net/ipv4/tcp.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the  BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              Implementation of the Transmission Control Protocol(TCP).
   7  *
   8  * Authors:     Ross Biro
   9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
  11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
  12  *              Florian La Roche, <flla@stud.uni-sb.de>
  13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
  15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
  16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
  17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  18  *              Jorge Cwik, <jorge@laser.satlink.net>
  19  *
  20  * Fixes:
  21  *              Alan Cox        :       Numerous verify_area() calls
  22  *              Alan Cox        :       Set the ACK bit on a reset
  23  *              Alan Cox        :       Stopped it crashing if it closed while
  24  *                                      sk->inuse=1 and was trying to connect
  25  *                                      (tcp_err()).
  26  *              Alan Cox        :       All icmp error handling was broken
  27  *                                      pointers passed where wrong and the
  28  *                                      socket was looked up backwards. Nobody
  29  *                                      tested any icmp error code obviously.
  30  *              Alan Cox        :       tcp_err() now handled properly. It
  31  *                                      wakes people on errors. poll
  32  *                                      behaves and the icmp error race
  33  *                                      has gone by moving it into sock.c
  34  *              Alan Cox        :       tcp_send_reset() fixed to work for
  35  *                                      everything not just packets for
  36  *                                      unknown sockets.
  37  *              Alan Cox        :       tcp option processing.
  38  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
  39  *                                      syn rule wrong]
  40  *              Herp Rosmanith  :       More reset fixes
  41  *              Alan Cox        :       No longer acks invalid rst frames.
  42  *                                      Acking any kind of RST is right out.
  43  *              Alan Cox        :       Sets an ignore me flag on an rst
  44  *                                      receive otherwise odd bits of prattle
  45  *                                      escape still
  46  *              Alan Cox        :       Fixed another acking RST frame bug.
  47  *                                      Should stop LAN workplace lockups.
  48  *              Alan Cox        :       Some tidyups using the new skb list
  49  *                                      facilities
  50  *              Alan Cox        :       sk->keepopen now seems to work
  51  *              Alan Cox        :       Pulls options out correctly on accepts
  52  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
  53  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
  54  *                                      bit to skb ops.
  55  *              Alan Cox        :       Tidied tcp_data to avoid a potential
  56  *                                      nasty.
  57  *              Alan Cox        :       Added some better commenting, as the
  58  *                                      tcp is hard to follow
  59  *              Alan Cox        :       Removed incorrect check for 20 * psh
  60  *      Michael O'Reilly        :       ack < copied bug fix.
  61  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
  62  *              Alan Cox        :       FIN with no memory -> CRASH
  63  *              Alan Cox        :       Added socket option proto entries.
  64  *                                      Also added awareness of them to accept.
  65  *              Alan Cox        :       Added TCP options (SOL_TCP)
  66  *              Alan Cox        :       Switched wakeup calls to callbacks,
  67  *                                      so the kernel can layer network
  68  *                                      sockets.
  69  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
  70  *              Alan Cox        :       Handle FIN (more) properly (we hope).
  71  *              Alan Cox        :       RST frames sent on unsynchronised
  72  *                                      state ack error.
  73  *              Alan Cox        :       Put in missing check for SYN bit.
  74  *              Alan Cox        :       Added tcp_select_window() aka NET2E
  75  *                                      window non shrink trick.
  76  *              Alan Cox        :       Added a couple of small NET2E timer
  77  *                                      fixes
  78  *              Charles Hedrick :       TCP fixes
  79  *              Toomas Tamm     :       TCP window fixes
  80  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
  81  *              Charles Hedrick :       Rewrote most of it to actually work
  82  *              Linus           :       Rewrote tcp_read() and URG handling
  83  *                                      completely
  84  *              Gerhard Koerting:       Fixed some missing timer handling
  85  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
  86  *              Gerhard Koerting:       PC/TCP workarounds
  87  *              Adam Caldwell   :       Assorted timer/timing errors
  88  *              Matthew Dillon  :       Fixed another RST bug
  89  *              Alan Cox        :       Move to kernel side addressing changes.
  90  *              Alan Cox        :       Beginning work on TCP fastpathing
  91  *                                      (not yet usable)
  92  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
  93  *              Alan Cox        :       TCP fast path debugging
  94  *              Alan Cox        :       Window clamping
  95  *              Michael Riepe   :       Bug in tcp_check()
  96  *              Matt Dillon     :       More TCP improvements and RST bug fixes
  97  *              Matt Dillon     :       Yet more small nasties remove from the
  98  *                                      TCP code (Be very nice to this man if
  99  *                                      tcp finally works 100%) 8)
 100  *              Alan Cox        :       BSD accept semantics.
 101  *              Alan Cox        :       Reset on closedown bug.
 102  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
 103  *              Michael Pall    :       Handle poll() after URG properly in
 104  *                                      all cases.
 105  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
 106  *                                      (multi URG PUSH broke rlogin).
 107  *              Michael Pall    :       Fix the multi URG PUSH problem in
 108  *                                      tcp_readable(), poll() after URG
 109  *                                      works now.
 110  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
 111  *                                      BSD api.
 112  *              Alan Cox        :       Changed the semantics of sk->socket to
 113  *                                      fix a race and a signal problem with
 114  *                                      accept() and async I/O.
 115  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
 116  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
 117  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
 118  *                                      clients/servers which listen in on
 119  *                                      fixed ports.
 120  *              Alan Cox        :       Cleaned the above up and shrank it to
 121  *                                      a sensible code size.
 122  *              Alan Cox        :       Self connect lockup fix.
 123  *              Alan Cox        :       No connect to multicast.
 124  *              Ross Biro       :       Close unaccepted children on master
 125  *                                      socket close.
 126  *              Alan Cox        :       Reset tracing code.
 127  *              Alan Cox        :       Spurious resets on shutdown.
 128  *              Alan Cox        :       Giant 15 minute/60 second timer error
 129  *              Alan Cox        :       Small whoops in polling before an
 130  *                                      accept.
 131  *              Alan Cox        :       Kept the state trace facility since
 132  *                                      it's handy for debugging.
 133  *              Alan Cox        :       More reset handler fixes.
 134  *              Alan Cox        :       Started rewriting the code based on
 135  *                                      the RFC's for other useful protocol
 136  *                                      references see: Comer, KA9Q NOS, and
 137  *                                      for a reference on the difference
 138  *                                      between specifications and how BSD
 139  *                                      works see the 4.4lite source.
 140  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
 141  *                                      close.
 142  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
 143  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
 144  *              Alan Cox        :       Reimplemented timers as per the RFC
 145  *                                      and using multiple timers for sanity.
 146  *              Alan Cox        :       Small bug fixes, and a lot of new
 147  *                                      comments.
 148  *              Alan Cox        :       Fixed dual reader crash by locking
 149  *                                      the buffers (much like datagram.c)
 150  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
 151  *                                      now gets fed up of retrying without
 152  *                                      (even a no space) answer.
 153  *              Alan Cox        :       Extracted closing code better
 154  *              Alan Cox        :       Fixed the closing state machine to
 155  *                                      resemble the RFC.
 156  *              Alan Cox        :       More 'per spec' fixes.
 157  *              Jorge Cwik      :       Even faster checksumming.
 158  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
 159  *                                      only frames. At least one pc tcp stack
 160  *                                      generates them.
 161  *              Alan Cox        :       Cache last socket.
 162  *              Alan Cox        :       Per route irtt.
 163  *              Matt Day        :       poll()->select() match BSD precisely on error
 164  *              Alan Cox        :       New buffers
 165  *              Marc Tamsky     :       Various sk->prot->retransmits and
 166  *                                      sk->retransmits misupdating fixed.
 167  *                                      Fixed tcp_write_timeout: stuck close,
 168  *                                      and TCP syn retries gets used now.
 169  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
 170  *                                      ack if state is TCP_CLOSED.
 171  *              Alan Cox        :       Look up device on a retransmit - routes may
 172  *                                      change. Doesn't yet cope with MSS shrink right
 173  *                                      but it's a start!
 174  *              Marc Tamsky     :       Closing in closing fixes.
 175  *              Mike Shaver     :       RFC1122 verifications.
 176  *              Alan Cox        :       rcv_saddr errors.
 177  *              Alan Cox        :       Block double connect().
 178  *              Alan Cox        :       Small hooks for enSKIP.
 179  *              Alexey Kuznetsov:       Path MTU discovery.
 180  *              Alan Cox        :       Support soft errors.
 181  *              Alan Cox        :       Fix MTU discovery pathological case
 182  *                                      when the remote claims no mtu!
 183  *              Marc Tamsky     :       TCP_CLOSE fix.
 184  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
 185  *                                      window but wrong (fixes NT lpd problems)
 186  *              Pedro Roque     :       Better TCP window handling, delayed ack.
 187  *              Joerg Reuter    :       No modification of locked buffers in
 188  *                                      tcp_do_retransmit()
 189  *              Eric Schenk     :       Changed receiver side silly window
 190  *                                      avoidance algorithm to BSD style
 191  *                                      algorithm. This doubles throughput
 192  *                                      against machines running Solaris,
 193  *                                      and seems to result in general
 194  *                                      improvement.
 195  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
 196  *      Willy Konynenberg       :       Transparent proxying support.
 197  *      Mike McLagan            :       Routing by source
 198  *              Keith Owens     :       Do proper merging with partial SKB's in
 199  *                                      tcp_do_sendmsg to avoid burstiness.
 200  *              Eric Schenk     :       Fix fast close down bug with
 201  *                                      shutdown() followed by close().
 202  *              Andi Kleen      :       Make poll agree with SIGIO
 203  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
 204  *                                      lingertime == 0 (RFC 793 ABORT Call)
 205  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
 206  *                                      csum_and_copy_from_user() if possible.
 207  *
 208  *              This program is free software; you can redistribute it and/or
 209  *              modify it under the terms of the GNU General Public License
 210  *              as published by the Free Software Foundation; either version
 211  *              2 of the License, or(at your option) any later version.
 212  *
 213  * Description of States:
 214  *
 215  *      TCP_SYN_SENT            sent a connection request, waiting for ack
 216  *
 217  *      TCP_SYN_RECV            received a connection request, sent ack,
 218  *                              waiting for final ack in three-way handshake.
 219  *
 220  *      TCP_ESTABLISHED         connection established
 221  *
 222  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
 223  *                              transmission of remaining buffered data
 224  *
 225  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
 226  *                              to shutdown
 227  *
 228  *      TCP_CLOSING             both sides have shutdown but we still have
 229  *                              data we have to finish sending
 230  *
 231  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
 232  *                              closed, can only be entered from FIN_WAIT2
 233  *                              or CLOSING.  Required because the other end
 234  *                              may not have gotten our last ACK causing it
 235  *                              to retransmit the data packet (which we ignore)
 236  *
 237  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
 238  *                              us to finish writing our data and to shutdown
 239  *                              (we have to close() to move on to LAST_ACK)
 240  *
 241  *      TCP_LAST_ACK            out side has shutdown after remote has
 242  *                              shutdown.  There may still be data in our
 243  *                              buffer that we have to finish sending
 244  *
 245  *      TCP_CLOSE               socket is finished
 246  */
 247
 248 #define pr_fmt(fmt) "TCP: " fmt
 249
 250 #include <crypto/hash.h>
 251 #include <linux/kernel.h>
 252 #include <linux/module.h>
 253 #include <linux/types.h>
 254 #include <linux/fcntl.h>
 255 #include <linux/poll.h>
 256 #include <linux/inet_diag.h>
 257 #include <linux/init.h>
 258 #include <linux/fs.h>
 259 #include <linux/skbuff.h>
 260 #include <linux/scatterlist.h>
 261 #include <linux/splice.h>
 262 #include <linux/net.h>
 263 #include <linux/socket.h>
 264 #include <linux/random.h>
 265 #include <linux/bootmem.h>
 266 #include <linux/highmem.h>
 267 #include <linux/swap.h>
 268 #include <linux/cache.h>
 269 #include <linux/err.h>
 270 #include <linux/time.h>
 271 #include <linux/slab.h>
 272 #include <linux/errqueue.h>
 273 #include <linux/static_key.h>
 274
 275 #include <net/icmp.h>
 276 #include <net/inet_common.h>
 277 #include <net/tcp.h>
 278 #include <net/xfrm.h>
 279 #include <net/ip.h>
 280 #include <net/sock.h>
 281
 282 #include <linux/uaccess.h>
 283 #include <asm/ioctls.h>
 284 #include <net/busy_poll.h>
 285
 286 struct percpu_counter tcp_orphan_count;
 287 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 288
 289 long sysctl_tcp_mem[3] __read_mostly;
 290 EXPORT_SYMBOL(sysctl_tcp_mem);
 291
 292 atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
 293 EXPORT_SYMBOL(tcp_memory_allocated);
 294
 295 #if IS_ENABLED(CONFIG_SMC)
 296 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
 297 EXPORT_SYMBOL(tcp_have_smc);
 298 #endif
 299
 300 /*
 301  * Current number of TCP sockets.
 302  */
 303 struct percpu_counter tcp_sockets_allocated;
 304 EXPORT_SYMBOL(tcp_sockets_allocated);
 305
 306 /*
 307  * TCP splice context
 308  */
 309 struct tcp_splice_state {
 310         struct pipe_inode_info *pipe;
 311         size_t len;
 312         unsigned int flags;
 313 };
 314
 315 /*
 316  * Pressure flag: try to collapse.
 317  * Technical note: it is used by multiple contexts non atomically.
 318  * All the __sk_mem_schedule() is of this nature: accounting
 319  * is strict, actions are advisory and have some latency.
