net/ipv4/tcp.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   4  *              operating system.  INET is implemented using the  BSD Socket
   5  *              interface as the means of communication with the user level.
   6  *
   7  *              Implementation of the Transmission Control Protocol(TCP).
   8  *
   9  * Authors:     Ross Biro
  10  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  11  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
  12  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
  13  *              Florian La Roche, <flla@stud.uni-sb.de>
  14  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  15  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
  16  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
  17  *              Matthew Dillon, <dillon@apollo.west.oic.com>
  18  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  19  *              Jorge Cwik, <jorge@laser.satlink.net>
  20  *
  21  * Fixes:
  22  *              Alan Cox        :       Numerous verify_area() calls
  23  *              Alan Cox        :       Set the ACK bit on a reset
  24  *              Alan Cox        :       Stopped it crashing if it closed while
  25  *                                      sk->inuse=1 and was trying to connect
  26  *                                      (tcp_err()).
  27  *              Alan Cox        :       All icmp error handling was broken
  28  *                                      pointers passed where wrong and the
  29  *                                      socket was looked up backwards. Nobody
  30  *                                      tested any icmp error code obviously.
  31  *              Alan Cox        :       tcp_err() now handled properly. It
  32  *                                      wakes people on errors. poll
  33  *                                      behaves and the icmp error race
  34  *                                      has gone by moving it into sock.c
  35  *              Alan Cox        :       tcp_send_reset() fixed to work for
  36  *                                      everything not just packets for
  37  *                                      unknown sockets.
  38  *              Alan Cox        :       tcp option processing.
  39  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
  40  *                                      syn rule wrong]
  41  *              Herp Rosmanith  :       More reset fixes
  42  *              Alan Cox        :       No longer acks invalid rst frames.
  43  *                                      Acking any kind of RST is right out.
  44  *              Alan Cox        :       Sets an ignore me flag on an rst
  45  *                                      receive otherwise odd bits of prattle
  46  *                                      escape still
  47  *              Alan Cox        :       Fixed another acking RST frame bug.
  48  *                                      Should stop LAN workplace lockups.
  49  *              Alan Cox        :       Some tidyups using the new skb list
  50  *                                      facilities
  51  *              Alan Cox        :       sk->keepopen now seems to work
  52  *              Alan Cox        :       Pulls options out correctly on accepts
  53  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
  54  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
  55  *                                      bit to skb ops.
  56  *              Alan Cox        :       Tidied tcp_data to avoid a potential
  57  *                                      nasty.
  58  *              Alan Cox        :       Added some better commenting, as the
  59  *                                      tcp is hard to follow
  60  *              Alan Cox        :       Removed incorrect check for 20 * psh
  61  *      Michael O'Reilly        :       ack < copied bug fix.
  62  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
  63  *              Alan Cox        :       FIN with no memory -> CRASH
  64  *              Alan Cox        :       Added socket option proto entries.
  65  *                                      Also added awareness of them to accept.
  66  *              Alan Cox        :       Added TCP options (SOL_TCP)
  67  *              Alan Cox        :       Switched wakeup calls to callbacks,
  68  *                                      so the kernel can layer network
  69  *                                      sockets.
  70  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
  71  *              Alan Cox        :       Handle FIN (more) properly (we hope).
  72  *              Alan Cox        :       RST frames sent on unsynchronised
  73  *                                      state ack error.
  74  *              Alan Cox        :       Put in missing check for SYN bit.
  75  *              Alan Cox        :       Added tcp_select_window() aka NET2E
  76  *                                      window non shrink trick.
  77  *              Alan Cox        :       Added a couple of small NET2E timer
  78  *                                      fixes
  79  *              Charles Hedrick :       TCP fixes
  80  *              Toomas Tamm     :       TCP window fixes
  81  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
  82  *              Charles Hedrick :       Rewrote most of it to actually work
  83  *              Linus           :       Rewrote tcp_read() and URG handling
  84  *                                      completely
  85  *              Gerhard Koerting:       Fixed some missing timer handling
  86  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
  87  *              Gerhard Koerting:       PC/TCP workarounds
  88  *              Adam Caldwell   :       Assorted timer/timing errors
  89  *              Matthew Dillon  :       Fixed another RST bug
  90  *              Alan Cox        :       Move to kernel side addressing changes.
  91  *              Alan Cox        :       Beginning work on TCP fastpathing
  92  *                                      (not yet usable)
  93  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
  94  *              Alan Cox        :       TCP fast path debugging
  95  *              Alan Cox        :       Window clamping
  96  *              Michael Riepe   :       Bug in tcp_check()
  97  *              Matt Dillon     :       More TCP improvements and RST bug fixes
  98  *              Matt Dillon     :       Yet more small nasties remove from the
  99  *                                      TCP code (Be very nice to this man if
 100  *                                      tcp finally works 100%) 8)
 101  *              Alan Cox        :       BSD accept semantics.
 102  *              Alan Cox        :       Reset on closedown bug.
 103  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
 104  *              Michael Pall    :       Handle poll() after URG properly in
 105  *                                      all cases.
 106  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
 107  *                                      (multi URG PUSH broke rlogin).
 108  *              Michael Pall    :       Fix the multi URG PUSH problem in
 109  *                                      tcp_readable(), poll() after URG
 110  *                                      works now.
 111  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
 112  *                                      BSD api.
 113  *              Alan Cox        :       Changed the semantics of sk->socket to
 114  *                                      fix a race and a signal problem with
 115  *                                      accept() and async I/O.
 116  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
 117  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
 118  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
 119  *                                      clients/servers which listen in on
 120  *                                      fixed ports.
 121  *              Alan Cox        :       Cleaned the above up and shrank it to
 122  *                                      a sensible code size.
 123  *              Alan Cox        :       Self connect lockup fix.
 124  *              Alan Cox        :       No connect to multicast.
 125  *              Ross Biro       :       Close unaccepted children on master
 126  *                                      socket close.
 127  *              Alan Cox        :       Reset tracing code.
 128  *              Alan Cox        :       Spurious resets on shutdown.
 129  *              Alan Cox        :       Giant 15 minute/60 second timer error
 130  *              Alan Cox        :       Small whoops in polling before an
 131  *                                      accept.
 132  *              Alan Cox        :       Kept the state trace facility since
 133  *                                      it's handy for debugging.
 134  *              Alan Cox        :       More reset handler fixes.
 135  *              Alan Cox        :       Started rewriting the code based on
 136  *                                      the RFC's for other useful protocol
 137  *                                      references see: Comer, KA9Q NOS, and
 138  *                                      for a reference on the difference
 139  *                                      between specifications and how BSD
 140  *                                      works see the 4.4lite source.
 141  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
 142  *                                      close.
 143  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
 144  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
 145  *              Alan Cox        :       Reimplemented timers as per the RFC
 146  *                                      and using multiple timers for sanity.
 147  *              Alan Cox        :       Small bug fixes, and a lot of new
 148  *                                      comments.
 149  *              Alan Cox        :       Fixed dual reader crash by locking
 150  *                                      the buffers (much like datagram.c)
 151  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
 152  *                                      now gets fed up of retrying without
 153  *                                      (even a no space) answer.
 154  *              Alan Cox        :       Extracted closing code better
 155  *              Alan Cox        :       Fixed the closing state machine to
 156  *                                      resemble the RFC.
 157  *              Alan Cox        :       More 'per spec' fixes.
 158  *              Jorge Cwik      :       Even faster checksumming.
 159  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
 160  *                                      only frames. At least one pc tcp stack
 161  *                                      generates them.
 162  *              Alan Cox        :       Cache last socket.
 163  *              Alan Cox        :       Per route irtt.
 164  *              Matt Day        :       poll()->select() match BSD precisely on error
 165  *              Alan Cox        :       New buffers
 166  *              Marc Tamsky     :       Various sk->prot->retransmits and
 167  *                                      sk->retransmits misupdating fixed.
 168  *                                      Fixed tcp_write_timeout: stuck close,
 169  *                                      and TCP syn retries gets used now.
 170  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
 171  *                                      ack if state is TCP_CLOSED.
 172  *              Alan Cox        :       Look up device on a retransmit - routes may
 173  *                                      change. Doesn't yet cope with MSS shrink right
 174  *                                      but it's a start!
 175  *              Marc Tamsky     :       Closing in closing fixes.
 176  *              Mike Shaver     :       RFC1122 verifications.
 177  *              Alan Cox        :       rcv_saddr errors.
 178  *              Alan Cox        :       Block double connect().
 179  *              Alan Cox        :       Small hooks for enSKIP.
 180  *              Alexey Kuznetsov:       Path MTU discovery.
 181  *              Alan Cox        :       Support soft errors.
 182  *              Alan Cox        :       Fix MTU discovery pathological case
 183  *                                      when the remote claims no mtu!
 184  *              Marc Tamsky     :       TCP_CLOSE fix.
 185  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
 186  *                                      window but wrong (fixes NT lpd problems)
 187  *              Pedro Roque     :       Better TCP window handling, delayed ack.
 188  *              Joerg Reuter    :       No modification of locked buffers in
 189  *                                      tcp_do_retransmit()
 190  *              Eric Schenk     :       Changed receiver side silly window
 191  *                                      avoidance algorithm to BSD style
 192  *                                      algorithm. This doubles throughput
 193  *                                      against machines running Solaris,
 194  *                                      and seems to result in general
 195  *                                      improvement.
 196  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
 197  *      Willy Konynenberg       :       Transparent proxying support.
 198  *      Mike McLagan            :       Routing by source
 199  *              Keith Owens     :       Do proper merging with partial SKB's in
 200  *                                      tcp_do_sendmsg to avoid burstiness.
 201  *              Eric Schenk     :       Fix fast close down bug with
 202  *                                      shutdown() followed by close().
 203  *              Andi Kleen      :       Make poll agree with SIGIO
 204  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
 205  *                                      lingertime == 0 (RFC 793 ABORT Call)
 206  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
 207  *                                      csum_and_copy_from_user() if possible.
 208  *
 209  * Description of States:
 210  *
 211  *      TCP_SYN_SENT            sent a connection request, waiting for ack
 212  *
 213  *      TCP_SYN_RECV            received a connection request, sent ack,
 214  *                              waiting for final ack in three-way handshake.
 215  *
 216  *      TCP_ESTABLISHED         connection established
 217  *
 218  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
 219  *                              transmission of remaining buffered data
 220  *
 221  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
 222  *                              to shutdown
 223  *
 224  *      TCP_CLOSING             both sides have shutdown but we still have
 225  *                              data we have to finish sending
 226  *
 227  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
 228  *                              closed, can only be entered from FIN_WAIT2
 229  *                              or CLOSING.  Required because the other end
 230  *                              may not have gotten our last ACK causing it
 231  *                              to retransmit the data packet (which we ignore)
 232  *
 233  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
 234  *                              us to finish writing our data and to shutdown
 235  *                              (we have to close() to move on to LAST_ACK)
 236  *
 237  *      TCP_LAST_ACK            out side has shutdown after remote has
 238  *                              shutdown.  There may still be data in our
 239  *                              buffer that we have to finish sending
 240  *
 241  *      TCP_CLOSE               socket is finished
 242  */
 243
 244 #define pr_fmt(fmt) "TCP: " fmt
 245
 246 #include <crypto/hash.h>
 247 #include <linux/kernel.h>
 248 #include <linux/module.h>
 249 #include <linux/types.h>
 250 #include <linux/fcntl.h>
 251 #include <linux/poll.h>
 252 #include <linux/inet_diag.h>
 253 #include <linux/init.h>
 254 #include <linux/fs.h>
 255 #include <linux/skbuff.h>
 256 #include <linux/scatterlist.h>
 257 #include <linux/splice.h>
 258 #include <linux/net.h>
 259 #include <linux/socket.h>
 260 #include <linux/random.h>
 261 #include <linux/memblock.h>
 262 #include <linux/highmem.h>
 263 #include <linux/swap.h>
 264 #include <linux/cache.h>
 265 #include <linux/err.h>
 266 #include <linux/time.h>
 267 #include <linux/slab.h>
 268 #include <linux/errqueue.h>
 269 #include <linux/static_key.h>
 270
 271 #include <net/icmp.h>
 272 #include <net/inet_common.h>
 273 #include <net/tcp.h>
 274 #include <net/xfrm.h>
 275 #include <net/ip.h>
 276 #include <net/sock.h>
 277
 278 #include <linux/uaccess.h>
 279 #include <asm/ioctls.h>
 280 #include <net/busy_poll.h>
 281
 282 struct percpu_counter tcp_orphan_count;
 283 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 284
 285 long sysctl_tcp_mem[3] __read_mostly;
 286 EXPORT_SYMBOL(sysctl_tcp_mem);
 287
 288 atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
 289 EXPORT_SYMBOL(tcp_memory_allocated);
 290
 291 #if IS_ENABLED(CONFIG_SMC)
 292 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
 293 EXPORT_SYMBOL(tcp_have_smc);
 294 #endif
 295
 296 /*
 297  * Current number of TCP sockets.
