mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
 883                 }
 884                 if (likely(mem_scheduled)) {
 885                         skb_reserve(skb, sk->sk_prot->max_header);
 886                         /*
 887                          * Make sure that we have exactly size bytes
 888                          * available to the caller, no more, no less.
 889                          */
 890                         skb->reserved_tailroom = skb->end - skb->tail - size;
 891                         INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
 892                         return skb;
 893                 }
 894                 __kfree_skb(skb);
 895         } else {
 896                 sk->sk_prot->enter_memory_pressure(sk);
 897                 sk_stream_moderate_sndbuf(sk);
 898         }
 899         return NULL;
 900 }
 901
 902 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 903                                        int large_allowed)
 904 {
 905         struct tcp_sock *tp = tcp_sk(sk);
 906         u32 new_size_goal, size_goal;
 907
 908         if (!large_allowed)
 909                 return mss_now;
 910
 911         /* Note : tcp_tso_autosize() will eventually split this later */
 912         new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
 913         new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
 914
 915         /* We try hard to avoid divides here */
 916         size_goal = tp->gso_segs * mss_now;
 917         if (unlikely(new_size_goal < size_goal ||
 918                      new_size_goal >= size_goal + mss_now)) {
 919                 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
 920                                      sk->sk_gso_max_segs);
 921                 size_goal = tp->gso_segs * mss_now;
 922         }
 923
 924         return max(size_goal, mss_now);
 925 }
 926
 927 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 928 {
 929         int mss_now;
 930
 931         mss_now = tcp_current_mss(sk);
 932         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 933
 934         return mss_now;
 935 }
 936
 937 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
 938                          size_t size, int flags)
 939 {
 940         struct tcp_sock *tp = tcp_sk(sk);
 941         int mss_now, size_goal;
 942         int err;
 943         ssize_t copied;
 944         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 945
 946         /* Wait for a connection to finish. One exception is TCP Fast Open
 947          * (passive side) where data is allowed to be sent before a connection
 948          * is fully established.
 949          */
 950         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
 951             !tcp_passive_fastopen(sk)) {
 952                 err = sk_stream_wait_connect(sk, &timeo);
 953                 if (err != 0)
 954                         goto out_err;
 955         }
 956
 957         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 958
 959         mss_now = tcp_send_mss(sk, &size_goal, flags);
 960         copied = 0;
 961
 962         err = -EPIPE;
 963         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 964                 goto out_err;
 965
 966         while (size > 0) {
 967                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 968                 int copy, i;
 969                 bool can_coalesce;
 970
 971                 if (!skb || (copy = size_goal - skb->len) <= 0 ||
 972                     !tcp_skb_can_collapse_to(skb)) {
 973 new_segment:
 974                         if (!sk_stream_memory_free(sk))
 975                                 goto wait_for_sndbuf;
 976
 977                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
 978                                         tcp_rtx_and_write_queues_empty(sk));
 979                         if (!skb)
 980                                 goto wait_for_memory;
 981
 982                         skb_entail(sk, skb);
 983                         copy = size_goal;
 984                 }
 985
 986                 if (copy > size)
 987                         copy = size;
 988
 989                 i = skb_shinfo(skb)->nr_frags;
 990                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 991                 if (!can_coalesce && i >= sysctl_max_skb_frags) {
 992                         tcp_mark_push(tp, skb);
 993                         goto new_segment;
 994                 }
 995                 if (!sk_wmem_schedule(sk, copy))
 996                         goto wait_for_memory;
 997
 998                 if (can_coalesce) {
 999                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1000                 } else {
1001                         get_page(page);
1002                         skb_fill_page_desc(skb, i, page, offset, copy);
1003                 }
1004
1005                 if (!(flags & MSG_NO_SHARED_FRAGS))
1006                         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1007
1008                 skb->len += copy;
1009                 skb->data_len += copy;
1010                 skb->truesize += copy;
1011                 sk->sk_wmem_queued += copy;
1012                 sk_mem_charge(sk, copy);
1013                 skb->ip_summed = CHECKSUM_PARTIAL;
1014                 tp->write_seq += copy;
1015                 TCP_SKB_CB(skb)->end_seq += copy;
1016                 tcp_skb_pcount_set(skb, 0);
1017
1018                 if (!copied)
1019                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1020
1021                 copied += copy;
1022                 offset += copy;
1023                 size -= copy;
1024                 if (!size)
1025                         goto out;
1026
1027                 if (skb->len < size_goal || (flags & MSG_OOB))
1028                         continue;
1029
1030                 if (forced_push(tp)) {
1031                         tcp_mark_push(tp, skb);
1032                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1033                 } else if (skb == tcp_send_head(sk))
1034                         tcp_push_one(sk, mss_now);
1035                 continue;
1036
1037 wait_for_sndbuf:
1038                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1039 wait_for_memory:
1040                 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1041                          TCP_NAGLE_PUSH, size_goal);
1042
1043                 err = sk_stream_wait_memory(sk, &timeo);
1044                 if (err != 0)
1045                         goto do_error;
1046
1047                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1048         }
1049
1050 out:
1051         if (copied) {
1052                 tcp_tx_timestamp(sk, sk->sk_tsflags);
1053                 if (!(flags & MSG_SENDPAGE_NOTLAST))
1054                         tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1055         }
1056         return copied;
1057
1058 do_error:
1059         if (copied)
1060                 goto out;
1061 out_err:
1062         /* make sure we wake any epoll edge trigger waiter */
1063         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1064                      err == -EAGAIN)) {
1065                 sk->sk_write_space(sk);
1066                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1067         }
1068         return sk_stream_error(sk, flags, err);
1069 }
1070 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1071
1072 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1073                         size_t size, int flags)
1074 {
1075         if (!(sk->sk_route_caps & NETIF_F_SG))
1076                 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1077
1078         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1079
1080         return do_tcp_sendpages(sk, page, offset, size, flags);
1081 }
1082 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1083
1084 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1085                  size_t size, int flags)
1086 {
1087         int ret;
1088
1089         lock_sock(sk);
1090         ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1091         release_sock(sk);
1092
1093         return ret;
1094 }
1095 EXPORT_SYMBOL(tcp_sendpage);
1096
1097 /* Do not bother using a page frag for very small frames.
1098  * But use this heuristic only for the first skb in write queue.
1099  *
1100  * Having no payload in skb->head allows better SACK shifting
1101  * in tcp_shift_skb_data(), reducing sack/rack overhead, because
1102  * write queue has less skbs.
1103  * Each skb can hold up to MAX_SKB_FRAGS * 32Kbytes, or ~0.5 MB.
1104  * This also speeds up tso_fragment(), since it wont fallback
1105  * to tcp_fragment().
1106  */
1107 static int linear_payload_sz(bool first_skb)
1108 {
1109         if (first_skb)
1110                 return SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1111         return 0;
1112 }
1113
1114 static int select_size(bool first_skb, bool zc)
1115 {
1116         if (zc)
1117                 return 0;
1118         return linear_payload_sz(first_skb);
1119 }
1120
1121 void tcp_free_fastopen_req(struct tcp_sock *tp)
1122 {
1123         if (tp->fastopen_req) {
1124                 kfree(tp->fastopen_req);
1125                 tp->fastopen_req = NULL;
1126         }
1127 }
1128
1129 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1130                                 int *copied, size_t size)
1131 {
1132         struct tcp_sock *tp = tcp_sk(sk);
1133         struct inet_sock *inet = inet_sk(sk);
1134         struct sockaddr *uaddr = msg->msg_name;
1135         int err, flags;
1136
1137         if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1138             (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1139              uaddr->sa_family == AF_UNSPEC))
1140                 return -EOPNOTSUPP;
1141         if (tp->fastopen_req)
1142                 return -EALREADY; /* Another Fast Open is in progress */
1143
1144         tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1145                                    sk->sk_allocation);
1146         if (unlikely(!tp->fastopen_req))
1147                 return -ENOBUFS;
1148         tp->fastopen_req->data = msg;
1149         tp->fastopen_req->size = size;
1150
1151         if (inet->defer_connect) {
1152                 err = tcp_connect(sk);
1153                 /* Same failure procedure as in tcp_v4/6_connect */
1154                 if (err) {
1155                         tcp_set_state(sk, TCP_CLOSE);
1156                         inet->inet_dport = 0;
1157                         sk->sk_route_caps = 0;
1158                 }
1159         }
1160         flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1161         err = __inet_stream_connect(sk->sk_socket, uaddr,
1162                                     msg->msg_namelen, flags, 1);
1163         /* fastopen_req could already be freed in __inet_stream_connect
1164          * if the connection times out or gets rst
1165          */
1166         if (tp->fastopen_req) {
1167                 *copied = tp->fastopen_req->copied;
1168                 tcp_free_fastopen_req(tp);
1169                 inet->defer_connect = 0;
1170         }
1171         return err;
1172 }
1173
1174 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1175 {
1176         struct tcp_sock *tp = tcp_sk(sk);
1177         struct ubuf_info *uarg = NULL;
1178         struct sk_buff *skb;
1179         struct sockcm_cookie sockc;
1180         int flags, err, copied = 0;
1181         int mss_now = 0, size_goal, copied_syn = 0;
1182         bool process_backlog = false;
1183         bool zc = false;
1184         long timeo;
1185
1186         flags = msg->msg_flags;
1187
1188         if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
1189                 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1190                         err = -EINVAL;
1191                         goto out_err;
1192                 }
1193
1194                 skb = tcp_write_queue_tail(sk);
1195                 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
1196                 if (!uarg) {
1197                         err = -ENOBUFS;
1198                         goto out_err;
1199                 }
1200
1201                 zc = sk->sk_route_caps & NETIF_F_SG;
1202                 if (!zc)
1203                         uarg->zerocopy = 0;
1204         }
1205
1206         if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1207             !tp->repair) {
1208                 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1209                 if (err == -EINPROGRESS && copied_syn > 0)
1210                         goto out;
1211                 else if (err)
1212                         goto out_err;
1213         }
1214
1215         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1216
1217         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1218
1219         /* Wait for a connection to finish. One exception is TCP Fast Open
1220          * (passive side) where data is allowed to be sent before a connection
1221          * is fully established.
1222          */
1223         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1224             !tcp_passive_fastopen(sk)) {
1225                 err = sk_stream_wait_connect(sk, &timeo);
1226                 if (err != 0)
1227                         goto do_error;
1228         }
1229
1230         if (unlikely(tp->repair)) {
1231                 if (tp->repair_queue == TCP_RECV_QUEUE) {
1232                         copied = tcp_send_rcvq(sk, msg, size);
1233                         goto out_nopush;
1234                 }
1235
1236                 err = -EINVAL;
1237                 if (tp->repair_queue == TCP_NO_QUEUE)
1238                         goto out_err;
1239
1240                 /* 'common' sending to sendq */
1241         }
1242
1243         sockcm_init(&sockc, sk);
1244         if (msg->msg_controllen) {
1245                 err = sock_cmsg_send(sk, msg, &sockc);
1246                 if (unlikely(err)) {
1247                         err = -EINVAL;
1248                         goto out_err;
1249                 }
1250         }
1251
1252         /* This should be in poll */
1253         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1254
1255         /* Ok commence sending. */
1256         copied = 0;
1257
1258 restart:
1259         mss_now = tcp_send_mss(sk, &size_goal, flags);
1260
1261         err = -EPIPE;
1262         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1263                 goto do_error;
1264
1265         while (msg_data_left(msg)) {
1266                 int copy = 0;
1267
1268                 skb = tcp_write_queue_tail(sk);
1269                 if (skb)
1270                         copy = size_goal - skb->len;
1271
1272                 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1273                         bool first_skb;
1274                         int linear;
1275
1276 new_segment:
1277                         if (!sk_stream_memory_free(sk))
1278                                 goto wait_for_sndbuf;
1279
1280                         if (process_backlog && sk_flush_backlog(sk)) {
1281                                 process_backlog = false;
1282                                 goto restart;
1283                         }
1284                         first_skb = tcp_rtx_and_write_queues_empty(sk);
1285                         linear = select_size(first_skb, zc);
1286                         skb = sk_stream_alloc_skb(sk, linear, sk->sk_allocation,
1287                                                   first_skb);
1288                         if (!skb)
1289                                 goto wait_for_memory;
1290
1291                         process_backlog = true;
1292                         skb->ip_summed = CHECKSUM_PARTIAL;
1293
1294                         skb_entail(sk, skb);
1295                         copy = size_goal;
1296
1297                         /* All packets are restored as if they have
1298                          * already been sent. skb_mstamp_ns isn't set to
1299                          * avoid wrong rtt estimation.
