net/ipv4/tcp.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the  BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              Implementation of the Transmission Control Protocol(TCP).
   7  *
   8  * Authors:     Ross Biro
   9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
  11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
  12  *              Florian La Roche, <flla@stud.uni-sb.de>
  13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
  15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
  16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
  17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  18  *              Jorge Cwik, <jorge@laser.satlink.net>
  19  *
  20  * Fixes:
  21  *              Alan Cox        :       Numerous verify_area() calls
  22  *              Alan Cox        :       Set the ACK bit on a reset
  23  *              Alan Cox        :       Stopped it crashing if it closed while
  24  *                                      sk->inuse=1 and was trying to connect
  25  *                                      (tcp_err()).
  26  *              Alan Cox        :       All icmp error handling was broken
  27  *                                      pointers passed where wrong and the
  28  *                                      socket was looked up backwards. Nobody
  29  *                                      tested any icmp error code obviously.
  30  *              Alan Cox        :       tcp_err() now handled properly. It
  31  *                                      wakes people on errors. poll
  32  *                                      behaves and the icmp error race
  33  *                                      has gone by moving it into sock.c
  34  *              Alan Cox        :       tcp_send_reset() fixed to work for
  35  *                                      everything not just packets for
  36  *                                      unknown sockets.
  37  *              Alan Cox        :       tcp option processing.
  38  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
  39  *                                      syn rule wrong]
  40  *              Herp Rosmanith  :       More reset fixes
  41  *              Alan Cox        :       No longer acks invalid rst frames.
  42  *                                      Acking any kind of RST is right out.
  43  *              Alan Cox        :       Sets an ignore me flag on an rst
  44  *                                      receive otherwise odd bits of prattle
  45  *                                      escape still
  46  *              Alan Cox        :       Fixed another acking RST frame bug.
  47  *                                      Should stop LAN workplace lockups.
  48  *              Alan Cox        :       Some tidyups using the new skb list
  49  *                                      facilities
  50  *              Alan Cox        :       sk->keepopen now seems to work
  51  *              Alan Cox        :       Pulls options out correctly on accepts
  52  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
  53  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
  54  *                                      bit to skb ops.
  55  *              Alan Cox        :       Tidied tcp_data to avoid a potential
  56  *                                      nasty.
  57  *              Alan Cox        :       Added some better commenting, as the
  58  *                                      tcp is hard to follow
  59  *              Alan Cox        :       Removed incorrect check for 20 * psh
  60  *      Michael O'Reilly        :       ack < copied bug fix.
  61  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
  62  *              Alan Cox        :       FIN with no memory -> CRASH
  63  *              Alan Cox        :       Added socket option proto entries.
  64  *                                      Also added awareness of them to accept.
  65  *              Alan Cox        :       Added TCP options (SOL_TCP)
  66  *              Alan Cox        :       Switched wakeup calls to callbacks,
  67  *                                      so the kernel can layer network
  68  *                                      sockets.
  69  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
  70  *              Alan Cox        :       Handle FIN (more) properly (we hope).
  71  *              Alan Cox        :       RST frames sent on unsynchronised
  72  *                                      state ack error.
  73  *              Alan Cox        :       Put in missing check for SYN bit.
  74  *              Alan Cox        :       Added tcp_select_window() aka NET2E
  75  *                                      window non shrink trick.
  76  *              Alan Cox        :       Added a couple of small NET2E timer
  77  *                                      fixes
  78  *              Charles Hedrick :       TCP fixes
  79  *              Toomas Tamm     :       TCP window fixes
  80  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
  81  *              Charles Hedrick :       Rewrote most of it to actually work
  82  *              Linus           :       Rewrote tcp_read() and URG handling
  83  *                                      completely
  84  *              Gerhard Koerting:       Fixed some missing timer handling
  85  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
  86  *              Gerhard Koerting:       PC/TCP workarounds
  87  *              Adam Caldwell   :       Assorted timer/timing errors
  88  *              Matthew Dillon  :       Fixed another RST bug
  89  *              Alan Cox        :       Move to kernel side addressing changes.
  90  *              Alan Cox        :       Beginning work on TCP fastpathing
  91  *                                      (not yet usable)
  92  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
  93  *              Alan Cox        :       TCP fast path debugging
  94  *              Alan Cox        :       Window clamping
  95  *              Michael Riepe   :       Bug in tcp_check()
  96  *              Matt Dillon     :       More TCP improvements and RST bug fixes
  97  *              Matt Dillon     :       Yet more small nasties remove from the
  98  *                                      TCP code (Be very nice to this man if
  99  *                                      tcp finally works 100%) 8)
 100  *              Alan Cox        :       BSD accept semantics.
 101  *              Alan Cox        :       Reset on closedown bug.
 102  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
 103  *              Michael Pall    :       Handle poll() after URG properly in
 104  *                                      all cases.
 105  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
 106  *                                      (multi URG PUSH broke rlogin).
 107  *              Michael Pall    :       Fix the multi URG PUSH problem in
 108  *                                      tcp_readable(), poll() after URG
 109  *                                      works now.
 110  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
 111  *                                      BSD api.
 112  *              Alan Cox        :       Changed the semantics of sk->socket to
 113  *                                      fix a race and a signal problem with
 114  *                                      accept() and async I/O.
 115  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
 116  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
 117  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
 118  *                                      clients/servers which listen in on
 119  *                                      fixed ports.
 120  *              Alan Cox        :       Cleaned the above up and shrank it to
 121  *                                      a sensible code size.
 122  *              Alan Cox        :       Self connect lockup fix.
 123  *              Alan Cox        :       No connect to multicast.
 124  *              Ross Biro       :       Close unaccepted children on master
 125  *                                      socket close.
 126  *              Alan Cox        :       Reset tracing code.
 127  *              Alan Cox        :       Spurious resets on shutdown.
 128  *              Alan Cox        :       Giant 15 minute/60 second timer error
 129  *              Alan Cox        :       Small whoops in polling before an
 130  *                                      accept.
 131  *              Alan Cox        :       Kept the state trace facility since
 132  *                                      it's handy for debugging.
 133  *              Alan Cox        :       More reset handler fixes.
 134  *              Alan Cox        :       Started rewriting the code based on
 135  *                                      the RFC's for other useful protocol
 136  *                                      references see: Comer, KA9Q NOS, and
 137  *                                      for a reference on the difference
 138  *                                      between specifications and how BSD
 139  *                                      works see the 4.4lite source.
 140  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
 141  *                                      close.
 142  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
 143  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
 144  *              Alan Cox        :       Reimplemented timers as per the RFC
 145  *                                      and using multiple timers for sanity.
 146  *              Alan Cox        :       Small bug fixes, and a lot of new
 147  *                                      comments.
 148  *              Alan Cox        :       Fixed dual reader crash by locking
 149  *                                      the buffers (much like datagram.c)
 150  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
 151  *                                      now gets fed up of retrying without
 152  *                                      (even a no space) answer.
 153  *              Alan Cox        :       Extracted closing code better
 154  *              Alan Cox        :       Fixed the closing state machine to
 155  *                                      resemble the RFC.
 156  *              Alan Cox        :       More 'per spec' fixes.
 157  *              Jorge Cwik      :       Even faster checksumming.
 158  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
 159  *                                      only frames. At least one pc tcp stack
 160  *                                      generates them.
 161  *              Alan Cox        :       Cache last socket.
 162  *              Alan Cox        :       Per route irtt.
 163  *              Matt Day        :       poll()->select() match BSD precisely on error
 164  *              Alan Cox        :       New buffers
 165  *              Marc Tamsky     :       Various sk->prot->retransmits and
 166  *                                      sk->retransmits misupdating fixed.
 167  *                                      Fixed tcp_write_timeout: stuck close,
 168  *                                      and TCP syn retries gets used now.
 169  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
 170  *                                      ack if state is TCP_CLOSED.
 171  *              Alan Cox        :       Look up device on a retransmit - routes may
 172  *                                      change. Doesn't yet cope with MSS shrink right
 173  *                                      but it's a start!
 174  *              Marc Tamsky     :       Closing in closing fixes.
 175  *              Mike Shaver     :       RFC1122 verifications.
 176  *              Alan Cox        :       rcv_saddr errors.
 177  *              Alan Cox        :       Block double connect().
 178  *              Alan Cox        :       Small hooks for enSKIP.
 179  *              Alexey Kuznetsov:       Path MTU discovery.
 180  *              Alan Cox        :       Support soft errors.
 181  *              Alan Cox        :       Fix MTU discovery pathological case
 182  *                                      when the remote claims no mtu!
 183  *              Marc Tamsky     :       TCP_CLOSE fix.
 184  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
 185  *                                      window but wrong (fixes NT lpd problems)
 186  *              Pedro Roque     :       Better TCP window handling, delayed ack.
 187  *              Joerg Reuter    :       No modification of locked buffers in
 188  *                                      tcp_do_retransmit()
 189  *              Eric Schenk     :       Changed receiver side silly window
 190  *                                      avoidance algorithm to BSD style
 191  *                                      algorithm. This doubles throughput
 192  *                                      against machines running Solaris,
 193  *                                      and seems to result in general
 194  *                                      improvement.
 195  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
 196  *      Willy Konynenberg       :       Transparent proxying support.
 197  *      Mike McLagan            :       Routing by source
 198  *              Keith Owens     :       Do proper merging with partial SKB's in
 199  *                                      tcp_do_sendmsg to avoid burstiness.
 200  *              Eric Schenk     :       Fix fast close down bug with
 201  *                                      shutdown() followed by close().
 202  *              Andi Kleen      :       Make poll agree with SIGIO
 203  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
 204  *                                      lingertime == 0 (RFC 793 ABORT Call)
 205  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
 206  *                                      csum_and_copy_from_user() if possible.
 207  *
 208  *              This program is free software; you can redistribute it and/or
 209  *              modify it under the terms of the GNU General Public License
 210  *              as published by the Free Software Foundation; either version
 211  *              2 of the License, or(at your option) any later version.
 212  *
 213  * Description of States:
 214  *
 215  *      TCP_SYN_SENT            sent a connection request, waiting for ack
 216  *
 217  *      TCP_SYN_RECV            received a connection request, sent ack,
 218  *                              waiting for final ack in three-way handshake.
 219  *
 220  *      TCP_ESTABLISHED         connection established
 221  *
 222  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
 223  *                              transmission of remaining buffered data
 224  *
 225  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
 226  *                              to shutdown
 227  *
 228  *      TCP_CLOSING             both sides have shutdown but we still have
 229  *                              data we have to finish sending
 230  *
 231  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
 232  *                              closed, can only be entered from FIN_WAIT2
 233  *                              or CLOSING.  Required because the other end
 234  *                              may not have gotten our last ACK causing it
 235  *                              to retransmit the data packet (which we ignore)
 236  *
 237  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
 238  *                              us to finish writing our data and to shutdown
 239  *                              (we have to close() to move on to LAST_ACK)
 240  *
 241  *      TCP_LAST_ACK            out side has shutdown after remote has
 242  *                              shutdown.  There may still be data in our
 243  *                              buffer that we have to finish sending
 244  *
 245  *      TCP_CLOSE               socket is finished
 246  */
 247
 248 #define pr_fmt(fmt) "TCP: " fmt
 249
 250 #include <crypto/hash.h>
 251 #include <linux/kernel.h>
 252 #include <linux/module.h>
 253 #include <linux/types.h>
 254 #include <linux/fcntl.h>
 255 #include <linux/poll.h>
 256 #include <linux/inet_diag.h>
 257 #include <linux/init.h>
 258 #include <linux/fs.h>
 259 #include <linux/skbuff.h>
 260 #include <linux/scatterlist.h>
 261 #include <linux/splice.h>
 262 #include <linux/net.h>
 263 #include <linux/socket.h>
 264 #include <linux/random.h>
 265 #include <linux/memblock.h>
 266 #include <linux/highmem.h>
 267 #include <linux/swap.h>
 268 #include <linux/cache.h>
 269 #include <linux/err.h>
 270 #include <linux/time.h>
 271 #include <linux/slab.h>
 272 #include <linux/errqueue.h>
 273 #include <linux/static_key.h>
 274
 275 #include <net/icmp.h>
 276 #include <net/inet_common.h>
 277 #include <net/tcp.h>
 278 #include <net/xfrm.h>
 279 #include <net/ip.h>
 280 #include <net/sock.h>
 281
 282 #include <linux/uaccess.h>
 283 #include <asm/ioctls.h>
 284 #include <net/busy_poll.h>
 285
 286 struct percpu_counter tcp_orphan_count;
 287 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 288
 289 long sysctl_tcp_mem[3] __read_mostly;
 290 EXPORT_SYMBOL(sysctl_tcp_mem);
 291
 292 atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
 293 EXPORT_SYMBOL(tcp_memory_allocated);
 294
 295 #if IS_ENABLED(CONFIG_SMC)
 296 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
 297 EXPORT_SYMBOL(tcp_have_smc);
 298 #endif
 299
 300 /*
 301  * Current number of TCP sockets.
 302  */
 303 struct percpu_counter tcp_sockets_allocated;
 304 EXPORT_SYMBOL(tcp_sockets_allocated);
 305
 306 /*
 307  * TCP splice context
 308  */
 309 struct tcp_splice_state {
 310         struct pipe_inode_info *pipe;
 311         size_t len;
 312         unsigned int flags;
 313 };
 314
 315 /*
 316  * Pressure flag: try to collapse.
 317  * Technical note: it is used by multiple contexts non atomically.
 318  * All the __sk_mem_schedule() is of this nature: accounting
 319  * is strict, actions are advisory and have some latency.
 320  */
 321 unsigned long tcp_memory_pressure __read_mostly;
 322 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
 323
 324 void tcp_enter_memory_pressure(struct sock *sk)
 325 {
 326         unsigned long val;
 327
 328         if (tcp_memory_pressure)
 329                 return;
 330         val = jiffies;
 331
 332         if (!val)
 333                 val--;
 334         if (!cmpxchg(&tcp_memory_pressure, 0, val))
 335                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
 336 }
 337 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
 338
 339 void tcp_leave_memory_pressure(struct sock *sk)
 340 {
 341         unsigned long val;
 342
 343         if (!tcp_memory_pressure)
 344                 return;
 345         val = xchg(&tcp_memory_pressure, 0);
 346         if (val)
 347                 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
 348                               jiffies_to_msecs(jiffies - val));
 349 }
 350 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
 351
 352 /* Convert seconds to retransmits based on initial and max timeout */
 353 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
 354 {
 355         u8 res = 0;
 356
 357         if (seconds > 0) {
 358                 int period = timeout;
 359
 360                 res = 1;
 361                 while (seconds > period && res < 255) {
 362                         res++;
 363                         timeout <<= 1;
 364                         if (timeout > rto_max)
 365                                 timeout = rto_max;
 366                         period += timeout;
 367                 }
 368         }
 369         return res;
 370 }
 371
 372 /* Convert retransmits to seconds based on initial and max timeout */
 373 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
 374 {
 375         int period = 0;
 376
 377         if (retrans > 0) {
 378                 period = timeout;
 379                 while (--retrans) {
 380                         timeout <<= 1;
 381                         if (timeout > rto_max)
 382                                 timeout = rto_max;
 383                         period += timeout;
 384                 }
 385         }
 386         return period;
 387 }
 388
 389 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
 390 {
 391         u32 rate = READ_ONCE(tp->rate_delivered);
 392         u32 intv = READ_ONCE(tp->rate_interval_us);
 393         u64 rate64 = 0;
 394
 395         if (rate && intv) {
 396                 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
 397                 do_div(rate64, intv);
 398         }
 399         return rate64;
 400 }
 401
 402 /* Address-family independent initialization for a tcp_sock.
 403  *
 404  * NOTE: A lot of things set to zero explicitly by call to
 405  *       sk_alloc() so need not be done here.