 320  */
 321 unsigned long tcp_memory_pressure __read_mostly;
 322 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
 323
 324 void tcp_enter_memory_pressure(struct sock *sk)
 325 {
 326         unsigned long val;
 327
 328         if (tcp_memory_pressure)
 329                 return;
 330         val = jiffies;
 331
 332         if (!val)
 333                 val--;
 334         if (!cmpxchg(&tcp_memory_pressure, 0, val))
 335                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
 336 }
 337 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
 338
 339 void tcp_leave_memory_pressure(struct sock *sk)
 340 {
 341         unsigned long val;
 342
 343         if (!tcp_memory_pressure)
 344                 return;
 345         val = xchg(&tcp_memory_pressure, 0);
 346         if (val)
 347                 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
 348                               jiffies_to_msecs(jiffies - val));
 349 }
 350 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
 351
 352 /* Convert seconds to retransmits based on initial and max timeout */
 353 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
 354 {
 355         u8 res = 0;
 356
 357         if (seconds > 0) {
 358                 int period = timeout;
 359
 360                 res = 1;
 361                 while (seconds > period && res < 255) {
 362                         res++;
 363                         timeout <<= 1;
 364                         if (timeout > rto_max)
 365                                 timeout = rto_max;
 366                         period += timeout;
 367                 }
 368         }
 369         return res;
 370 }
 371
 372 /* Convert retransmits to seconds based on initial and max timeout */
 373 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
 374 {
 375         int period = 0;
 376
 377         if (retrans > 0) {
 378                 period = timeout;
 379                 while (--retrans) {
 380                         timeout <<= 1;
 381                         if (timeout > rto_max)
 382                                 timeout = rto_max;
 383                         period += timeout;
 384                 }
 385         }
 386         return period;
 387 }
 388
 389 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
 390 {
 391         u32 rate = READ_ONCE(tp->rate_delivered);
 392         u32 intv = READ_ONCE(tp->rate_interval_us);
 393         u64 rate64 = 0;
 394
 395         if (rate && intv) {
 396                 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
 397                 do_div(rate64, intv);
 398         }
 399         return rate64;
 400 }
 401
 402 /* Address-family independent initialization for a tcp_sock.
 403  *
 404  * NOTE: A lot of things set to zero explicitly by call to
 405  *       sk_alloc() so need not be done here.
 406  */
 407 void tcp_init_sock(struct sock *sk)
 408 {
 409         struct inet_connection_sock *icsk = inet_csk(sk);
 410         struct tcp_sock *tp = tcp_sk(sk);
 411
 412         tp->out_of_order_queue = RB_ROOT;
 413         sk->tcp_rtx_queue = RB_ROOT;
 414         tcp_init_xmit_timers(sk);
 415         INIT_LIST_HEAD(&tp->tsq_node);
 416         INIT_LIST_HEAD(&tp->tsorted_sent_queue);
 417
 418         icsk->icsk_rto = TCP_TIMEOUT_INIT;
 419         tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
 420         minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
 421
 422         /* So many TCP implementations out there (incorrectly) count the
 423          * initial SYN frame in their delayed-ACK and congestion control
 424          * algorithms that we must have the following bandaid to talk
 425          * efficiently to them.  -DaveM
 426          */
 427         tp->snd_cwnd = TCP_INIT_CWND;
 428
 429         /* There's a bubble in the pipe until at least the first ACK. */
 430         tp->app_limited = ~0U;
 431
 432         /* See draft-stevens-tcpca-spec-01 for discussion of the
 433          * initialization of these values.
 434          */
 435         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 436         tp->snd_cwnd_clamp = ~0;
 437         tp->mss_cache = TCP_MSS_DEFAULT;
 438
 439         tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
 440         tcp_assign_congestion_control(sk);
 441
 442         tp->tsoffset = 0;
 443         tp->rack.reo_wnd_steps = 1;
 444
 445         sk->sk_state = TCP_CLOSE;
 446
 447         sk->sk_write_space = sk_stream_write_space;
 448         sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
 449
 450         icsk->icsk_sync_mss = tcp_sync_mss;
 451
 452         sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
 453         sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
 454
 455         sk_sockets_allocated_inc(sk);
 456         sk->sk_route_forced_caps = NETIF_F_GSO;
 457 }
 458 EXPORT_SYMBOL(tcp_init_sock);
 459
 460 void tcp_init_transfer(struct sock *sk, int bpf_op)
 461 {
 462         struct inet_connection_sock *icsk = inet_csk(sk);
 463
 464         tcp_mtup_init(sk);
 465         icsk->icsk_af_ops->rebuild_header(sk);
 466         tcp_init_metrics(sk);
 467         tcp_call_bpf(sk, bpf_op, 0, NULL);
 468         tcp_init_congestion_control(sk);
 469         tcp_init_buffer_space(sk);
 470 }
 471
 472 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
 473 {
 474         struct sk_buff *skb = tcp_write_queue_tail(sk);
 475
 476         if (tsflags && skb) {
 477                 struct skb_shared_info *shinfo = skb_shinfo(skb);
 478                 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 479
 480                 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
 481                 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
 482                         tcb->txstamp_ack = 1;
 483                 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
 484                         shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
 485         }
 486 }
 487
 488 static inline bool tcp_stream_is_readable(const struct tcp_sock *tp,
 489                                           int target, struct sock *sk)
 490 {
 491         return (tp->rcv_nxt - tp->copied_seq >= target) ||
 492                 (sk->sk_prot->stream_memory_read ?
 493                 sk->sk_prot->stream_memory_read(sk) : false);
 494 }
 495
 496 /*
 497  *      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, sk_sleep(sk), 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                                 if (!sock_flag(sk, SOCK_DONE))
 821                                         ret = -ENOTCONN;
 822                                 break;
 823                         }
 824                         if (!timeo) {
 825                                 ret = -EAGAIN;
 826                                 break;
 827                         }
 828                         /* if __tcp_splice_read() got nothing while we have
 829                          * an skb in receive queue, we do not want to loop.
 830                          * This might happen with URG data.
 831                          */
 832                         if (!skb_queue_empty(&sk->sk_receive_queue))
 833                                 break;
 834                         sk_wait_data(sk, &timeo, NULL);
 835                         if (signal_pending(current)) {
 836                                 ret = sock_intr_errno(timeo);
 837                                 break;
 838                         }
 839                         continue;
 840                 }
 841                 tss.len -= ret;
 842                 spliced += ret;
 843
 844                 if (!timeo)
 845                         break;
 846                 release_sock(sk);
 847                 lock_sock(sk);
 848
 849                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 850                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 851                     signal_pending(current))
 852                         break;
 853         }
 854
 855         release_sock(sk);
 856
 857         if (spliced)
 858                 return spliced;
 859
 860         return ret;
 861 }
 862 EXPORT_SYMBOL(tcp_splice_read);
 863
 864 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
 865                                     bool force_schedule)
 866 {
 867         struct sk_buff *skb;
 868
 869         /* The TCP header must be at least 32-bit aligned.  */
 870         size = ALIGN(size, 4);
 871
 872         if (unlikely(tcp_under_memory_pressure(sk)))
 873                 sk_mem_reclaim_partial(sk);
 874
 875         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 876         if (likely(skb)) {
 877                 bool mem_scheduled;
 878
 879                 if (force_schedule) {
 880                         mem_scheduled = true;
 881                         sk_forced_mem_schedule(sk, skb->truesize);
 882                 } else {
 883                         mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
 884                 }
 885                 if (likely(mem_scheduled)) {
 886                         skb_reserve(skb, sk->sk_prot->max_header);
 887                         /*
 888                          * Make sure that we have exactly size bytes
 889                          * available to the caller, no more, no less.
 890                          */
 891                         skb->reserved_tailroom = skb->end - skb->tail - size;
 892                         INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
 893                         return skb;
 894                 }
 895                 __kfree_skb(skb);
 896         } else {
 897                 sk->sk_prot->enter_memory_pressure(sk);
 898                 sk_stream_moderate_sndbuf(sk);
 899         }
 900         return NULL;
 901 }
 902
 903 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 904                                        int large_allowed)
 905 {
 906         struct tcp_sock *tp = tcp_sk(sk);
 907         u32 new_size_goal, size_goal;
 908
 909         if (!large_allowed)
 910                 return mss_now;
 911
 912         /* Note : tcp_tso_autosize() will eventually split this later */
 913         new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
 914         new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
 915
 916         /* We try hard to avoid divides here */
 917         size_goal = tp->gso_segs * mss_now;
 918         if (unlikely(new_size_goal < size_goal ||
 919                      new_size_goal >= size_goal + mss_now)) {
 920                 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
 921                                      sk->sk_gso_max_segs);
 922                 size_goal = tp->gso_segs * mss_now;
 923         }
 924
 925         return max(size_goal, mss_now);
 926 }
 927
 928 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 929 {
 930         int mss_now;
 931
 932         mss_now = tcp_current_mss(sk);
 933         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 934
 935         return mss_now;
 936 }
 937
 938 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
 939                          size_t size, int flags)
 940 {
 941         struct tcp_sock *tp = tcp_sk(sk);
 942         int mss_now, size_goal;
 943         int err;
 944         ssize_t copied;
 945         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 946
 947         /* Wait for a connection to finish. One exception is TCP Fast Open
 948          * (passive side) where data is allowed to be sent before a connection
 949          * is fully established.
 950          */
 951         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
 952             !tcp_passive_fastopen(sk)) {
 953                 err = sk_stream_wait_connect(sk, &timeo);
 954                 if (err != 0)
 955                         goto out_err;
 956         }
 957
 958         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 959
 960         mss_now = tcp_send_mss(sk, &size_goal, flags);
 961         copied = 0;
 962
 963         err = -EPIPE;
 964         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 965                 goto out_err;
 966
 967         while (size > 0) {
 968                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 969                 int copy, i;
 970                 bool can_coalesce;
 971
 972                 if (!skb || (copy = size_goal - skb->len) <= 0 ||
 973                     !tcp_skb_can_collapse_to(skb)) {
 974 new_segment:
 975                         if (!sk_stream_memory_free(sk))
 976                                 goto wait_for_sndbuf;
 977
 978                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
 979                                         tcp_rtx_and_write_queues_empty(sk));
 980                         if (!skb)
 981                                 goto wait_for_memory;
 982
 983                         skb_entail(sk, skb);
 984                         copy = size_goal;
 985                 }
 986
 987                 if (copy > size)
 988                         copy = size;
 989
 990                 i = skb_shinfo(skb)->nr_frags;
 991                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 992                 if (!can_coalesce && i >= sysctl_max_skb_frags) {
 993                         tcp_mark_push(tp, skb);
 994                         goto new_segment;
 995                 }
 996                 if (!sk_wmem_schedule(sk, copy))
 997                         goto wait_for_memory;
 998
 999                 if (can_coalesce) {
1000                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1001                 } else {
1002                         get_page(page);
1003                         skb_fill_page_desc(skb, i, page, offset, copy);
1004                 }
1005
1006                 if (!(flags & MSG_NO_SHARED_FRAGS))
1007                         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1008
1009                 skb->len += copy;
1010                 skb->data_len += copy;
1011                 skb->truesize += copy;
1012                 sk->sk_wmem_queued += copy;
1013                 sk_mem_charge(sk, copy);
1014                 skb->ip_summed = CHECKSUM_PARTIAL;
1015                 tp->write_seq += copy;
1016                 TCP_SKB_CB(skb)->end_seq += copy;
1017                 tcp_skb_pcount_set(skb, 0);
1018
1019                 if (!copied)
1020                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1021
1022                 copied += copy;
1023                 offset += copy;
1024                 size -= copy;
1025                 if (!size)
1026                         goto out;
1027
1028                 if (skb->len < size_goal || (flags & MSG_OOB))
1029                         continue;
1030
1031                 if (forced_push(tp)) {
1032                         tcp_mark_push(tp, skb);
1033                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1034                 } else if (skb == tcp_send_head(sk))
1035                         tcp_push_one(sk, mss_now);
1036                 continue;
1037
1038 wait_for_sndbuf:
1039                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1040 wait_for_memory:
1041                 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1042                          TCP_NAGLE_PUSH, size_goal);
1043
1044                 err = sk_stream_wait_memory(sk, &timeo);
1045                 if (err != 0)
1046                         goto do_error;
1047
1048                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1049         }
1050
1051 out:
1052         if (copied) {
1053                 tcp_tx_timestamp(sk, sk->sk_tsflags);
1054                 if (!(flags & MSG_SENDPAGE_NOTLAST))
1055                         tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1056         }
1057         return copied;
1058
1059 do_error:
1060         if (copied)
1061                 goto out;
1062 out_err:
1063         /* make sure we wake any epoll edge trigger waiter */
1064         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1065                      err == -EAGAIN)) {
1066                 sk->sk_write_space(sk);
1067                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1068         }
1069         return sk_stream_error(sk, flags, err);
1070 }
1071 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1072
1073 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1074                         size_t size, int flags)
1075 {
1076         if (!(sk->sk_route_caps & NETIF_F_SG))
1077                 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1078
1079         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1080
1081         return do_tcp_sendpages(sk, page, offset, size, flags);
1082 }
1083 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1084
1085 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1086                  size_t size, int flags)
1087 {
1088         int ret;
1089
1090         lock_sock(sk);
1091         ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1092         release_sock(sk);
1093
1094         return ret;
1095 }
1096 EXPORT_SYMBOL(tcp_sendpage);
1097
1098 /* Do not bother using a page frag for very small frames.
1099  * But use this heuristic only for the first skb in write queue.
1100  *
1101  * Having no payload in skb->head allows better SACK shifting
1102  * in tcp_shift_skb_data(), reducing sack/rack overhead, because
1103  * write queue has less skbs.
1104  * Each skb can hold up to MAX_SKB_FRAGS * 32Kbytes, or ~0.5 MB.
1105  * This also speeds up tso_fragment(), since it wont fallback
1106  * to tcp_fragment().
