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 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
 461 {
 462         struct sk_buff *skb = tcp_write_queue_tail(sk);
 463
 464         if (tsflags && skb) {
 465                 struct skb_shared_info *shinfo = skb_shinfo(skb);
 466                 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 467
 468                 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
 469                 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
 470                         tcb->txstamp_ack = 1;
 471                 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
 472                         shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
 473         }
 474 }
 475
 476 static inline bool tcp_stream_is_readable(const struct tcp_sock *tp,
 477                                           int target, struct sock *sk)
 478 {
 479         return (tp->rcv_nxt - tp->copied_seq >= target) ||
 480                 (sk->sk_prot->stream_memory_read ?
 481                 sk->sk_prot->stream_memory_read(sk) : false);
 482 }
 483
 484 /*
 485  *      Wait for a TCP event.
 486  *
 487  *      Note that we don't need to lock the socket, as the upper poll layers
 488  *      take care of normal races (between the test and the event) and we don't
 489  *      go look at any of the socket buffers directly.
 490  */
 491 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 492 {
 493         __poll_t mask;
 494         struct sock *sk = sock->sk;
 495         const struct tcp_sock *tp = tcp_sk(sk);
 496         int state;
 497
 498         sock_poll_wait(file, sock, wait);
 499
 500         state = inet_sk_state_load(sk);
 501         if (state == TCP_LISTEN)
 502                 return inet_csk_listen_poll(sk);
 503
 504         /* Socket is not locked. We are protected from async events
 505          * by poll logic and correct handling of state changes
 506          * made by other threads is impossible in any case.
 507          */
 508
 509         mask = 0;
 510
 511         /*
 512          * EPOLLHUP is certainly not done right. But poll() doesn't
 513          * have a notion of HUP in just one direction, and for a
 514          * socket the read side is more interesting.
 515          *
 516          * Some poll() documentation says that EPOLLHUP is incompatible
 517          * with the EPOLLOUT/POLLWR flags, so somebody should check this
 518          * all. But careful, it tends to be safer to return too many
 519          * bits than too few, and you can easily break real applications
 520          * if you don't tell them that something has hung up!
 521          *
 522          * Check-me.
 523          *
 524          * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
 525          * our fs/select.c). It means that after we received EOF,
 526          * poll always returns immediately, making impossible poll() on write()
 527          * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
 528          * if and only if shutdown has been made in both directions.
 529          * Actually, it is interesting to look how Solaris and DUX
 530          * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
 531          * then we could set it on SND_SHUTDOWN. BTW examples given
 532          * in Stevens' books assume exactly this behaviour, it explains
 533          * why EPOLLHUP is incompatible with EPOLLOUT.  --ANK
 534          *
 535          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 536          * blocking on fresh not-connected or disconnected socket. --ANK
 537          */
 538         if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
 539                 mask |= EPOLLHUP;
 540         if (sk->sk_shutdown & RCV_SHUTDOWN)
 541                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
 542
 543         /* Connected or passive Fast Open socket? */
 544         if (state != TCP_SYN_SENT &&
 545             (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
 546                 int target = sock_rcvlowat(sk, 0, INT_MAX);
 547
 548                 if (tp->urg_seq == tp->copied_seq &&
 549                     !sock_flag(sk, SOCK_URGINLINE) &&
 550                     tp->urg_data)
 551                         target++;
 552
 553                 if (tcp_stream_is_readable(tp, target, sk))
 554                         mask |= EPOLLIN | EPOLLRDNORM;
 555
 556                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 557                         if (sk_stream_is_writeable(sk)) {
 558                                 mask |= EPOLLOUT | EPOLLWRNORM;
 559                         } else {  /* send SIGIO later */
 560                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 561                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 562
 563                                 /* Race breaker. If space is freed after
 564                                  * wspace test but before the flags are set,
 565                                  * IO signal will be lost. Memory barrier
 566                                  * pairs with the input side.
 567                                  */
 568                                 smp_mb__after_atomic();
 569                                 if (sk_stream_is_writeable(sk))
 570                                         mask |= EPOLLOUT | EPOLLWRNORM;
 571                         }
 572                 } else
 573                         mask |= EPOLLOUT | EPOLLWRNORM;
 574
 575                 if (tp->urg_data & TCP_URG_VALID)
 576                         mask |= EPOLLPRI;
 577         } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
 578                 /* Active TCP fastopen socket with defer_connect
 579                  * Return EPOLLOUT so application can call write()
 580                  * in order for kernel to generate SYN+data
 581                  */
 582                 mask |= EPOLLOUT | EPOLLWRNORM;
 583         }
 584         /* This barrier is coupled with smp_wmb() in tcp_reset() */
 585         smp_rmb();
 586         if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
 587                 mask |= EPOLLERR;
 588
 589         return mask;
 590 }
 591 EXPORT_SYMBOL(tcp_poll);
 592
 593 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 594 {
 595         struct tcp_sock *tp = tcp_sk(sk);
 596         int answ;
 597         bool slow;
 598
 599         switch (cmd) {
 600         case SIOCINQ:
 601                 if (sk->sk_state == TCP_LISTEN)
 602                         return -EINVAL;
 603
 604                 slow = lock_sock_fast(sk);
 605                 answ = tcp_inq(sk);
 606                 unlock_sock_fast(sk, slow);
 607                 break;
 608         case SIOCATMARK:
 609                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 610                 break;
 611         case SIOCOUTQ:
 612                 if (sk->sk_state == TCP_LISTEN)
 613                         return -EINVAL;
 614
 615                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 616                         answ = 0;
 617                 else
 618                         answ = tp->write_seq - tp->snd_una;
 619                 break;
 620         case SIOCOUTQNSD:
 621                 if (sk->sk_state == TCP_LISTEN)
 622                         return -EINVAL;
 623
 624                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 625                         answ = 0;
 626                 else
 627                         answ = tp->write_seq - tp->snd_nxt;
 628                 break;
 629         default:
 630                 return -ENOIOCTLCMD;
 631         }
 632
 633         return put_user(answ, (int __user *)arg);
 634 }
 635 EXPORT_SYMBOL(tcp_ioctl);
 636
 637 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 638 {
 639         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 640         tp->pushed_seq = tp->write_seq;
 641 }
 642
 643 static inline bool forced_push(const struct tcp_sock *tp)
 644 {
 645         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 646 }
 647
 648 static void skb_entail(struct sock *sk, struct sk_buff *skb)
 649 {
 650         struct tcp_sock *tp = tcp_sk(sk);
 651         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 652
 653         skb->csum    = 0;
 654         tcb->seq     = tcb->end_seq = tp->write_seq;
 655         tcb->tcp_flags = TCPHDR_ACK;
 656         tcb->sacked  = 0;
 657         __skb_header_release(skb);
 658         tcp_add_write_queue_tail(sk, skb);
 659         sk->sk_wmem_queued += skb->truesize;
 660         sk_mem_charge(sk, skb->truesize);
 661         if (tp->nonagle & TCP_NAGLE_PUSH)
 662                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 663
 664         tcp_slow_start_after_idle_check(sk);
 665 }
 666
 667 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
 668 {
 669         if (flags & MSG_OOB)
 670                 tp->snd_up = tp->write_seq;
 671 }
 672
 673 /* If a not yet filled skb is pushed, do not send it if
 674  * we have data packets in Qdisc or NIC queues :
 675  * Because TX completion will happen shortly, it gives a chance
 676  * to coalesce future sendmsg() payload into this skb, without
 677  * need for a timer, and with no latency trade off.
 678  * As packets containing data payload have a bigger truesize
 679  * than pure acks (dataless) packets, the last checks prevent
 680  * autocorking if we only have an ACK in Qdisc/NIC queues,
 681  * or if TX completion was delayed after we processed ACK packet.
 682  */
 683 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
 684                                 int size_goal)
 685 {
 686         return skb->len < size_goal &&
 687                sock_net(sk)->ipv4.sysctl_tcp_autocorking &&
 688                !tcp_rtx_queue_empty(sk) &&
 689                refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
 690 }
 691
 692 static void tcp_push(struct sock *sk, int flags, int mss_now,
 693                      int nonagle, int size_goal)
 694 {
 695         struct tcp_sock *tp = tcp_sk(sk);
 696         struct sk_buff *skb;
 697
 698         skb = tcp_write_queue_tail(sk);
 699         if (!skb)
 700                 return;
 701         if (!(flags & MSG_MORE) || forced_push(tp))
 702                 tcp_mark_push(tp, skb);
 703
 704         tcp_mark_urg(tp, flags);
 705
 706         if (tcp_should_autocork(sk, skb, size_goal)) {
 707
 708                 /* avoid atomic op if TSQ_THROTTLED bit is already set */
 709                 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
 710                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
 711                         set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
 712                 }
 713                 /* It is possible TX completion already happened
 714                  * before we set TSQ_THROTTLED.
 715                  */
 716                 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
 717                         return;
 718         }
 719
 720         if (flags & MSG_MORE)
 721                 nonagle = TCP_NAGLE_CORK;
 722
 723         __tcp_push_pending_frames(sk, mss_now, nonagle);
 724 }
 725
 726 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 727                                 unsigned int offset, size_t len)
 728 {
 729         struct tcp_splice_state *tss = rd_desc->arg.data;
 730         int ret;
 731
 732         ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
 733                               min(rd_desc->count, len), tss->flags);
 734         if (ret > 0)
 735                 rd_desc->count -= ret;
 736         return ret;
 737 }
 738
 739 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 740 {
 741         /* Store TCP splice context information in read_descriptor_t. */
 742         read_descriptor_t rd_desc = {
 743                 .arg.data = tss,
 744                 .count    = tss->len,
 745         };
 746
 747         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 748 }
 749
 750 /**
 751  *  tcp_splice_read - splice data from TCP socket to a pipe
 752  * @sock:       socket to splice from
 753  * @ppos:       position (not valid)
 754  * @pipe:       pipe to splice to
 755  * @len:        number of bytes to splice
 756  * @flags:      splice modifier flags
 757  *
 758  * Description:
 759  *    Will read pages from given socket and fill them into a pipe.
 760  *
 761  **/
 762 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 763                         struct pipe_inode_info *pipe, size_t len,
 764                         unsigned int flags)
 765 {
 766         struct sock *sk = sock->sk;
 767         struct tcp_splice_state tss = {
 768                 .pipe = pipe,
 769                 .len = len,
 770                 .flags = flags,
 771         };
 772         long timeo;
 773         ssize_t spliced;
 774         int ret;
 775
 776         sock_rps_record_flow(sk);
 777         /*
 778          * We can't seek on a socket input
 779          */
 780         if (unlikely(*ppos))
 781                 return -ESPIPE;
 782
 783         ret = spliced = 0;
 784
 785         lock_sock(sk);
 786
 787         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 788         while (tss.len) {
 789                 ret = __tcp_splice_read(sk, &tss);
 790                 if (ret < 0)
 791                         break;
 792                 else if (!ret) {
 793                         if (spliced)
 794                                 break;
 795                         if (sock_flag(sk, SOCK_DONE))
 796                                 break;
 797                         if (sk->sk_err) {
 798                                 ret = sock_error(sk);
 799                                 break;
 800                         }
 801                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 802                                 break;
 803                         if (sk->sk_state == TCP_CLOSE) {
 804                                 /*
 805                                  * This occurs when user tries to read
 806                                  * from never connected socket.
 807                                  */
 808                                 ret = -ENOTCONN;
 809                                 break;
 810                         }
 811                         if (!timeo) {
 812                                 ret = -EAGAIN;
 813                                 break;
 814                         }
 815                         /* if __tcp_splice_read() got nothing while we have
 816                          * an skb in receive queue, we do not want to loop.
 817                          * This might happen with URG data.
 818                          */
 819                         if (!skb_queue_empty(&sk->sk_receive_queue))
 820                                 break;
 821                         sk_wait_data(sk, &timeo, NULL);
 822                         if (signal_pending(current)) {
 823                                 ret = sock_intr_errno(timeo);
 824                                 break;
 825                         }
 826                         continue;
 827                 }
 828                 tss.len -= ret;
 829                 spliced += ret;
 830
 831                 if (!timeo)
 832                         break;
 833                 release_sock(sk);
 834                 lock_sock(sk);
 835
 836                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 837                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 838                     signal_pending(current))
 839                         break;
 840         }
 841
 842         release_sock(sk);
 843
 844         if (spliced)
 845                 return spliced;
 846
 847         return ret;
 848 }
 849 EXPORT_SYMBOL(tcp_splice_read);
 850
 851 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
 852                                     bool force_schedule)
 853 {
 854         struct sk_buff *skb;
 855
 856         if (likely(!size)) {
 857                 skb = sk->sk_tx_skb_cache;
 858                 if (skb && !skb_cloned(skb)) {
 859                         skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
 860                         sk->sk_tx_skb_cache = NULL;
 861                         pskb_trim(skb, 0);
 862                         INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
 863                         skb_shinfo(skb)->tx_flags = 0;
 864                         memset(TCP_SKB_CB(skb), 0, sizeof(struct tcp_skb_cb));
 865                         return skb;
 866                 }
 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         if (IS_ENABLED(CONFIG_DEBUG_VM) &&
 947             WARN_ONCE(PageSlab(page), "page must not be a Slab one"))
 948                 return -EINVAL;
 949
 950         /* Wait for a connection to finish. One exception is TCP Fast Open
 951          * (passive side) where data is allowed to be sent before a connection
 952          * is fully established.