 406  */
 407 void tcp_init_sock(struct sock *sk)
 408 {
 409         struct inet_connection_sock *icsk = inet_csk(sk);
 410         struct tcp_sock *tp = tcp_sk(sk);
 411
 412         tp->out_of_order_queue = RB_ROOT;
 413         sk->tcp_rtx_queue = RB_ROOT;
 414         tcp_init_xmit_timers(sk);
 415         INIT_LIST_HEAD(&tp->tsq_node);
 416         INIT_LIST_HEAD(&tp->tsorted_sent_queue);
 417
 418         icsk->icsk_rto = TCP_TIMEOUT_INIT;
 419         tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
 420         minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
 421
 422         /* So many TCP implementations out there (incorrectly) count the
 423          * initial SYN frame in their delayed-ACK and congestion control
 424          * algorithms that we must have the following bandaid to talk
 425          * efficiently to them.  -DaveM
 426          */
 427         tp->snd_cwnd = TCP_INIT_CWND;
 428
 429         /* There's a bubble in the pipe until at least the first ACK. */
 430         tp->app_limited = ~0U;
 431
 432         /* See draft-stevens-tcpca-spec-01 for discussion of the
 433          * initialization of these values.
 434          */
 435         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 436         tp->snd_cwnd_clamp = ~0;
 437         tp->mss_cache = TCP_MSS_DEFAULT;
 438
 439         tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
 440         tcp_assign_congestion_control(sk);
 441
 442         tp->tsoffset = 0;
 443         tp->rack.reo_wnd_steps = 1;
 444
 445         sk->sk_state = TCP_CLOSE;
 446
 447         sk->sk_write_space = sk_stream_write_space;
 448         sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
 449
 450         icsk->icsk_sync_mss = tcp_sync_mss;
 451
 452         sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
 453         sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
 454
 455         sk_sockets_allocated_inc(sk);
 456         sk->sk_route_forced_caps = NETIF_F_GSO;
 457 }
 458 EXPORT_SYMBOL(tcp_init_sock);
 459
 460 void tcp_init_transfer(struct sock *sk, int bpf_op)
 461 {
 462         struct inet_connection_sock *icsk = inet_csk(sk);
 463
 464         tcp_mtup_init(sk);
 465         icsk->icsk_af_ops->rebuild_header(sk);
 466         tcp_init_metrics(sk);
 467         tcp_call_bpf(sk, bpf_op, 0, NULL);
 468         tcp_init_congestion_control(sk);
 469         tcp_init_buffer_space(sk);
 470 }
 471
 472 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
 473 {
 474         struct sk_buff *skb = tcp_write_queue_tail(sk);
 475
 476         if (tsflags && skb) {
 477                 struct skb_shared_info *shinfo = skb_shinfo(skb);
 478                 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 479
 480                 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
 481                 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
 482                         tcb->txstamp_ack = 1;
 483                 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
 484                         shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
 485         }
 486 }
 487
 488 static inline bool tcp_stream_is_readable(const struct tcp_sock *tp,
 489                                           int target, struct sock *sk)
 490 {
 491         return (tp->rcv_nxt - tp->copied_seq >= target) ||
 492                 (sk->sk_prot->stream_memory_read ?
 493                 sk->sk_prot->stream_memory_read(sk) : false);
 494 }
 495
 496 /*
 497  *      Wait for a TCP event.
 498  *
 499  *      Note that we don't need to lock the socket, as the upper poll layers
 500  *      take care of normal races (between the test and the event) and we don't
 501  *      go look at any of the socket buffers directly.
 502  */
 503 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 504 {
 505         __poll_t mask;
 506         struct sock *sk = sock->sk;
 507         const struct tcp_sock *tp = tcp_sk(sk);
 508         int state;
 509
 510         sock_poll_wait(file, sock, wait);
 511
 512         state = inet_sk_state_load(sk);
 513         if (state == TCP_LISTEN)
 514                 return inet_csk_listen_poll(sk);
 515
 516         /* Socket is not locked. We are protected from async events
 517          * by poll logic and correct handling of state changes
 518          * made by other threads is impossible in any case.
 519          */
 520
 521         mask = 0;
 522
 523         /*
 524          * EPOLLHUP is certainly not done right. But poll() doesn't
 525          * have a notion of HUP in just one direction, and for a
 526          * socket the read side is more interesting.
 527          *
 528          * Some poll() documentation says that EPOLLHUP is incompatible
 529          * with the EPOLLOUT/POLLWR flags, so somebody should check this
 530          * all. But careful, it tends to be safer to return too many
 531          * bits than too few, and you can easily break real applications
 532          * if you don't tell them that something has hung up!
 533          *
 534          * Check-me.
 535          *
 536          * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
 537          * our fs/select.c). It means that after we received EOF,
 538          * poll always returns immediately, making impossible poll() on write()
 539          * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
 540          * if and only if shutdown has been made in both directions.
 541          * Actually, it is interesting to look how Solaris and DUX
 542          * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
 543          * then we could set it on SND_SHUTDOWN. BTW examples given
 544          * in Stevens' books assume exactly this behaviour, it explains
 545          * why EPOLLHUP is incompatible with EPOLLOUT.  --ANK
 546          *
 547          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 548          * blocking on fresh not-connected or disconnected socket. --ANK
 549          */
 550         if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
 551                 mask |= EPOLLHUP;
 552         if (sk->sk_shutdown & RCV_SHUTDOWN)
 553                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
 554
 555         /* Connected or passive Fast Open socket? */
 556         if (state != TCP_SYN_SENT &&
 557             (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
 558                 int target = sock_rcvlowat(sk, 0, INT_MAX);
 559
 560                 if (tp->urg_seq == tp->copied_seq &&
 561                     !sock_flag(sk, SOCK_URGINLINE) &&
 562                     tp->urg_data)
 563                         target++;
 564
 565                 if (tcp_stream_is_readable(tp, target, sk))
 566                         mask |= EPOLLIN | EPOLLRDNORM;
 567
 568                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 569                         if (sk_stream_is_writeable(sk)) {
 570                                 mask |= EPOLLOUT | EPOLLWRNORM;
 571                         } else {  /* send SIGIO later */
 572                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 573                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 574
 575                                 /* Race breaker. If space is freed after
 576                                  * wspace test but before the flags are set,
 577                                  * IO signal will be lost. Memory barrier
 578                                  * pairs with the input side.
 579                                  */
 580                                 smp_mb__after_atomic();
 581                                 if (sk_stream_is_writeable(sk))
 582                                         mask |= EPOLLOUT | EPOLLWRNORM;
 583                         }
 584                 } else
 585                         mask |= EPOLLOUT | EPOLLWRNORM;
 586
 587                 if (tp->urg_data & TCP_URG_VALID)
 588                         mask |= EPOLLPRI;
 589         } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
 590                 /* Active TCP fastopen socket with defer_connect
 591                  * Return EPOLLOUT so application can call write()
 592                  * in order for kernel to generate SYN+data
 593                  */
 594                 mask |= EPOLLOUT | EPOLLWRNORM;
 595         }
 596         /* This barrier is coupled with smp_wmb() in tcp_reset() */
 597         smp_rmb();
 598         if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
 599                 mask |= EPOLLERR;
 600
 601         return mask;
 602 }
 603 EXPORT_SYMBOL(tcp_poll);
 604
 605 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 606 {
 607         struct tcp_sock *tp = tcp_sk(sk);
 608         int answ;
 609         bool slow;
 610
 611         switch (cmd) {
 612         case SIOCINQ:
 613                 if (sk->sk_state == TCP_LISTEN)
 614                         return -EINVAL;
 615
 616                 slow = lock_sock_fast(sk);
 617                 answ = tcp_inq(sk);
 618                 unlock_sock_fast(sk, slow);
 619                 break;
 620         case SIOCATMARK:
 621                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 622                 break;
 623         case SIOCOUTQ:
 624                 if (sk->sk_state == TCP_LISTEN)
 625                         return -EINVAL;
 626
 627                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 628                         answ = 0;
 629                 else
 630                         answ = tp->write_seq - tp->snd_una;
 631                 break;
 632         case SIOCOUTQNSD:
 633                 if (sk->sk_state == TCP_LISTEN)
 634                         return -EINVAL;
 635
 636                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 637                         answ = 0;
 638                 else
 639                         answ = tp->write_seq - tp->snd_nxt;
 640                 break;
 641         default:
 642                 return -ENOIOCTLCMD;
 643         }
 644
 645         return put_user(answ, (int __user *)arg);
 646 }
 647 EXPORT_SYMBOL(tcp_ioctl);
 648
 649 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 650 {
 651         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 652         tp->pushed_seq = tp->write_seq;
 653 }
 654
 655 static inline bool forced_push(const struct tcp_sock *tp)
 656 {
 657         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 658 }
 659
 660 static void skb_entail(struct sock *sk, struct sk_buff *skb)
 661 {
 662         struct tcp_sock *tp = tcp_sk(sk);
 663         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 664
 665         skb->csum    = 0;
 666         tcb->seq     = tcb->end_seq = tp->write_seq;
 667         tcb->tcp_flags = TCPHDR_ACK;
 668         tcb->sacked  = 0;
 669         __skb_header_release(skb);
 670         tcp_add_write_queue_tail(sk, skb);
 671         sk->sk_wmem_queued += skb->truesize;
 672         sk_mem_charge(sk, skb->truesize);
 673         if (tp->nonagle & TCP_NAGLE_PUSH)
 674                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 675
 676         tcp_slow_start_after_idle_check(sk);
 677 }
 678
 679 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
 680 {
 681         if (flags & MSG_OOB)
 682                 tp->snd_up = tp->write_seq;
 683 }
 684
 685 /* If a not yet filled skb is pushed, do not send it if
 686  * we have data packets in Qdisc or NIC queues :
 687  * Because TX completion will happen shortly, it gives a chance
 688  * to coalesce future sendmsg() payload into this skb, without
 689  * need for a timer, and with no latency trade off.
 690  * As packets containing data payload have a bigger truesize
 691  * than pure acks (dataless) packets, the last checks prevent
 692  * autocorking if we only have an ACK in Qdisc/NIC queues,
 693  * or if TX completion was delayed after we processed ACK packet.
 694  */
 695 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
 696                                 int size_goal)
 697 {
 698         return skb->len < size_goal &&
 699                sock_net(sk)->ipv4.sysctl_tcp_autocorking &&
 700                !tcp_rtx_queue_empty(sk) &&
 701                refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
 702 }
 703
 704 static void tcp_push(struct sock *sk, int flags, int mss_now,
 705                      int nonagle, int size_goal)
 706 {
 707         struct tcp_sock *tp = tcp_sk(sk);
 708         struct sk_buff *skb;
 709
 710         skb = tcp_write_queue_tail(sk);
 711         if (!skb)
 712                 return;
 713         if (!(flags & MSG_MORE) || forced_push(tp))
 714                 tcp_mark_push(tp, skb);
 715
 716         tcp_mark_urg(tp, flags);
 717
 718         if (tcp_should_autocork(sk, skb, size_goal)) {
 719
 720                 /* avoid atomic op if TSQ_THROTTLED bit is already set */
 721                 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
 722                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
 723                         set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
 724                 }
 725                 /* It is possible TX completion already happened
 726                  * before we set TSQ_THROTTLED.
 727                  */
 728                 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
 729                         return;
 730         }
 731
 732         if (flags & MSG_MORE)
 733                 nonagle = TCP_NAGLE_CORK;
 734
 735         __tcp_push_pending_frames(sk, mss_now, nonagle);
 736 }
 737
 738 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 739                                 unsigned int offset, size_t len)
 740 {
 741         struct tcp_splice_state *tss = rd_desc->arg.data;
 742         int ret;
 743
 744         ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
 745                               min(rd_desc->count, len), tss->flags);
 746         if (ret > 0)
 747                 rd_desc->count -= ret;
 748         return ret;
 749 }
 750
 751 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 752 {
 753         /* Store TCP splice context information in read_descriptor_t. */
 754         read_descriptor_t rd_desc = {
 755                 .arg.data = tss,
 756                 .count    = tss->len,
 757         };
 758
 759         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 760 }
 761
 762 /**
 763  *  tcp_splice_read - splice data from TCP socket to a pipe
 764  * @sock:       socket to splice from
 765  * @ppos:       position (not valid)
 766  * @pipe:       pipe to splice to
 767  * @len:        number of bytes to splice
 768  * @flags:      splice modifier flags
 769  *
 770  * Description:
 771  *    Will read pages from given socket and fill them into a pipe.
 772  *
 773  **/
 774 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 775                         struct pipe_inode_info *pipe, size_t len,
 776                         unsigned int flags)
 777 {
 778         struct sock *sk = sock->sk;
 779         struct tcp_splice_state tss = {
 780                 .pipe = pipe,
 781                 .len = len,
 782                 .flags = flags,
 783         };
 784         long timeo;
 785         ssize_t spliced;
 786         int ret;
 787
 788         sock_rps_record_flow(sk);
 789         /*
 790          * We can't seek on a socket input
 791          */
 792         if (unlikely(*ppos))
 793                 return -ESPIPE;
 794
 795         ret = spliced = 0;
 796
 797         lock_sock(sk);
 798
 799         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 800         while (tss.len) {
 801                 ret = __tcp_splice_read(sk, &tss);
 802                 if (ret < 0)
 803                         break;
 804                 else if (!ret) {
 805                         if (spliced)
 806                                 break;
 807                         if (sock_flag(sk, SOCK_DONE))
 808                                 break;
 809                         if (sk->sk_err) {
 810                                 ret = sock_error(sk);
 811                                 break;
 812                         }
 813                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 814                                 break;
 815                         if (sk->sk_state == TCP_CLOSE) {
 816                                 /*
 817                                  * This occurs when user tries to read
 818                                  * from never connected socket.
 819                                  */
 820                                 ret = -ENOTCONN;
 821                                 break;
 822                         }
 823                         if (!timeo) {
 824                                 ret = -EAGAIN;
 825                                 break;
 826                         }
 827                         /* if __tcp_splice_read() got nothing while we have
 828                          * an skb in receive queue, we do not want to loop.
 829                          * This might happen with URG data.
 830                          */
 831                         if (!skb_queue_empty(&sk->sk_receive_queue))
 832                                 break;
 833                         sk_wait_data(sk, &timeo, NULL);
 834                         if (signal_pending(current)) {
 835                                 ret = sock_intr_errno(timeo);
 836                                 break;
 837                         }
 838                         continue;
 839                 }
 840                 tss.len -= ret;
 841                 spliced += ret;
 842
 843                 if (!timeo)
 844                         break;
 845                 release_sock(sk);
 846                 lock_sock(sk);
 847
 848                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 849                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 850                     signal_pending(current))
 851                         break;
 852         }
 853
 854         release_sock(sk);
 855
 856         if (spliced)
 857                 return spliced;
 858
 859         return ret;
 860 }
 861 EXPORT_SYMBOL(tcp_splice_read);
 862
 863 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
 864                                     bool force_schedule)
 865 {
 866         struct sk_buff *skb;
 867
 868         if (likely(!size)) {
 869                 skb = sk->sk_tx_skb_cache;
 870                 if (skb && !skb_cloned(skb)) {
 871                         skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
 872                         sk->sk_tx_skb_cache = NULL;
 873                         pskb_trim(skb, 0);
 874                         INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
 875                         skb_shinfo(skb)->tx_flags = 0;
 876                         memset(TCP_SKB_CB(skb), 0, sizeof(struct tcp_skb_cb));
 877                         return skb;
 878                 }
 879         }
 880         /* The TCP header must be at least 32-bit aligned.  */
 881         size = ALIGN(size, 4);
 882
 883         if (unlikely(tcp_under_memory_pressure(sk)))
 884                 sk_mem_reclaim_partial(sk);
 885
 886         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 887         if (likely(skb)) {
 888                 bool mem_scheduled;
 889
 890                 if (force_schedule) {
 891                         mem_scheduled = true;
 892                         sk_forced_mem_schedule(sk, skb->truesize);
 893                 } else {
 894                         mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
 895                 }
 896                 if (likely(mem_scheduled)) {
 897                         skb_reserve(skb, sk->sk_prot->max_header);
 898                         /*
 899                          * Make sure that we have exactly size bytes
 900                          * available to the caller, no more, no less.