 298  */
 299 struct percpu_counter tcp_sockets_allocated;
 300 EXPORT_SYMBOL(tcp_sockets_allocated);
 301
 302 /*
 303  * TCP splice context
 304  */
 305 struct tcp_splice_state {
 306         struct pipe_inode_info *pipe;
 307         size_t len;
 308         unsigned int flags;
 309 };
 310
 311 /*
 312  * Pressure flag: try to collapse.
 313  * Technical note: it is used by multiple contexts non atomically.
 314  * All the __sk_mem_schedule() is of this nature: accounting
 315  * is strict, actions are advisory and have some latency.
 316  */
 317 unsigned long tcp_memory_pressure __read_mostly;
 318 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
 319
 320 DEFINE_STATIC_KEY_FALSE(tcp_rx_skb_cache_key);
 321 EXPORT_SYMBOL(tcp_rx_skb_cache_key);
 322
 323 DEFINE_STATIC_KEY_FALSE(tcp_tx_skb_cache_key);
 324
 325 void tcp_enter_memory_pressure(struct sock *sk)
 326 {
 327         unsigned long val;
 328
 329         if (tcp_memory_pressure)
 330                 return;
 331         val = jiffies;
 332
 333         if (!val)
 334                 val--;
 335         if (!cmpxchg(&tcp_memory_pressure, 0, val))
 336                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
 337 }
 338 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
 339
 340 void tcp_leave_memory_pressure(struct sock *sk)
 341 {
 342         unsigned long val;
 343
 344         if (!tcp_memory_pressure)
 345                 return;
 346         val = xchg(&tcp_memory_pressure, 0);
 347         if (val)
 348                 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
 349                               jiffies_to_msecs(jiffies - val));
 350 }
 351 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
 352
 353 /* Convert seconds to retransmits based on initial and max timeout */
 354 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
 355 {
 356         u8 res = 0;
 357
 358         if (seconds > 0) {
 359                 int period = timeout;
 360
 361                 res = 1;
 362                 while (seconds > period && res < 255) {
 363                         res++;
 364                         timeout <<= 1;
 365                         if (timeout > rto_max)
 366                                 timeout = rto_max;
 367                         period += timeout;
 368                 }
 369         }
 370         return res;
 371 }
 372
 373 /* Convert retransmits to seconds based on initial and max timeout */
 374 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
 375 {
 376         int period = 0;
 377
 378         if (retrans > 0) {
 379                 period = timeout;
 380                 while (--retrans) {
 381                         timeout <<= 1;
 382                         if (timeout > rto_max)
 383                                 timeout = rto_max;
 384                         period += timeout;
 385                 }
 386         }
 387         return period;
 388 }
 389
 390 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
 391 {
 392         u32 rate = READ_ONCE(tp->rate_delivered);
 393         u32 intv = READ_ONCE(tp->rate_interval_us);
 394         u64 rate64 = 0;
 395
 396         if (rate && intv) {
 397                 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
 398                 do_div(rate64, intv);
 399         }
 400         return rate64;
 401 }
 402
 403 /* Address-family independent initialization for a tcp_sock.
 404  *
 405  * NOTE: A lot of things set to zero explicitly by call to
 406  *       sk_alloc() so need not be done here.
 407  */
 408 void tcp_init_sock(struct sock *sk)
 409 {
 410         struct inet_connection_sock *icsk = inet_csk(sk);
 411         struct tcp_sock *tp = tcp_sk(sk);
 412
 413         tp->out_of_order_queue = RB_ROOT;
 414         sk->tcp_rtx_queue = RB_ROOT;
 415         tcp_init_xmit_timers(sk);
 416         INIT_LIST_HEAD(&tp->tsq_node);
 417         INIT_LIST_HEAD(&tp->tsorted_sent_queue);
 418
 419         icsk->icsk_rto = TCP_TIMEOUT_INIT;
 420         tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
 421         minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
 422
 423         /* So many TCP implementations out there (incorrectly) count the
 424          * initial SYN frame in their delayed-ACK and congestion control
 425          * algorithms that we must have the following bandaid to talk
 426          * efficiently to them.  -DaveM
 427          */
 428         tp->snd_cwnd = TCP_INIT_CWND;
 429
 430         /* There's a bubble in the pipe until at least the first ACK. */
 431         tp->app_limited = ~0U;
 432
 433         /* See draft-stevens-tcpca-spec-01 for discussion of the
 434          * initialization of these values.
 435          */
 436         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 437         tp->snd_cwnd_clamp = ~0;
 438         tp->mss_cache = TCP_MSS_DEFAULT;
 439
 440         tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
 441         tcp_assign_congestion_control(sk);
 442
 443         tp->tsoffset = 0;
 444         tp->rack.reo_wnd_steps = 1;
 445
 446         sk->sk_state = TCP_CLOSE;
 447
 448         sk->sk_write_space = sk_stream_write_space;
 449         sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
 450
 451         icsk->icsk_sync_mss = tcp_sync_mss;
 452
 453         sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
 454         sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
 455
 456         sk_sockets_allocated_inc(sk);
 457         sk->sk_route_forced_caps = NETIF_F_GSO;
 458 }
 459 EXPORT_SYMBOL(tcp_init_sock);
 460
 461 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
 462 {
 463         struct sk_buff *skb = tcp_write_queue_tail(sk);
 464
 465         if (tsflags && skb) {
 466                 struct skb_shared_info *shinfo = skb_shinfo(skb);
 467                 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 468
 469                 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
 470                 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
 471                         tcb->txstamp_ack = 1;
 472                 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
 473                         shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
 474         }
 475 }
 476
 477 static inline bool tcp_stream_is_readable(const struct tcp_sock *tp,
 478                                           int target, struct sock *sk)
 479 {
 480         return (tp->rcv_nxt - tp->copied_seq >= target) ||
 481                 (sk->sk_prot->stream_memory_read ?
 482                 sk->sk_prot->stream_memory_read(sk) : false);
 483 }
 484
 485 /*
 486  *      Wait for a TCP event.
 487  *
 488  *      Note that we don't need to lock the socket, as the upper poll layers
 489  *      take care of normal races (between the test and the event) and we don't
 490  *      go look at any of the socket buffers directly.
 491  */
 492 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 493 {
 494         __poll_t mask;
 495         struct sock *sk = sock->sk;
 496         const struct tcp_sock *tp = tcp_sk(sk);
 497         int state;
 498
 499         sock_poll_wait(file, sock, wait);
 500
 501         state = inet_sk_state_load(sk);
 502         if (state == TCP_LISTEN)
 503                 return inet_csk_listen_poll(sk);
 504
 505         /* Socket is not locked. We are protected from async events
 506          * by poll logic and correct handling of state changes
 507          * made by other threads is impossible in any case.
 508          */
 509
 510         mask = 0;
 511
 512         /*
 513          * EPOLLHUP is certainly not done right. But poll() doesn't
 514          * have a notion of HUP in just one direction, and for a
 515          * socket the read side is more interesting.
 516          *
 517          * Some poll() documentation says that EPOLLHUP is incompatible
 518          * with the EPOLLOUT/POLLWR flags, so somebody should check this
 519          * all. But careful, it tends to be safer to return too many
 520          * bits than too few, and you can easily break real applications
 521          * if you don't tell them that something has hung up!
 522          *
 523          * Check-me.
 524          *
 525          * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
 526          * our fs/select.c). It means that after we received EOF,
 527          * poll always returns immediately, making impossible poll() on write()
 528          * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
 529          * if and only if shutdown has been made in both directions.
 530          * Actually, it is interesting to look how Solaris and DUX
 531          * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
 532          * then we could set it on SND_SHUTDOWN. BTW examples given
 533          * in Stevens' books assume exactly this behaviour, it explains
 534          * why EPOLLHUP is incompatible with EPOLLOUT.  --ANK
 535          *
 536          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 537          * blocking on fresh not-connected or disconnected socket. --ANK
 538          */
 539         if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
 540                 mask |= EPOLLHUP;
 541         if (sk->sk_shutdown & RCV_SHUTDOWN)
 542                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
 543
 544         /* Connected or passive Fast Open socket? */
 545         if (state != TCP_SYN_SENT &&
 546             (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
 547                 int target = sock_rcvlowat(sk, 0, INT_MAX);
 548
 549                 if (tp->urg_seq == tp->copied_seq &&
 550                     !sock_flag(sk, SOCK_URGINLINE) &&
 551                     tp->urg_data)
 552                         target++;
 553
 554                 if (tcp_stream_is_readable(tp, target, sk))
 555                         mask |= EPOLLIN | EPOLLRDNORM;
 556
 557                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 558                         if (sk_stream_is_writeable(sk)) {
 559                                 mask |= EPOLLOUT | EPOLLWRNORM;
 560                         } else {  /* send SIGIO later */
 561                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 562                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 563
 564                                 /* Race breaker. If space is freed after
 565                                  * wspace test but before the flags are set,
 566                                  * IO signal will be lost. Memory barrier
 567                                  * pairs with the input side.
 568                                  */
 569                                 smp_mb__after_atomic();
 570                                 if (sk_stream_is_writeable(sk))
 571                                         mask |= EPOLLOUT | EPOLLWRNORM;
 572                         }
 573                 } else
 574                         mask |= EPOLLOUT | EPOLLWRNORM;
 575
 576                 if (tp->urg_data & TCP_URG_VALID)
 577                         mask |= EPOLLPRI;
 578         } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
 579                 /* Active TCP fastopen socket with defer_connect
 580                  * Return EPOLLOUT so application can call write()
 581                  * in order for kernel to generate SYN+data
 582                  */
 583                 mask |= EPOLLOUT | EPOLLWRNORM;
 584         }
 585         /* This barrier is coupled with smp_wmb() in tcp_reset() */
 586         smp_rmb();
 587         if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
 588                 mask |= EPOLLERR;
 589
 590         return mask;
 591 }
 592 EXPORT_SYMBOL(tcp_poll);
 593
 594 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 595 {
 596         struct tcp_sock *tp = tcp_sk(sk);
 597         int answ;
 598         bool slow;
 599
 600         switch (cmd) {
 601         case SIOCINQ:
 602                 if (sk->sk_state == TCP_LISTEN)
 603                         return -EINVAL;
 604
 605                 slow = lock_sock_fast(sk);
 606                 answ = tcp_inq(sk);
 607                 unlock_sock_fast(sk, slow);
 608                 break;
 609         case SIOCATMARK:
 610                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 611                 break;
 612         case SIOCOUTQ:
 613                 if (sk->sk_state == TCP_LISTEN)
 614                         return -EINVAL;
 615
 616                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 617                         answ = 0;
 618                 else
 619                         answ = tp->write_seq - tp->snd_una;
 620                 break;
 621         case SIOCOUTQNSD:
 622                 if (sk->sk_state == TCP_LISTEN)
 623                         return -EINVAL;
 624
 625                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 626                         answ = 0;
 627                 else
 628                         answ = tp->write_seq - tp->snd_nxt;
 629                 break;
 630         default:
 631                 return -ENOIOCTLCMD;
 632         }
 633
 634         return put_user(answ, (int __user *)arg);
 635 }
 636 EXPORT_SYMBOL(tcp_ioctl);
 637
 638 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 639 {
 640         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 641         tp->pushed_seq = tp->write_seq;
 642 }
 643
 644 static inline bool forced_push(const struct tcp_sock *tp)
 645 {
 646         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 647 }
 648
 649 static void skb_entail(struct sock *sk, struct sk_buff *skb)
 650 {
 651         struct tcp_sock *tp = tcp_sk(sk);
 652         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 653
 654         skb->csum    = 0;
 655         tcb->seq     = tcb->end_seq = tp->write_seq;
 656         tcb->tcp_flags = TCPHDR_ACK;
 657         tcb->sacked  = 0;
 658         __skb_header_release(skb);
 659         tcp_add_write_queue_tail(sk, skb);
 660         sk->sk_wmem_queued += skb->truesize;
 661         sk_mem_charge(sk, skb->truesize);
 662         if (tp->nonagle & TCP_NAGLE_PUSH)
 663                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 664
 665         tcp_slow_start_after_idle_check(sk);
 666 }
 667
 668 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
 669 {
 670         if (flags & MSG_OOB)
 671                 tp->snd_up = tp->write_seq;
 672 }
 673
 674 /* If a not yet filled skb is pushed, do not send it if
 675  * we have data packets in Qdisc or NIC queues :
 676  * Because TX completion will happen shortly, it gives a chance
 677  * to coalesce future sendmsg() payload into this skb, without
 678  * need for a timer, and with no latency trade off.
 679  * As packets containing data payload have a bigger truesize
 680  * than pure acks (dataless) packets, the last checks prevent
 681  * autocorking if we only have an ACK in Qdisc/NIC queues,
 682  * or if TX completion was delayed after we processed ACK packet.
 683  */
 684 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
 685                                 int size_goal)
 686 {
 687         return skb->len < size_goal &&
 688                sock_net(sk)->ipv4.sysctl_tcp_autocorking &&
 689                !tcp_rtx_queue_empty(sk) &&
 690                refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
 691 }
 692
 693 static void tcp_push(struct sock *sk, int flags, int mss_now,
 694                      int nonagle, int size_goal)
 695 {
 696         struct tcp_sock *tp = tcp_sk(sk);
 697         struct sk_buff *skb;
 698
 699         skb = tcp_write_queue_tail(sk);
 700         if (!skb)
 701                 return;
 702         if (!(flags & MSG_MORE) || forced_push(tp))
 703                 tcp_mark_push(tp, skb);
 704
 705         tcp_mark_urg(tp, flags);
 706
 707         if (tcp_should_autocork(sk, skb, size_goal)) {
 708
 709                 /* avoid atomic op if TSQ_THROTTLED bit is already set */
 710                 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
 711                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
 712                         set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
 713                 }
 714                 /* It is possible TX completion already happened
 715                  * before we set TSQ_THROTTLED.