 953          */
 954         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
 955             !tcp_passive_fastopen(sk)) {
 956                 err = sk_stream_wait_connect(sk, &timeo);
 957                 if (err != 0)
 958                         goto out_err;
 959         }
 960
 961         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 962
 963         mss_now = tcp_send_mss(sk, &size_goal, flags);
 964         copied = 0;
 965
 966         err = -EPIPE;
 967         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 968                 goto out_err;
 969
 970         while (size > 0) {
 971                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 972                 int copy, i;
 973                 bool can_coalesce;
 974
 975                 if (!skb || (copy = size_goal - skb->len) <= 0 ||
 976                     !tcp_skb_can_collapse_to(skb)) {
 977 new_segment:
 978                         if (!sk_stream_memory_free(sk))
 979                                 goto wait_for_sndbuf;
 980
 981                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
 982                                         tcp_rtx_and_write_queues_empty(sk));
 983                         if (!skb)
 984                                 goto wait_for_memory;
 985
 986                         skb_entail(sk, skb);
 987                         copy = size_goal;
 988                 }
 989
 990                 if (copy > size)
 991                         copy = size;
 992
 993                 i = skb_shinfo(skb)->nr_frags;
 994                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 995                 if (!can_coalesce && i >= sysctl_max_skb_frags) {
 996                         tcp_mark_push(tp, skb);
 997                         goto new_segment;
 998                 }
 999                 if (!sk_wmem_schedule(sk, copy))
1000                         goto wait_for_memory;
1001
1002                 if (can_coalesce) {
1003                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1004                 } else {
1005                         get_page(page);
1006                         skb_fill_page_desc(skb, i, page, offset, copy);
1007                 }
1008
1009                 if (!(flags & MSG_NO_SHARED_FRAGS))
1010                         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1011
1012                 skb->len += copy;
1013                 skb->data_len += copy;
1014                 skb->truesize += copy;
1015                 sk->sk_wmem_queued += copy;
1016                 sk_mem_charge(sk, copy);
1017                 skb->ip_summed = CHECKSUM_PARTIAL;
1018                 tp->write_seq += copy;
1019                 TCP_SKB_CB(skb)->end_seq += copy;
1020                 tcp_skb_pcount_set(skb, 0);
1021
1022                 if (!copied)
1023                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1024
1025                 copied += copy;
1026                 offset += copy;
1027                 size -= copy;
1028                 if (!size)
1029                         goto out;
1030
1031                 if (skb->len < size_goal || (flags & MSG_OOB))
1032                         continue;
1033
1034                 if (forced_push(tp)) {
1035                         tcp_mark_push(tp, skb);
1036                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1037                 } else if (skb == tcp_send_head(sk))
1038                         tcp_push_one(sk, mss_now);
1039                 continue;
1040
1041 wait_for_sndbuf:
1042                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1043 wait_for_memory:
1044                 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1045                          TCP_NAGLE_PUSH, size_goal);
1046
1047                 err = sk_stream_wait_memory(sk, &timeo);
1048                 if (err != 0)
1049                         goto do_error;
1050
1051                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1052         }
1053
1054 out:
1055         if (copied) {
1056                 tcp_tx_timestamp(sk, sk->sk_tsflags);
1057                 if (!(flags & MSG_SENDPAGE_NOTLAST))
1058                         tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1059         }
1060         return copied;
1061
1062 do_error:
1063         if (copied)
1064                 goto out;
1065 out_err:
1066         /* make sure we wake any epoll edge trigger waiter */
1067         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1068                      err == -EAGAIN)) {
1069                 sk->sk_write_space(sk);
1070                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1071         }
1072         return sk_stream_error(sk, flags, err);
1073 }
1074 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1075
1076 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1077                         size_t size, int flags)
1078 {
1079         if (!(sk->sk_route_caps & NETIF_F_SG))
1080                 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1081
1082         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1083
1084         return do_tcp_sendpages(sk, page, offset, size, flags);
1085 }
1086 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1087
1088 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1089                  size_t size, int flags)
1090 {
1091         int ret;
1092
1093         lock_sock(sk);
1094         ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1095         release_sock(sk);
1096
1097         return ret;
1098 }
1099 EXPORT_SYMBOL(tcp_sendpage);
1100
1101 void tcp_free_fastopen_req(struct tcp_sock *tp)
1102 {
1103         if (tp->fastopen_req) {
1104                 kfree(tp->fastopen_req);
1105                 tp->fastopen_req = NULL;
1106         }
1107 }
1108
1109 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1110                                 int *copied, size_t size,
1111                                 struct ubuf_info *uarg)
1112 {
1113         struct tcp_sock *tp = tcp_sk(sk);
1114         struct inet_sock *inet = inet_sk(sk);
1115         struct sockaddr *uaddr = msg->msg_name;
1116         int err, flags;
1117
1118         if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1119             (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1120              uaddr->sa_family == AF_UNSPEC))
1121                 return -EOPNOTSUPP;
1122         if (tp->fastopen_req)
1123                 return -EALREADY; /* Another Fast Open is in progress */
1124
1125         tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1126                                    sk->sk_allocation);
1127         if (unlikely(!tp->fastopen_req))
1128                 return -ENOBUFS;
1129         tp->fastopen_req->data = msg;
1130         tp->fastopen_req->size = size;
1131         tp->fastopen_req->uarg = uarg;
1132
1133         if (inet->defer_connect) {
1134                 err = tcp_connect(sk);
1135                 /* Same failure procedure as in tcp_v4/6_connect */
1136                 if (err) {
1137                         tcp_set_state(sk, TCP_CLOSE);
1138                         inet->inet_dport = 0;
1139                         sk->sk_route_caps = 0;
1140                 }
1141         }
1142         flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1143         err = __inet_stream_connect(sk->sk_socket, uaddr,
1144                                     msg->msg_namelen, flags, 1);
1145         /* fastopen_req could already be freed in __inet_stream_connect
1146          * if the connection times out or gets rst
1147          */
1148         if (tp->fastopen_req) {
1149                 *copied = tp->fastopen_req->copied;
1150                 tcp_free_fastopen_req(tp);
1151                 inet->defer_connect = 0;
1152         }
1153         return err;
1154 }
1155
1156 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1157 {
1158         struct tcp_sock *tp = tcp_sk(sk);
1159         struct ubuf_info *uarg = NULL;
1160         struct sk_buff *skb;
1161         struct sockcm_cookie sockc;
1162         int flags, err, copied = 0;
1163         int mss_now = 0, size_goal, copied_syn = 0;
1164         bool process_backlog = false;
1165         bool zc = false;
1166         long timeo;
1167
1168         flags = msg->msg_flags;
1169
1170         if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
1171                 skb = tcp_write_queue_tail(sk);
1172                 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
1173                 if (!uarg) {
1174                         err = -ENOBUFS;
1175                         goto out_err;
1176                 }
1177
1178                 zc = sk->sk_route_caps & NETIF_F_SG;
1179                 if (!zc)
1180                         uarg->zerocopy = 0;
1181         }
1182
1183         if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1184             !tp->repair) {
1185                 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1186                 if (err == -EINPROGRESS && copied_syn > 0)
1187                         goto out;
1188                 else if (err)
1189                         goto out_err;
1190         }
1191
1192         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1193
1194         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1195
1196         /* Wait for a connection to finish. One exception is TCP Fast Open
1197          * (passive side) where data is allowed to be sent before a connection
1198          * is fully established.
1199          */
1200         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1201             !tcp_passive_fastopen(sk)) {
1202                 err = sk_stream_wait_connect(sk, &timeo);
1203                 if (err != 0)
1204                         goto do_error;
1205         }
1206
1207         if (unlikely(tp->repair)) {
1208                 if (tp->repair_queue == TCP_RECV_QUEUE) {
1209                         copied = tcp_send_rcvq(sk, msg, size);
1210                         goto out_nopush;
1211                 }
1212
1213                 err = -EINVAL;
1214                 if (tp->repair_queue == TCP_NO_QUEUE)
1215                         goto out_err;
1216
1217                 /* 'common' sending to sendq */
1218         }
1219
1220         sockcm_init(&sockc, sk);
1221         if (msg->msg_controllen) {
1222                 err = sock_cmsg_send(sk, msg, &sockc);
1223                 if (unlikely(err)) {
1224                         err = -EINVAL;
1225                         goto out_err;
1226                 }
1227         }
1228
1229         /* This should be in poll */
1230         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1231
1232         /* Ok commence sending. */
1233         copied = 0;
1234
1235 restart:
1236         mss_now = tcp_send_mss(sk, &size_goal, flags);
1237
1238         err = -EPIPE;
1239         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1240                 goto do_error;
1241
1242         while (msg_data_left(msg)) {
1243                 int copy = 0;
1244
1245                 skb = tcp_write_queue_tail(sk);
1246                 if (skb)
1247                         copy = size_goal - skb->len;
1248
1249                 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1250                         bool first_skb;
1251
1252 new_segment:
1253                         if (!sk_stream_memory_free(sk))
1254                                 goto wait_for_sndbuf;
1255
1256                         if (process_backlog && sk_flush_backlog(sk)) {
1257                                 process_backlog = false;
1258                                 goto restart;
1259                         }
1260                         first_skb = tcp_rtx_and_write_queues_empty(sk);
1261                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
1262                                                   first_skb);
1263                         if (!skb)
1264                                 goto wait_for_memory;
1265
1266                         process_backlog = true;
1267                         skb->ip_summed = CHECKSUM_PARTIAL;
1268
1269                         skb_entail(sk, skb);
1270                         copy = size_goal;
1271
1272                         /* All packets are restored as if they have
1273                          * already been sent. skb_mstamp_ns isn't set to
1274                          * avoid wrong rtt estimation.
1275                          */
1276                         if (tp->repair)
1277                                 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1278                 }
1279
1280                 /* Try to append data to the end of skb. */
1281                 if (copy > msg_data_left(msg))
1282                         copy = msg_data_left(msg);
1283
1284                 /* Where to copy to? */
1285                 if (skb_availroom(skb) > 0 && !zc) {
1286                         /* We have some space in skb head. Superb! */
1287                         copy = min_t(int, copy, skb_availroom(skb));
1288                         err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1289                         if (err)
1290                                 goto do_fault;
1291                 } else if (!zc) {
1292                         bool merge = true;
1293                         int i = skb_shinfo(skb)->nr_frags;
1294                         struct page_frag *pfrag = sk_page_frag(sk);
1295
1296                         if (!sk_page_frag_refill(sk, pfrag))
1297                                 goto wait_for_memory;
1298
1299                         if (!skb_can_coalesce(skb, i, pfrag->page,
1300                                               pfrag->offset)) {
1301                                 if (i >= sysctl_max_skb_frags) {
1302                                         tcp_mark_push(tp, skb);
1303                                         goto new_segment;
1304                                 }
1305                                 merge = false;
1306                         }
1307
1308                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1309
1310                         if (!sk_wmem_schedule(sk, copy))
1311                                 goto wait_for_memory;
1312
1313                         err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1314                                                        pfrag->page,
1315                                                        pfrag->offset,
1316                                                        copy);
1317                         if (err)
1318                                 goto do_error;
1319
1320                         /* Update the skb. */
1321                         if (merge) {
1322                                 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1323                         } else {
1324                                 skb_fill_page_desc(skb, i, pfrag->page,
1325                                                    pfrag->offset, copy);
1326                                 page_ref_inc(pfrag->page);
1327                         }
1328                         pfrag->offset += copy;
1329                 } else {
1330                         err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1331                         if (err == -EMSGSIZE || err == -EEXIST) {
1332                                 tcp_mark_push(tp, skb);
1333                                 goto new_segment;
1334                         }
1335                         if (err < 0)
1336                                 goto do_error;
1337                         copy = err;
1338                 }
1339
1340                 if (!copied)
1341                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1342
1343                 tp->write_seq += copy;
1344                 TCP_SKB_CB(skb)->end_seq += copy;
1345                 tcp_skb_pcount_set(skb, 0);
1346
1347                 copied += copy;
1348                 if (!msg_data_left(msg)) {
1349                         if (unlikely(flags & MSG_EOR))
1350                                 TCP_SKB_CB(skb)->eor = 1;
1351                         goto out;
1352                 }
1353
1354                 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1355                         continue;
1356
1357                 if (forced_push(tp)) {
1358                         tcp_mark_push(tp, skb);
1359                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1360                 } else if (skb == tcp_send_head(sk))
1361                         tcp_push_one(sk, mss_now);
1362                 continue;
1363
1364 wait_for_sndbuf:
1365                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1366 wait_for_memory:
1367                 if (copied)
1368                         tcp_push(sk, flags & ~MSG_MORE, mss_now,
1369                                  TCP_NAGLE_PUSH, size_goal);
1370
1371                 err = sk_stream_wait_memory(sk, &timeo);
1372                 if (err != 0)
1373                         goto do_error;
1374
1375                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1376         }
1377
1378 out:
1379         if (copied) {
1380                 tcp_tx_timestamp(sk, sockc.tsflags);
1381                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1382         }
1383 out_nopush:
1384         sock_zerocopy_put(uarg);
1385         return copied + copied_syn;
1386
1387 do_fault:
1388         if (!skb->len) {
1389                 tcp_unlink_write_queue(skb, sk);
1390                 /* It is the one place in all of TCP, except connection
1391                  * reset, where we can be unlinking the send_head.