 901                          */
 902                         skb->reserved_tailroom = skb->end - skb->tail - size;
 903                         INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
 904                         return skb;
 905                 }
 906                 __kfree_skb(skb);
 907         } else {
 908                 sk->sk_prot->enter_memory_pressure(sk);
 909                 sk_stream_moderate_sndbuf(sk);
 910         }
 911         return NULL;
 912 }
 913
 914 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 915                                        int large_allowed)
 916 {
 917         struct tcp_sock *tp = tcp_sk(sk);
 918         u32 new_size_goal, size_goal;
 919
 920         if (!large_allowed)
 921                 return mss_now;
 922
 923         /* Note : tcp_tso_autosize() will eventually split this later */
 924         new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
 925         new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
 926
 927         /* We try hard to avoid divides here */
 928         size_goal = tp->gso_segs * mss_now;
 929         if (unlikely(new_size_goal < size_goal ||
 930                      new_size_goal >= size_goal + mss_now)) {
 931                 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
 932                                      sk->sk_gso_max_segs);
 933                 size_goal = tp->gso_segs * mss_now;
 934         }
 935
 936         return max(size_goal, mss_now);
 937 }
 938
 939 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 940 {
 941         int mss_now;
 942
 943         mss_now = tcp_current_mss(sk);
 944         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 945
 946         return mss_now;
 947 }
 948
 949 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
 950                          size_t size, int flags)
 951 {
 952         struct tcp_sock *tp = tcp_sk(sk);
 953         int mss_now, size_goal;
 954         int err;
 955         ssize_t copied;
 956         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 957
 958         if (IS_ENABLED(CONFIG_DEBUG_VM) &&
 959             WARN_ONCE(PageSlab(page), "page must not be a Slab one"))
 960                 return -EINVAL;
 961
 962         /* Wait for a connection to finish. One exception is TCP Fast Open
 963          * (passive side) where data is allowed to be sent before a connection
 964          * is fully established.
 965          */
 966         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
 967             !tcp_passive_fastopen(sk)) {
 968                 err = sk_stream_wait_connect(sk, &timeo);
 969                 if (err != 0)
 970                         goto out_err;
 971         }
 972
 973         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 974
 975         mss_now = tcp_send_mss(sk, &size_goal, flags);
 976         copied = 0;
 977
 978         err = -EPIPE;
 979         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 980                 goto out_err;
 981
 982         while (size > 0) {
 983                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 984                 int copy, i;
 985                 bool can_coalesce;
 986
 987                 if (!skb || (copy = size_goal - skb->len) <= 0 ||
 988                     !tcp_skb_can_collapse_to(skb)) {
 989 new_segment:
 990                         if (!sk_stream_memory_free(sk))
 991                                 goto wait_for_sndbuf;
 992
 993                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
 994                                         tcp_rtx_and_write_queues_empty(sk));
 995                         if (!skb)
 996                                 goto wait_for_memory;
 997
 998                         skb_entail(sk, skb);
 999                         copy = size_goal;
1000                 }
1001
1002                 if (copy > size)
1003                         copy = size;
1004
1005                 i = skb_shinfo(skb)->nr_frags;
1006                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
1007                 if (!can_coalesce && i >= sysctl_max_skb_frags) {
1008                         tcp_mark_push(tp, skb);
1009                         goto new_segment;
1010                 }
1011                 if (!sk_wmem_schedule(sk, copy))
1012                         goto wait_for_memory;
1013
1014                 if (can_coalesce) {
1015                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1016                 } else {
1017                         get_page(page);
1018                         skb_fill_page_desc(skb, i, page, offset, copy);
1019                 }
1020
1021                 if (!(flags & MSG_NO_SHARED_FRAGS))
1022                         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1023
1024                 skb->len += copy;
1025                 skb->data_len += copy;
1026                 skb->truesize += copy;
1027                 sk->sk_wmem_queued += copy;
1028                 sk_mem_charge(sk, copy);
1029                 skb->ip_summed = CHECKSUM_PARTIAL;
1030                 tp->write_seq += copy;
1031                 TCP_SKB_CB(skb)->end_seq += copy;
1032                 tcp_skb_pcount_set(skb, 0);
1033
1034                 if (!copied)
1035                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1036
1037                 copied += copy;
1038                 offset += copy;
1039                 size -= copy;
1040                 if (!size)
1041                         goto out;
1042
1043                 if (skb->len < size_goal || (flags & MSG_OOB))
1044                         continue;
1045
1046                 if (forced_push(tp)) {
1047                         tcp_mark_push(tp, skb);
1048                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1049                 } else if (skb == tcp_send_head(sk))
1050                         tcp_push_one(sk, mss_now);
1051                 continue;
1052
1053 wait_for_sndbuf:
1054                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1055 wait_for_memory:
1056                 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1057                          TCP_NAGLE_PUSH, size_goal);
1058
1059                 err = sk_stream_wait_memory(sk, &timeo);
1060                 if (err != 0)
1061                         goto do_error;
1062
1063                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1064         }
1065
1066 out:
1067         if (copied) {
1068                 tcp_tx_timestamp(sk, sk->sk_tsflags);
1069                 if (!(flags & MSG_SENDPAGE_NOTLAST))
1070                         tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1071         }
1072         return copied;
1073
1074 do_error:
1075         if (copied)
1076                 goto out;
1077 out_err:
1078         /* make sure we wake any epoll edge trigger waiter */
1079         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1080                      err == -EAGAIN)) {
1081                 sk->sk_write_space(sk);
1082                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1083         }
1084         return sk_stream_error(sk, flags, err);
1085 }
1086 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1087
1088 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1089                         size_t size, int flags)
1090 {
1091         if (!(sk->sk_route_caps & NETIF_F_SG))
1092                 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1093
1094         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1095
1096         return do_tcp_sendpages(sk, page, offset, size, flags);
1097 }
1098 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1099
1100 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1101                  size_t size, int flags)
1102 {
1103         int ret;
1104
1105         lock_sock(sk);
1106         ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1107         release_sock(sk);
1108
1109         return ret;
1110 }
1111 EXPORT_SYMBOL(tcp_sendpage);
1112
1113 void tcp_free_fastopen_req(struct tcp_sock *tp)
1114 {
1115         if (tp->fastopen_req) {
1116                 kfree(tp->fastopen_req);
1117                 tp->fastopen_req = NULL;
1118         }
1119 }
1120
1121 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1122                                 int *copied, size_t size,
1123                                 struct ubuf_info *uarg)
1124 {
1125         struct tcp_sock *tp = tcp_sk(sk);
1126         struct inet_sock *inet = inet_sk(sk);
1127         struct sockaddr *uaddr = msg->msg_name;
1128         int err, flags;
1129
1130         if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1131             (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1132              uaddr->sa_family == AF_UNSPEC))
1133                 return -EOPNOTSUPP;
1134         if (tp->fastopen_req)
1135                 return -EALREADY; /* Another Fast Open is in progress */
1136
1137         tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1138                                    sk->sk_allocation);
1139         if (unlikely(!tp->fastopen_req))
1140                 return -ENOBUFS;
1141         tp->fastopen_req->data = msg;
1142         tp->fastopen_req->size = size;
1143         tp->fastopen_req->uarg = uarg;
1144
1145         if (inet->defer_connect) {
1146                 err = tcp_connect(sk);
1147                 /* Same failure procedure as in tcp_v4/6_connect */
1148                 if (err) {
1149                         tcp_set_state(sk, TCP_CLOSE);
1150                         inet->inet_dport = 0;
1151                         sk->sk_route_caps = 0;
1152                 }
1153         }
1154         flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1155         err = __inet_stream_connect(sk->sk_socket, uaddr,
1156                                     msg->msg_namelen, flags, 1);
1157         /* fastopen_req could already be freed in __inet_stream_connect
1158          * if the connection times out or gets rst
1159          */
1160         if (tp->fastopen_req) {
1161                 *copied = tp->fastopen_req->copied;
1162                 tcp_free_fastopen_req(tp);
1163                 inet->defer_connect = 0;
1164         }
1165         return err;
1166 }
1167
1168 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1169 {
1170         struct tcp_sock *tp = tcp_sk(sk);
1171         struct ubuf_info *uarg = NULL;
1172         struct sk_buff *skb;
1173         struct sockcm_cookie sockc;
1174         int flags, err, copied = 0;
1175         int mss_now = 0, size_goal, copied_syn = 0;
1176         bool process_backlog = false;
1177         bool zc = false;
1178         long timeo;
1179
1180         flags = msg->msg_flags;
1181
1182         if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
1183                 skb = tcp_write_queue_tail(sk);
1184                 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
1185                 if (!uarg) {
1186                         err = -ENOBUFS;
1187                         goto out_err;
1188                 }
1189
1190                 zc = sk->sk_route_caps & NETIF_F_SG;
1191                 if (!zc)
1192                         uarg->zerocopy = 0;
1193         }
1194
1195         if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1196             !tp->repair) {
1197                 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1198                 if (err == -EINPROGRESS && copied_syn > 0)
1199                         goto out;
1200                 else if (err)
1201                         goto out_err;
1202         }
1203
1204         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1205
1206         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1207
1208         /* Wait for a connection to finish. One exception is TCP Fast Open
1209          * (passive side) where data is allowed to be sent before a connection
1210          * is fully established.
1211          */
1212         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1213             !tcp_passive_fastopen(sk)) {
1214                 err = sk_stream_wait_connect(sk, &timeo);
1215                 if (err != 0)
1216                         goto do_error;
1217         }
1218
1219         if (unlikely(tp->repair)) {
1220                 if (tp->repair_queue == TCP_RECV_QUEUE) {
1221                         copied = tcp_send_rcvq(sk, msg, size);
1222                         goto out_nopush;
1223                 }
1224
1225                 err = -EINVAL;
1226                 if (tp->repair_queue == TCP_NO_QUEUE)
1227                         goto out_err;
1228
1229                 /* 'common' sending to sendq */
1230         }
1231
1232         sockcm_init(&sockc, sk);
1233         if (msg->msg_controllen) {
1234                 err = sock_cmsg_send(sk, msg, &sockc);
1235                 if (unlikely(err)) {
1236                         err = -EINVAL;
1237                         goto out_err;
1238                 }
1239         }
1240
1241         /* This should be in poll */
1242         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1243
1244         /* Ok commence sending. */
1245         copied = 0;
1246
1247 restart:
1248         mss_now = tcp_send_mss(sk, &size_goal, flags);
1249
1250         err = -EPIPE;
1251         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1252                 goto do_error;
1253
1254         while (msg_data_left(msg)) {
1255                 int copy = 0;
1256
1257                 skb = tcp_write_queue_tail(sk);
1258                 if (skb)
1259                         copy = size_goal - skb->len;
1260
1261                 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1262                         bool first_skb;
1263
1264 new_segment:
1265                         if (!sk_stream_memory_free(sk))
1266                                 goto wait_for_sndbuf;
1267
1268                         if (process_backlog && sk_flush_backlog(sk)) {
1269                                 process_backlog = false;
1270                                 goto restart;
1271                         }
1272                         first_skb = tcp_rtx_and_write_queues_empty(sk);
1273                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
1274                                                   first_skb);
1275                         if (!skb)
1276                                 goto wait_for_memory;
1277
1278                         process_backlog = true;
1279                         skb->ip_summed = CHECKSUM_PARTIAL;
1280
1281                         skb_entail(sk, skb);
1282                         copy = size_goal;
1283
1284                         /* All packets are restored as if they have
1285                          * already been sent. skb_mstamp_ns isn't set to
1286                          * avoid wrong rtt estimation.
1287                          */
1288                         if (tp->repair)
1289                                 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1290                 }
1291
1292                 /* Try to append data to the end of skb. */
1293                 if (copy > msg_data_left(msg))
1294                         copy = msg_data_left(msg);
1295
1296                 /* Where to copy to? */
1297                 if (skb_availroom(skb) > 0 && !zc) {
1298                         /* We have some space in skb head. Superb! */
1299                         copy = min_t(int, copy, skb_availroom(skb));
1300                         err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1301                         if (err)
1302                                 goto do_fault;
1303                 } else if (!zc) {
1304                         bool merge = true;
1305                         int i = skb_shinfo(skb)->nr_frags;
1306                         struct page_frag *pfrag = sk_page_frag(sk);
1307
1308                         if (!sk_page_frag_refill(sk, pfrag))
1309                                 goto wait_for_memory;
1310
1311                         if (!skb_can_coalesce(skb, i, pfrag->page,
1312                                               pfrag->offset)) {
1313                                 if (i >= sysctl_max_skb_frags) {
1314                                         tcp_mark_push(tp, skb);
1315                                         goto new_segment;
1316                                 }
1317                                 merge = false;
1318                         }
1319
1320                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1321
1322                         if (!sk_wmem_schedule(sk, copy))
1323                                 goto wait_for_memory;
1324
1325                         err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1326                                                        pfrag->page,
1327                                                        pfrag->offset,
1328                                                        copy);
1329                         if (err)
1330                                 goto do_error;
1331
1332                         /* Update the skb. */
1333                         if (merge) {
1334                                 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1335                         } else {
1336                                 skb_fill_page_desc(skb, i, pfrag->page,
1337                                                    pfrag->offset, copy);
1338                                 page_ref_inc(pfrag->page);
1339                         }
1340                         pfrag->offset += copy;
1341                 } else {
1342                         err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1343                         if (err == -EMSGSIZE || err == -EEXIST) {
1344                                 tcp_mark_push(tp, skb);
1345                                 goto new_segment;
1346                         }
1347                         if (err < 0)
1348                                 goto do_error;
1349                         copy = err;
1350                 }
1351
1352                 if (!copied)
1353                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1354
1355                 tp->write_seq += copy;
1356                 TCP_SKB_CB(skb)->end_seq += copy;
1357                 tcp_skb_pcount_set(skb, 0);
1358
1359                 copied += copy;
1360                 if (!msg_data_left(msg)) {
1361                         if (unlikely(flags & MSG_EOR))
1362                                 TCP_SKB_CB(skb)->eor = 1;
1363                         goto out;
1364                 }
1365
1366                 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1367                         continue;
1368
1369                 if (forced_push(tp)) {
1370                         tcp_mark_push(tp, skb);
1371                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1372                 } else if (skb == tcp_send_head(sk))
1373                         tcp_push_one(sk, mss_now);
1374                 continue;
1375
1376 wait_for_sndbuf:
1377                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1378 wait_for_memory:
1379                 if (copied)
1380                         tcp_push(sk, flags & ~MSG_MORE, mss_now,
1381                                  TCP_NAGLE_PUSH, size_goal);
1382
1383                 err = sk_stream_wait_memory(sk, &timeo);
1384                 if (err != 0)
1385                         goto do_error;
1386
1387                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1388         }
1389
1390 out:
1391         if (copied) {
1392                 tcp_tx_timestamp(sk, sockc.tsflags);
1393                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1394         }
1395 out_nopush:
1396         sock_zerocopy_put(uarg);
1397         return copied + copied_syn;
1398
1399 do_fault:
1400         if (!skb->len) {
1401                 tcp_unlink_write_queue(skb, sk);
1402                 /* It is the one place in all of TCP, except connection
1403                  * reset, where we can be unlinking the send_head.
1404                  */
1405                 if (tcp_write_queue_empty(sk))
1406                         tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
1407                 sk_wmem_free_skb(sk, skb);
1408         }
1409
1410 do_error:
1411         if (copied + copied_syn)
1412                 goto out;
1413 out_err:
1414         sock_zerocopy_put_abort(uarg, true);
1415         err = sk_stream_error(sk, flags, err);
1416         /* make sure we wake any epoll edge trigger waiter */
1417         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1418                      err == -EAGAIN)) {
1419                 sk->sk_write_space(sk);
1420                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1421         }
1422         return err;
1423 }
1424 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1425
1426 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1427 {
1428         int ret;
1429
1430         lock_sock(sk);
1431         ret = tcp_sendmsg_locked(sk, msg, size);
1432         release_sock(sk);
1433
1434         return ret;
1435 }
1436 EXPORT_SYMBOL(tcp_sendmsg);
1437
1438 /*
1439  *      Handle reading urgent data. BSD has very simple semantics for
1440  *      this, no blocking and very strange errors 8)
1441  */
1442
1443 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1444 {
1445         struct tcp_sock *tp = tcp_sk(sk);
1446
1447         /* No URG data to read. */
1448         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1449             tp->urg_data == TCP_URG_READ)
1450                 return -EINVAL; /* Yes this is right ! */
1451
1452         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1453                 return -ENOTCONN;
1454
1455         if (tp->urg_data & TCP_URG_VALID) {
1456                 int err = 0;
1457                 char c = tp->urg_data;
1458
1459                 if (!(flags & MSG_PEEK))
1460                         tp->urg_data = TCP_URG_READ;
1461
1462                 /* Read urgent data. */
1463                 msg->msg_flags |= MSG_OOB;
1464
1465                 if (len > 0) {
1466                         if (!(flags & MSG_TRUNC))
1467                                 err = memcpy_to_msg(msg, &c, 1);
1468                         len = 1;
1469                 } else
1470                         msg->msg_flags |= MSG_TRUNC;
1471
1472                 return err ? -EFAULT : len;
1473         }
1474
1475         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1476                 return 0;
1477
1478         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1479          * the available implementations agree in this case:
1480          * this call should never block, independent of the
1481          * blocking state of the socket.