1107  */
1108 static int linear_payload_sz(bool first_skb)
1109 {
1110         if (first_skb)
1111                 return SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1112         return 0;
1113 }
1114
1115 static int select_size(bool first_skb, bool zc)
1116 {
1117         if (zc)
1118                 return 0;
1119         return linear_payload_sz(first_skb);
1120 }
1121
1122 void tcp_free_fastopen_req(struct tcp_sock *tp)
1123 {
1124         if (tp->fastopen_req) {
1125                 kfree(tp->fastopen_req);
1126                 tp->fastopen_req = NULL;
1127         }
1128 }
1129
1130 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1131                                 int *copied, size_t size)
1132 {
1133         struct tcp_sock *tp = tcp_sk(sk);
1134         struct inet_sock *inet = inet_sk(sk);
1135         struct sockaddr *uaddr = msg->msg_name;
1136         int err, flags;
1137
1138         if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1139             (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1140              uaddr->sa_family == AF_UNSPEC))
1141                 return -EOPNOTSUPP;
1142         if (tp->fastopen_req)
1143                 return -EALREADY; /* Another Fast Open is in progress */
1144
1145         tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1146                                    sk->sk_allocation);
1147         if (unlikely(!tp->fastopen_req))
1148                 return -ENOBUFS;
1149         tp->fastopen_req->data = msg;
1150         tp->fastopen_req->size = size;
1151
1152         if (inet->defer_connect) {
1153                 err = tcp_connect(sk);
1154                 /* Same failure procedure as in tcp_v4/6_connect */
1155                 if (err) {
1156                         tcp_set_state(sk, TCP_CLOSE);
1157                         inet->inet_dport = 0;
1158                         sk->sk_route_caps = 0;
1159                 }
1160         }
1161         flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1162         err = __inet_stream_connect(sk->sk_socket, uaddr,
1163                                     msg->msg_namelen, flags, 1);
1164         /* fastopen_req could already be freed in __inet_stream_connect
1165          * if the connection times out or gets rst
1166          */
1167         if (tp->fastopen_req) {
1168                 *copied = tp->fastopen_req->copied;
1169                 tcp_free_fastopen_req(tp);
1170                 inet->defer_connect = 0;
1171         }
1172         return err;
1173 }
1174
1175 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1176 {
1177         struct tcp_sock *tp = tcp_sk(sk);
1178         struct ubuf_info *uarg = NULL;
1179         struct sk_buff *skb;
1180         struct sockcm_cookie sockc;
1181         int flags, err, copied = 0;
1182         int mss_now = 0, size_goal, copied_syn = 0;
1183         bool process_backlog = false;
1184         bool zc = false;
1185         long timeo;
1186
1187         flags = msg->msg_flags;
1188
1189         if (flags & MSG_ZEROCOPY && size) {
1190                 if (sk->sk_state != TCP_ESTABLISHED) {
1191                         err = -EINVAL;
1192                         goto out_err;
1193                 }
1194
1195                 skb = tcp_write_queue_tail(sk);
1196                 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
1197                 if (!uarg) {
1198                         err = -ENOBUFS;
1199                         goto out_err;
1200                 }
1201
1202                 zc = sk->sk_route_caps & NETIF_F_SG;
1203                 if (!zc)
1204                         uarg->zerocopy = 0;
1205         }
1206
1207         if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1208             !tp->repair) {
1209                 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1210                 if (err == -EINPROGRESS && copied_syn > 0)
1211                         goto out;
1212                 else if (err)
1213                         goto out_err;
1214         }
1215
1216         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1217
1218         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1219
1220         /* Wait for a connection to finish. One exception is TCP Fast Open
1221          * (passive side) where data is allowed to be sent before a connection
1222          * is fully established.
1223          */
1224         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1225             !tcp_passive_fastopen(sk)) {
1226                 err = sk_stream_wait_connect(sk, &timeo);
1227                 if (err != 0)
1228                         goto do_error;
1229         }
1230
1231         if (unlikely(tp->repair)) {
1232                 if (tp->repair_queue == TCP_RECV_QUEUE) {
1233                         copied = tcp_send_rcvq(sk, msg, size);
1234                         goto out_nopush;
1235                 }
1236
1237                 err = -EINVAL;
1238                 if (tp->repair_queue == TCP_NO_QUEUE)
1239                         goto out_err;
1240
1241                 /* 'common' sending to sendq */
1242         }
1243
1244         sockc.tsflags = sk->sk_tsflags;
1245         if (msg->msg_controllen) {
1246                 err = sock_cmsg_send(sk, msg, &sockc);
1247                 if (unlikely(err)) {
1248                         err = -EINVAL;
1249                         goto out_err;
1250                 }
1251         }
1252
1253         /* This should be in poll */
1254         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1255
1256         /* Ok commence sending. */
1257         copied = 0;
1258
1259 restart:
1260         mss_now = tcp_send_mss(sk, &size_goal, flags);
1261
1262         err = -EPIPE;
1263         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1264                 goto do_error;
1265
1266         while (msg_data_left(msg)) {
1267                 int copy = 0;
1268
1269                 skb = tcp_write_queue_tail(sk);
1270                 if (skb)
1271                         copy = size_goal - skb->len;
1272
1273                 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1274                         bool first_skb;
1275                         int linear;
1276
1277 new_segment:
1278                         /* Allocate new segment. If the interface is SG,
1279                          * allocate skb fitting to single page.
1280                          */
1281                         if (!sk_stream_memory_free(sk))
1282                                 goto wait_for_sndbuf;
1283
1284                         if (process_backlog && sk_flush_backlog(sk)) {
1285                                 process_backlog = false;
1286                                 goto restart;
1287                         }
1288                         first_skb = tcp_rtx_and_write_queues_empty(sk);
1289                         linear = select_size(first_skb, zc);
1290                         skb = sk_stream_alloc_skb(sk, linear, sk->sk_allocation,
1291                                                   first_skb);
1292                         if (!skb)
1293                                 goto wait_for_memory;
1294
1295                         process_backlog = true;
1296                         skb->ip_summed = CHECKSUM_PARTIAL;
1297
1298                         skb_entail(sk, skb);
1299                         copy = size_goal;
1300
1301                         /* All packets are restored as if they have
1302                          * already been sent. skb_mstamp isn't set to
1303                          * avoid wrong rtt estimation.
1304                          */
1305                         if (tp->repair)
1306                                 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1307                 }
1308
1309                 /* Try to append data to the end of skb. */
1310                 if (copy > msg_data_left(msg))
1311                         copy = msg_data_left(msg);
1312
1313                 /* Where to copy to? */
1314                 if (skb_availroom(skb) > 0 && !zc) {
1315                         /* We have some space in skb head. Superb! */
1316                         copy = min_t(int, copy, skb_availroom(skb));
1317                         err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1318                         if (err)
1319                                 goto do_fault;
1320                 } else if (!zc) {
1321                         bool merge = true;
1322                         int i = skb_shinfo(skb)->nr_frags;
1323                         struct page_frag *pfrag = sk_page_frag(sk);
1324
1325                         if (!sk_page_frag_refill(sk, pfrag))
1326                                 goto wait_for_memory;
1327
1328                         if (!skb_can_coalesce(skb, i, pfrag->page,
1329                                               pfrag->offset)) {
1330                                 if (i >= sysctl_max_skb_frags) {
1331                                         tcp_mark_push(tp, skb);
1332                                         goto new_segment;
1333                                 }
1334                                 merge = false;
1335                         }
1336
1337                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1338
1339                         if (!sk_wmem_schedule(sk, copy))
1340                                 goto wait_for_memory;
1341
1342                         err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1343                                                        pfrag->page,
1344                                                        pfrag->offset,
1345                                                        copy);
1346                         if (err)
1347                                 goto do_error;
1348
1349                         /* Update the skb. */
1350                         if (merge) {
1351                                 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1352                         } else {
1353                                 skb_fill_page_desc(skb, i, pfrag->page,
1354                                                    pfrag->offset, copy);
1355                                 page_ref_inc(pfrag->page);
1356                         }
1357                         pfrag->offset += copy;
1358                 } else {
1359                         err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1360                         if (err == -EMSGSIZE || err == -EEXIST) {
1361                                 tcp_mark_push(tp, skb);
1362                                 goto new_segment;
1363                         }
1364                         if (err < 0)
1365                                 goto do_error;
1366                         copy = err;
1367                 }
1368
1369                 if (!copied)
1370                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1371
1372                 tp->write_seq += copy;
1373                 TCP_SKB_CB(skb)->end_seq += copy;
1374                 tcp_skb_pcount_set(skb, 0);
1375
1376                 copied += copy;
1377                 if (!msg_data_left(msg)) {
1378                         if (unlikely(flags & MSG_EOR))
1379                                 TCP_SKB_CB(skb)->eor = 1;
1380                         goto out;
1381                 }
1382
1383                 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1384                         continue;
1385
1386                 if (forced_push(tp)) {
1387                         tcp_mark_push(tp, skb);
1388                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1389                 } else if (skb == tcp_send_head(sk))
1390                         tcp_push_one(sk, mss_now);
1391                 continue;
1392
1393 wait_for_sndbuf:
1394                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1395 wait_for_memory:
1396                 if (copied)
1397                         tcp_push(sk, flags & ~MSG_MORE, mss_now,
1398                                  TCP_NAGLE_PUSH, size_goal);
1399
1400                 err = sk_stream_wait_memory(sk, &timeo);
1401                 if (err != 0)
1402                         goto do_error;
1403
1404                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1405         }
1406
1407 out:
1408         if (copied) {
1409                 tcp_tx_timestamp(sk, sockc.tsflags);
1410                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1411         }
1412 out_nopush:
1413         sock_zerocopy_put(uarg);
1414         return copied + copied_syn;
1415
1416 do_fault:
1417         if (!skb->len) {
1418                 tcp_unlink_write_queue(skb, sk);
1419                 /* It is the one place in all of TCP, except connection
1420                  * reset, where we can be unlinking the send_head.
1421                  */
1422                 tcp_check_send_head(sk, skb);
1423                 sk_wmem_free_skb(sk, skb);
1424         }
1425
1426 do_error:
1427         if (copied + copied_syn)
1428                 goto out;
1429 out_err:
1430         sock_zerocopy_put_abort(uarg);
1431         err = sk_stream_error(sk, flags, err);
1432         /* make sure we wake any epoll edge trigger waiter */
1433         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1434                      err == -EAGAIN)) {
1435                 sk->sk_write_space(sk);
1436                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1437         }
1438         return err;
1439 }
1440 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1441
1442 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1443 {
1444         int ret;
1445
1446         lock_sock(sk);
1447         ret = tcp_sendmsg_locked(sk, msg, size);
1448         release_sock(sk);
1449
1450         return ret;
1451 }
1452 EXPORT_SYMBOL(tcp_sendmsg);
1453
1454 /*
1455  *      Handle reading urgent data. BSD has very simple semantics for
1456  *      this, no blocking and very strange errors 8)
1457  */
1458
1459 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1460 {
1461         struct tcp_sock *tp = tcp_sk(sk);
1462
1463         /* No URG data to read. */
1464         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1465             tp->urg_data == TCP_URG_READ)
1466                 return -EINVAL; /* Yes this is right ! */
1467
1468         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1469                 return -ENOTCONN;
1470
1471         if (tp->urg_data & TCP_URG_VALID) {
1472                 int err = 0;
1473                 char c = tp->urg_data;
1474
1475                 if (!(flags & MSG_PEEK))
1476                         tp->urg_data = TCP_URG_READ;
1477
1478                 /* Read urgent data. */
1479                 msg->msg_flags |= MSG_OOB;
1480
1481                 if (len > 0) {
1482                         if (!(flags & MSG_TRUNC))
1483                                 err = memcpy_to_msg(msg, &c, 1);
1484                         len = 1;
1485                 } else
1486                         msg->msg_flags |= MSG_TRUNC;
1487
1488                 return err ? -EFAULT : len;
1489         }
1490
1491         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1492                 return 0;
1493
1494         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1495          * the available implementations agree in this case:
1496          * this call should never block, independent of the
1497          * blocking state of the socket.
1498          * Mike <pall@rz.uni-karlsruhe.de>
1499          */
1500         return -EAGAIN;
1501 }
1502
1503 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1504 {
1505         struct sk_buff *skb;
1506         int copied = 0, err = 0;
1507
1508         /* XXX -- need to support SO_PEEK_OFF */
1509
1510         skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1511                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1512                 if (err)
1513                         return err;
1514                 copied += skb->len;
1515         }
1516
1517         skb_queue_walk(&sk->sk_write_queue, skb) {
1518                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1519                 if (err)
1520                         break;
1521
1522                 copied += skb->len;
1523         }
1524
1525         return err ?: copied;
1526 }
1527
1528 /* Clean up the receive buffer for full frames taken by the user,
1529  * then send an ACK if necessary.  COPIED is the number of bytes
1530  * tcp_recvmsg has given to the user so far, it speeds up the
1531  * calculation of whether or not we must ACK for the sake of
1532  * a window update.
1533  */
1534 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1535 {
1536         struct tcp_sock *tp = tcp_sk(sk);
1537         bool time_to_ack = false;
1538
1539         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1540
1541         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1542              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1543              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1544
1545         if (inet_csk_ack_scheduled(sk)) {
1546                 const struct inet_connection_sock *icsk = inet_csk(sk);
1547                    /* Delayed ACKs frequently hit locked sockets during bulk
1548                     * receive. */
1549                 if (icsk->icsk_ack.blocked ||
1550                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1551                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1552                     /*
1553                      * If this read emptied read buffer, we send ACK, if
1554                      * connection is not bidirectional, user drained
1555                      * receive buffer and there was a small segment
1556                      * in queue.
1557                      */
1558                     (copied > 0 &&
1559                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1560                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1561                        !icsk->icsk_ack.pingpong)) &&
1562                       !atomic_read(&sk->sk_rmem_alloc)))
1563                         time_to_ack = true;
1564         }
1565
1566         /* We send an ACK if we can now advertise a non-zero window
1567          * which has been raised "significantly".
1568          *
1569          * Even if window raised up to infinity, do not send window open ACK
1570          * in states, where we will not receive more. It is useless.
1571          */
1572         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1573                 __u32 rcv_window_now = tcp_receive_window(tp);
1574
1575                 /* Optimize, __tcp_select_window() is not cheap. */
1576                 if (2*rcv_window_now <= tp->window_clamp) {
1577                         __u32 new_window = __tcp_select_window(sk);
1578
1579                         /* Send ACK now, if this read freed lots of space
1580                          * in our buffer. Certainly, new_window is new window.
1581                          * We can advertise it now, if it is not less than current one.
1582                          * "Lots" means "at least twice" here.
1583                          */
1584                         if (new_window && new_window >= 2 * rcv_window_now)
1585                                 time_to_ack = true;
1586                 }
1587         }
1588         if (time_to_ack)
1589                 tcp_send_ack(sk);
1590 }
1591
1592 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1593 {
1594         struct sk_buff *skb;
1595         u32 offset;
1596
1597         while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1598                 offset = seq - TCP_SKB_CB(skb)->seq;
1599                 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1600                         pr_err_once("%s: found a SYN, please report !\n", __func__);
1601                         offset--;
1602                 }
1603                 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1604                         *off = offset;
1605                         return skb;
1606                 }
1607                 /* This looks weird, but this can happen if TCP collapsing
1608                  * splitted a fat GRO packet, while we released socket lock
1609                  * in skb_splice_bits()
1610                  */
1611                 sk_eat_skb(sk, skb);
1612         }
1613         return NULL;
1614 }
1615
1616 /*
1617  * This routine provides an alternative to tcp_recvmsg() for routines
1618  * that would like to handle copying from skbuffs directly in 'sendfile'
1619  * fashion.
1620  * Note:
1621  *      - It is assumed that the socket was locked by the caller.
1622  *      - The routine does not block.
1623  *      - At present, there is no support for reading OOB data
1624  *        or for 'peeking' the socket using this routine
1625  *        (although both would be easy to implement).
