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 (READ_ONCE(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 (!READ_ONCE(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         int avail = READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->copied_seq);
 492
 493         if (avail > 0) {
 494                 if (avail >= target)
 495                         return true;
 496                 if (tcp_rmem_pressure(sk))
 497                         return true;
 498                 if (tcp_receive_window(tp) <= inet_csk(sk)->icsk_ack.rcv_mss)
 499                         return true;
 500         }
 501         if (sk->sk_prot->stream_memory_read)
 502                 return sk->sk_prot->stream_memory_read(sk);
 503         return false;
 504 }
 505
 506 /*
 507  *      Wait for a TCP event.
 508  *
 509  *      Note that we don't need to lock the socket, as the upper poll layers
 510  *      take care of normal races (between the test and the event) and we don't
 511  *      go look at any of the socket buffers directly.
 512  */
 513 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 514 {
 515         __poll_t mask;
 516         struct sock *sk = sock->sk;
 517         const struct tcp_sock *tp = tcp_sk(sk);
 518         int state;
 519
 520         sock_poll_wait(file, sock, wait);
 521
 522         state = inet_sk_state_load(sk);
 523         if (state == TCP_LISTEN)
 524                 return inet_csk_listen_poll(sk);
 525
 526         /* Socket is not locked. We are protected from async events
 527          * by poll logic and correct handling of state changes
 528          * made by other threads is impossible in any case.
 529          */
 530
 531         mask = 0;
 532
 533         /*
 534          * EPOLLHUP is certainly not done right. But poll() doesn't
 535          * have a notion of HUP in just one direction, and for a
 536          * socket the read side is more interesting.
 537          *
 538          * Some poll() documentation says that EPOLLHUP is incompatible
 539          * with the EPOLLOUT/POLLWR flags, so somebody should check this
 540          * all. But careful, it tends to be safer to return too many
 541          * bits than too few, and you can easily break real applications
 542          * if you don't tell them that something has hung up!
 543          *
 544          * Check-me.
 545          *
 546          * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
 547          * our fs/select.c). It means that after we received EOF,
 548          * poll always returns immediately, making impossible poll() on write()
 549          * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
 550          * if and only if shutdown has been made in both directions.
 551          * Actually, it is interesting to look how Solaris and DUX
 552          * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
 553          * then we could set it on SND_SHUTDOWN. BTW examples given
 554          * in Stevens' books assume exactly this behaviour, it explains
 555          * why EPOLLHUP is incompatible with EPOLLOUT.  --ANK
 556          *
 557          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 558          * blocking on fresh not-connected or disconnected socket. --ANK
 559          */
 560         if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
 561                 mask |= EPOLLHUP;
 562         if (sk->sk_shutdown & RCV_SHUTDOWN)
 563                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
 564
 565         /* Connected or passive Fast Open socket? */
 566         if (state != TCP_SYN_SENT &&
 567             (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
 568                 int target = sock_rcvlowat(sk, 0, INT_MAX);
 569
 570                 if (tp->urg_seq == READ_ONCE(tp->copied_seq) &&
 571                     !sock_flag(sk, SOCK_URGINLINE) &&
 572                     tp->urg_data)
 573                         target++;
 574
 575                 if (tcp_stream_is_readable(tp, target, sk))
 576                         mask |= EPOLLIN | EPOLLRDNORM;
 577
 578                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 579                         if (sk_stream_is_writeable(sk)) {
 580                                 mask |= EPOLLOUT | EPOLLWRNORM;
 581                         } else {  /* send SIGIO later */
 582                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 583                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 584
 585                                 /* Race breaker. If space is freed after
 586                                  * wspace test but before the flags are set,
 587                                  * IO signal will be lost. Memory barrier
 588                                  * pairs with the input side.
 589                                  */
 590                                 smp_mb__after_atomic();
 591                                 if (sk_stream_is_writeable(sk))
 592                                         mask |= EPOLLOUT | EPOLLWRNORM;
 593                         }
 594                 } else
 595                         mask |= EPOLLOUT | EPOLLWRNORM;
 596
 597                 if (tp->urg_data & TCP_URG_VALID)
 598                         mask |= EPOLLPRI;
 599         } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
 600                 /* Active TCP fastopen socket with defer_connect
 601                  * Return EPOLLOUT so application can call write()
 602                  * in order for kernel to generate SYN+data
 603                  */
 604                 mask |= EPOLLOUT | EPOLLWRNORM;
 605         }
 606         /* This barrier is coupled with smp_wmb() in tcp_reset() */
 607         smp_rmb();
 608         if (sk->sk_err || !skb_queue_empty_lockless(&sk->sk_error_queue))
 609                 mask |= EPOLLERR;
 610
 611         return mask;
 612 }
 613 EXPORT_SYMBOL(tcp_poll);
 614
 615 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 616 {
 617         struct tcp_sock *tp = tcp_sk(sk);
 618         int answ;
 619         bool slow;
 620
 621         switch (cmd) {
 622         case SIOCINQ:
 623                 if (sk->sk_state == TCP_LISTEN)
 624                         return -EINVAL;
 625
 626                 slow = lock_sock_fast(sk);
 627                 answ = tcp_inq(sk);
 628                 unlock_sock_fast(sk, slow);
 629                 break;
 630         case SIOCATMARK:
 631                 answ = tp->urg_data && tp->urg_seq == READ_ONCE(tp->copied_seq);
 632                 break;
 633         case SIOCOUTQ:
 634                 if (sk->sk_state == TCP_LISTEN)
 635                         return -EINVAL;
 636
 637                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 638                         answ = 0;
 639                 else
 640                         answ = READ_ONCE(tp->write_seq) - tp->snd_una;
 641                 break;
 642         case SIOCOUTQNSD:
 643                 if (sk->sk_state == TCP_LISTEN)
 644                         return -EINVAL;
 645
 646                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 647                         answ = 0;
 648                 else
 649                         answ = READ_ONCE(tp->write_seq) - tp->snd_nxt;
 650                 break;
 651         default:
 652                 return -ENOIOCTLCMD;
 653         }
 654
 655         return put_user(answ, (int __user *)arg);
 656 }
 657 EXPORT_SYMBOL(tcp_ioctl);
 658
 659 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 660 {
 661         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 662         tp->pushed_seq = tp->write_seq;
 663 }
 664
 665 static inline bool forced_push(const struct tcp_sock *tp)
 666 {
 667         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 668 }
 669
 670 static void skb_entail(struct sock *sk, struct sk_buff *skb)
 671 {
 672         struct tcp_sock *tp = tcp_sk(sk);
 673         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 674
 675         skb->csum    = 0;
 676         tcb->seq     = tcb->end_seq = tp->write_seq;
 677         tcb->tcp_flags = TCPHDR_ACK;
 678         tcb->sacked  = 0;
 679         __skb_header_release(skb);
 680         tcp_add_write_queue_tail(sk, skb);
 681         sk->sk_wmem_queued += skb->truesize;
 682         sk_mem_charge(sk, skb->truesize);
 683         if (tp->nonagle & TCP_NAGLE_PUSH)
 684                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 685
 686         tcp_slow_start_after_idle_check(sk);
 687 }
 688
 689 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
 690 {
 691         if (flags & MSG_OOB)
 692                 tp->snd_up = tp->write_seq;
 693 }
 694
 695 /* If a not yet filled skb is pushed, do not send it if
 696  * we have data packets in Qdisc or NIC queues :
 697  * Because TX completion will happen shortly, it gives a chance
 698  * to coalesce future sendmsg() payload into this skb, without
 699  * need for a timer, and with no latency trade off.
 700  * As packets containing data payload have a bigger truesize
 701  * than pure acks (dataless) packets, the last checks prevent
 702  * autocorking if we only have an ACK in Qdisc/NIC queues,
 703  * or if TX completion was delayed after we processed ACK packet.
 704  */
 705 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
 706                                 int size_goal)
 707 {
 708         return skb->len < size_goal &&
 709                sock_net(sk)->ipv4.sysctl_tcp_autocorking &&
 710                !tcp_rtx_queue_empty(sk) &&
 711                refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
 712 }
 713
 714 static void tcp_push(struct sock *sk, int flags, int mss_now,
 715                      int nonagle, int size_goal)
 716 {
 717         struct tcp_sock *tp = tcp_sk(sk);
 718         struct sk_buff *skb;
 719
 720         skb = tcp_write_queue_tail(sk);
 721         if (!skb)
 722                 return;
 723         if (!(flags & MSG_MORE) || forced_push(tp))
 724                 tcp_mark_push(tp, skb);
 725
 726         tcp_mark_urg(tp, flags);
 727
 728         if (tcp_should_autocork(sk, skb, size_goal)) {
 729
 730                 /* avoid atomic op if TSQ_THROTTLED bit is already set */
 731                 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
 732                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
 733                         set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
 734                 }
 735                 /* It is possible TX completion already happened
 736                  * before we set TSQ_THROTTLED.
 737                  */
 738                 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
 739                         return;
 740         }
 741
 742         if (flags & MSG_MORE)
 743                 nonagle = TCP_NAGLE_CORK;
 744
 745         __tcp_push_pending_frames(sk, mss_now, nonagle);
 746 }
 747
 748 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 749                                 unsigned int offset, size_t len)
 750 {
 751         struct tcp_splice_state *tss = rd_desc->arg.data;
 752         int ret;
 753
 754         ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
 755                               min(rd_desc->count, len), tss->flags);
 756         if (ret > 0)
 757                 rd_desc->count -= ret;
 758         return ret;
 759 }
 760
 761 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 762 {
 763         /* Store TCP splice context information in read_descriptor_t. */
 764         read_descriptor_t rd_desc = {
 765                 .arg.data = tss,
 766                 .count    = tss->len,
 767         };
 768
 769         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 770 }
 771
 772 /**
 773  *  tcp_splice_read - splice data from TCP socket to a pipe
 774  * @sock:       socket to splice from
 775  * @ppos:       position (not valid)
 776  * @pipe:       pipe to splice to
 777  * @len:        number of bytes to splice
 778  * @flags:      splice modifier flags
 779  *
 780  * Description:
 781  *    Will read pages from given socket and fill them into a pipe.
 782  *
 783  **/
 784 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 785                         struct pipe_inode_info *pipe, size_t len,
 786                         unsigned int flags)
 787 {
 788         struct sock *sk = sock->sk;
 789         struct tcp_splice_state tss = {
 790                 .pipe = pipe,
 791                 .len = len,
 792                 .flags = flags,
 793         };
 794         long timeo;
 795         ssize_t spliced;
 796         int ret;
 797
 798         sock_rps_record_flow(sk);
 799         /*
 800          * We can't seek on a socket input
 801          */
 802         if (unlikely(*ppos))
 803                 return -ESPIPE;
 804
 805         ret = spliced = 0;
 806
 807         lock_sock(sk);
 808
 809         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 810         while (tss.len) {
 811                 ret = __tcp_splice_read(sk, &tss);
 812                 if (ret < 0)
 813                         break;
 814                 else if (!ret) {
 815                         if (spliced)
 816                                 break;
 817                         if (sock_flag(sk, SOCK_DONE))
 818                                 break;
 819                         if (sk->sk_err) {
 820                                 ret = sock_error(sk);
 821                                 break;
 822                         }
 823                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 824                                 break;
 825                         if (sk->sk_state == TCP_CLOSE) {
 826                                 /*
 827                                  * This occurs when user tries to read
 828                                  * from never connected socket.
 829                                  */
 830                                 ret = -ENOTCONN;
 831                                 break;
 832                         }
 833                         if (!timeo) {
 834                                 ret = -EAGAIN;
 835                                 break;
 836                         }
 837                         /* if __tcp_splice_read() got nothing while we have
 838                          * an skb in receive queue, we do not want to loop.
 839                          * This might happen with URG data.
 840                          */
 841                         if (!skb_queue_empty(&sk->sk_receive_queue))
 842                                 break;
 843                         sk_wait_data(sk, &timeo, NULL);
 844                         if (signal_pending(current)) {
 845                                 ret = sock_intr_errno(timeo);
 846                                 break;
 847                         }
 848                         continue;
 849                 }
 850                 tss.len -= ret;
 851                 spliced += ret;
 852
 853                 if (!timeo)
 854                         break;
 855                 release_sock(sk);
 856                 lock_sock(sk);
 857
 858                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 859                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 860                     signal_pending(current))
 861                         break;
 862         }
 863
 864         release_sock(sk);
 865
 866         if (spliced)
 867                 return spliced;
 868
 869         return ret;
 870 }
 871 EXPORT_SYMBOL(tcp_splice_read);
 872
 873 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
 874                                     bool force_schedule)
 875 {
 876         struct sk_buff *skb;
 877
 878         /* The TCP header must be at least 32-bit aligned.  */
 879         size = ALIGN(size, 4);
 880
 881         if (unlikely(tcp_under_memory_pressure(sk)))
 882                 sk_mem_reclaim_partial(sk);
 883
 884         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 885         if (likely(skb)) {
 886                 bool mem_scheduled;
 887
 888                 if (force_schedule) {
 889                         mem_scheduled = true;
 890                         sk_forced_mem_schedule(sk, skb->truesize);
 891                 } else {
 892                         mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
 893                 }
 894                 if (likely(mem_scheduled)) {
 895                         skb_reserve(skb, sk->sk_prot->max_header);
 896                         /*
 897                          * Make sure that we have exactly size bytes
 898                          * available to the caller, no more, no less.