1482          * Mike <pall@rz.uni-karlsruhe.de>
1483          */
1484         return -EAGAIN;
1485 }
1486
1487 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1488 {
1489         struct sk_buff *skb;
1490         int copied = 0, err = 0;
1491
1492         /* XXX -- need to support SO_PEEK_OFF */
1493
1494         skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1495                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1496                 if (err)
1497                         return err;
1498                 copied += skb->len;
1499         }
1500
1501         skb_queue_walk(&sk->sk_write_queue, skb) {
1502                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1503                 if (err)
1504                         break;
1505
1506                 copied += skb->len;
1507         }
1508
1509         return err ?: copied;
1510 }
1511
1512 /* Clean up the receive buffer for full frames taken by the user,
1513  * then send an ACK if necessary.  COPIED is the number of bytes
1514  * tcp_recvmsg has given to the user so far, it speeds up the
1515  * calculation of whether or not we must ACK for the sake of
1516  * a window update.
1517  */
1518 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1519 {
1520         struct tcp_sock *tp = tcp_sk(sk);
1521         bool time_to_ack = false;
1522
1523         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1524
1525         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1526              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1527              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1528
1529         if (inet_csk_ack_scheduled(sk)) {
1530                 const struct inet_connection_sock *icsk = inet_csk(sk);
1531                    /* Delayed ACKs frequently hit locked sockets during bulk
1532                     * receive. */
1533                 if (icsk->icsk_ack.blocked ||
1534                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1535                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1536                     /*
1537                      * If this read emptied read buffer, we send ACK, if
1538                      * connection is not bidirectional, user drained
1539                      * receive buffer and there was a small segment
1540                      * in queue.
1541                      */
1542                     (copied > 0 &&
1543                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1544                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1545                        !inet_csk_in_pingpong_mode(sk))) &&
1546                       !atomic_read(&sk->sk_rmem_alloc)))
1547                         time_to_ack = true;
1548         }
1549
1550         /* We send an ACK if we can now advertise a non-zero window
1551          * which has been raised "significantly".
1552          *
1553          * Even if window raised up to infinity, do not send window open ACK
1554          * in states, where we will not receive more. It is useless.
1555          */
1556         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1557                 __u32 rcv_window_now = tcp_receive_window(tp);
1558
1559                 /* Optimize, __tcp_select_window() is not cheap. */
1560                 if (2*rcv_window_now <= tp->window_clamp) {
1561                         __u32 new_window = __tcp_select_window(sk);
1562
1563                         /* Send ACK now, if this read freed lots of space
1564                          * in our buffer. Certainly, new_window is new window.
1565                          * We can advertise it now, if it is not less than current one.
1566                          * "Lots" means "at least twice" here.
1567                          */
1568                         if (new_window && new_window >= 2 * rcv_window_now)
1569                                 time_to_ack = true;
1570                 }
1571         }
1572         if (time_to_ack)
1573                 tcp_send_ack(sk);
1574 }
1575
1576 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1577 {
1578         struct sk_buff *skb;
1579         u32 offset;
1580
1581         while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1582                 offset = seq - TCP_SKB_CB(skb)->seq;
1583                 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1584                         pr_err_once("%s: found a SYN, please report !\n", __func__);
1585                         offset--;
1586                 }
1587                 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1588                         *off = offset;
1589                         return skb;
1590                 }
1591                 /* This looks weird, but this can happen if TCP collapsing
1592                  * splitted a fat GRO packet, while we released socket lock
1593                  * in skb_splice_bits()
1594                  */
1595                 sk_eat_skb(sk, skb);
1596         }
1597         return NULL;
1598 }
1599
1600 /*
1601  * This routine provides an alternative to tcp_recvmsg() for routines
1602  * that would like to handle copying from skbuffs directly in 'sendfile'
1603  * fashion.
1604  * Note:
1605  *      - It is assumed that the socket was locked by the caller.
1606  *      - The routine does not block.
1607  *      - At present, there is no support for reading OOB data
1608  *        or for 'peeking' the socket using this routine
1609  *        (although both would be easy to implement).
1610  */
1611 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1612                   sk_read_actor_t recv_actor)
1613 {
1614         struct sk_buff *skb;
1615         struct tcp_sock *tp = tcp_sk(sk);
1616         u32 seq = tp->copied_seq;
1617         u32 offset;
1618         int copied = 0;
1619
1620         if (sk->sk_state == TCP_LISTEN)
1621                 return -ENOTCONN;
1622         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1623                 if (offset < skb->len) {
1624                         int used;
1625                         size_t len;
1626
1627                         len = skb->len - offset;
1628                         /* Stop reading if we hit a patch of urgent data */
1629                         if (tp->urg_data) {
1630                                 u32 urg_offset = tp->urg_seq - seq;
1631                                 if (urg_offset < len)
1632                                         len = urg_offset;
1633                                 if (!len)
1634                                         break;
1635                         }
1636                         used = recv_actor(desc, skb, offset, len);
1637                         if (used <= 0) {
1638                                 if (!copied)
1639                                         copied = used;
1640                                 break;
1641                         } else if (used <= len) {
1642                                 seq += used;
1643                                 copied += used;
1644                                 offset += used;
1645                         }
1646                         /* If recv_actor drops the lock (e.g. TCP splice
1647                          * receive) the skb pointer might be invalid when
1648                          * getting here: tcp_collapse might have deleted it
1649                          * while aggregating skbs from the socket queue.
1650                          */
1651                         skb = tcp_recv_skb(sk, seq - 1, &offset);
1652                         if (!skb)
1653                                 break;
1654                         /* TCP coalescing might have appended data to the skb.
1655                          * Try to splice more frags
1656                          */
1657                         if (offset + 1 != skb->len)
1658                                 continue;
1659                 }
1660                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1661                         sk_eat_skb(sk, skb);
1662                         ++seq;
1663                         break;
1664                 }
1665                 sk_eat_skb(sk, skb);
1666                 if (!desc->count)
1667                         break;
1668                 tp->copied_seq = seq;
1669         }
1670         tp->copied_seq = seq;
1671
1672         tcp_rcv_space_adjust(sk);
1673
1674         /* Clean up data we have read: This will do ACK frames. */
1675         if (copied > 0) {
1676                 tcp_recv_skb(sk, seq, &offset);
1677                 tcp_cleanup_rbuf(sk, copied);
1678         }
1679         return copied;
1680 }
1681 EXPORT_SYMBOL(tcp_read_sock);
1682
1683 int tcp_peek_len(struct socket *sock)
1684 {
1685         return tcp_inq(sock->sk);
1686 }
1687 EXPORT_SYMBOL(tcp_peek_len);
1688
1689 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1690 int tcp_set_rcvlowat(struct sock *sk, int val)
1691 {
1692         int cap;
1693
1694         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1695                 cap = sk->sk_rcvbuf >> 1;
1696         else
1697                 cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1;
1698         val = min(val, cap);
1699         sk->sk_rcvlowat = val ? : 1;
1700
1701         /* Check if we need to signal EPOLLIN right now */
1702         tcp_data_ready(sk);
1703
1704         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1705                 return 0;
1706
1707         val <<= 1;
1708         if (val > sk->sk_rcvbuf) {
1709                 sk->sk_rcvbuf = val;
1710                 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1711         }
1712         return 0;
1713 }
1714 EXPORT_SYMBOL(tcp_set_rcvlowat);
1715
1716 #ifdef CONFIG_MMU
1717 static const struct vm_operations_struct tcp_vm_ops = {
1718 };
1719
1720 int tcp_mmap(struct file *file, struct socket *sock,
1721              struct vm_area_struct *vma)
1722 {
1723         if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1724                 return -EPERM;
1725         vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1726
1727         /* Instruct vm_insert_page() to not down_read(mmap_sem) */
1728         vma->vm_flags |= VM_MIXEDMAP;
1729
1730         vma->vm_ops = &tcp_vm_ops;
1731         return 0;
1732 }
1733 EXPORT_SYMBOL(tcp_mmap);
1734
1735 static int tcp_zerocopy_receive(struct sock *sk,
1736                                 struct tcp_zerocopy_receive *zc)
1737 {
1738         unsigned long address = (unsigned long)zc->address;
1739         const skb_frag_t *frags = NULL;
1740         u32 length = 0, seq, offset;
1741         struct vm_area_struct *vma;
1742         struct sk_buff *skb = NULL;
1743         struct tcp_sock *tp;
1744         int inq;
1745         int ret;
1746
1747         if (address & (PAGE_SIZE - 1) || address != zc->address)
1748                 return -EINVAL;
1749
1750         if (sk->sk_state == TCP_LISTEN)
1751                 return -ENOTCONN;
1752
1753         sock_rps_record_flow(sk);
1754
1755         down_read(&current->mm->mmap_sem);
1756
1757         ret = -EINVAL;
1758         vma = find_vma(current->mm, address);
1759         if (!vma || vma->vm_start > address || vma->vm_ops != &tcp_vm_ops)
1760                 goto out;
1761         zc->length = min_t(unsigned long, zc->length, vma->vm_end - address);
1762
1763         tp = tcp_sk(sk);
1764         seq = tp->copied_seq;
1765         inq = tcp_inq(sk);
1766         zc->length = min_t(u32, zc->length, inq);
1767         zc->length &= ~(PAGE_SIZE - 1);
1768         if (zc->length) {
1769                 zap_page_range(vma, address, zc->length);
1770                 zc->recv_skip_hint = 0;
1771         } else {
1772                 zc->recv_skip_hint = inq;
1773         }
1774         ret = 0;
1775         while (length + PAGE_SIZE <= zc->length) {
1776                 if (zc->recv_skip_hint < PAGE_SIZE) {
1777                         if (skb) {
1778                                 skb = skb->next;
1779                                 offset = seq - TCP_SKB_CB(skb)->seq;
1780                         } else {
1781                                 skb = tcp_recv_skb(sk, seq, &offset);
1782                         }
1783
1784                         zc->recv_skip_hint = skb->len - offset;
1785                         offset -= skb_headlen(skb);
1786                         if ((int)offset < 0 || skb_has_frag_list(skb))
1787                                 break;
1788                         frags = skb_shinfo(skb)->frags;
1789                         while (offset) {
1790                                 if (frags->size > offset)
1791                                         goto out;
1792                                 offset -= frags->size;
1793                                 frags++;
1794                         }
1795                 }
1796                 if (frags->size != PAGE_SIZE || frags->page_offset) {
1797                         int remaining = zc->recv_skip_hint;
1798
1799                         while (remaining && (frags->size != PAGE_SIZE ||
1800                                              frags->page_offset)) {
1801                                 remaining -= frags->size;
1802                                 frags++;
1803                         }
1804                         zc->recv_skip_hint -= remaining;
1805                         break;
1806                 }
1807                 ret = vm_insert_page(vma, address + length,
1808                                      skb_frag_page(frags));
1809                 if (ret)
1810                         break;
1811                 length += PAGE_SIZE;
1812                 seq += PAGE_SIZE;
1813                 zc->recv_skip_hint -= PAGE_SIZE;
1814                 frags++;
1815         }
1816 out:
1817         up_read(&current->mm->mmap_sem);
1818         if (length) {
1819                 tp->copied_seq = seq;
1820                 tcp_rcv_space_adjust(sk);
1821
1822                 /* Clean up data we have read: This will do ACK frames. */
1823                 tcp_recv_skb(sk, seq, &offset);
1824                 tcp_cleanup_rbuf(sk, length);
1825                 ret = 0;
1826                 if (length == zc->length)
1827                         zc->recv_skip_hint = 0;
1828         } else {
1829                 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
1830                         ret = -EIO;
1831         }
1832         zc->length = length;
1833         return ret;
1834 }
1835 #endif
1836
1837 static void tcp_update_recv_tstamps(struct sk_buff *skb,
1838                                     struct scm_timestamping_internal *tss)
1839 {
1840         if (skb->tstamp)
1841                 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1842         else
1843                 tss->ts[0] = (struct timespec64) {0};
1844
1845         if (skb_hwtstamps(skb)->hwtstamp)
1846                 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1847         else
1848                 tss->ts[2] = (struct timespec64) {0};
1849 }
1850
1851 /* Similar to __sock_recv_timestamp, but does not require an skb */
1852 static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
1853                                struct scm_timestamping_internal *tss)
1854 {
1855         int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
1856         bool has_timestamping = false;
1857
1858         if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
1859                 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
1860                         if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
1861                                 if (new_tstamp) {
1862                                         struct __kernel_timespec kts = {tss->ts[0].tv_sec, tss->ts[0].tv_nsec};
1863
1864                                         put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
1865                                                  sizeof(kts), &kts);
1866                                 } else {
1867                                         struct timespec ts_old = timespec64_to_timespec(tss->ts[0]);
1868
1869                                         put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
1870                                                  sizeof(ts_old), &ts_old);
1871                                 }
1872                         } else {
1873                                 if (new_tstamp) {
1874                                         struct __kernel_sock_timeval stv;
1875
1876                                         stv.tv_sec = tss->ts[0].tv_sec;
1877                                         stv.tv_usec = tss->ts[0].tv_nsec / 1000;
1878                                         put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
1879                                                  sizeof(stv), &stv);
1880                                 } else {
1881                                         struct __kernel_old_timeval tv;
1882
1883                                         tv.tv_sec = tss->ts[0].tv_sec;
1884                                         tv.tv_usec = tss->ts[0].tv_nsec / 1000;
1885                                         put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
1886                                                  sizeof(tv), &tv);
1887                                 }
1888                         }
1889                 }
1890
1891                 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
1892                         has_timestamping = true;
1893                 else
1894                         tss->ts[0] = (struct timespec64) {0};
1895         }
1896
1897         if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
1898                 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
1899                         has_timestamping = true;
1900                 else
1901                         tss->ts[2] = (struct timespec64) {0};
1902         }
1903
1904         if (has_timestamping) {
1905                 tss->ts[1] = (struct timespec64) {0};
1906                 if (sock_flag(sk, SOCK_TSTAMP_NEW))
1907                         put_cmsg_scm_timestamping64(msg, tss);
1908                 else
1909                         put_cmsg_scm_timestamping(msg, tss);
1910         }
1911 }
1912
1913 static int tcp_inq_hint(struct sock *sk)
1914 {
1915         const struct tcp_sock *tp = tcp_sk(sk);
1916         u32 copied_seq = READ_ONCE(tp->copied_seq);
1917         u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
1918         int inq;
1919
1920         inq = rcv_nxt - copied_seq;
1921         if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
1922                 lock_sock(sk);
1923                 inq = tp->rcv_nxt - tp->copied_seq;
1924                 release_sock(sk);
1925         }
1926         /* After receiving a FIN, tell the user-space to continue reading
1927          * by returning a non-zero inq.
1928          */
1929         if (inq == 0 && sock_flag(sk, SOCK_DONE))
1930                 inq = 1;
1931         return inq;
1932 }
1933
1934 /*
1935  *      This routine copies from a sock struct into the user buffer.
1936  *
1937  *      Technical note: in 2.3 we work on _locked_ socket, so that
1938  *      tricks with *seq access order and skb->users are not required.
1939  *      Probably, code can be easily improved even more.