1392                  */
1393                 if (tcp_write_queue_empty(sk))
1394                         tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
1395                 sk_wmem_free_skb(sk, skb);
1396         }
1397
1398 do_error:
1399         if (copied + copied_syn)
1400                 goto out;
1401 out_err:
1402         sock_zerocopy_put_abort(uarg, true);
1403         err = sk_stream_error(sk, flags, err);
1404         /* make sure we wake any epoll edge trigger waiter */
1405         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1406                      err == -EAGAIN)) {
1407                 sk->sk_write_space(sk);
1408                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1409         }
1410         return err;
1411 }
1412 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1413
1414 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1415 {
1416         int ret;
1417
1418         lock_sock(sk);
1419         ret = tcp_sendmsg_locked(sk, msg, size);
1420         release_sock(sk);
1421
1422         return ret;
1423 }
1424 EXPORT_SYMBOL(tcp_sendmsg);
1425
1426 /*
1427  *      Handle reading urgent data. BSD has very simple semantics for
1428  *      this, no blocking and very strange errors 8)
1429  */
1430
1431 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1432 {
1433         struct tcp_sock *tp = tcp_sk(sk);
1434
1435         /* No URG data to read. */
1436         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1437             tp->urg_data == TCP_URG_READ)
1438                 return -EINVAL; /* Yes this is right ! */
1439
1440         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1441                 return -ENOTCONN;
1442
1443         if (tp->urg_data & TCP_URG_VALID) {
1444                 int err = 0;
1445                 char c = tp->urg_data;
1446
1447                 if (!(flags & MSG_PEEK))
1448                         tp->urg_data = TCP_URG_READ;
1449
1450                 /* Read urgent data. */
1451                 msg->msg_flags |= MSG_OOB;
1452
1453                 if (len > 0) {
1454                         if (!(flags & MSG_TRUNC))
1455                                 err = memcpy_to_msg(msg, &c, 1);
1456                         len = 1;
1457                 } else
1458                         msg->msg_flags |= MSG_TRUNC;
1459
1460                 return err ? -EFAULT : len;
1461         }
1462
1463         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1464                 return 0;
1465
1466         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1467          * the available implementations agree in this case:
1468          * this call should never block, independent of the
1469          * blocking state of the socket.
1470          * Mike <pall@rz.uni-karlsruhe.de>
1471          */
1472         return -EAGAIN;
1473 }
1474
1475 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1476 {
1477         struct sk_buff *skb;
1478         int copied = 0, err = 0;
1479
1480         /* XXX -- need to support SO_PEEK_OFF */
1481
1482         skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1483                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1484                 if (err)
1485                         return err;
1486                 copied += skb->len;
1487         }
1488
1489         skb_queue_walk(&sk->sk_write_queue, skb) {
1490                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1491                 if (err)
1492                         break;
1493
1494                 copied += skb->len;
1495         }
1496
1497         return err ?: copied;
1498 }
1499
1500 /* Clean up the receive buffer for full frames taken by the user,
1501  * then send an ACK if necessary.  COPIED is the number of bytes
1502  * tcp_recvmsg has given to the user so far, it speeds up the
1503  * calculation of whether or not we must ACK for the sake of
1504  * a window update.
1505  */
1506 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1507 {
1508         struct tcp_sock *tp = tcp_sk(sk);
1509         bool time_to_ack = false;
1510
1511         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1512
1513         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1514              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1515              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1516
1517         if (inet_csk_ack_scheduled(sk)) {
1518                 const struct inet_connection_sock *icsk = inet_csk(sk);
1519                    /* Delayed ACKs frequently hit locked sockets during bulk
1520                     * receive. */
1521                 if (icsk->icsk_ack.blocked ||
1522                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1523                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1524                     /*
1525                      * If this read emptied read buffer, we send ACK, if
1526                      * connection is not bidirectional, user drained
1527                      * receive buffer and there was a small segment
1528                      * in queue.
1529                      */
1530                     (copied > 0 &&
1531                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1532                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1533                        !inet_csk_in_pingpong_mode(sk))) &&
1534                       !atomic_read(&sk->sk_rmem_alloc)))
1535                         time_to_ack = true;
1536         }
1537
1538         /* We send an ACK if we can now advertise a non-zero window
1539          * which has been raised "significantly".
1540          *
1541          * Even if window raised up to infinity, do not send window open ACK
1542          * in states, where we will not receive more. It is useless.
1543          */
1544         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1545                 __u32 rcv_window_now = tcp_receive_window(tp);
1546
1547                 /* Optimize, __tcp_select_window() is not cheap. */
1548                 if (2*rcv_window_now <= tp->window_clamp) {
1549                         __u32 new_window = __tcp_select_window(sk);
1550
1551                         /* Send ACK now, if this read freed lots of space
1552                          * in our buffer. Certainly, new_window is new window.
1553                          * We can advertise it now, if it is not less than current one.
1554                          * "Lots" means "at least twice" here.
1555                          */
1556                         if (new_window && new_window >= 2 * rcv_window_now)
1557                                 time_to_ack = true;
1558                 }
1559         }
1560         if (time_to_ack)
1561                 tcp_send_ack(sk);
1562 }
1563
1564 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1565 {
1566         struct sk_buff *skb;
1567         u32 offset;
1568
1569         while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1570                 offset = seq - TCP_SKB_CB(skb)->seq;
1571                 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1572                         pr_err_once("%s: found a SYN, please report !\n", __func__);
1573                         offset--;
1574                 }
1575                 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1576                         *off = offset;
1577                         return skb;
1578                 }
1579                 /* This looks weird, but this can happen if TCP collapsing
1580                  * splitted a fat GRO packet, while we released socket lock
1581                  * in skb_splice_bits()
1582                  */
1583                 sk_eat_skb(sk, skb);
1584         }
1585         return NULL;
1586 }
1587
1588 /*
1589  * This routine provides an alternative to tcp_recvmsg() for routines
1590  * that would like to handle copying from skbuffs directly in 'sendfile'
1591  * fashion.
1592  * Note:
1593  *      - It is assumed that the socket was locked by the caller.
1594  *      - The routine does not block.
1595  *      - At present, there is no support for reading OOB data
1596  *        or for 'peeking' the socket using this routine
1597  *        (although both would be easy to implement).
1598  */
1599 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1600                   sk_read_actor_t recv_actor)
1601 {
1602         struct sk_buff *skb;
1603         struct tcp_sock *tp = tcp_sk(sk);
1604         u32 seq = tp->copied_seq;
1605         u32 offset;
1606         int copied = 0;
1607
1608         if (sk->sk_state == TCP_LISTEN)
1609                 return -ENOTCONN;
1610         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1611                 if (offset < skb->len) {
1612                         int used;
1613                         size_t len;
1614
1615                         len = skb->len - offset;
1616                         /* Stop reading if we hit a patch of urgent data */
1617                         if (tp->urg_data) {
1618                                 u32 urg_offset = tp->urg_seq - seq;
1619                                 if (urg_offset < len)
1620                                         len = urg_offset;
1621                                 if (!len)
1622                                         break;
1623                         }
1624                         used = recv_actor(desc, skb, offset, len);
1625                         if (used <= 0) {
1626                                 if (!copied)
1627                                         copied = used;
1628                                 break;
1629                         } else if (used <= len) {
1630                                 seq += used;
1631                                 copied += used;
1632                                 offset += used;
1633                         }
1634                         /* If recv_actor drops the lock (e.g. TCP splice
1635                          * receive) the skb pointer might be invalid when
1636                          * getting here: tcp_collapse might have deleted it
1637                          * while aggregating skbs from the socket queue.
1638                          */
1639                         skb = tcp_recv_skb(sk, seq - 1, &offset);
1640                         if (!skb)
1641                                 break;
1642                         /* TCP coalescing might have appended data to the skb.
1643                          * Try to splice more frags
1644                          */
1645                         if (offset + 1 != skb->len)
1646                                 continue;
1647                 }
1648                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1649                         sk_eat_skb(sk, skb);
1650                         ++seq;
1651                         break;
1652                 }
1653                 sk_eat_skb(sk, skb);
1654                 if (!desc->count)
1655                         break;
1656                 tp->copied_seq = seq;
1657         }
1658         tp->copied_seq = seq;
1659
1660         tcp_rcv_space_adjust(sk);
1661
1662         /* Clean up data we have read: This will do ACK frames. */
1663         if (copied > 0) {
1664                 tcp_recv_skb(sk, seq, &offset);
1665                 tcp_cleanup_rbuf(sk, copied);
1666         }
1667         return copied;
1668 }
1669 EXPORT_SYMBOL(tcp_read_sock);
1670
1671 int tcp_peek_len(struct socket *sock)
1672 {
1673         return tcp_inq(sock->sk);
1674 }
1675 EXPORT_SYMBOL(tcp_peek_len);
1676
1677 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1678 int tcp_set_rcvlowat(struct sock *sk, int val)
1679 {
1680         int cap;
1681
1682         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1683                 cap = sk->sk_rcvbuf >> 1;
1684         else
1685                 cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1;
1686         val = min(val, cap);
1687         sk->sk_rcvlowat = val ? : 1;
1688
1689         /* Check if we need to signal EPOLLIN right now */
1690         tcp_data_ready(sk);
1691
1692         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1693                 return 0;
1694
1695         val <<= 1;
1696         if (val > sk->sk_rcvbuf) {
1697                 sk->sk_rcvbuf = val;
1698                 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1699         }
1700         return 0;
1701 }
1702 EXPORT_SYMBOL(tcp_set_rcvlowat);
1703
1704 #ifdef CONFIG_MMU
1705 static const struct vm_operations_struct tcp_vm_ops = {
1706 };
1707
1708 int tcp_mmap(struct file *file, struct socket *sock,
1709              struct vm_area_struct *vma)
1710 {
1711         if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1712                 return -EPERM;
1713         vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1714
1715         /* Instruct vm_insert_page() to not down_read(mmap_sem) */
1716         vma->vm_flags |= VM_MIXEDMAP;
1717
1718         vma->vm_ops = &tcp_vm_ops;
1719         return 0;
1720 }
1721 EXPORT_SYMBOL(tcp_mmap);
1722
1723 static int tcp_zerocopy_receive(struct sock *sk,
1724                                 struct tcp_zerocopy_receive *zc)
1725 {
1726         unsigned long address = (unsigned long)zc->address;
1727         const skb_frag_t *frags = NULL;
1728         u32 length = 0, seq, offset;
1729         struct vm_area_struct *vma;
1730         struct sk_buff *skb = NULL;
1731         struct tcp_sock *tp;
1732         int inq;
1733         int ret;
1734
1735         if (address & (PAGE_SIZE - 1) || address != zc->address)
1736                 return -EINVAL;
1737
1738         if (sk->sk_state == TCP_LISTEN)
1739                 return -ENOTCONN;
1740
1741         sock_rps_record_flow(sk);
1742
1743         down_read(&current->mm->mmap_sem);
1744
1745         ret = -EINVAL;
1746         vma = find_vma(current->mm, address);
1747         if (!vma || vma->vm_start > address || vma->vm_ops != &tcp_vm_ops)
1748                 goto out;
1749         zc->length = min_t(unsigned long, zc->length, vma->vm_end - address);
1750
1751         tp = tcp_sk(sk);
1752         seq = tp->copied_seq;
1753         inq = tcp_inq(sk);
1754         zc->length = min_t(u32, zc->length, inq);
1755         zc->length &= ~(PAGE_SIZE - 1);
1756         if (zc->length) {
1757                 zap_page_range(vma, address, zc->length);
1758                 zc->recv_skip_hint = 0;
1759         } else {
1760                 zc->recv_skip_hint = inq;
1761         }
1762         ret = 0;
1763         while (length + PAGE_SIZE <= zc->length) {
1764                 if (zc->recv_skip_hint < PAGE_SIZE) {
1765                         if (skb) {
1766                                 skb = skb->next;
1767                                 offset = seq - TCP_SKB_CB(skb)->seq;
1768                         } else {
1769                                 skb = tcp_recv_skb(sk, seq, &offset);
1770                         }
1771
1772                         zc->recv_skip_hint = skb->len - offset;
1773                         offset -= skb_headlen(skb);
1774                         if ((int)offset < 0 || skb_has_frag_list(skb))
1775                                 break;
1776                         frags = skb_shinfo(skb)->frags;
1777                         while (offset) {
1778                                 if (frags->size > offset)
1779                                         goto out;
1780                                 offset -= frags->size;
1781                                 frags++;
1782                         }
1783                 }
1784                 if (frags->size != PAGE_SIZE || frags->page_offset) {
1785                         int remaining = zc->recv_skip_hint;
1786
1787                         while (remaining && (frags->size != PAGE_SIZE ||
1788                                              frags->page_offset)) {
1789                                 remaining -= frags->size;
1790                                 frags++;
1791                         }
1792                         zc->recv_skip_hint -= remaining;
1793                         break;
1794                 }
1795                 ret = vm_insert_page(vma, address + length,
1796                                      skb_frag_page(frags));
1797                 if (ret)
1798                         break;
1799                 length += PAGE_SIZE;
1800                 seq += PAGE_SIZE;
1801                 zc->recv_skip_hint -= PAGE_SIZE;
1802                 frags++;
1803         }
1804 out:
1805         up_read(&current->mm->mmap_sem);
1806         if (length) {
1807                 tp->copied_seq = seq;
1808                 tcp_rcv_space_adjust(sk);
1809
1810                 /* Clean up data we have read: This will do ACK frames. */
1811                 tcp_recv_skb(sk, seq, &offset);
1812                 tcp_cleanup_rbuf(sk, length);
1813                 ret = 0;
1814                 if (length == zc->length)
1815                         zc->recv_skip_hint = 0;
1816         } else {
1817                 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
1818                         ret = -EIO;
1819         }
1820         zc->length = length;
1821         return ret;
1822 }
1823 #endif
1824
1825 static void tcp_update_recv_tstamps(struct sk_buff *skb,
1826                                     struct scm_timestamping_internal *tss)
1827 {
1828         if (skb->tstamp)
1829                 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1830         else
1831                 tss->ts[0] = (struct timespec64) {0};
1832
1833         if (skb_hwtstamps(skb)->hwtstamp)
1834                 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1835         else
1836                 tss->ts[2] = (struct timespec64) {0};
1837 }
1838
1839 /* Similar to __sock_recv_timestamp, but does not require an skb */
1840 static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
1841                                struct scm_timestamping_internal *tss)
1842 {
1843         int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
1844         bool has_timestamping = false;
1845
1846         if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
1847                 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
1848                         if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
1849                                 if (new_tstamp) {
1850                                         struct __kernel_timespec kts = {tss->ts[0].tv_sec, tss->ts[0].tv_nsec};
1851
1852                                         put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
1853                                                  sizeof(kts), &kts);
1854                                 } else {
1855                                         struct timespec ts_old = timespec64_to_timespec(tss->ts[0]);
1856
1857                                         put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
1858                                                  sizeof(ts_old), &ts_old);
1859                                 }
1860                         } else {
1861                                 if (new_tstamp) {
1862                                         struct __kernel_sock_timeval stv;
1863
1864                                         stv.tv_sec = tss->ts[0].tv_sec;
1865                                         stv.tv_usec = tss->ts[0].tv_nsec / 1000;
1866                                         put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
1867                                                  sizeof(stv), &stv);
1868                                 } else {
1869                                         struct __kernel_old_timeval tv;
1870
1871                                         tv.tv_sec = tss->ts[0].tv_sec;
1872                                         tv.tv_usec = tss->ts[0].tv_nsec / 1000;
1873                                         put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
1874                                                  sizeof(tv), &tv);
1875                                 }
1876                         }
1877                 }
1878
1879                 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
1880                         has_timestamping = true;
1881                 else
1882                         tss->ts[0] = (struct timespec64) {0};
1883         }
1884
1885         if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
1886                 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
1887                         has_timestamping = true;
1888                 else
1889                         tss->ts[2] = (struct timespec64) {0};
1890         }
1891
1892         if (has_timestamping) {
1893                 tss->ts[1] = (struct timespec64) {0};
1894                 if (sock_flag(sk, SOCK_TSTAMP_NEW))
1895                         put_cmsg_scm_timestamping64(msg, tss);
1896                 else
1897                         put_cmsg_scm_timestamping(msg, tss);
1898         }
1899 }
1900
1901 static int tcp_inq_hint(struct sock *sk)
1902 {
1903         const struct tcp_sock *tp = tcp_sk(sk);
1904         u32 copied_seq = READ_ONCE(tp->copied_seq);
1905         u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
1906         int inq;
1907
1908         inq = rcv_nxt - copied_seq;
1909         if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
1910                 lock_sock(sk);
1911                 inq = tp->rcv_nxt - tp->copied_seq;
1912                 release_sock(sk);
1913         }
1914         /* After receiving a FIN, tell the user-space to continue reading
1915          * by returning a non-zero inq.