1626  */
1627 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1628                   sk_read_actor_t recv_actor)
1629 {
1630         struct sk_buff *skb;
1631         struct tcp_sock *tp = tcp_sk(sk);
1632         u32 seq = tp->copied_seq;
1633         u32 offset;
1634         int copied = 0;
1635
1636         if (sk->sk_state == TCP_LISTEN)
1637                 return -ENOTCONN;
1638         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1639                 if (offset < skb->len) {
1640                         int used;
1641                         size_t len;
1642
1643                         len = skb->len - offset;
1644                         /* Stop reading if we hit a patch of urgent data */
1645                         if (tp->urg_data) {
1646                                 u32 urg_offset = tp->urg_seq - seq;
1647                                 if (urg_offset < len)
1648                                         len = urg_offset;
1649                                 if (!len)
1650                                         break;
1651                         }
1652                         used = recv_actor(desc, skb, offset, len);
1653                         if (used <= 0) {
1654                                 if (!copied)
1655                                         copied = used;
1656                                 break;
1657                         } else if (used <= len) {
1658                                 seq += used;
1659                                 copied += used;
1660                                 offset += used;
1661                         }
1662                         /* If recv_actor drops the lock (e.g. TCP splice
1663                          * receive) the skb pointer might be invalid when
1664                          * getting here: tcp_collapse might have deleted it
1665                          * while aggregating skbs from the socket queue.
1666                          */
1667                         skb = tcp_recv_skb(sk, seq - 1, &offset);
1668                         if (!skb)
1669                                 break;
1670                         /* TCP coalescing might have appended data to the skb.
1671                          * Try to splice more frags
1672                          */
1673                         if (offset + 1 != skb->len)
1674                                 continue;
1675                 }
1676                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1677                         sk_eat_skb(sk, skb);
1678                         ++seq;
1679                         break;
1680                 }
1681                 sk_eat_skb(sk, skb);
1682                 if (!desc->count)
1683                         break;
1684                 tp->copied_seq = seq;
1685         }
1686         tp->copied_seq = seq;
1687
1688         tcp_rcv_space_adjust(sk);
1689
1690         /* Clean up data we have read: This will do ACK frames. */
1691         if (copied > 0) {
1692                 tcp_recv_skb(sk, seq, &offset);
1693                 tcp_cleanup_rbuf(sk, copied);
1694         }
1695         return copied;
1696 }
1697 EXPORT_SYMBOL(tcp_read_sock);
1698
1699 int tcp_peek_len(struct socket *sock)
1700 {
1701         return tcp_inq(sock->sk);
1702 }
1703 EXPORT_SYMBOL(tcp_peek_len);
1704
1705 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1706 int tcp_set_rcvlowat(struct sock *sk, int val)
1707 {
1708         int cap;
1709
1710         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1711                 cap = sk->sk_rcvbuf >> 1;
1712         else
1713                 cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1;
1714         val = min(val, cap);
1715         sk->sk_rcvlowat = val ? : 1;
1716
1717         /* Check if we need to signal EPOLLIN right now */
1718         tcp_data_ready(sk);
1719
1720         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1721                 return 0;
1722
1723         val <<= 1;
1724         if (val > sk->sk_rcvbuf) {
1725                 sk->sk_rcvbuf = val;
1726                 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1727         }
1728         return 0;
1729 }
1730 EXPORT_SYMBOL(tcp_set_rcvlowat);
1731
1732 #ifdef CONFIG_MMU
1733 static const struct vm_operations_struct tcp_vm_ops = {
1734 };
1735
1736 int tcp_mmap(struct file *file, struct socket *sock,
1737              struct vm_area_struct *vma)
1738 {
1739         if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1740                 return -EPERM;
1741         vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1742
1743         /* Instruct vm_insert_page() to not down_read(mmap_sem) */
1744         vma->vm_flags |= VM_MIXEDMAP;
1745
1746         vma->vm_ops = &tcp_vm_ops;
1747         return 0;
1748 }
1749 EXPORT_SYMBOL(tcp_mmap);
1750
1751 static int tcp_zerocopy_receive(struct sock *sk,
1752                                 struct tcp_zerocopy_receive *zc)
1753 {
1754         unsigned long address = (unsigned long)zc->address;
1755         const skb_frag_t *frags = NULL;
1756         u32 length = 0, seq, offset;
1757         struct vm_area_struct *vma;
1758         struct sk_buff *skb = NULL;
1759         struct tcp_sock *tp;
1760         int ret;
1761
1762         if (address & (PAGE_SIZE - 1) || address != zc->address)
1763                 return -EINVAL;
1764
1765         if (sk->sk_state == TCP_LISTEN)
1766                 return -ENOTCONN;
1767
1768         sock_rps_record_flow(sk);
1769
1770         down_read(&current->mm->mmap_sem);
1771
1772         ret = -EINVAL;
1773         vma = find_vma(current->mm, address);
1774         if (!vma || vma->vm_start > address || vma->vm_ops != &tcp_vm_ops)
1775                 goto out;
1776         zc->length = min_t(unsigned long, zc->length, vma->vm_end - address);
1777
1778         tp = tcp_sk(sk);
1779         seq = tp->copied_seq;
1780         zc->length = min_t(u32, zc->length, tcp_inq(sk));
1781         zc->length &= ~(PAGE_SIZE - 1);
1782
1783         zap_page_range(vma, address, zc->length);
1784
1785         zc->recv_skip_hint = 0;
1786         ret = 0;
1787         while (length + PAGE_SIZE <= zc->length) {
1788                 if (zc->recv_skip_hint < PAGE_SIZE) {
1789                         if (skb) {
1790                                 skb = skb->next;
1791                                 offset = seq - TCP_SKB_CB(skb)->seq;
1792                         } else {
1793                                 skb = tcp_recv_skb(sk, seq, &offset);
1794                         }
1795
1796                         zc->recv_skip_hint = skb->len - offset;
1797                         offset -= skb_headlen(skb);
1798                         if ((int)offset < 0 || skb_has_frag_list(skb))
1799                                 break;
1800                         frags = skb_shinfo(skb)->frags;
1801                         while (offset) {
1802                                 if (frags->size > offset)
1803                                         goto out;
1804                                 offset -= frags->size;
1805                                 frags++;
1806                         }
1807                 }
1808                 if (frags->size != PAGE_SIZE || frags->page_offset)
1809                         break;
1810                 ret = vm_insert_page(vma, address + length,
1811                                      skb_frag_page(frags));
1812                 if (ret)
1813                         break;
1814                 length += PAGE_SIZE;
1815                 seq += PAGE_SIZE;
1816                 zc->recv_skip_hint -= PAGE_SIZE;
1817                 frags++;
1818         }
1819 out:
1820         up_read(&current->mm->mmap_sem);
1821         if (length) {
1822                 tp->copied_seq = seq;
1823                 tcp_rcv_space_adjust(sk);
1824
1825                 /* Clean up data we have read: This will do ACK frames. */
1826                 tcp_recv_skb(sk, seq, &offset);
1827                 tcp_cleanup_rbuf(sk, length);
1828                 ret = 0;
1829                 if (length == zc->length)
1830                         zc->recv_skip_hint = 0;
1831         } else {
1832                 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
1833                         ret = -EIO;
1834         }
1835         zc->length = length;
1836         return ret;
1837 }
1838 #endif
1839
1840 static void tcp_update_recv_tstamps(struct sk_buff *skb,
1841                                     struct scm_timestamping *tss)
1842 {
1843         if (skb->tstamp)
1844                 tss->ts[0] = ktime_to_timespec(skb->tstamp);
1845         else
1846                 tss->ts[0] = (struct timespec) {0};
1847
1848         if (skb_hwtstamps(skb)->hwtstamp)
1849                 tss->ts[2] = ktime_to_timespec(skb_hwtstamps(skb)->hwtstamp);
1850         else
1851                 tss->ts[2] = (struct timespec) {0};
1852 }
1853
1854 /* Similar to __sock_recv_timestamp, but does not require an skb */
1855 static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
1856                                struct scm_timestamping *tss)
1857 {
1858         struct timeval tv;
1859         bool has_timestamping = false;
1860
1861         if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
1862                 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
1863                         if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
1864                                 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPNS,
1865                                          sizeof(tss->ts[0]), &tss->ts[0]);
1866                         } else {
1867                                 tv.tv_sec = tss->ts[0].tv_sec;
1868                                 tv.tv_usec = tss->ts[0].tv_nsec / 1000;
1869
1870                                 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMP,
1871                                          sizeof(tv), &tv);
1872                         }
1873                 }
1874
1875                 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
1876                         has_timestamping = true;
1877                 else
1878                         tss->ts[0] = (struct timespec) {0};
1879         }
1880
1881         if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
1882                 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
1883                         has_timestamping = true;
1884                 else
1885                         tss->ts[2] = (struct timespec) {0};
1886         }
1887
1888         if (has_timestamping) {
1889                 tss->ts[1] = (struct timespec) {0};
1890                 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPING,
1891                          sizeof(*tss), tss);
1892         }
1893 }
1894
1895 static int tcp_inq_hint(struct sock *sk)
1896 {
1897         const struct tcp_sock *tp = tcp_sk(sk);
1898         u32 copied_seq = READ_ONCE(tp->copied_seq);
1899         u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
1900         int inq;
1901
1902         inq = rcv_nxt - copied_seq;
1903         if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
1904                 lock_sock(sk);
1905                 inq = tp->rcv_nxt - tp->copied_seq;
1906                 release_sock(sk);
1907         }
1908         return inq;
1909 }
1910
1911 /*
1912  *      This routine copies from a sock struct into the user buffer.
1913  *
1914  *      Technical note: in 2.3 we work on _locked_ socket, so that
1915  *      tricks with *seq access order and skb->users are not required.
1916  *      Probably, code can be easily improved even more.
1917  */
1918
1919 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1920                 int flags, int *addr_len)
1921 {
1922         struct tcp_sock *tp = tcp_sk(sk);
1923         int copied = 0;
1924         u32 peek_seq;
1925         u32 *seq;
1926         unsigned long used;
1927         int err, inq;
1928         int target;             /* Read at least this many bytes */
1929         long timeo;
1930         struct sk_buff *skb, *last;
1931         u32 urg_hole = 0;
1932         struct scm_timestamping tss;
1933         bool has_tss = false;
1934         bool has_cmsg;
1935
1936         if (unlikely(flags & MSG_ERRQUEUE))
1937                 return inet_recv_error(sk, msg, len, addr_len);
1938
1939         if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1940             (sk->sk_state == TCP_ESTABLISHED))
1941                 sk_busy_loop(sk, nonblock);
1942
1943         lock_sock(sk);
1944
1945         err = -ENOTCONN;
1946         if (sk->sk_state == TCP_LISTEN)
1947                 goto out;
1948
1949         has_cmsg = tp->recvmsg_inq;
1950         timeo = sock_rcvtimeo(sk, nonblock);
1951
1952         /* Urgent data needs to be handled specially. */
1953         if (flags & MSG_OOB)
1954                 goto recv_urg;
1955
1956         if (unlikely(tp->repair)) {
1957                 err = -EPERM;
1958                 if (!(flags & MSG_PEEK))
1959                         goto out;
1960
1961                 if (tp->repair_queue == TCP_SEND_QUEUE)
1962                         goto recv_sndq;
1963
1964                 err = -EINVAL;
1965                 if (tp->repair_queue == TCP_NO_QUEUE)
1966                         goto out;
1967
1968                 /* 'common' recv queue MSG_PEEK-ing */
1969         }
1970
1971         seq = &tp->copied_seq;
1972         if (flags & MSG_PEEK) {
1973                 peek_seq = tp->copied_seq;
1974                 seq = &peek_seq;
1975         }
1976
1977         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1978
1979         do {
1980                 u32 offset;
1981
1982                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1983                 if (tp->urg_data && tp->urg_seq == *seq) {
1984                         if (copied)
1985                                 break;
1986                         if (signal_pending(current)) {
1987                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1988                                 break;
1989                         }
1990                 }
1991
1992                 /* Next get a buffer. */
1993
1994                 last = skb_peek_tail(&sk->sk_receive_queue);
1995                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1996                         last = skb;
1997                         /* Now that we have two receive queues this
1998                          * shouldn't happen.
1999                          */
2000                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2001                                  "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2002                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2003                                  flags))
2004                                 break;
2005
2006                         offset = *seq - TCP_SKB_CB(skb)->seq;
2007                         if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2008                                 pr_err_once("%s: found a SYN, please report !\n", __func__);
2009                                 offset--;
2010                         }
2011                         if (offset < skb->len)
2012                                 goto found_ok_skb;
2013                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2014                                 goto found_fin_ok;
2015                         WARN(!(flags & MSG_PEEK),
2016                              "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2017                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2018                 }
2019
2020                 /* Well, if we have backlog, try to process it now yet. */
2021
2022                 if (copied >= target && !sk->sk_backlog.tail)
2023                         break;
2024
2025                 if (copied) {
2026                         if (sk->sk_err ||
2027                             sk->sk_state == TCP_CLOSE ||
2028                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
2029                             !timeo ||
2030                             signal_pending(current))
2031                                 break;
2032                 } else {
2033                         if (sock_flag(sk, SOCK_DONE))
2034                                 break;
2035
2036                         if (sk->sk_err) {
2037                                 copied = sock_error(sk);
2038                                 break;
2039                         }
2040
2041                         if (sk->sk_shutdown & RCV_SHUTDOWN)
2042                                 break;
2043
2044                         if (sk->sk_state == TCP_CLOSE) {
2045                                 if (!sock_flag(sk, SOCK_DONE)) {
2046                                         /* This occurs when user tries to read
2047                                          * from never connected socket.