1940  */
1941
1942 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1943                 int flags, int *addr_len)
1944 {
1945         struct tcp_sock *tp = tcp_sk(sk);
1946         int copied = 0;
1947         u32 peek_seq;
1948         u32 *seq;
1949         unsigned long used;
1950         int err, inq;
1951         int target;             /* Read at least this many bytes */
1952         long timeo;
1953         struct sk_buff *skb, *last;
1954         u32 urg_hole = 0;
1955         struct scm_timestamping_internal tss;
1956         bool has_tss = false;
1957         bool has_cmsg;
1958
1959         if (unlikely(flags & MSG_ERRQUEUE))
1960                 return inet_recv_error(sk, msg, len, addr_len);
1961
1962         if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1963             (sk->sk_state == TCP_ESTABLISHED))
1964                 sk_busy_loop(sk, nonblock);
1965
1966         lock_sock(sk);
1967
1968         err = -ENOTCONN;
1969         if (sk->sk_state == TCP_LISTEN)
1970                 goto out;
1971
1972         has_cmsg = tp->recvmsg_inq;
1973         timeo = sock_rcvtimeo(sk, nonblock);
1974
1975         /* Urgent data needs to be handled specially. */
1976         if (flags & MSG_OOB)
1977                 goto recv_urg;
1978
1979         if (unlikely(tp->repair)) {
1980                 err = -EPERM;
1981                 if (!(flags & MSG_PEEK))
1982                         goto out;
1983
1984                 if (tp->repair_queue == TCP_SEND_QUEUE)
1985                         goto recv_sndq;
1986
1987                 err = -EINVAL;
1988                 if (tp->repair_queue == TCP_NO_QUEUE)
1989                         goto out;
1990
1991                 /* 'common' recv queue MSG_PEEK-ing */
1992         }
1993
1994         seq = &tp->copied_seq;
1995         if (flags & MSG_PEEK) {
1996                 peek_seq = tp->copied_seq;
1997                 seq = &peek_seq;
1998         }
1999
2000         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2001
2002         do {
2003                 u32 offset;
2004
2005                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2006                 if (tp->urg_data && tp->urg_seq == *seq) {
2007                         if (copied)
2008                                 break;
2009                         if (signal_pending(current)) {
2010                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2011                                 break;
2012                         }
2013                 }
2014
2015                 /* Next get a buffer. */
2016
2017                 last = skb_peek_tail(&sk->sk_receive_queue);
2018                 skb_queue_walk(&sk->sk_receive_queue, skb) {
2019                         last = skb;
2020                         /* Now that we have two receive queues this
2021                          * shouldn't happen.
2022                          */
2023                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2024                                  "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2025                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2026                                  flags))
2027                                 break;
2028
2029                         offset = *seq - TCP_SKB_CB(skb)->seq;
2030                         if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2031                                 pr_err_once("%s: found a SYN, please report !\n", __func__);
2032                                 offset--;
2033                         }
2034                         if (offset < skb->len)
2035                                 goto found_ok_skb;
2036                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2037                                 goto found_fin_ok;
2038                         WARN(!(flags & MSG_PEEK),
2039                              "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2040                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2041                 }
2042
2043                 /* Well, if we have backlog, try to process it now yet. */
2044
2045                 if (copied >= target && !sk->sk_backlog.tail)
2046                         break;
2047
2048                 if (copied) {
2049                         if (sk->sk_err ||
2050                             sk->sk_state == TCP_CLOSE ||
2051                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
2052                             !timeo ||
2053                             signal_pending(current))
2054                                 break;
2055                 } else {
2056                         if (sock_flag(sk, SOCK_DONE))
2057                                 break;
2058
2059                         if (sk->sk_err) {
2060                                 copied = sock_error(sk);
2061                                 break;
2062                         }
2063
2064                         if (sk->sk_shutdown & RCV_SHUTDOWN)
2065                                 break;
2066
2067                         if (sk->sk_state == TCP_CLOSE) {
2068                                 /* This occurs when user tries to read
2069                                  * from never connected socket.
2070                                  */
2071                                 copied = -ENOTCONN;
2072                                 break;
2073                         }
2074
2075                         if (!timeo) {
2076                                 copied = -EAGAIN;
2077                                 break;
2078                         }
2079
2080                         if (signal_pending(current)) {
2081                                 copied = sock_intr_errno(timeo);
2082                                 break;
2083                         }
2084                 }
2085
2086                 tcp_cleanup_rbuf(sk, copied);
2087
2088                 if (copied >= target) {
2089                         /* Do not sleep, just process backlog. */
2090                         release_sock(sk);
2091                         lock_sock(sk);
2092                 } else {
2093                         sk_wait_data(sk, &timeo, last);
2094                 }
2095
2096                 if ((flags & MSG_PEEK) &&
2097                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
2098                         net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2099                                             current->comm,
2100                                             task_pid_nr(current));
2101                         peek_seq = tp->copied_seq;
2102                 }
2103                 continue;
2104
2105 found_ok_skb:
2106                 /* Ok so how much can we use? */
2107                 used = skb->len - offset;
2108                 if (len < used)
2109                         used = len;
2110
2111                 /* Do we have urgent data here? */
2112                 if (tp->urg_data) {
2113                         u32 urg_offset = tp->urg_seq - *seq;
2114                         if (urg_offset < used) {
2115                                 if (!urg_offset) {
2116                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
2117                                                 ++*seq;
2118                                                 urg_hole++;
2119                                                 offset++;
2120                                                 used--;
2121                                                 if (!used)
2122                                                         goto skip_copy;
2123                                         }
2124                                 } else
2125                                         used = urg_offset;
2126                         }
2127                 }
2128
2129                 if (!(flags & MSG_TRUNC)) {
2130                         err = skb_copy_datagram_msg(skb, offset, msg, used);
2131                         if (err) {
2132                                 /* Exception. Bailout! */
2133                                 if (!copied)
2134                                         copied = -EFAULT;
2135                                 break;
2136                         }
2137                 }
2138
2139                 *seq += used;
2140                 copied += used;
2141                 len -= used;
2142
2143                 tcp_rcv_space_adjust(sk);
2144
2145 skip_copy:
2146                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
2147                         tp->urg_data = 0;
2148                         tcp_fast_path_check(sk);
2149                 }
2150                 if (used + offset < skb->len)
2151                         continue;
2152
2153                 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2154                         tcp_update_recv_tstamps(skb, &tss);
2155                         has_tss = true;
2156                         has_cmsg = true;
2157                 }
2158                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2159                         goto found_fin_ok;
2160                 if (!(flags & MSG_PEEK))
2161                         sk_eat_skb(sk, skb);
2162                 continue;
2163
2164 found_fin_ok:
2165                 /* Process the FIN. */
2166                 ++*seq;
2167                 if (!(flags & MSG_PEEK))
2168                         sk_eat_skb(sk, skb);
2169                 break;
2170         } while (len > 0);
2171
2172         /* According to UNIX98, msg_name/msg_namelen are ignored
2173          * on connected socket. I was just happy when found this 8) --ANK
2174          */
2175
2176         /* Clean up data we have read: This will do ACK frames. */
2177         tcp_cleanup_rbuf(sk, copied);
2178
2179         release_sock(sk);
2180
2181         if (has_cmsg) {
2182                 if (has_tss)
2183                         tcp_recv_timestamp(msg, sk, &tss);
2184                 if (tp->recvmsg_inq) {
2185                         inq = tcp_inq_hint(sk);
2186                         put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2187                 }
2188         }
2189
2190         return copied;
2191
2192 out:
2193         release_sock(sk);
2194         return err;
2195
2196 recv_urg:
2197         err = tcp_recv_urg(sk, msg, len, flags);
2198         goto out;
2199
2200 recv_sndq:
2201         err = tcp_peek_sndq(sk, msg, len);
2202         goto out;
2203 }
2204 EXPORT_SYMBOL(tcp_recvmsg);
2205
2206 void tcp_set_state(struct sock *sk, int state)
2207 {
2208         int oldstate = sk->sk_state;
2209
2210         /* We defined a new enum for TCP states that are exported in BPF
2211          * so as not force the internal TCP states to be frozen. The
2212          * following checks will detect if an internal state value ever
2213          * differs from the BPF value. If this ever happens, then we will
2214          * need to remap the internal value to the BPF value before calling
2215          * tcp_call_bpf_2arg.
2216          */
2217         BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2218         BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2219         BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2220         BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2221         BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2222         BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2223         BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2224         BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2225         BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2226         BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2227         BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2228         BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2229         BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2230
2231         if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2232                 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2233
2234         switch (state) {
2235         case TCP_ESTABLISHED:
2236                 if (oldstate != TCP_ESTABLISHED)
2237                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2238                 break;
2239
2240         case TCP_CLOSE:
2241                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2242                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2243
2244                 sk->sk_prot->unhash(sk);
2245                 if (inet_csk(sk)->icsk_bind_hash &&
2246                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2247                         inet_put_port(sk);
2248                 /* fall through */
2249         default:
2250                 if (oldstate == TCP_ESTABLISHED)
2251                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2252         }
2253
2254         /* Change state AFTER socket is unhashed to avoid closed
2255          * socket sitting in hash tables.
2256          */
2257         inet_sk_state_store(sk, state);
2258 }
2259 EXPORT_SYMBOL_GPL(tcp_set_state);
2260
2261 /*
2262  *      State processing on a close. This implements the state shift for
2263  *      sending our FIN frame. Note that we only send a FIN for some
2264  *      states. A shutdown() may have already sent the FIN, or we may be
2265  *      closed.
2266  */
2267
2268 static const unsigned char new_state[16] = {
2269   /* current state:        new state:      action:      */
2270   [0 /* (Invalid) */]   = TCP_CLOSE,
2271   [TCP_ESTABLISHED]     = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2272   [TCP_SYN_SENT]        = TCP_CLOSE,
2273   [TCP_SYN_RECV]        = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2274   [TCP_FIN_WAIT1]       = TCP_FIN_WAIT1,
2275   [TCP_FIN_WAIT2]       = TCP_FIN_WAIT2,
2276   [TCP_TIME_WAIT]       = TCP_CLOSE,
2277   [TCP_CLOSE]           = TCP_CLOSE,
2278   [TCP_CLOSE_WAIT]      = TCP_LAST_ACK  | TCP_ACTION_FIN,
2279   [TCP_LAST_ACK]        = TCP_LAST_ACK,
2280   [TCP_LISTEN]          = TCP_CLOSE,
2281   [TCP_CLOSING]         = TCP_CLOSING,
2282   [TCP_NEW_SYN_RECV]    = TCP_CLOSE,    /* should not happen ! */
2283 };
2284
2285 static int tcp_close_state(struct sock *sk)
2286 {
2287         int next = (int)new_state[sk->sk_state];
2288         int ns = next & TCP_STATE_MASK;
2289
2290         tcp_set_state(sk, ns);
2291
2292         return next & TCP_ACTION_FIN;
2293 }
2294
2295 /*
2296  *      Shutdown the sending side of a connection. Much like close except
2297  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2298  */
2299
2300 void tcp_shutdown(struct sock *sk, int how)
2301 {
2302         /*      We need to grab some memory, and put together a FIN,
2303          *      and then put it into the queue to be sent.
2304          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2305          */
2306         if (!(how & SEND_SHUTDOWN))
2307                 return;
2308
2309         /* If we've already sent a FIN, or it's a closed state, skip this. */
2310         if ((1 << sk->sk_state) &
2311             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2312              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2313                 /* Clear out any half completed packets.  FIN if needed. */
2314                 if (tcp_close_state(sk))
2315                         tcp_send_fin(sk);
2316         }
2317 }
2318 EXPORT_SYMBOL(tcp_shutdown);
2319
2320 bool tcp_check_oom(struct sock *sk, int shift)
2321 {
2322         bool too_many_orphans, out_of_socket_memory;
2323
2324         too_many_orphans = tcp_too_many_orphans(sk, shift);
2325         out_of_socket_memory = tcp_out_of_memory(sk);
2326
2327         if (too_many_orphans)
2328                 net_info_ratelimited("too many orphaned sockets\n");
2329         if (out_of_socket_memory)
2330                 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2331         return too_many_orphans || out_of_socket_memory;
2332 }
2333
2334 void tcp_close(struct sock *sk, long timeout)
2335 {
2336         struct sk_buff *skb;
2337         int data_was_unread = 0;
2338         int state;
2339
2340         lock_sock(sk);
2341         sk->sk_shutdown = SHUTDOWN_MASK;
2342
2343         if (sk->sk_state == TCP_LISTEN) {
2344                 tcp_set_state(sk, TCP_CLOSE);
2345
2346                 /* Special case. */
2347                 inet_csk_listen_stop(sk);
2348
2349                 goto adjudge_to_death;
2350         }
2351
2352         /*  We need to flush the recv. buffs.  We do this only on the
2353          *  descriptor close, not protocol-sourced closes, because the
2354          *  reader process may not have drained the data yet!
2355          */
2356         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2357                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2358
2359                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2360                         len--;
2361                 data_was_unread += len;
2362                 __kfree_skb(skb);
2363         }
2364
2365         sk_mem_reclaim(sk);
2366
2367         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2368         if (sk->sk_state == TCP_CLOSE)
2369                 goto adjudge_to_death;
2370
2371         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2372          * data was lost. To witness the awful effects of the old behavior of
2373          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2374          * GET in an FTP client, suspend the process, wait for the client to
2375          * advertise a zero window, then kill -9 the FTP client, wheee...
2376          * Note: timeout is always zero in such a case.
2377          */
2378         if (unlikely(tcp_sk(sk)->repair)) {
2379                 sk->sk_prot->disconnect(sk, 0);
2380         } else if (data_was_unread) {
2381                 /* Unread data was tossed, zap the connection. */
2382                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2383                 tcp_set_state(sk, TCP_CLOSE);
2384                 tcp_send_active_reset(sk, sk->sk_allocation);
2385         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2386                 /* Check zero linger _after_ checking for unread data. */
2387                 sk->sk_prot->disconnect(sk, 0);
2388                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2389         } else if (tcp_close_state(sk)) {
2390                 /* We FIN if the application ate all the data before
2391                  * zapping the connection.
2392                  */
2393
2394                 /* RED-PEN. Formally speaking, we have broken TCP state
2395                  * machine. State transitions:
2396                  *
2397                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2398                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2399                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2400                  *
2401                  * are legal only when FIN has been sent (i.e. in window),
2402                  * rather than queued out of window. Purists blame.
2403                  *
2404                  * F.e. "RFC state" is ESTABLISHED,
2405                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2406                  *
2407                  * The visible declinations are that sometimes
2408                  * we enter time-wait state, when it is not required really
2409                  * (harmless), do not send active resets, when they are
2410                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2411                  * they look as CLOSING or LAST_ACK for Linux)
2412                  * Probably, I missed some more holelets.
2413                  *                                              --ANK
2414                  * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2415                  * in a single packet! (May consider it later but will
2416                  * probably need API support or TCP_CORK SYN-ACK until
2417                  * data is written and socket is closed.)
2418                  */
2419                 tcp_send_fin(sk);
2420         }
2421
2422         sk_stream_wait_close(sk, timeout);
2423
2424 adjudge_to_death:
2425         state = sk->sk_state;
2426         sock_hold(sk);
2427         sock_orphan(sk);
2428
2429         local_bh_disable();
2430         bh_lock_sock(sk);
2431         /* remove backlog if any, without releasing ownership. */
2432         __release_sock(sk);
2433
2434         percpu_counter_inc(sk->sk_prot->orphan_count);
2435
2436         /* Have we already been destroyed by a softirq or backlog? */
2437         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2438                 goto out;
2439
2440         /*      This is a (useful) BSD violating of the RFC. There is a
2441          *      problem with TCP as specified in that the other end could
2442          *      keep a socket open forever with no application left this end.
2443          *      We use a 1 minute timeout (about the same as BSD) then kill
2444          *      our end. If they send after that then tough - BUT: long enough
2445          *      that we won't make the old 4*rto = almost no time - whoops
2446          *      reset mistake.