1300                          */
1301                         if (tp->repair)
1302                                 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1303                 }
1304
1305                 /* Try to append data to the end of skb. */
1306                 if (copy > msg_data_left(msg))
1307                         copy = msg_data_left(msg);
1308
1309                 /* Where to copy to? */
1310                 if (skb_availroom(skb) > 0 && !zc) {
1311                         /* We have some space in skb head. Superb! */
1312                         copy = min_t(int, copy, skb_availroom(skb));
1313                         err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1314                         if (err)
1315                                 goto do_fault;
1316                 } else if (!zc) {
1317                         bool merge = true;
1318                         int i = skb_shinfo(skb)->nr_frags;
1319                         struct page_frag *pfrag = sk_page_frag(sk);
1320
1321                         if (!sk_page_frag_refill(sk, pfrag))
1322                                 goto wait_for_memory;
1323
1324                         if (!skb_can_coalesce(skb, i, pfrag->page,
1325                                               pfrag->offset)) {
1326                                 if (i >= sysctl_max_skb_frags) {
1327                                         tcp_mark_push(tp, skb);
1328                                         goto new_segment;
1329                                 }
1330                                 merge = false;
1331                         }
1332
1333                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1334
1335                         if (!sk_wmem_schedule(sk, copy))
1336                                 goto wait_for_memory;
1337
1338                         err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1339                                                        pfrag->page,
1340                                                        pfrag->offset,
1341                                                        copy);
1342                         if (err)
1343                                 goto do_error;
1344
1345                         /* Update the skb. */
1346                         if (merge) {
1347                                 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1348                         } else {
1349                                 skb_fill_page_desc(skb, i, pfrag->page,
1350                                                    pfrag->offset, copy);
1351                                 page_ref_inc(pfrag->page);
1352                         }
1353                         pfrag->offset += copy;
1354                 } else {
1355                         err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1356                         if (err == -EMSGSIZE || err == -EEXIST) {
1357                                 tcp_mark_push(tp, skb);
1358                                 goto new_segment;
1359                         }
1360                         if (err < 0)
1361                                 goto do_error;
1362                         copy = err;
1363                 }
1364
1365                 if (!copied)
1366                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1367
1368                 tp->write_seq += copy;
1369                 TCP_SKB_CB(skb)->end_seq += copy;
1370                 tcp_skb_pcount_set(skb, 0);
1371
1372                 copied += copy;
1373                 if (!msg_data_left(msg)) {
1374                         if (unlikely(flags & MSG_EOR))
1375                                 TCP_SKB_CB(skb)->eor = 1;
1376                         goto out;
1377                 }
1378
1379                 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1380                         continue;
1381
1382                 if (forced_push(tp)) {
1383                         tcp_mark_push(tp, skb);
1384                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1385                 } else if (skb == tcp_send_head(sk))
1386                         tcp_push_one(sk, mss_now);
1387                 continue;
1388
1389 wait_for_sndbuf:
1390                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1391 wait_for_memory:
1392                 if (copied)
1393                         tcp_push(sk, flags & ~MSG_MORE, mss_now,
1394                                  TCP_NAGLE_PUSH, size_goal);
1395
1396                 err = sk_stream_wait_memory(sk, &timeo);
1397                 if (err != 0)
1398                         goto do_error;
1399
1400                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1401         }
1402
1403 out:
1404         if (copied) {
1405                 tcp_tx_timestamp(sk, sockc.tsflags);
1406                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1407         }
1408 out_nopush:
1409         sock_zerocopy_put(uarg);
1410         return copied + copied_syn;
1411
1412 do_fault:
1413         if (!skb->len) {
1414                 tcp_unlink_write_queue(skb, sk);
1415                 /* It is the one place in all of TCP, except connection
1416                  * reset, where we can be unlinking the send_head.
1417                  */
1418                 tcp_check_send_head(sk, skb);
1419                 sk_wmem_free_skb(sk, skb);
1420         }
1421
1422 do_error:
1423         if (copied + copied_syn)
1424                 goto out;
1425 out_err:
1426         sock_zerocopy_put_abort(uarg, true);
1427         err = sk_stream_error(sk, flags, err);
1428         /* make sure we wake any epoll edge trigger waiter */
1429         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1430                      err == -EAGAIN)) {
1431                 sk->sk_write_space(sk);
1432                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1433         }
1434         return err;
1435 }
1436 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1437
1438 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1439 {
1440         int ret;
1441
1442         lock_sock(sk);
1443         ret = tcp_sendmsg_locked(sk, msg, size);
1444         release_sock(sk);
1445
1446         return ret;
1447 }
1448 EXPORT_SYMBOL(tcp_sendmsg);
1449
1450 /*
1451  *      Handle reading urgent data. BSD has very simple semantics for
1452  *      this, no blocking and very strange errors 8)
1453  */
1454
1455 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1456 {
1457         struct tcp_sock *tp = tcp_sk(sk);
1458
1459         /* No URG data to read. */
1460         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1461             tp->urg_data == TCP_URG_READ)
1462                 return -EINVAL; /* Yes this is right ! */
1463
1464         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1465                 return -ENOTCONN;
1466
1467         if (tp->urg_data & TCP_URG_VALID) {
1468                 int err = 0;
1469                 char c = tp->urg_data;
1470
1471                 if (!(flags & MSG_PEEK))
1472                         tp->urg_data = TCP_URG_READ;
1473
1474                 /* Read urgent data. */
1475                 msg->msg_flags |= MSG_OOB;
1476
1477                 if (len > 0) {
1478                         if (!(flags & MSG_TRUNC))
1479                                 err = memcpy_to_msg(msg, &c, 1);
1480                         len = 1;
1481                 } else
1482                         msg->msg_flags |= MSG_TRUNC;
1483
1484                 return err ? -EFAULT : len;
1485         }
1486
1487         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1488                 return 0;
1489
1490         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1491          * the available implementations agree in this case:
1492          * this call should never block, independent of the
1493          * blocking state of the socket.
1494          * Mike <pall@rz.uni-karlsruhe.de>
1495          */
1496         return -EAGAIN;
1497 }
1498
1499 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1500 {
1501         struct sk_buff *skb;
1502         int copied = 0, err = 0;
1503
1504         /* XXX -- need to support SO_PEEK_OFF */
1505
1506         skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1507                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1508                 if (err)
1509                         return err;
1510                 copied += skb->len;
1511         }
1512
1513         skb_queue_walk(&sk->sk_write_queue, skb) {
1514                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1515                 if (err)
1516                         break;
1517
1518                 copied += skb->len;
1519         }
1520
1521         return err ?: copied;
1522 }
1523
1524 /* Clean up the receive buffer for full frames taken by the user,
1525  * then send an ACK if necessary.  COPIED is the number of bytes
1526  * tcp_recvmsg has given to the user so far, it speeds up the
1527  * calculation of whether or not we must ACK for the sake of
1528  * a window update.
1529  */
1530 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1531 {
1532         struct tcp_sock *tp = tcp_sk(sk);
1533         bool time_to_ack = false;
1534
1535         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1536
1537         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1538              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1539              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1540
1541         if (inet_csk_ack_scheduled(sk)) {
1542                 const struct inet_connection_sock *icsk = inet_csk(sk);
1543                    /* Delayed ACKs frequently hit locked sockets during bulk
1544                     * receive. */
1545                 if (icsk->icsk_ack.blocked ||
1546                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1547                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1548                     /*
1549                      * If this read emptied read buffer, we send ACK, if
1550                      * connection is not bidirectional, user drained
1551                      * receive buffer and there was a small segment
1552                      * in queue.
1553                      */
1554                     (copied > 0 &&
1555                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1556                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1557                        !icsk->icsk_ack.pingpong)) &&
1558                       !atomic_read(&sk->sk_rmem_alloc)))
1559                         time_to_ack = true;
1560         }
1561
1562         /* We send an ACK if we can now advertise a non-zero window
1563          * which has been raised "significantly".
1564          *
1565          * Even if window raised up to infinity, do not send window open ACK
1566          * in states, where we will not receive more. It is useless.
1567          */
1568         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1569                 __u32 rcv_window_now = tcp_receive_window(tp);
1570
1571                 /* Optimize, __tcp_select_window() is not cheap. */
1572                 if (2*rcv_window_now <= tp->window_clamp) {
1573                         __u32 new_window = __tcp_select_window(sk);
1574
1575                         /* Send ACK now, if this read freed lots of space
1576                          * in our buffer. Certainly, new_window is new window.
1577                          * We can advertise it now, if it is not less than current one.
1578                          * "Lots" means "at least twice" here.
1579                          */
1580                         if (new_window && new_window >= 2 * rcv_window_now)
1581                                 time_to_ack = true;
1582                 }
1583         }
1584         if (time_to_ack)
1585                 tcp_send_ack(sk);
1586 }
1587
1588 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1589 {
1590         struct sk_buff *skb;
1591         u32 offset;
1592
1593         while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1594                 offset = seq - TCP_SKB_CB(skb)->seq;
1595                 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1596                         pr_err_once("%s: found a SYN, please report !\n", __func__);
1597                         offset--;
1598                 }
1599                 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1600                         *off = offset;
1601                         return skb;
1602                 }
1603                 /* This looks weird, but this can happen if TCP collapsing
1604                  * splitted a fat GRO packet, while we released socket lock
1605                  * in skb_splice_bits()
1606                  */
1607                 sk_eat_skb(sk, skb);
1608         }
1609         return NULL;
1610 }
1611
1612 /*
1613  * This routine provides an alternative to tcp_recvmsg() for routines
1614  * that would like to handle copying from skbuffs directly in 'sendfile'
1615  * fashion.
1616  * Note:
1617  *      - It is assumed that the socket was locked by the caller.
1618  *      - The routine does not block.
1619  *      - At present, there is no support for reading OOB data
1620  *        or for 'peeking' the socket using this routine
1621  *        (although both would be easy to implement).
1622  */
1623 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1624                   sk_read_actor_t recv_actor)
1625 {
1626         struct sk_buff *skb;
1627         struct tcp_sock *tp = tcp_sk(sk);
1628         u32 seq = tp->copied_seq;
1629         u32 offset;
1630         int copied = 0;
1631
1632         if (sk->sk_state == TCP_LISTEN)
1633                 return -ENOTCONN;
1634         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1635                 if (offset < skb->len) {
1636                         int used;
1637                         size_t len;
1638
1639                         len = skb->len - offset;
1640                         /* Stop reading if we hit a patch of urgent data */
1641                         if (tp->urg_data) {
1642                                 u32 urg_offset = tp->urg_seq - seq;
1643                                 if (urg_offset < len)
1644                                         len = urg_offset;
1645                                 if (!len)
1646                                         break;
1647                         }
1648                         used = recv_actor(desc, skb, offset, len);
1649                         if (used <= 0) {
1650                                 if (!copied)
1651                                         copied = used;
1652                                 break;
1653                         } else if (used <= len) {
1654                                 seq += used;
1655                                 copied += used;
1656                                 offset += used;
1657                         }
1658                         /* If recv_actor drops the lock (e.g. TCP splice
1659                          * receive) the skb pointer might be invalid when
1660                          * getting here: tcp_collapse might have deleted it
1661                          * while aggregating skbs from the socket queue.
1662                          */
1663                         skb = tcp_recv_skb(sk, seq - 1, &offset);
1664                         if (!skb)
1665                                 break;
1666                         /* TCP coalescing might have appended data to the skb.