 899                          */
 900                         skb->reserved_tailroom = skb->end - skb->tail - size;
 901                         INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
 902                         return skb;
 903                 }
 904                 __kfree_skb(skb);
 905         } else {
 906                 sk->sk_prot->enter_memory_pressure(sk);
 907                 sk_stream_moderate_sndbuf(sk);
 908         }
 909         return NULL;
 910 }
 911
 912 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 913                                        int large_allowed)
 914 {
 915         struct tcp_sock *tp = tcp_sk(sk);
 916         u32 new_size_goal, size_goal;
 917
 918         if (!large_allowed)
 919                 return mss_now;
 920
 921         /* Note : tcp_tso_autosize() will eventually split this later */
 922         new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
 923         new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
 924
 925         /* We try hard to avoid divides here */
 926         size_goal = tp->gso_segs * mss_now;
 927         if (unlikely(new_size_goal < size_goal ||
 928                      new_size_goal >= size_goal + mss_now)) {
 929                 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
 930                                      sk->sk_gso_max_segs);
 931                 size_goal = tp->gso_segs * mss_now;
 932         }
 933
 934         return max(size_goal, mss_now);
 935 }
 936
 937 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 938 {
 939         int mss_now;
 940
 941         mss_now = tcp_current_mss(sk);
 942         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 943
 944         return mss_now;
 945 }
 946
 947 /* In some cases, both sendpage() and sendmsg() could have added
 948  * an skb to the write queue, but failed adding payload on it.
 949  * We need to remove it to consume less memory, but more
 950  * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
 951  * users.
 952  */
 953 static void tcp_remove_empty_skb(struct sock *sk, struct sk_buff *skb)
 954 {
 955         if (skb && !skb->len) {
 956                 tcp_unlink_write_queue(skb, sk);
 957                 if (tcp_write_queue_empty(sk))
 958                         tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
 959                 sk_wmem_free_skb(sk, skb);
 960         }
 961 }
 962
 963 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
 964                          size_t size, int flags)
 965 {
 966         struct tcp_sock *tp = tcp_sk(sk);
 967         int mss_now, size_goal;
 968         int err;
 969         ssize_t copied;
 970         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 971
 972         /* Wait for a connection to finish. One exception is TCP Fast Open
 973          * (passive side) where data is allowed to be sent before a connection
 974          * is fully established.
 975          */
 976         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
 977             !tcp_passive_fastopen(sk)) {
 978                 err = sk_stream_wait_connect(sk, &timeo);
 979                 if (err != 0)
 980                         goto out_err;
 981         }
 982
 983         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 984
 985         mss_now = tcp_send_mss(sk, &size_goal, flags);
 986         copied = 0;
 987
 988         err = -EPIPE;
 989         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 990                 goto out_err;
 991
 992         while (size > 0) {
 993                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 994                 int copy, i;
 995                 bool can_coalesce;
 996
 997                 if (!skb || (copy = size_goal - skb->len) <= 0 ||
 998                     !tcp_skb_can_collapse_to(skb)) {
 999 new_segment:
1000                         if (!sk_stream_memory_free(sk))
1001                                 goto wait_for_sndbuf;
1002
1003                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
1004                                         tcp_rtx_and_write_queues_empty(sk));
1005                         if (!skb)
1006                                 goto wait_for_memory;
1007
1008                         skb_entail(sk, skb);
1009                         copy = size_goal;
1010                 }
1011
1012                 if (copy > size)
1013                         copy = size;
1014
1015                 i = skb_shinfo(skb)->nr_frags;
1016                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
1017                 if (!can_coalesce && i >= sysctl_max_skb_frags) {
1018                         tcp_mark_push(tp, skb);
1019                         goto new_segment;
1020                 }
1021                 if (!sk_wmem_schedule(sk, copy))
1022                         goto wait_for_memory;
1023
1024                 if (can_coalesce) {
1025                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1026                 } else {
1027                         get_page(page);
1028                         skb_fill_page_desc(skb, i, page, offset, copy);
1029                 }
1030
1031                 if (!(flags & MSG_NO_SHARED_FRAGS))
1032                         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1033
1034                 skb->len += copy;
1035                 skb->data_len += copy;
1036                 skb->truesize += copy;
1037                 sk->sk_wmem_queued += copy;
1038                 sk_mem_charge(sk, copy);
1039                 skb->ip_summed = CHECKSUM_PARTIAL;
1040                 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1041                 TCP_SKB_CB(skb)->end_seq += copy;
1042                 tcp_skb_pcount_set(skb, 0);
1043
1044                 if (!copied)
1045                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1046
1047                 copied += copy;
1048                 offset += copy;
1049                 size -= copy;
1050                 if (!size)
1051                         goto out;
1052
1053                 if (skb->len < size_goal || (flags & MSG_OOB))
1054                         continue;
1055
1056                 if (forced_push(tp)) {
1057                         tcp_mark_push(tp, skb);
1058                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1059                 } else if (skb == tcp_send_head(sk))
1060                         tcp_push_one(sk, mss_now);
1061                 continue;
1062
1063 wait_for_sndbuf:
1064                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1065 wait_for_memory:
1066                 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1067                          TCP_NAGLE_PUSH, size_goal);
1068
1069                 err = sk_stream_wait_memory(sk, &timeo);
1070                 if (err != 0)
1071                         goto do_error;
1072
1073                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1074         }
1075
1076 out:
1077         if (copied) {
1078                 tcp_tx_timestamp(sk, sk->sk_tsflags);
1079                 if (!(flags & MSG_SENDPAGE_NOTLAST))
1080                         tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1081         }
1082         return copied;
1083
1084 do_error:
1085         tcp_remove_empty_skb(sk, tcp_write_queue_tail(sk));
1086         if (copied)
1087                 goto out;
1088 out_err:
1089         /* make sure we wake any epoll edge trigger waiter */
1090         if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1091                 sk->sk_write_space(sk);
1092                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1093         }
1094         return sk_stream_error(sk, flags, err);
1095 }
1096 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1097
1098 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1099                         size_t size, int flags)
1100 {
1101         if (!(sk->sk_route_caps & NETIF_F_SG))
1102                 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1103
1104         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1105
1106         return do_tcp_sendpages(sk, page, offset, size, flags);
1107 }
1108 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1109
1110 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1111                  size_t size, int flags)
1112 {
1113         int ret;
1114
1115         lock_sock(sk);
1116         ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1117         release_sock(sk);
1118
1119         return ret;
1120 }
1121 EXPORT_SYMBOL(tcp_sendpage);
1122
1123 /* Do not bother using a page frag for very small frames.
1124  * But use this heuristic only for the first skb in write queue.
1125  *
1126  * Having no payload in skb->head allows better SACK shifting
1127  * in tcp_shift_skb_data(), reducing sack/rack overhead, because
1128  * write queue has less skbs.
1129  * Each skb can hold up to MAX_SKB_FRAGS * 32Kbytes, or ~0.5 MB.
1130  * This also speeds up tso_fragment(), since it wont fallback
1131  * to tcp_fragment().
1132  */
1133 static int linear_payload_sz(bool first_skb)
1134 {
1135         if (first_skb)
1136                 return SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1137         return 0;
1138 }
1139
1140 static int select_size(bool first_skb, bool zc)
1141 {
1142         if (zc)
1143                 return 0;
1144         return linear_payload_sz(first_skb);
1145 }
1146
1147 void tcp_free_fastopen_req(struct tcp_sock *tp)
1148 {
1149         if (tp->fastopen_req) {
1150                 kfree(tp->fastopen_req);
1151                 tp->fastopen_req = NULL;
1152         }
1153 }
1154
1155 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1156                                 int *copied, size_t size)
1157 {
1158         struct tcp_sock *tp = tcp_sk(sk);
1159         struct inet_sock *inet = inet_sk(sk);
1160         struct sockaddr *uaddr = msg->msg_name;
1161         int err, flags;
1162
1163         if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1164             (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1165              uaddr->sa_family == AF_UNSPEC))
1166                 return -EOPNOTSUPP;
1167         if (tp->fastopen_req)
1168                 return -EALREADY; /* Another Fast Open is in progress */
1169
1170         tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1171                                    sk->sk_allocation);
1172         if (unlikely(!tp->fastopen_req))
1173                 return -ENOBUFS;
1174         tp->fastopen_req->data = msg;
1175         tp->fastopen_req->size = size;
1176
1177         if (inet->defer_connect) {
1178                 err = tcp_connect(sk);
1179                 /* Same failure procedure as in tcp_v4/6_connect */
1180                 if (err) {
1181                         tcp_set_state(sk, TCP_CLOSE);
1182                         inet->inet_dport = 0;
1183                         sk->sk_route_caps = 0;
1184                 }
1185         }
1186         flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1187         err = __inet_stream_connect(sk->sk_socket, uaddr,
1188                                     msg->msg_namelen, flags, 1);
1189         /* fastopen_req could already be freed in __inet_stream_connect
1190          * if the connection times out or gets rst
1191          */
1192         if (tp->fastopen_req) {
1193                 *copied = tp->fastopen_req->copied;
1194                 tcp_free_fastopen_req(tp);
1195                 inet->defer_connect = 0;
1196         }
1197         return err;
1198 }
1199
1200 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1201 {
1202         struct tcp_sock *tp = tcp_sk(sk);
1203         struct ubuf_info *uarg = NULL;
1204         struct sk_buff *skb;
1205         struct sockcm_cookie sockc;
1206         int flags, err, copied = 0;
1207         int mss_now = 0, size_goal, copied_syn = 0;
1208         bool process_backlog = false;
1209         bool zc = false;
1210         long timeo;
1211
1212         flags = msg->msg_flags;
1213
1214         if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
1215                 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1216                         err = -EINVAL;
1217                         goto out_err;
1218                 }
1219
1220                 skb = tcp_write_queue_tail(sk);
1221                 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
1222                 if (!uarg) {
1223                         err = -ENOBUFS;
1224                         goto out_err;
1225                 }
1226
1227                 zc = sk->sk_route_caps & NETIF_F_SG;
1228                 if (!zc)
1229                         uarg->zerocopy = 0;
1230         }
1231
1232         if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1233             !tp->repair) {
1234                 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1235                 if (err == -EINPROGRESS && copied_syn > 0)
1236                         goto out;
1237                 else if (err)
1238                         goto out_err;
1239         }
1240
1241         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1242
1243         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1244
1245         /* Wait for a connection to finish. One exception is TCP Fast Open
1246          * (passive side) where data is allowed to be sent before a connection
1247          * is fully established.
1248          */
1249         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1250             !tcp_passive_fastopen(sk)) {
1251                 err = sk_stream_wait_connect(sk, &timeo);
1252                 if (err != 0)
1253                         goto do_error;
1254         }
1255
1256         if (unlikely(tp->repair)) {
1257                 if (tp->repair_queue == TCP_RECV_QUEUE) {
1258                         copied = tcp_send_rcvq(sk, msg, size);
1259                         goto out_nopush;
1260                 }
1261
1262                 err = -EINVAL;
1263                 if (tp->repair_queue == TCP_NO_QUEUE)
1264                         goto out_err;
1265
1266                 /* 'common' sending to sendq */
1267         }
1268
1269         sockcm_init(&sockc, sk);
1270         if (msg->msg_controllen) {
1271                 err = sock_cmsg_send(sk, msg, &sockc);
1272                 if (unlikely(err)) {
1273                         err = -EINVAL;
1274                         goto out_err;
1275                 }
1276         }
1277
1278         /* This should be in poll */
1279         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1280
1281         /* Ok commence sending. */
1282         copied = 0;
1283
1284 restart:
1285         mss_now = tcp_send_mss(sk, &size_goal, flags);
1286
1287         err = -EPIPE;
1288         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1289                 goto do_error;
1290
1291         while (msg_data_left(msg)) {
1292                 int copy = 0;
1293
1294                 skb = tcp_write_queue_tail(sk);
1295                 if (skb)
1296                         copy = size_goal - skb->len;
1297
1298                 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1299                         bool first_skb;
1300                         int linear;
1301
1302 new_segment:
1303                         if (!sk_stream_memory_free(sk))
1304                                 goto wait_for_sndbuf;
1305
1306                         if (process_backlog && sk_flush_backlog(sk)) {
1307                                 process_backlog = false;
1308                                 goto restart;
1309                         }
1310                         first_skb = tcp_rtx_and_write_queues_empty(sk);
1311                         linear = select_size(first_skb, zc);
1312                         skb = sk_stream_alloc_skb(sk, linear, sk->sk_allocation,
1313                                                   first_skb);
1314                         if (!skb)
1315                                 goto wait_for_memory;
1316
1317                         process_backlog = true;
1318                         skb->ip_summed = CHECKSUM_PARTIAL;
1319
1320                         skb_entail(sk, skb);
1321                         copy = size_goal;
1322
1323                         /* All packets are restored as if they have
1324                          * already been sent. skb_mstamp isn't set to
1325                          * avoid wrong rtt estimation.