2048                                          */
2049                                         copied = -ENOTCONN;
2050                                         break;
2051                                 }
2052                                 break;
2053                         }
2054
2055                         if (!timeo) {
2056                                 copied = -EAGAIN;
2057                                 break;
2058                         }
2059
2060                         if (signal_pending(current)) {
2061                                 copied = sock_intr_errno(timeo);
2062                                 break;
2063                         }
2064                 }
2065
2066                 tcp_cleanup_rbuf(sk, copied);
2067
2068                 if (copied >= target) {
2069                         /* Do not sleep, just process backlog. */
2070                         release_sock(sk);
2071                         lock_sock(sk);
2072                 } else {
2073                         sk_wait_data(sk, &timeo, last);
2074                 }
2075
2076                 if ((flags & MSG_PEEK) &&
2077                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
2078                         net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2079                                             current->comm,
2080                                             task_pid_nr(current));
2081                         peek_seq = tp->copied_seq;
2082                 }
2083                 continue;
2084
2085         found_ok_skb:
2086                 /* Ok so how much can we use? */
2087                 used = skb->len - offset;
2088                 if (len < used)
2089                         used = len;
2090
2091                 /* Do we have urgent data here? */
2092                 if (tp->urg_data) {
2093                         u32 urg_offset = tp->urg_seq - *seq;
2094                         if (urg_offset < used) {
2095                                 if (!urg_offset) {
2096                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
2097                                                 ++*seq;
2098                                                 urg_hole++;
2099                                                 offset++;
2100                                                 used--;
2101                                                 if (!used)
2102                                                         goto skip_copy;
2103                                         }
2104                                 } else
2105                                         used = urg_offset;
2106                         }
2107                 }
2108
2109                 if (!(flags & MSG_TRUNC)) {
2110                         err = skb_copy_datagram_msg(skb, offset, msg, used);
2111                         if (err) {
2112                                 /* Exception. Bailout! */
2113                                 if (!copied)
2114                                         copied = -EFAULT;
2115                                 break;
2116                         }
2117                 }
2118
2119                 *seq += used;
2120                 copied += used;
2121                 len -= used;
2122
2123                 tcp_rcv_space_adjust(sk);
2124
2125 skip_copy:
2126                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
2127                         tp->urg_data = 0;
2128                         tcp_fast_path_check(sk);
2129                 }
2130                 if (used + offset < skb->len)
2131                         continue;
2132
2133                 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2134                         tcp_update_recv_tstamps(skb, &tss);
2135                         has_tss = true;
2136                         has_cmsg = true;
2137                 }
2138                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2139                         goto found_fin_ok;
2140                 if (!(flags & MSG_PEEK))
2141                         sk_eat_skb(sk, skb);
2142                 continue;
2143
2144         found_fin_ok:
2145                 /* Process the FIN. */
2146                 ++*seq;
2147                 if (!(flags & MSG_PEEK))
2148                         sk_eat_skb(sk, skb);
2149                 break;
2150         } while (len > 0);
2151
2152         /* According to UNIX98, msg_name/msg_namelen are ignored
2153          * on connected socket. I was just happy when found this 8) --ANK
2154          */
2155
2156         /* Clean up data we have read: This will do ACK frames. */
2157         tcp_cleanup_rbuf(sk, copied);
2158
2159         release_sock(sk);
2160
2161         if (has_cmsg) {
2162                 if (has_tss)
2163                         tcp_recv_timestamp(msg, sk, &tss);
2164                 if (tp->recvmsg_inq) {
2165                         inq = tcp_inq_hint(sk);
2166                         put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2167                 }
2168         }
2169
2170         return copied;
2171
2172 out:
2173         release_sock(sk);
2174         return err;
2175
2176 recv_urg:
2177         err = tcp_recv_urg(sk, msg, len, flags);
2178         goto out;
2179
2180 recv_sndq:
2181         err = tcp_peek_sndq(sk, msg, len);
2182         goto out;
2183 }
2184 EXPORT_SYMBOL(tcp_recvmsg);
2185
2186 void tcp_set_state(struct sock *sk, int state)
2187 {
2188         int oldstate = sk->sk_state;
2189
2190         /* We defined a new enum for TCP states that are exported in BPF
2191          * so as not force the internal TCP states to be frozen. The
2192          * following checks will detect if an internal state value ever
2193          * differs from the BPF value. If this ever happens, then we will
2194          * need to remap the internal value to the BPF value before calling
2195          * tcp_call_bpf_2arg.
2196          */
2197         BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2198         BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2199         BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2200         BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2201         BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2202         BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2203         BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2204         BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2205         BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2206         BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2207         BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2208         BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2209         BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2210
2211         if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2212                 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2213
2214         switch (state) {
2215         case TCP_ESTABLISHED:
2216                 if (oldstate != TCP_ESTABLISHED)
2217                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2218                 break;
2219
2220         case TCP_CLOSE:
2221                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2222                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2223
2224                 sk->sk_prot->unhash(sk);
2225                 if (inet_csk(sk)->icsk_bind_hash &&
2226                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2227                         inet_put_port(sk);
2228                 /* fall through */
2229         default:
2230                 if (oldstate == TCP_ESTABLISHED)
2231                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2232         }
2233
2234         /* Change state AFTER socket is unhashed to avoid closed
2235          * socket sitting in hash tables.
2236          */
2237         inet_sk_state_store(sk, state);
2238
2239 #ifdef STATE_TRACE
2240         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
2241 #endif
2242 }
2243 EXPORT_SYMBOL_GPL(tcp_set_state);
2244
2245 /*
2246  *      State processing on a close. This implements the state shift for
2247  *      sending our FIN frame. Note that we only send a FIN for some
2248  *      states. A shutdown() may have already sent the FIN, or we may be
2249  *      closed.
2250  */
2251
2252 static const unsigned char new_state[16] = {
2253   /* current state:        new state:      action:      */
2254   [0 /* (Invalid) */]   = TCP_CLOSE,
2255   [TCP_ESTABLISHED]     = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2256   [TCP_SYN_SENT]        = TCP_CLOSE,
2257   [TCP_SYN_RECV]        = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2258   [TCP_FIN_WAIT1]       = TCP_FIN_WAIT1,
2259   [TCP_FIN_WAIT2]       = TCP_FIN_WAIT2,
2260   [TCP_TIME_WAIT]       = TCP_CLOSE,
2261   [TCP_CLOSE]           = TCP_CLOSE,
2262   [TCP_CLOSE_WAIT]      = TCP_LAST_ACK  | TCP_ACTION_FIN,
2263   [TCP_LAST_ACK]        = TCP_LAST_ACK,
2264   [TCP_LISTEN]          = TCP_CLOSE,
2265   [TCP_CLOSING]         = TCP_CLOSING,
2266   [TCP_NEW_SYN_RECV]    = TCP_CLOSE,    /* should not happen ! */
2267 };
2268
2269 static int tcp_close_state(struct sock *sk)
2270 {
2271         int next = (int)new_state[sk->sk_state];
2272         int ns = next & TCP_STATE_MASK;
2273
2274         tcp_set_state(sk, ns);
2275
2276         return next & TCP_ACTION_FIN;
2277 }
2278
2279 /*
2280  *      Shutdown the sending side of a connection. Much like close except
2281  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2282  */
2283
2284 void tcp_shutdown(struct sock *sk, int how)
2285 {
2286         /*      We need to grab some memory, and put together a FIN,
2287          *      and then put it into the queue to be sent.
2288          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2289          */
2290         if (!(how & SEND_SHUTDOWN))
2291                 return;
2292
2293         /* If we've already sent a FIN, or it's a closed state, skip this. */
2294         if ((1 << sk->sk_state) &
2295             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2296              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2297                 /* Clear out any half completed packets.  FIN if needed. */
2298                 if (tcp_close_state(sk))
2299                         tcp_send_fin(sk);
2300         }
2301 }
2302 EXPORT_SYMBOL(tcp_shutdown);
2303
2304 bool tcp_check_oom(struct sock *sk, int shift)
2305 {
2306         bool too_many_orphans, out_of_socket_memory;
2307
2308         too_many_orphans = tcp_too_many_orphans(sk, shift);
2309         out_of_socket_memory = tcp_out_of_memory(sk);
2310
2311         if (too_many_orphans)
2312                 net_info_ratelimited("too many orphaned sockets\n");
2313         if (out_of_socket_memory)
2314                 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2315         return too_many_orphans || out_of_socket_memory;
2316 }
2317
2318 void tcp_close(struct sock *sk, long timeout)
2319 {
2320         struct sk_buff *skb;
2321         int data_was_unread = 0;
2322         int state;
2323
2324         lock_sock(sk);
2325         sk->sk_shutdown = SHUTDOWN_MASK;
2326
2327         if (sk->sk_state == TCP_LISTEN) {
2328                 tcp_set_state(sk, TCP_CLOSE);
2329
2330                 /* Special case. */
2331                 inet_csk_listen_stop(sk);
2332
2333                 goto adjudge_to_death;
2334         }
2335
2336         /*  We need to flush the recv. buffs.  We do this only on the
2337          *  descriptor close, not protocol-sourced closes, because the
2338          *  reader process may not have drained the data yet!
2339          */
2340         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2341                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2342
2343                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2344                         len--;
2345                 data_was_unread += len;
2346                 __kfree_skb(skb);
2347         }
2348
2349         sk_mem_reclaim(sk);
2350
2351         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2352         if (sk->sk_state == TCP_CLOSE)
2353                 goto adjudge_to_death;
2354
2355         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2356          * data was lost. To witness the awful effects of the old behavior of
2357          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2358          * GET in an FTP client, suspend the process, wait for the client to
2359          * advertise a zero window, then kill -9 the FTP client, wheee...
2360          * Note: timeout is always zero in such a case.
2361          */
2362         if (unlikely(tcp_sk(sk)->repair)) {
2363                 sk->sk_prot->disconnect(sk, 0);
2364         } else if (data_was_unread) {
2365                 /* Unread data was tossed, zap the connection. */
2366                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2367                 tcp_set_state(sk, TCP_CLOSE);
2368                 tcp_send_active_reset(sk, sk->sk_allocation);
2369         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2370                 /* Check zero linger _after_ checking for unread data. */
2371                 sk->sk_prot->disconnect(sk, 0);
2372                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2373         } else if (tcp_close_state(sk)) {
2374                 /* We FIN if the application ate all the data before
2375                  * zapping the connection.
2376                  */
2377
2378                 /* RED-PEN. Formally speaking, we have broken TCP state
2379                  * machine. State transitions:
2380                  *
2381                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2382                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2383                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2384                  *
2385                  * are legal only when FIN has been sent (i.e. in window),
2386                  * rather than queued out of window. Purists blame.
2387                  *
2388                  * F.e. "RFC state" is ESTABLISHED,
2389                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2390                  *
2391                  * The visible declinations are that sometimes
2392                  * we enter time-wait state, when it is not required really
2393                  * (harmless), do not send active resets, when they are
2394                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2395                  * they look as CLOSING or LAST_ACK for Linux)
2396                  * Probably, I missed some more holelets.
2397                  *                                              --ANK
2398                  * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2399                  * in a single packet! (May consider it later but will
2400                  * probably need API support or TCP_CORK SYN-ACK until
2401                  * data is written and socket is closed.)
2402                  */
2403                 tcp_send_fin(sk);
2404         }
2405
2406         sk_stream_wait_close(sk, timeout);
2407
2408 adjudge_to_death:
2409         state = sk->sk_state;
2410         sock_hold(sk);
2411         sock_orphan(sk);
2412
2413         /* It is the last release_sock in its life. It will remove backlog. */
2414         release_sock(sk);
2415
2416
2417         /* Now socket is owned by kernel and we acquire BH lock
2418          *  to finish close. No need to check for user refs.
2419          */
2420         local_bh_disable();
2421         bh_lock_sock(sk);
2422         WARN_ON(sock_owned_by_user(sk));
2423
2424         percpu_counter_inc(sk->sk_prot->orphan_count);
2425
2426         /* Have we already been destroyed by a softirq or backlog? */
2427         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2428                 goto out;
2429
2430         /*      This is a (useful) BSD violating of the RFC. There is a
2431          *      problem with TCP as specified in that the other end could
2432          *      keep a socket open forever with no application left this end.
2433          *      We use a 1 minute timeout (about the same as BSD) then kill
2434          *      our end. If they send after that then tough - BUT: long enough
2435          *      that we won't make the old 4*rto = almost no time - whoops
2436          *      reset mistake.
2437          *
2438          *      Nope, it was not mistake. It is really desired behaviour
2439          *      f.e. on http servers, when such sockets are useless, but
2440          *      consume significant resources. Let's do it with special
2441          *      linger2 option.                                 --ANK
2442          */
2443
2444         if (sk->sk_state == TCP_FIN_WAIT2) {
2445                 struct tcp_sock *tp = tcp_sk(sk);
2446                 if (tp->linger2 < 0) {
2447                         tcp_set_state(sk, TCP_CLOSE);
2448                         tcp_send_active_reset(sk, GFP_ATOMIC);
2449                         __NET_INC_STATS(sock_net(sk),
2450                                         LINUX_MIB_TCPABORTONLINGER);
2451                 } else {
2452                         const int tmo = tcp_fin_time(sk);
2453
2454                         if (tmo > TCP_TIMEWAIT_LEN) {
2455                                 inet_csk_reset_keepalive_timer(sk,
2456                                                 tmo - TCP_TIMEWAIT_LEN);
2457                         } else {
2458                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2459                                 goto out;
2460                         }
2461                 }
2462         }
2463         if (sk->sk_state != TCP_CLOSE) {
2464                 sk_mem_reclaim(sk);
2465                 if (tcp_check_oom(sk, 0)) {
2466                         tcp_set_state(sk, TCP_CLOSE);
2467                         tcp_send_active_reset(sk, GFP_ATOMIC);
2468                         __NET_INC_STATS(sock_net(sk),
2469                                         LINUX_MIB_TCPABORTONMEMORY);
2470                 } else if (!check_net(sock_net(sk))) {
2471                         /* Not possible to send reset; just close */
2472                         tcp_set_state(sk, TCP_CLOSE);
2473                 }
2474         }
2475
2476         if (sk->sk_state == TCP_CLOSE) {
2477                 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2478                 /* We could get here with a non-NULL req if the socket is
2479                  * aborted (e.g., closed with unread data) before 3WHS
2480                  * finishes.