2447          *
2448          *      Nope, it was not mistake. It is really desired behaviour
2449          *      f.e. on http servers, when such sockets are useless, but
2450          *      consume significant resources. Let's do it with special
2451          *      linger2 option.                                 --ANK
2452          */
2453
2454         if (sk->sk_state == TCP_FIN_WAIT2) {
2455                 struct tcp_sock *tp = tcp_sk(sk);
2456                 if (tp->linger2 < 0) {
2457                         tcp_set_state(sk, TCP_CLOSE);
2458                         tcp_send_active_reset(sk, GFP_ATOMIC);
2459                         __NET_INC_STATS(sock_net(sk),
2460                                         LINUX_MIB_TCPABORTONLINGER);
2461                 } else {
2462                         const int tmo = tcp_fin_time(sk);
2463
2464                         if (tmo > TCP_TIMEWAIT_LEN) {
2465                                 inet_csk_reset_keepalive_timer(sk,
2466                                                 tmo - TCP_TIMEWAIT_LEN);
2467                         } else {
2468                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2469                                 goto out;
2470                         }
2471                 }
2472         }
2473         if (sk->sk_state != TCP_CLOSE) {
2474                 sk_mem_reclaim(sk);
2475                 if (tcp_check_oom(sk, 0)) {
2476                         tcp_set_state(sk, TCP_CLOSE);
2477                         tcp_send_active_reset(sk, GFP_ATOMIC);
2478                         __NET_INC_STATS(sock_net(sk),
2479                                         LINUX_MIB_TCPABORTONMEMORY);
2480                 } else if (!check_net(sock_net(sk))) {
2481                         /* Not possible to send reset; just close */
2482                         tcp_set_state(sk, TCP_CLOSE);
2483                 }
2484         }
2485
2486         if (sk->sk_state == TCP_CLOSE) {
2487                 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2488                 /* We could get here with a non-NULL req if the socket is
2489                  * aborted (e.g., closed with unread data) before 3WHS
2490                  * finishes.
2491                  */
2492                 if (req)
2493                         reqsk_fastopen_remove(sk, req, false);
2494                 inet_csk_destroy_sock(sk);
2495         }
2496         /* Otherwise, socket is reprieved until protocol close. */
2497
2498 out:
2499         bh_unlock_sock(sk);
2500         local_bh_enable();
2501         release_sock(sk);
2502         sock_put(sk);
2503 }
2504 EXPORT_SYMBOL(tcp_close);
2505
2506 /* These states need RST on ABORT according to RFC793 */
2507
2508 static inline bool tcp_need_reset(int state)
2509 {
2510         return (1 << state) &
2511                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2512                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2513 }
2514
2515 static void tcp_rtx_queue_purge(struct sock *sk)
2516 {
2517         struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2518
2519         while (p) {
2520                 struct sk_buff *skb = rb_to_skb(p);
2521
2522                 p = rb_next(p);
2523                 /* Since we are deleting whole queue, no need to
2524                  * list_del(&skb->tcp_tsorted_anchor)
2525                  */
2526                 tcp_rtx_queue_unlink(skb, sk);
2527                 sk_wmem_free_skb(sk, skb);
2528         }
2529 }
2530
2531 void tcp_write_queue_purge(struct sock *sk)
2532 {
2533         struct sk_buff *skb;
2534
2535         tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2536         while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2537                 tcp_skb_tsorted_anchor_cleanup(skb);
2538                 sk_wmem_free_skb(sk, skb);
2539         }
2540         tcp_rtx_queue_purge(sk);
2541         skb = sk->sk_tx_skb_cache;
2542         if (skb) {
2543                 __kfree_skb(skb);
2544                 sk->sk_tx_skb_cache = NULL;
2545         }
2546         INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2547         sk_mem_reclaim(sk);
2548         tcp_clear_all_retrans_hints(tcp_sk(sk));
2549         tcp_sk(sk)->packets_out = 0;
2550         inet_csk(sk)->icsk_backoff = 0;
2551 }
2552
2553 int tcp_disconnect(struct sock *sk, int flags)
2554 {
2555         struct inet_sock *inet = inet_sk(sk);
2556         struct inet_connection_sock *icsk = inet_csk(sk);
2557         struct tcp_sock *tp = tcp_sk(sk);
2558         int old_state = sk->sk_state;
2559
2560         if (old_state != TCP_CLOSE)
2561                 tcp_set_state(sk, TCP_CLOSE);
2562
2563         /* ABORT function of RFC793 */
2564         if (old_state == TCP_LISTEN) {
2565                 inet_csk_listen_stop(sk);
2566         } else if (unlikely(tp->repair)) {
2567                 sk->sk_err = ECONNABORTED;
2568         } else if (tcp_need_reset(old_state) ||
2569                    (tp->snd_nxt != tp->write_seq &&
2570                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2571                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2572                  * states
2573                  */
2574                 tcp_send_active_reset(sk, gfp_any());
2575                 sk->sk_err = ECONNRESET;
2576         } else if (old_state == TCP_SYN_SENT)
2577                 sk->sk_err = ECONNRESET;
2578
2579         tcp_clear_xmit_timers(sk);
2580         __skb_queue_purge(&sk->sk_receive_queue);
2581         if (sk->sk_rx_skb_cache) {
2582                 __kfree_skb(sk->sk_rx_skb_cache);
2583                 sk->sk_rx_skb_cache = NULL;
2584         }
2585         tp->copied_seq = tp->rcv_nxt;
2586         tp->urg_data = 0;
2587         tcp_write_queue_purge(sk);
2588         tcp_fastopen_active_disable_ofo_check(sk);
2589         skb_rbtree_purge(&tp->out_of_order_queue);
2590
2591         inet->inet_dport = 0;
2592
2593         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2594                 inet_reset_saddr(sk);
2595
2596         sk->sk_shutdown = 0;
2597         sock_reset_flag(sk, SOCK_DONE);
2598         tp->srtt_us = 0;
2599         tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
2600         tp->rcv_rtt_last_tsecr = 0;
2601         tp->write_seq += tp->max_window + 2;
2602         if (tp->write_seq == 0)
2603                 tp->write_seq = 1;
2604         icsk->icsk_backoff = 0;
2605         tp->snd_cwnd = 2;
2606         icsk->icsk_probes_out = 0;
2607         icsk->icsk_rto = TCP_TIMEOUT_INIT;
2608         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2609         tp->snd_cwnd = TCP_INIT_CWND;
2610         tp->snd_cwnd_cnt = 0;
2611         tp->window_clamp = 0;
2612         tp->delivered_ce = 0;
2613         tcp_set_ca_state(sk, TCP_CA_Open);
2614         tp->is_sack_reneg = 0;
2615         tcp_clear_retrans(tp);
2616         inet_csk_delack_init(sk);
2617         /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2618          * issue in __tcp_select_window()
2619          */
2620         icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2621         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2622         __sk_dst_reset(sk);
2623         dst_release(sk->sk_rx_dst);
2624         sk->sk_rx_dst = NULL;
2625         tcp_saved_syn_free(tp);
2626         tp->compressed_ack = 0;
2627         tp->bytes_sent = 0;
2628         tp->bytes_retrans = 0;
2629         tp->duplicate_sack[0].start_seq = 0;
2630         tp->duplicate_sack[0].end_seq = 0;
2631         tp->dsack_dups = 0;
2632         tp->reord_seen = 0;
2633         tp->retrans_out = 0;
2634         tp->sacked_out = 0;
2635         tp->tlp_high_seq = 0;
2636         tp->last_oow_ack_time = 0;
2637         /* There's a bubble in the pipe until at least the first ACK. */
2638         tp->app_limited = ~0U;
2639         tp->rack.mstamp = 0;
2640         tp->rack.advanced = 0;
2641         tp->rack.reo_wnd_steps = 1;
2642         tp->rack.last_delivered = 0;
2643         tp->rack.reo_wnd_persist = 0;
2644         tp->rack.dsack_seen = 0;
2645         tp->syn_data_acked = 0;
2646         tp->rx_opt.saw_tstamp = 0;
2647         tp->rx_opt.dsack = 0;
2648         tp->rx_opt.num_sacks = 0;
2649
2650
2651         /* Clean up fastopen related fields */
2652         tcp_free_fastopen_req(tp);
2653         inet->defer_connect = 0;
2654
2655         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2656
2657         if (sk->sk_frag.page) {
2658                 put_page(sk->sk_frag.page);
2659                 sk->sk_frag.page = NULL;
2660                 sk->sk_frag.offset = 0;
2661         }
2662
2663         sk->sk_error_report(sk);
2664         return 0;
2665 }
2666 EXPORT_SYMBOL(tcp_disconnect);
2667
2668 static inline bool tcp_can_repair_sock(const struct sock *sk)
2669 {
2670         return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2671                 (sk->sk_state != TCP_LISTEN);
2672 }
2673
2674 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2675 {
2676         struct tcp_repair_window opt;
2677
2678         if (!tp->repair)
2679                 return -EPERM;
2680
2681         if (len != sizeof(opt))
2682                 return -EINVAL;
2683
2684         if (copy_from_user(&opt, optbuf, sizeof(opt)))
2685                 return -EFAULT;
2686
2687         if (opt.max_window < opt.snd_wnd)
2688                 return -EINVAL;
2689
2690         if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2691                 return -EINVAL;
2692
2693         if (after(opt.rcv_wup, tp->rcv_nxt))
2694                 return -EINVAL;
2695
2696         tp->snd_wl1     = opt.snd_wl1;
2697         tp->snd_wnd     = opt.snd_wnd;
2698         tp->max_window  = opt.max_window;
2699
2700         tp->rcv_wnd     = opt.rcv_wnd;
2701         tp->rcv_wup     = opt.rcv_wup;
2702
2703         return 0;
2704 }
2705
2706 static int tcp_repair_options_est(struct sock *sk,
2707                 struct tcp_repair_opt __user *optbuf, unsigned int len)
2708 {
2709         struct tcp_sock *tp = tcp_sk(sk);
2710         struct tcp_repair_opt opt;
2711
2712         while (len >= sizeof(opt)) {
2713                 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2714                         return -EFAULT;
2715
2716                 optbuf++;
2717                 len -= sizeof(opt);
2718
2719                 switch (opt.opt_code) {
2720                 case TCPOPT_MSS:
2721                         tp->rx_opt.mss_clamp = opt.opt_val;
2722                         tcp_mtup_init(sk);
2723                         break;
2724                 case TCPOPT_WINDOW:
2725                         {
2726                                 u16 snd_wscale = opt.opt_val & 0xFFFF;
2727                                 u16 rcv_wscale = opt.opt_val >> 16;
2728
2729                                 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
2730                                         return -EFBIG;
2731
2732                                 tp->rx_opt.snd_wscale = snd_wscale;
2733                                 tp->rx_opt.rcv_wscale = rcv_wscale;
2734                                 tp->rx_opt.wscale_ok = 1;
2735                         }
2736                         break;
2737                 case TCPOPT_SACK_PERM:
2738                         if (opt.opt_val != 0)
2739                                 return -EINVAL;
2740
2741                         tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2742                         break;
2743                 case TCPOPT_TIMESTAMP:
2744                         if (opt.opt_val != 0)
2745                                 return -EINVAL;
2746
2747                         tp->rx_opt.tstamp_ok = 1;
2748                         break;
2749                 }
2750         }
2751
2752         return 0;
2753 }
2754
2755 /*
2756  *      Socket option code for TCP.
2757  */
2758 static int do_tcp_setsockopt(struct sock *sk, int level,
2759                 int optname, char __user *optval, unsigned int optlen)
2760 {
2761         struct tcp_sock *tp = tcp_sk(sk);
2762         struct inet_connection_sock *icsk = inet_csk(sk);
2763         struct net *net = sock_net(sk);
2764         int val;
2765         int err = 0;
2766
2767         /* These are data/string values, all the others are ints */
2768         switch (optname) {
2769         case TCP_CONGESTION: {
2770                 char name[TCP_CA_NAME_MAX];
2771
2772                 if (optlen < 1)
2773                         return -EINVAL;
2774
2775                 val = strncpy_from_user(name, optval,
2776                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2777                 if (val < 0)
2778                         return -EFAULT;
2779                 name[val] = 0;
2780
2781                 lock_sock(sk);
2782                 err = tcp_set_congestion_control(sk, name, true, true);
2783                 release_sock(sk);
2784                 return err;
2785         }
2786         case TCP_ULP: {
2787                 char name[TCP_ULP_NAME_MAX];
2788
2789                 if (optlen < 1)
2790                         return -EINVAL;
2791
2792                 val = strncpy_from_user(name, optval,
2793                                         min_t(long, TCP_ULP_NAME_MAX - 1,
2794                                               optlen));
2795                 if (val < 0)
2796                         return -EFAULT;
2797                 name[val] = 0;
2798
2799                 lock_sock(sk);
2800                 err = tcp_set_ulp(sk, name);
2801                 release_sock(sk);
2802                 return err;
2803         }
2804         case TCP_FASTOPEN_KEY: {
2805                 __u8 key[TCP_FASTOPEN_KEY_LENGTH];
2806
2807                 if (optlen != sizeof(key))
2808                         return -EINVAL;
2809
2810                 if (copy_from_user(key, optval, optlen))
2811                         return -EFAULT;
2812
2813                 return tcp_fastopen_reset_cipher(net, sk, key, sizeof(key));
2814         }
2815         default:
2816                 /* fallthru */
2817                 break;
2818         }
2819
2820         if (optlen < sizeof(int))
2821                 return -EINVAL;
2822
2823         if (get_user(val, (int __user *)optval))
2824                 return -EFAULT;
2825
2826         lock_sock(sk);
2827
2828         switch (optname) {
2829         case TCP_MAXSEG:
2830                 /* Values greater than interface MTU won't take effect. However
2831                  * at the point when this call is done we typically don't yet
2832                  * know which interface is going to be used
2833                  */
2834                 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
2835                         err = -EINVAL;
2836                         break;
2837                 }
2838                 tp->rx_opt.user_mss = val;
2839                 break;
2840
2841         case TCP_NODELAY:
2842                 if (val) {
2843                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2844                          * this option on corked socket is remembered, but
2845                          * it is not activated until cork is cleared.
2846                          *
2847                          * However, when TCP_NODELAY is set we make
2848                          * an explicit push, which overrides even TCP_CORK
2849                          * for currently queued segments.
2850                          */
2851                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2852                         tcp_push_pending_frames(sk);
2853                 } else {
2854                         tp->nonagle &= ~TCP_NAGLE_OFF;
2855                 }
2856                 break;
2857
2858         case TCP_THIN_LINEAR_TIMEOUTS:
2859                 if (val < 0 || val > 1)
2860                         err = -EINVAL;
2861                 else
2862                         tp->thin_lto = val;
2863                 break;
2864
2865         case TCP_THIN_DUPACK:
2866                 if (val < 0 || val > 1)
2867                         err = -EINVAL;
2868                 break;
2869
2870         case TCP_REPAIR:
2871                 if (!tcp_can_repair_sock(sk))
2872                         err = -EPERM;
2873                 else if (val == TCP_REPAIR_ON) {
2874                         tp->repair = 1;
2875                         sk->sk_reuse = SK_FORCE_REUSE;
2876                         tp->repair_queue = TCP_NO_QUEUE;
2877                 } else if (val == TCP_REPAIR_OFF) {
2878                         tp->repair = 0;
2879                         sk->sk_reuse = SK_NO_REUSE;
2880                         tcp_send_window_probe(sk);
2881                 } else if (val == TCP_REPAIR_OFF_NO_WP) {
2882                         tp->repair = 0;
2883                         sk->sk_reuse = SK_NO_REUSE;
2884                 } else
2885                         err = -EINVAL;
2886
2887                 break;
2888
2889         case TCP_REPAIR_QUEUE:
2890                 if (!tp->repair)
2891                         err = -EPERM;
2892                 else if ((unsigned int)val < TCP_QUEUES_NR)
2893                         tp->repair_queue = val;
2894                 else
2895                         err = -EINVAL;
2896                 break;
2897
2898         case TCP_QUEUE_SEQ:
2899                 if (sk->sk_state != TCP_CLOSE)
2900                         err = -EPERM;
2901                 else if (tp->repair_queue == TCP_SEND_QUEUE)
2902                         tp->write_seq = val;
2903                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2904                         tp->rcv_nxt = val;
2905                 else
2906                         err = -EINVAL;
2907                 break;
2908
2909         case TCP_REPAIR_OPTIONS:
2910                 if (!tp->repair)
2911                         err = -EINVAL;
2912                 else if (sk->sk_state == TCP_ESTABLISHED)
2913                         err = tcp_repair_options_est(sk,
2914                                         (struct tcp_repair_opt __user *)optval,
2915                                         optlen);
2916                 else
2917                         err = -EPERM;
2918                 break;
2919
2920         case TCP_CORK:
2921                 /* When set indicates to always queue non-full frames.