1326                          */
1327                         if (tp->repair)
1328                                 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1329                 }
1330
1331                 /* Try to append data to the end of skb. */
1332                 if (copy > msg_data_left(msg))
1333                         copy = msg_data_left(msg);
1334
1335                 /* Where to copy to? */
1336                 if (skb_availroom(skb) > 0 && !zc) {
1337                         /* We have some space in skb head. Superb! */
1338                         copy = min_t(int, copy, skb_availroom(skb));
1339                         err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1340                         if (err)
1341                                 goto do_fault;
1342                 } else if (!zc) {
1343                         bool merge = true;
1344                         int i = skb_shinfo(skb)->nr_frags;
1345                         struct page_frag *pfrag = sk_page_frag(sk);
1346
1347                         if (!sk_page_frag_refill(sk, pfrag))
1348                                 goto wait_for_memory;
1349
1350                         if (!skb_can_coalesce(skb, i, pfrag->page,
1351                                               pfrag->offset)) {
1352                                 if (i >= sysctl_max_skb_frags) {
1353                                         tcp_mark_push(tp, skb);
1354                                         goto new_segment;
1355                                 }
1356                                 merge = false;
1357                         }
1358
1359                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1360
1361                         if (!sk_wmem_schedule(sk, copy))
1362                                 goto wait_for_memory;
1363
1364                         err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1365                                                        pfrag->page,
1366                                                        pfrag->offset,
1367                                                        copy);
1368                         if (err)
1369                                 goto do_error;
1370
1371                         /* Update the skb. */
1372                         if (merge) {
1373                                 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1374                         } else {
1375                                 skb_fill_page_desc(skb, i, pfrag->page,
1376                                                    pfrag->offset, copy);
1377                                 page_ref_inc(pfrag->page);
1378                         }
1379                         pfrag->offset += copy;
1380                 } else {
1381                         err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1382                         if (err == -EMSGSIZE || err == -EEXIST) {
1383                                 tcp_mark_push(tp, skb);
1384                                 goto new_segment;
1385                         }
1386                         if (err < 0)
1387                                 goto do_error;
1388                         copy = err;
1389                 }
1390
1391                 if (!copied)
1392                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1393
1394                 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1395                 TCP_SKB_CB(skb)->end_seq += copy;
1396                 tcp_skb_pcount_set(skb, 0);
1397
1398                 copied += copy;
1399                 if (!msg_data_left(msg)) {
1400                         if (unlikely(flags & MSG_EOR))
1401                                 TCP_SKB_CB(skb)->eor = 1;
1402                         goto out;
1403                 }
1404
1405                 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1406                         continue;
1407
1408                 if (forced_push(tp)) {
1409                         tcp_mark_push(tp, skb);
1410                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1411                 } else if (skb == tcp_send_head(sk))
1412                         tcp_push_one(sk, mss_now);
1413                 continue;
1414
1415 wait_for_sndbuf:
1416                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1417 wait_for_memory:
1418                 if (copied)
1419                         tcp_push(sk, flags & ~MSG_MORE, mss_now,
1420                                  TCP_NAGLE_PUSH, size_goal);
1421
1422                 err = sk_stream_wait_memory(sk, &timeo);
1423                 if (err != 0)
1424                         goto do_error;
1425
1426                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1427         }
1428
1429 out:
1430         if (copied) {
1431                 tcp_tx_timestamp(sk, sockc.tsflags);
1432                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1433         }
1434 out_nopush:
1435         sock_zerocopy_put(uarg);
1436         return copied + copied_syn;
1437
1438 do_error:
1439         skb = tcp_write_queue_tail(sk);
1440 do_fault:
1441         tcp_remove_empty_skb(sk, skb);
1442
1443         if (copied + copied_syn)
1444                 goto out;
1445 out_err:
1446         sock_zerocopy_put_abort(uarg);
1447         err = sk_stream_error(sk, flags, err);
1448         /* make sure we wake any epoll edge trigger waiter */
1449         if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1450                 sk->sk_write_space(sk);
1451                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1452         }
1453         return err;
1454 }
1455 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1456
1457 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1458 {
1459         int ret;
1460
1461         lock_sock(sk);
1462         ret = tcp_sendmsg_locked(sk, msg, size);
1463         release_sock(sk);
1464
1465         return ret;
1466 }
1467 EXPORT_SYMBOL(tcp_sendmsg);
1468
1469 /*
1470  *      Handle reading urgent data. BSD has very simple semantics for
1471  *      this, no blocking and very strange errors 8)
1472  */
1473
1474 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1475 {
1476         struct tcp_sock *tp = tcp_sk(sk);
1477
1478         /* No URG data to read. */
1479         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1480             tp->urg_data == TCP_URG_READ)
1481                 return -EINVAL; /* Yes this is right ! */
1482
1483         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1484                 return -ENOTCONN;
1485
1486         if (tp->urg_data & TCP_URG_VALID) {
1487                 int err = 0;
1488                 char c = tp->urg_data;
1489
1490                 if (!(flags & MSG_PEEK))
1491                         tp->urg_data = TCP_URG_READ;
1492
1493                 /* Read urgent data. */
1494                 msg->msg_flags |= MSG_OOB;
1495
1496                 if (len > 0) {
1497                         if (!(flags & MSG_TRUNC))
1498                                 err = memcpy_to_msg(msg, &c, 1);
1499                         len = 1;
1500                 } else
1501                         msg->msg_flags |= MSG_TRUNC;
1502
1503                 return err ? -EFAULT : len;
1504         }
1505
1506         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1507                 return 0;
1508
1509         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1510          * the available implementations agree in this case:
1511          * this call should never block, independent of the
1512          * blocking state of the socket.
1513          * Mike <pall@rz.uni-karlsruhe.de>
1514          */
1515         return -EAGAIN;
1516 }
1517
1518 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1519 {
1520         struct sk_buff *skb;
1521         int copied = 0, err = 0;
1522
1523         /* XXX -- need to support SO_PEEK_OFF */
1524
1525         skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1526                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1527                 if (err)
1528                         return err;
1529                 copied += skb->len;
1530         }
1531
1532         skb_queue_walk(&sk->sk_write_queue, skb) {
1533                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1534                 if (err)
1535                         break;
1536
1537                 copied += skb->len;
1538         }
1539
1540         return err ?: copied;
1541 }
1542
1543 /* Clean up the receive buffer for full frames taken by the user,
1544  * then send an ACK if necessary.  COPIED is the number of bytes
1545  * tcp_recvmsg has given to the user so far, it speeds up the
1546  * calculation of whether or not we must ACK for the sake of
1547  * a window update.
1548  */
1549 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1550 {
1551         struct tcp_sock *tp = tcp_sk(sk);
1552         bool time_to_ack = false;
1553
1554         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1555
1556         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1557              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1558              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1559
1560         if (inet_csk_ack_scheduled(sk)) {
1561                 const struct inet_connection_sock *icsk = inet_csk(sk);
1562                    /* Delayed ACKs frequently hit locked sockets during bulk
1563                     * receive. */
1564                 if (icsk->icsk_ack.blocked ||
1565                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1566                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1567                     /*
1568                      * If this read emptied read buffer, we send ACK, if
1569                      * connection is not bidirectional, user drained
1570                      * receive buffer and there was a small segment
1571                      * in queue.
1572                      */
1573                     (copied > 0 &&
1574                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1575                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1576                        !icsk->icsk_ack.pingpong)) &&
1577                       !atomic_read(&sk->sk_rmem_alloc)))
1578                         time_to_ack = true;
1579         }
1580
1581         /* We send an ACK if we can now advertise a non-zero window
1582          * which has been raised "significantly".
1583          *
1584          * Even if window raised up to infinity, do not send window open ACK
1585          * in states, where we will not receive more. It is useless.
1586          */
1587         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1588                 __u32 rcv_window_now = tcp_receive_window(tp);
1589
1590                 /* Optimize, __tcp_select_window() is not cheap. */
1591                 if (2*rcv_window_now <= tp->window_clamp) {
1592                         __u32 new_window = __tcp_select_window(sk);
1593
1594                         /* Send ACK now, if this read freed lots of space
1595                          * in our buffer. Certainly, new_window is new window.
1596                          * We can advertise it now, if it is not less than current one.
1597                          * "Lots" means "at least twice" here.
1598                          */
1599                         if (new_window && new_window >= 2 * rcv_window_now)
1600                                 time_to_ack = true;
1601                 }
1602         }
1603         if (time_to_ack)
1604                 tcp_send_ack(sk);
1605 }
1606
1607 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1608 {
1609         struct sk_buff *skb;
1610         u32 offset;
1611
1612         while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1613                 offset = seq - TCP_SKB_CB(skb)->seq;
1614                 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1615                         pr_err_once("%s: found a SYN, please report !\n", __func__);
1616                         offset--;
1617                 }
1618                 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1619                         *off = offset;
1620                         return skb;
1621                 }
1622                 /* This looks weird, but this can happen if TCP collapsing
1623                  * splitted a fat GRO packet, while we released socket lock
1624                  * in skb_splice_bits()
1625                  */
1626                 sk_eat_skb(sk, skb);
1627         }
1628         return NULL;
1629 }
1630
1631 /*
1632  * This routine provides an alternative to tcp_recvmsg() for routines
1633  * that would like to handle copying from skbuffs directly in 'sendfile'
1634  * fashion.
1635  * Note:
1636  *      - It is assumed that the socket was locked by the caller.
1637  *      - The routine does not block.
1638  *      - At present, there is no support for reading OOB data
1639  *        or for 'peeking' the socket using this routine
1640  *        (although both would be easy to implement).
1641  */
1642 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1643                   sk_read_actor_t recv_actor)
1644 {
1645         struct sk_buff *skb;
1646         struct tcp_sock *tp = tcp_sk(sk);
1647         u32 seq = tp->copied_seq;
1648         u32 offset;
1649         int copied = 0;
1650
1651         if (sk->sk_state == TCP_LISTEN)
1652                 return -ENOTCONN;
1653         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1654                 if (offset < skb->len) {
1655                         int used;
1656                         size_t len;
1657
1658                         len = skb->len - offset;
1659                         /* Stop reading if we hit a patch of urgent data */
1660                         if (tp->urg_data) {
1661                                 u32 urg_offset = tp->urg_seq - seq;
1662                                 if (urg_offset < len)
1663                                         len = urg_offset;
1664                                 if (!len)
1665                                         break;
1666                         }
1667                         used = recv_actor(desc, skb, offset, len);
1668                         if (used <= 0) {
1669                                 if (!copied)
1670                                         copied = used;
1671                                 break;
1672                         } else if (used <= len) {
1673                                 seq += used;
1674                                 copied += used;
1675                                 offset += used;
1676                         }
1677                         /* If recv_actor drops the lock (e.g. TCP splice
1678                          * receive) the skb pointer might be invalid when
1679                          * getting here: tcp_collapse might have deleted it
1680                          * while aggregating skbs from the socket queue.
1681                          */
1682                         skb = tcp_recv_skb(sk, seq - 1, &offset);
1683                         if (!skb)
1684                                 break;
1685                         /* TCP coalescing might have appended data to the skb.
1686                          * Try to splice more frags
1687                          */
1688                         if (offset + 1 != skb->len)
1689                                 continue;
1690                 }
1691                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1692                         sk_eat_skb(sk, skb);
1693                         ++seq;
1694                         break;
1695                 }
1696                 sk_eat_skb(sk, skb);
1697                 if (!desc->count)
1698                         break;
1699                 WRITE_ONCE(tp->copied_seq, seq);
1700         }
1701         WRITE_ONCE(tp->copied_seq, seq);
1702
1703         tcp_rcv_space_adjust(sk);
1704
1705         /* Clean up data we have read: This will do ACK frames. */
1706         if (copied > 0) {
1707                 tcp_recv_skb(sk, seq, &offset);
1708                 tcp_cleanup_rbuf(sk, copied);
1709         }
1710         return copied;
1711 }
1712 EXPORT_SYMBOL(tcp_read_sock);
1713
1714 int tcp_peek_len(struct socket *sock)
1715 {
1716         return tcp_inq(sock->sk);
1717 }
1718 EXPORT_SYMBOL(tcp_peek_len);
1719
1720 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1721 int tcp_set_rcvlowat(struct sock *sk, int val)
1722 {
1723         int cap;
1724
1725         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1726                 cap = sk->sk_rcvbuf >> 1;
1727         else
1728                 cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1;
1729         val = min(val, cap);
1730         sk->sk_rcvlowat = val ? : 1;
1731
1732         /* Check if we need to signal EPOLLIN right now */
1733         tcp_data_ready(sk);
1734
1735         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1736                 return 0;
1737
1738         val <<= 1;
1739         if (val > sk->sk_rcvbuf) {
1740                 sk->sk_rcvbuf = val;
1741                 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1742         }
1743         return 0;
1744 }
1745 EXPORT_SYMBOL(tcp_set_rcvlowat);
1746
1747 #ifdef CONFIG_MMU
1748 static const struct vm_operations_struct tcp_vm_ops = {
1749 };
1750
1751 int tcp_mmap(struct file *file, struct socket *sock,
1752              struct vm_area_struct *vma)
1753 {
1754         if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1755                 return -EPERM;
1756         vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1757
1758         /* Instruct vm_insert_page() to not down_read(mmap_sem) */
1759         vma->vm_flags |= VM_MIXEDMAP;
1760
1761         vma->vm_ops = &tcp_vm_ops;
1762         return 0;
1763 }
1764 EXPORT_SYMBOL(tcp_mmap);
1765
1766 static int tcp_zerocopy_receive(struct sock *sk,
1767                                 struct tcp_zerocopy_receive *zc)
1768 {
1769         unsigned long address = (unsigned long)zc->address;
1770         const skb_frag_t *frags = NULL;
1771         u32 length = 0, seq, offset;
1772         struct vm_area_struct *vma;
1773         struct sk_buff *skb = NULL;
1774         struct tcp_sock *tp;
1775         int ret;
1776
1777         if (address & (PAGE_SIZE - 1) || address != zc->address)
1778                 return -EINVAL;
1779
1780         if (sk->sk_state == TCP_LISTEN)
1781                 return -ENOTCONN;
1782
1783         sock_rps_record_flow(sk);
1784
1785         down_read(&current->mm->mmap_sem);
1786
1787         vma = find_vma(current->mm, address);
1788         if (!vma || vma->vm_start > address || vma->vm_ops != &tcp_vm_ops) {
1789                 up_read(&current->mm->mmap_sem);
1790                 return -EINVAL;
1791         }
1792         zc->length = min_t(unsigned long, zc->length, vma->vm_end - address);
1793
1794         tp = tcp_sk(sk);
1795         seq = tp->copied_seq;
1796         zc->length = min_t(u32, zc->length, tcp_inq(sk));
1797         zc->length &= ~(PAGE_SIZE - 1);
1798
1799         zap_page_range(vma, address, zc->length);
1800
1801         zc->recv_skip_hint = 0;
1802         ret = 0;
1803         while (length + PAGE_SIZE <= zc->length) {
1804                 if (zc->recv_skip_hint < PAGE_SIZE) {
1805                         if (skb) {
1806                                 skb = skb->next;
1807                                 offset = seq - TCP_SKB_CB(skb)->seq;
1808                         } else {
1809                                 skb = tcp_recv_skb(sk, seq, &offset);
1810                         }
1811
1812                         zc->recv_skip_hint = skb->len - offset;
1813                         offset -= skb_headlen(skb);
1814                         if ((int)offset < 0 || skb_has_frag_list(skb))
1815                                 break;
1816                         frags = skb_shinfo(skb)->frags;
1817                         while (offset) {
1818                                 if (frags->size > offset)
1819                                         goto out;
1820                                 offset -= frags->size;
1821                                 frags++;
1822                         }
1823                 }
1824                 if (frags->size != PAGE_SIZE || frags->page_offset)
1825                         break;
1826                 ret = vm_insert_page(vma, address + length,
1827                                      skb_frag_page(frags));
1828                 if (ret)
1829                         break;
1830                 length += PAGE_SIZE;
1831                 seq += PAGE_SIZE;
1832                 zc->recv_skip_hint -= PAGE_SIZE;
1833                 frags++;
1834         }
1835 out:
1836         up_read(&current->mm->mmap_sem);
1837         if (length) {
1838                 WRITE_ONCE(tp->copied_seq, seq);
1839                 tcp_rcv_space_adjust(sk);
1840
1841                 /* Clean up data we have read: This will do ACK frames. */
1842                 tcp_recv_skb(sk, seq, &offset);
1843                 tcp_cleanup_rbuf(sk, length);
1844                 ret = 0;
1845                 if (length == zc->length)
1846                         zc->recv_skip_hint = 0;
1847         } else {
1848                 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
1849                         ret = -EIO;
1850         }
1851         zc->length = length;
1852         return ret;
1853 }
1854 #endif
1855
1856 static void tcp_update_recv_tstamps(struct sk_buff *skb,
1857                                     struct scm_timestamping *tss)
1858 {
1859         if (skb->tstamp)
1860                 tss->ts[0] = ktime_to_timespec(skb->tstamp);
1861         else
1862                 tss->ts[0] = (struct timespec) {0};
1863
1864         if (skb_hwtstamps(skb)->hwtstamp)
1865                 tss->ts[2] = ktime_to_timespec(skb_hwtstamps(skb)->hwtstamp);
1866         else
1867                 tss->ts[2] = (struct timespec) {0};
1868 }
1869
1870 /* Similar to __sock_recv_timestamp, but does not require an skb */
1871 static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
1872                                struct scm_timestamping *tss)
1873 {
1874         struct timeval tv;
1875         bool has_timestamping = false;
1876
1877         if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
1878                 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
1879                         if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
1880                                 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPNS,
1881                                          sizeof(tss->ts[0]), &tss->ts[0]);
1882                         } else {
1883                                 tv.tv_sec = tss->ts[0].tv_sec;
1884                                 tv.tv_usec = tss->ts[0].tv_nsec / 1000;
1885
1886                                 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMP,
1887                                          sizeof(tv), &tv);
1888                         }
1889                 }
1890
1891                 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
1892                         has_timestamping = true;
1893                 else
1894                         tss->ts[0] = (struct timespec) {0};
1895         }
1896
1897         if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
1898                 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
1899                         has_timestamping = true;
1900                 else
1901                         tss->ts[2] = (struct timespec) {0};
1902         }
1903
1904         if (has_timestamping) {
1905                 tss->ts[1] = (struct timespec) {0};
1906                 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPING,
1907                          sizeof(*tss), tss);
1908         }
1909 }
1910
1911 static int tcp_inq_hint(struct sock *sk)
1912 {
1913         const struct tcp_sock *tp = tcp_sk(sk);
1914         u32 copied_seq = READ_ONCE(tp->copied_seq);
1915         u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
1916         int inq;
1917
1918         inq = rcv_nxt - copied_seq;
1919         if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
1920                 lock_sock(sk);
1921                 inq = tp->rcv_nxt - tp->copied_seq;
1922                 release_sock(sk);
1923         }
1924         /* After receiving a FIN, tell the user-space to continue reading
1925          * by returning a non-zero inq.