2481                  */
2482                 if (req)
2483                         reqsk_fastopen_remove(sk, req, false);
2484                 inet_csk_destroy_sock(sk);
2485         }
2486         /* Otherwise, socket is reprieved until protocol close. */
2487
2488 out:
2489         bh_unlock_sock(sk);
2490         local_bh_enable();
2491         sock_put(sk);
2492 }
2493 EXPORT_SYMBOL(tcp_close);
2494
2495 /* These states need RST on ABORT according to RFC793 */
2496
2497 static inline bool tcp_need_reset(int state)
2498 {
2499         return (1 << state) &
2500                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2501                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2502 }
2503
2504 static void tcp_rtx_queue_purge(struct sock *sk)
2505 {
2506         struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2507
2508         while (p) {
2509                 struct sk_buff *skb = rb_to_skb(p);
2510
2511                 p = rb_next(p);
2512                 /* Since we are deleting whole queue, no need to
2513                  * list_del(&skb->tcp_tsorted_anchor)
2514                  */
2515                 tcp_rtx_queue_unlink(skb, sk);
2516                 sk_wmem_free_skb(sk, skb);
2517         }
2518 }
2519
2520 void tcp_write_queue_purge(struct sock *sk)
2521 {
2522         struct sk_buff *skb;
2523
2524         tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2525         while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2526                 tcp_skb_tsorted_anchor_cleanup(skb);
2527                 sk_wmem_free_skb(sk, skb);
2528         }
2529         tcp_rtx_queue_purge(sk);
2530         INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2531         sk_mem_reclaim(sk);
2532         tcp_clear_all_retrans_hints(tcp_sk(sk));
2533         tcp_sk(sk)->packets_out = 0;
2534 }
2535
2536 int tcp_disconnect(struct sock *sk, int flags)
2537 {
2538         struct inet_sock *inet = inet_sk(sk);
2539         struct inet_connection_sock *icsk = inet_csk(sk);
2540         struct tcp_sock *tp = tcp_sk(sk);
2541         int err = 0;
2542         int old_state = sk->sk_state;
2543
2544         if (old_state != TCP_CLOSE)
2545                 tcp_set_state(sk, TCP_CLOSE);
2546
2547         /* ABORT function of RFC793 */
2548         if (old_state == TCP_LISTEN) {
2549                 inet_csk_listen_stop(sk);
2550         } else if (unlikely(tp->repair)) {
2551                 sk->sk_err = ECONNABORTED;
2552         } else if (tcp_need_reset(old_state) ||
2553                    (tp->snd_nxt != tp->write_seq &&
2554                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2555                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2556                  * states
2557                  */
2558                 tcp_send_active_reset(sk, gfp_any());
2559                 sk->sk_err = ECONNRESET;
2560         } else if (old_state == TCP_SYN_SENT)
2561                 sk->sk_err = ECONNRESET;
2562
2563         tcp_clear_xmit_timers(sk);
2564         __skb_queue_purge(&sk->sk_receive_queue);
2565         tp->copied_seq = tp->rcv_nxt;
2566         tp->urg_data = 0;
2567         tcp_write_queue_purge(sk);
2568         tcp_fastopen_active_disable_ofo_check(sk);
2569         skb_rbtree_purge(&tp->out_of_order_queue);
2570
2571         inet->inet_dport = 0;
2572
2573         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2574                 inet_reset_saddr(sk);
2575
2576         sk->sk_shutdown = 0;
2577         sock_reset_flag(sk, SOCK_DONE);
2578         tp->srtt_us = 0;
2579         tp->write_seq += tp->max_window + 2;
2580         if (tp->write_seq == 0)
2581                 tp->write_seq = 1;
2582         icsk->icsk_backoff = 0;
2583         tp->snd_cwnd = 2;
2584         icsk->icsk_probes_out = 0;
2585         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2586         tp->snd_cwnd_cnt = 0;
2587         tp->window_clamp = 0;
2588         tp->delivered_ce = 0;
2589         tcp_set_ca_state(sk, TCP_CA_Open);
2590         tp->is_sack_reneg = 0;
2591         tcp_clear_retrans(tp);
2592         inet_csk_delack_init(sk);
2593         /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2594          * issue in __tcp_select_window()
2595          */
2596         icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2597         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2598         __sk_dst_reset(sk);
2599         dst_release(sk->sk_rx_dst);
2600         sk->sk_rx_dst = NULL;
2601         tcp_saved_syn_free(tp);
2602         tp->compressed_ack = 0;
2603
2604         /* Clean up fastopen related fields */
2605         tcp_free_fastopen_req(tp);
2606         inet->defer_connect = 0;
2607
2608         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2609
2610         if (sk->sk_frag.page) {
2611                 put_page(sk->sk_frag.page);
2612                 sk->sk_frag.page = NULL;
2613                 sk->sk_frag.offset = 0;
2614         }
2615
2616         sk->sk_error_report(sk);
2617         return err;
2618 }
2619 EXPORT_SYMBOL(tcp_disconnect);
2620
2621 static inline bool tcp_can_repair_sock(const struct sock *sk)
2622 {
2623         return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2624                 (sk->sk_state != TCP_LISTEN);
2625 }
2626
2627 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2628 {
2629         struct tcp_repair_window opt;
2630
2631         if (!tp->repair)
2632                 return -EPERM;
2633
2634         if (len != sizeof(opt))
2635                 return -EINVAL;
2636
2637         if (copy_from_user(&opt, optbuf, sizeof(opt)))
2638                 return -EFAULT;
2639
2640         if (opt.max_window < opt.snd_wnd)
2641                 return -EINVAL;
2642
2643         if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2644                 return -EINVAL;
2645
2646         if (after(opt.rcv_wup, tp->rcv_nxt))
2647                 return -EINVAL;
2648
2649         tp->snd_wl1     = opt.snd_wl1;
2650         tp->snd_wnd     = opt.snd_wnd;
2651         tp->max_window  = opt.max_window;
2652
2653         tp->rcv_wnd     = opt.rcv_wnd;
2654         tp->rcv_wup     = opt.rcv_wup;
2655
2656         return 0;
2657 }
2658
2659 static int tcp_repair_options_est(struct sock *sk,
2660                 struct tcp_repair_opt __user *optbuf, unsigned int len)
2661 {
2662         struct tcp_sock *tp = tcp_sk(sk);
2663         struct tcp_repair_opt opt;
2664
2665         while (len >= sizeof(opt)) {
2666                 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2667                         return -EFAULT;
2668
2669                 optbuf++;
2670                 len -= sizeof(opt);
2671
2672                 switch (opt.opt_code) {
2673                 case TCPOPT_MSS:
2674                         tp->rx_opt.mss_clamp = opt.opt_val;
2675                         tcp_mtup_init(sk);
2676                         break;
2677                 case TCPOPT_WINDOW:
2678                         {
2679                                 u16 snd_wscale = opt.opt_val & 0xFFFF;
2680                                 u16 rcv_wscale = opt.opt_val >> 16;
2681
2682                                 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
2683                                         return -EFBIG;
2684
2685                                 tp->rx_opt.snd_wscale = snd_wscale;
2686                                 tp->rx_opt.rcv_wscale = rcv_wscale;
2687                                 tp->rx_opt.wscale_ok = 1;
2688                         }
2689                         break;
2690                 case TCPOPT_SACK_PERM:
2691                         if (opt.opt_val != 0)
2692                                 return -EINVAL;
2693
2694                         tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2695                         break;
2696                 case TCPOPT_TIMESTAMP:
2697                         if (opt.opt_val != 0)
2698                                 return -EINVAL;
2699
2700                         tp->rx_opt.tstamp_ok = 1;
2701                         break;
2702                 }
2703         }
2704
2705         return 0;
2706 }
2707
2708 /*
2709  *      Socket option code for TCP.
2710  */
2711 static int do_tcp_setsockopt(struct sock *sk, int level,
2712                 int optname, char __user *optval, unsigned int optlen)
2713 {
2714         struct tcp_sock *tp = tcp_sk(sk);
2715         struct inet_connection_sock *icsk = inet_csk(sk);
2716         struct net *net = sock_net(sk);
2717         int val;
2718         int err = 0;
2719
2720         /* These are data/string values, all the others are ints */
2721         switch (optname) {
2722         case TCP_CONGESTION: {
2723                 char name[TCP_CA_NAME_MAX];
2724
2725                 if (optlen < 1)
2726                         return -EINVAL;
2727
2728                 val = strncpy_from_user(name, optval,
2729                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2730                 if (val < 0)
2731                         return -EFAULT;
2732                 name[val] = 0;
2733
2734                 lock_sock(sk);
2735                 err = tcp_set_congestion_control(sk, name, true, true);
2736                 release_sock(sk);
2737                 return err;
2738         }
2739         case TCP_ULP: {
2740                 char name[TCP_ULP_NAME_MAX];
2741
2742                 if (optlen < 1)
2743                         return -EINVAL;
2744
2745                 val = strncpy_from_user(name, optval,
2746                                         min_t(long, TCP_ULP_NAME_MAX - 1,
2747                                               optlen));
2748                 if (val < 0)
2749                         return -EFAULT;
2750                 name[val] = 0;
2751
2752                 lock_sock(sk);
2753                 err = tcp_set_ulp(sk, name);
2754                 release_sock(sk);
2755                 return err;
2756         }
2757         case TCP_FASTOPEN_KEY: {
2758                 __u8 key[TCP_FASTOPEN_KEY_LENGTH];
2759
2760                 if (optlen != sizeof(key))
2761                         return -EINVAL;
2762
2763                 if (copy_from_user(key, optval, optlen))
2764                         return -EFAULT;
2765
2766                 return tcp_fastopen_reset_cipher(net, sk, key, sizeof(key));
2767         }
2768         default:
2769                 /* fallthru */
2770                 break;
2771         }
2772
2773         if (optlen < sizeof(int))
2774                 return -EINVAL;
2775
2776         if (get_user(val, (int __user *)optval))
2777                 return -EFAULT;
2778
2779         lock_sock(sk);
2780
2781         switch (optname) {
2782         case TCP_MAXSEG:
2783                 /* Values greater than interface MTU won't take effect. However
2784                  * at the point when this call is done we typically don't yet
2785                  * know which interface is going to be used
2786                  */
2787                 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
2788                         err = -EINVAL;
2789                         break;
2790                 }
2791                 tp->rx_opt.user_mss = val;
2792                 break;
2793
2794         case TCP_NODELAY:
2795                 if (val) {
2796                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2797                          * this option on corked socket is remembered, but
2798                          * it is not activated until cork is cleared.
2799                          *
2800                          * However, when TCP_NODELAY is set we make
2801                          * an explicit push, which overrides even TCP_CORK
2802                          * for currently queued segments.
2803                          */
2804                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2805                         tcp_push_pending_frames(sk);
2806                 } else {
2807                         tp->nonagle &= ~TCP_NAGLE_OFF;
2808                 }
2809                 break;
2810
2811         case TCP_THIN_LINEAR_TIMEOUTS:
2812                 if (val < 0 || val > 1)
2813                         err = -EINVAL;
2814                 else
2815                         tp->thin_lto = val;
2816                 break;
2817
2818         case TCP_THIN_DUPACK:
2819                 if (val < 0 || val > 1)
2820                         err = -EINVAL;
2821                 break;
2822
2823         case TCP_REPAIR:
2824                 if (!tcp_can_repair_sock(sk))
2825                         err = -EPERM;
2826                 else if (val == TCP_REPAIR_ON) {
2827                         tp->repair = 1;
2828                         sk->sk_reuse = SK_FORCE_REUSE;
2829                         tp->repair_queue = TCP_NO_QUEUE;
2830                 } else if (val == TCP_REPAIR_OFF) {
2831                         tp->repair = 0;
2832                         sk->sk_reuse = SK_NO_REUSE;
2833                         tcp_send_window_probe(sk);
2834                 } else if (val == TCP_REPAIR_OFF_NO_WP) {
2835                         tp->repair = 0;
2836                         sk->sk_reuse = SK_NO_REUSE;
2837                 } else
2838                         err = -EINVAL;
2839
2840                 break;
2841
2842         case TCP_REPAIR_QUEUE:
2843                 if (!tp->repair)
2844                         err = -EPERM;
2845                 else if ((unsigned int)val < TCP_QUEUES_NR)
2846                         tp->repair_queue = val;
2847                 else
2848                         err = -EINVAL;
2849                 break;
2850
2851         case TCP_QUEUE_SEQ:
2852                 if (sk->sk_state != TCP_CLOSE)
2853                         err = -EPERM;
2854                 else if (tp->repair_queue == TCP_SEND_QUEUE)
2855                         tp->write_seq = val;
2856                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2857                         tp->rcv_nxt = val;
2858                 else
2859                         err = -EINVAL;
2860                 break;
2861
2862         case TCP_REPAIR_OPTIONS:
2863                 if (!tp->repair)
2864                         err = -EINVAL;
2865                 else if (sk->sk_state == TCP_ESTABLISHED)
2866                         err = tcp_repair_options_est(sk,
2867                                         (struct tcp_repair_opt __user *)optval,
2868                                         optlen);
2869                 else
2870                         err = -EPERM;
2871                 break;
2872
2873         case TCP_CORK:
2874                 /* When set indicates to always queue non-full frames.
2875                  * Later the user clears this option and we transmit
2876                  * any pending partial frames in the queue.  This is
2877                  * meant to be used alongside sendfile() to get properly
2878                  * filled frames when the user (for example) must write
2879                  * out headers with a write() call first and then use
2880                  * sendfile to send out the data parts.
2881                  *
2882                  * TCP_CORK can be set together with TCP_NODELAY and it is
2883                  * stronger than TCP_NODELAY.
2884                  */
2885                 if (val) {
2886                         tp->nonagle |= TCP_NAGLE_CORK;
2887                 } else {
2888                         tp->nonagle &= ~TCP_NAGLE_CORK;
2889                         if (tp->nonagle&TCP_NAGLE_OFF)
2890                                 tp->nonagle |= TCP_NAGLE_PUSH;
2891                         tcp_push_pending_frames(sk);
2892                 }
2893                 break;
2894
2895         case TCP_KEEPIDLE:
2896                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2897                         err = -EINVAL;
2898                 else {
2899                         tp->keepalive_time = val * HZ;
2900                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2901                             !((1 << sk->sk_state) &
2902                               (TCPF_CLOSE | TCPF_LISTEN))) {
2903                                 u32 elapsed = keepalive_time_elapsed(tp);
2904                                 if (tp->keepalive_time > elapsed)
2905                                         elapsed = tp->keepalive_time - elapsed;
2906                                 else
2907                                         elapsed = 0;
2908                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2909                         }
2910                 }
2911                 break;
2912         case TCP_KEEPINTVL:
2913                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2914                         err = -EINVAL;
2915                 else
2916                         tp->keepalive_intvl = val * HZ;
2917                 break;
2918         case TCP_KEEPCNT:
2919                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2920                         err = -EINVAL;
2921                 else
2922                         tp->keepalive_probes = val;
2923                 break;
2924         case TCP_SYNCNT:
2925                 if (val < 1 || val > MAX_TCP_SYNCNT)
2926                         err = -EINVAL;
2927                 else
2928                         icsk->icsk_syn_retries = val;
2929                 break;
2930
2931         case TCP_SAVE_SYN:
2932                 if (val < 0 || val > 1)
2933                         err = -EINVAL;
2934                 else
2935                         tp->save_syn = val;
2936                 break;
2937
2938         case TCP_LINGER2:
2939                 if (val < 0)
2940                         tp->linger2 = -1;
2941                 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
2942                         tp->linger2 = 0;
2943                 else
2944                         tp->linger2 = val * HZ;
2945                 break;
2946
2947         case TCP_DEFER_ACCEPT:
2948                 /* Translate value in seconds to number of retransmits */
2949                 icsk->icsk_accept_queue.rskq_defer_accept =
2950                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2951                                         TCP_RTO_MAX / HZ);
2952                 break;
2953
2954         case TCP_WINDOW_CLAMP:
2955                 if (!val) {
2956                         if (sk->sk_state != TCP_CLOSE) {
2957                                 err = -EINVAL;
2958                                 break;
2959                         }
2960                         tp->window_clamp = 0;
2961                 } else
2962                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2963                                                 SOCK_MIN_RCVBUF / 2 : val;
2964                 break;
2965
2966         case TCP_QUICKACK:
2967                 if (!val) {
2968                         icsk->icsk_ack.pingpong = 1;
2969                 } else {
2970                         icsk->icsk_ack.pingpong = 0;
2971                         if ((1 << sk->sk_state) &
2972                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2973                             inet_csk_ack_scheduled(sk)) {
2974                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2975                                 tcp_cleanup_rbuf(sk, 1);
2976                                 if (!(val & 1))
2977                                         icsk->icsk_ack.pingpong = 1;
2978                         }
2979                 }
2980                 break;
2981
2982 #ifdef CONFIG_TCP_MD5SIG
2983         case TCP_MD5SIG:
2984         case TCP_MD5SIG_EXT:
2985                 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
2986                         err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
2987                 else
2988                         err = -EINVAL;
2989                 break;
2990 #endif
2991         case TCP_USER_TIMEOUT:
2992                 /* Cap the max time in ms TCP will retry or probe the window
2993                  * before giving up and aborting (ETIMEDOUT) a connection.