1667                          * Try to splice more frags
1668                          */
1669                         if (offset + 1 != skb->len)
1670                                 continue;
1671                 }
1672                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1673                         sk_eat_skb(sk, skb);
1674                         ++seq;
1675                         break;
1676                 }
1677                 sk_eat_skb(sk, skb);
1678                 if (!desc->count)
1679                         break;
1680                 tp->copied_seq = seq;
1681         }
1682         tp->copied_seq = seq;
1683
1684         tcp_rcv_space_adjust(sk);
1685
1686         /* Clean up data we have read: This will do ACK frames. */
1687         if (copied > 0) {
1688                 tcp_recv_skb(sk, seq, &offset);
1689                 tcp_cleanup_rbuf(sk, copied);
1690         }
1691         return copied;
1692 }
1693 EXPORT_SYMBOL(tcp_read_sock);
1694
1695 int tcp_peek_len(struct socket *sock)
1696 {
1697         return tcp_inq(sock->sk);
1698 }
1699 EXPORT_SYMBOL(tcp_peek_len);
1700
1701 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1702 int tcp_set_rcvlowat(struct sock *sk, int val)
1703 {
1704         int cap;
1705
1706         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1707                 cap = sk->sk_rcvbuf >> 1;
1708         else
1709                 cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1;
1710         val = min(val, cap);
1711         sk->sk_rcvlowat = val ? : 1;
1712
1713         /* Check if we need to signal EPOLLIN right now */
1714         tcp_data_ready(sk);
1715
1716         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1717                 return 0;
1718
1719         val <<= 1;
1720         if (val > sk->sk_rcvbuf) {
1721                 sk->sk_rcvbuf = val;
1722                 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1723         }
1724         return 0;
1725 }
1726 EXPORT_SYMBOL(tcp_set_rcvlowat);
1727
1728 #ifdef CONFIG_MMU
1729 static const struct vm_operations_struct tcp_vm_ops = {
1730 };
1731
1732 int tcp_mmap(struct file *file, struct socket *sock,
1733              struct vm_area_struct *vma)
1734 {
1735         if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1736                 return -EPERM;
1737         vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1738
1739         /* Instruct vm_insert_page() to not down_read(mmap_sem) */
1740         vma->vm_flags |= VM_MIXEDMAP;
1741
1742         vma->vm_ops = &tcp_vm_ops;
1743         return 0;
1744 }
1745 EXPORT_SYMBOL(tcp_mmap);
1746
1747 static int tcp_zerocopy_receive(struct sock *sk,
1748                                 struct tcp_zerocopy_receive *zc)
1749 {
1750         unsigned long address = (unsigned long)zc->address;
1751         const skb_frag_t *frags = NULL;
1752         u32 length = 0, seq, offset;
1753         struct vm_area_struct *vma;
1754         struct sk_buff *skb = NULL;
1755         struct tcp_sock *tp;
1756         int inq;
1757         int ret;
1758
1759         if (address & (PAGE_SIZE - 1) || address != zc->address)
1760                 return -EINVAL;
1761
1762         if (sk->sk_state == TCP_LISTEN)
1763                 return -ENOTCONN;
1764
1765         sock_rps_record_flow(sk);
1766
1767         down_read(&current->mm->mmap_sem);
1768
1769         ret = -EINVAL;
1770         vma = find_vma(current->mm, address);
1771         if (!vma || vma->vm_start > address || vma->vm_ops != &tcp_vm_ops)
1772                 goto out;
1773         zc->length = min_t(unsigned long, zc->length, vma->vm_end - address);
1774
1775         tp = tcp_sk(sk);
1776         seq = tp->copied_seq;
1777         inq = tcp_inq(sk);
1778         zc->length = min_t(u32, zc->length, inq);
1779         zc->length &= ~(PAGE_SIZE - 1);
1780         if (zc->length) {
1781                 zap_page_range(vma, address, zc->length);
1782                 zc->recv_skip_hint = 0;
1783         } else {
1784                 zc->recv_skip_hint = inq;
1785         }
1786         ret = 0;
1787         while (length + PAGE_SIZE <= zc->length) {
1788                 if (zc->recv_skip_hint < PAGE_SIZE) {
1789                         if (skb) {
1790                                 skb = skb->next;
1791                                 offset = seq - TCP_SKB_CB(skb)->seq;
1792                         } else {
1793                                 skb = tcp_recv_skb(sk, seq, &offset);
1794                         }
1795
1796                         zc->recv_skip_hint = skb->len - offset;
1797                         offset -= skb_headlen(skb);
1798                         if ((int)offset < 0 || skb_has_frag_list(skb))
1799                                 break;
1800                         frags = skb_shinfo(skb)->frags;
1801                         while (offset) {
1802                                 if (frags->size > offset)
1803                                         goto out;
1804                                 offset -= frags->size;
1805                                 frags++;
1806                         }
1807                 }
1808                 if (frags->size != PAGE_SIZE || frags->page_offset) {
1809                         int remaining = zc->recv_skip_hint;
1810
1811                         while (remaining && (frags->size != PAGE_SIZE ||
1812                                              frags->page_offset)) {
1813                                 remaining -= frags->size;
1814                                 frags++;
1815                         }
1816                         zc->recv_skip_hint -= remaining;
1817                         break;
1818                 }
1819                 ret = vm_insert_page(vma, address + length,
1820                                      skb_frag_page(frags));
1821                 if (ret)
1822                         break;
1823                 length += PAGE_SIZE;
1824                 seq += PAGE_SIZE;
1825                 zc->recv_skip_hint -= PAGE_SIZE;
1826                 frags++;
1827         }
1828 out:
1829         up_read(&current->mm->mmap_sem);
1830         if (length) {
1831                 tp->copied_seq = seq;
1832                 tcp_rcv_space_adjust(sk);
1833
1834                 /* Clean up data we have read: This will do ACK frames. */
1835                 tcp_recv_skb(sk, seq, &offset);
1836                 tcp_cleanup_rbuf(sk, length);
1837                 ret = 0;
1838                 if (length == zc->length)
1839                         zc->recv_skip_hint = 0;
1840         } else {
1841                 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
1842                         ret = -EIO;
1843         }
1844         zc->length = length;
1845         return ret;
1846 }
1847 #endif
1848
1849 static void tcp_update_recv_tstamps(struct sk_buff *skb,
1850                                     struct scm_timestamping *tss)
1851 {
1852         if (skb->tstamp)
1853                 tss->ts[0] = ktime_to_timespec(skb->tstamp);
1854         else
1855                 tss->ts[0] = (struct timespec) {0};
1856
1857         if (skb_hwtstamps(skb)->hwtstamp)
1858                 tss->ts[2] = ktime_to_timespec(skb_hwtstamps(skb)->hwtstamp);
1859         else
1860                 tss->ts[2] = (struct timespec) {0};
1861 }
1862
1863 /* Similar to __sock_recv_timestamp, but does not require an skb */
1864 static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
1865                                struct scm_timestamping *tss)
1866 {
1867         struct timeval tv;
1868         bool has_timestamping = false;
1869
1870         if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
1871                 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
1872                         if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
1873                                 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPNS,
1874                                          sizeof(tss->ts[0]), &tss->ts[0]);
1875                         } else {
1876                                 tv.tv_sec = tss->ts[0].tv_sec;
1877                                 tv.tv_usec = tss->ts[0].tv_nsec / 1000;
1878
1879                                 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMP,
1880                                          sizeof(tv), &tv);
1881                         }
1882                 }
1883
1884                 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
1885                         has_timestamping = true;
1886                 else
1887                         tss->ts[0] = (struct timespec) {0};
1888         }
1889
1890         if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
1891                 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
1892                         has_timestamping = true;
1893                 else
1894                         tss->ts[2] = (struct timespec) {0};
1895         }
1896
1897         if (has_timestamping) {
1898                 tss->ts[1] = (struct timespec) {0};
1899                 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPING,
1900                          sizeof(*tss), tss);
1901         }
1902 }
1903
1904 static int tcp_inq_hint(struct sock *sk)
1905 {
1906         const struct tcp_sock *tp = tcp_sk(sk);
1907         u32 copied_seq = READ_ONCE(tp->copied_seq);
1908         u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
1909         int inq;
1910
1911         inq = rcv_nxt - copied_seq;
1912         if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
1913                 lock_sock(sk);
1914                 inq = tp->rcv_nxt - tp->copied_seq;
1915                 release_sock(sk);
1916         }
1917         return inq;
1918 }
1919
1920 /*
1921  *      This routine copies from a sock struct into the user buffer.
1922  *
1923  *      Technical note: in 2.3 we work on _locked_ socket, so that
1924  *      tricks with *seq access order and skb->users are not required.
1925  *      Probably, code can be easily improved even more.
1926  */
1927
1928 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1929                 int flags, int *addr_len)
1930 {
1931         struct tcp_sock *tp = tcp_sk(sk);
1932         int copied = 0;
1933         u32 peek_seq;
1934         u32 *seq;
1935         unsigned long used;
1936         int err, inq;
1937         int target;             /* Read at least this many bytes */
1938         long timeo;
1939         struct sk_buff *skb, *last;
1940         u32 urg_hole = 0;
1941         struct scm_timestamping tss;
1942         bool has_tss = false;
1943         bool has_cmsg;
1944
1945         if (unlikely(flags & MSG_ERRQUEUE))
1946                 return inet_recv_error(sk, msg, len, addr_len);
1947
1948         if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1949             (sk->sk_state == TCP_ESTABLISHED))
1950                 sk_busy_loop(sk, nonblock);
1951
1952         lock_sock(sk);
1953
1954         err = -ENOTCONN;
1955         if (sk->sk_state == TCP_LISTEN)
1956                 goto out;
1957
1958         has_cmsg = tp->recvmsg_inq;
1959         timeo = sock_rcvtimeo(sk, nonblock);
1960
1961         /* Urgent data needs to be handled specially. */
1962         if (flags & MSG_OOB)
1963                 goto recv_urg;
1964
1965         if (unlikely(tp->repair)) {
1966                 err = -EPERM;
1967                 if (!(flags & MSG_PEEK))
1968                         goto out;
1969
1970                 if (tp->repair_queue == TCP_SEND_QUEUE)
1971                         goto recv_sndq;
1972
1973                 err = -EINVAL;
1974                 if (tp->repair_queue == TCP_NO_QUEUE)
1975                         goto out;
1976
1977                 /* 'common' recv queue MSG_PEEK-ing */
1978         }
1979
1980         seq = &tp->copied_seq;
1981         if (flags & MSG_PEEK) {
1982                 peek_seq = tp->copied_seq;
1983                 seq = &peek_seq;
1984         }
1985
1986         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1987
1988         do {
1989                 u32 offset;
1990
1991                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1992                 if (tp->urg_data && tp->urg_seq == *seq) {
1993                         if (copied)
1994                                 break;
1995                         if (signal_pending(current)) {
1996                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1997                                 break;
1998                         }
1999                 }
2000
2001                 /* Next get a buffer. */
2002
2003                 last = skb_peek_tail(&sk->sk_receive_queue);
2004                 skb_queue_walk(&sk->sk_receive_queue, skb) {
2005                         last = skb;
2006                         /* Now that we have two receive queues this
2007                          * shouldn't happen.
2008                          */
2009                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2010                                  "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2011                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2012                                  flags))
2013                                 break;
2014
2015                         offset = *seq - TCP_SKB_CB(skb)->seq;
2016                         if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2017                                 pr_err_once("%s: found a SYN, please report !\n", __func__);
2018                                 offset--;
2019                         }
2020                         if (offset < skb->len)
2021                                 goto found_ok_skb;
2022                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2023                                 goto found_fin_ok;
2024                         WARN(!(flags & MSG_PEEK),
2025                              "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2026                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2027                 }
2028
2029                 /* Well, if we have backlog, try to process it now yet. */
2030
2031                 if (copied >= target && !sk->sk_backlog.tail)
2032                         break;
2033
2034                 if (copied) {
2035                         if (sk->sk_err ||
2036                             sk->sk_state == TCP_CLOSE ||
2037                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
2038                             !timeo ||
2039                             signal_pending(current))
2040                                 break;
2041                 } else {
2042                         if (sock_flag(sk, SOCK_DONE))
2043                                 break;
2044
2045                         if (sk->sk_err) {
2046                                 copied = sock_error(sk);
2047                                 break;
2048                         }
2049
2050                         if (sk->sk_shutdown & RCV_SHUTDOWN)
2051                                 break;
2052
2053                         if (sk->sk_state == TCP_CLOSE) {
2054                                 /* This occurs when user tries to read
2055                                  * from never connected socket.