1926          */
1927         if (inq == 0 && sock_flag(sk, SOCK_DONE))
1928                 inq = 1;
1929         return inq;
1930 }
1931
1932 /*
1933  *      This routine copies from a sock struct into the user buffer.
1934  *
1935  *      Technical note: in 2.3 we work on _locked_ socket, so that
1936  *      tricks with *seq access order and skb->users are not required.
1937  *      Probably, code can be easily improved even more.
1938  */
1939
1940 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1941                 int flags, int *addr_len)
1942 {
1943         struct tcp_sock *tp = tcp_sk(sk);
1944         int copied = 0;
1945         u32 peek_seq;
1946         u32 *seq;
1947         unsigned long used;
1948         int err, inq;
1949         int target;             /* Read at least this many bytes */
1950         long timeo;
1951         struct sk_buff *skb, *last;
1952         u32 urg_hole = 0;
1953         struct scm_timestamping tss;
1954         bool has_tss = false;
1955         bool has_cmsg;
1956
1957         if (unlikely(flags & MSG_ERRQUEUE))
1958                 return inet_recv_error(sk, msg, len, addr_len);
1959
1960         if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue) &&
1961             (sk->sk_state == TCP_ESTABLISHED))
1962                 sk_busy_loop(sk, nonblock);
1963
1964         lock_sock(sk);
1965
1966         err = -ENOTCONN;
1967         if (sk->sk_state == TCP_LISTEN)
1968                 goto out;
1969
1970         has_cmsg = tp->recvmsg_inq;
1971         timeo = sock_rcvtimeo(sk, nonblock);
1972
1973         /* Urgent data needs to be handled specially. */
1974         if (flags & MSG_OOB)
1975                 goto recv_urg;
1976
1977         if (unlikely(tp->repair)) {
1978                 err = -EPERM;
1979                 if (!(flags & MSG_PEEK))
1980                         goto out;
1981
1982                 if (tp->repair_queue == TCP_SEND_QUEUE)
1983                         goto recv_sndq;
1984
1985                 err = -EINVAL;
1986                 if (tp->repair_queue == TCP_NO_QUEUE)
1987                         goto out;
1988
1989                 /* 'common' recv queue MSG_PEEK-ing */
1990         }
1991
1992         seq = &tp->copied_seq;
1993         if (flags & MSG_PEEK) {
1994                 peek_seq = tp->copied_seq;
1995                 seq = &peek_seq;
1996         }
1997
1998         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1999
2000         do {
2001                 u32 offset;
2002
2003                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2004                 if (tp->urg_data && tp->urg_seq == *seq) {
2005                         if (copied)
2006                                 break;
2007                         if (signal_pending(current)) {
2008                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2009                                 break;
2010                         }
2011                 }
2012
2013                 /* Next get a buffer. */
2014
2015                 last = skb_peek_tail(&sk->sk_receive_queue);
2016                 skb_queue_walk(&sk->sk_receive_queue, skb) {
2017                         last = skb;
2018                         /* Now that we have two receive queues this
2019                          * shouldn't happen.
2020                          */
2021                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2022                                  "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2023                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2024                                  flags))
2025                                 break;
2026
2027                         offset = *seq - TCP_SKB_CB(skb)->seq;
2028                         if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2029                                 pr_err_once("%s: found a SYN, please report !\n", __func__);
2030                                 offset--;
2031                         }
2032                         if (offset < skb->len)
2033                                 goto found_ok_skb;
2034                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2035                                 goto found_fin_ok;
2036                         WARN(!(flags & MSG_PEEK),
2037                              "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2038                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2039                 }
2040
2041                 /* Well, if we have backlog, try to process it now yet. */
2042
2043                 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2044                         break;
2045
2046                 if (copied) {
2047                         if (sk->sk_err ||
2048                             sk->sk_state == TCP_CLOSE ||
2049                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
2050                             !timeo ||
2051                             signal_pending(current))
2052                                 break;
2053                 } else {
2054                         if (sock_flag(sk, SOCK_DONE))
2055                                 break;
2056
2057                         if (sk->sk_err) {
2058                                 copied = sock_error(sk);
2059                                 break;
2060                         }
2061
2062                         if (sk->sk_shutdown & RCV_SHUTDOWN)
2063                                 break;
2064
2065                         if (sk->sk_state == TCP_CLOSE) {
2066                                 /* This occurs when user tries to read
2067                                  * from never connected socket.
2068                                  */
2069                                 copied = -ENOTCONN;
2070                                 break;
2071                         }
2072
2073                         if (!timeo) {
2074                                 copied = -EAGAIN;
2075                                 break;
2076                         }
2077
2078                         if (signal_pending(current)) {
2079                                 copied = sock_intr_errno(timeo);
2080                                 break;
2081                         }
2082                 }
2083
2084                 tcp_cleanup_rbuf(sk, copied);
2085
2086                 if (copied >= target) {
2087                         /* Do not sleep, just process backlog. */
2088                         release_sock(sk);
2089                         lock_sock(sk);
2090                 } else {
2091                         sk_wait_data(sk, &timeo, last);
2092                 }
2093
2094                 if ((flags & MSG_PEEK) &&
2095                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
2096                         net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2097                                             current->comm,
2098                                             task_pid_nr(current));
2099                         peek_seq = tp->copied_seq;
2100                 }
2101                 continue;
2102
2103         found_ok_skb:
2104                 /* Ok so how much can we use? */
2105                 used = skb->len - offset;
2106                 if (len < used)
2107                         used = len;
2108
2109                 /* Do we have urgent data here? */
2110                 if (tp->urg_data) {
2111                         u32 urg_offset = tp->urg_seq - *seq;
2112                         if (urg_offset < used) {
2113                                 if (!urg_offset) {
2114                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
2115                                                 WRITE_ONCE(*seq, *seq + 1);
2116                                                 urg_hole++;
2117                                                 offset++;
2118                                                 used--;
2119                                                 if (!used)
2120                                                         goto skip_copy;
2121                                         }
2122                                 } else
2123                                         used = urg_offset;
2124                         }
2125                 }
2126
2127                 if (!(flags & MSG_TRUNC)) {
2128                         err = skb_copy_datagram_msg(skb, offset, msg, used);
2129                         if (err) {
2130                                 /* Exception. Bailout! */
2131                                 if (!copied)
2132                                         copied = -EFAULT;
2133                                 break;
2134                         }
2135                 }
2136
2137                 WRITE_ONCE(*seq, *seq + used);
2138                 copied += used;
2139                 len -= used;
2140
2141                 tcp_rcv_space_adjust(sk);
2142
2143 skip_copy:
2144                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
2145                         tp->urg_data = 0;
2146                         tcp_fast_path_check(sk);
2147                 }
2148
2149                 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2150                         tcp_update_recv_tstamps(skb, &tss);
2151                         has_tss = true;
2152                         has_cmsg = true;
2153                 }
2154
2155                 if (used + offset < skb->len)
2156                         continue;
2157
2158                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2159                         goto found_fin_ok;
2160                 if (!(flags & MSG_PEEK))
2161                         sk_eat_skb(sk, skb);
2162                 continue;
2163
2164         found_fin_ok:
2165                 /* Process the FIN. */
2166                 WRITE_ONCE(*seq, *seq + 1);
2167                 if (!(flags & MSG_PEEK))
2168                         sk_eat_skb(sk, skb);
2169                 break;
2170         } while (len > 0);
2171
2172         /* According to UNIX98, msg_name/msg_namelen are ignored
2173          * on connected socket. I was just happy when found this 8) --ANK
2174          */
2175
2176         /* Clean up data we have read: This will do ACK frames. */
2177         tcp_cleanup_rbuf(sk, copied);
2178
2179         release_sock(sk);
2180
2181         if (has_cmsg) {
2182                 if (has_tss)
2183                         tcp_recv_timestamp(msg, sk, &tss);
2184                 if (tp->recvmsg_inq) {
2185                         inq = tcp_inq_hint(sk);
2186                         put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2187                 }
2188         }
2189
2190         return copied;
2191
2192 out:
2193         release_sock(sk);
2194         return err;
2195
2196 recv_urg:
2197         err = tcp_recv_urg(sk, msg, len, flags);
2198         goto out;
2199
2200 recv_sndq:
2201         err = tcp_peek_sndq(sk, msg, len);
2202         goto out;
2203 }
2204 EXPORT_SYMBOL(tcp_recvmsg);
2205
2206 void tcp_set_state(struct sock *sk, int state)
2207 {
2208         int oldstate = sk->sk_state;
2209
2210         /* We defined a new enum for TCP states that are exported in BPF
2211          * so as not force the internal TCP states to be frozen. The
2212          * following checks will detect if an internal state value ever
2213          * differs from the BPF value. If this ever happens, then we will
2214          * need to remap the internal value to the BPF value before calling
2215          * tcp_call_bpf_2arg.
2216          */
2217         BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2218         BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2219         BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2220         BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2221         BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2222         BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2223         BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2224         BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2225         BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2226         BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2227         BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2228         BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2229         BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2230
2231         if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2232                 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2233
2234         switch (state) {
2235         case TCP_ESTABLISHED:
2236                 if (oldstate != TCP_ESTABLISHED)
2237                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2238                 break;
2239
2240         case TCP_CLOSE:
2241                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2242                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2243
2244                 sk->sk_prot->unhash(sk);
2245                 if (inet_csk(sk)->icsk_bind_hash &&
2246                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2247                         inet_put_port(sk);
2248                 /* fall through */
2249         default:
2250                 if (oldstate == TCP_ESTABLISHED)
2251                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2252         }
2253
2254         /* Change state AFTER socket is unhashed to avoid closed
2255          * socket sitting in hash tables.
2256          */
2257         inet_sk_state_store(sk, state);
2258
2259 #ifdef STATE_TRACE
2260         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
2261 #endif
2262 }
2263 EXPORT_SYMBOL_GPL(tcp_set_state);
2264
2265 /*
2266  *      State processing on a close. This implements the state shift for
2267  *      sending our FIN frame. Note that we only send a FIN for some
2268  *      states. A shutdown() may have already sent the FIN, or we may be
2269  *      closed.