 716                  */
 717                 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
 718                         return;
 719         }
 720
 721         if (flags & MSG_MORE)
 722                 nonagle = TCP_NAGLE_CORK;
 723
 724         __tcp_push_pending_frames(sk, mss_now, nonagle);
 725 }
 726
 727 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 728                                 unsigned int offset, size_t len)
 729 {
 730         struct tcp_splice_state *tss = rd_desc->arg.data;
 731         int ret;
 732
 733         ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
 734                               min(rd_desc->count, len), tss->flags);
 735         if (ret > 0)
 736                 rd_desc->count -= ret;
 737         return ret;
 738 }
 739
 740 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 741 {
 742         /* Store TCP splice context information in read_descriptor_t. */
 743         read_descriptor_t rd_desc = {
 744                 .arg.data = tss,
 745                 .count    = tss->len,
 746         };
 747
 748         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 749 }
 750
 751 /**
 752  *  tcp_splice_read - splice data from TCP socket to a pipe
 753  * @sock:       socket to splice from
 754  * @ppos:       position (not valid)
 755  * @pipe:       pipe to splice to
 756  * @len:        number of bytes to splice
 757  * @flags:      splice modifier flags
 758  *
 759  * Description:
 760  *    Will read pages from given socket and fill them into a pipe.
 761  *
 762  **/
 763 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 764                         struct pipe_inode_info *pipe, size_t len,
 765                         unsigned int flags)
 766 {
 767         struct sock *sk = sock->sk;
 768         struct tcp_splice_state tss = {
 769                 .pipe = pipe,
 770                 .len = len,
 771                 .flags = flags,
 772         };
 773         long timeo;
 774         ssize_t spliced;
 775         int ret;
 776
 777         sock_rps_record_flow(sk);
 778         /*
 779          * We can't seek on a socket input
 780          */
 781         if (unlikely(*ppos))
 782                 return -ESPIPE;
 783
 784         ret = spliced = 0;
 785
 786         lock_sock(sk);
 787
 788         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 789         while (tss.len) {
 790                 ret = __tcp_splice_read(sk, &tss);
 791                 if (ret < 0)
 792                         break;
 793                 else if (!ret) {
 794                         if (spliced)
 795                                 break;
 796                         if (sock_flag(sk, SOCK_DONE))
 797                                 break;
 798                         if (sk->sk_err) {
 799                                 ret = sock_error(sk);
 800                                 break;
 801                         }
 802                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 803                                 break;
 804                         if (sk->sk_state == TCP_CLOSE) {
 805                                 /*
 806                                  * This occurs when user tries to read
 807                                  * from never connected socket.
 808                                  */
 809                                 ret = -ENOTCONN;
 810                                 break;
 811                         }
 812                         if (!timeo) {
 813                                 ret = -EAGAIN;
 814                                 break;
 815                         }
 816                         /* if __tcp_splice_read() got nothing while we have
 817                          * an skb in receive queue, we do not want to loop.
 818                          * This might happen with URG data.
 819                          */
 820                         if (!skb_queue_empty(&sk->sk_receive_queue))
 821                                 break;
 822                         sk_wait_data(sk, &timeo, NULL);
 823                         if (signal_pending(current)) {
 824                                 ret = sock_intr_errno(timeo);
 825                                 break;
 826                         }
 827                         continue;
 828                 }
 829                 tss.len -= ret;
 830                 spliced += ret;
 831
 832                 if (!timeo)
 833                         break;
 834                 release_sock(sk);
 835                 lock_sock(sk);
 836
 837                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 838                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 839                     signal_pending(current))
 840                         break;
 841         }
 842
 843         release_sock(sk);
 844
 845         if (spliced)
 846                 return spliced;
 847
 848         return ret;
 849 }
 850 EXPORT_SYMBOL(tcp_splice_read);
 851
 852 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
 853                                     bool force_schedule)
 854 {
 855         struct sk_buff *skb;
 856
 857         if (likely(!size)) {
 858                 skb = sk->sk_tx_skb_cache;
 859                 if (skb) {
 860                         skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
 861                         sk->sk_tx_skb_cache = NULL;
 862                         pskb_trim(skb, 0);
 863                         INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
 864                         skb_shinfo(skb)->tx_flags = 0;
 865                         memset(TCP_SKB_CB(skb), 0, sizeof(struct tcp_skb_cb));
 866                         return skb;
 867                 }
 868         }
 869         /* The TCP header must be at least 32-bit aligned.  */
 870         size = ALIGN(size, 4);
 871
 872         if (unlikely(tcp_under_memory_pressure(sk)))
 873                 sk_mem_reclaim_partial(sk);
 874
 875         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 876         if (likely(skb)) {
 877                 bool mem_scheduled;
 878
 879                 if (force_schedule) {
 880                         mem_scheduled = true;
 881                         sk_forced_mem_schedule(sk, skb->truesize);
 882                 } else {
 883                         mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
 884                 }
 885                 if (likely(mem_scheduled)) {
 886                         skb_reserve(skb, sk->sk_prot->max_header);
 887                         /*
 888                          * Make sure that we have exactly size bytes
 889                          * available to the caller, no more, no less.
 890                          */
 891                         skb->reserved_tailroom = skb->end - skb->tail - size;
 892                         INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
 893                         return skb;
 894                 }
 895                 __kfree_skb(skb);
 896         } else {
 897                 sk->sk_prot->enter_memory_pressure(sk);
 898                 sk_stream_moderate_sndbuf(sk);
 899         }
 900         return NULL;
 901 }
 902
 903 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 904                                        int large_allowed)
 905 {
 906         struct tcp_sock *tp = tcp_sk(sk);
 907         u32 new_size_goal, size_goal;
 908
 909         if (!large_allowed)
 910                 return mss_now;
 911
 912         /* Note : tcp_tso_autosize() will eventually split this later */
 913         new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
 914         new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
 915
 916         /* We try hard to avoid divides here */
 917         size_goal = tp->gso_segs * mss_now;
 918         if (unlikely(new_size_goal < size_goal ||
 919                      new_size_goal >= size_goal + mss_now)) {
 920                 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
 921                                      sk->sk_gso_max_segs);
 922                 size_goal = tp->gso_segs * mss_now;
 923         }
 924
 925         return max(size_goal, mss_now);
 926 }
 927
 928 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 929 {
 930         int mss_now;
 931
 932         mss_now = tcp_current_mss(sk);
 933         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 934
 935         return mss_now;
 936 }
 937
 938 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
 939                          size_t size, int flags)
 940 {
 941         struct tcp_sock *tp = tcp_sk(sk);
 942         int mss_now, size_goal;
 943         int err;
 944         ssize_t copied;
 945         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 946
 947         if (IS_ENABLED(CONFIG_DEBUG_VM) &&
 948             WARN_ONCE(PageSlab(page), "page must not be a Slab one"))
 949                 return -EINVAL;
 950
 951         /* Wait for a connection to finish. One exception is TCP Fast Open
 952          * (passive side) where data is allowed to be sent before a connection
 953          * is fully established.
 954          */
 955         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
 956             !tcp_passive_fastopen(sk)) {
 957                 err = sk_stream_wait_connect(sk, &timeo);
 958                 if (err != 0)
 959                         goto out_err;
 960         }
 961
 962         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 963
 964         mss_now = tcp_send_mss(sk, &size_goal, flags);
 965         copied = 0;
 966
 967         err = -EPIPE;
 968         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 969                 goto out_err;
 970
 971         while (size > 0) {
 972                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 973                 int copy, i;
 974                 bool can_coalesce;
 975
 976                 if (!skb || (copy = size_goal - skb->len) <= 0 ||
 977                     !tcp_skb_can_collapse_to(skb)) {
 978 new_segment:
 979                         if (!sk_stream_memory_free(sk))
 980                                 goto wait_for_sndbuf;
 981
 982                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
 983                                         tcp_rtx_and_write_queues_empty(sk));
 984                         if (!skb)
 985                                 goto wait_for_memory;
 986
 987                         skb_entail(sk, skb);
 988                         copy = size_goal;
 989                 }
 990
 991                 if (copy > size)
 992                         copy = size;
 993
 994                 i = skb_shinfo(skb)->nr_frags;
 995                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 996                 if (!can_coalesce && i >= sysctl_max_skb_frags) {
 997                         tcp_mark_push(tp, skb);
 998                         goto new_segment;
 999                 }
1000                 if (!sk_wmem_schedule(sk, copy))
1001                         goto wait_for_memory;
1002
1003                 if (can_coalesce) {
1004                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1005                 } else {
1006                         get_page(page);
1007                         skb_fill_page_desc(skb, i, page, offset, copy);
1008                 }
1009
1010                 if (!(flags & MSG_NO_SHARED_FRAGS))
1011                         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1012
1013                 skb->len += copy;
1014                 skb->data_len += copy;
1015                 skb->truesize += copy;
1016                 sk->sk_wmem_queued += copy;
1017                 sk_mem_charge(sk, copy);
1018                 skb->ip_summed = CHECKSUM_PARTIAL;
1019                 tp->write_seq += copy;
1020                 TCP_SKB_CB(skb)->end_seq += copy;
1021                 tcp_skb_pcount_set(skb, 0);
1022
1023                 if (!copied)
1024                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1025
1026                 copied += copy;
1027                 offset += copy;
1028                 size -= copy;
1029                 if (!size)
1030                         goto out;
1031
1032                 if (skb->len < size_goal || (flags & MSG_OOB))
1033                         continue;
1034
1035                 if (forced_push(tp)) {
1036                         tcp_mark_push(tp, skb);
1037                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1038                 } else if (skb == tcp_send_head(sk))
1039                         tcp_push_one(sk, mss_now);
1040                 continue;
1041
1042 wait_for_sndbuf:
1043                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1044 wait_for_memory:
1045                 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1046                          TCP_NAGLE_PUSH, size_goal);
1047
1048                 err = sk_stream_wait_memory(sk, &timeo);
1049                 if (err != 0)
1050                         goto do_error;
1051
1052                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1053         }
1054
1055 out:
1056         if (copied) {
1057                 tcp_tx_timestamp(sk, sk->sk_tsflags);
1058                 if (!(flags & MSG_SENDPAGE_NOTLAST))
1059                         tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1060         }
1061         return copied;
1062
1063 do_error:
1064         if (copied)
1065                 goto out;
1066 out_err:
1067         /* make sure we wake any epoll edge trigger waiter */
1068         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1069                      err == -EAGAIN)) {
1070                 sk->sk_write_space(sk);
1071                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1072         }
1073         return sk_stream_error(sk, flags, err);
1074 }
1075 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1076
1077 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1078                         size_t size, int flags)
1079 {
1080         if (!(sk->sk_route_caps & NETIF_F_SG))
1081                 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1082
1083         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1084
1085         return do_tcp_sendpages(sk, page, offset, size, flags);
1086 }
1087 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1088
1089 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1090                  size_t size, int flags)
1091 {
1092         int ret;
1093
1094         lock_sock(sk);
1095         ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1096         release_sock(sk);
1097
1098         return ret;
1099 }
1100 EXPORT_SYMBOL(tcp_sendpage);
1101
1102 void tcp_free_fastopen_req(struct tcp_sock *tp)
1103 {
1104         if (tp->fastopen_req) {
1105                 kfree(tp->fastopen_req);
1106                 tp->fastopen_req = NULL;
1107         }
1108 }
1109
1110 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1111                                 int *copied, size_t size,
1112                                 struct ubuf_info *uarg)
1113 {
1114         struct tcp_sock *tp = tcp_sk(sk);
1115         struct inet_sock *inet = inet_sk(sk);
1116         struct sockaddr *uaddr = msg->msg_name;
1117         int err, flags;
1118
1119         if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1120             (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1121              uaddr->sa_family == AF_UNSPEC))
1122                 return -EOPNOTSUPP;
1123         if (tp->fastopen_req)
1124                 return -EALREADY; /* Another Fast Open is in progress */
1125
1126         tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1127                                    sk->sk_allocation);
1128         if (unlikely(!tp->fastopen_req))
1129                 return -ENOBUFS;
1130         tp->fastopen_req->data = msg;
1131         tp->fastopen_req->size = size;
1132         tp->fastopen_req->uarg = uarg;
1133
1134         if (inet->defer_connect) {
1135                 err = tcp_connect(sk);
1136                 /* Same failure procedure as in tcp_v4/6_connect */
1137                 if (err) {
1138                         tcp_set_state(sk, TCP_CLOSE);
1139                         inet->inet_dport = 0;
1140                         sk->sk_route_caps = 0;
1141                 }
1142         }
1143         flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1144         err = __inet_stream_connect(sk->sk_socket, uaddr,
1145                                     msg->msg_namelen, flags, 1);
1146         /* fastopen_req could already be freed in __inet_stream_connect
1147          * if the connection times out or gets rst
1148          */
1149         if (tp->fastopen_req) {
1150                 *copied = tp->fastopen_req->copied;
1151                 tcp_free_fastopen_req(tp);
1152                 inet->defer_connect = 0;
1153         }
1154         return err;
1155 }
1156
1157 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1158 {
1159         struct tcp_sock *tp = tcp_sk(sk);
1160         struct ubuf_info *uarg = NULL;
1161         struct sk_buff *skb;
1162         struct sockcm_cookie sockc;
1163         int flags, err, copied = 0;
1164         int mss_now = 0, size_goal, copied_syn = 0;
1165         bool process_backlog = false;
1166         bool zc = false;
1167         long timeo;
1168
1169         flags = msg->msg_flags;
1170
1171         if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
1172                 skb = tcp_write_queue_tail(sk);
1173                 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
1174                 if (!uarg) {
1175                         err = -ENOBUFS;
1176                         goto out_err;
1177                 }
1178
1179                 zc = sk->sk_route_caps & NETIF_F_SG;
1180                 if (!zc)
1181                         uarg->zerocopy = 0;
1182         }
1183
1184         if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1185             !tp->repair) {
1186                 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1187                 if (err == -EINPROGRESS && copied_syn > 0)
1188                         goto out;
1189                 else if (err)
1190                         goto out_err;
1191         }
1192
1193         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1194
1195         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1196
1197         /* Wait for a connection to finish. One exception is TCP Fast Open
1198          * (passive side) where data is allowed to be sent before a connection
1199          * is fully established.