2056                                  */
2057                                 copied = -ENOTCONN;
2058                                 break;
2059                         }
2060
2061                         if (!timeo) {
2062                                 copied = -EAGAIN;
2063                                 break;
2064                         }
2065
2066                         if (signal_pending(current)) {
2067                                 copied = sock_intr_errno(timeo);
2068                                 break;
2069                         }
2070                 }
2071
2072                 tcp_cleanup_rbuf(sk, copied);
2073
2074                 if (copied >= target) {
2075                         /* Do not sleep, just process backlog. */
2076                         release_sock(sk);
2077                         lock_sock(sk);
2078                 } else {
2079                         sk_wait_data(sk, &timeo, last);
2080                 }
2081
2082                 if ((flags & MSG_PEEK) &&
2083                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
2084                         net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2085                                             current->comm,
2086                                             task_pid_nr(current));
2087                         peek_seq = tp->copied_seq;
2088                 }
2089                 continue;
2090
2091 found_ok_skb:
2092                 /* Ok so how much can we use? */
2093                 used = skb->len - offset;
2094                 if (len < used)
2095                         used = len;
2096
2097                 /* Do we have urgent data here? */
2098                 if (tp->urg_data) {
2099                         u32 urg_offset = tp->urg_seq - *seq;
2100                         if (urg_offset < used) {
2101                                 if (!urg_offset) {
2102                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
2103                                                 ++*seq;
2104                                                 urg_hole++;
2105                                                 offset++;
2106                                                 used--;
2107                                                 if (!used)
2108                                                         goto skip_copy;
2109                                         }
2110                                 } else
2111                                         used = urg_offset;
2112                         }
2113                 }
2114
2115                 if (!(flags & MSG_TRUNC)) {
2116                         err = skb_copy_datagram_msg(skb, offset, msg, used);
2117                         if (err) {
2118                                 /* Exception. Bailout! */
2119                                 if (!copied)
2120                                         copied = -EFAULT;
2121                                 break;
2122                         }
2123                 }
2124
2125                 *seq += used;
2126                 copied += used;
2127                 len -= used;
2128
2129                 tcp_rcv_space_adjust(sk);
2130
2131 skip_copy:
2132                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
2133                         tp->urg_data = 0;
2134                         tcp_fast_path_check(sk);
2135                 }
2136                 if (used + offset < skb->len)
2137                         continue;
2138
2139                 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2140                         tcp_update_recv_tstamps(skb, &tss);
2141                         has_tss = true;
2142                         has_cmsg = true;
2143                 }
2144                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2145                         goto found_fin_ok;
2146                 if (!(flags & MSG_PEEK))
2147                         sk_eat_skb(sk, skb);
2148                 continue;
2149
2150 found_fin_ok:
2151                 /* Process the FIN. */
2152                 ++*seq;
2153                 if (!(flags & MSG_PEEK))
2154                         sk_eat_skb(sk, skb);
2155                 break;
2156         } while (len > 0);
2157
2158         /* According to UNIX98, msg_name/msg_namelen are ignored
2159          * on connected socket. I was just happy when found this 8) --ANK
2160          */
2161
2162         /* Clean up data we have read: This will do ACK frames. */
2163         tcp_cleanup_rbuf(sk, copied);
2164
2165         release_sock(sk);
2166
2167         if (has_cmsg) {
2168                 if (has_tss)
2169                         tcp_recv_timestamp(msg, sk, &tss);
2170                 if (tp->recvmsg_inq) {
2171                         inq = tcp_inq_hint(sk);
2172                         put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2173                 }
2174         }
2175
2176         return copied;
2177
2178 out:
2179         release_sock(sk);
2180         return err;
2181
2182 recv_urg:
2183         err = tcp_recv_urg(sk, msg, len, flags);
2184         goto out;
2185
2186 recv_sndq:
2187         err = tcp_peek_sndq(sk, msg, len);
2188         goto out;
2189 }
2190 EXPORT_SYMBOL(tcp_recvmsg);
2191
2192 void tcp_set_state(struct sock *sk, int state)
2193 {
2194         int oldstate = sk->sk_state;
2195
2196         /* We defined a new enum for TCP states that are exported in BPF
2197          * so as not force the internal TCP states to be frozen. The
2198          * following checks will detect if an internal state value ever
2199          * differs from the BPF value. If this ever happens, then we will
2200          * need to remap the internal value to the BPF value before calling
2201          * tcp_call_bpf_2arg.
2202          */
2203         BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2204         BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2205         BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2206         BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2207         BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2208         BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2209         BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2210         BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2211         BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2212         BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2213         BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2214         BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2215         BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2216
2217         if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2218                 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2219
2220         switch (state) {
2221         case TCP_ESTABLISHED:
2222                 if (oldstate != TCP_ESTABLISHED)
2223                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2224                 break;
2225
2226         case TCP_CLOSE:
2227                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2228                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2229
2230                 sk->sk_prot->unhash(sk);
2231                 if (inet_csk(sk)->icsk_bind_hash &&
2232                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2233                         inet_put_port(sk);
2234                 /* fall through */
2235         default:
2236                 if (oldstate == TCP_ESTABLISHED)
2237                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2238         }
2239
2240         /* Change state AFTER socket is unhashed to avoid closed
2241          * socket sitting in hash tables.
2242          */
2243         inet_sk_state_store(sk, state);
2244 }
2245 EXPORT_SYMBOL_GPL(tcp_set_state);
2246
2247 /*
2248  *      State processing on a close. This implements the state shift for
2249  *      sending our FIN frame. Note that we only send a FIN for some
2250  *      states. A shutdown() may have already sent the FIN, or we may be
2251  *      closed.
2252  */
2253
2254 static const unsigned char new_state[16] = {
2255   /* current state:        new state:      action:      */
2256   [0 /* (Invalid) */]   = TCP_CLOSE,
2257   [TCP_ESTABLISHED]     = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2258   [TCP_SYN_SENT]        = TCP_CLOSE,
2259   [TCP_SYN_RECV]        = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2260   [TCP_FIN_WAIT1]       = TCP_FIN_WAIT1,
2261   [TCP_FIN_WAIT2]       = TCP_FIN_WAIT2,
2262   [TCP_TIME_WAIT]       = TCP_CLOSE,
2263   [TCP_CLOSE]           = TCP_CLOSE,
2264   [TCP_CLOSE_WAIT]      = TCP_LAST_ACK  | TCP_ACTION_FIN,
2265   [TCP_LAST_ACK]        = TCP_LAST_ACK,
2266   [TCP_LISTEN]          = TCP_CLOSE,
2267   [TCP_CLOSING]         = TCP_CLOSING,
2268   [TCP_NEW_SYN_RECV]    = TCP_CLOSE,    /* should not happen ! */
2269 };
2270
2271 static int tcp_close_state(struct sock *sk)
2272 {
2273         int next = (int)new_state[sk->sk_state];
2274         int ns = next & TCP_STATE_MASK;
2275
2276         tcp_set_state(sk, ns);
2277
2278         return next & TCP_ACTION_FIN;
2279 }
2280
2281 /*
2282  *      Shutdown the sending side of a connection. Much like close except
2283  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2284  */
2285
2286 void tcp_shutdown(struct sock *sk, int how)
2287 {
2288         /*      We need to grab some memory, and put together a FIN,
2289          *      and then put it into the queue to be sent.
2290          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2291          */
2292         if (!(how & SEND_SHUTDOWN))
2293                 return;
2294
2295         /* If we've already sent a FIN, or it's a closed state, skip this. */
2296         if ((1 << sk->sk_state) &
2297             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2298              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2299                 /* Clear out any half completed packets.  FIN if needed. */
2300                 if (tcp_close_state(sk))
2301                         tcp_send_fin(sk);
2302         }
2303 }
2304 EXPORT_SYMBOL(tcp_shutdown);
2305
2306 bool tcp_check_oom(struct sock *sk, int shift)
2307 {
2308         bool too_many_orphans, out_of_socket_memory;
2309
2310         too_many_orphans = tcp_too_many_orphans(sk, shift);
2311         out_of_socket_memory = tcp_out_of_memory(sk);
2312
2313         if (too_many_orphans)
2314                 net_info_ratelimited("too many orphaned sockets\n");
2315         if (out_of_socket_memory)
2316                 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2317         return too_many_orphans || out_of_socket_memory;
2318 }
2319
2320 void tcp_close(struct sock *sk, long timeout)
2321 {
2322         struct sk_buff *skb;
2323         int data_was_unread = 0;
2324         int state;
2325
2326         lock_sock(sk);
2327         sk->sk_shutdown = SHUTDOWN_MASK;
2328
2329         if (sk->sk_state == TCP_LISTEN) {
2330                 tcp_set_state(sk, TCP_CLOSE);
2331
2332                 /* Special case. */
2333                 inet_csk_listen_stop(sk);
2334
2335                 goto adjudge_to_death;
2336         }
2337
2338         /*  We need to flush the recv. buffs.  We do this only on the
2339          *  descriptor close, not protocol-sourced closes, because the
2340          *  reader process may not have drained the data yet!
2341          */
2342         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2343                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2344
2345                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2346                         len--;
2347                 data_was_unread += len;
2348                 __kfree_skb(skb);
2349         }
2350
2351         sk_mem_reclaim(sk);
2352
2353         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2354         if (sk->sk_state == TCP_CLOSE)
2355                 goto adjudge_to_death;
2356
2357         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2358          * data was lost. To witness the awful effects of the old behavior of
2359          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2360          * GET in an FTP client, suspend the process, wait for the client to
2361          * advertise a zero window, then kill -9 the FTP client, wheee...
2362          * Note: timeout is always zero in such a case.
2363          */
2364         if (unlikely(tcp_sk(sk)->repair)) {
2365                 sk->sk_prot->disconnect(sk, 0);
2366         } else if (data_was_unread) {
2367                 /* Unread data was tossed, zap the connection. */
2368                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2369                 tcp_set_state(sk, TCP_CLOSE);
2370                 tcp_send_active_reset(sk, sk->sk_allocation);
2371         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2372                 /* Check zero linger _after_ checking for unread data. */
2373                 sk->sk_prot->disconnect(sk, 0);
2374                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2375         } else if (tcp_close_state(sk)) {
2376                 /* We FIN if the application ate all the data before
2377                  * zapping the connection.
2378                  */
2379
2380                 /* RED-PEN. Formally speaking, we have broken TCP state
2381                  * machine. State transitions:
2382                  *
2383                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2384                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2385                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2386                  *
2387                  * are legal only when FIN has been sent (i.e. in window),
2388                  * rather than queued out of window. Purists blame.
2389                  *
2390                  * F.e. "RFC state" is ESTABLISHED,
2391                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2392                  *
2393                  * The visible declinations are that sometimes
2394                  * we enter time-wait state, when it is not required really
2395                  * (harmless), do not send active resets, when they are
2396                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2397                  * they look as CLOSING or LAST_ACK for Linux)
2398                  * Probably, I missed some more holelets.
2399                  *                                              --ANK
2400                  * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2401                  * in a single packet! (May consider it later but will
2402                  * probably need API support or TCP_CORK SYN-ACK until
2403                  * data is written and socket is closed.)
2404                  */
2405                 tcp_send_fin(sk);
2406         }
2407
2408         sk_stream_wait_close(sk, timeout);
2409
2410 adjudge_to_death:
2411         state = sk->sk_state;
2412         sock_hold(sk);
2413         sock_orphan(sk);
2414
2415         local_bh_disable();
2416         bh_lock_sock(sk);
2417         /* remove backlog if any, without releasing ownership. */
2418         __release_sock(sk);
2419
2420         percpu_counter_inc(sk->sk_prot->orphan_count);
2421
2422         /* Have we already been destroyed by a softirq or backlog? */
2423         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2424                 goto out;
2425
2426         /*      This is a (useful) BSD violating of the RFC. There is a
2427          *      problem with TCP as specified in that the other end could
2428          *      keep a socket open forever with no application left this end.