2922                  * Later the user clears this option and we transmit
2923                  * any pending partial frames in the queue.  This is
2924                  * meant to be used alongside sendfile() to get properly
2925                  * filled frames when the user (for example) must write
2926                  * out headers with a write() call first and then use
2927                  * sendfile to send out the data parts.
2928                  *
2929                  * TCP_CORK can be set together with TCP_NODELAY and it is
2930                  * stronger than TCP_NODELAY.
2931                  */
2932                 if (val) {
2933                         tp->nonagle |= TCP_NAGLE_CORK;
2934                 } else {
2935                         tp->nonagle &= ~TCP_NAGLE_CORK;
2936                         if (tp->nonagle&TCP_NAGLE_OFF)
2937                                 tp->nonagle |= TCP_NAGLE_PUSH;
2938                         tcp_push_pending_frames(sk);
2939                 }
2940                 break;
2941
2942         case TCP_KEEPIDLE:
2943                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2944                         err = -EINVAL;
2945                 else {
2946                         tp->keepalive_time = val * HZ;
2947                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2948                             !((1 << sk->sk_state) &
2949                               (TCPF_CLOSE | TCPF_LISTEN))) {
2950                                 u32 elapsed = keepalive_time_elapsed(tp);
2951                                 if (tp->keepalive_time > elapsed)
2952                                         elapsed = tp->keepalive_time - elapsed;
2953                                 else
2954                                         elapsed = 0;
2955                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2956                         }
2957                 }
2958                 break;
2959         case TCP_KEEPINTVL:
2960                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2961                         err = -EINVAL;
2962                 else
2963                         tp->keepalive_intvl = val * HZ;
2964                 break;
2965         case TCP_KEEPCNT:
2966                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2967                         err = -EINVAL;
2968                 else
2969                         tp->keepalive_probes = val;
2970                 break;
2971         case TCP_SYNCNT:
2972                 if (val < 1 || val > MAX_TCP_SYNCNT)
2973                         err = -EINVAL;
2974                 else
2975                         icsk->icsk_syn_retries = val;
2976                 break;
2977
2978         case TCP_SAVE_SYN:
2979                 if (val < 0 || val > 1)
2980                         err = -EINVAL;
2981                 else
2982                         tp->save_syn = val;
2983                 break;
2984
2985         case TCP_LINGER2:
2986                 if (val < 0)
2987                         tp->linger2 = -1;
2988                 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
2989                         tp->linger2 = 0;
2990                 else
2991                         tp->linger2 = val * HZ;
2992                 break;
2993
2994         case TCP_DEFER_ACCEPT:
2995                 /* Translate value in seconds to number of retransmits */
2996                 icsk->icsk_accept_queue.rskq_defer_accept =
2997                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2998                                         TCP_RTO_MAX / HZ);
2999                 break;
3000
3001         case TCP_WINDOW_CLAMP:
3002                 if (!val) {
3003                         if (sk->sk_state != TCP_CLOSE) {
3004                                 err = -EINVAL;
3005                                 break;
3006                         }
3007                         tp->window_clamp = 0;
3008                 } else
3009                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3010                                                 SOCK_MIN_RCVBUF / 2 : val;
3011                 break;
3012
3013         case TCP_QUICKACK:
3014                 if (!val) {
3015                         inet_csk_enter_pingpong_mode(sk);
3016                 } else {
3017                         inet_csk_exit_pingpong_mode(sk);
3018                         if ((1 << sk->sk_state) &
3019                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3020                             inet_csk_ack_scheduled(sk)) {
3021                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
3022                                 tcp_cleanup_rbuf(sk, 1);
3023                                 if (!(val & 1))
3024                                         inet_csk_enter_pingpong_mode(sk);
3025                         }
3026                 }
3027                 break;
3028
3029 #ifdef CONFIG_TCP_MD5SIG
3030         case TCP_MD5SIG:
3031         case TCP_MD5SIG_EXT:
3032                 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
3033                         err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3034                 else
3035                         err = -EINVAL;
3036                 break;
3037 #endif
3038         case TCP_USER_TIMEOUT:
3039                 /* Cap the max time in ms TCP will retry or probe the window
3040                  * before giving up and aborting (ETIMEDOUT) a connection.
3041                  */
3042                 if (val < 0)
3043                         err = -EINVAL;
3044                 else
3045                         icsk->icsk_user_timeout = val;
3046                 break;
3047
3048         case TCP_FASTOPEN:
3049                 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3050                     TCPF_LISTEN))) {
3051                         tcp_fastopen_init_key_once(net);
3052
3053                         fastopen_queue_tune(sk, val);
3054                 } else {
3055                         err = -EINVAL;
3056                 }
3057                 break;
3058         case TCP_FASTOPEN_CONNECT:
3059                 if (val > 1 || val < 0) {
3060                         err = -EINVAL;
3061                 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
3062                         if (sk->sk_state == TCP_CLOSE)
3063                                 tp->fastopen_connect = val;
3064                         else
3065                                 err = -EINVAL;
3066                 } else {
3067                         err = -EOPNOTSUPP;
3068                 }
3069                 break;
3070         case TCP_FASTOPEN_NO_COOKIE:
3071                 if (val > 1 || val < 0)
3072                         err = -EINVAL;
3073                 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3074                         err = -EINVAL;
3075                 else
3076                         tp->fastopen_no_cookie = val;
3077                 break;
3078         case TCP_TIMESTAMP:
3079                 if (!tp->repair)
3080                         err = -EPERM;
3081                 else
3082                         tp->tsoffset = val - tcp_time_stamp_raw();
3083                 break;
3084         case TCP_REPAIR_WINDOW:
3085                 err = tcp_repair_set_window(tp, optval, optlen);
3086                 break;
3087         case TCP_NOTSENT_LOWAT:
3088                 tp->notsent_lowat = val;
3089                 sk->sk_write_space(sk);
3090                 break;
3091         case TCP_INQ:
3092                 if (val > 1 || val < 0)
3093                         err = -EINVAL;
3094                 else
3095                         tp->recvmsg_inq = val;
3096                 break;
3097         default:
3098                 err = -ENOPROTOOPT;
3099                 break;
3100         }
3101
3102         release_sock(sk);
3103         return err;
3104 }
3105
3106 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
3107                    unsigned int optlen)
3108 {
3109         const struct inet_connection_sock *icsk = inet_csk(sk);
3110
3111         if (level != SOL_TCP)
3112                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3113                                                      optval, optlen);
3114         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3115 }
3116 EXPORT_SYMBOL(tcp_setsockopt);
3117
3118 #ifdef CONFIG_COMPAT
3119 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
3120                           char __user *optval, unsigned int optlen)
3121 {
3122         if (level != SOL_TCP)
3123                 return inet_csk_compat_setsockopt(sk, level, optname,
3124                                                   optval, optlen);
3125         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3126 }
3127 EXPORT_SYMBOL(compat_tcp_setsockopt);
3128 #endif
3129
3130 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3131                                       struct tcp_info *info)
3132 {
3133         u64 stats[__TCP_CHRONO_MAX], total = 0;
3134         enum tcp_chrono i;
3135
3136         for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3137                 stats[i] = tp->chrono_stat[i - 1];
3138                 if (i == tp->chrono_type)
3139                         stats[i] += tcp_jiffies32 - tp->chrono_start;
3140                 stats[i] *= USEC_PER_SEC / HZ;
3141                 total += stats[i];
3142         }
3143
3144         info->tcpi_busy_time = total;
3145         info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3146         info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3147 }
3148
3149 /* Return information about state of tcp endpoint in API format. */
3150 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3151 {
3152         const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3153         const struct inet_connection_sock *icsk = inet_csk(sk);
3154         unsigned long rate;
3155         u32 now;
3156         u64 rate64;
3157         bool slow;
3158
3159         memset(info, 0, sizeof(*info));
3160         if (sk->sk_type != SOCK_STREAM)
3161                 return;
3162
3163         info->tcpi_state = inet_sk_state_load(sk);
3164
3165         /* Report meaningful fields for all TCP states, including listeners */
3166         rate = READ_ONCE(sk->sk_pacing_rate);
3167         rate64 = (rate != ~0UL) ? rate : ~0ULL;
3168         info->tcpi_pacing_rate = rate64;
3169
3170         rate = READ_ONCE(sk->sk_max_pacing_rate);
3171         rate64 = (rate != ~0UL) ? rate : ~0ULL;
3172         info->tcpi_max_pacing_rate = rate64;
3173
3174         info->tcpi_reordering = tp->reordering;
3175         info->tcpi_snd_cwnd = tp->snd_cwnd;
3176
3177         if (info->tcpi_state == TCP_LISTEN) {
3178                 /* listeners aliased fields :
3179                  * tcpi_unacked -> Number of children ready for accept()
3180                  * tcpi_sacked  -> max backlog
3181                  */
3182                 info->tcpi_unacked = sk->sk_ack_backlog;
3183                 info->tcpi_sacked = sk->sk_max_ack_backlog;
3184                 return;
3185         }
3186
3187         slow = lock_sock_fast(sk);
3188
3189         info->tcpi_ca_state = icsk->icsk_ca_state;
3190         info->tcpi_retransmits = icsk->icsk_retransmits;
3191         info->tcpi_probes = icsk->icsk_probes_out;
3192         info->tcpi_backoff = icsk->icsk_backoff;
3193
3194         if (tp->rx_opt.tstamp_ok)
3195                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3196         if (tcp_is_sack(tp))
3197                 info->tcpi_options |= TCPI_OPT_SACK;
3198         if (tp->rx_opt.wscale_ok) {
3199                 info->tcpi_options |= TCPI_OPT_WSCALE;
3200                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3201                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3202         }
3203
3204         if (tp->ecn_flags & TCP_ECN_OK)
3205                 info->tcpi_options |= TCPI_OPT_ECN;
3206         if (tp->ecn_flags & TCP_ECN_SEEN)
3207                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3208         if (tp->syn_data_acked)
3209                 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3210
3211         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3212         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3213         info->tcpi_snd_mss = tp->mss_cache;
3214         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3215
3216         info->tcpi_unacked = tp->packets_out;
3217         info->tcpi_sacked = tp->sacked_out;
3218
3219         info->tcpi_lost = tp->lost_out;
3220         info->tcpi_retrans = tp->retrans_out;
3221
3222         now = tcp_jiffies32;
3223         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3224         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3225         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3226
3227         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3228         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3229         info->tcpi_rtt = tp->srtt_us >> 3;
3230         info->tcpi_rttvar = tp->mdev_us >> 2;
3231         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3232         info->tcpi_advmss = tp->advmss;
3233
3234         info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3235         info->tcpi_rcv_space = tp->rcvq_space.space;
3236
3237         info->tcpi_total_retrans = tp->total_retrans;
3238
3239         info->tcpi_bytes_acked = tp->bytes_acked;
3240         info->tcpi_bytes_received = tp->bytes_received;
3241         info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3242         tcp_get_info_chrono_stats(tp, info);
3243
3244         info->tcpi_segs_out = tp->segs_out;
3245         info->tcpi_segs_in = tp->segs_in;
3246
3247         info->tcpi_min_rtt = tcp_min_rtt(tp);
3248         info->tcpi_data_segs_in = tp->data_segs_in;
3249         info->tcpi_data_segs_out = tp->data_segs_out;
3250
3251         info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3252         rate64 = tcp_compute_delivery_rate(tp);
3253         if (rate64)
3254                 info->tcpi_delivery_rate = rate64;
3255         info->tcpi_delivered = tp->delivered;
3256         info->tcpi_delivered_ce = tp->delivered_ce;
3257         info->tcpi_bytes_sent = tp->bytes_sent;
3258         info->tcpi_bytes_retrans = tp->bytes_retrans;
3259         info->tcpi_dsack_dups = tp->dsack_dups;
3260         info->tcpi_reord_seen = tp->reord_seen;
3261         unlock_sock_fast(sk, slow);
3262 }
3263 EXPORT_SYMBOL_GPL(tcp_get_info);
3264
3265 static size_t tcp_opt_stats_get_size(void)
3266 {
3267         return
3268                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3269                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3270                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3271                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3272                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3273                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3274                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3275                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3276                 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3277                 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3278                 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3279                 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3280                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3281                 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3282                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3283                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3284                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3285                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3286                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3287                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3288                 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3289                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3290                 0;
3291 }
3292
3293 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk)
3294 {
3295         const struct tcp_sock *tp = tcp_sk(sk);
3296         struct sk_buff *stats;
3297         struct tcp_info info;
3298         unsigned long rate;
3299         u64 rate64;
3300
3301         stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3302         if (!stats)
3303                 return NULL;
3304
3305         tcp_get_info_chrono_stats(tp, &info);
3306         nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3307                           info.tcpi_busy_time, TCP_NLA_PAD);
3308         nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3309                           info.tcpi_rwnd_limited, TCP_NLA_PAD);
3310         nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3311                           info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3312         nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3313                           tp->data_segs_out, TCP_NLA_PAD);
3314         nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3315                           tp->total_retrans, TCP_NLA_PAD);
3316
3317         rate = READ_ONCE(sk->sk_pacing_rate);
3318         rate64 = (rate != ~0UL) ? rate : ~0ULL;
3319         nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3320
3321         rate64 = tcp_compute_delivery_rate(tp);
3322         nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3323
3324         nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3325         nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3326         nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3327
3328         nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3329         nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3330         nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3331         nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3332         nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3333
3334         nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3335         nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3336
3337         nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3338                           TCP_NLA_PAD);
3339         nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3340                           TCP_NLA_PAD);
3341         nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3342         nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3343         nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3344
3345         return stats;
3346 }
3347
3348 static int do_tcp_getsockopt(struct sock *sk, int level,
3349                 int optname, char __user *optval, int __user *optlen)
3350 {
3351         struct inet_connection_sock *icsk = inet_csk(sk);
3352         struct tcp_sock *tp = tcp_sk(sk);
3353         struct net *net = sock_net(sk);
3354         int val, len;
3355
3356         if (get_user(len, optlen))
3357                 return -EFAULT;
3358
3359         len = min_t(unsigned int, len, sizeof(int));
3360
3361         if (len < 0)
3362                 return -EINVAL;
3363
3364         switch (optname) {
3365         case TCP_MAXSEG:
3366                 val = tp->mss_cache;
3367                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3368                         val = tp->rx_opt.user_mss;
3369                 if (tp->repair)
3370                         val = tp->rx_opt.mss_clamp;
3371                 break;
3372         case TCP_NODELAY:
3373                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3374                 break;
3375         case TCP_CORK:
3376                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3377                 break;
3378         case TCP_KEEPIDLE:
3379                 val = keepalive_time_when(tp) / HZ;
3380                 break;
3381         case TCP_KEEPINTVL:
3382                 val = keepalive_intvl_when(tp) / HZ;
3383                 break;
3384         case TCP_KEEPCNT:
3385                 val = keepalive_probes(tp);
3386                 break;
3387         case TCP_SYNCNT:
3388                 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3389                 break;
3390         case TCP_LINGER2:
3391                 val = tp->linger2;
3392                 if (val >= 0)
3393                         val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3394                 break;
3395         case TCP_DEFER_ACCEPT:
3396                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3397                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3398                 break;
3399         case TCP_WINDOW_CLAMP:
3400                 val = tp->window_clamp;
3401                 break;
3402         case TCP_INFO: {
3403                 struct tcp_info info;
3404
3405                 if (get_user(len, optlen))
3406                         return -EFAULT;
3407
3408                 tcp_get_info(sk, &info);
3409
3410                 len = min_t(unsigned int, len, sizeof(info));
3411                 if (put_user(len, optlen))
3412                         return -EFAULT;