2270  */
2271
2272 static const unsigned char new_state[16] = {
2273   /* current state:        new state:      action:      */
2274   [0 /* (Invalid) */]   = TCP_CLOSE,
2275   [TCP_ESTABLISHED]     = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2276   [TCP_SYN_SENT]        = TCP_CLOSE,
2277   [TCP_SYN_RECV]        = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2278   [TCP_FIN_WAIT1]       = TCP_FIN_WAIT1,
2279   [TCP_FIN_WAIT2]       = TCP_FIN_WAIT2,
2280   [TCP_TIME_WAIT]       = TCP_CLOSE,
2281   [TCP_CLOSE]           = TCP_CLOSE,
2282   [TCP_CLOSE_WAIT]      = TCP_LAST_ACK  | TCP_ACTION_FIN,
2283   [TCP_LAST_ACK]        = TCP_LAST_ACK,
2284   [TCP_LISTEN]          = TCP_CLOSE,
2285   [TCP_CLOSING]         = TCP_CLOSING,
2286   [TCP_NEW_SYN_RECV]    = TCP_CLOSE,    /* should not happen ! */
2287 };
2288
2289 static int tcp_close_state(struct sock *sk)
2290 {
2291         int next = (int)new_state[sk->sk_state];
2292         int ns = next & TCP_STATE_MASK;
2293
2294         tcp_set_state(sk, ns);
2295
2296         return next & TCP_ACTION_FIN;
2297 }
2298
2299 /*
2300  *      Shutdown the sending side of a connection. Much like close except
2301  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2302  */
2303
2304 void tcp_shutdown(struct sock *sk, int how)
2305 {
2306         /*      We need to grab some memory, and put together a FIN,
2307          *      and then put it into the queue to be sent.
2308          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2309          */
2310         if (!(how & SEND_SHUTDOWN))
2311                 return;
2312
2313         /* If we've already sent a FIN, or it's a closed state, skip this. */
2314         if ((1 << sk->sk_state) &
2315             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2316              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2317                 /* Clear out any half completed packets.  FIN if needed. */
2318                 if (tcp_close_state(sk))
2319                         tcp_send_fin(sk);
2320         }
2321 }
2322 EXPORT_SYMBOL(tcp_shutdown);
2323
2324 bool tcp_check_oom(struct sock *sk, int shift)
2325 {
2326         bool too_many_orphans, out_of_socket_memory;
2327
2328         too_many_orphans = tcp_too_many_orphans(sk, shift);
2329         out_of_socket_memory = tcp_out_of_memory(sk);
2330
2331         if (too_many_orphans)
2332                 net_info_ratelimited("too many orphaned sockets\n");
2333         if (out_of_socket_memory)
2334                 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2335         return too_many_orphans || out_of_socket_memory;
2336 }
2337
2338 void tcp_close(struct sock *sk, long timeout)
2339 {
2340         struct sk_buff *skb;
2341         int data_was_unread = 0;
2342         int state;
2343
2344         lock_sock(sk);
2345         sk->sk_shutdown = SHUTDOWN_MASK;
2346
2347         if (sk->sk_state == TCP_LISTEN) {
2348                 tcp_set_state(sk, TCP_CLOSE);
2349
2350                 /* Special case. */
2351                 inet_csk_listen_stop(sk);
2352
2353                 goto adjudge_to_death;
2354         }
2355
2356         /*  We need to flush the recv. buffs.  We do this only on the
2357          *  descriptor close, not protocol-sourced closes, because the
2358          *  reader process may not have drained the data yet!
2359          */
2360         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2361                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2362
2363                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2364                         len--;
2365                 data_was_unread += len;
2366                 __kfree_skb(skb);
2367         }
2368
2369         sk_mem_reclaim(sk);
2370
2371         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2372         if (sk->sk_state == TCP_CLOSE)
2373                 goto adjudge_to_death;
2374
2375         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2376          * data was lost. To witness the awful effects of the old behavior of
2377          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2378          * GET in an FTP client, suspend the process, wait for the client to
2379          * advertise a zero window, then kill -9 the FTP client, wheee...
2380          * Note: timeout is always zero in such a case.
2381          */
2382         if (unlikely(tcp_sk(sk)->repair)) {
2383                 sk->sk_prot->disconnect(sk, 0);
2384         } else if (data_was_unread) {
2385                 /* Unread data was tossed, zap the connection. */
2386                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2387                 tcp_set_state(sk, TCP_CLOSE);
2388                 tcp_send_active_reset(sk, sk->sk_allocation);
2389         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2390                 /* Check zero linger _after_ checking for unread data. */
2391                 sk->sk_prot->disconnect(sk, 0);
2392                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2393         } else if (tcp_close_state(sk)) {
2394                 /* We FIN if the application ate all the data before
2395                  * zapping the connection.
2396                  */
2397
2398                 /* RED-PEN. Formally speaking, we have broken TCP state
2399                  * machine. State transitions:
2400                  *
2401                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2402                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2403                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2404                  *
2405                  * are legal only when FIN has been sent (i.e. in window),
2406                  * rather than queued out of window. Purists blame.
2407                  *
2408                  * F.e. "RFC state" is ESTABLISHED,
2409                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2410                  *
2411                  * The visible declinations are that sometimes
2412                  * we enter time-wait state, when it is not required really
2413                  * (harmless), do not send active resets, when they are
2414                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2415                  * they look as CLOSING or LAST_ACK for Linux)
2416                  * Probably, I missed some more holelets.
2417                  *                                              --ANK
2418                  * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2419                  * in a single packet! (May consider it later but will
2420                  * probably need API support or TCP_CORK SYN-ACK until
2421                  * data is written and socket is closed.)
2422                  */
2423                 tcp_send_fin(sk);
2424         }
2425
2426         sk_stream_wait_close(sk, timeout);
2427
2428 adjudge_to_death:
2429         state = sk->sk_state;
2430         sock_hold(sk);
2431         sock_orphan(sk);
2432
2433         local_bh_disable();
2434         bh_lock_sock(sk);
2435         /* remove backlog if any, without releasing ownership. */
2436         __release_sock(sk);
2437
2438         percpu_counter_inc(sk->sk_prot->orphan_count);
2439
2440         /* Have we already been destroyed by a softirq or backlog? */
2441         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2442                 goto out;
2443
2444         /*      This is a (useful) BSD violating of the RFC. There is a
2445          *      problem with TCP as specified in that the other end could
2446          *      keep a socket open forever with no application left this end.
2447          *      We use a 1 minute timeout (about the same as BSD) then kill
2448          *      our end. If they send after that then tough - BUT: long enough
2449          *      that we won't make the old 4*rto = almost no time - whoops
2450          *      reset mistake.
2451          *
2452          *      Nope, it was not mistake. It is really desired behaviour
2453          *      f.e. on http servers, when such sockets are useless, but
2454          *      consume significant resources. Let's do it with special
2455          *      linger2 option.                                 --ANK
2456          */
2457
2458         if (sk->sk_state == TCP_FIN_WAIT2) {
2459                 struct tcp_sock *tp = tcp_sk(sk);
2460                 if (tp->linger2 < 0) {
2461                         tcp_set_state(sk, TCP_CLOSE);
2462                         tcp_send_active_reset(sk, GFP_ATOMIC);
2463                         __NET_INC_STATS(sock_net(sk),
2464                                         LINUX_MIB_TCPABORTONLINGER);
2465                 } else {
2466                         const int tmo = tcp_fin_time(sk);
2467
2468                         if (tmo > TCP_TIMEWAIT_LEN) {
2469                                 inet_csk_reset_keepalive_timer(sk,
2470                                                 tmo - TCP_TIMEWAIT_LEN);
2471                         } else {
2472                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2473                                 goto out;
2474                         }
2475                 }
2476         }
2477         if (sk->sk_state != TCP_CLOSE) {
2478                 sk_mem_reclaim(sk);
2479                 if (tcp_check_oom(sk, 0)) {
2480                         tcp_set_state(sk, TCP_CLOSE);
2481                         tcp_send_active_reset(sk, GFP_ATOMIC);
2482                         __NET_INC_STATS(sock_net(sk),
2483                                         LINUX_MIB_TCPABORTONMEMORY);
2484                 } else if (!check_net(sock_net(sk))) {
2485                         /* Not possible to send reset; just close */
2486                         tcp_set_state(sk, TCP_CLOSE);
2487                 }
2488         }
2489
2490         if (sk->sk_state == TCP_CLOSE) {
2491                 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2492                 /* We could get here with a non-NULL req if the socket is
2493                  * aborted (e.g., closed with unread data) before 3WHS
2494                  * finishes.
2495                  */
2496                 if (req)
2497                         reqsk_fastopen_remove(sk, req, false);
2498                 inet_csk_destroy_sock(sk);
2499         }
2500         /* Otherwise, socket is reprieved until protocol close. */
2501
2502 out:
2503         bh_unlock_sock(sk);
2504         local_bh_enable();
2505         release_sock(sk);
2506         sock_put(sk);
2507 }
2508 EXPORT_SYMBOL(tcp_close);
2509
2510 /* These states need RST on ABORT according to RFC793 */
2511
2512 static inline bool tcp_need_reset(int state)
2513 {
2514         return (1 << state) &
2515                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2516                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2517 }
2518
2519 static void tcp_rtx_queue_purge(struct sock *sk)
2520 {
2521         struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2522
2523         tcp_sk(sk)->highest_sack = NULL;
2524         while (p) {
2525                 struct sk_buff *skb = rb_to_skb(p);
2526
2527                 p = rb_next(p);
2528                 /* Since we are deleting whole queue, no need to
2529                  * list_del(&skb->tcp_tsorted_anchor)
2530                  */
2531                 tcp_rtx_queue_unlink(skb, sk);
2532                 sk_wmem_free_skb(sk, skb);
2533         }
2534 }
2535
2536 void tcp_write_queue_purge(struct sock *sk)
2537 {
2538         struct sk_buff *skb;
2539
2540         tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2541         while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2542                 tcp_skb_tsorted_anchor_cleanup(skb);
2543                 sk_wmem_free_skb(sk, skb);
2544         }
2545         tcp_rtx_queue_purge(sk);
2546         INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2547         sk_mem_reclaim(sk);
2548         tcp_clear_all_retrans_hints(tcp_sk(sk));
2549         tcp_sk(sk)->packets_out = 0;
2550         inet_csk(sk)->icsk_backoff = 0;
2551 }
2552
2553 int tcp_disconnect(struct sock *sk, int flags)
2554 {
2555         struct inet_sock *inet = inet_sk(sk);
2556         struct inet_connection_sock *icsk = inet_csk(sk);
2557         struct tcp_sock *tp = tcp_sk(sk);
2558         int old_state = sk->sk_state;
2559         u32 seq;
2560
2561         if (old_state != TCP_CLOSE)
2562                 tcp_set_state(sk, TCP_CLOSE);
2563
2564         /* ABORT function of RFC793 */
2565         if (old_state == TCP_LISTEN) {
2566                 inet_csk_listen_stop(sk);
2567         } else if (unlikely(tp->repair)) {
2568                 sk->sk_err = ECONNABORTED;
2569         } else if (tcp_need_reset(old_state) ||
2570                    (tp->snd_nxt != tp->write_seq &&
2571                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2572                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2573                  * states
2574                  */
2575                 tcp_send_active_reset(sk, gfp_any());
2576                 sk->sk_err = ECONNRESET;
2577         } else if (old_state == TCP_SYN_SENT)
2578                 sk->sk_err = ECONNRESET;
2579
2580         tcp_clear_xmit_timers(sk);
2581         __skb_queue_purge(&sk->sk_receive_queue);
2582         WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
2583         tp->urg_data = 0;
2584         tcp_write_queue_purge(sk);
2585         tcp_fastopen_active_disable_ofo_check(sk);
2586         skb_rbtree_purge(&tp->out_of_order_queue);
2587
2588         inet->inet_dport = 0;
2589
2590         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2591                 inet_reset_saddr(sk);
2592
2593         sk->sk_shutdown = 0;
2594         sock_reset_flag(sk, SOCK_DONE);
2595         tp->srtt_us = 0;
2596         tp->rcv_rtt_last_tsecr = 0;
2597
2598         seq = tp->write_seq + tp->max_window + 2;
2599         if (!seq)
2600                 seq = 1;
2601         WRITE_ONCE(tp->write_seq, seq);
2602
2603         tp->snd_cwnd = 2;
2604         icsk->icsk_probes_out = 0;
2605         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2606         tp->snd_cwnd_cnt = 0;
2607         tp->window_clamp = 0;
2608         tp->delivered = 0;
2609         tp->delivered_ce = 0;
2610         if (icsk->icsk_ca_ops->release)
2611                 icsk->icsk_ca_ops->release(sk);
2612         memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
2613         tcp_set_ca_state(sk, TCP_CA_Open);
2614         tp->is_sack_reneg = 0;
2615         tcp_clear_retrans(tp);
2616         tp->total_retrans = 0;
2617         inet_csk_delack_init(sk);
2618         /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2619          * issue in __tcp_select_window()
2620          */
2621         icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2622         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2623         __sk_dst_reset(sk);
2624         dst_release(sk->sk_rx_dst);
2625         sk->sk_rx_dst = NULL;
2626         tcp_saved_syn_free(tp);
2627         tp->compressed_ack = 0;
2628         tp->segs_in = 0;
2629         tp->segs_out = 0;
2630         tp->bytes_sent = 0;
2631         tp->bytes_acked = 0;
2632         tp->bytes_received = 0;
2633         tp->bytes_retrans = 0;
2634         tp->data_segs_in = 0;
2635         tp->data_segs_out = 0;
2636         tp->dsack_dups = 0;
2637         tp->reord_seen = 0;
2638
2639         /* Clean up fastopen related fields */
2640         tcp_free_fastopen_req(tp);
2641         inet->defer_connect = 0;
2642
2643         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2644
2645         if (sk->sk_frag.page) {
2646                 put_page(sk->sk_frag.page);
2647                 sk->sk_frag.page = NULL;
2648                 sk->sk_frag.offset = 0;
2649         }
2650
2651         sk->sk_error_report(sk);
2652         return 0;
2653 }
2654 EXPORT_SYMBOL(tcp_disconnect);
2655
2656 static inline bool tcp_can_repair_sock(const struct sock *sk)
2657 {
2658         return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2659                 (sk->sk_state != TCP_LISTEN);
2660 }
2661
2662 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2663 {
2664         struct tcp_repair_window opt;
2665
2666         if (!tp->repair)
2667                 return -EPERM;
2668
2669         if (len != sizeof(opt))
2670                 return -EINVAL;
2671
2672         if (copy_from_user(&opt, optbuf, sizeof(opt)))
2673                 return -EFAULT;
2674
2675         if (opt.max_window < opt.snd_wnd)
2676                 return -EINVAL;
2677
2678         if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2679                 return -EINVAL;
2680
2681         if (after(opt.rcv_wup, tp->rcv_nxt))
2682                 return -EINVAL;
2683
2684         tp->snd_wl1     = opt.snd_wl1;
2685         tp->snd_wnd     = opt.snd_wnd;
2686         tp->max_window  = opt.max_window;
2687
2688         tp->rcv_wnd     = opt.rcv_wnd;
2689         tp->rcv_wup     = opt.rcv_wup;
2690
2691         return 0;
2692 }
2693
2694 static int tcp_repair_options_est(struct sock *sk,
2695                 struct tcp_repair_opt __user *optbuf, unsigned int len)
2696 {
2697         struct tcp_sock *tp = tcp_sk(sk);
2698         struct tcp_repair_opt opt;
2699
2700         while (len >= sizeof(opt)) {
2701                 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2702                         return -EFAULT;
2703
2704                 optbuf++;
2705                 len -= sizeof(opt);
2706
2707                 switch (opt.opt_code) {
2708                 case TCPOPT_MSS:
2709                         tp->rx_opt.mss_clamp = opt.opt_val;
2710                         tcp_mtup_init(sk);
2711                         break;
2712                 case TCPOPT_WINDOW:
2713                         {
2714                                 u16 snd_wscale = opt.opt_val & 0xFFFF;
2715                                 u16 rcv_wscale = opt.opt_val >> 16;
2716
2717                                 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
2718                                         return -EFBIG;
2719
2720                                 tp->rx_opt.snd_wscale = snd_wscale;
2721                                 tp->rx_opt.rcv_wscale = rcv_wscale;
2722                                 tp->rx_opt.wscale_ok = 1;
2723                         }
2724                         break;
2725                 case TCPOPT_SACK_PERM:
2726                         if (opt.opt_val != 0)
2727                                 return -EINVAL;
2728
2729                         tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2730                         break;
2731                 case TCPOPT_TIMESTAMP:
2732                         if (opt.opt_val != 0)
2733                                 return -EINVAL;
2734
2735                         tp->rx_opt.tstamp_ok = 1;
2736                         break;
2737                 }
2738         }
2739
2740         return 0;
2741 }
2742
2743 /*
2744  *      Socket option code for TCP.