2429          *      We use a 1 minute timeout (about the same as BSD) then kill
2430          *      our end. If they send after that then tough - BUT: long enough
2431          *      that we won't make the old 4*rto = almost no time - whoops
2432          *      reset mistake.
2433          *
2434          *      Nope, it was not mistake. It is really desired behaviour
2435          *      f.e. on http servers, when such sockets are useless, but
2436          *      consume significant resources. Let's do it with special
2437          *      linger2 option.                                 --ANK
2438          */
2439
2440         if (sk->sk_state == TCP_FIN_WAIT2) {
2441                 struct tcp_sock *tp = tcp_sk(sk);
2442                 if (tp->linger2 < 0) {
2443                         tcp_set_state(sk, TCP_CLOSE);
2444                         tcp_send_active_reset(sk, GFP_ATOMIC);
2445                         __NET_INC_STATS(sock_net(sk),
2446                                         LINUX_MIB_TCPABORTONLINGER);
2447                 } else {
2448                         const int tmo = tcp_fin_time(sk);
2449
2450                         if (tmo > TCP_TIMEWAIT_LEN) {
2451                                 inet_csk_reset_keepalive_timer(sk,
2452                                                 tmo - TCP_TIMEWAIT_LEN);
2453                         } else {
2454                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2455                                 goto out;
2456                         }
2457                 }
2458         }
2459         if (sk->sk_state != TCP_CLOSE) {
2460                 sk_mem_reclaim(sk);
2461                 if (tcp_check_oom(sk, 0)) {
2462                         tcp_set_state(sk, TCP_CLOSE);
2463                         tcp_send_active_reset(sk, GFP_ATOMIC);
2464                         __NET_INC_STATS(sock_net(sk),
2465                                         LINUX_MIB_TCPABORTONMEMORY);
2466                 } else if (!check_net(sock_net(sk))) {
2467                         /* Not possible to send reset; just close */
2468                         tcp_set_state(sk, TCP_CLOSE);
2469                 }
2470         }
2471
2472         if (sk->sk_state == TCP_CLOSE) {
2473                 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2474                 /* We could get here with a non-NULL req if the socket is
2475                  * aborted (e.g., closed with unread data) before 3WHS
2476                  * finishes.
2477                  */
2478                 if (req)
2479                         reqsk_fastopen_remove(sk, req, false);
2480                 inet_csk_destroy_sock(sk);
2481         }
2482         /* Otherwise, socket is reprieved until protocol close. */
2483
2484 out:
2485         bh_unlock_sock(sk);
2486         local_bh_enable();
2487         release_sock(sk);
2488         sock_put(sk);
2489 }
2490 EXPORT_SYMBOL(tcp_close);
2491
2492 /* These states need RST on ABORT according to RFC793 */
2493
2494 static inline bool tcp_need_reset(int state)
2495 {
2496         return (1 << state) &
2497                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2498                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2499 }
2500
2501 static void tcp_rtx_queue_purge(struct sock *sk)
2502 {
2503         struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2504
2505         while (p) {
2506                 struct sk_buff *skb = rb_to_skb(p);
2507
2508                 p = rb_next(p);
2509                 /* Since we are deleting whole queue, no need to
2510                  * list_del(&skb->tcp_tsorted_anchor)
2511                  */
2512                 tcp_rtx_queue_unlink(skb, sk);
2513                 sk_wmem_free_skb(sk, skb);
2514         }
2515 }
2516
2517 void tcp_write_queue_purge(struct sock *sk)
2518 {
2519         struct sk_buff *skb;
2520
2521         tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2522         while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2523                 tcp_skb_tsorted_anchor_cleanup(skb);
2524                 sk_wmem_free_skb(sk, skb);
2525         }
2526         tcp_rtx_queue_purge(sk);
2527         INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2528         sk_mem_reclaim(sk);
2529         tcp_clear_all_retrans_hints(tcp_sk(sk));
2530         tcp_sk(sk)->packets_out = 0;
2531         inet_csk(sk)->icsk_backoff = 0;
2532 }
2533
2534 int tcp_disconnect(struct sock *sk, int flags)
2535 {
2536         struct inet_sock *inet = inet_sk(sk);
2537         struct inet_connection_sock *icsk = inet_csk(sk);
2538         struct tcp_sock *tp = tcp_sk(sk);
2539         int old_state = sk->sk_state;
2540
2541         if (old_state != TCP_CLOSE)
2542                 tcp_set_state(sk, TCP_CLOSE);
2543
2544         /* ABORT function of RFC793 */
2545         if (old_state == TCP_LISTEN) {
2546                 inet_csk_listen_stop(sk);
2547         } else if (unlikely(tp->repair)) {
2548                 sk->sk_err = ECONNABORTED;
2549         } else if (tcp_need_reset(old_state) ||
2550                    (tp->snd_nxt != tp->write_seq &&
2551                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2552                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2553                  * states
2554                  */
2555                 tcp_send_active_reset(sk, gfp_any());
2556                 sk->sk_err = ECONNRESET;
2557         } else if (old_state == TCP_SYN_SENT)
2558                 sk->sk_err = ECONNRESET;
2559
2560         tcp_clear_xmit_timers(sk);
2561         __skb_queue_purge(&sk->sk_receive_queue);
2562         tp->copied_seq = tp->rcv_nxt;
2563         tp->urg_data = 0;
2564         tcp_write_queue_purge(sk);
2565         tcp_fastopen_active_disable_ofo_check(sk);
2566         skb_rbtree_purge(&tp->out_of_order_queue);
2567
2568         inet->inet_dport = 0;
2569
2570         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2571                 inet_reset_saddr(sk);
2572
2573         sk->sk_shutdown = 0;
2574         sock_reset_flag(sk, SOCK_DONE);
2575         tp->srtt_us = 0;
2576         tp->rcv_rtt_last_tsecr = 0;
2577         tp->write_seq += tp->max_window + 2;
2578         if (tp->write_seq == 0)
2579                 tp->write_seq = 1;
2580         tp->snd_cwnd = 2;
2581         icsk->icsk_probes_out = 0;
2582         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2583         tp->snd_cwnd_cnt = 0;
2584         tp->window_clamp = 0;
2585         tp->delivered_ce = 0;
2586         tcp_set_ca_state(sk, TCP_CA_Open);
2587         tp->is_sack_reneg = 0;
2588         tcp_clear_retrans(tp);
2589         inet_csk_delack_init(sk);
2590         /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2591          * issue in __tcp_select_window()
2592          */
2593         icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2594         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2595         __sk_dst_reset(sk);
2596         dst_release(sk->sk_rx_dst);
2597         sk->sk_rx_dst = NULL;
2598         tcp_saved_syn_free(tp);
2599         tp->compressed_ack = 0;
2600         tp->bytes_sent = 0;
2601         tp->bytes_retrans = 0;
2602         tp->duplicate_sack[0].start_seq = 0;
2603         tp->duplicate_sack[0].end_seq = 0;
2604         tp->dsack_dups = 0;
2605         tp->reord_seen = 0;
2606
2607         /* Clean up fastopen related fields */
2608         tcp_free_fastopen_req(tp);
2609         inet->defer_connect = 0;
2610
2611         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2612
2613         if (sk->sk_frag.page) {
2614                 put_page(sk->sk_frag.page);
2615                 sk->sk_frag.page = NULL;
2616                 sk->sk_frag.offset = 0;
2617         }
2618
2619         sk->sk_error_report(sk);
2620         return 0;
2621 }
2622 EXPORT_SYMBOL(tcp_disconnect);
2623
2624 static inline bool tcp_can_repair_sock(const struct sock *sk)
2625 {
2626         return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2627                 (sk->sk_state != TCP_LISTEN);
2628 }
2629
2630 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2631 {
2632         struct tcp_repair_window opt;
2633
2634         if (!tp->repair)
2635                 return -EPERM;
2636
2637         if (len != sizeof(opt))
2638                 return -EINVAL;
2639
2640         if (copy_from_user(&opt, optbuf, sizeof(opt)))
2641                 return -EFAULT;
2642
2643         if (opt.max_window < opt.snd_wnd)
2644                 return -EINVAL;
2645
2646         if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2647                 return -EINVAL;
2648
2649         if (after(opt.rcv_wup, tp->rcv_nxt))
2650                 return -EINVAL;
2651
2652         tp->snd_wl1     = opt.snd_wl1;
2653         tp->snd_wnd     = opt.snd_wnd;
2654         tp->max_window  = opt.max_window;
2655
2656         tp->rcv_wnd     = opt.rcv_wnd;
2657         tp->rcv_wup     = opt.rcv_wup;
2658
2659         return 0;
2660 }
2661
2662 static int tcp_repair_options_est(struct sock *sk,
2663                 struct tcp_repair_opt __user *optbuf, unsigned int len)
2664 {
2665         struct tcp_sock *tp = tcp_sk(sk);
2666         struct tcp_repair_opt opt;
2667
2668         while (len >= sizeof(opt)) {
2669                 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2670                         return -EFAULT;
2671
2672                 optbuf++;
2673                 len -= sizeof(opt);
2674
2675                 switch (opt.opt_code) {
2676                 case TCPOPT_MSS:
2677                         tp->rx_opt.mss_clamp = opt.opt_val;
2678                         tcp_mtup_init(sk);
2679                         break;
2680                 case TCPOPT_WINDOW:
2681                         {
2682                                 u16 snd_wscale = opt.opt_val & 0xFFFF;
2683                                 u16 rcv_wscale = opt.opt_val >> 16;
2684
2685                                 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
2686                                         return -EFBIG;
2687
2688                                 tp->rx_opt.snd_wscale = snd_wscale;
2689                                 tp->rx_opt.rcv_wscale = rcv_wscale;
2690                                 tp->rx_opt.wscale_ok = 1;
2691                         }
2692                         break;
2693                 case TCPOPT_SACK_PERM:
2694                         if (opt.opt_val != 0)
2695                                 return -EINVAL;
2696
2697                         tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2698                         break;
2699                 case TCPOPT_TIMESTAMP:
2700                         if (opt.opt_val != 0)
2701                                 return -EINVAL;
2702
2703                         tp->rx_opt.tstamp_ok = 1;
2704                         break;
2705                 }
2706         }
2707
2708         return 0;
2709 }
2710
2711 /*
2712  *      Socket option code for TCP.
2713  */
2714 static int do_tcp_setsockopt(struct sock *sk, int level,
2715                 int optname, char __user *optval, unsigned int optlen)
2716 {
2717         struct tcp_sock *tp = tcp_sk(sk);
2718         struct inet_connection_sock *icsk = inet_csk(sk);
2719         struct net *net = sock_net(sk);
2720         int val;
2721         int err = 0;
2722
2723         /* These are data/string values, all the others are ints */
2724         switch (optname) {
2725         case TCP_CONGESTION: {
2726                 char name[TCP_CA_NAME_MAX];
2727
2728                 if (optlen < 1)
2729                         return -EINVAL;
2730
2731                 val = strncpy_from_user(name, optval,
2732                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2733                 if (val < 0)
2734                         return -EFAULT;
2735                 name[val] = 0;
2736
2737                 lock_sock(sk);
2738                 err = tcp_set_congestion_control(sk, name, true, true);
2739                 release_sock(sk);
2740                 return err;
2741         }
2742         case TCP_ULP: {
2743                 char name[TCP_ULP_NAME_MAX];
2744
2745                 if (optlen < 1)
2746                         return -EINVAL;
2747
2748                 val = strncpy_from_user(name, optval,
2749                                         min_t(long, TCP_ULP_NAME_MAX - 1,
2750                                               optlen));
2751                 if (val < 0)
2752                         return -EFAULT;
2753                 name[val] = 0;
2754
2755                 lock_sock(sk);
2756                 err = tcp_set_ulp(sk, name);
2757                 release_sock(sk);
2758                 return err;
2759         }
2760         case TCP_FASTOPEN_KEY: {
2761                 __u8 key[TCP_FASTOPEN_KEY_LENGTH];
2762
2763                 if (optlen != sizeof(key))
2764                         return -EINVAL;
2765
2766                 if (copy_from_user(key, optval, optlen))
2767                         return -EFAULT;
2768
2769                 return tcp_fastopen_reset_cipher(net, sk, key, sizeof(key));
2770         }
2771         default:
2772                 /* fallthru */
2773                 break;
2774         }
2775
2776         if (optlen < sizeof(int))
2777                 return -EINVAL;
2778
2779         if (get_user(val, (int __user *)optval))
2780                 return -EFAULT;
2781
2782         lock_sock(sk);
2783
2784         switch (optname) {
2785         case TCP_MAXSEG:
2786                 /* Values greater than interface MTU won't take effect. However
2787                  * at the point when this call is done we typically don't yet
2788                  * know which interface is going to be used
2789                  */
2790                 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
2791                         err = -EINVAL;
2792                         break;
2793                 }
2794                 tp->rx_opt.user_mss = val;
2795                 break;
2796
2797         case TCP_NODELAY:
2798                 if (val) {
2799                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2800                          * this option on corked socket is remembered, but
2801                          * it is not activated until cork is cleared.