2994                  */
2995                 if (val < 0)
2996                         err = -EINVAL;
2997                 else
2998                         icsk->icsk_user_timeout = msecs_to_jiffies(val);
2999                 break;
3000
3001         case TCP_FASTOPEN:
3002                 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3003                     TCPF_LISTEN))) {
3004                         tcp_fastopen_init_key_once(net);
3005
3006                         fastopen_queue_tune(sk, val);
3007                 } else {
3008                         err = -EINVAL;
3009                 }
3010                 break;
3011         case TCP_FASTOPEN_CONNECT:
3012                 if (val > 1 || val < 0) {
3013                         err = -EINVAL;
3014                 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
3015                         if (sk->sk_state == TCP_CLOSE)
3016                                 tp->fastopen_connect = val;
3017                         else
3018                                 err = -EINVAL;
3019                 } else {
3020                         err = -EOPNOTSUPP;
3021                 }
3022                 break;
3023         case TCP_FASTOPEN_NO_COOKIE:
3024                 if (val > 1 || val < 0)
3025                         err = -EINVAL;
3026                 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3027                         err = -EINVAL;
3028                 else
3029                         tp->fastopen_no_cookie = val;
3030                 break;
3031         case TCP_TIMESTAMP:
3032                 if (!tp->repair)
3033                         err = -EPERM;
3034                 else
3035                         tp->tsoffset = val - tcp_time_stamp_raw();
3036                 break;
3037         case TCP_REPAIR_WINDOW:
3038                 err = tcp_repair_set_window(tp, optval, optlen);
3039                 break;
3040         case TCP_NOTSENT_LOWAT:
3041                 tp->notsent_lowat = val;
3042                 sk->sk_write_space(sk);
3043                 break;
3044         case TCP_INQ:
3045                 if (val > 1 || val < 0)
3046                         err = -EINVAL;
3047                 else
3048                         tp->recvmsg_inq = val;
3049                 break;
3050         default:
3051                 err = -ENOPROTOOPT;
3052                 break;
3053         }
3054
3055         release_sock(sk);
3056         return err;
3057 }
3058
3059 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
3060                    unsigned int optlen)
3061 {
3062         const struct inet_connection_sock *icsk = inet_csk(sk);
3063
3064         if (level != SOL_TCP)
3065                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3066                                                      optval, optlen);
3067         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3068 }
3069 EXPORT_SYMBOL(tcp_setsockopt);
3070
3071 #ifdef CONFIG_COMPAT
3072 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
3073                           char __user *optval, unsigned int optlen)
3074 {
3075         if (level != SOL_TCP)
3076                 return inet_csk_compat_setsockopt(sk, level, optname,
3077                                                   optval, optlen);
3078         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3079 }
3080 EXPORT_SYMBOL(compat_tcp_setsockopt);
3081 #endif
3082
3083 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3084                                       struct tcp_info *info)
3085 {
3086         u64 stats[__TCP_CHRONO_MAX], total = 0;
3087         enum tcp_chrono i;
3088
3089         for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3090                 stats[i] = tp->chrono_stat[i - 1];
3091                 if (i == tp->chrono_type)
3092                         stats[i] += tcp_jiffies32 - tp->chrono_start;
3093                 stats[i] *= USEC_PER_SEC / HZ;
3094                 total += stats[i];
3095         }
3096
3097         info->tcpi_busy_time = total;
3098         info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3099         info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3100 }
3101
3102 /* Return information about state of tcp endpoint in API format. */
3103 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3104 {
3105         const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3106         const struct inet_connection_sock *icsk = inet_csk(sk);
3107         u32 now;
3108         u64 rate64;
3109         bool slow;
3110         u32 rate;
3111
3112         memset(info, 0, sizeof(*info));
3113         if (sk->sk_type != SOCK_STREAM)
3114                 return;
3115
3116         info->tcpi_state = inet_sk_state_load(sk);
3117
3118         /* Report meaningful fields for all TCP states, including listeners */
3119         rate = READ_ONCE(sk->sk_pacing_rate);
3120         rate64 = rate != ~0U ? rate : ~0ULL;
3121         info->tcpi_pacing_rate = rate64;
3122
3123         rate = READ_ONCE(sk->sk_max_pacing_rate);
3124         rate64 = rate != ~0U ? rate : ~0ULL;
3125         info->tcpi_max_pacing_rate = rate64;
3126
3127         info->tcpi_reordering = tp->reordering;
3128         info->tcpi_snd_cwnd = tp->snd_cwnd;
3129
3130         if (info->tcpi_state == TCP_LISTEN) {
3131                 /* listeners aliased fields :
3132                  * tcpi_unacked -> Number of children ready for accept()
3133                  * tcpi_sacked  -> max backlog
3134                  */
3135                 info->tcpi_unacked = sk->sk_ack_backlog;
3136                 info->tcpi_sacked = sk->sk_max_ack_backlog;
3137                 return;
3138         }
3139
3140         slow = lock_sock_fast(sk);
3141
3142         info->tcpi_ca_state = icsk->icsk_ca_state;
3143         info->tcpi_retransmits = icsk->icsk_retransmits;
3144         info->tcpi_probes = icsk->icsk_probes_out;
3145         info->tcpi_backoff = icsk->icsk_backoff;
3146
3147         if (tp->rx_opt.tstamp_ok)
3148                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3149         if (tcp_is_sack(tp))
3150                 info->tcpi_options |= TCPI_OPT_SACK;
3151         if (tp->rx_opt.wscale_ok) {
3152                 info->tcpi_options |= TCPI_OPT_WSCALE;
3153                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3154                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3155         }
3156
3157         if (tp->ecn_flags & TCP_ECN_OK)
3158                 info->tcpi_options |= TCPI_OPT_ECN;
3159         if (tp->ecn_flags & TCP_ECN_SEEN)
3160                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3161         if (tp->syn_data_acked)
3162                 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3163
3164         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3165         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3166         info->tcpi_snd_mss = tp->mss_cache;
3167         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3168
3169         info->tcpi_unacked = tp->packets_out;
3170         info->tcpi_sacked = tp->sacked_out;
3171
3172         info->tcpi_lost = tp->lost_out;
3173         info->tcpi_retrans = tp->retrans_out;
3174
3175         now = tcp_jiffies32;
3176         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3177         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3178         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3179
3180         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3181         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3182         info->tcpi_rtt = tp->srtt_us >> 3;
3183         info->tcpi_rttvar = tp->mdev_us >> 2;
3184         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3185         info->tcpi_advmss = tp->advmss;
3186
3187         info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3188         info->tcpi_rcv_space = tp->rcvq_space.space;
3189
3190         info->tcpi_total_retrans = tp->total_retrans;
3191
3192         info->tcpi_bytes_acked = tp->bytes_acked;
3193         info->tcpi_bytes_received = tp->bytes_received;
3194         info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3195         tcp_get_info_chrono_stats(tp, info);
3196
3197         info->tcpi_segs_out = tp->segs_out;
3198         info->tcpi_segs_in = tp->segs_in;
3199
3200         info->tcpi_min_rtt = tcp_min_rtt(tp);
3201         info->tcpi_data_segs_in = tp->data_segs_in;
3202         info->tcpi_data_segs_out = tp->data_segs_out;
3203
3204         info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3205         rate64 = tcp_compute_delivery_rate(tp);
3206         if (rate64)
3207                 info->tcpi_delivery_rate = rate64;
3208         info->tcpi_delivered = tp->delivered;
3209         info->tcpi_delivered_ce = tp->delivered_ce;
3210         unlock_sock_fast(sk, slow);
3211 }
3212 EXPORT_SYMBOL_GPL(tcp_get_info);
3213
3214 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk)
3215 {
3216         const struct tcp_sock *tp = tcp_sk(sk);
3217         struct sk_buff *stats;
3218         struct tcp_info info;
3219         u64 rate64;
3220         u32 rate;
3221
3222         stats = alloc_skb(7 * nla_total_size_64bit(sizeof(u64)) +
3223                           7 * nla_total_size(sizeof(u32)) +
3224                           3 * nla_total_size(sizeof(u8)), GFP_ATOMIC);
3225         if (!stats)
3226                 return NULL;
3227
3228         tcp_get_info_chrono_stats(tp, &info);
3229         nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3230                           info.tcpi_busy_time, TCP_NLA_PAD);
3231         nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3232                           info.tcpi_rwnd_limited, TCP_NLA_PAD);
3233         nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3234                           info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3235         nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3236                           tp->data_segs_out, TCP_NLA_PAD);
3237         nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3238                           tp->total_retrans, TCP_NLA_PAD);
3239
3240         rate = READ_ONCE(sk->sk_pacing_rate);
3241         rate64 = rate != ~0U ? rate : ~0ULL;
3242         nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3243
3244         rate64 = tcp_compute_delivery_rate(tp);
3245         nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3246
3247         nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3248         nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3249         nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3250
3251         nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3252         nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3253         nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3254         nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3255         nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3256
3257         nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3258         nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3259
3260         return stats;
3261 }
3262
3263 static int do_tcp_getsockopt(struct sock *sk, int level,
3264                 int optname, char __user *optval, int __user *optlen)
3265 {
3266         struct inet_connection_sock *icsk = inet_csk(sk);
3267         struct tcp_sock *tp = tcp_sk(sk);
3268         struct net *net = sock_net(sk);
3269         int val, len;
3270
3271         if (get_user(len, optlen))
3272                 return -EFAULT;
3273
3274         len = min_t(unsigned int, len, sizeof(int));
3275
3276         if (len < 0)
3277                 return -EINVAL;
3278
3279         switch (optname) {
3280         case TCP_MAXSEG:
3281                 val = tp->mss_cache;
3282                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3283                         val = tp->rx_opt.user_mss;
3284                 if (tp->repair)
3285                         val = tp->rx_opt.mss_clamp;
3286                 break;
3287         case TCP_NODELAY:
3288                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3289                 break;
3290         case TCP_CORK:
3291                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3292                 break;
3293         case TCP_KEEPIDLE:
3294                 val = keepalive_time_when(tp) / HZ;
3295                 break;
3296         case TCP_KEEPINTVL:
3297                 val = keepalive_intvl_when(tp) / HZ;
3298                 break;
3299         case TCP_KEEPCNT:
3300                 val = keepalive_probes(tp);
3301                 break;
3302         case TCP_SYNCNT:
3303                 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3304                 break;
3305         case TCP_LINGER2:
3306                 val = tp->linger2;
3307                 if (val >= 0)
3308                         val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3309                 break;
3310         case TCP_DEFER_ACCEPT:
3311                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3312                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3313                 break;
3314         case TCP_WINDOW_CLAMP:
3315                 val = tp->window_clamp;
3316                 break;
3317         case TCP_INFO: {
3318                 struct tcp_info info;
3319
3320                 if (get_user(len, optlen))
3321                         return -EFAULT;
3322
3323                 tcp_get_info(sk, &info);
3324
3325                 len = min_t(unsigned int, len, sizeof(info));
3326                 if (put_user(len, optlen))
3327                         return -EFAULT;
3328                 if (copy_to_user(optval, &info, len))
3329                         return -EFAULT;
3330                 return 0;
3331         }
3332         case TCP_CC_INFO: {
3333                 const struct tcp_congestion_ops *ca_ops;
3334                 union tcp_cc_info info;
3335                 size_t sz = 0;
3336                 int attr;
3337
3338                 if (get_user(len, optlen))
3339                         return -EFAULT;
3340
3341                 ca_ops = icsk->icsk_ca_ops;
3342                 if (ca_ops && ca_ops->get_info)
3343                         sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3344
3345                 len = min_t(unsigned int, len, sz);
3346                 if (put_user(len, optlen))
3347                         return -EFAULT;
3348                 if (copy_to_user(optval, &info, len))
3349                         return -EFAULT;
3350                 return 0;
3351         }
3352         case TCP_QUICKACK:
3353                 val = !icsk->icsk_ack.pingpong;
3354                 break;
3355
3356         case TCP_CONGESTION:
3357                 if (get_user(len, optlen))
3358                         return -EFAULT;
3359                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
3360                 if (put_user(len, optlen))
3361                         return -EFAULT;
3362                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3363                         return -EFAULT;
3364                 return 0;
3365
3366         case TCP_ULP:
3367                 if (get_user(len, optlen))
3368                         return -EFAULT;
3369                 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
3370                 if (!icsk->icsk_ulp_ops) {
3371                         if (put_user(0, optlen))
3372                                 return -EFAULT;
3373                         return 0;
3374                 }
3375                 if (put_user(len, optlen))
3376                         return -EFAULT;
3377                 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
3378                         return -EFAULT;
3379                 return 0;
3380
3381         case TCP_FASTOPEN_KEY: {
3382                 __u8 key[TCP_FASTOPEN_KEY_LENGTH];
3383                 struct tcp_fastopen_context *ctx;
3384
3385                 if (get_user(len, optlen))
3386                         return -EFAULT;
3387
3388                 rcu_read_lock();
3389                 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
3390                 if (ctx)
3391                         memcpy(key, ctx->key, sizeof(key));
3392                 else
3393                         len = 0;
3394                 rcu_read_unlock();
3395
3396                 len = min_t(unsigned int, len, sizeof(key));
3397                 if (put_user(len, optlen))
3398                         return -EFAULT;
3399                 if (copy_to_user(optval, key, len))
3400                         return -EFAULT;
3401                 return 0;
3402         }
3403         case TCP_THIN_LINEAR_TIMEOUTS:
3404                 val = tp->thin_lto;
3405                 break;
3406
3407         case TCP_THIN_DUPACK:
3408                 val = 0;
3409                 break;
3410
3411         case TCP_REPAIR:
3412                 val = tp->repair;
3413                 break;
3414
3415         case TCP_REPAIR_QUEUE:
3416                 if (tp->repair)
3417                         val = tp->repair_queue;
3418                 else
3419                         return -EINVAL;
3420                 break;
3421
3422         case TCP_REPAIR_WINDOW: {
3423                 struct tcp_repair_window opt;
3424
3425                 if (get_user(len, optlen))
3426                         return -EFAULT;
3427