1916          */
1917         if (inq == 0 && sock_flag(sk, SOCK_DONE))
1918                 inq = 1;
1919         return inq;
1920 }
1921
1922 /*
1923  *      This routine copies from a sock struct into the user buffer.
1924  *
1925  *      Technical note: in 2.3 we work on _locked_ socket, so that
1926  *      tricks with *seq access order and skb->users are not required.
1927  *      Probably, code can be easily improved even more.
1928  */
1929
1930 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1931                 int flags, int *addr_len)
1932 {
1933         struct tcp_sock *tp = tcp_sk(sk);
1934         int copied = 0;
1935         u32 peek_seq;
1936         u32 *seq;
1937         unsigned long used;
1938         int err, inq;
1939         int target;             /* Read at least this many bytes */
1940         long timeo;
1941         struct sk_buff *skb, *last;
1942         u32 urg_hole = 0;
1943         struct scm_timestamping_internal tss;
1944         bool has_tss = false;
1945         bool has_cmsg;
1946
1947         if (unlikely(flags & MSG_ERRQUEUE))
1948                 return inet_recv_error(sk, msg, len, addr_len);
1949
1950         if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1951             (sk->sk_state == TCP_ESTABLISHED))
1952                 sk_busy_loop(sk, nonblock);
1953
1954         lock_sock(sk);
1955
1956         err = -ENOTCONN;
1957         if (sk->sk_state == TCP_LISTEN)
1958                 goto out;
1959
1960         has_cmsg = tp->recvmsg_inq;
1961         timeo = sock_rcvtimeo(sk, nonblock);
1962
1963         /* Urgent data needs to be handled specially. */
1964         if (flags & MSG_OOB)
1965                 goto recv_urg;
1966
1967         if (unlikely(tp->repair)) {
1968                 err = -EPERM;
1969                 if (!(flags & MSG_PEEK))
1970                         goto out;
1971
1972                 if (tp->repair_queue == TCP_SEND_QUEUE)
1973                         goto recv_sndq;
1974
1975                 err = -EINVAL;
1976                 if (tp->repair_queue == TCP_NO_QUEUE)
1977                         goto out;
1978
1979                 /* 'common' recv queue MSG_PEEK-ing */
1980         }
1981
1982         seq = &tp->copied_seq;
1983         if (flags & MSG_PEEK) {
1984                 peek_seq = tp->copied_seq;
1985                 seq = &peek_seq;
1986         }
1987
1988         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1989
1990         do {
1991                 u32 offset;
1992
1993                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1994                 if (tp->urg_data && tp->urg_seq == *seq) {
1995                         if (copied)
1996                                 break;
1997                         if (signal_pending(current)) {
1998                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1999                                 break;
2000                         }
2001                 }
2002
2003                 /* Next get a buffer. */
2004
2005                 last = skb_peek_tail(&sk->sk_receive_queue);
2006                 skb_queue_walk(&sk->sk_receive_queue, skb) {
2007                         last = skb;
2008                         /* Now that we have two receive queues this
2009                          * shouldn't happen.
2010                          */
2011                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2012                                  "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2013                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2014                                  flags))
2015                                 break;
2016
2017                         offset = *seq - TCP_SKB_CB(skb)->seq;
2018                         if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2019                                 pr_err_once("%s: found a SYN, please report !\n", __func__);
2020                                 offset--;
2021                         }
2022                         if (offset < skb->len)
2023                                 goto found_ok_skb;
2024                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2025                                 goto found_fin_ok;
2026                         WARN(!(flags & MSG_PEEK),
2027                              "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2028                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2029                 }
2030
2031                 /* Well, if we have backlog, try to process it now yet. */
2032
2033                 if (copied >= target && !sk->sk_backlog.tail)
2034                         break;
2035
2036                 if (copied) {
2037                         if (sk->sk_err ||
2038                             sk->sk_state == TCP_CLOSE ||
2039                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
2040                             !timeo ||
2041                             signal_pending(current))
2042                                 break;
2043                 } else {
2044                         if (sock_flag(sk, SOCK_DONE))
2045                                 break;
2046
2047                         if (sk->sk_err) {
2048                                 copied = sock_error(sk);
2049                                 break;
2050                         }
2051
2052                         if (sk->sk_shutdown & RCV_SHUTDOWN)
2053                                 break;
2054
2055                         if (sk->sk_state == TCP_CLOSE) {
2056                                 /* This occurs when user tries to read
2057                                  * from never connected socket.
2058                                  */
2059                                 copied = -ENOTCONN;
2060                                 break;
2061                         }
2062
2063                         if (!timeo) {
2064                                 copied = -EAGAIN;
2065                                 break;
2066                         }
2067
2068                         if (signal_pending(current)) {
2069                                 copied = sock_intr_errno(timeo);
2070                                 break;
2071                         }
2072                 }
2073
2074                 tcp_cleanup_rbuf(sk, copied);
2075
2076                 if (copied >= target) {
2077                         /* Do not sleep, just process backlog. */
2078                         release_sock(sk);
2079                         lock_sock(sk);
2080                 } else {
2081                         sk_wait_data(sk, &timeo, last);
2082                 }
2083
2084                 if ((flags & MSG_PEEK) &&
2085                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
2086                         net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2087                                             current->comm,
2088                                             task_pid_nr(current));
2089                         peek_seq = tp->copied_seq;
2090                 }
2091                 continue;
2092
2093 found_ok_skb:
2094                 /* Ok so how much can we use? */
2095                 used = skb->len - offset;
2096                 if (len < used)
2097                         used = len;
2098
2099                 /* Do we have urgent data here? */
2100                 if (tp->urg_data) {
2101                         u32 urg_offset = tp->urg_seq - *seq;
2102                         if (urg_offset < used) {
2103                                 if (!urg_offset) {
2104                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
2105                                                 ++*seq;
2106                                                 urg_hole++;
2107                                                 offset++;
2108                                                 used--;
2109                                                 if (!used)
2110                                                         goto skip_copy;
2111                                         }
2112                                 } else
2113                                         used = urg_offset;
2114                         }
2115                 }
2116
2117                 if (!(flags & MSG_TRUNC)) {
2118                         err = skb_copy_datagram_msg(skb, offset, msg, used);
2119                         if (err) {
2120                                 /* Exception. Bailout! */
2121                                 if (!copied)
2122                                         copied = -EFAULT;
2123                                 break;
2124                         }
2125                 }
2126
2127                 *seq += used;
2128                 copied += used;
2129                 len -= used;
2130
2131                 tcp_rcv_space_adjust(sk);
2132
2133 skip_copy:
2134                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
2135                         tp->urg_data = 0;
2136                         tcp_fast_path_check(sk);
2137                 }
2138                 if (used + offset < skb->len)
2139                         continue;
2140
2141                 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2142                         tcp_update_recv_tstamps(skb, &tss);
2143                         has_tss = true;
2144                         has_cmsg = true;
2145                 }
2146                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2147                         goto found_fin_ok;
2148                 if (!(flags & MSG_PEEK))
2149                         sk_eat_skb(sk, skb);
2150                 continue;
2151
2152 found_fin_ok:
2153                 /* Process the FIN. */
2154                 ++*seq;
2155                 if (!(flags & MSG_PEEK))
2156                         sk_eat_skb(sk, skb);
2157                 break;
2158         } while (len > 0);
2159
2160         /* According to UNIX98, msg_name/msg_namelen are ignored
2161          * on connected socket. I was just happy when found this 8) --ANK
2162          */
2163
2164         /* Clean up data we have read: This will do ACK frames. */
2165         tcp_cleanup_rbuf(sk, copied);
2166
2167         release_sock(sk);
2168
2169         if (has_cmsg) {
2170                 if (has_tss)
2171                         tcp_recv_timestamp(msg, sk, &tss);
2172                 if (tp->recvmsg_inq) {
2173                         inq = tcp_inq_hint(sk);
2174                         put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2175                 }
2176         }
2177
2178         return copied;
2179
2180 out:
2181         release_sock(sk);
2182         return err;
2183
2184 recv_urg:
2185         err = tcp_recv_urg(sk, msg, len, flags);
2186         goto out;
2187
2188 recv_sndq:
2189         err = tcp_peek_sndq(sk, msg, len);
2190         goto out;
2191 }
2192 EXPORT_SYMBOL(tcp_recvmsg);
2193
2194 void tcp_set_state(struct sock *sk, int state)
2195 {
2196         int oldstate = sk->sk_state;
2197
2198         /* We defined a new enum for TCP states that are exported in BPF
2199          * so as not force the internal TCP states to be frozen. The
2200          * following checks will detect if an internal state value ever
2201          * differs from the BPF value. If this ever happens, then we will
2202          * need to remap the internal value to the BPF value before calling
2203          * tcp_call_bpf_2arg.
2204          */
2205         BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2206         BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2207         BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2208         BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2209         BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2210         BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2211         BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2212         BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2213         BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2214         BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2215         BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2216         BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2217         BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2218
2219         if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2220                 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2221
2222         switch (state) {
2223         case TCP_ESTABLISHED:
2224                 if (oldstate != TCP_ESTABLISHED)
2225                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2226                 break;
2227
2228         case TCP_CLOSE:
2229                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2230                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2231
2232                 sk->sk_prot->unhash(sk);
2233                 if (inet_csk(sk)->icsk_bind_hash &&
2234                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2235                         inet_put_port(sk);
2236                 /* fall through */
2237         default:
2238                 if (oldstate == TCP_ESTABLISHED)
2239                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2240         }
2241
2242         /* Change state AFTER socket is unhashed to avoid closed
2243          * socket sitting in hash tables.
2244          */
2245         inet_sk_state_store(sk, state);
2246 }
2247 EXPORT_SYMBOL_GPL(tcp_set_state);
2248
2249 /*
2250  *      State processing on a close. This implements the state shift for
2251  *      sending our FIN frame. Note that we only send a FIN for some
2252  *      states. A shutdown() may have already sent the FIN, or we may be
2253  *      closed.
2254  */
2255
2256 static const unsigned char new_state[16] = {
2257   /* current state:        new state:      action:      */
2258   [0 /* (Invalid) */]   = TCP_CLOSE,
2259   [TCP_ESTABLISHED]     = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2260   [TCP_SYN_SENT]        = TCP_CLOSE,
2261   [TCP_SYN_RECV]        = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2262   [TCP_FIN_WAIT1]       = TCP_FIN_WAIT1,
2263   [TCP_FIN_WAIT2]       = TCP_FIN_WAIT2,
2264   [TCP_TIME_WAIT]       = TCP_CLOSE,
2265   [TCP_CLOSE]           = TCP_CLOSE,
2266   [TCP_CLOSE_WAIT]      = TCP_LAST_ACK  | TCP_ACTION_FIN,
2267   [TCP_LAST_ACK]        = TCP_LAST_ACK,
2268   [TCP_LISTEN]          = TCP_CLOSE,
2269   [TCP_CLOSING]         = TCP_CLOSING,
2270   [TCP_NEW_SYN_RECV]    = TCP_CLOSE,    /* should not happen ! */
2271 };
2272
2273 static int tcp_close_state(struct sock *sk)
2274 {
2275         int next = (int)new_state[sk->sk_state];
2276         int ns = next & TCP_STATE_MASK;
2277
2278         tcp_set_state(sk, ns);
2279
2280         return next & TCP_ACTION_FIN;
2281 }
2282
2283 /*
2284  *      Shutdown the sending side of a connection. Much like close except
2285  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2286  */
2287
2288 void tcp_shutdown(struct sock *sk, int how)
2289 {
2290         /*      We need to grab some memory, and put together a FIN,
2291          *      and then put it into the queue to be sent.