2802                          *
2803                          * However, when TCP_NODELAY is set we make
2804                          * an explicit push, which overrides even TCP_CORK
2805                          * for currently queued segments.
2806                          */
2807                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2808                         tcp_push_pending_frames(sk);
2809                 } else {
2810                         tp->nonagle &= ~TCP_NAGLE_OFF;
2811                 }
2812                 break;
2813
2814         case TCP_THIN_LINEAR_TIMEOUTS:
2815                 if (val < 0 || val > 1)
2816                         err = -EINVAL;
2817                 else
2818                         tp->thin_lto = val;
2819                 break;
2820
2821         case TCP_THIN_DUPACK:
2822                 if (val < 0 || val > 1)
2823                         err = -EINVAL;
2824                 break;
2825
2826         case TCP_REPAIR:
2827                 if (!tcp_can_repair_sock(sk))
2828                         err = -EPERM;
2829                 else if (val == TCP_REPAIR_ON) {
2830                         tp->repair = 1;
2831                         sk->sk_reuse = SK_FORCE_REUSE;
2832                         tp->repair_queue = TCP_NO_QUEUE;
2833                 } else if (val == TCP_REPAIR_OFF) {
2834                         tp->repair = 0;
2835                         sk->sk_reuse = SK_NO_REUSE;
2836                         tcp_send_window_probe(sk);
2837                 } else if (val == TCP_REPAIR_OFF_NO_WP) {
2838                         tp->repair = 0;
2839                         sk->sk_reuse = SK_NO_REUSE;
2840                 } else
2841                         err = -EINVAL;
2842
2843                 break;
2844
2845         case TCP_REPAIR_QUEUE:
2846                 if (!tp->repair)
2847                         err = -EPERM;
2848                 else if ((unsigned int)val < TCP_QUEUES_NR)
2849                         tp->repair_queue = val;
2850                 else
2851                         err = -EINVAL;
2852                 break;
2853
2854         case TCP_QUEUE_SEQ:
2855                 if (sk->sk_state != TCP_CLOSE)
2856                         err = -EPERM;
2857                 else if (tp->repair_queue == TCP_SEND_QUEUE)
2858                         tp->write_seq = val;
2859                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2860                         tp->rcv_nxt = val;
2861                 else
2862                         err = -EINVAL;
2863                 break;
2864
2865         case TCP_REPAIR_OPTIONS:
2866                 if (!tp->repair)
2867                         err = -EINVAL;
2868                 else if (sk->sk_state == TCP_ESTABLISHED)
2869                         err = tcp_repair_options_est(sk,
2870                                         (struct tcp_repair_opt __user *)optval,
2871                                         optlen);
2872                 else
2873                         err = -EPERM;
2874                 break;
2875
2876         case TCP_CORK:
2877                 /* When set indicates to always queue non-full frames.
2878                  * Later the user clears this option and we transmit
2879                  * any pending partial frames in the queue.  This is
2880                  * meant to be used alongside sendfile() to get properly
2881                  * filled frames when the user (for example) must write
2882                  * out headers with a write() call first and then use
2883                  * sendfile to send out the data parts.
2884                  *
2885                  * TCP_CORK can be set together with TCP_NODELAY and it is
2886                  * stronger than TCP_NODELAY.
2887                  */
2888                 if (val) {
2889                         tp->nonagle |= TCP_NAGLE_CORK;
2890                 } else {
2891                         tp->nonagle &= ~TCP_NAGLE_CORK;
2892                         if (tp->nonagle&TCP_NAGLE_OFF)
2893                                 tp->nonagle |= TCP_NAGLE_PUSH;
2894                         tcp_push_pending_frames(sk);
2895                 }
2896                 break;
2897
2898         case TCP_KEEPIDLE:
2899                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2900                         err = -EINVAL;
2901                 else {
2902                         tp->keepalive_time = val * HZ;
2903                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2904                             !((1 << sk->sk_state) &
2905                               (TCPF_CLOSE | TCPF_LISTEN))) {
2906                                 u32 elapsed = keepalive_time_elapsed(tp);
2907                                 if (tp->keepalive_time > elapsed)
2908                                         elapsed = tp->keepalive_time - elapsed;
2909                                 else
2910                                         elapsed = 0;
2911                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2912                         }
2913                 }
2914                 break;
2915         case TCP_KEEPINTVL:
2916                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2917                         err = -EINVAL;
2918                 else
2919                         tp->keepalive_intvl = val * HZ;
2920                 break;
2921         case TCP_KEEPCNT:
2922                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2923                         err = -EINVAL;
2924                 else
2925                         tp->keepalive_probes = val;
2926                 break;
2927         case TCP_SYNCNT:
2928                 if (val < 1 || val > MAX_TCP_SYNCNT)
2929                         err = -EINVAL;
2930                 else
2931                         icsk->icsk_syn_retries = val;
2932                 break;
2933
2934         case TCP_SAVE_SYN:
2935                 if (val < 0 || val > 1)
2936                         err = -EINVAL;
2937                 else
2938                         tp->save_syn = val;
2939                 break;
2940
2941         case TCP_LINGER2:
2942                 if (val < 0)
2943                         tp->linger2 = -1;
2944                 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
2945                         tp->linger2 = 0;
2946                 else
2947                         tp->linger2 = val * HZ;
2948                 break;
2949
2950         case TCP_DEFER_ACCEPT:
2951                 /* Translate value in seconds to number of retransmits */
2952                 icsk->icsk_accept_queue.rskq_defer_accept =
2953                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2954                                         TCP_RTO_MAX / HZ);
2955                 break;
2956
2957         case TCP_WINDOW_CLAMP:
2958                 if (!val) {
2959                         if (sk->sk_state != TCP_CLOSE) {
2960                                 err = -EINVAL;
2961                                 break;
2962                         }
2963                         tp->window_clamp = 0;
2964                 } else
2965                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2966                                                 SOCK_MIN_RCVBUF / 2 : val;
2967                 break;
2968
2969         case TCP_QUICKACK:
2970                 if (!val) {
2971                         icsk->icsk_ack.pingpong = 1;
2972                 } else {
2973                         icsk->icsk_ack.pingpong = 0;
2974                         if ((1 << sk->sk_state) &
2975                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2976                             inet_csk_ack_scheduled(sk)) {
2977                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2978                                 tcp_cleanup_rbuf(sk, 1);
2979                                 if (!(val & 1))
2980                                         icsk->icsk_ack.pingpong = 1;
2981                         }
2982                 }
2983                 break;
2984
2985 #ifdef CONFIG_TCP_MD5SIG
2986         case TCP_MD5SIG:
2987         case TCP_MD5SIG_EXT:
2988                 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
2989                         err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
2990                 else
2991                         err = -EINVAL;
2992                 break;
2993 #endif
2994         case TCP_USER_TIMEOUT:
2995                 /* Cap the max time in ms TCP will retry or probe the window
2996                  * before giving up and aborting (ETIMEDOUT) a connection.
2997                  */
2998                 if (val < 0)
2999                         err = -EINVAL;
3000                 else
3001                         icsk->icsk_user_timeout = val;
3002                 break;
3003
3004         case TCP_FASTOPEN:
3005                 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3006                     TCPF_LISTEN))) {
3007                         tcp_fastopen_init_key_once(net);
3008
3009                         fastopen_queue_tune(sk, val);
3010                 } else {
3011                         err = -EINVAL;
3012                 }
3013                 break;
3014         case TCP_FASTOPEN_CONNECT:
3015                 if (val > 1 || val < 0) {
3016                         err = -EINVAL;
3017                 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
3018                         if (sk->sk_state == TCP_CLOSE)
3019                                 tp->fastopen_connect = val;
3020                         else
3021                                 err = -EINVAL;
3022                 } else {
3023                         err = -EOPNOTSUPP;
3024                 }
3025                 break;
3026         case TCP_FASTOPEN_NO_COOKIE:
3027                 if (val > 1 || val < 0)
3028                         err = -EINVAL;
3029                 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3030                         err = -EINVAL;
3031                 else
3032                         tp->fastopen_no_cookie = val;
3033                 break;
3034         case TCP_TIMESTAMP:
3035                 if (!tp->repair)
3036                         err = -EPERM;
3037                 else
3038                         tp->tsoffset = val - tcp_time_stamp_raw();
3039                 break;
3040         case TCP_REPAIR_WINDOW:
3041                 err = tcp_repair_set_window(tp, optval, optlen);
3042                 break;
3043         case TCP_NOTSENT_LOWAT:
3044                 tp->notsent_lowat = val;
3045                 sk->sk_write_space(sk);
3046                 break;
3047         case TCP_INQ:
3048                 if (val > 1 || val < 0)
3049                         err = -EINVAL;
3050                 else
3051                         tp->recvmsg_inq = val;
3052                 break;
3053         default:
3054                 err = -ENOPROTOOPT;
3055                 break;
3056         }
3057
3058         release_sock(sk);
3059         return err;
3060 }
3061
3062 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
3063                    unsigned int optlen)
3064 {
3065         const struct inet_connection_sock *icsk = inet_csk(sk);
3066
3067         if (level != SOL_TCP)
3068                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3069                                                      optval, optlen);
3070         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3071 }
3072 EXPORT_SYMBOL(tcp_setsockopt);
3073
3074 #ifdef CONFIG_COMPAT
3075 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
3076                           char __user *optval, unsigned int optlen)
3077 {
3078         if (level != SOL_TCP)
3079                 return inet_csk_compat_setsockopt(sk, level, optname,
3080                                                   optval, optlen);
3081         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3082 }
3083 EXPORT_SYMBOL(compat_tcp_setsockopt);
3084 #endif
3085
3086 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3087                                       struct tcp_info *info)
3088 {
3089         u64 stats[__TCP_CHRONO_MAX], total = 0;
3090         enum tcp_chrono i;
3091
3092         for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3093                 stats[i] = tp->chrono_stat[i - 1];
3094                 if (i == tp->chrono_type)
3095                         stats[i] += tcp_jiffies32 - tp->chrono_start;
3096                 stats[i] *= USEC_PER_SEC / HZ;
3097                 total += stats[i];
3098         }
3099
3100         info->tcpi_busy_time = total;
3101         info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3102         info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3103 }
3104
3105 /* Return information about state of tcp endpoint in API format. */
3106 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3107 {
3108         const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3109         const struct inet_connection_sock *icsk = inet_csk(sk);
3110         unsigned long rate;
3111         u32 now;
3112         u64 rate64;
3113         bool slow;
3114
3115         memset(info, 0, sizeof(*info));
3116         if (sk->sk_type != SOCK_STREAM)
3117                 return;
3118
3119         info->tcpi_state = inet_sk_state_load(sk);
3120
3121         /* Report meaningful fields for all TCP states, including listeners */
3122         rate = READ_ONCE(sk->sk_pacing_rate);
3123         rate64 = (rate != ~0UL) ? rate : ~0ULL;
3124         info->tcpi_pacing_rate = rate64;
3125
3126         rate = READ_ONCE(sk->sk_max_pacing_rate);
3127         rate64 = (rate != ~0UL) ? rate : ~0ULL;
3128         info->tcpi_max_pacing_rate = rate64;
3129
3130         info->tcpi_reordering = tp->reordering;
3131         info->tcpi_snd_cwnd = tp->snd_cwnd;
3132
3133         if (info->tcpi_state == TCP_LISTEN) {
3134                 /* listeners aliased fields :
3135                  * tcpi_unacked -> Number of children ready for accept()
3136                  * tcpi_sacked  -> max backlog
3137                  */
3138                 info->tcpi_unacked = sk->sk_ack_backlog;
3139                 info->tcpi_sacked = sk->sk_max_ack_backlog;
3140                 return;
3141         }
3142
3143         slow = lock_sock_fast(sk);
3144
3145         info->tcpi_ca_state = icsk->icsk_ca_state;
3146         info->tcpi_retransmits = icsk->icsk_retransmits;
3147         info->tcpi_probes = icsk->icsk_probes_out;
3148         info->tcpi_backoff = icsk->icsk_backoff;
3149
3150         if (tp->rx_opt.tstamp_ok)
3151                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3152         if (tcp_is_sack(tp))
3153                 info->tcpi_options |= TCPI_OPT_SACK;
3154         if (tp->rx_opt.wscale_ok) {
3155                 info->tcpi_options |= TCPI_OPT_WSCALE;
3156                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3157                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3158         }
3159
3160         if (tp->ecn_flags & TCP_ECN_OK)
3161                 info->tcpi_options |= TCPI_OPT_ECN;
3162         if (tp->ecn_flags & TCP_ECN_SEEN)
3163                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3164         if (tp->syn_data_acked)