1200          */
1201         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1202             !tcp_passive_fastopen(sk)) {
1203                 err = sk_stream_wait_connect(sk, &timeo);
1204                 if (err != 0)
1205                         goto do_error;
1206         }
1207
1208         if (unlikely(tp->repair)) {
1209                 if (tp->repair_queue == TCP_RECV_QUEUE) {
1210                         copied = tcp_send_rcvq(sk, msg, size);
1211                         goto out_nopush;
1212                 }
1213
1214                 err = -EINVAL;
1215                 if (tp->repair_queue == TCP_NO_QUEUE)
1216                         goto out_err;
1217
1218                 /* 'common' sending to sendq */
1219         }
1220
1221         sockcm_init(&sockc, sk);
1222         if (msg->msg_controllen) {
1223                 err = sock_cmsg_send(sk, msg, &sockc);
1224                 if (unlikely(err)) {
1225                         err = -EINVAL;
1226                         goto out_err;
1227                 }
1228         }
1229
1230         /* This should be in poll */
1231         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1232
1233         /* Ok commence sending. */
1234         copied = 0;
1235
1236 restart:
1237         mss_now = tcp_send_mss(sk, &size_goal, flags);
1238
1239         err = -EPIPE;
1240         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1241                 goto do_error;
1242
1243         while (msg_data_left(msg)) {
1244                 int copy = 0;
1245
1246                 skb = tcp_write_queue_tail(sk);
1247                 if (skb)
1248                         copy = size_goal - skb->len;
1249
1250                 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1251                         bool first_skb;
1252
1253 new_segment:
1254                         if (!sk_stream_memory_free(sk))
1255                                 goto wait_for_sndbuf;
1256
1257                         if (process_backlog && sk_flush_backlog(sk)) {
1258                                 process_backlog = false;
1259                                 goto restart;
1260                         }
1261                         first_skb = tcp_rtx_and_write_queues_empty(sk);
1262                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
1263                                                   first_skb);
1264                         if (!skb)
1265                                 goto wait_for_memory;
1266
1267                         process_backlog = true;
1268                         skb->ip_summed = CHECKSUM_PARTIAL;
1269
1270                         skb_entail(sk, skb);
1271                         copy = size_goal;
1272
1273                         /* All packets are restored as if they have
1274                          * already been sent. skb_mstamp_ns isn't set to
1275                          * avoid wrong rtt estimation.
1276                          */
1277                         if (tp->repair)
1278                                 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1279                 }
1280
1281                 /* Try to append data to the end of skb. */
1282                 if (copy > msg_data_left(msg))
1283                         copy = msg_data_left(msg);
1284
1285                 /* Where to copy to? */
1286                 if (skb_availroom(skb) > 0 && !zc) {
1287                         /* We have some space in skb head. Superb! */
1288                         copy = min_t(int, copy, skb_availroom(skb));
1289                         err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1290                         if (err)
1291                                 goto do_fault;
1292                 } else if (!zc) {
1293                         bool merge = true;
1294                         int i = skb_shinfo(skb)->nr_frags;
1295                         struct page_frag *pfrag = sk_page_frag(sk);
1296
1297                         if (!sk_page_frag_refill(sk, pfrag))
1298                                 goto wait_for_memory;
1299
1300                         if (!skb_can_coalesce(skb, i, pfrag->page,
1301                                               pfrag->offset)) {
1302                                 if (i >= sysctl_max_skb_frags) {
1303                                         tcp_mark_push(tp, skb);
1304                                         goto new_segment;
1305                                 }
1306                                 merge = false;
1307                         }
1308
1309                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1310
1311                         if (!sk_wmem_schedule(sk, copy))
1312                                 goto wait_for_memory;
1313
1314                         err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1315                                                        pfrag->page,
1316                                                        pfrag->offset,
1317                                                        copy);
1318                         if (err)
1319                                 goto do_error;
1320
1321                         /* Update the skb. */
1322                         if (merge) {
1323                                 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1324                         } else {
1325                                 skb_fill_page_desc(skb, i, pfrag->page,
1326                                                    pfrag->offset, copy);
1327                                 page_ref_inc(pfrag->page);
1328                         }
1329                         pfrag->offset += copy;
1330                 } else {
1331                         err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1332                         if (err == -EMSGSIZE || err == -EEXIST) {
1333                                 tcp_mark_push(tp, skb);
1334                                 goto new_segment;
1335                         }
1336                         if (err < 0)
1337                                 goto do_error;
1338                         copy = err;
1339                 }
1340
1341                 if (!copied)
1342                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1343
1344                 tp->write_seq += copy;
1345                 TCP_SKB_CB(skb)->end_seq += copy;
1346                 tcp_skb_pcount_set(skb, 0);
1347
1348                 copied += copy;
1349                 if (!msg_data_left(msg)) {
1350                         if (unlikely(flags & MSG_EOR))
1351                                 TCP_SKB_CB(skb)->eor = 1;
1352                         goto out;
1353                 }
1354
1355                 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1356                         continue;
1357
1358                 if (forced_push(tp)) {
1359                         tcp_mark_push(tp, skb);
1360                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1361                 } else if (skb == tcp_send_head(sk))
1362                         tcp_push_one(sk, mss_now);
1363                 continue;
1364
1365 wait_for_sndbuf:
1366                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1367 wait_for_memory:
1368                 if (copied)
1369                         tcp_push(sk, flags & ~MSG_MORE, mss_now,
1370                                  TCP_NAGLE_PUSH, size_goal);
1371
1372                 err = sk_stream_wait_memory(sk, &timeo);
1373                 if (err != 0)
1374                         goto do_error;
1375
1376                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1377         }
1378
1379 out:
1380         if (copied) {
1381                 tcp_tx_timestamp(sk, sockc.tsflags);
1382                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1383         }
1384 out_nopush:
1385         sock_zerocopy_put(uarg);
1386         return copied + copied_syn;
1387
1388 do_fault:
1389         if (!skb->len) {
1390                 tcp_unlink_write_queue(skb, sk);
1391                 /* It is the one place in all of TCP, except connection
1392                  * reset, where we can be unlinking the send_head.
1393                  */
1394                 if (tcp_write_queue_empty(sk))
1395                         tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
1396                 sk_wmem_free_skb(sk, skb);
1397         }
1398
1399 do_error:
1400         if (copied + copied_syn)
1401                 goto out;
1402 out_err:
1403         sock_zerocopy_put_abort(uarg, true);
1404         err = sk_stream_error(sk, flags, err);
1405         /* make sure we wake any epoll edge trigger waiter */
1406         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1407                      err == -EAGAIN)) {
1408                 sk->sk_write_space(sk);
1409                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1410         }
1411         return err;
1412 }
1413 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1414
1415 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1416 {
1417         int ret;
1418
1419         lock_sock(sk);
1420         ret = tcp_sendmsg_locked(sk, msg, size);
1421         release_sock(sk);
1422
1423         return ret;
1424 }
1425 EXPORT_SYMBOL(tcp_sendmsg);
1426
1427 /*
1428  *      Handle reading urgent data. BSD has very simple semantics for
1429  *      this, no blocking and very strange errors 8)
1430  */
1431
1432 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1433 {
1434         struct tcp_sock *tp = tcp_sk(sk);
1435
1436         /* No URG data to read. */
1437         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1438             tp->urg_data == TCP_URG_READ)
1439                 return -EINVAL; /* Yes this is right ! */
1440
1441         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1442                 return -ENOTCONN;
1443
1444         if (tp->urg_data & TCP_URG_VALID) {
1445                 int err = 0;
1446                 char c = tp->urg_data;
1447
1448                 if (!(flags & MSG_PEEK))
1449                         tp->urg_data = TCP_URG_READ;
1450
1451                 /* Read urgent data. */
1452                 msg->msg_flags |= MSG_OOB;
1453
1454                 if (len > 0) {
1455                         if (!(flags & MSG_TRUNC))
1456                                 err = memcpy_to_msg(msg, &c, 1);
1457                         len = 1;
1458                 } else
1459                         msg->msg_flags |= MSG_TRUNC;
1460
1461                 return err ? -EFAULT : len;
1462         }
1463
1464         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1465                 return 0;
1466
1467         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1468          * the available implementations agree in this case:
1469          * this call should never block, independent of the
1470          * blocking state of the socket.
1471          * Mike <pall@rz.uni-karlsruhe.de>
1472          */
1473         return -EAGAIN;
1474 }
1475
1476 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1477 {
1478         struct sk_buff *skb;
1479         int copied = 0, err = 0;
1480
1481         /* XXX -- need to support SO_PEEK_OFF */
1482
1483         skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1484                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1485                 if (err)
1486                         return err;
1487                 copied += skb->len;
1488         }
1489
1490         skb_queue_walk(&sk->sk_write_queue, skb) {
1491                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1492                 if (err)
1493                         break;
1494
1495                 copied += skb->len;
1496         }
1497
1498         return err ?: copied;
1499 }
1500
1501 /* Clean up the receive buffer for full frames taken by the user,
1502  * then send an ACK if necessary.  COPIED is the number of bytes
1503  * tcp_recvmsg has given to the user so far, it speeds up the
1504  * calculation of whether or not we must ACK for the sake of
1505  * a window update.
1506  */
1507 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1508 {
1509         struct tcp_sock *tp = tcp_sk(sk);
1510         bool time_to_ack = false;
1511
1512         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1513
1514         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1515              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1516              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1517
1518         if (inet_csk_ack_scheduled(sk)) {
1519                 const struct inet_connection_sock *icsk = inet_csk(sk);
1520                    /* Delayed ACKs frequently hit locked sockets during bulk
1521                     * receive. */
1522                 if (icsk->icsk_ack.blocked ||
1523                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1524                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1525                     /*
1526                      * If this read emptied read buffer, we send ACK, if
1527                      * connection is not bidirectional, user drained
1528                      * receive buffer and there was a small segment
1529                      * in queue.
1530                      */
1531                     (copied > 0 &&
1532                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1533                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1534                        !inet_csk_in_pingpong_mode(sk))) &&
1535                       !atomic_read(&sk->sk_rmem_alloc)))
1536                         time_to_ack = true;
1537         }
1538
1539         /* We send an ACK if we can now advertise a non-zero window
1540          * which has been raised "significantly".
1541          *
1542          * Even if window raised up to infinity, do not send window open ACK
1543          * in states, where we will not receive more. It is useless.
1544          */
1545         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1546                 __u32 rcv_window_now = tcp_receive_window(tp);
1547
1548                 /* Optimize, __tcp_select_window() is not cheap. */
1549                 if (2*rcv_window_now <= tp->window_clamp) {
1550                         __u32 new_window = __tcp_select_window(sk);
1551
1552                         /* Send ACK now, if this read freed lots of space
1553                          * in our buffer. Certainly, new_window is new window.
1554                          * We can advertise it now, if it is not less than current one.
1555                          * "Lots" means "at least twice" here.
1556                          */
1557                         if (new_window && new_window >= 2 * rcv_window_now)
1558                                 time_to_ack = true;
1559                 }
1560         }
1561         if (time_to_ack)
1562                 tcp_send_ack(sk);
1563 }
1564
1565 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1566 {
1567         struct sk_buff *skb;
1568         u32 offset;
1569
1570         while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1571                 offset = seq - TCP_SKB_CB(skb)->seq;
1572                 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1573                         pr_err_once("%s: found a SYN, please report !\n", __func__);
1574                         offset--;
1575                 }
1576                 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1577                         *off = offset;
1578                         return skb;
1579                 }
1580                 /* This looks weird, but this can happen if TCP collapsing
1581                  * splitted a fat GRO packet, while we released socket lock
1582                  * in skb_splice_bits()
1583                  */
1584                 sk_eat_skb(sk, skb);
1585         }
1586         return NULL;
1587 }
1588
1589 /*
1590  * This routine provides an alternative to tcp_recvmsg() for routines
1591  * that would like to handle copying from skbuffs directly in 'sendfile'
1592  * fashion.