2292          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2293          */
2294         if (!(how & SEND_SHUTDOWN))
2295                 return;
2296
2297         /* If we've already sent a FIN, or it's a closed state, skip this. */
2298         if ((1 << sk->sk_state) &
2299             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2300              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2301                 /* Clear out any half completed packets.  FIN if needed. */
2302                 if (tcp_close_state(sk))
2303                         tcp_send_fin(sk);
2304         }
2305 }
2306 EXPORT_SYMBOL(tcp_shutdown);
2307
2308 bool tcp_check_oom(struct sock *sk, int shift)
2309 {
2310         bool too_many_orphans, out_of_socket_memory;
2311
2312         too_many_orphans = tcp_too_many_orphans(sk, shift);
2313         out_of_socket_memory = tcp_out_of_memory(sk);
2314
2315         if (too_many_orphans)
2316                 net_info_ratelimited("too many orphaned sockets\n");
2317         if (out_of_socket_memory)
2318                 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2319         return too_many_orphans || out_of_socket_memory;
2320 }
2321
2322 void tcp_close(struct sock *sk, long timeout)
2323 {
2324         struct sk_buff *skb;
2325         int data_was_unread = 0;
2326         int state;
2327
2328         lock_sock(sk);
2329         sk->sk_shutdown = SHUTDOWN_MASK;
2330
2331         if (sk->sk_state == TCP_LISTEN) {
2332                 tcp_set_state(sk, TCP_CLOSE);
2333
2334                 /* Special case. */
2335                 inet_csk_listen_stop(sk);
2336
2337                 goto adjudge_to_death;
2338         }
2339
2340         /*  We need to flush the recv. buffs.  We do this only on the
2341          *  descriptor close, not protocol-sourced closes, because the
2342          *  reader process may not have drained the data yet!
2343          */
2344         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2345                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2346
2347                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2348                         len--;
2349                 data_was_unread += len;
2350                 __kfree_skb(skb);
2351         }
2352
2353         sk_mem_reclaim(sk);
2354
2355         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2356         if (sk->sk_state == TCP_CLOSE)
2357                 goto adjudge_to_death;
2358
2359         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2360          * data was lost. To witness the awful effects of the old behavior of
2361          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2362          * GET in an FTP client, suspend the process, wait for the client to
2363          * advertise a zero window, then kill -9 the FTP client, wheee...
2364          * Note: timeout is always zero in such a case.
2365          */
2366         if (unlikely(tcp_sk(sk)->repair)) {
2367                 sk->sk_prot->disconnect(sk, 0);
2368         } else if (data_was_unread) {
2369                 /* Unread data was tossed, zap the connection. */
2370                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2371                 tcp_set_state(sk, TCP_CLOSE);
2372                 tcp_send_active_reset(sk, sk->sk_allocation);
2373         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2374                 /* Check zero linger _after_ checking for unread data. */
2375                 sk->sk_prot->disconnect(sk, 0);
2376                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2377         } else if (tcp_close_state(sk)) {
2378                 /* We FIN if the application ate all the data before
2379                  * zapping the connection.
2380                  */
2381
2382                 /* RED-PEN. Formally speaking, we have broken TCP state
2383                  * machine. State transitions:
2384                  *
2385                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2386                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2387                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2388                  *
2389                  * are legal only when FIN has been sent (i.e. in window),
2390                  * rather than queued out of window. Purists blame.
2391                  *
2392                  * F.e. "RFC state" is ESTABLISHED,
2393                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2394                  *
2395                  * The visible declinations are that sometimes
2396                  * we enter time-wait state, when it is not required really
2397                  * (harmless), do not send active resets, when they are
2398                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2399                  * they look as CLOSING or LAST_ACK for Linux)
2400                  * Probably, I missed some more holelets.
2401                  *                                              --ANK
2402                  * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2403                  * in a single packet! (May consider it later but will
2404                  * probably need API support or TCP_CORK SYN-ACK until
2405                  * data is written and socket is closed.)
2406                  */
2407                 tcp_send_fin(sk);
2408         }
2409
2410         sk_stream_wait_close(sk, timeout);
2411
2412 adjudge_to_death:
2413         state = sk->sk_state;
2414         sock_hold(sk);
2415         sock_orphan(sk);
2416
2417         local_bh_disable();
2418         bh_lock_sock(sk);
2419         /* remove backlog if any, without releasing ownership. */
2420         __release_sock(sk);
2421
2422         percpu_counter_inc(sk->sk_prot->orphan_count);
2423
2424         /* Have we already been destroyed by a softirq or backlog? */
2425         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2426                 goto out;
2427
2428         /*      This is a (useful) BSD violating of the RFC. There is a
2429          *      problem with TCP as specified in that the other end could
2430          *      keep a socket open forever with no application left this end.
2431          *      We use a 1 minute timeout (about the same as BSD) then kill
2432          *      our end. If they send after that then tough - BUT: long enough
2433          *      that we won't make the old 4*rto = almost no time - whoops
2434          *      reset mistake.
2435          *
2436          *      Nope, it was not mistake. It is really desired behaviour
2437          *      f.e. on http servers, when such sockets are useless, but
2438          *      consume significant resources. Let's do it with special
2439          *      linger2 option.                                 --ANK
2440          */
2441
2442         if (sk->sk_state == TCP_FIN_WAIT2) {
2443                 struct tcp_sock *tp = tcp_sk(sk);
2444                 if (tp->linger2 < 0) {
2445                         tcp_set_state(sk, TCP_CLOSE);
2446                         tcp_send_active_reset(sk, GFP_ATOMIC);
2447                         __NET_INC_STATS(sock_net(sk),
2448                                         LINUX_MIB_TCPABORTONLINGER);
2449                 } else {
2450                         const int tmo = tcp_fin_time(sk);
2451
2452                         if (tmo > TCP_TIMEWAIT_LEN) {
2453                                 inet_csk_reset_keepalive_timer(sk,
2454                                                 tmo - TCP_TIMEWAIT_LEN);
2455                         } else {
2456                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2457                                 goto out;
2458                         }
2459                 }
2460         }
2461         if (sk->sk_state != TCP_CLOSE) {
2462                 sk_mem_reclaim(sk);
2463                 if (tcp_check_oom(sk, 0)) {
2464                         tcp_set_state(sk, TCP_CLOSE);
2465                         tcp_send_active_reset(sk, GFP_ATOMIC);
2466                         __NET_INC_STATS(sock_net(sk),
2467                                         LINUX_MIB_TCPABORTONMEMORY);
2468                 } else if (!check_net(sock_net(sk))) {
2469                         /* Not possible to send reset; just close */
2470                         tcp_set_state(sk, TCP_CLOSE);
2471                 }
2472         }
2473
2474         if (sk->sk_state == TCP_CLOSE) {
2475                 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2476                 /* We could get here with a non-NULL req if the socket is
2477                  * aborted (e.g., closed with unread data) before 3WHS
2478                  * finishes.
2479                  */
2480                 if (req)
2481                         reqsk_fastopen_remove(sk, req, false);
2482                 inet_csk_destroy_sock(sk);
2483         }
2484         /* Otherwise, socket is reprieved until protocol close. */
2485
2486 out:
2487         bh_unlock_sock(sk);
2488         local_bh_enable();
2489         release_sock(sk);
2490         sock_put(sk);
2491 }
2492 EXPORT_SYMBOL(tcp_close);
2493
2494 /* These states need RST on ABORT according to RFC793 */
2495
2496 static inline bool tcp_need_reset(int state)
2497 {
2498         return (1 << state) &
2499                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2500                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2501 }
2502
2503 static void tcp_rtx_queue_purge(struct sock *sk)
2504 {
2505         struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2506
2507         while (p) {
2508                 struct sk_buff *skb = rb_to_skb(p);
2509
2510                 p = rb_next(p);
2511                 /* Since we are deleting whole queue, no need to
2512                  * list_del(&skb->tcp_tsorted_anchor)
2513                  */
2514                 tcp_rtx_queue_unlink(skb, sk);
2515                 sk_wmem_free_skb(sk, skb);
2516         }
2517 }
2518
2519 void tcp_write_queue_purge(struct sock *sk)
2520 {
2521         struct sk_buff *skb;
2522
2523         tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2524         while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2525                 tcp_skb_tsorted_anchor_cleanup(skb);
2526                 sk_wmem_free_skb(sk, skb);
2527         }
2528         tcp_rtx_queue_purge(sk);
2529         skb = sk->sk_tx_skb_cache;
2530         if (skb) {
2531                 __kfree_skb(skb);
2532                 sk->sk_tx_skb_cache = NULL;
2533         }
2534         INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2535         sk_mem_reclaim(sk);
2536         tcp_clear_all_retrans_hints(tcp_sk(sk));
2537         tcp_sk(sk)->packets_out = 0;
2538         inet_csk(sk)->icsk_backoff = 0;
2539 }
2540
2541 int tcp_disconnect(struct sock *sk, int flags)
2542 {
2543         struct inet_sock *inet = inet_sk(sk);
2544         struct inet_connection_sock *icsk = inet_csk(sk);
2545         struct tcp_sock *tp = tcp_sk(sk);
2546         int old_state = sk->sk_state;
2547
2548         if (old_state != TCP_CLOSE)
2549                 tcp_set_state(sk, TCP_CLOSE);
2550
2551         /* ABORT function of RFC793 */
2552         if (old_state == TCP_LISTEN) {
2553                 inet_csk_listen_stop(sk);
2554         } else if (unlikely(tp->repair)) {
2555                 sk->sk_err = ECONNABORTED;
2556         } else if (tcp_need_reset(old_state) ||
2557                    (tp->snd_nxt != tp->write_seq &&
2558                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2559                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2560                  * states
2561                  */
2562                 tcp_send_active_reset(sk, gfp_any());
2563                 sk->sk_err = ECONNRESET;
2564         } else if (old_state == TCP_SYN_SENT)
2565                 sk->sk_err = ECONNRESET;
2566
2567         tcp_clear_xmit_timers(sk);
2568         __skb_queue_purge(&sk->sk_receive_queue);
2569         if (sk->sk_rx_skb_cache) {
2570                 __kfree_skb(sk->sk_rx_skb_cache);
2571                 sk->sk_rx_skb_cache = NULL;
2572         }
2573         tp->copied_seq = tp->rcv_nxt;
2574         tp->urg_data = 0;
2575         tcp_write_queue_purge(sk);
2576         tcp_fastopen_active_disable_ofo_check(sk);
2577         skb_rbtree_purge(&tp->out_of_order_queue);
2578
2579         inet->inet_dport = 0;
2580
2581         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2582                 inet_reset_saddr(sk);
2583
2584         sk->sk_shutdown = 0;
2585         sock_reset_flag(sk, SOCK_DONE);
2586         tp->srtt_us = 0;
2587         tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
2588         tp->rcv_rtt_last_tsecr = 0;
2589         tp->write_seq += tp->max_window + 2;
2590         if (tp->write_seq == 0)
2591                 tp->write_seq = 1;
2592         icsk->icsk_backoff = 0;
2593         tp->snd_cwnd = 2;
2594         icsk->icsk_probes_out = 0;
2595         icsk->icsk_rto = TCP_TIMEOUT_INIT;
2596         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2597         tp->snd_cwnd = TCP_INIT_CWND;
2598         tp->snd_cwnd_cnt = 0;
2599         tp->window_clamp = 0;
2600         tp->delivered_ce = 0;
2601         tcp_set_ca_state(sk, TCP_CA_Open);
2602         tp->is_sack_reneg = 0;
2603         tcp_clear_retrans(tp);
2604         inet_csk_delack_init(sk);
2605         /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2606          * issue in __tcp_select_window()
2607          */
2608         icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2609         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2610         __sk_dst_reset(sk);
2611         dst_release(sk->sk_rx_dst);
2612         sk->sk_rx_dst = NULL;
2613         tcp_saved_syn_free(tp);
2614         tp->compressed_ack = 0;
2615         tp->bytes_sent = 0;
2616         tp->bytes_retrans = 0;
2617         tp->duplicate_sack[0].start_seq = 0;
2618         tp->duplicate_sack[0].end_seq = 0;
2619         tp->dsack_dups = 0;
2620         tp->reord_seen = 0;
2621         tp->retrans_out = 0;
2622         tp->sacked_out = 0;
2623         tp->tlp_high_seq = 0;
2624         tp->last_oow_ack_time = 0;
2625         /* There's a bubble in the pipe until at least the first ACK. */
2626         tp->app_limited = ~0U;
2627         tp->rack.mstamp = 0;
2628         tp->rack.advanced = 0;
2629         tp->rack.reo_wnd_steps = 1;
2630         tp->rack.last_delivered = 0;
2631         tp->rack.reo_wnd_persist = 0;
2632         tp->rack.dsack_seen = 0;
2633         tp->syn_data_acked = 0;
2634         tp->rx_opt.saw_tstamp = 0;
2635         tp->rx_opt.dsack = 0;
2636         tp->rx_opt.num_sacks = 0;
2637
2638
2639         /* Clean up fastopen related fields */
2640         tcp_free_fastopen_req(tp);
2641         inet->defer_connect = 0;
2642
2643         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2644
2645         if (sk->sk_frag.page) {
2646                 put_page(sk->sk_frag.page);
2647                 sk->sk_frag.page = NULL;
2648                 sk->sk_frag.offset = 0;
2649         }
2650
2651         sk->sk_error_report(sk);
2652         return 0;
2653 }
2654 EXPORT_SYMBOL(tcp_disconnect);
2655
2656 static inline bool tcp_can_repair_sock(const struct sock *sk)
2657 {
2658         return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2659                 (sk->sk_state != TCP_LISTEN);
2660 }
2661
2662 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2663 {
2664         struct tcp_repair_window opt;
2665
2666         if (!tp->repair)
2667                 return -EPERM;
2668
2669         if (len != sizeof(opt))
2670                 return -EINVAL;
2671
2672         if (copy_from_user(&opt, optbuf, sizeof(opt)))
2673                 return -EFAULT;
2674
2675         if (opt.max_window < opt.snd_wnd)
2676                 return -EINVAL;
2677
2678         if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2679                 return -EINVAL;
2680
2681         if (after(opt.rcv_wup, tp->rcv_nxt))
2682                 return -EINVAL;
2683
2684         tp->snd_wl1     = opt.snd_wl1;
2685         tp->snd_wnd     = opt.snd_wnd;
2686         tp->max_window  = opt.max_window;
2687
2688         tp->rcv_wnd     = opt.rcv_wnd;
2689         tp->rcv_wup     = opt.rcv_wup;
2690
2691         return 0;
2692 }
2693
2694 static int tcp_repair_options_est(struct sock *sk,
2695                 struct tcp_repair_opt __user *optbuf, unsigned int len)
2696 {
2697         struct tcp_sock *tp = tcp_sk(sk);
2698         struct tcp_repair_opt opt;
2699
2700         while (len >= sizeof(opt)) {
2701                 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2702                         return -EFAULT;
2703
2704                 optbuf++;
2705                 len -= sizeof(opt);
2706
2707                 switch (opt.opt_code) {
2708                 case TCPOPT_MSS:
2709                         tp->rx_opt.mss_clamp = opt.opt_val;
2710                         tcp_mtup_init(sk);
2711                         break;
2712                 case TCPOPT_WINDOW:
2713                         {
2714                                 u16 snd_wscale = opt.opt_val & 0xFFFF;
2715                                 u16 rcv_wscale = opt.opt_val >> 16;
2716
2717                                 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
2718                                         return -EFBIG;
2719
2720                                 tp->rx_opt.snd_wscale = snd_wscale;
2721                                 tp->rx_opt.rcv_wscale = rcv_wscale;
2722                                 tp->rx_opt.wscale_ok = 1;
2723                         }
2724                         break;
2725                 case TCPOPT_SACK_PERM:
2726                         if (opt.opt_val != 0)
2727                                 return -EINVAL;
2728
2729                         tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2730                         break;
2731                 case TCPOPT_TIMESTAMP:
2732                         if (opt.opt_val != 0)
2733                                 return -EINVAL;
2734
2735                         tp->rx_opt.tstamp_ok = 1;
2736                         break;
2737                 }
2738         }
2739
2740         return 0;
2741 }
2742
2743 /*
2744  *      Socket option code for TCP.