2745  */
2746 static int do_tcp_setsockopt(struct sock *sk, int level,
2747                 int optname, char __user *optval, unsigned int optlen)
2748 {
2749         struct tcp_sock *tp = tcp_sk(sk);
2750         struct inet_connection_sock *icsk = inet_csk(sk);
2751         struct net *net = sock_net(sk);
2752         int val;
2753         int err = 0;
2754
2755         /* These are data/string values, all the others are ints */
2756         switch (optname) {
2757         case TCP_CONGESTION: {
2758                 char name[TCP_CA_NAME_MAX];
2759
2760                 if (optlen < 1)
2761                         return -EINVAL;
2762
2763                 val = strncpy_from_user(name, optval,
2764                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2765                 if (val < 0)
2766                         return -EFAULT;
2767                 name[val] = 0;
2768
2769                 lock_sock(sk);
2770                 err = tcp_set_congestion_control(sk, name, true, true,
2771                                                  ns_capable(sock_net(sk)->user_ns,
2772                                                             CAP_NET_ADMIN));
2773                 release_sock(sk);
2774                 return err;
2775         }
2776         case TCP_ULP: {
2777                 char name[TCP_ULP_NAME_MAX];
2778
2779                 if (optlen < 1)
2780                         return -EINVAL;
2781
2782                 val = strncpy_from_user(name, optval,
2783                                         min_t(long, TCP_ULP_NAME_MAX - 1,
2784                                               optlen));
2785                 if (val < 0)
2786                         return -EFAULT;
2787                 name[val] = 0;
2788
2789                 lock_sock(sk);
2790                 err = tcp_set_ulp(sk, name);
2791                 release_sock(sk);
2792                 return err;
2793         }
2794         case TCP_FASTOPEN_KEY: {
2795                 __u8 key[TCP_FASTOPEN_KEY_LENGTH];
2796
2797                 if (optlen != sizeof(key))
2798                         return -EINVAL;
2799
2800                 if (copy_from_user(key, optval, optlen))
2801                         return -EFAULT;
2802
2803                 return tcp_fastopen_reset_cipher(net, sk, key, sizeof(key));
2804         }
2805         default:
2806                 /* fallthru */
2807                 break;
2808         }
2809
2810         if (optlen < sizeof(int))
2811                 return -EINVAL;
2812
2813         if (get_user(val, (int __user *)optval))
2814                 return -EFAULT;
2815
2816         lock_sock(sk);
2817
2818         switch (optname) {
2819         case TCP_MAXSEG:
2820                 /* Values greater than interface MTU won't take effect. However
2821                  * at the point when this call is done we typically don't yet
2822                  * know which interface is going to be used
2823                  */
2824                 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
2825                         err = -EINVAL;
2826                         break;
2827                 }
2828                 tp->rx_opt.user_mss = val;
2829                 break;
2830
2831         case TCP_NODELAY:
2832                 if (val) {
2833                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2834                          * this option on corked socket is remembered, but
2835                          * it is not activated until cork is cleared.
2836                          *
2837                          * However, when TCP_NODELAY is set we make
2838                          * an explicit push, which overrides even TCP_CORK
2839                          * for currently queued segments.
2840                          */
2841                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2842                         tcp_push_pending_frames(sk);
2843                 } else {
2844                         tp->nonagle &= ~TCP_NAGLE_OFF;
2845                 }
2846                 break;
2847
2848         case TCP_THIN_LINEAR_TIMEOUTS:
2849                 if (val < 0 || val > 1)
2850                         err = -EINVAL;
2851                 else
2852                         tp->thin_lto = val;
2853                 break;
2854
2855         case TCP_THIN_DUPACK:
2856                 if (val < 0 || val > 1)
2857                         err = -EINVAL;
2858                 break;
2859
2860         case TCP_REPAIR:
2861                 if (!tcp_can_repair_sock(sk))
2862                         err = -EPERM;
2863                 else if (val == TCP_REPAIR_ON) {
2864                         tp->repair = 1;
2865                         sk->sk_reuse = SK_FORCE_REUSE;
2866                         tp->repair_queue = TCP_NO_QUEUE;
2867                 } else if (val == TCP_REPAIR_OFF) {
2868                         tp->repair = 0;
2869                         sk->sk_reuse = SK_NO_REUSE;
2870                         tcp_send_window_probe(sk);
2871                 } else if (val == TCP_REPAIR_OFF_NO_WP) {
2872                         tp->repair = 0;
2873                         sk->sk_reuse = SK_NO_REUSE;
2874                 } else
2875                         err = -EINVAL;
2876
2877                 break;
2878
2879         case TCP_REPAIR_QUEUE:
2880                 if (!tp->repair)
2881                         err = -EPERM;
2882                 else if ((unsigned int)val < TCP_QUEUES_NR)
2883                         tp->repair_queue = val;
2884                 else
2885                         err = -EINVAL;
2886                 break;
2887
2888         case TCP_QUEUE_SEQ:
2889                 if (sk->sk_state != TCP_CLOSE) {
2890                         err = -EPERM;
2891                 } else if (tp->repair_queue == TCP_SEND_QUEUE) {
2892                         if (!tcp_rtx_queue_empty(sk))
2893                                 err = -EPERM;
2894                         else
2895                                 WRITE_ONCE(tp->write_seq, val);
2896                 } else if (tp->repair_queue == TCP_RECV_QUEUE) {
2897                         if (tp->rcv_nxt != tp->copied_seq) {
2898                                 err = -EPERM;
2899                         } else {
2900                                 WRITE_ONCE(tp->rcv_nxt, val);
2901                                 WRITE_ONCE(tp->copied_seq, val);
2902                         }
2903                 } else {
2904                         err = -EINVAL;
2905                 }
2906                 break;
2907
2908         case TCP_REPAIR_OPTIONS:
2909                 if (!tp->repair)
2910                         err = -EINVAL;
2911                 else if (sk->sk_state == TCP_ESTABLISHED)
2912                         err = tcp_repair_options_est(sk,
2913                                         (struct tcp_repair_opt __user *)optval,
2914                                         optlen);
2915                 else
2916                         err = -EPERM;
2917                 break;
2918
2919         case TCP_CORK:
2920                 /* When set indicates to always queue non-full frames.
2921                  * Later the user clears this option and we transmit
2922                  * any pending partial frames in the queue.  This is
2923                  * meant to be used alongside sendfile() to get properly
2924                  * filled frames when the user (for example) must write
2925                  * out headers with a write() call first and then use
2926                  * sendfile to send out the data parts.
2927                  *
2928                  * TCP_CORK can be set together with TCP_NODELAY and it is
2929                  * stronger than TCP_NODELAY.
2930                  */
2931                 if (val) {
2932                         tp->nonagle |= TCP_NAGLE_CORK;
2933                 } else {
2934                         tp->nonagle &= ~TCP_NAGLE_CORK;
2935                         if (tp->nonagle&TCP_NAGLE_OFF)
2936                                 tp->nonagle |= TCP_NAGLE_PUSH;
2937                         tcp_push_pending_frames(sk);
2938                 }
2939                 break;
2940
2941         case TCP_KEEPIDLE:
2942                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2943                         err = -EINVAL;
2944                 else {
2945                         tp->keepalive_time = val * HZ;
2946                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2947                             !((1 << sk->sk_state) &
2948                               (TCPF_CLOSE | TCPF_LISTEN))) {
2949                                 u32 elapsed = keepalive_time_elapsed(tp);
2950                                 if (tp->keepalive_time > elapsed)
2951                                         elapsed = tp->keepalive_time - elapsed;
2952                                 else
2953                                         elapsed = 0;
2954                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2955                         }
2956                 }
2957                 break;
2958         case TCP_KEEPINTVL:
2959                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2960                         err = -EINVAL;
2961                 else
2962                         tp->keepalive_intvl = val * HZ;
2963                 break;
2964         case TCP_KEEPCNT:
2965                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2966                         err = -EINVAL;
2967                 else
2968                         tp->keepalive_probes = val;
2969                 break;
2970         case TCP_SYNCNT:
2971                 if (val < 1 || val > MAX_TCP_SYNCNT)
2972                         err = -EINVAL;
2973                 else
2974                         icsk->icsk_syn_retries = val;
2975                 break;
2976
2977         case TCP_SAVE_SYN:
2978                 if (val < 0 || val > 1)
2979                         err = -EINVAL;
2980                 else
2981                         tp->save_syn = val;
2982                 break;
2983
2984         case TCP_LINGER2:
2985                 if (val < 0)
2986                         tp->linger2 = -1;
2987                 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
2988                         tp->linger2 = 0;
2989                 else
2990                         tp->linger2 = val * HZ;
2991                 break;
2992
2993         case TCP_DEFER_ACCEPT:
2994                 /* Translate value in seconds to number of retransmits */
2995                 icsk->icsk_accept_queue.rskq_defer_accept =
2996                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2997                                         TCP_RTO_MAX / HZ);
2998                 break;
2999
3000         case TCP_WINDOW_CLAMP:
3001                 if (!val) {
3002                         if (sk->sk_state != TCP_CLOSE) {
3003                                 err = -EINVAL;
3004                                 break;
3005                         }
3006                         tp->window_clamp = 0;
3007                 } else
3008                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3009                                                 SOCK_MIN_RCVBUF / 2 : val;
3010                 break;
3011
3012         case TCP_QUICKACK:
3013                 if (!val) {
3014                         icsk->icsk_ack.pingpong = 1;
3015                 } else {
3016                         icsk->icsk_ack.pingpong = 0;
3017                         if ((1 << sk->sk_state) &
3018                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3019                             inet_csk_ack_scheduled(sk)) {
3020                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
3021                                 tcp_cleanup_rbuf(sk, 1);
3022                                 if (!(val & 1))
3023                                         icsk->icsk_ack.pingpong = 1;
3024                         }
3025                 }
3026                 break;
3027
3028 #ifdef CONFIG_TCP_MD5SIG
3029         case TCP_MD5SIG:
3030         case TCP_MD5SIG_EXT:
3031                 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3032                 break;
3033 #endif
3034         case TCP_USER_TIMEOUT:
3035                 /* Cap the max time in ms TCP will retry or probe the window
3036                  * before giving up and aborting (ETIMEDOUT) a connection.