1593  * Note:
1594  *      - It is assumed that the socket was locked by the caller.
1595  *      - The routine does not block.
1596  *      - At present, there is no support for reading OOB data
1597  *        or for 'peeking' the socket using this routine
1598  *        (although both would be easy to implement).
1599  */
1600 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1601                   sk_read_actor_t recv_actor)
1602 {
1603         struct sk_buff *skb;
1604         struct tcp_sock *tp = tcp_sk(sk);
1605         u32 seq = tp->copied_seq;
1606         u32 offset;
1607         int copied = 0;
1608
1609         if (sk->sk_state == TCP_LISTEN)
1610                 return -ENOTCONN;
1611         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1612                 if (offset < skb->len) {
1613                         int used;
1614                         size_t len;
1615
1616                         len = skb->len - offset;
1617                         /* Stop reading if we hit a patch of urgent data */
1618                         if (tp->urg_data) {
1619                                 u32 urg_offset = tp->urg_seq - seq;
1620                                 if (urg_offset < len)
1621                                         len = urg_offset;
1622                                 if (!len)
1623                                         break;
1624                         }
1625                         used = recv_actor(desc, skb, offset, len);
1626                         if (used <= 0) {
1627                                 if (!copied)
1628                                         copied = used;
1629                                 break;
1630                         } else if (used <= len) {
1631                                 seq += used;
1632                                 copied += used;
1633                                 offset += used;
1634                         }
1635                         /* If recv_actor drops the lock (e.g. TCP splice
1636                          * receive) the skb pointer might be invalid when
1637                          * getting here: tcp_collapse might have deleted it
1638                          * while aggregating skbs from the socket queue.
1639                          */
1640                         skb = tcp_recv_skb(sk, seq - 1, &offset);
1641                         if (!skb)
1642                                 break;
1643                         /* TCP coalescing might have appended data to the skb.
1644                          * Try to splice more frags
1645                          */
1646                         if (offset + 1 != skb->len)
1647                                 continue;
1648                 }
1649                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1650                         sk_eat_skb(sk, skb);
1651                         ++seq;
1652                         break;
1653                 }
1654                 sk_eat_skb(sk, skb);
1655                 if (!desc->count)
1656                         break;
1657                 tp->copied_seq = seq;
1658         }
1659         tp->copied_seq = seq;
1660
1661         tcp_rcv_space_adjust(sk);
1662
1663         /* Clean up data we have read: This will do ACK frames. */
1664         if (copied > 0) {
1665                 tcp_recv_skb(sk, seq, &offset);
1666                 tcp_cleanup_rbuf(sk, copied);
1667         }
1668         return copied;
1669 }
1670 EXPORT_SYMBOL(tcp_read_sock);
1671
1672 int tcp_peek_len(struct socket *sock)
1673 {
1674         return tcp_inq(sock->sk);
1675 }
1676 EXPORT_SYMBOL(tcp_peek_len);
1677
1678 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1679 int tcp_set_rcvlowat(struct sock *sk, int val)
1680 {
1681         int cap;
1682
1683         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1684                 cap = sk->sk_rcvbuf >> 1;
1685         else
1686                 cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1;
1687         val = min(val, cap);
1688         sk->sk_rcvlowat = val ? : 1;
1689
1690         /* Check if we need to signal EPOLLIN right now */
1691         tcp_data_ready(sk);
1692
1693         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1694                 return 0;
1695
1696         val <<= 1;
1697         if (val > sk->sk_rcvbuf) {
1698                 sk->sk_rcvbuf = val;
1699                 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1700         }
1701         return 0;
1702 }
1703 EXPORT_SYMBOL(tcp_set_rcvlowat);
1704
1705 #ifdef CONFIG_MMU
1706 static const struct vm_operations_struct tcp_vm_ops = {
1707 };
1708
1709 int tcp_mmap(struct file *file, struct socket *sock,
1710              struct vm_area_struct *vma)
1711 {
1712         if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1713                 return -EPERM;
1714         vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1715
1716         /* Instruct vm_insert_page() to not down_read(mmap_sem) */
1717         vma->vm_flags |= VM_MIXEDMAP;
1718
1719         vma->vm_ops = &tcp_vm_ops;
1720         return 0;
1721 }
1722 EXPORT_SYMBOL(tcp_mmap);
1723
1724 static int tcp_zerocopy_receive(struct sock *sk,
1725                                 struct tcp_zerocopy_receive *zc)
1726 {
1727         unsigned long address = (unsigned long)zc->address;
1728         const skb_frag_t *frags = NULL;
1729         u32 length = 0, seq, offset;
1730         struct vm_area_struct *vma;
1731         struct sk_buff *skb = NULL;
1732         struct tcp_sock *tp;
1733         int inq;
1734         int ret;
1735
1736         if (address & (PAGE_SIZE - 1) || address != zc->address)
1737                 return -EINVAL;
1738
1739         if (sk->sk_state == TCP_LISTEN)
1740                 return -ENOTCONN;
1741
1742         sock_rps_record_flow(sk);
1743
1744         down_read(&current->mm->mmap_sem);
1745
1746         ret = -EINVAL;
1747         vma = find_vma(current->mm, address);
1748         if (!vma || vma->vm_start > address || vma->vm_ops != &tcp_vm_ops)
1749                 goto out;
1750         zc->length = min_t(unsigned long, zc->length, vma->vm_end - address);
1751
1752         tp = tcp_sk(sk);
1753         seq = tp->copied_seq;
1754         inq = tcp_inq(sk);
1755         zc->length = min_t(u32, zc->length, inq);
1756         zc->length &= ~(PAGE_SIZE - 1);
1757         if (zc->length) {
1758                 zap_page_range(vma, address, zc->length);
1759                 zc->recv_skip_hint = 0;
1760         } else {
1761                 zc->recv_skip_hint = inq;
1762         }
1763         ret = 0;
1764         while (length + PAGE_SIZE <= zc->length) {
1765                 if (zc->recv_skip_hint < PAGE_SIZE) {
1766                         if (skb) {
1767                                 skb = skb->next;
1768                                 offset = seq - TCP_SKB_CB(skb)->seq;
1769                         } else {
1770                                 skb = tcp_recv_skb(sk, seq, &offset);
1771                         }
1772
1773                         zc->recv_skip_hint = skb->len - offset;
1774                         offset -= skb_headlen(skb);
1775                         if ((int)offset < 0 || skb_has_frag_list(skb))
1776                                 break;
1777                         frags = skb_shinfo(skb)->frags;
1778                         while (offset) {
1779                                 if (frags->size > offset)
1780                                         goto out;
1781                                 offset -= frags->size;
1782                                 frags++;
1783                         }
1784                 }
1785                 if (frags->size != PAGE_SIZE || frags->page_offset) {
1786                         int remaining = zc->recv_skip_hint;
1787
1788                         while (remaining && (frags->size != PAGE_SIZE ||
1789                                              frags->page_offset)) {
1790                                 remaining -= frags->size;
1791                                 frags++;
1792                         }
1793                         zc->recv_skip_hint -= remaining;
1794                         break;
1795                 }
1796                 ret = vm_insert_page(vma, address + length,
1797                                      skb_frag_page(frags));
1798                 if (ret)
1799                         break;
1800                 length += PAGE_SIZE;
1801                 seq += PAGE_SIZE;
1802                 zc->recv_skip_hint -= PAGE_SIZE;
1803                 frags++;
1804         }
1805 out:
1806         up_read(&current->mm->mmap_sem);
1807         if (length) {
1808                 tp->copied_seq = seq;
1809                 tcp_rcv_space_adjust(sk);
1810
1811                 /* Clean up data we have read: This will do ACK frames. */
1812                 tcp_recv_skb(sk, seq, &offset);
1813                 tcp_cleanup_rbuf(sk, length);
1814                 ret = 0;
1815                 if (length == zc->length)
1816                         zc->recv_skip_hint = 0;
1817         } else {
1818                 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
1819                         ret = -EIO;
1820         }
1821         zc->length = length;
1822         return ret;
1823 }
1824 #endif
1825
1826 static void tcp_update_recv_tstamps(struct sk_buff *skb,
1827                                     struct scm_timestamping_internal *tss)
1828 {
1829         if (skb->tstamp)
1830                 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1831         else
1832                 tss->ts[0] = (struct timespec64) {0};
1833
1834         if (skb_hwtstamps(skb)->hwtstamp)
1835                 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1836         else
1837                 tss->ts[2] = (struct timespec64) {0};
1838 }
1839
1840 /* Similar to __sock_recv_timestamp, but does not require an skb */
1841 static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
1842                                struct scm_timestamping_internal *tss)
1843 {
1844         int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
1845         bool has_timestamping = false;
1846
1847         if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
1848                 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
1849                         if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
1850                                 if (new_tstamp) {
1851                                         struct __kernel_timespec kts = {tss->ts[0].tv_sec, tss->ts[0].tv_nsec};
1852
1853                                         put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
1854                                                  sizeof(kts), &kts);
1855                                 } else {
1856                                         struct timespec ts_old = timespec64_to_timespec(tss->ts[0]);
1857
1858                                         put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
1859                                                  sizeof(ts_old), &ts_old);
1860                                 }
1861                         } else {
1862                                 if (new_tstamp) {
1863                                         struct __kernel_sock_timeval stv;
1864
1865                                         stv.tv_sec = tss->ts[0].tv_sec;
1866                                         stv.tv_usec = tss->ts[0].tv_nsec / 1000;
1867                                         put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
1868                                                  sizeof(stv), &stv);
1869                                 } else {
1870                                         struct __kernel_old_timeval tv;
1871
1872                                         tv.tv_sec = tss->ts[0].tv_sec;
1873                                         tv.tv_usec = tss->ts[0].tv_nsec / 1000;
1874                                         put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
1875                                                  sizeof(tv), &tv);
1876                                 }
1877                         }
1878                 }
1879
1880                 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
1881                         has_timestamping = true;
1882                 else
1883                         tss->ts[0] = (struct timespec64) {0};
1884         }
1885
1886         if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
1887                 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
1888                         has_timestamping = true;
1889                 else
1890                         tss->ts[2] = (struct timespec64) {0};
1891         }
1892
1893         if (has_timestamping) {
1894                 tss->ts[1] = (struct timespec64) {0};
1895                 if (sock_flag(sk, SOCK_TSTAMP_NEW))
1896                         put_cmsg_scm_timestamping64(msg, tss);
1897                 else
1898                         put_cmsg_scm_timestamping(msg, tss);
1899         }
1900 }
1901
1902 static int tcp_inq_hint(struct sock *sk)
1903 {
1904         const struct tcp_sock *tp = tcp_sk(sk);
1905         u32 copied_seq = READ_ONCE(tp->copied_seq);
1906         u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
1907         int inq;
1908
1909         inq = rcv_nxt - copied_seq;
1910         if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
1911                 lock_sock(sk);
1912                 inq = tp->rcv_nxt - tp->copied_seq;
1913                 release_sock(sk);
1914         }
1915         /* After receiving a FIN, tell the user-space to continue reading
1916          * by returning a non-zero inq.
1917          */
1918         if (inq == 0 && sock_flag(sk, SOCK_DONE))
1919                 inq = 1;
1920         return inq;
1921 }
1922
1923 /*
1924  *      This routine copies from a sock struct into the user buffer.
1925  *
1926  *      Technical note: in 2.3 we work on _locked_ socket, so that
1927  *      tricks with *seq access order and skb->users are not required.
1928  *      Probably, code can be easily improved even more.
1929  */
1930
1931 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1932                 int flags, int *addr_len)
1933 {
1934         struct tcp_sock *tp = tcp_sk(sk);
1935         int copied = 0;
1936         u32 peek_seq;
1937         u32 *seq;
1938         unsigned long used;
1939         int err, inq;
1940         int target;             /* Read at least this many bytes */
1941         long timeo;
1942         struct sk_buff *skb, *last;
1943         u32 urg_hole = 0;
1944         struct scm_timestamping_internal tss;
1945         bool has_tss = false;
1946         bool has_cmsg;
1947
1948         if (unlikely(flags & MSG_ERRQUEUE))
1949                 return inet_recv_error(sk, msg, len, addr_len);
1950
1951         if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1952             (sk->sk_state == TCP_ESTABLISHED))
1953                 sk_busy_loop(sk, nonblock);
1954
1955         lock_sock(sk);
1956
1957         err = -ENOTCONN;
1958         if (sk->sk_state == TCP_LISTEN)
1959                 goto out;
1960
1961         has_cmsg = tp->recvmsg_inq;
1962         timeo = sock_rcvtimeo(sk, nonblock);
1963
1964         /* Urgent data needs to be handled specially. */
1965         if (flags & MSG_OOB)
1966                 goto recv_urg;
1967
1968         if (unlikely(tp->repair)) {
1969                 err = -EPERM;
1970                 if (!(flags & MSG_PEEK))
1971                         goto out;
1972
1973                 if (tp->repair_queue == TCP_SEND_QUEUE)
1974                         goto recv_sndq;
1975
1976                 err = -EINVAL;
1977                 if (tp->repair_queue == TCP_NO_QUEUE)
1978                         goto out;
1979
1980                 /* 'common' recv queue MSG_PEEK-ing */
1981         }
1982
1983         seq = &tp->copied_seq;
1984         if (flags & MSG_PEEK) {
1985                 peek_seq = tp->copied_seq;
1986                 seq = &peek_seq;
1987         }
1988
1989         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1990
1991         do {
1992                 u32 offset;
1993
1994                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1995                 if (tp->urg_data && tp->urg_seq == *seq) {
1996                         if (copied)
1997                                 break;
1998                         if (signal_pending(current)) {
1999                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2000                                 break;
2001                         }
2002                 }
2003
2004                 /* Next get a buffer. */
2005
2006                 last = skb_peek_tail(&sk->sk_receive_queue);
2007                 skb_queue_walk(&sk->sk_receive_queue, skb) {
2008                         last = skb;
2009                         /* Now that we have two receive queues this
2010                          * shouldn't happen.