3165                 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3166
3167         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3168         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3169         info->tcpi_snd_mss = tp->mss_cache;
3170         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3171
3172         info->tcpi_unacked = tp->packets_out;
3173         info->tcpi_sacked = tp->sacked_out;
3174
3175         info->tcpi_lost = tp->lost_out;
3176         info->tcpi_retrans = tp->retrans_out;
3177
3178         now = tcp_jiffies32;
3179         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3180         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3181         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3182
3183         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3184         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3185         info->tcpi_rtt = tp->srtt_us >> 3;
3186         info->tcpi_rttvar = tp->mdev_us >> 2;
3187         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3188         info->tcpi_advmss = tp->advmss;
3189
3190         info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3191         info->tcpi_rcv_space = tp->rcvq_space.space;
3192
3193         info->tcpi_total_retrans = tp->total_retrans;
3194
3195         info->tcpi_bytes_acked = tp->bytes_acked;
3196         info->tcpi_bytes_received = tp->bytes_received;
3197         info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3198         tcp_get_info_chrono_stats(tp, info);
3199
3200         info->tcpi_segs_out = tp->segs_out;
3201         info->tcpi_segs_in = tp->segs_in;
3202
3203         info->tcpi_min_rtt = tcp_min_rtt(tp);
3204         info->tcpi_data_segs_in = tp->data_segs_in;
3205         info->tcpi_data_segs_out = tp->data_segs_out;
3206
3207         info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3208         rate64 = tcp_compute_delivery_rate(tp);
3209         if (rate64)
3210                 info->tcpi_delivery_rate = rate64;
3211         info->tcpi_delivered = tp->delivered;
3212         info->tcpi_delivered_ce = tp->delivered_ce;
3213         info->tcpi_bytes_sent = tp->bytes_sent;
3214         info->tcpi_bytes_retrans = tp->bytes_retrans;
3215         info->tcpi_dsack_dups = tp->dsack_dups;
3216         info->tcpi_reord_seen = tp->reord_seen;
3217         unlock_sock_fast(sk, slow);
3218 }
3219 EXPORT_SYMBOL_GPL(tcp_get_info);
3220
3221 static size_t tcp_opt_stats_get_size(void)
3222 {
3223         return
3224                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3225                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3226                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3227                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3228                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3229                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3230                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3231                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3232                 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3233                 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3234                 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3235                 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3236                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3237                 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3238                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3239                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3240                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3241                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3242                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3243                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3244                 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3245                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3246                 0;
3247 }
3248
3249 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk)
3250 {
3251         const struct tcp_sock *tp = tcp_sk(sk);
3252         struct sk_buff *stats;
3253         struct tcp_info info;
3254         unsigned long rate;
3255         u64 rate64;
3256
3257         stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3258         if (!stats)
3259                 return NULL;
3260
3261         tcp_get_info_chrono_stats(tp, &info);
3262         nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3263                           info.tcpi_busy_time, TCP_NLA_PAD);
3264         nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3265                           info.tcpi_rwnd_limited, TCP_NLA_PAD);
3266         nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3267                           info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3268         nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3269                           tp->data_segs_out, TCP_NLA_PAD);
3270         nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3271                           tp->total_retrans, TCP_NLA_PAD);
3272
3273         rate = READ_ONCE(sk->sk_pacing_rate);
3274         rate64 = (rate != ~0UL) ? rate : ~0ULL;
3275         nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3276
3277         rate64 = tcp_compute_delivery_rate(tp);
3278         nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3279
3280         nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3281         nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3282         nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3283
3284         nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3285         nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3286         nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3287         nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3288         nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3289
3290         nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3291         nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3292
3293         nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3294                           TCP_NLA_PAD);
3295         nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3296                           TCP_NLA_PAD);
3297         nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3298         nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3299         nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3300
3301         return stats;
3302 }
3303
3304 static int do_tcp_getsockopt(struct sock *sk, int level,
3305                 int optname, char __user *optval, int __user *optlen)
3306 {
3307         struct inet_connection_sock *icsk = inet_csk(sk);
3308         struct tcp_sock *tp = tcp_sk(sk);
3309         struct net *net = sock_net(sk);
3310         int val, len;
3311
3312         if (get_user(len, optlen))
3313                 return -EFAULT;
3314
3315         len = min_t(unsigned int, len, sizeof(int));
3316
3317         if (len < 0)
3318                 return -EINVAL;
3319
3320         switch (optname) {
3321         case TCP_MAXSEG:
3322                 val = tp->mss_cache;
3323                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3324                         val = tp->rx_opt.user_mss;
3325                 if (tp->repair)
3326                         val = tp->rx_opt.mss_clamp;
3327                 break;
3328         case TCP_NODELAY:
3329                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3330                 break;
3331         case TCP_CORK:
3332                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3333                 break;
3334         case TCP_KEEPIDLE:
3335                 val = keepalive_time_when(tp) / HZ;
3336                 break;
3337         case TCP_KEEPINTVL:
3338                 val = keepalive_intvl_when(tp) / HZ;
3339                 break;
3340         case TCP_KEEPCNT:
3341                 val = keepalive_probes(tp);
3342                 break;
3343         case TCP_SYNCNT:
3344                 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3345                 break;
3346         case TCP_LINGER2:
3347                 val = tp->linger2;
3348                 if (val >= 0)
3349                         val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3350                 break;
3351         case TCP_DEFER_ACCEPT:
3352                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3353                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3354                 break;
3355         case TCP_WINDOW_CLAMP:
3356                 val = tp->window_clamp;
3357                 break;
3358         case TCP_INFO: {
3359                 struct tcp_info info;
3360
3361                 if (get_user(len, optlen))
3362                         return -EFAULT;
3363
3364                 tcp_get_info(sk, &info);
3365
3366                 len = min_t(unsigned int, len, sizeof(info));
3367                 if (put_user(len, optlen))
3368                         return -EFAULT;
3369                 if (copy_to_user(optval, &info, len))
3370                         return -EFAULT;
3371                 return 0;
3372         }
3373         case TCP_CC_INFO: {
3374                 const struct tcp_congestion_ops *ca_ops;
3375                 union tcp_cc_info info;
3376                 size_t sz = 0;
3377                 int attr;
3378
3379                 if (get_user(len, optlen))
3380                         return -EFAULT;
3381
3382                 ca_ops = icsk->icsk_ca_ops;
3383                 if (ca_ops && ca_ops->get_info)
3384                         sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3385
3386                 len = min_t(unsigned int, len, sz);
3387                 if (put_user(len, optlen))
3388                         return -EFAULT;
3389                 if (copy_to_user(optval, &info, len))
3390                         return -EFAULT;
3391                 return 0;
3392         }
3393         case TCP_QUICKACK:
3394                 val = !icsk->icsk_ack.pingpong;
3395                 break;
3396
3397         case TCP_CONGESTION:
3398                 if (get_user(len, optlen))
3399                         return -EFAULT;
3400                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
3401                 if (put_user(len, optlen))
3402                         return -EFAULT;
3403                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3404                         return -EFAULT;
3405                 return 0;
3406
3407         case TCP_ULP:
3408                 if (get_user(len, optlen))
3409                         return -EFAULT;
3410                 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
3411                 if (!icsk->icsk_ulp_ops) {
3412                         if (put_user(0, optlen))
3413                                 return -EFAULT;
3414                         return 0;
3415                 }
3416                 if (put_user(len, optlen))
3417                         return -EFAULT;
3418                 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
3419                         return -EFAULT;
3420                 return 0;
3421
3422         case TCP_FASTOPEN_KEY: {
3423                 __u8 key[TCP_FASTOPEN_KEY_LENGTH];
3424                 struct tcp_fastopen_context *ctx;
3425
3426                 if (get_user(len, optlen))
3427                         return -EFAULT;
3428
3429                 rcu_read_lock();
3430                 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
3431                 if (ctx)
3432                         memcpy(key, ctx->key, sizeof(key));
3433                 else
3434                         len = 0;
3435                 rcu_read_unlock();
3436
3437                 len = min_t(unsigned int, len, sizeof(key));
3438                 if (put_user(len, optlen))
3439                         return -EFAULT;
3440                 if (copy_to_user(optval, key, len))
3441                         return -EFAULT;
3442                 return 0;
3443         }
3444         case TCP_THIN_LINEAR_TIMEOUTS:
3445                 val = tp->thin_lto;
3446                 break;
3447
3448         case TCP_THIN_DUPACK:
3449                 val = 0;
3450                 break;
3451
3452         case TCP_REPAIR:
3453                 val = tp->repair;
3454                 break;
3455
3456         case TCP_REPAIR_QUEUE:
3457                 if (tp->repair)
3458                         val = tp->repair_queue;
3459                 else
3460                         return -EINVAL;
3461                 break;
3462
3463         case TCP_REPAIR_WINDOW: {
3464                 struct tcp_repair_window opt;
3465
3466                 if (get_user(len, optlen))
3467                         return -EFAULT;
3468
3469                 if (len != sizeof(opt))
3470                         return -EINVAL;
3471
3472                 if (!tp->repair)
3473                         return -EPERM;
3474
3475                 opt.snd_wl1     = tp->snd_wl1;
3476                 opt.snd_wnd     = tp->snd_wnd;
3477                 opt.max_window  = tp->max_window;
3478                 opt.rcv_wnd     = tp->rcv_wnd;
3479                 opt.rcv_wup     = tp->rcv_wup;
3480
3481                 if (copy_to_user(optval, &opt, len))
3482                         return -EFAULT;
3483                 return 0;
3484         }
3485         case TCP_QUEUE_SEQ:
3486                 if (tp->repair_queue == TCP_SEND_QUEUE)
3487                         val = tp->write_seq;
3488                 else if (tp->repair_queue == TCP_RECV_QUEUE)
3489                         val = tp->rcv_nxt;
3490                 else
3491                         return -EINVAL;
3492                 break;
3493
3494         case TCP_USER_TIMEOUT:
3495                 val = icsk->icsk_user_timeout;
3496                 break;
3497
3498         case TCP_FASTOPEN:
3499                 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3500                 break;
3501
3502         case TCP_FASTOPEN_CONNECT:
3503                 val = tp->fastopen_connect;
3504                 break;
3505
3506         case TCP_FASTOPEN_NO_COOKIE:
3507                 val = tp->fastopen_no_cookie;
3508                 break;
3509
3510         case TCP_TIMESTAMP:
3511                 val = tcp_time_stamp_raw() + tp->tsoffset;
3512                 break;
3513         case TCP_NOTSENT_LOWAT:
3514                 val = tp->notsent_lowat;
3515                 break;
3516         case TCP_INQ:
3517                 val = tp->recvmsg_inq;
3518                 break;
3519         case TCP_SAVE_SYN:
3520                 val = tp->save_syn;
3521                 break;
3522         case TCP_SAVED_SYN: {
3523                 if (get_user(len, optlen))
3524                         return -EFAULT;
3525
3526                 lock_sock(sk);
3527                 if (tp->saved_syn) {
3528                         if (len < tp->saved_syn[0]) {
3529                                 if (put_user(tp->saved_syn[0], optlen)) {
3530                                         release_sock(sk);