2745  */
2746 static int do_tcp_setsockopt(struct sock *sk, int level,
2747                 int optname, char __user *optval, unsigned int optlen)
2748 {
2749         struct tcp_sock *tp = tcp_sk(sk);
2750         struct inet_connection_sock *icsk = inet_csk(sk);
2751         struct net *net = sock_net(sk);
2752         int val;
2753         int err = 0;
2754
2755         /* These are data/string values, all the others are ints */
2756         switch (optname) {
2757         case TCP_CONGESTION: {
2758                 char name[TCP_CA_NAME_MAX];
2759
2760                 if (optlen < 1)
2761                         return -EINVAL;
2762
2763                 val = strncpy_from_user(name, optval,
2764                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2765                 if (val < 0)
2766                         return -EFAULT;
2767                 name[val] = 0;
2768
2769                 lock_sock(sk);
2770                 err = tcp_set_congestion_control(sk, name, true, true);
2771                 release_sock(sk);
2772                 return err;
2773         }
2774         case TCP_ULP: {
2775                 char name[TCP_ULP_NAME_MAX];
2776
2777                 if (optlen < 1)
2778                         return -EINVAL;
2779
2780                 val = strncpy_from_user(name, optval,
2781                                         min_t(long, TCP_ULP_NAME_MAX - 1,
2782                                               optlen));
2783                 if (val < 0)
2784                         return -EFAULT;
2785                 name[val] = 0;
2786
2787                 lock_sock(sk);
2788                 err = tcp_set_ulp(sk, name);
2789                 release_sock(sk);
2790                 return err;
2791         }
2792         case TCP_FASTOPEN_KEY: {
2793                 __u8 key[TCP_FASTOPEN_KEY_LENGTH];
2794
2795                 if (optlen != sizeof(key))
2796                         return -EINVAL;
2797
2798                 if (copy_from_user(key, optval, optlen))
2799                         return -EFAULT;
2800
2801                 return tcp_fastopen_reset_cipher(net, sk, key, sizeof(key));
2802         }
2803         default:
2804                 /* fallthru */
2805                 break;
2806         }
2807
2808         if (optlen < sizeof(int))
2809                 return -EINVAL;
2810
2811         if (get_user(val, (int __user *)optval))
2812                 return -EFAULT;
2813
2814         lock_sock(sk);
2815
2816         switch (optname) {
2817         case TCP_MAXSEG:
2818                 /* Values greater than interface MTU won't take effect. However
2819                  * at the point when this call is done we typically don't yet
2820                  * know which interface is going to be used
2821                  */
2822                 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
2823                         err = -EINVAL;
2824                         break;
2825                 }
2826                 tp->rx_opt.user_mss = val;
2827                 break;
2828
2829         case TCP_NODELAY:
2830                 if (val) {
2831                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2832                          * this option on corked socket is remembered, but
2833                          * it is not activated until cork is cleared.
2834                          *
2835                          * However, when TCP_NODELAY is set we make
2836                          * an explicit push, which overrides even TCP_CORK
2837                          * for currently queued segments.
2838                          */
2839                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2840                         tcp_push_pending_frames(sk);
2841                 } else {
2842                         tp->nonagle &= ~TCP_NAGLE_OFF;
2843                 }
2844                 break;
2845
2846         case TCP_THIN_LINEAR_TIMEOUTS:
2847                 if (val < 0 || val > 1)
2848                         err = -EINVAL;
2849                 else
2850                         tp->thin_lto = val;
2851                 break;
2852
2853         case TCP_THIN_DUPACK:
2854                 if (val < 0 || val > 1)
2855                         err = -EINVAL;
2856                 break;
2857
2858         case TCP_REPAIR:
2859                 if (!tcp_can_repair_sock(sk))
2860                         err = -EPERM;
2861                 else if (val == TCP_REPAIR_ON) {
2862                         tp->repair = 1;
2863                         sk->sk_reuse = SK_FORCE_REUSE;
2864                         tp->repair_queue = TCP_NO_QUEUE;
2865                 } else if (val == TCP_REPAIR_OFF) {
2866                         tp->repair = 0;
2867                         sk->sk_reuse = SK_NO_REUSE;
2868                         tcp_send_window_probe(sk);
2869                 } else if (val == TCP_REPAIR_OFF_NO_WP) {
2870                         tp->repair = 0;
2871                         sk->sk_reuse = SK_NO_REUSE;
2872                 } else
2873                         err = -EINVAL;
2874
2875                 break;
2876
2877         case TCP_REPAIR_QUEUE:
2878                 if (!tp->repair)
2879                         err = -EPERM;
2880                 else if ((unsigned int)val < TCP_QUEUES_NR)
2881                         tp->repair_queue = val;
2882                 else
2883                         err = -EINVAL;
2884                 break;
2885
2886         case TCP_QUEUE_SEQ:
2887                 if (sk->sk_state != TCP_CLOSE)
2888                         err = -EPERM;
2889                 else if (tp->repair_queue == TCP_SEND_QUEUE)
2890                         tp->write_seq = val;
2891                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2892                         tp->rcv_nxt = val;
2893                 else
2894                         err = -EINVAL;
2895                 break;
2896
2897         case TCP_REPAIR_OPTIONS:
2898                 if (!tp->repair)
2899                         err = -EINVAL;
2900                 else if (sk->sk_state == TCP_ESTABLISHED)
2901                         err = tcp_repair_options_est(sk,
2902                                         (struct tcp_repair_opt __user *)optval,
2903                                         optlen);
2904                 else
2905                         err = -EPERM;
2906                 break;
2907
2908         case TCP_CORK:
2909                 /* When set indicates to always queue non-full frames.
2910                  * Later the user clears this option and we transmit
2911                  * any pending partial frames in the queue.  This is
2912                  * meant to be used alongside sendfile() to get properly
2913                  * filled frames when the user (for example) must write
2914                  * out headers with a write() call first and then use
2915                  * sendfile to send out the data parts.
2916                  *
2917                  * TCP_CORK can be set together with TCP_NODELAY and it is
2918                  * stronger than TCP_NODELAY.
2919                  */
2920                 if (val) {
2921                         tp->nonagle |= TCP_NAGLE_CORK;
2922                 } else {
2923                         tp->nonagle &= ~TCP_NAGLE_CORK;
2924                         if (tp->nonagle&TCP_NAGLE_OFF)
2925                                 tp->nonagle |= TCP_NAGLE_PUSH;
2926                         tcp_push_pending_frames(sk);
2927                 }
2928                 break;
2929
2930         case TCP_KEEPIDLE:
2931                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2932                         err = -EINVAL;
2933                 else {
2934                         tp->keepalive_time = val * HZ;
2935                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2936                             !((1 << sk->sk_state) &
2937                               (TCPF_CLOSE | TCPF_LISTEN))) {
2938                                 u32 elapsed = keepalive_time_elapsed(tp);
2939                                 if (tp->keepalive_time > elapsed)
2940                                         elapsed = tp->keepalive_time - elapsed;
2941                                 else
2942                                         elapsed = 0;
2943                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2944                         }
2945                 }
2946                 break;
2947         case TCP_KEEPINTVL:
2948                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2949                         err = -EINVAL;
2950                 else
2951                         tp->keepalive_intvl = val * HZ;
2952                 break;
2953         case TCP_KEEPCNT:
2954                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2955                         err = -EINVAL;
2956                 else
2957                         tp->keepalive_probes = val;
2958                 break;
2959         case TCP_SYNCNT:
2960                 if (val < 1 || val > MAX_TCP_SYNCNT)
2961                         err = -EINVAL;
2962                 else
2963                         icsk->icsk_syn_retries = val;
2964                 break;
2965
2966         case TCP_SAVE_SYN:
2967                 if (val < 0 || val > 1)
2968                         err = -EINVAL;
2969                 else
2970                         tp->save_syn = val;
2971                 break;
2972
2973         case TCP_LINGER2:
2974                 if (val < 0)
2975                         tp->linger2 = -1;
2976                 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
2977                         tp->linger2 = 0;
2978                 else
2979                         tp->linger2 = val * HZ;
2980                 break;
2981
2982         case TCP_DEFER_ACCEPT:
2983                 /* Translate value in seconds to number of retransmits */
2984                 icsk->icsk_accept_queue.rskq_defer_accept =
2985                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2986                                         TCP_RTO_MAX / HZ);
2987                 break;
2988
2989         case TCP_WINDOW_CLAMP:
2990                 if (!val) {
2991                         if (sk->sk_state != TCP_CLOSE) {
2992                                 err = -EINVAL;
2993                                 break;
2994                         }
2995                         tp->window_clamp = 0;
2996                 } else
2997                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2998                                                 SOCK_MIN_RCVBUF / 2 : val;
2999                 break;
3000
3001         case TCP_QUICKACK:
3002                 if (!val) {
3003                         inet_csk_enter_pingpong_mode(sk);
3004                 } else {
3005                         inet_csk_exit_pingpong_mode(sk);
3006                         if ((1 << sk->sk_state) &
3007                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3008                             inet_csk_ack_scheduled(sk)) {
3009                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
3010                                 tcp_cleanup_rbuf(sk, 1);
3011                                 if (!(val & 1))
3012                                         inet_csk_enter_pingpong_mode(sk);
3013                         }
3014                 }
3015                 break;
3016
3017 #ifdef CONFIG_TCP_MD5SIG
3018         case TCP_MD5SIG:
3019         case TCP_MD5SIG_EXT:
3020                 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
3021                         err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3022                 else
3023                         err = -EINVAL;
3024                 break;
3025 #endif
3026         case TCP_USER_TIMEOUT:
3027                 /* Cap the max time in ms TCP will retry or probe the window
3028                  * before giving up and aborting (ETIMEDOUT) a connection.