2011                          */
2012                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2013                                  "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2014                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2015                                  flags))
2016                                 break;
2017
2018                         offset = *seq - TCP_SKB_CB(skb)->seq;
2019                         if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2020                                 pr_err_once("%s: found a SYN, please report !\n", __func__);
2021                                 offset--;
2022                         }
2023                         if (offset < skb->len)
2024                                 goto found_ok_skb;
2025                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2026                                 goto found_fin_ok;
2027                         WARN(!(flags & MSG_PEEK),
2028                              "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2029                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2030                 }
2031
2032                 /* Well, if we have backlog, try to process it now yet. */
2033
2034                 if (copied >= target && !sk->sk_backlog.tail)
2035                         break;
2036
2037                 if (copied) {
2038                         if (sk->sk_err ||
2039                             sk->sk_state == TCP_CLOSE ||
2040                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
2041                             !timeo ||
2042                             signal_pending(current))
2043                                 break;
2044                 } else {
2045                         if (sock_flag(sk, SOCK_DONE))
2046                                 break;
2047
2048                         if (sk->sk_err) {
2049                                 copied = sock_error(sk);
2050                                 break;
2051                         }
2052
2053                         if (sk->sk_shutdown & RCV_SHUTDOWN)
2054                                 break;
2055
2056                         if (sk->sk_state == TCP_CLOSE) {
2057                                 /* This occurs when user tries to read
2058                                  * from never connected socket.
2059                                  */
2060                                 copied = -ENOTCONN;
2061                                 break;
2062                         }
2063
2064                         if (!timeo) {
2065                                 copied = -EAGAIN;
2066                                 break;
2067                         }
2068
2069                         if (signal_pending(current)) {
2070                                 copied = sock_intr_errno(timeo);
2071                                 break;
2072                         }
2073                 }
2074
2075                 tcp_cleanup_rbuf(sk, copied);
2076
2077                 if (copied >= target) {
2078                         /* Do not sleep, just process backlog. */
2079                         release_sock(sk);
2080                         lock_sock(sk);
2081                 } else {
2082                         sk_wait_data(sk, &timeo, last);
2083                 }
2084
2085                 if ((flags & MSG_PEEK) &&
2086                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
2087                         net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2088                                             current->comm,
2089                                             task_pid_nr(current));
2090                         peek_seq = tp->copied_seq;
2091                 }
2092                 continue;
2093
2094 found_ok_skb:
2095                 /* Ok so how much can we use? */
2096                 used = skb->len - offset;
2097                 if (len < used)
2098                         used = len;
2099
2100                 /* Do we have urgent data here? */
2101                 if (tp->urg_data) {
2102                         u32 urg_offset = tp->urg_seq - *seq;
2103                         if (urg_offset < used) {
2104                                 if (!urg_offset) {
2105                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
2106                                                 ++*seq;
2107                                                 urg_hole++;
2108                                                 offset++;
2109                                                 used--;
2110                                                 if (!used)
2111                                                         goto skip_copy;
2112                                         }
2113                                 } else
2114                                         used = urg_offset;
2115                         }
2116                 }
2117
2118                 if (!(flags & MSG_TRUNC)) {
2119                         err = skb_copy_datagram_msg(skb, offset, msg, used);
2120                         if (err) {
2121                                 /* Exception. Bailout! */
2122                                 if (!copied)
2123                                         copied = -EFAULT;
2124                                 break;
2125                         }
2126                 }
2127
2128                 *seq += used;
2129                 copied += used;
2130                 len -= used;
2131
2132                 tcp_rcv_space_adjust(sk);
2133
2134 skip_copy:
2135                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
2136                         tp->urg_data = 0;
2137                         tcp_fast_path_check(sk);
2138                 }
2139                 if (used + offset < skb->len)
2140                         continue;
2141
2142                 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2143                         tcp_update_recv_tstamps(skb, &tss);
2144                         has_tss = true;
2145                         has_cmsg = true;
2146                 }
2147                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2148                         goto found_fin_ok;
2149                 if (!(flags & MSG_PEEK))
2150                         sk_eat_skb(sk, skb);
2151                 continue;
2152
2153 found_fin_ok:
2154                 /* Process the FIN. */
2155                 ++*seq;
2156                 if (!(flags & MSG_PEEK))
2157                         sk_eat_skb(sk, skb);
2158                 break;
2159         } while (len > 0);
2160
2161         /* According to UNIX98, msg_name/msg_namelen are ignored
2162          * on connected socket. I was just happy when found this 8) --ANK
2163          */
2164
2165         /* Clean up data we have read: This will do ACK frames. */
2166         tcp_cleanup_rbuf(sk, copied);
2167
2168         release_sock(sk);
2169
2170         if (has_cmsg) {
2171                 if (has_tss)
2172                         tcp_recv_timestamp(msg, sk, &tss);
2173                 if (tp->recvmsg_inq) {
2174                         inq = tcp_inq_hint(sk);
2175                         put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2176                 }
2177         }
2178
2179         return copied;
2180
2181 out:
2182         release_sock(sk);
2183         return err;
2184
2185 recv_urg:
2186         err = tcp_recv_urg(sk, msg, len, flags);
2187         goto out;
2188
2189 recv_sndq:
2190         err = tcp_peek_sndq(sk, msg, len);
2191         goto out;
2192 }
2193 EXPORT_SYMBOL(tcp_recvmsg);
2194
2195 void tcp_set_state(struct sock *sk, int state)
2196 {
2197         int oldstate = sk->sk_state;
2198
2199         /* We defined a new enum for TCP states that are exported in BPF
2200          * so as not force the internal TCP states to be frozen. The
2201          * following checks will detect if an internal state value ever
2202          * differs from the BPF value. If this ever happens, then we will
2203          * need to remap the internal value to the BPF value before calling
2204          * tcp_call_bpf_2arg.
2205          */
2206         BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2207         BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2208         BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2209         BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2210         BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2211         BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2212         BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2213         BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2214         BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2215         BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2216         BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2217         BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2218         BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2219
2220         if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2221                 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2222
2223         switch (state) {
2224         case TCP_ESTABLISHED:
2225                 if (oldstate != TCP_ESTABLISHED)
2226                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2227                 break;
2228
2229         case TCP_CLOSE:
2230                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2231                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2232
2233                 sk->sk_prot->unhash(sk);
2234                 if (inet_csk(sk)->icsk_bind_hash &&
2235                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2236                         inet_put_port(sk);
2237                 /* fall through */
2238         default:
2239                 if (oldstate == TCP_ESTABLISHED)
2240                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2241         }
2242
2243         /* Change state AFTER socket is unhashed to avoid closed
2244          * socket sitting in hash tables.
2245          */
2246         inet_sk_state_store(sk, state);
2247 }
2248 EXPORT_SYMBOL_GPL(tcp_set_state);
2249
2250 /*
2251  *      State processing on a close. This implements the state shift for
2252  *      sending our FIN frame. Note that we only send a FIN for some
2253  *      states. A shutdown() may have already sent the FIN, or we may be
2254  *      closed.
2255  */
2256
2257 static const unsigned char new_state[16] = {
2258   /* current state:        new state:      action:      */
2259   [0 /* (Invalid) */]   = TCP_CLOSE,
2260   [TCP_ESTABLISHED]     = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2261   [TCP_SYN_SENT]        = TCP_CLOSE,
2262   [TCP_SYN_RECV]        = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2263   [TCP_FIN_WAIT1]       = TCP_FIN_WAIT1,
2264   [TCP_FIN_WAIT2]       = TCP_FIN_WAIT2,
2265   [TCP_TIME_WAIT]       = TCP_CLOSE,
2266   [TCP_CLOSE]           = TCP_CLOSE,
2267   [TCP_CLOSE_WAIT]      = TCP_LAST_ACK  | TCP_ACTION_FIN,
2268   [TCP_LAST_ACK]        = TCP_LAST_ACK,
2269   [TCP_LISTEN]          = TCP_CLOSE,
2270   [TCP_CLOSING]         = TCP_CLOSING,
2271   [TCP_NEW_SYN_RECV]    = TCP_CLOSE,    /* should not happen ! */
2272 };
2273
2274 static int tcp_close_state(struct sock *sk)
2275 {
2276         int next = (int)new_state[sk->sk_state];
2277         int ns = next & TCP_STATE_MASK;
2278
2279         tcp_set_state(sk, ns);
2280
2281         return next & TCP_ACTION_FIN;
2282 }
2283
2284 /*
2285  *      Shutdown the sending side of a connection. Much like close except
2286  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2287  */
2288
2289 void tcp_shutdown(struct sock *sk, int how)
2290 {
2291         /*      We need to grab some memory, and put together a FIN,
2292          *      and then put it into the queue to be sent.
2293          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2294          */
2295         if (!(how & SEND_SHUTDOWN))
2296                 return;
2297
2298         /* If we've already sent a FIN, or it's a closed state, skip this. */
2299         if ((1 << sk->sk_state) &
2300             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2301              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2302                 /* Clear out any half completed packets.  FIN if needed. */
2303                 if (tcp_close_state(sk))
2304                         tcp_send_fin(sk);
2305         }
2306 }
2307 EXPORT_SYMBOL(tcp_shutdown);
2308
2309 bool tcp_check_oom(struct sock *sk, int shift)
2310 {
2311         bool too_many_orphans, out_of_socket_memory;
2312
2313         too_many_orphans = tcp_too_many_orphans(sk, shift);
2314         out_of_socket_memory = tcp_out_of_memory(sk);
2315
2316         if (too_many_orphans)
2317                 net_info_ratelimited("too many orphaned sockets\n");
2318         if (out_of_socket_memory)
2319                 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2320         return too_many_orphans || out_of_socket_memory;
2321 }
2322
2323 void tcp_close(struct sock *sk, long timeout)
2324 {
2325         struct sk_buff *skb;
2326         int data_was_unread = 0;
2327         int state;
2328
2329         lock_sock(sk);
2330         sk->sk_shutdown = SHUTDOWN_MASK;
2331
2332         if (sk->sk_state == TCP_LISTEN) {
2333                 tcp_set_state(sk, TCP_CLOSE);
2334
2335                 /* Special case. */
2336                 inet_csk_listen_stop(sk);
2337
2338                 goto adjudge_to_death;
2339         }
2340
2341         /*  We need to flush the recv. buffs.  We do this only on the
2342          *  descriptor close, not protocol-sourced closes, because the
2343          *  reader process may not have drained the data yet!
2344          */
2345         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2346                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2347
2348                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2349                         len--;
2350                 data_was_unread += len;
2351                 __kfree_skb(skb);
2352         }
2353
2354         sk_mem_reclaim(sk);
2355
2356         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2357         if (sk->sk_state == TCP_CLOSE)
2358                 goto adjudge_to_death;
2359
2360         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2361          * data was lost. To witness the awful effects of the old behavior of
2362          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2363          * GET in an FTP client, suspend the process, wait for the client to
2364          * advertise a zero window, then kill -9 the FTP client, wheee...
2365          * Note: timeout is always zero in such a case.
2366          */
2367         if (unlikely(tcp_sk(sk)->repair)) {
2368                 sk->sk_prot->disconnect(sk, 0);
2369         } else if (data_was_unread) {
2370                 /* Unread data was tossed, zap the connection. */
2371                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2372                 tcp_set_state(sk, TCP_CLOSE);
2373                 tcp_send_active_reset(sk, sk->sk_allocation);
2374         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2375                 /* Check zero linger _after_ checking for unread data. */
2376                 sk->sk_prot->disconnect(sk, 0);
2377                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2378         } else if (tcp_close_state(sk)) {
2379                 /* We FIN if the application ate all the data before
2380                  * zapping the connection.
2381                  */
2382
2383                 /* RED-PEN. Formally speaking, we have broken TCP state
2384                  * machine. State transitions:
2385                  *
2386                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2387                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2388                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2389                  *
2390                  * are legal only when FIN has been sent (i.e. in window),
2391                  * rather than queued out of window. Purists blame.