3037                  */
3038                 if (val < 0)
3039                         err = -EINVAL;
3040                 else
3041                         icsk->icsk_user_timeout = val;
3042                 break;
3043
3044         case TCP_FASTOPEN:
3045                 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3046                     TCPF_LISTEN))) {
3047                         tcp_fastopen_init_key_once(net);
3048
3049                         fastopen_queue_tune(sk, val);
3050                 } else {
3051                         err = -EINVAL;
3052                 }
3053                 break;
3054         case TCP_FASTOPEN_CONNECT:
3055                 if (val > 1 || val < 0) {
3056                         err = -EINVAL;
3057                 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
3058                         if (sk->sk_state == TCP_CLOSE)
3059                                 tp->fastopen_connect = val;
3060                         else
3061                                 err = -EINVAL;
3062                 } else {
3063                         err = -EOPNOTSUPP;
3064                 }
3065                 break;
3066         case TCP_FASTOPEN_NO_COOKIE:
3067                 if (val > 1 || val < 0)
3068                         err = -EINVAL;
3069                 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3070                         err = -EINVAL;
3071                 else
3072                         tp->fastopen_no_cookie = val;
3073                 break;
3074         case TCP_TIMESTAMP:
3075                 if (!tp->repair)
3076                         err = -EPERM;
3077                 else
3078                         tp->tsoffset = val - tcp_time_stamp_raw();
3079                 break;
3080         case TCP_REPAIR_WINDOW:
3081                 err = tcp_repair_set_window(tp, optval, optlen);
3082                 break;
3083         case TCP_NOTSENT_LOWAT:
3084                 tp->notsent_lowat = val;
3085                 sk->sk_write_space(sk);
3086                 break;
3087         case TCP_INQ:
3088                 if (val > 1 || val < 0)
3089                         err = -EINVAL;
3090                 else
3091                         tp->recvmsg_inq = val;
3092                 break;
3093         default:
3094                 err = -ENOPROTOOPT;
3095                 break;
3096         }
3097
3098         release_sock(sk);
3099         return err;
3100 }
3101
3102 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
3103                    unsigned int optlen)
3104 {
3105         const struct inet_connection_sock *icsk = inet_csk(sk);
3106
3107         if (level != SOL_TCP)
3108                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3109                                                      optval, optlen);
3110         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3111 }
3112 EXPORT_SYMBOL(tcp_setsockopt);
3113
3114 #ifdef CONFIG_COMPAT
3115 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
3116                           char __user *optval, unsigned int optlen)
3117 {
3118         if (level != SOL_TCP)
3119                 return inet_csk_compat_setsockopt(sk, level, optname,
3120                                                   optval, optlen);
3121         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3122 }
3123 EXPORT_SYMBOL(compat_tcp_setsockopt);
3124 #endif
3125
3126 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3127                                       struct tcp_info *info)
3128 {
3129         u64 stats[__TCP_CHRONO_MAX], total = 0;
3130         enum tcp_chrono i;
3131
3132         for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3133                 stats[i] = tp->chrono_stat[i - 1];
3134                 if (i == tp->chrono_type)
3135                         stats[i] += tcp_jiffies32 - tp->chrono_start;
3136                 stats[i] *= USEC_PER_SEC / HZ;
3137                 total += stats[i];
3138         }
3139
3140         info->tcpi_busy_time = total;
3141         info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3142         info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3143 }
3144
3145 /* Return information about state of tcp endpoint in API format. */
3146 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3147 {
3148         const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3149         const struct inet_connection_sock *icsk = inet_csk(sk);
3150         u32 now;
3151         u64 rate64;
3152         bool slow;
3153         u32 rate;
3154
3155         memset(info, 0, sizeof(*info));
3156         if (sk->sk_type != SOCK_STREAM)
3157                 return;
3158
3159         info->tcpi_state = inet_sk_state_load(sk);
3160
3161         /* Report meaningful fields for all TCP states, including listeners */
3162         rate = READ_ONCE(sk->sk_pacing_rate);
3163         rate64 = rate != ~0U ? rate : ~0ULL;
3164         info->tcpi_pacing_rate = rate64;
3165
3166         rate = READ_ONCE(sk->sk_max_pacing_rate);
3167         rate64 = rate != ~0U ? rate : ~0ULL;
3168         info->tcpi_max_pacing_rate = rate64;
3169
3170         info->tcpi_reordering = tp->reordering;
3171         info->tcpi_snd_cwnd = tp->snd_cwnd;
3172
3173         if (info->tcpi_state == TCP_LISTEN) {
3174                 /* listeners aliased fields :
3175                  * tcpi_unacked -> Number of children ready for accept()
3176                  * tcpi_sacked  -> max backlog
3177                  */
3178                 info->tcpi_unacked = sk->sk_ack_backlog;
3179                 info->tcpi_sacked = sk->sk_max_ack_backlog;
3180                 return;
3181         }
3182
3183         slow = lock_sock_fast(sk);
3184
3185         info->tcpi_ca_state = icsk->icsk_ca_state;
3186         info->tcpi_retransmits = icsk->icsk_retransmits;
3187         info->tcpi_probes = icsk->icsk_probes_out;
3188         info->tcpi_backoff = icsk->icsk_backoff;
3189
3190         if (tp->rx_opt.tstamp_ok)
3191                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3192         if (tcp_is_sack(tp))
3193                 info->tcpi_options |= TCPI_OPT_SACK;
3194         if (tp->rx_opt.wscale_ok) {
3195                 info->tcpi_options |= TCPI_OPT_WSCALE;
3196                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3197                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3198         }
3199
3200         if (tp->ecn_flags & TCP_ECN_OK)
3201                 info->tcpi_options |= TCPI_OPT_ECN;
3202         if (tp->ecn_flags & TCP_ECN_SEEN)
3203                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3204         if (tp->syn_data_acked)
3205                 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3206
3207         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3208         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3209         info->tcpi_snd_mss = tp->mss_cache;
3210         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3211
3212         info->tcpi_unacked = tp->packets_out;
3213         info->tcpi_sacked = tp->sacked_out;
3214
3215         info->tcpi_lost = tp->lost_out;
3216         info->tcpi_retrans = tp->retrans_out;
3217
3218         now = tcp_jiffies32;
3219         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3220         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3221         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3222
3223         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3224         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3225         info->tcpi_rtt = tp->srtt_us >> 3;
3226         info->tcpi_rttvar = tp->mdev_us >> 2;
3227         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3228         info->tcpi_advmss = tp->advmss;
3229
3230         info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3231         info->tcpi_rcv_space = tp->rcvq_space.space;
3232
3233         info->tcpi_total_retrans = tp->total_retrans;
3234
3235         info->tcpi_bytes_acked = tp->bytes_acked;
3236         info->tcpi_bytes_received = tp->bytes_received;
3237         info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3238         tcp_get_info_chrono_stats(tp, info);
3239
3240         info->tcpi_segs_out = tp->segs_out;
3241         info->tcpi_segs_in = tp->segs_in;
3242
3243         info->tcpi_min_rtt = tcp_min_rtt(tp);
3244         info->tcpi_data_segs_in = tp->data_segs_in;
3245         info->tcpi_data_segs_out = tp->data_segs_out;
3246
3247         info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3248         rate64 = tcp_compute_delivery_rate(tp);
3249         if (rate64)
3250                 info->tcpi_delivery_rate = rate64;
3251         info->tcpi_delivered = tp->delivered;
3252         info->tcpi_delivered_ce = tp->delivered_ce;
3253         info->tcpi_bytes_sent = tp->bytes_sent;
3254         info->tcpi_bytes_retrans = tp->bytes_retrans;
3255         info->tcpi_dsack_dups = tp->dsack_dups;
3256         info->tcpi_reord_seen = tp->reord_seen;
3257         unlock_sock_fast(sk, slow);
3258 }
3259 EXPORT_SYMBOL_GPL(tcp_get_info);
3260
3261 static size_t tcp_opt_stats_get_size(void)
3262 {
3263         return
3264                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3265                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3266                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3267                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3268                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3269                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3270                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3271                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3272                 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3273                 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3274                 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3275                 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3276                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3277                 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3278                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3279                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3280                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3281                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3282                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3283                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3284                 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3285                 0;
3286 }
3287
3288 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk)
3289 {
3290         const struct tcp_sock *tp = tcp_sk(sk);
3291         struct sk_buff *stats;
3292         struct tcp_info info;
3293         u64 rate64;
3294         u32 rate;
3295
3296         stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3297         if (!stats)
3298                 return NULL;
3299
3300         tcp_get_info_chrono_stats(tp, &info);
3301         nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3302                           info.tcpi_busy_time, TCP_NLA_PAD);
3303         nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3304                           info.tcpi_rwnd_limited, TCP_NLA_PAD);
3305         nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3306                           info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3307         nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3308                           tp->data_segs_out, TCP_NLA_PAD);
3309         nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3310                           tp->total_retrans, TCP_NLA_PAD);
3311
3312         rate = READ_ONCE(sk->sk_pacing_rate);
3313         rate64 = rate != ~0U ? rate : ~0ULL;
3314         nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3315
3316         rate64 = tcp_compute_delivery_rate(tp);
3317         nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3318
3319         nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3320         nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3321         nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3322
3323         nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3324         nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3325         nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3326         nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3327         nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3328
3329         nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3330         nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3331
3332         nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3333                           TCP_NLA_PAD);
3334         nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3335                           TCP_NLA_PAD);
3336         nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3337         nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3338
3339         return stats;
3340 }
3341
3342 static int do_tcp_getsockopt(struct sock *sk, int level,
3343                 int optname, char __user *optval, int __user *optlen)
3344 {
3345         struct inet_connection_sock *icsk = inet_csk(sk);
3346         struct tcp_sock *tp = tcp_sk(sk);
3347         struct net *net = sock_net(sk);
3348         int val, len;
3349
3350         if (get_user(len, optlen))
3351                 return -EFAULT;
3352
3353         len = min_t(unsigned int, len, sizeof(int));
3354
3355         if (len < 0)
3356                 return -EINVAL;
3357
3358         switch (optname) {
3359         case TCP_MAXSEG:
3360                 val = tp->mss_cache;
3361                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3362                         val = tp->rx_opt.user_mss;
3363                 if (tp->repair)
3364                         val = tp->rx_opt.mss_clamp;
3365                 break;
3366         case TCP_NODELAY:
3367                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3368                 break;
3369         case TCP_CORK:
3370                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3371                 break;
3372         case TCP_KEEPIDLE:
3373                 val = keepalive_time_when(tp) / HZ;
3374                 break;
3375         case TCP_KEEPINTVL:
3376                 val = keepalive_intvl_when(tp) / HZ;
3377                 break;
3378         case TCP_KEEPCNT:
3379                 val = keepalive_probes(tp);
3380                 break;
3381         case TCP_SYNCNT:
3382                 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3383                 break;
3384         case TCP_LINGER2:
3385                 val = tp->linger2;
3386                 if (val >= 0)
3387                         val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3388                 break;
3389         case TCP_DEFER_ACCEPT:
3390                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3391                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3392                 break;
3393         case TCP_WINDOW_CLAMP:
3394                 val = tp->window_clamp;
3395                 break;
3396         case TCP_INFO: {
3397                 struct tcp_info info;
3398
3399                 if (get_user(len, optlen))
3400                         return -EFAULT;
3401
3402                 tcp_get_info(sk, &info);
3403
3404                 len = min_t(unsigned int, len, sizeof(info));
3405                 if (put_user(len, optlen))
3406                         return -EFAULT;
3407                 if (copy_to_user(optval, &info, len))
3408                         return -EFAULT;
3409                 return 0;
3410         }
3411         case TCP_CC_INFO: {
3412                 const struct tcp_congestion_ops *ca_ops;
3413                 union tcp_cc_info info;
3414                 size_t sz = 0;
3415                 int attr;
3416
3417                 if (get_user(len, optlen))
3418                         return -EFAULT;
3419
3420                 ca_ops = icsk->icsk_ca_ops;
3421                 if (ca_ops && ca_ops->get_info)
3422                         sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3423
3424                 len = min_t(unsigned int, len, sz);
3425                 if (put_user(len, optlen))
3426                         return -EFAULT;
3427                 if (copy_to_user(optval, &info, len))
3428                         return -EFAULT;
3429                 return 0;
3430         }
3431         case TCP_QUICKACK:
3432                 val = !icsk->icsk_ack.pingpong;
3433                 break;
3434
3435         case TCP_CONGESTION:
3436                 if (get_user(len, optlen))
3437                         return -EFAULT;
3438                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
3439                 if (put_user(len, optlen))
3440                         return -EFAULT;
3441                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3442                         return -EFAULT;
3443                 return 0;
3444
3445         case TCP_ULP:
3446                 if (get_user(len, optlen))
3447                         return -EFAULT;
3448                 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
3449                 if (!icsk->icsk_ulp_ops) {
3450                         if (put_user(0, optlen))
3451                                 return -EFAULT;
3452                         return 0;
3453                 }
3454                 if (put_user(len, optlen))
3455                         return -EFAULT;
3456                 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
3457                         return -EFAULT;
3458                 return 0;
3459
3460         case TCP_FASTOPEN_KEY: {
3461                 __u8 key[TCP_FASTOPEN_KEY_LENGTH];
3462                 struct tcp_fastopen_context *ctx;
3463
3464                 if (get_user(len, optlen))
3465                         return -EFAULT;
3466
3467                 rcu_read_lock();
3468                 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
3469                 if (ctx)
3470                         memcpy(key, ctx->key, sizeof(key));
3471                 else
3472                         len = 0;
3473                 rcu_read_unlock();
3474
3475                 len = min_t(unsigned int, len, sizeof(key));
3476                 if (put_user(len, optlen))