3029                  */
3030                 if (val < 0)
3031                         err = -EINVAL;
3032                 else
3033                         icsk->icsk_user_timeout = val;
3034                 break;
3035
3036         case TCP_FASTOPEN:
3037                 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3038                     TCPF_LISTEN))) {
3039                         tcp_fastopen_init_key_once(net);
3040
3041                         fastopen_queue_tune(sk, val);
3042                 } else {
3043                         err = -EINVAL;
3044                 }
3045                 break;
3046         case TCP_FASTOPEN_CONNECT:
3047                 if (val > 1 || val < 0) {
3048                         err = -EINVAL;
3049                 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
3050                         if (sk->sk_state == TCP_CLOSE)
3051                                 tp->fastopen_connect = val;
3052                         else
3053                                 err = -EINVAL;
3054                 } else {
3055                         err = -EOPNOTSUPP;
3056                 }
3057                 break;
3058         case TCP_FASTOPEN_NO_COOKIE:
3059                 if (val > 1 || val < 0)
3060                         err = -EINVAL;
3061                 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3062                         err = -EINVAL;
3063                 else
3064                         tp->fastopen_no_cookie = val;
3065                 break;
3066         case TCP_TIMESTAMP:
3067                 if (!tp->repair)
3068                         err = -EPERM;
3069                 else
3070                         tp->tsoffset = val - tcp_time_stamp_raw();
3071                 break;
3072         case TCP_REPAIR_WINDOW:
3073                 err = tcp_repair_set_window(tp, optval, optlen);
3074                 break;
3075         case TCP_NOTSENT_LOWAT:
3076                 tp->notsent_lowat = val;
3077                 sk->sk_write_space(sk);
3078                 break;
3079         case TCP_INQ:
3080                 if (val > 1 || val < 0)
3081                         err = -EINVAL;
3082                 else
3083                         tp->recvmsg_inq = val;
3084                 break;
3085         default:
3086                 err = -ENOPROTOOPT;
3087                 break;
3088         }
3089
3090         release_sock(sk);
3091         return err;
3092 }
3093
3094 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
3095                    unsigned int optlen)
3096 {
3097         const struct inet_connection_sock *icsk = inet_csk(sk);
3098
3099         if (level != SOL_TCP)
3100                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3101                                                      optval, optlen);
3102         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3103 }
3104 EXPORT_SYMBOL(tcp_setsockopt);
3105
3106 #ifdef CONFIG_COMPAT
3107 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
3108                           char __user *optval, unsigned int optlen)
3109 {
3110         if (level != SOL_TCP)
3111                 return inet_csk_compat_setsockopt(sk, level, optname,
3112                                                   optval, optlen);
3113         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3114 }
3115 EXPORT_SYMBOL(compat_tcp_setsockopt);
3116 #endif
3117
3118 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3119                                       struct tcp_info *info)
3120 {
3121         u64 stats[__TCP_CHRONO_MAX], total = 0;
3122         enum tcp_chrono i;
3123
3124         for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3125                 stats[i] = tp->chrono_stat[i - 1];
3126                 if (i == tp->chrono_type)
3127                         stats[i] += tcp_jiffies32 - tp->chrono_start;
3128                 stats[i] *= USEC_PER_SEC / HZ;
3129                 total += stats[i];
3130         }
3131
3132         info->tcpi_busy_time = total;
3133         info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3134         info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3135 }
3136
3137 /* Return information about state of tcp endpoint in API format. */
3138 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3139 {
3140         const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3141         const struct inet_connection_sock *icsk = inet_csk(sk);
3142         unsigned long rate;
3143         u32 now;
3144         u64 rate64;
3145         bool slow;
3146
3147         memset(info, 0, sizeof(*info));
3148         if (sk->sk_type != SOCK_STREAM)
3149                 return;
3150
3151         info->tcpi_state = inet_sk_state_load(sk);
3152
3153         /* Report meaningful fields for all TCP states, including listeners */
3154         rate = READ_ONCE(sk->sk_pacing_rate);
3155         rate64 = (rate != ~0UL) ? rate : ~0ULL;
3156         info->tcpi_pacing_rate = rate64;
3157
3158         rate = READ_ONCE(sk->sk_max_pacing_rate);
3159         rate64 = (rate != ~0UL) ? rate : ~0ULL;
3160         info->tcpi_max_pacing_rate = rate64;
3161
3162         info->tcpi_reordering = tp->reordering;
3163         info->tcpi_snd_cwnd = tp->snd_cwnd;
3164
3165         if (info->tcpi_state == TCP_LISTEN) {
3166                 /* listeners aliased fields :
3167                  * tcpi_unacked -> Number of children ready for accept()
3168                  * tcpi_sacked  -> max backlog
3169                  */
3170                 info->tcpi_unacked = sk->sk_ack_backlog;
3171                 info->tcpi_sacked = sk->sk_max_ack_backlog;
3172                 return;
3173         }
3174
3175         slow = lock_sock_fast(sk);
3176
3177         info->tcpi_ca_state = icsk->icsk_ca_state;
3178         info->tcpi_retransmits = icsk->icsk_retransmits;
3179         info->tcpi_probes = icsk->icsk_probes_out;
3180         info->tcpi_backoff = icsk->icsk_backoff;
3181
3182         if (tp->rx_opt.tstamp_ok)
3183                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3184         if (tcp_is_sack(tp))
3185                 info->tcpi_options |= TCPI_OPT_SACK;
3186         if (tp->rx_opt.wscale_ok) {
3187                 info->tcpi_options |= TCPI_OPT_WSCALE;
3188                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3189                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3190         }
3191
3192         if (tp->ecn_flags & TCP_ECN_OK)
3193                 info->tcpi_options |= TCPI_OPT_ECN;
3194         if (tp->ecn_flags & TCP_ECN_SEEN)
3195                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3196         if (tp->syn_data_acked)
3197                 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3198
3199         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3200         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3201         info->tcpi_snd_mss = tp->mss_cache;
3202         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3203
3204         info->tcpi_unacked = tp->packets_out;
3205         info->tcpi_sacked = tp->sacked_out;
3206
3207         info->tcpi_lost = tp->lost_out;
3208         info->tcpi_retrans = tp->retrans_out;
3209
3210         now = tcp_jiffies32;
3211         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3212         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3213         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3214
3215         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3216         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3217         info->tcpi_rtt = tp->srtt_us >> 3;
3218         info->tcpi_rttvar = tp->mdev_us >> 2;
3219         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3220         info->tcpi_advmss = tp->advmss;
3221
3222         info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3223         info->tcpi_rcv_space = tp->rcvq_space.space;
3224
3225         info->tcpi_total_retrans = tp->total_retrans;
3226
3227         info->tcpi_bytes_acked = tp->bytes_acked;
3228         info->tcpi_bytes_received = tp->bytes_received;
3229         info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3230         tcp_get_info_chrono_stats(tp, info);
3231
3232         info->tcpi_segs_out = tp->segs_out;
3233         info->tcpi_segs_in = tp->segs_in;
3234
3235         info->tcpi_min_rtt = tcp_min_rtt(tp);
3236         info->tcpi_data_segs_in = tp->data_segs_in;
3237         info->tcpi_data_segs_out = tp->data_segs_out;
3238
3239         info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3240         rate64 = tcp_compute_delivery_rate(tp);
3241         if (rate64)
3242                 info->tcpi_delivery_rate = rate64;
3243         info->tcpi_delivered = tp->delivered;
3244         info->tcpi_delivered_ce = tp->delivered_ce;
3245         info->tcpi_bytes_sent = tp->bytes_sent;
3246         info->tcpi_bytes_retrans = tp->bytes_retrans;
3247         info->tcpi_dsack_dups = tp->dsack_dups;
3248         info->tcpi_reord_seen = tp->reord_seen;
3249         unlock_sock_fast(sk, slow);
3250 }
3251 EXPORT_SYMBOL_GPL(tcp_get_info);
3252
3253 static size_t tcp_opt_stats_get_size(void)
3254 {
3255         return
3256                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3257                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3258                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3259                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3260                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3261                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3262                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3263                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3264                 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3265                 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3266                 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3267                 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3268                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3269                 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3270                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3271                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3272                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3273                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3274                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3275                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3276                 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3277                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3278                 0;
3279 }
3280
3281 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk)
3282 {
3283         const struct tcp_sock *tp = tcp_sk(sk);
3284         struct sk_buff *stats;
3285         struct tcp_info info;
3286         unsigned long rate;
3287         u64 rate64;
3288
3289         stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3290         if (!stats)
3291                 return NULL;
3292
3293         tcp_get_info_chrono_stats(tp, &info);
3294         nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3295                           info.tcpi_busy_time, TCP_NLA_PAD);
3296         nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3297                           info.tcpi_rwnd_limited, TCP_NLA_PAD);
3298         nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3299                           info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3300         nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3301                           tp->data_segs_out, TCP_NLA_PAD);
3302         nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3303                           tp->total_retrans, TCP_NLA_PAD);
3304
3305         rate = READ_ONCE(sk->sk_pacing_rate);
3306         rate64 = (rate != ~0UL) ? rate : ~0ULL;
3307         nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3308
3309         rate64 = tcp_compute_delivery_rate(tp);
3310         nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3311
3312         nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3313         nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3314         nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3315
3316         nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3317         nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3318         nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3319         nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3320         nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3321
3322         nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3323         nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3324
3325         nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3326                           TCP_NLA_PAD);
3327         nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3328                           TCP_NLA_PAD);
3329         nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3330         nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3331         nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3332
3333         return stats;
3334 }
3335
3336 static int do_tcp_getsockopt(struct sock *sk, int level,
3337                 int optname, char __user *optval, int __user *optlen)
3338 {
3339         struct inet_connection_sock *icsk = inet_csk(sk);
3340         struct tcp_sock *tp = tcp_sk(sk);
3341         struct net *net = sock_net(sk);
3342         int val, len;
3343
3344         if (get_user(len, optlen))
3345                 return -EFAULT;
3346
3347         len = min_t(unsigned int, len, sizeof(int));
3348
3349         if (len < 0)
3350                 return -EINVAL;
3351
3352         switch (optname) {
3353         case TCP_MAXSEG:
3354                 val = tp->mss_cache;
3355                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3356                         val = tp->rx_opt.user_mss;
3357                 if (tp->repair)
3358                         val = tp->rx_opt.mss_clamp;
3359                 break;
3360         case TCP_NODELAY:
3361                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3362                 break;
3363         case TCP_CORK:
3364                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3365                 break;
3366         case TCP_KEEPIDLE:
3367                 val = keepalive_time_when(tp) / HZ;
3368                 break;
3369         case TCP_KEEPINTVL:
3370                 val = keepalive_intvl_when(tp) / HZ;
3371                 break;
3372         case TCP_KEEPCNT:
3373                 val = keepalive_probes(tp);
3374                 break;
3375         case TCP_SYNCNT:
3376                 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3377                 break;
3378         case TCP_LINGER2:
3379                 val = tp->linger2;
3380                 if (val >= 0)
3381                         val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3382                 break;
3383         case TCP_DEFER_ACCEPT:
3384                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3385                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3386                 break;
3387         case TCP_WINDOW_CLAMP:
3388                 val = tp->window_clamp;
3389                 break;
3390         case TCP_INFO: {
3391                 struct tcp_info info;
3392
3393                 if (get_user(len, optlen))
3394                         return -EFAULT;
3395
3396                 tcp_get_info(sk, &info);
3397
3398                 len = min_t(unsigned int, len, sizeof(info));
3399                 if (put_user(len, optlen))
3400                         return -EFAULT;
3401                 if (copy_to_user(optval, &info, len))
3402                         return -EFAULT;
3403                 return 0;
3404         }
3405         case TCP_CC_INFO: {
3406                 const struct tcp_congestion_ops *ca_ops;
3407                 union tcp_cc_info info;
3408                 size_t sz = 0;
3409                 int attr;
3410
3411                 if (get_user(len, optlen))
3412                         return -EFAULT;
3413
3414                 ca_ops = icsk->icsk_ca_ops;
3415                 if (ca_ops && ca_ops->get_info)
3416                         sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3417
3418                 len = min_t(unsigned int, len, sz);
3419                 if (put_user(len, optlen))
3420                         return -EFAULT;
3421                 if (copy_to_user(optval, &info, len))
3422                         return -EFAULT;
3423                 return 0;
3424         }
3425         case TCP_QUICKACK:
3426                 val = !inet_csk_in_pingpong_mode(sk);
3427                 break;
3428
3429         case TCP_CONGESTION:
3430                 if (get_user(len, optlen))
3431                         return -EFAULT;
3432                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
3433                 if (put_user(len, optlen))
3434                         return -EFAULT;
3435                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3436                         return -EFAULT;
3437                 return 0;
3438
3439         case TCP_ULP:
3440                 if (get_user(len, optlen))
3441                         return -EFAULT;
3442                 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
3443                 if (!icsk->icsk_ulp_ops) {
3444                         if (put_user(0, optlen))
3445                                 return -EFAULT;
3446                         return 0;
3447                 }
3448                 if (put_user(len, optlen))
3449                         return -EFAULT;
3450                 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
3451                         return -EFAULT;
3452                 return 0;
3453
3454         case TCP_FASTOPEN_KEY: {
3455                 __u8 key[TCP_FASTOPEN_KEY_LENGTH];
3456                 struct tcp_fastopen_context *ctx;
3457
3458                 if (get_user(len, optlen))
3459                         return -EFAULT;
3460
3461                 rcu_read_lock();
3462                 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
3463                 if (ctx)
3464                         memcpy(key, ctx->key, sizeof(key));
3465                 else
3466                         len = 0;
3467                 rcu_read_unlock();
3468