2392                  *
2393                  * F.e. "RFC state" is ESTABLISHED,
2394                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2395                  *
2396                  * The visible declinations are that sometimes
2397                  * we enter time-wait state, when it is not required really
2398                  * (harmless), do not send active resets, when they are
2399                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2400                  * they look as CLOSING or LAST_ACK for Linux)
2401                  * Probably, I missed some more holelets.
2402                  *                                              --ANK
2403                  * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2404                  * in a single packet! (May consider it later but will
2405                  * probably need API support or TCP_CORK SYN-ACK until
2406                  * data is written and socket is closed.)
2407                  */
2408                 tcp_send_fin(sk);
2409         }
2410
2411         sk_stream_wait_close(sk, timeout);
2412
2413 adjudge_to_death:
2414         state = sk->sk_state;
2415         sock_hold(sk);
2416         sock_orphan(sk);
2417
2418         local_bh_disable();
2419         bh_lock_sock(sk);
2420         /* remove backlog if any, without releasing ownership. */
2421         __release_sock(sk);
2422
2423         percpu_counter_inc(sk->sk_prot->orphan_count);
2424
2425         /* Have we already been destroyed by a softirq or backlog? */
2426         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2427                 goto out;
2428
2429         /*      This is a (useful) BSD violating of the RFC. There is a
2430          *      problem with TCP as specified in that the other end could
2431          *      keep a socket open forever with no application left this end.
2432          *      We use a 1 minute timeout (about the same as BSD) then kill
2433          *      our end. If they send after that then tough - BUT: long enough
2434          *      that we won't make the old 4*rto = almost no time - whoops
2435          *      reset mistake.
2436          *
2437          *      Nope, it was not mistake. It is really desired behaviour
2438          *      f.e. on http servers, when such sockets are useless, but
2439          *      consume significant resources. Let's do it with special
2440          *      linger2 option.                                 --ANK
2441          */
2442
2443         if (sk->sk_state == TCP_FIN_WAIT2) {
2444                 struct tcp_sock *tp = tcp_sk(sk);
2445                 if (tp->linger2 < 0) {
2446                         tcp_set_state(sk, TCP_CLOSE);
2447                         tcp_send_active_reset(sk, GFP_ATOMIC);
2448                         __NET_INC_STATS(sock_net(sk),
2449                                         LINUX_MIB_TCPABORTONLINGER);
2450                 } else {
2451                         const int tmo = tcp_fin_time(sk);
2452
2453                         if (tmo > TCP_TIMEWAIT_LEN) {
2454                                 inet_csk_reset_keepalive_timer(sk,
2455                                                 tmo - TCP_TIMEWAIT_LEN);
2456                         } else {
2457                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2458                                 goto out;
2459                         }
2460                 }
2461         }
2462         if (sk->sk_state != TCP_CLOSE) {
2463                 sk_mem_reclaim(sk);
2464                 if (tcp_check_oom(sk, 0)) {
2465                         tcp_set_state(sk, TCP_CLOSE);
2466                         tcp_send_active_reset(sk, GFP_ATOMIC);
2467                         __NET_INC_STATS(sock_net(sk),
2468                                         LINUX_MIB_TCPABORTONMEMORY);
2469                 } else if (!check_net(sock_net(sk))) {
2470                         /* Not possible to send reset; just close */
2471                         tcp_set_state(sk, TCP_CLOSE);
2472                 }
2473         }
2474
2475         if (sk->sk_state == TCP_CLOSE) {
2476                 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2477                 /* We could get here with a non-NULL req if the socket is
2478                  * aborted (e.g., closed with unread data) before 3WHS
2479                  * finishes.
2480                  */
2481                 if (req)
2482                         reqsk_fastopen_remove(sk, req, false);
2483                 inet_csk_destroy_sock(sk);
2484         }
2485         /* Otherwise, socket is reprieved until protocol close. */
2486
2487 out:
2488         bh_unlock_sock(sk);
2489         local_bh_enable();
2490         release_sock(sk);
2491         sock_put(sk);
2492 }
2493 EXPORT_SYMBOL(tcp_close);
2494
2495 /* These states need RST on ABORT according to RFC793 */
2496
2497 static inline bool tcp_need_reset(int state)
2498 {
2499         return (1 << state) &
2500                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2501                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2502 }
2503
2504 static void tcp_rtx_queue_purge(struct sock *sk)
2505 {
2506         struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2507
2508         while (p) {
2509                 struct sk_buff *skb = rb_to_skb(p);
2510
2511                 p = rb_next(p);
2512                 /* Since we are deleting whole queue, no need to
2513                  * list_del(&skb->tcp_tsorted_anchor)
2514                  */
2515                 tcp_rtx_queue_unlink(skb, sk);
2516                 sk_wmem_free_skb(sk, skb);
2517         }
2518 }
2519
2520 void tcp_write_queue_purge(struct sock *sk)
2521 {
2522         struct sk_buff *skb;
2523
2524         tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2525         while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2526                 tcp_skb_tsorted_anchor_cleanup(skb);
2527                 sk_wmem_free_skb(sk, skb);
2528         }
2529         tcp_rtx_queue_purge(sk);
2530         skb = sk->sk_tx_skb_cache;
2531         if (skb) {
2532                 __kfree_skb(skb);
2533                 sk->sk_tx_skb_cache = NULL;
2534         }
2535         INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2536         sk_mem_reclaim(sk);
2537         tcp_clear_all_retrans_hints(tcp_sk(sk));
2538         tcp_sk(sk)->packets_out = 0;
2539         inet_csk(sk)->icsk_backoff = 0;
2540 }
2541
2542 int tcp_disconnect(struct sock *sk, int flags)
2543 {
2544         struct inet_sock *inet = inet_sk(sk);
2545         struct inet_connection_sock *icsk = inet_csk(sk);
2546         struct tcp_sock *tp = tcp_sk(sk);
2547         int old_state = sk->sk_state;
2548
2549         if (old_state != TCP_CLOSE)
2550                 tcp_set_state(sk, TCP_CLOSE);
2551
2552         /* ABORT function of RFC793 */
2553         if (old_state == TCP_LISTEN) {
2554                 inet_csk_listen_stop(sk);
2555         } else if (unlikely(tp->repair)) {
2556                 sk->sk_err = ECONNABORTED;
2557         } else if (tcp_need_reset(old_state) ||
2558                    (tp->snd_nxt != tp->write_seq &&
2559                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2560                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2561                  * states
2562                  */
2563                 tcp_send_active_reset(sk, gfp_any());
2564                 sk->sk_err = ECONNRESET;
2565         } else if (old_state == TCP_SYN_SENT)
2566                 sk->sk_err = ECONNRESET;
2567
2568         tcp_clear_xmit_timers(sk);
2569         __skb_queue_purge(&sk->sk_receive_queue);
2570         if (sk->sk_rx_skb_cache) {
2571                 __kfree_skb(sk->sk_rx_skb_cache);
2572                 sk->sk_rx_skb_cache = NULL;
2573         }
2574         tp->copied_seq = tp->rcv_nxt;
2575         tp->urg_data = 0;
2576         tcp_write_queue_purge(sk);
2577         tcp_fastopen_active_disable_ofo_check(sk);
2578         skb_rbtree_purge(&tp->out_of_order_queue);
2579
2580         inet->inet_dport = 0;
2581
2582         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2583                 inet_reset_saddr(sk);
2584
2585         sk->sk_shutdown = 0;
2586         sock_reset_flag(sk, SOCK_DONE);
2587         tp->srtt_us = 0;
2588         tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
2589         tp->rcv_rtt_last_tsecr = 0;
2590         tp->write_seq += tp->max_window + 2;
2591         if (tp->write_seq == 0)
2592                 tp->write_seq = 1;
2593         icsk->icsk_backoff = 0;
2594         tp->snd_cwnd = 2;
2595         icsk->icsk_probes_out = 0;
2596         icsk->icsk_rto = TCP_TIMEOUT_INIT;
2597         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2598         tp->snd_cwnd = TCP_INIT_CWND;
2599         tp->snd_cwnd_cnt = 0;
2600         tp->window_clamp = 0;
2601         tp->delivered_ce = 0;
2602         tcp_set_ca_state(sk, TCP_CA_Open);
2603         tp->is_sack_reneg = 0;
2604         tcp_clear_retrans(tp);
2605         inet_csk_delack_init(sk);
2606         /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2607          * issue in __tcp_select_window()
2608          */
2609         icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2610         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2611         __sk_dst_reset(sk);
2612         dst_release(sk->sk_rx_dst);
2613         sk->sk_rx_dst = NULL;
2614         tcp_saved_syn_free(tp);
2615         tp->compressed_ack = 0;
2616         tp->bytes_sent = 0;
2617         tp->bytes_acked = 0;
2618         tp->bytes_received = 0;
2619         tp->bytes_retrans = 0;
2620         tp->duplicate_sack[0].start_seq = 0;
2621         tp->duplicate_sack[0].end_seq = 0;
2622         tp->dsack_dups = 0;
2623         tp->reord_seen = 0;
2624         tp->retrans_out = 0;
2625         tp->sacked_out = 0;
2626         tp->tlp_high_seq = 0;
2627         tp->last_oow_ack_time = 0;
2628         /* There's a bubble in the pipe until at least the first ACK. */
2629         tp->app_limited = ~0U;
2630         tp->rack.mstamp = 0;
2631         tp->rack.advanced = 0;
2632         tp->rack.reo_wnd_steps = 1;
2633         tp->rack.last_delivered = 0;
2634         tp->rack.reo_wnd_persist = 0;
2635         tp->rack.dsack_seen = 0;
2636         tp->syn_data_acked = 0;
2637         tp->rx_opt.saw_tstamp = 0;
2638         tp->rx_opt.dsack = 0;
2639         tp->rx_opt.num_sacks = 0;
2640
2641
2642         /* Clean up fastopen related fields */
2643         tcp_free_fastopen_req(tp);
2644         inet->defer_connect = 0;
2645
2646         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2647
2648         if (sk->sk_frag.page) {
2649                 put_page(sk->sk_frag.page);
2650                 sk->sk_frag.page = NULL;
2651                 sk->sk_frag.offset = 0;
2652         }
2653
2654         sk->sk_error_report(sk);
2655         return 0;
2656 }
2657 EXPORT_SYMBOL(tcp_disconnect);
2658
2659 static inline bool tcp_can_repair_sock(const struct sock *sk)
2660 {
2661         return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2662                 (sk->sk_state != TCP_LISTEN);
2663 }
2664
2665 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2666 {
2667         struct tcp_repair_window opt;
2668
2669         if (!tp->repair)
2670                 return -EPERM;
2671
2672         if (len != sizeof(opt))
2673                 return -EINVAL;
2674
2675         if (copy_from_user(&opt, optbuf, sizeof(opt)))
2676                 return -EFAULT;
2677
2678         if (opt.max_window < opt.snd_wnd)
2679                 return -EINVAL;
2680
2681         if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2682                 return -EINVAL;
2683
2684         if (after(opt.rcv_wup, tp->rcv_nxt))
2685                 return -EINVAL;
2686
2687         tp->snd_wl1     = opt.snd_wl1;
2688         tp->snd_wnd     = opt.snd_wnd;
2689         tp->max_window  = opt.max_window;
2690
2691         tp->rcv_wnd     = opt.rcv_wnd;
2692         tp->rcv_wup     = opt.rcv_wup;
2693
2694         return 0;
2695 }
2696
2697 static int tcp_repair_options_est(struct sock *sk,
2698                 struct tcp_repair_opt __user *optbuf, unsigned int len)
2699 {
2700         struct tcp_sock *tp = tcp_sk(sk);
2701         struct tcp_repair_opt opt;
2702
2703         while (len >= sizeof(opt)) {
2704                 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2705                         return -EFAULT;
2706
2707                 optbuf++;
2708                 len -= sizeof(opt);
2709
2710                 switch (opt.opt_code) {
2711                 case TCPOPT_MSS:
2712                         tp->rx_opt.mss_clamp = opt.opt_val;
2713                         tcp_mtup_init(sk);
2714                         break;
2715                 case TCPOPT_WINDOW:
2716                         {
2717                                 u16 snd_wscale = opt.opt_val & 0xFFFF;
2718                                 u16 rcv_wscale = opt.opt_val >> 16;
2719
2720                                 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
2721                                         return -EFBIG;
2722
2723                                 tp->rx_opt.snd_wscale = snd_wscale;
2724                                 tp->rx_opt.rcv_wscale = rcv_wscale;
2725                                 tp->rx_opt.wscale_ok = 1;
2726                         }
2727                         break;
2728                 case TCPOPT_SACK_PERM:
2729                         if (opt.opt_val != 0)
2730                                 return -EINVAL;
2731
2732                         tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2733                         break;
2734                 case TCPOPT_TIMESTAMP:
2735                         if (opt.opt_val != 0)
2736                                 return -EINVAL;
2737
2738                         tp->rx_opt.tstamp_ok = 1;
2739                         break;
2740                 }
2741         }
2742
2743         return 0;
2744 }
2745
2746 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
2747 EXPORT_SYMBOL(tcp_tx_delay_enabled);
2748
2749 static void tcp_enable_tx_delay(void)
2750 {
2751         if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
2752                 static int __tcp_tx_delay_enabled = 0;
2753
2754                 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
2755                         static_branch_enable(&tcp_tx_delay_enabled);
2756                         pr_info("TCP_TX_DELAY enabled\n");
2757                 }
2758         }
2759 }
2760
2761 /*
2762  *      Socket option code for TCP.