3413                 if (copy_to_user(optval, &info, len))
3414                         return -EFAULT;
3415                 return 0;
3416         }
3417         case TCP_CC_INFO: {
3418                 const struct tcp_congestion_ops *ca_ops;
3419                 union tcp_cc_info info;
3420                 size_t sz = 0;
3421                 int attr;
3422
3423                 if (get_user(len, optlen))
3424                         return -EFAULT;
3425
3426                 ca_ops = icsk->icsk_ca_ops;
3427                 if (ca_ops && ca_ops->get_info)
3428                         sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3429
3430                 len = min_t(unsigned int, len, sz);
3431                 if (put_user(len, optlen))
3432                         return -EFAULT;
3433                 if (copy_to_user(optval, &info, len))
3434                         return -EFAULT;
3435                 return 0;
3436         }
3437         case TCP_QUICKACK:
3438                 val = !inet_csk_in_pingpong_mode(sk);
3439                 break;
3440
3441         case TCP_CONGESTION:
3442                 if (get_user(len, optlen))
3443                         return -EFAULT;
3444                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
3445                 if (put_user(len, optlen))
3446                         return -EFAULT;
3447                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3448                         return -EFAULT;
3449                 return 0;
3450
3451         case TCP_ULP:
3452                 if (get_user(len, optlen))
3453                         return -EFAULT;
3454                 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
3455                 if (!icsk->icsk_ulp_ops) {
3456                         if (put_user(0, optlen))
3457                                 return -EFAULT;
3458                         return 0;
3459                 }
3460                 if (put_user(len, optlen))
3461                         return -EFAULT;
3462                 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
3463                         return -EFAULT;
3464                 return 0;
3465
3466         case TCP_FASTOPEN_KEY: {
3467                 __u8 key[TCP_FASTOPEN_KEY_LENGTH];
3468                 struct tcp_fastopen_context *ctx;
3469
3470                 if (get_user(len, optlen))
3471                         return -EFAULT;
3472
3473                 rcu_read_lock();
3474                 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
3475                 if (ctx)
3476                         memcpy(key, ctx->key, sizeof(key));
3477                 else
3478                         len = 0;
3479                 rcu_read_unlock();
3480
3481                 len = min_t(unsigned int, len, sizeof(key));
3482                 if (put_user(len, optlen))
3483                         return -EFAULT;
3484                 if (copy_to_user(optval, key, len))
3485                         return -EFAULT;
3486                 return 0;
3487         }
3488         case TCP_THIN_LINEAR_TIMEOUTS:
3489                 val = tp->thin_lto;
3490                 break;
3491
3492         case TCP_THIN_DUPACK:
3493                 val = 0;
3494                 break;
3495
3496         case TCP_REPAIR:
3497                 val = tp->repair;
3498                 break;
3499
3500         case TCP_REPAIR_QUEUE:
3501                 if (tp->repair)
3502                         val = tp->repair_queue;
3503                 else
3504                         return -EINVAL;
3505                 break;
3506
3507         case TCP_REPAIR_WINDOW: {
3508                 struct tcp_repair_window opt;
3509
3510                 if (get_user(len, optlen))
3511                         return -EFAULT;
3512
3513                 if (len != sizeof(opt))
3514                         return -EINVAL;
3515
3516                 if (!tp->repair)
3517                         return -EPERM;
3518
3519                 opt.snd_wl1     = tp->snd_wl1;
3520                 opt.snd_wnd     = tp->snd_wnd;
3521                 opt.max_window  = tp->max_window;
3522                 opt.rcv_wnd     = tp->rcv_wnd;
3523                 opt.rcv_wup     = tp->rcv_wup;
3524
3525                 if (copy_to_user(optval, &opt, len))
3526                         return -EFAULT;
3527                 return 0;
3528         }
3529         case TCP_QUEUE_SEQ:
3530                 if (tp->repair_queue == TCP_SEND_QUEUE)
3531                         val = tp->write_seq;
3532                 else if (tp->repair_queue == TCP_RECV_QUEUE)
3533                         val = tp->rcv_nxt;
3534                 else
3535                         return -EINVAL;
3536                 break;
3537
3538         case TCP_USER_TIMEOUT:
3539                 val = icsk->icsk_user_timeout;
3540                 break;
3541
3542         case TCP_FASTOPEN:
3543                 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3544                 break;
3545
3546         case TCP_FASTOPEN_CONNECT:
3547                 val = tp->fastopen_connect;
3548                 break;
3549
3550         case TCP_FASTOPEN_NO_COOKIE:
3551                 val = tp->fastopen_no_cookie;
3552                 break;
3553
3554         case TCP_TIMESTAMP:
3555                 val = tcp_time_stamp_raw() + tp->tsoffset;
3556                 break;
3557         case TCP_NOTSENT_LOWAT:
3558                 val = tp->notsent_lowat;
3559                 break;
3560         case TCP_INQ:
3561                 val = tp->recvmsg_inq;
3562                 break;
3563         case TCP_SAVE_SYN:
3564                 val = tp->save_syn;
3565                 break;
3566         case TCP_SAVED_SYN: {
3567                 if (get_user(len, optlen))
3568                         return -EFAULT;
3569
3570                 lock_sock(sk);
3571                 if (tp->saved_syn) {
3572                         if (len < tp->saved_syn[0]) {
3573                                 if (put_user(tp->saved_syn[0], optlen)) {
3574                                         release_sock(sk);
3575                                         return -EFAULT;
3576                                 }
3577                                 release_sock(sk);
3578                                 return -EINVAL;
3579                         }
3580                         len = tp->saved_syn[0];
3581                         if (put_user(len, optlen)) {
3582                                 release_sock(sk);
3583                                 return -EFAULT;
3584                         }
3585                         if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3586                                 release_sock(sk);
3587                                 return -EFAULT;
3588                         }
3589                         tcp_saved_syn_free(tp);
3590                         release_sock(sk);
3591                 } else {
3592                         release_sock(sk);
3593                         len = 0;
3594                         if (put_user(len, optlen))
3595                                 return -EFAULT;
3596                 }
3597                 return 0;
3598         }
3599 #ifdef CONFIG_MMU
3600         case TCP_ZEROCOPY_RECEIVE: {
3601                 struct tcp_zerocopy_receive zc;
3602                 int err;
3603
3604                 if (get_user(len, optlen))
3605                         return -EFAULT;
3606                 if (len != sizeof(zc))
3607                         return -EINVAL;
3608                 if (copy_from_user(&zc, optval, len))
3609                         return -EFAULT;
3610                 lock_sock(sk);
3611                 err = tcp_zerocopy_receive(sk, &zc);
3612                 release_sock(sk);
3613                 if (!err && copy_to_user(optval, &zc, len))
3614                         err = -EFAULT;
3615                 return err;
3616         }
3617 #endif
3618         default:
3619                 return -ENOPROTOOPT;
3620         }
3621
3622         if (put_user(len, optlen))
3623                 return -EFAULT;
3624         if (copy_to_user(optval, &val, len))
3625                 return -EFAULT;
3626         return 0;
3627 }
3628
3629 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3630                    int __user *optlen)
3631 {
3632         struct inet_connection_sock *icsk = inet_csk(sk);
3633
3634         if (level != SOL_TCP)
3635                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3636                                                      optval, optlen);
3637         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3638 }
3639 EXPORT_SYMBOL(tcp_getsockopt);
3640
3641 #ifdef CONFIG_COMPAT
3642 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3643                           char __user *optval, int __user *optlen)
3644 {
3645         if (level != SOL_TCP)
3646                 return inet_csk_compat_getsockopt(sk, level, optname,
3647                                                   optval, optlen);
3648         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3649 }
3650 EXPORT_SYMBOL(compat_tcp_getsockopt);
3651 #endif
3652
3653 #ifdef CONFIG_TCP_MD5SIG
3654 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3655 static DEFINE_MUTEX(tcp_md5sig_mutex);
3656 static bool tcp_md5sig_pool_populated = false;
3657
3658 static void __tcp_alloc_md5sig_pool(void)
3659 {
3660         struct crypto_ahash *hash;
3661         int cpu;
3662
3663         hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3664         if (IS_ERR(hash))
3665                 return;
3666
3667         for_each_possible_cpu(cpu) {
3668                 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3669                 struct ahash_request *req;
3670
3671                 if (!scratch) {
3672                         scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3673                                                sizeof(struct tcphdr),
3674                                                GFP_KERNEL,
3675                                                cpu_to_node(cpu));
3676                         if (!scratch)
3677                                 return;
3678                         per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3679                 }
3680                 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3681                         continue;
3682
3683                 req = ahash_request_alloc(hash, GFP_KERNEL);
3684                 if (!req)
3685                         return;
3686
3687                 ahash_request_set_callback(req, 0, NULL, NULL);
3688
3689                 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3690         }
3691         /* before setting tcp_md5sig_pool_populated, we must commit all writes
3692          * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3693          */
3694         smp_wmb();
3695         tcp_md5sig_pool_populated = true;
3696 }
3697
3698 bool tcp_alloc_md5sig_pool(void)
3699 {
3700         if (unlikely(!tcp_md5sig_pool_populated)) {
3701                 mutex_lock(&tcp_md5sig_mutex);
3702
3703                 if (!tcp_md5sig_pool_populated) {
3704                         __tcp_alloc_md5sig_pool();
3705                         if (tcp_md5sig_pool_populated)
3706                                 static_branch_inc(&tcp_md5_needed);
3707                 }
3708
3709                 mutex_unlock(&tcp_md5sig_mutex);
3710         }
3711         return tcp_md5sig_pool_populated;
3712 }
3713 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3714
3715
3716 /**
3717  *      tcp_get_md5sig_pool - get md5sig_pool for this user
3718  *
3719  *      We use percpu structure, so if we succeed, we exit with preemption
3720  *      and BH disabled, to make sure another thread or softirq handling
3721  *      wont try to get same context.
3722  */
3723 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3724 {
3725         local_bh_disable();
3726
3727         if (tcp_md5sig_pool_populated) {
3728                 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3729                 smp_rmb();
3730                 return this_cpu_ptr(&tcp_md5sig_pool);
3731         }
3732         local_bh_enable();
3733         return NULL;
3734 }
3735 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3736
3737 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3738                           const struct sk_buff *skb, unsigned int header_len)
3739 {
3740         struct scatterlist sg;
3741         const struct tcphdr *tp = tcp_hdr(skb);
3742         struct ahash_request *req = hp->md5_req;
3743         unsigned int i;
3744         const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3745                                            skb_headlen(skb) - header_len : 0;
3746         const struct skb_shared_info *shi = skb_shinfo(skb);
3747         struct sk_buff *frag_iter;
3748
3749         sg_init_table(&sg, 1);
3750
3751         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3752         ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3753         if (crypto_ahash_update(req))
3754                 return 1;
3755
3756         for (i = 0; i < shi->nr_frags; ++i) {
3757                 const struct skb_frag_struct *f = &shi->frags[i];
3758                 unsigned int offset = f->page_offset;
3759                 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3760
3761                 sg_set_page(&sg, page, skb_frag_size(f),
3762                             offset_in_page(offset));
3763                 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3764                 if (crypto_ahash_update(req))
3765                         return 1;
3766         }
3767
3768         skb_walk_frags(skb, frag_iter)
3769                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3770                         return 1;
3771
3772         return 0;
3773 }
3774 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3775
3776 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3777 {
3778         struct scatterlist sg;
3779
3780         sg_init_one(&sg, key->key, key->keylen);
3781         ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen);
3782         return crypto_ahash_update(hp->md5_req);
3783 }
3784 EXPORT_SYMBOL(tcp_md5_hash_key);
3785
3786 #endif
3787
3788 void tcp_done(struct sock *sk)
3789 {
3790         struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3791
3792         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3793                 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3794
3795         tcp_set_state(sk, TCP_CLOSE);
3796         tcp_clear_xmit_timers(sk);
3797         if (req)
3798                 reqsk_fastopen_remove(sk, req, false);
3799
3800         sk->sk_shutdown = SHUTDOWN_MASK;
3801
3802         if (!sock_flag(sk, SOCK_DEAD))
3803                 sk->sk_state_change(sk);
3804         else
3805                 inet_csk_destroy_sock(sk);
3806 }
3807 EXPORT_SYMBOL_GPL(tcp_done);
3808
3809 int tcp_abort(struct sock *sk, int err)
3810 {
3811         if (!sk_fullsock(sk)) {
3812                 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3813                         struct request_sock *req = inet_reqsk(sk);
3814
3815                         local_bh_disable();
3816                         inet_csk_reqsk_queue_drop(req->rsk_listener, req);
3817                         local_bh_enable();
3818                         return 0;
3819                 }
3820                 return -EOPNOTSUPP;
3821         }
3822
3823         /* Don't race with userspace socket closes such as tcp_close. */
3824         lock_sock(sk);
3825
3826         if (sk->sk_state == TCP_LISTEN) {
3827                 tcp_set_state(sk, TCP_CLOSE);
3828                 inet_csk_listen_stop(sk);
3829         }
3830
3831         /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3832         local_bh_disable();
3833         bh_lock_sock(sk);
3834
3835         if (!sock_flag(sk, SOCK_DEAD)) {
3836                 sk->sk_err = err;
3837                 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3838                 smp_wmb();
3839                 sk->sk_error_report(sk);
3840                 if (tcp_need_reset(sk->sk_state))
3841                         tcp_send_active_reset(sk, GFP_ATOMIC);
3842                 tcp_done(sk);
3843         }
3844
3845         bh_unlock_sock(sk);
3846         local_bh_enable();
3847         tcp_write_queue_purge(sk);
3848         release_sock(sk);
3849         return 0;
3850 }
3851 EXPORT_SYMBOL_GPL(tcp_abort);
3852
3853 extern struct tcp_congestion_ops tcp_reno;
3854
3855 static __initdata unsigned long thash_entries;
3856 static int __init set_thash_entries(char *str)
3857 {
3858         ssize_t ret;
3859
3860         if (!str)
3861                 return 0;
3862
3863         ret = kstrtoul(str, 0, &thash_entries);
3864         if (ret)
3865                 return 0;
3866
3867         return 1;
3868 }
3869 __setup("thash_entries=", set_thash_entries);
3870
3871 static void __init tcp_init_mem(void)
3872 {
3873         unsigned long limit = nr_free_buffer_pages() / 16;
3874
3875         limit = max(limit, 128UL);
3876         sysctl_tcp_mem[0] = limit / 4 * 3;              /* 4.68 % */
3877         sysctl_tcp_mem[1] = limit;                      /* 6.25 % */
3878         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;      /* 9.37 % */
3879 }
3880
3881 void __init tcp_init(void)
3882 {
3883         int max_rshare, max_wshare, cnt;
3884         unsigned long limit;
3885         unsigned int i;
3886
3887         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3888                      FIELD_SIZEOF(struct sk_buff, cb));
3889
3890         percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3891         percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3892         inet_hashinfo_init(&tcp_hashinfo);
3893         inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
3894                             thash_entries, 21,  /* one slot per 2 MB*/
3895                             0, 64 * 1024);
3896         tcp_hashinfo.bind_bucket_cachep =
3897                 kmem_cache_create("tcp_bind_bucket",
3898                                   sizeof(struct inet_bind_bucket), 0,
3899                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3900
3901         /* Size and allocate the main established and bind bucket
3902          * hash tables.
3903          *
3904          * The methodology is similar to that of the buffer cache.
3905          */
3906         tcp_hashinfo.ehash =
3907                 alloc_large_system_hash("TCP established",
3908                                         sizeof(struct inet_ehash_bucket),
3909                                         thash_entries,
3910                                         17, /* one slot per 128 KB of memory */
3911                                         0,
3912                                         NULL,
3913                                         &tcp_hashinfo.ehash_mask,
3914                                         0,
3915                                         thash_entries ? 0 : 512 * 1024);
3916         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3917                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3918
3919         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3920                 panic("TCP: failed to alloc ehash_locks");
3921         tcp_hashinfo.bhash =
3922                 alloc_large_system_hash("TCP bind",
3923                                         sizeof(struct inet_bind_hashbucket),
3924                                         tcp_hashinfo.ehash_mask + 1,
3925                                         17, /* one slot per 128 KB of memory */
3926                                         0,
3927                                         &tcp_hashinfo.bhash_size,
3928                                         NULL,
3929                                         0,
3930                                         64 * 1024);
3931         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3932         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3933                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3934                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3935         }
3936
3937
3938         cnt = tcp_hashinfo.ehash_mask + 1;
3939         sysctl_tcp_max_orphans = cnt / 2;
3940
3941         tcp_init_mem();
3942         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3943         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3944         max_wshare = min(4UL*1024*1024, limit);
3945         max_rshare = min(6UL*1024*1024, limit);
3946
3947         init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3948         init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
3949         init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3950
3951         init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3952         init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
3953         init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
3954
3955         pr_info("Hash tables configured (established %u bind %u)\n",
3956                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3957
3958         tcp_v4_init();
3959         tcp_metrics_init();
3960         BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3961         tcp_tasklet_init();
3962 }