3469                 len = min_t(unsigned int, len, sizeof(key));
3470                 if (put_user(len, optlen))
3471                         return -EFAULT;
3472                 if (copy_to_user(optval, key, len))
3473                         return -EFAULT;
3474                 return 0;
3475         }
3476         case TCP_THIN_LINEAR_TIMEOUTS:
3477                 val = tp->thin_lto;
3478                 break;
3479
3480         case TCP_THIN_DUPACK:
3481                 val = 0;
3482                 break;
3483
3484         case TCP_REPAIR:
3485                 val = tp->repair;
3486                 break;
3487
3488         case TCP_REPAIR_QUEUE:
3489                 if (tp->repair)
3490                         val = tp->repair_queue;
3491                 else
3492                         return -EINVAL;
3493                 break;
3494
3495         case TCP_REPAIR_WINDOW: {
3496                 struct tcp_repair_window opt;
3497
3498                 if (get_user(len, optlen))
3499                         return -EFAULT;
3500
3501                 if (len != sizeof(opt))
3502                         return -EINVAL;
3503
3504                 if (!tp->repair)
3505                         return -EPERM;
3506
3507                 opt.snd_wl1     = tp->snd_wl1;
3508                 opt.snd_wnd     = tp->snd_wnd;
3509                 opt.max_window  = tp->max_window;
3510                 opt.rcv_wnd     = tp->rcv_wnd;
3511                 opt.rcv_wup     = tp->rcv_wup;
3512
3513                 if (copy_to_user(optval, &opt, len))
3514                         return -EFAULT;
3515                 return 0;
3516         }
3517         case TCP_QUEUE_SEQ:
3518                 if (tp->repair_queue == TCP_SEND_QUEUE)
3519                         val = tp->write_seq;
3520                 else if (tp->repair_queue == TCP_RECV_QUEUE)
3521                         val = tp->rcv_nxt;
3522                 else
3523                         return -EINVAL;
3524                 break;
3525
3526         case TCP_USER_TIMEOUT:
3527                 val = icsk->icsk_user_timeout;
3528                 break;
3529
3530         case TCP_FASTOPEN:
3531                 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3532                 break;
3533
3534         case TCP_FASTOPEN_CONNECT:
3535                 val = tp->fastopen_connect;
3536                 break;
3537
3538         case TCP_FASTOPEN_NO_COOKIE:
3539                 val = tp->fastopen_no_cookie;
3540                 break;
3541
3542         case TCP_TIMESTAMP:
3543                 val = tcp_time_stamp_raw() + tp->tsoffset;
3544                 break;
3545         case TCP_NOTSENT_LOWAT:
3546                 val = tp->notsent_lowat;
3547                 break;
3548         case TCP_INQ:
3549                 val = tp->recvmsg_inq;
3550                 break;
3551         case TCP_SAVE_SYN:
3552                 val = tp->save_syn;
3553                 break;
3554         case TCP_SAVED_SYN: {
3555                 if (get_user(len, optlen))
3556                         return -EFAULT;
3557
3558                 lock_sock(sk);
3559                 if (tp->saved_syn) {
3560                         if (len < tp->saved_syn[0]) {
3561                                 if (put_user(tp->saved_syn[0], optlen)) {
3562                                         release_sock(sk);
3563                                         return -EFAULT;
3564                                 }
3565                                 release_sock(sk);
3566                                 return -EINVAL;
3567                         }
3568                         len = tp->saved_syn[0];
3569                         if (put_user(len, optlen)) {
3570                                 release_sock(sk);
3571                                 return -EFAULT;
3572                         }
3573                         if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3574                                 release_sock(sk);
3575                                 return -EFAULT;
3576                         }
3577                         tcp_saved_syn_free(tp);
3578                         release_sock(sk);
3579                 } else {
3580                         release_sock(sk);
3581                         len = 0;
3582                         if (put_user(len, optlen))
3583                                 return -EFAULT;
3584                 }
3585                 return 0;
3586         }
3587 #ifdef CONFIG_MMU
3588         case TCP_ZEROCOPY_RECEIVE: {
3589                 struct tcp_zerocopy_receive zc;
3590                 int err;
3591
3592                 if (get_user(len, optlen))
3593                         return -EFAULT;
3594                 if (len != sizeof(zc))
3595                         return -EINVAL;
3596                 if (copy_from_user(&zc, optval, len))
3597                         return -EFAULT;
3598                 lock_sock(sk);
3599                 err = tcp_zerocopy_receive(sk, &zc);
3600                 release_sock(sk);
3601                 if (!err && copy_to_user(optval, &zc, len))
3602                         err = -EFAULT;
3603                 return err;
3604         }
3605 #endif
3606         default:
3607                 return -ENOPROTOOPT;
3608         }
3609
3610         if (put_user(len, optlen))
3611                 return -EFAULT;
3612         if (copy_to_user(optval, &val, len))
3613                 return -EFAULT;
3614         return 0;
3615 }
3616
3617 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3618                    int __user *optlen)
3619 {
3620         struct inet_connection_sock *icsk = inet_csk(sk);
3621
3622         if (level != SOL_TCP)
3623                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3624                                                      optval, optlen);
3625         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3626 }
3627 EXPORT_SYMBOL(tcp_getsockopt);
3628
3629 #ifdef CONFIG_COMPAT
3630 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3631                           char __user *optval, int __user *optlen)
3632 {
3633         if (level != SOL_TCP)
3634                 return inet_csk_compat_getsockopt(sk, level, optname,
3635                                                   optval, optlen);
3636         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3637 }
3638 EXPORT_SYMBOL(compat_tcp_getsockopt);
3639 #endif
3640
3641 #ifdef CONFIG_TCP_MD5SIG
3642 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3643 static DEFINE_MUTEX(tcp_md5sig_mutex);
3644 static bool tcp_md5sig_pool_populated = false;
3645
3646 static void __tcp_alloc_md5sig_pool(void)
3647 {
3648         struct crypto_ahash *hash;
3649         int cpu;
3650
3651         hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3652         if (IS_ERR(hash))
3653                 return;
3654
3655         for_each_possible_cpu(cpu) {
3656                 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3657                 struct ahash_request *req;
3658
3659                 if (!scratch) {
3660                         scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3661                                                sizeof(struct tcphdr),
3662                                                GFP_KERNEL,
3663                                                cpu_to_node(cpu));
3664                         if (!scratch)
3665                                 return;
3666                         per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3667                 }
3668                 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3669                         continue;
3670
3671                 req = ahash_request_alloc(hash, GFP_KERNEL);
3672                 if (!req)
3673                         return;
3674
3675                 ahash_request_set_callback(req, 0, NULL, NULL);
3676
3677                 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3678         }
3679         /* before setting tcp_md5sig_pool_populated, we must commit all writes
3680          * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3681          */
3682         smp_wmb();
3683         tcp_md5sig_pool_populated = true;
3684 }
3685
3686 bool tcp_alloc_md5sig_pool(void)
3687 {
3688         if (unlikely(!tcp_md5sig_pool_populated)) {
3689                 mutex_lock(&tcp_md5sig_mutex);
3690
3691                 if (!tcp_md5sig_pool_populated) {
3692                         __tcp_alloc_md5sig_pool();
3693                         if (tcp_md5sig_pool_populated)
3694                                 static_branch_inc(&tcp_md5_needed);
3695                 }
3696
3697                 mutex_unlock(&tcp_md5sig_mutex);
3698         }
3699         return tcp_md5sig_pool_populated;
3700 }
3701 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3702
3703
3704 /**
3705  *      tcp_get_md5sig_pool - get md5sig_pool for this user
3706  *
3707  *      We use percpu structure, so if we succeed, we exit with preemption
3708  *      and BH disabled, to make sure another thread or softirq handling
3709  *      wont try to get same context.
3710  */
3711 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3712 {
3713         local_bh_disable();
3714
3715         if (tcp_md5sig_pool_populated) {
3716                 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3717                 smp_rmb();
3718                 return this_cpu_ptr(&tcp_md5sig_pool);
3719         }
3720         local_bh_enable();
3721         return NULL;
3722 }
3723 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3724
3725 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3726                           const struct sk_buff *skb, unsigned int header_len)
3727 {
3728         struct scatterlist sg;
3729         const struct tcphdr *tp = tcp_hdr(skb);
3730         struct ahash_request *req = hp->md5_req;
3731         unsigned int i;
3732         const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3733                                            skb_headlen(skb) - header_len : 0;
3734         const struct skb_shared_info *shi = skb_shinfo(skb);
3735         struct sk_buff *frag_iter;
3736
3737         sg_init_table(&sg, 1);
3738
3739         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3740         ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3741         if (crypto_ahash_update(req))
3742                 return 1;
3743
3744         for (i = 0; i < shi->nr_frags; ++i) {
3745                 const struct skb_frag_struct *f = &shi->frags[i];
3746                 unsigned int offset = f->page_offset;
3747                 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3748
3749                 sg_set_page(&sg, page, skb_frag_size(f),
3750                             offset_in_page(offset));
3751                 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3752                 if (crypto_ahash_update(req))
3753                         return 1;
3754         }
3755
3756         skb_walk_frags(skb, frag_iter)
3757                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3758                         return 1;
3759
3760         return 0;
3761 }
3762 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3763
3764 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3765 {
3766         struct scatterlist sg;
3767
3768         sg_init_one(&sg, key->key, key->keylen);
3769         ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen);
3770         return crypto_ahash_update(hp->md5_req);
3771 }
3772 EXPORT_SYMBOL(tcp_md5_hash_key);
3773
3774 #endif
3775
3776 void tcp_done(struct sock *sk)
3777 {
3778         struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3779
3780         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3781                 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3782
3783         tcp_set_state(sk, TCP_CLOSE);
3784         tcp_clear_xmit_timers(sk);
3785         if (req)
3786                 reqsk_fastopen_remove(sk, req, false);
3787
3788         sk->sk_shutdown = SHUTDOWN_MASK;
3789
3790         if (!sock_flag(sk, SOCK_DEAD))
3791                 sk->sk_state_change(sk);
3792         else
3793                 inet_csk_destroy_sock(sk);
3794 }
3795 EXPORT_SYMBOL_GPL(tcp_done);
3796
3797 int tcp_abort(struct sock *sk, int err)
3798 {
3799         if (!sk_fullsock(sk)) {
3800                 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3801                         struct request_sock *req = inet_reqsk(sk);
3802
3803                         local_bh_disable();
3804                         inet_csk_reqsk_queue_drop(req->rsk_listener, req);
3805                         local_bh_enable();
3806                         return 0;
3807                 }
3808                 return -EOPNOTSUPP;
3809         }
3810
3811         /* Don't race with userspace socket closes such as tcp_close. */
3812         lock_sock(sk);
3813
3814         if (sk->sk_state == TCP_LISTEN) {
3815                 tcp_set_state(sk, TCP_CLOSE);
3816                 inet_csk_listen_stop(sk);
3817         }
3818
3819         /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3820         local_bh_disable();
3821         bh_lock_sock(sk);
3822
3823         if (!sock_flag(sk, SOCK_DEAD)) {
3824                 sk->sk_err = err;
3825                 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3826                 smp_wmb();
3827                 sk->sk_error_report(sk);
3828                 if (tcp_need_reset(sk->sk_state))
3829                         tcp_send_active_reset(sk, GFP_ATOMIC);
3830                 tcp_done(sk);
3831         }
3832
3833         bh_unlock_sock(sk);
3834         local_bh_enable();
3835         tcp_write_queue_purge(sk);
3836         release_sock(sk);
3837         return 0;
3838 }
3839 EXPORT_SYMBOL_GPL(tcp_abort);
3840
3841 extern struct tcp_congestion_ops tcp_reno;
3842
3843 static __initdata unsigned long thash_entries;
3844 static int __init set_thash_entries(char *str)
3845 {
3846         ssize_t ret;
3847
3848         if (!str)
3849                 return 0;
3850
3851         ret = kstrtoul(str, 0, &thash_entries);
3852         if (ret)
3853                 return 0;
3854
3855         return 1;
3856 }
3857 __setup("thash_entries=", set_thash_entries);
3858
3859 static void __init tcp_init_mem(void)
3860 {
3861         unsigned long limit = nr_free_buffer_pages() / 16;
3862
3863         limit = max(limit, 128UL);
3864         sysctl_tcp_mem[0] = limit / 4 * 3;              /* 4.68 % */
3865         sysctl_tcp_mem[1] = limit;                      /* 6.25 % */
3866         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;      /* 9.37 % */
3867 }
3868
3869 void __init tcp_init(void)
3870 {
3871         int max_rshare, max_wshare, cnt;
3872         unsigned long limit;
3873         unsigned int i;
3874
3875         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3876                      FIELD_SIZEOF(struct sk_buff, cb));
3877
3878         percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3879         percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3880         inet_hashinfo_init(&tcp_hashinfo);
3881         inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
3882                             thash_entries, 21,  /* one slot per 2 MB*/
3883                             0, 64 * 1024);
3884         tcp_hashinfo.bind_bucket_cachep =
3885                 kmem_cache_create("tcp_bind_bucket",
3886                                   sizeof(struct inet_bind_bucket), 0,
3887                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3888
3889         /* Size and allocate the main established and bind bucket
3890          * hash tables.
3891          *
3892          * The methodology is similar to that of the buffer cache.
3893          */
3894         tcp_hashinfo.ehash =
3895                 alloc_large_system_hash("TCP established",
3896                                         sizeof(struct inet_ehash_bucket),
3897                                         thash_entries,
3898                                         17, /* one slot per 128 KB of memory */
3899                                         0,
3900                                         NULL,
3901                                         &tcp_hashinfo.ehash_mask,
3902                                         0,
3903                                         thash_entries ? 0 : 512 * 1024);
3904         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3905                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3906
3907         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3908                 panic("TCP: failed to alloc ehash_locks");
3909         tcp_hashinfo.bhash =
3910                 alloc_large_system_hash("TCP bind",
3911                                         sizeof(struct inet_bind_hashbucket),
3912                                         tcp_hashinfo.ehash_mask + 1,
3913                                         17, /* one slot per 128 KB of memory */
3914                                         0,
3915                                         &tcp_hashinfo.bhash_size,
3916                                         NULL,
3917                                         0,
3918                                         64 * 1024);
3919         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3920         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3921                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3922                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3923         }
3924
3925
3926         cnt = tcp_hashinfo.ehash_mask + 1;
3927         sysctl_tcp_max_orphans = cnt / 2;
3928
3929         tcp_init_mem();
3930         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3931         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3932         max_wshare = min(4UL*1024*1024, limit);
3933         max_rshare = min(6UL*1024*1024, limit);
3934
3935         init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3936         init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
3937         init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3938
3939         init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3940         init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
3941         init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
3942
3943         pr_info("Hash tables configured (established %u bind %u)\n",
3944                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3945
3946         tcp_v4_init();
3947         tcp_metrics_init();
3948         BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3949         tcp_tasklet_init();
3950 }