3428                 if (len != sizeof(opt))
3429                         return -EINVAL;
3430
3431                 if (!tp->repair)
3432                         return -EPERM;
3433
3434                 opt.snd_wl1     = tp->snd_wl1;
3435                 opt.snd_wnd     = tp->snd_wnd;
3436                 opt.max_window  = tp->max_window;
3437                 opt.rcv_wnd     = tp->rcv_wnd;
3438                 opt.rcv_wup     = tp->rcv_wup;
3439
3440                 if (copy_to_user(optval, &opt, len))
3441                         return -EFAULT;
3442                 return 0;
3443         }
3444         case TCP_QUEUE_SEQ:
3445                 if (tp->repair_queue == TCP_SEND_QUEUE)
3446                         val = tp->write_seq;
3447                 else if (tp->repair_queue == TCP_RECV_QUEUE)
3448                         val = tp->rcv_nxt;
3449                 else
3450                         return -EINVAL;
3451                 break;
3452
3453         case TCP_USER_TIMEOUT:
3454                 val = jiffies_to_msecs(icsk->icsk_user_timeout);
3455                 break;
3456
3457         case TCP_FASTOPEN:
3458                 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3459                 break;
3460
3461         case TCP_FASTOPEN_CONNECT:
3462                 val = tp->fastopen_connect;
3463                 break;
3464
3465         case TCP_FASTOPEN_NO_COOKIE:
3466                 val = tp->fastopen_no_cookie;
3467                 break;
3468
3469         case TCP_TIMESTAMP:
3470                 val = tcp_time_stamp_raw() + tp->tsoffset;
3471                 break;
3472         case TCP_NOTSENT_LOWAT:
3473                 val = tp->notsent_lowat;
3474                 break;
3475         case TCP_INQ:
3476                 val = tp->recvmsg_inq;
3477                 break;
3478         case TCP_SAVE_SYN:
3479                 val = tp->save_syn;
3480                 break;
3481         case TCP_SAVED_SYN: {
3482                 if (get_user(len, optlen))
3483                         return -EFAULT;
3484
3485                 lock_sock(sk);
3486                 if (tp->saved_syn) {
3487                         if (len < tp->saved_syn[0]) {
3488                                 if (put_user(tp->saved_syn[0], optlen)) {
3489                                         release_sock(sk);
3490                                         return -EFAULT;
3491                                 }
3492                                 release_sock(sk);
3493                                 return -EINVAL;
3494                         }
3495                         len = tp->saved_syn[0];
3496                         if (put_user(len, optlen)) {
3497                                 release_sock(sk);
3498                                 return -EFAULT;
3499                         }
3500                         if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3501                                 release_sock(sk);
3502                                 return -EFAULT;
3503                         }
3504                         tcp_saved_syn_free(tp);
3505                         release_sock(sk);
3506                 } else {
3507                         release_sock(sk);
3508                         len = 0;
3509                         if (put_user(len, optlen))
3510                                 return -EFAULT;
3511                 }
3512                 return 0;
3513         }
3514 #ifdef CONFIG_MMU
3515         case TCP_ZEROCOPY_RECEIVE: {
3516                 struct tcp_zerocopy_receive zc;
3517                 int err;
3518
3519                 if (get_user(len, optlen))
3520                         return -EFAULT;
3521                 if (len != sizeof(zc))
3522                         return -EINVAL;
3523                 if (copy_from_user(&zc, optval, len))
3524                         return -EFAULT;
3525                 lock_sock(sk);
3526                 err = tcp_zerocopy_receive(sk, &zc);
3527                 release_sock(sk);
3528                 if (!err && copy_to_user(optval, &zc, len))
3529                         err = -EFAULT;
3530                 return err;
3531         }
3532 #endif
3533         default:
3534                 return -ENOPROTOOPT;
3535         }
3536
3537         if (put_user(len, optlen))
3538                 return -EFAULT;
3539         if (copy_to_user(optval, &val, len))
3540                 return -EFAULT;
3541         return 0;
3542 }
3543
3544 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3545                    int __user *optlen)
3546 {
3547         struct inet_connection_sock *icsk = inet_csk(sk);
3548
3549         if (level != SOL_TCP)
3550                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3551                                                      optval, optlen);
3552         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3553 }
3554 EXPORT_SYMBOL(tcp_getsockopt);
3555
3556 #ifdef CONFIG_COMPAT
3557 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3558                           char __user *optval, int __user *optlen)
3559 {
3560         if (level != SOL_TCP)
3561                 return inet_csk_compat_getsockopt(sk, level, optname,
3562                                                   optval, optlen);
3563         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3564 }
3565 EXPORT_SYMBOL(compat_tcp_getsockopt);
3566 #endif
3567
3568 #ifdef CONFIG_TCP_MD5SIG
3569 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3570 static DEFINE_MUTEX(tcp_md5sig_mutex);
3571 static bool tcp_md5sig_pool_populated = false;
3572
3573 static void __tcp_alloc_md5sig_pool(void)
3574 {
3575         struct crypto_ahash *hash;
3576         int cpu;
3577
3578         hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3579         if (IS_ERR(hash))
3580                 return;
3581
3582         for_each_possible_cpu(cpu) {
3583                 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3584                 struct ahash_request *req;
3585
3586                 if (!scratch) {
3587                         scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3588                                                sizeof(struct tcphdr),
3589                                                GFP_KERNEL,
3590                                                cpu_to_node(cpu));
3591                         if (!scratch)
3592                                 return;
3593                         per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3594                 }
3595                 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3596                         continue;
3597
3598                 req = ahash_request_alloc(hash, GFP_KERNEL);
3599                 if (!req)
3600                         return;
3601
3602                 ahash_request_set_callback(req, 0, NULL, NULL);
3603
3604                 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3605         }
3606         /* before setting tcp_md5sig_pool_populated, we must commit all writes
3607          * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3608          */
3609         smp_wmb();
3610         tcp_md5sig_pool_populated = true;
3611 }
3612
3613 bool tcp_alloc_md5sig_pool(void)
3614 {
3615         if (unlikely(!tcp_md5sig_pool_populated)) {
3616                 mutex_lock(&tcp_md5sig_mutex);
3617
3618                 if (!tcp_md5sig_pool_populated)
3619                         __tcp_alloc_md5sig_pool();
3620
3621                 mutex_unlock(&tcp_md5sig_mutex);
3622         }
3623         return tcp_md5sig_pool_populated;
3624 }
3625 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3626
3627
3628 /**
3629  *      tcp_get_md5sig_pool - get md5sig_pool for this user
3630  *
3631  *      We use percpu structure, so if we succeed, we exit with preemption
3632  *      and BH disabled, to make sure another thread or softirq handling
3633  *      wont try to get same context.
3634  */
3635 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3636 {
3637         local_bh_disable();
3638
3639         if (tcp_md5sig_pool_populated) {
3640                 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3641                 smp_rmb();
3642                 return this_cpu_ptr(&tcp_md5sig_pool);
3643         }
3644         local_bh_enable();
3645         return NULL;
3646 }
3647 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3648
3649 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3650                           const struct sk_buff *skb, unsigned int header_len)
3651 {
3652         struct scatterlist sg;
3653         const struct tcphdr *tp = tcp_hdr(skb);
3654         struct ahash_request *req = hp->md5_req;
3655         unsigned int i;
3656         const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3657                                            skb_headlen(skb) - header_len : 0;
3658         const struct skb_shared_info *shi = skb_shinfo(skb);
3659         struct sk_buff *frag_iter;
3660
3661         sg_init_table(&sg, 1);
3662
3663         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3664         ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3665         if (crypto_ahash_update(req))
3666                 return 1;
3667
3668         for (i = 0; i < shi->nr_frags; ++i) {
3669                 const struct skb_frag_struct *f = &shi->frags[i];
3670                 unsigned int offset = f->page_offset;
3671                 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3672
3673                 sg_set_page(&sg, page, skb_frag_size(f),
3674                             offset_in_page(offset));
3675                 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3676                 if (crypto_ahash_update(req))
3677                         return 1;
3678         }
3679
3680         skb_walk_frags(skb, frag_iter)
3681                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3682                         return 1;
3683
3684         return 0;
3685 }
3686 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3687
3688 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3689 {
3690         struct scatterlist sg;
3691
3692         sg_init_one(&sg, key->key, key->keylen);
3693         ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen);
3694         return crypto_ahash_update(hp->md5_req);
3695 }
3696 EXPORT_SYMBOL(tcp_md5_hash_key);
3697
3698 #endif
3699
3700 void tcp_done(struct sock *sk)
3701 {
3702         struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3703
3704         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3705                 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3706
3707         tcp_set_state(sk, TCP_CLOSE);
3708         tcp_clear_xmit_timers(sk);
3709         if (req)
3710                 reqsk_fastopen_remove(sk, req, false);
3711
3712         sk->sk_shutdown = SHUTDOWN_MASK;
3713
3714         if (!sock_flag(sk, SOCK_DEAD))
3715                 sk->sk_state_change(sk);
3716         else
3717                 inet_csk_destroy_sock(sk);
3718 }
3719 EXPORT_SYMBOL_GPL(tcp_done);
3720
3721 int tcp_abort(struct sock *sk, int err)
3722 {
3723         if (!sk_fullsock(sk)) {
3724                 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3725                         struct request_sock *req = inet_reqsk(sk);
3726
3727                         local_bh_disable();
3728                         inet_csk_reqsk_queue_drop(req->rsk_listener, req);
3729                         local_bh_enable();
3730                         return 0;
3731                 }
3732                 return -EOPNOTSUPP;
3733         }
3734
3735         /* Don't race with userspace socket closes such as tcp_close. */
3736         lock_sock(sk);
3737
3738         if (sk->sk_state == TCP_LISTEN) {
3739                 tcp_set_state(sk, TCP_CLOSE);
3740                 inet_csk_listen_stop(sk);
3741         }
3742
3743         /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3744         local_bh_disable();
3745         bh_lock_sock(sk);
3746
3747         if (!sock_flag(sk, SOCK_DEAD)) {
3748                 sk->sk_err = err;
3749                 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3750                 smp_wmb();
3751                 sk->sk_error_report(sk);
3752                 if (tcp_need_reset(sk->sk_state))
3753                         tcp_send_active_reset(sk, GFP_ATOMIC);
3754                 tcp_done(sk);
3755         }
3756
3757         bh_unlock_sock(sk);
3758         local_bh_enable();
3759         tcp_write_queue_purge(sk);
3760         release_sock(sk);
3761         return 0;
3762 }
3763 EXPORT_SYMBOL_GPL(tcp_abort);
3764
3765 extern struct tcp_congestion_ops tcp_reno;
3766
3767 static __initdata unsigned long thash_entries;
3768 static int __init set_thash_entries(char *str)
3769 {
3770         ssize_t ret;
3771
3772         if (!str)
3773                 return 0;
3774
3775         ret = kstrtoul(str, 0, &thash_entries);
3776         if (ret)
3777                 return 0;
3778
3779         return 1;
3780 }
3781 __setup("thash_entries=", set_thash_entries);
3782
3783 static void __init tcp_init_mem(void)
3784 {
3785         unsigned long limit = nr_free_buffer_pages() / 16;
3786
3787         limit = max(limit, 128UL);
3788         sysctl_tcp_mem[0] = limit / 4 * 3;              /* 4.68 % */
3789         sysctl_tcp_mem[1] = limit;                      /* 6.25 % */
3790         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;      /* 9.37 % */
3791 }
3792
3793 void __init tcp_init(void)
3794 {
3795         int max_rshare, max_wshare, cnt;
3796         unsigned long limit;
3797         unsigned int i;
3798
3799         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3800                      FIELD_SIZEOF(struct sk_buff, cb));
3801
3802         percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3803         percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3804         inet_hashinfo_init(&tcp_hashinfo);
3805         inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
3806                             thash_entries, 21,  /* one slot per 2 MB*/
3807                             0, 64 * 1024);
3808         tcp_hashinfo.bind_bucket_cachep =
3809                 kmem_cache_create("tcp_bind_bucket",
3810                                   sizeof(struct inet_bind_bucket), 0,
3811                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3812
3813         /* Size and allocate the main established and bind bucket
3814          * hash tables.
3815          *
3816          * The methodology is similar to that of the buffer cache.
3817          */
3818         tcp_hashinfo.ehash =
3819                 alloc_large_system_hash("TCP established",
3820                                         sizeof(struct inet_ehash_bucket),
3821                                         thash_entries,
3822                                         17, /* one slot per 128 KB of memory */
3823                                         0,
3824                                         NULL,
3825                                         &tcp_hashinfo.ehash_mask,
3826                                         0,
3827                                         thash_entries ? 0 : 512 * 1024);
3828         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3829                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3830
3831         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3832                 panic("TCP: failed to alloc ehash_locks");
3833         tcp_hashinfo.bhash =
3834                 alloc_large_system_hash("TCP bind",
3835                                         sizeof(struct inet_bind_hashbucket),
3836                                         tcp_hashinfo.ehash_mask + 1,
3837                                         17, /* one slot per 128 KB of memory */
3838                                         0,
3839                                         &tcp_hashinfo.bhash_size,
3840                                         NULL,
3841                                         0,
3842                                         64 * 1024);
3843         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3844         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3845                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3846                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3847         }
3848
3849
3850         cnt = tcp_hashinfo.ehash_mask + 1;
3851         sysctl_tcp_max_orphans = cnt / 2;
3852
3853         tcp_init_mem();
3854         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3855         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3856         max_wshare = min(4UL*1024*1024, limit);
3857         max_rshare = min(6UL*1024*1024, limit);
3858
3859         init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3860         init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
3861         init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3862
3863         init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3864         init_net.ipv4.sysctl_tcp_rmem[1] = 87380;
3865         init_net.ipv4.sysctl_tcp_rmem[2] = max(87380, max_rshare);
3866
3867         pr_info("Hash tables configured (established %u bind %u)\n",
3868                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3869
3870         tcp_v4_init();
3871         tcp_metrics_init();
3872         BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3873         tcp_tasklet_init();
3874 }