3531                                         return -EFAULT;
3532                                 }
3533                                 release_sock(sk);
3534                                 return -EINVAL;
3535                         }
3536                         len = tp->saved_syn[0];
3537                         if (put_user(len, optlen)) {
3538                                 release_sock(sk);
3539                                 return -EFAULT;
3540                         }
3541                         if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3542                                 release_sock(sk);
3543                                 return -EFAULT;
3544                         }
3545                         tcp_saved_syn_free(tp);
3546                         release_sock(sk);
3547                 } else {
3548                         release_sock(sk);
3549                         len = 0;
3550                         if (put_user(len, optlen))
3551                                 return -EFAULT;
3552                 }
3553                 return 0;
3554         }
3555 #ifdef CONFIG_MMU
3556         case TCP_ZEROCOPY_RECEIVE: {
3557                 struct tcp_zerocopy_receive zc;
3558                 int err;
3559
3560                 if (get_user(len, optlen))
3561                         return -EFAULT;
3562                 if (len != sizeof(zc))
3563                         return -EINVAL;
3564                 if (copy_from_user(&zc, optval, len))
3565                         return -EFAULT;
3566                 lock_sock(sk);
3567                 err = tcp_zerocopy_receive(sk, &zc);
3568                 release_sock(sk);
3569                 if (!err && copy_to_user(optval, &zc, len))
3570                         err = -EFAULT;
3571                 return err;
3572         }
3573 #endif
3574         default:
3575                 return -ENOPROTOOPT;
3576         }
3577
3578         if (put_user(len, optlen))
3579                 return -EFAULT;
3580         if (copy_to_user(optval, &val, len))
3581                 return -EFAULT;
3582         return 0;
3583 }
3584
3585 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3586                    int __user *optlen)
3587 {
3588         struct inet_connection_sock *icsk = inet_csk(sk);
3589
3590         if (level != SOL_TCP)
3591                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3592                                                      optval, optlen);
3593         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3594 }
3595 EXPORT_SYMBOL(tcp_getsockopt);
3596
3597 #ifdef CONFIG_COMPAT
3598 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3599                           char __user *optval, int __user *optlen)
3600 {
3601         if (level != SOL_TCP)
3602                 return inet_csk_compat_getsockopt(sk, level, optname,
3603                                                   optval, optlen);
3604         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3605 }
3606 EXPORT_SYMBOL(compat_tcp_getsockopt);
3607 #endif
3608
3609 #ifdef CONFIG_TCP_MD5SIG
3610 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3611 static DEFINE_MUTEX(tcp_md5sig_mutex);
3612 static bool tcp_md5sig_pool_populated = false;
3613
3614 static void __tcp_alloc_md5sig_pool(void)
3615 {
3616         struct crypto_ahash *hash;
3617         int cpu;
3618
3619         hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3620         if (IS_ERR(hash))
3621                 return;
3622
3623         for_each_possible_cpu(cpu) {
3624                 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3625                 struct ahash_request *req;
3626
3627                 if (!scratch) {
3628                         scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3629                                                sizeof(struct tcphdr),
3630                                                GFP_KERNEL,
3631                                                cpu_to_node(cpu));
3632                         if (!scratch)
3633                                 return;
3634                         per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3635                 }
3636                 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3637                         continue;
3638
3639                 req = ahash_request_alloc(hash, GFP_KERNEL);
3640                 if (!req)
3641                         return;
3642
3643                 ahash_request_set_callback(req, 0, NULL, NULL);
3644
3645                 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3646         }
3647         /* before setting tcp_md5sig_pool_populated, we must commit all writes
3648          * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3649          */
3650         smp_wmb();
3651         tcp_md5sig_pool_populated = true;
3652 }
3653
3654 bool tcp_alloc_md5sig_pool(void)
3655 {
3656         if (unlikely(!tcp_md5sig_pool_populated)) {
3657                 mutex_lock(&tcp_md5sig_mutex);
3658
3659                 if (!tcp_md5sig_pool_populated) {
3660                         __tcp_alloc_md5sig_pool();
3661                         if (tcp_md5sig_pool_populated)
3662                                 static_key_slow_inc(&tcp_md5_needed);
3663                 }
3664
3665                 mutex_unlock(&tcp_md5sig_mutex);
3666         }
3667         return tcp_md5sig_pool_populated;
3668 }
3669 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3670
3671
3672 /**
3673  *      tcp_get_md5sig_pool - get md5sig_pool for this user
3674  *
3675  *      We use percpu structure, so if we succeed, we exit with preemption
3676  *      and BH disabled, to make sure another thread or softirq handling
3677  *      wont try to get same context.
3678  */
3679 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3680 {
3681         local_bh_disable();
3682
3683         if (tcp_md5sig_pool_populated) {
3684                 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3685                 smp_rmb();
3686                 return this_cpu_ptr(&tcp_md5sig_pool);
3687         }
3688         local_bh_enable();
3689         return NULL;
3690 }
3691 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3692
3693 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3694                           const struct sk_buff *skb, unsigned int header_len)
3695 {
3696         struct scatterlist sg;
3697         const struct tcphdr *tp = tcp_hdr(skb);
3698         struct ahash_request *req = hp->md5_req;
3699         unsigned int i;
3700         const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3701                                            skb_headlen(skb) - header_len : 0;
3702         const struct skb_shared_info *shi = skb_shinfo(skb);
3703         struct sk_buff *frag_iter;
3704
3705         sg_init_table(&sg, 1);
3706
3707         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3708         ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3709         if (crypto_ahash_update(req))
3710                 return 1;
3711
3712         for (i = 0; i < shi->nr_frags; ++i) {
3713                 const struct skb_frag_struct *f = &shi->frags[i];
3714                 unsigned int offset = f->page_offset;
3715                 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3716
3717                 sg_set_page(&sg, page, skb_frag_size(f),
3718                             offset_in_page(offset));
3719                 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3720                 if (crypto_ahash_update(req))
3721                         return 1;
3722         }
3723
3724         skb_walk_frags(skb, frag_iter)
3725                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3726                         return 1;
3727
3728         return 0;
3729 }
3730 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3731
3732 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3733 {
3734         struct scatterlist sg;
3735
3736         sg_init_one(&sg, key->key, key->keylen);
3737         ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen);
3738         return crypto_ahash_update(hp->md5_req);
3739 }
3740 EXPORT_SYMBOL(tcp_md5_hash_key);
3741
3742 #endif
3743
3744 void tcp_done(struct sock *sk)
3745 {
3746         struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3747
3748         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3749                 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3750
3751         tcp_set_state(sk, TCP_CLOSE);
3752         tcp_clear_xmit_timers(sk);
3753         if (req)
3754                 reqsk_fastopen_remove(sk, req, false);
3755
3756         sk->sk_shutdown = SHUTDOWN_MASK;
3757
3758         if (!sock_flag(sk, SOCK_DEAD))
3759                 sk->sk_state_change(sk);
3760         else
3761                 inet_csk_destroy_sock(sk);
3762 }
3763 EXPORT_SYMBOL_GPL(tcp_done);
3764
3765 int tcp_abort(struct sock *sk, int err)
3766 {
3767         if (!sk_fullsock(sk)) {
3768                 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3769                         struct request_sock *req = inet_reqsk(sk);
3770
3771                         local_bh_disable();
3772                         inet_csk_reqsk_queue_drop(req->rsk_listener, req);
3773                         local_bh_enable();
3774                         return 0;
3775                 }
3776                 return -EOPNOTSUPP;
3777         }
3778
3779         /* Don't race with userspace socket closes such as tcp_close. */
3780         lock_sock(sk);
3781
3782         if (sk->sk_state == TCP_LISTEN) {
3783                 tcp_set_state(sk, TCP_CLOSE);
3784                 inet_csk_listen_stop(sk);
3785         }
3786
3787         /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3788         local_bh_disable();
3789         bh_lock_sock(sk);
3790
3791         if (!sock_flag(sk, SOCK_DEAD)) {
3792                 sk->sk_err = err;
3793                 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3794                 smp_wmb();
3795                 sk->sk_error_report(sk);
3796                 if (tcp_need_reset(sk->sk_state))
3797                         tcp_send_active_reset(sk, GFP_ATOMIC);
3798                 tcp_done(sk);
3799         }
3800
3801         bh_unlock_sock(sk);
3802         local_bh_enable();
3803         tcp_write_queue_purge(sk);
3804         release_sock(sk);
3805         return 0;
3806 }
3807 EXPORT_SYMBOL_GPL(tcp_abort);
3808
3809 extern struct tcp_congestion_ops tcp_reno;
3810
3811 static __initdata unsigned long thash_entries;
3812 static int __init set_thash_entries(char *str)
3813 {
3814         ssize_t ret;
3815
3816         if (!str)
3817                 return 0;
3818
3819         ret = kstrtoul(str, 0, &thash_entries);
3820         if (ret)
3821                 return 0;
3822
3823         return 1;
3824 }
3825 __setup("thash_entries=", set_thash_entries);
3826
3827 static void __init tcp_init_mem(void)
3828 {
3829         unsigned long limit = nr_free_buffer_pages() / 16;
3830
3831         limit = max(limit, 128UL);
3832         sysctl_tcp_mem[0] = limit / 4 * 3;              /* 4.68 % */
3833         sysctl_tcp_mem[1] = limit;                      /* 6.25 % */
3834         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;      /* 9.37 % */
3835 }
3836
3837 void __init tcp_init(void)
3838 {
3839         int max_rshare, max_wshare, cnt;
3840         unsigned long limit;
3841         unsigned int i;
3842
3843         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3844                      FIELD_SIZEOF(struct sk_buff, cb));
3845
3846         percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3847         percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3848         inet_hashinfo_init(&tcp_hashinfo);
3849         inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
3850                             thash_entries, 21,  /* one slot per 2 MB*/
3851                             0, 64 * 1024);
3852         tcp_hashinfo.bind_bucket_cachep =
3853                 kmem_cache_create("tcp_bind_bucket",
3854                                   sizeof(struct inet_bind_bucket), 0,
3855                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3856
3857         /* Size and allocate the main established and bind bucket
3858          * hash tables.
3859          *
3860          * The methodology is similar to that of the buffer cache.
3861          */
3862         tcp_hashinfo.ehash =
3863                 alloc_large_system_hash("TCP established",
3864                                         sizeof(struct inet_ehash_bucket),
3865                                         thash_entries,
3866                                         17, /* one slot per 128 KB of memory */
3867                                         0,
3868                                         NULL,
3869                                         &tcp_hashinfo.ehash_mask,
3870                                         0,
3871                                         thash_entries ? 0 : 512 * 1024);
3872         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3873                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3874
3875         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3876                 panic("TCP: failed to alloc ehash_locks");
3877         tcp_hashinfo.bhash =
3878                 alloc_large_system_hash("TCP bind",
3879                                         sizeof(struct inet_bind_hashbucket),
3880                                         tcp_hashinfo.ehash_mask + 1,
3881                                         17, /* one slot per 128 KB of memory */
3882                                         0,
3883                                         &tcp_hashinfo.bhash_size,
3884                                         NULL,
3885                                         0,
3886                                         64 * 1024);
3887         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3888         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3889                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3890                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3891         }
3892
3893
3894         cnt = tcp_hashinfo.ehash_mask + 1;
3895         sysctl_tcp_max_orphans = cnt / 2;
3896
3897         tcp_init_mem();
3898         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3899         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3900         max_wshare = min(4UL*1024*1024, limit);
3901         max_rshare = min(6UL*1024*1024, limit);
3902
3903         init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3904         init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
3905         init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3906
3907         init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3908         init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
3909         init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
3910
3911         pr_info("Hash tables configured (established %u bind %u)\n",
3912                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3913
3914         tcp_v4_init();
3915         tcp_metrics_init();
3916         BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3917         tcp_tasklet_init();
3918 }