3477                         return -EFAULT;
3478                 if (copy_to_user(optval, key, len))
3479                         return -EFAULT;
3480                 return 0;
3481         }
3482         case TCP_THIN_LINEAR_TIMEOUTS:
3483                 val = tp->thin_lto;
3484                 break;
3485
3486         case TCP_THIN_DUPACK:
3487                 val = 0;
3488                 break;
3489
3490         case TCP_REPAIR:
3491                 val = tp->repair;
3492                 break;
3493
3494         case TCP_REPAIR_QUEUE:
3495                 if (tp->repair)
3496                         val = tp->repair_queue;
3497                 else
3498                         return -EINVAL;
3499                 break;
3500
3501         case TCP_REPAIR_WINDOW: {
3502                 struct tcp_repair_window opt;
3503
3504                 if (get_user(len, optlen))
3505                         return -EFAULT;
3506
3507                 if (len != sizeof(opt))
3508                         return -EINVAL;
3509
3510                 if (!tp->repair)
3511                         return -EPERM;
3512
3513                 opt.snd_wl1     = tp->snd_wl1;
3514                 opt.snd_wnd     = tp->snd_wnd;
3515                 opt.max_window  = tp->max_window;
3516                 opt.rcv_wnd     = tp->rcv_wnd;
3517                 opt.rcv_wup     = tp->rcv_wup;
3518
3519                 if (copy_to_user(optval, &opt, len))
3520                         return -EFAULT;
3521                 return 0;
3522         }
3523         case TCP_QUEUE_SEQ:
3524                 if (tp->repair_queue == TCP_SEND_QUEUE)
3525                         val = tp->write_seq;
3526                 else if (tp->repair_queue == TCP_RECV_QUEUE)
3527                         val = tp->rcv_nxt;
3528                 else
3529                         return -EINVAL;
3530                 break;
3531
3532         case TCP_USER_TIMEOUT:
3533                 val = icsk->icsk_user_timeout;
3534                 break;
3535
3536         case TCP_FASTOPEN:
3537                 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3538                 break;
3539
3540         case TCP_FASTOPEN_CONNECT:
3541                 val = tp->fastopen_connect;
3542                 break;
3543
3544         case TCP_FASTOPEN_NO_COOKIE:
3545                 val = tp->fastopen_no_cookie;
3546                 break;
3547
3548         case TCP_TIMESTAMP:
3549                 val = tcp_time_stamp_raw() + tp->tsoffset;
3550                 break;
3551         case TCP_NOTSENT_LOWAT:
3552                 val = tp->notsent_lowat;
3553                 break;
3554         case TCP_INQ:
3555                 val = tp->recvmsg_inq;
3556                 break;
3557         case TCP_SAVE_SYN:
3558                 val = tp->save_syn;
3559                 break;
3560         case TCP_SAVED_SYN: {
3561                 if (get_user(len, optlen))
3562                         return -EFAULT;
3563
3564                 lock_sock(sk);
3565                 if (tp->saved_syn) {
3566                         if (len < tp->saved_syn[0]) {
3567                                 if (put_user(tp->saved_syn[0], optlen)) {
3568                                         release_sock(sk);
3569                                         return -EFAULT;
3570                                 }
3571                                 release_sock(sk);
3572                                 return -EINVAL;
3573                         }
3574                         len = tp->saved_syn[0];
3575                         if (put_user(len, optlen)) {
3576                                 release_sock(sk);
3577                                 return -EFAULT;
3578                         }
3579                         if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3580                                 release_sock(sk);
3581                                 return -EFAULT;
3582                         }
3583                         tcp_saved_syn_free(tp);
3584                         release_sock(sk);
3585                 } else {
3586                         release_sock(sk);
3587                         len = 0;
3588                         if (put_user(len, optlen))
3589                                 return -EFAULT;
3590                 }
3591                 return 0;
3592         }
3593 #ifdef CONFIG_MMU
3594         case TCP_ZEROCOPY_RECEIVE: {
3595                 struct tcp_zerocopy_receive zc;
3596                 int err;
3597
3598                 if (get_user(len, optlen))
3599                         return -EFAULT;
3600                 if (len != sizeof(zc))
3601                         return -EINVAL;
3602                 if (copy_from_user(&zc, optval, len))
3603                         return -EFAULT;
3604                 lock_sock(sk);
3605                 err = tcp_zerocopy_receive(sk, &zc);
3606                 release_sock(sk);
3607                 if (!err && copy_to_user(optval, &zc, len))
3608                         err = -EFAULT;
3609                 return err;
3610         }
3611 #endif
3612         default:
3613                 return -ENOPROTOOPT;
3614         }
3615
3616         if (put_user(len, optlen))
3617                 return -EFAULT;
3618         if (copy_to_user(optval, &val, len))
3619                 return -EFAULT;
3620         return 0;
3621 }
3622
3623 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3624                    int __user *optlen)
3625 {
3626         struct inet_connection_sock *icsk = inet_csk(sk);
3627
3628         if (level != SOL_TCP)
3629                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3630                                                      optval, optlen);
3631         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3632 }
3633 EXPORT_SYMBOL(tcp_getsockopt);
3634
3635 #ifdef CONFIG_COMPAT
3636 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3637                           char __user *optval, int __user *optlen)
3638 {
3639         if (level != SOL_TCP)
3640                 return inet_csk_compat_getsockopt(sk, level, optname,
3641                                                   optval, optlen);
3642         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3643 }
3644 EXPORT_SYMBOL(compat_tcp_getsockopt);
3645 #endif
3646
3647 #ifdef CONFIG_TCP_MD5SIG
3648 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3649 static DEFINE_MUTEX(tcp_md5sig_mutex);
3650 static bool tcp_md5sig_pool_populated = false;
3651
3652 static void __tcp_alloc_md5sig_pool(void)
3653 {
3654         struct crypto_ahash *hash;
3655         int cpu;
3656
3657         hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3658         if (IS_ERR(hash))
3659                 return;
3660
3661         for_each_possible_cpu(cpu) {
3662                 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3663                 struct ahash_request *req;
3664
3665                 if (!scratch) {
3666                         scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3667                                                sizeof(struct tcphdr),
3668                                                GFP_KERNEL,
3669                                                cpu_to_node(cpu));
3670                         if (!scratch)
3671                                 return;
3672                         per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3673                 }
3674                 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3675                         continue;
3676
3677                 req = ahash_request_alloc(hash, GFP_KERNEL);
3678                 if (!req)
3679                         return;
3680
3681                 ahash_request_set_callback(req, 0, NULL, NULL);
3682
3683                 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3684         }
3685         /* before setting tcp_md5sig_pool_populated, we must commit all writes
3686          * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3687          */
3688         smp_wmb();
3689         tcp_md5sig_pool_populated = true;
3690 }
3691
3692 bool tcp_alloc_md5sig_pool(void)
3693 {
3694         if (unlikely(!tcp_md5sig_pool_populated)) {
3695                 mutex_lock(&tcp_md5sig_mutex);
3696
3697                 if (!tcp_md5sig_pool_populated)
3698                         __tcp_alloc_md5sig_pool();
3699
3700                 mutex_unlock(&tcp_md5sig_mutex);
3701         }
3702         return tcp_md5sig_pool_populated;
3703 }
3704 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3705
3706
3707 /**
3708  *      tcp_get_md5sig_pool - get md5sig_pool for this user
3709  *
3710  *      We use percpu structure, so if we succeed, we exit with preemption
3711  *      and BH disabled, to make sure another thread or softirq handling
3712  *      wont try to get same context.
3713  */
3714 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3715 {
3716         local_bh_disable();
3717
3718         if (tcp_md5sig_pool_populated) {
3719                 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3720                 smp_rmb();
3721                 return this_cpu_ptr(&tcp_md5sig_pool);
3722         }
3723         local_bh_enable();
3724         return NULL;
3725 }
3726 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3727
3728 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3729                           const struct sk_buff *skb, unsigned int header_len)
3730 {
3731         struct scatterlist sg;
3732         const struct tcphdr *tp = tcp_hdr(skb);
3733         struct ahash_request *req = hp->md5_req;
3734         unsigned int i;
3735         const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3736                                            skb_headlen(skb) - header_len : 0;
3737         const struct skb_shared_info *shi = skb_shinfo(skb);
3738         struct sk_buff *frag_iter;
3739
3740         sg_init_table(&sg, 1);
3741
3742         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3743         ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3744         if (crypto_ahash_update(req))
3745                 return 1;
3746
3747         for (i = 0; i < shi->nr_frags; ++i) {
3748                 const struct skb_frag_struct *f = &shi->frags[i];
3749                 unsigned int offset = f->page_offset;
3750                 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3751
3752                 sg_set_page(&sg, page, skb_frag_size(f),
3753                             offset_in_page(offset));
3754                 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3755                 if (crypto_ahash_update(req))
3756                         return 1;
3757         }
3758
3759         skb_walk_frags(skb, frag_iter)
3760                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3761                         return 1;
3762
3763         return 0;
3764 }
3765 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3766
3767 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3768 {
3769         u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
3770         struct scatterlist sg;
3771
3772         sg_init_one(&sg, key->key, keylen);
3773         ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
3774
3775         /* tcp_md5_do_add() might change key->key under us */
3776         return crypto_ahash_update(hp->md5_req);
3777 }
3778 EXPORT_SYMBOL(tcp_md5_hash_key);
3779
3780 #endif
3781
3782 void tcp_done(struct sock *sk)
3783 {
3784         struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3785
3786         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3787                 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3788
3789         tcp_set_state(sk, TCP_CLOSE);
3790         tcp_clear_xmit_timers(sk);
3791         if (req)
3792                 reqsk_fastopen_remove(sk, req, false);
3793
3794         sk->sk_shutdown = SHUTDOWN_MASK;
3795
3796         if (!sock_flag(sk, SOCK_DEAD))
3797                 sk->sk_state_change(sk);
3798         else
3799                 inet_csk_destroy_sock(sk);
3800 }
3801 EXPORT_SYMBOL_GPL(tcp_done);
3802
3803 int tcp_abort(struct sock *sk, int err)
3804 {
3805         if (!sk_fullsock(sk)) {
3806                 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3807                         struct request_sock *req = inet_reqsk(sk);
3808
3809                         local_bh_disable();
3810                         inet_csk_reqsk_queue_drop(req->rsk_listener, req);
3811                         local_bh_enable();
3812                         return 0;
3813                 }
3814                 return -EOPNOTSUPP;
3815         }
3816
3817         /* Don't race with userspace socket closes such as tcp_close. */
3818         lock_sock(sk);
3819
3820         if (sk->sk_state == TCP_LISTEN) {
3821                 tcp_set_state(sk, TCP_CLOSE);
3822                 inet_csk_listen_stop(sk);
3823         }
3824
3825         /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3826         local_bh_disable();
3827         bh_lock_sock(sk);
3828
3829         if (!sock_flag(sk, SOCK_DEAD)) {
3830                 sk->sk_err = err;
3831                 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3832                 smp_wmb();
3833                 sk->sk_error_report(sk);
3834                 if (tcp_need_reset(sk->sk_state))
3835                         tcp_send_active_reset(sk, GFP_ATOMIC);
3836                 tcp_done(sk);
3837         }
3838
3839         bh_unlock_sock(sk);
3840         local_bh_enable();
3841         tcp_write_queue_purge(sk);
3842         release_sock(sk);
3843         return 0;
3844 }
3845 EXPORT_SYMBOL_GPL(tcp_abort);
3846
3847 extern struct tcp_congestion_ops tcp_reno;
3848
3849 static __initdata unsigned long thash_entries;
3850 static int __init set_thash_entries(char *str)
3851 {
3852         ssize_t ret;
3853
3854         if (!str)
3855                 return 0;
3856
3857         ret = kstrtoul(str, 0, &thash_entries);
3858         if (ret)
3859                 return 0;
3860
3861         return 1;
3862 }
3863 __setup("thash_entries=", set_thash_entries);
3864
3865 static void __init tcp_init_mem(void)
3866 {
3867         unsigned long limit = nr_free_buffer_pages() / 16;
3868
3869         limit = max(limit, 128UL);
3870         sysctl_tcp_mem[0] = limit / 4 * 3;              /* 4.68 % */
3871         sysctl_tcp_mem[1] = limit;                      /* 6.25 % */
3872         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;      /* 9.37 % */
3873 }
3874
3875 void __init tcp_init(void)
3876 {
3877         int max_rshare, max_wshare, cnt;
3878         unsigned long limit;
3879         unsigned int i;
3880
3881         BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
3882         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3883                      FIELD_SIZEOF(struct sk_buff, cb));
3884
3885         percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3886         percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3887         inet_hashinfo_init(&tcp_hashinfo);
3888         inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
3889                             thash_entries, 21,  /* one slot per 2 MB*/
3890                             0, 64 * 1024);
3891         tcp_hashinfo.bind_bucket_cachep =
3892                 kmem_cache_create("tcp_bind_bucket",
3893                                   sizeof(struct inet_bind_bucket), 0,
3894                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3895
3896         /* Size and allocate the main established and bind bucket
3897          * hash tables.
3898          *
3899          * The methodology is similar to that of the buffer cache.
3900          */
3901         tcp_hashinfo.ehash =
3902                 alloc_large_system_hash("TCP established",
3903                                         sizeof(struct inet_ehash_bucket),
3904                                         thash_entries,
3905                                         17, /* one slot per 128 KB of memory */
3906                                         0,
3907                                         NULL,
3908                                         &tcp_hashinfo.ehash_mask,
3909                                         0,
3910                                         thash_entries ? 0 : 512 * 1024);
3911         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3912                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3913
3914         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3915                 panic("TCP: failed to alloc ehash_locks");
3916         tcp_hashinfo.bhash =
3917                 alloc_large_system_hash("TCP bind",
3918                                         sizeof(struct inet_bind_hashbucket),
3919                                         tcp_hashinfo.ehash_mask + 1,
3920                                         17, /* one slot per 128 KB of memory */
3921                                         0,
3922                                         &tcp_hashinfo.bhash_size,
3923                                         NULL,
3924                                         0,
3925                                         64 * 1024);
3926         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3927         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3928                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3929                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3930         }
3931
3932
3933         cnt = tcp_hashinfo.ehash_mask + 1;
3934         sysctl_tcp_max_orphans = cnt / 2;
3935
3936         tcp_init_mem();
3937         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3938         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3939         max_wshare = min(4UL*1024*1024, limit);
3940         max_rshare = min(6UL*1024*1024, limit);
3941
3942         init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3943         init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
3944         init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3945
3946         init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3947         init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
3948         init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
3949
3950         pr_info("Hash tables configured (established %u bind %u)\n",
3951                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3952
3953         tcp_v4_init();
3954         tcp_metrics_init();
3955         BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3956         tcp_tasklet_init();
3957 }