2763  */
2764 static int do_tcp_setsockopt(struct sock *sk, int level,
2765                 int optname, char __user *optval, unsigned int optlen)
2766 {
2767         struct tcp_sock *tp = tcp_sk(sk);
2768         struct inet_connection_sock *icsk = inet_csk(sk);
2769         struct net *net = sock_net(sk);
2770         int val;
2771         int err = 0;
2772
2773         /* These are data/string values, all the others are ints */
2774         switch (optname) {
2775         case TCP_CONGESTION: {
2776                 char name[TCP_CA_NAME_MAX];
2777
2778                 if (optlen < 1)
2779                         return -EINVAL;
2780
2781                 val = strncpy_from_user(name, optval,
2782                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2783                 if (val < 0)
2784                         return -EFAULT;
2785                 name[val] = 0;
2786
2787                 lock_sock(sk);
2788                 err = tcp_set_congestion_control(sk, name, true, true);
2789                 release_sock(sk);
2790                 return err;
2791         }
2792         case TCP_ULP: {
2793                 char name[TCP_ULP_NAME_MAX];
2794
2795                 if (optlen < 1)
2796                         return -EINVAL;
2797
2798                 val = strncpy_from_user(name, optval,
2799                                         min_t(long, TCP_ULP_NAME_MAX - 1,
2800                                               optlen));
2801                 if (val < 0)
2802                         return -EFAULT;
2803                 name[val] = 0;
2804
2805                 lock_sock(sk);
2806                 err = tcp_set_ulp(sk, name);
2807                 release_sock(sk);
2808                 return err;
2809         }
2810         case TCP_FASTOPEN_KEY: {
2811                 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
2812                 __u8 *backup_key = NULL;
2813
2814                 /* Allow a backup key as well to facilitate key rotation
2815                  * First key is the active one.
2816                  */
2817                 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
2818                     optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
2819                         return -EINVAL;
2820
2821                 if (copy_from_user(key, optval, optlen))
2822                         return -EFAULT;
2823
2824                 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
2825                         backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
2826
2827                 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
2828         }
2829         default:
2830                 /* fallthru */
2831                 break;
2832         }
2833
2834         if (optlen < sizeof(int))
2835                 return -EINVAL;
2836
2837         if (get_user(val, (int __user *)optval))
2838                 return -EFAULT;
2839
2840         lock_sock(sk);
2841
2842         switch (optname) {
2843         case TCP_MAXSEG:
2844                 /* Values greater than interface MTU won't take effect. However
2845                  * at the point when this call is done we typically don't yet
2846                  * know which interface is going to be used
2847                  */
2848                 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
2849                         err = -EINVAL;
2850                         break;
2851                 }
2852                 tp->rx_opt.user_mss = val;
2853                 break;
2854
2855         case TCP_NODELAY:
2856                 if (val) {
2857                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2858                          * this option on corked socket is remembered, but
2859                          * it is not activated until cork is cleared.
2860                          *
2861                          * However, when TCP_NODELAY is set we make
2862                          * an explicit push, which overrides even TCP_CORK
2863                          * for currently queued segments.
2864                          */
2865                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2866                         tcp_push_pending_frames(sk);
2867                 } else {
2868                         tp->nonagle &= ~TCP_NAGLE_OFF;
2869                 }
2870                 break;
2871
2872         case TCP_THIN_LINEAR_TIMEOUTS:
2873                 if (val < 0 || val > 1)
2874                         err = -EINVAL;
2875                 else
2876                         tp->thin_lto = val;
2877                 break;
2878
2879         case TCP_THIN_DUPACK:
2880                 if (val < 0 || val > 1)
2881                         err = -EINVAL;
2882                 break;
2883
2884         case TCP_REPAIR:
2885                 if (!tcp_can_repair_sock(sk))
2886                         err = -EPERM;
2887                 else if (val == TCP_REPAIR_ON) {
2888                         tp->repair = 1;
2889                         sk->sk_reuse = SK_FORCE_REUSE;
2890                         tp->repair_queue = TCP_NO_QUEUE;
2891                 } else if (val == TCP_REPAIR_OFF) {
2892                         tp->repair = 0;
2893                         sk->sk_reuse = SK_NO_REUSE;
2894                         tcp_send_window_probe(sk);
2895                 } else if (val == TCP_REPAIR_OFF_NO_WP) {
2896                         tp->repair = 0;
2897                         sk->sk_reuse = SK_NO_REUSE;
2898                 } else
2899                         err = -EINVAL;
2900
2901                 break;
2902
2903         case TCP_REPAIR_QUEUE:
2904                 if (!tp->repair)
2905                         err = -EPERM;
2906                 else if ((unsigned int)val < TCP_QUEUES_NR)
2907                         tp->repair_queue = val;
2908                 else
2909                         err = -EINVAL;
2910                 break;
2911
2912         case TCP_QUEUE_SEQ:
2913                 if (sk->sk_state != TCP_CLOSE)
2914                         err = -EPERM;
2915                 else if (tp->repair_queue == TCP_SEND_QUEUE)
2916                         tp->write_seq = val;
2917                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2918                         tp->rcv_nxt = val;
2919                 else
2920                         err = -EINVAL;
2921                 break;
2922
2923         case TCP_REPAIR_OPTIONS:
2924                 if (!tp->repair)
2925                         err = -EINVAL;
2926                 else if (sk->sk_state == TCP_ESTABLISHED)
2927                         err = tcp_repair_options_est(sk,
2928                                         (struct tcp_repair_opt __user *)optval,
2929                                         optlen);
2930                 else
2931                         err = -EPERM;
2932                 break;
2933
2934         case TCP_CORK:
2935                 /* When set indicates to always queue non-full frames.
2936                  * Later the user clears this option and we transmit
2937                  * any pending partial frames in the queue.  This is
2938                  * meant to be used alongside sendfile() to get properly
2939                  * filled frames when the user (for example) must write
2940                  * out headers with a write() call first and then use
2941                  * sendfile to send out the data parts.
2942                  *
2943                  * TCP_CORK can be set together with TCP_NODELAY and it is
2944                  * stronger than TCP_NODELAY.
2945                  */
2946                 if (val) {
2947                         tp->nonagle |= TCP_NAGLE_CORK;
2948                 } else {
2949                         tp->nonagle &= ~TCP_NAGLE_CORK;
2950                         if (tp->nonagle&TCP_NAGLE_OFF)
2951                                 tp->nonagle |= TCP_NAGLE_PUSH;
2952                         tcp_push_pending_frames(sk);
2953                 }
2954                 break;
2955
2956         case TCP_KEEPIDLE:
2957                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2958                         err = -EINVAL;
2959                 else {
2960                         tp->keepalive_time = val * HZ;
2961                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2962                             !((1 << sk->sk_state) &
2963                               (TCPF_CLOSE | TCPF_LISTEN))) {
2964                                 u32 elapsed = keepalive_time_elapsed(tp);
2965                                 if (tp->keepalive_time > elapsed)
2966                                         elapsed = tp->keepalive_time - elapsed;
2967                                 else
2968                                         elapsed = 0;
2969                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2970                         }
2971                 }
2972                 break;
2973         case TCP_KEEPINTVL:
2974                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2975                         err = -EINVAL;
2976                 else
2977                         tp->keepalive_intvl = val * HZ;
2978                 break;
2979         case TCP_KEEPCNT:
2980                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2981                         err = -EINVAL;
2982                 else
2983                         tp->keepalive_probes = val;
2984                 break;
2985         case TCP_SYNCNT:
2986                 if (val < 1 || val > MAX_TCP_SYNCNT)
2987                         err = -EINVAL;
2988                 else
2989                         icsk->icsk_syn_retries = val;
2990                 break;
2991
2992         case TCP_SAVE_SYN:
2993                 if (val < 0 || val > 1)
2994                         err = -EINVAL;
2995                 else
2996                         tp->save_syn = val;
2997                 break;
2998
2999         case TCP_LINGER2:
3000                 if (val < 0)
3001                         tp->linger2 = -1;
3002                 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
3003                         tp->linger2 = 0;
3004                 else
3005                         tp->linger2 = val * HZ;
3006                 break;
3007
3008         case TCP_DEFER_ACCEPT:
3009                 /* Translate value in seconds to number of retransmits */
3010                 icsk->icsk_accept_queue.rskq_defer_accept =
3011                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3012                                         TCP_RTO_MAX / HZ);
3013                 break;
3014
3015         case TCP_WINDOW_CLAMP:
3016                 if (!val) {
3017                         if (sk->sk_state != TCP_CLOSE) {
3018                                 err = -EINVAL;
3019                                 break;
3020                         }
3021                         tp->window_clamp = 0;
3022                 } else
3023                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3024                                                 SOCK_MIN_RCVBUF / 2 : val;
3025                 break;
3026
3027         case TCP_QUICKACK:
3028                 if (!val) {
3029                         inet_csk_enter_pingpong_mode(sk);
3030                 } else {
3031                         inet_csk_exit_pingpong_mode(sk);
3032                         if ((1 << sk->sk_state) &
3033                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3034                             inet_csk_ack_scheduled(sk)) {
3035                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
3036                                 tcp_cleanup_rbuf(sk, 1);
3037                                 if (!(val & 1))
3038                                         inet_csk_enter_pingpong_mode(sk);
3039                         }
3040                 }
3041                 break;
3042
3043 #ifdef CONFIG_TCP_MD5SIG
3044         case TCP_MD5SIG:
3045         case TCP_MD5SIG_EXT:
3046                 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
3047                         err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3048                 else
3049                         err = -EINVAL;
3050                 break;
3051 #endif
3052         case TCP_USER_TIMEOUT:
3053                 /* Cap the max time in ms TCP will retry or probe the window
3054                  * before giving up and aborting (ETIMEDOUT) a connection.
3055                  */
3056                 if (val < 0)
3057                         err = -EINVAL;
3058                 else
3059                         icsk->icsk_user_timeout = val;
3060                 break;
3061
3062         case TCP_FASTOPEN:
3063                 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3064                     TCPF_LISTEN))) {
3065                         tcp_fastopen_init_key_once(net);
3066
3067                         fastopen_queue_tune(sk, val);
3068                 } else {
3069                         err = -EINVAL;
3070                 }
3071                 break;
3072         case TCP_FASTOPEN_CONNECT:
3073                 if (val > 1 || val < 0) {
3074                         err = -EINVAL;
3075                 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
3076                         if (sk->sk_state == TCP_CLOSE)
3077                                 tp->fastopen_connect = val;
3078                         else
3079                                 err = -EINVAL;
3080                 } else {
3081                         err = -EOPNOTSUPP;
3082                 }
3083                 break;
3084         case TCP_FASTOPEN_NO_COOKIE:
3085                 if (val > 1 || val < 0)
3086                         err = -EINVAL;
3087                 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3088                         err = -EINVAL;
3089                 else
3090                         tp->fastopen_no_cookie = val;
3091                 break;
3092         case TCP_TIMESTAMP:
3093                 if (!tp->repair)
3094                         err = -EPERM;
3095                 else
3096                         tp->tsoffset = val - tcp_time_stamp_raw();
3097                 break;
3098         case TCP_REPAIR_WINDOW:
3099                 err = tcp_repair_set_window(tp, optval, optlen);
3100                 break;
3101         case TCP_NOTSENT_LOWAT:
3102                 tp->notsent_lowat = val;
3103                 sk->sk_write_space(sk);
3104                 break;
3105         case TCP_INQ:
3106                 if (val > 1 || val < 0)
3107                         err = -EINVAL;
3108                 else
3109                         tp->recvmsg_inq = val;
3110                 break;
3111         case TCP_TX_DELAY:
3112                 if (val)
3113                         tcp_enable_tx_delay();
3114                 tp->tcp_tx_delay = val;
3115                 break;
3116         default:
3117                 err = -ENOPROTOOPT;
3118                 break;
3119         }
3120
3121         release_sock(sk);
3122         return err;
3123 }
3124
3125 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
3126                    unsigned int optlen)
3127 {
3128         const struct inet_connection_sock *icsk = inet_csk(sk);
3129
3130         if (level != SOL_TCP)
3131                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3132                                                      optval, optlen);
3133         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3134 }
3135 EXPORT_SYMBOL(tcp_setsockopt);
3136
3137 #ifdef CONFIG_COMPAT
3138 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
3139                           char __user *optval, unsigned int optlen)
3140 {
3141         if (level != SOL_TCP)
3142                 return inet_csk_compat_setsockopt(sk, level, optname,
3143                                                   optval, optlen);
3144         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3145 }
3146 EXPORT_SYMBOL(compat_tcp_setsockopt);
3147 #endif
3148
3149 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3150                                       struct tcp_info *info)
3151 {
3152         u64 stats[__TCP_CHRONO_MAX], total = 0;
3153         enum tcp_chrono i;
3154
3155         for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3156                 stats[i] = tp->chrono_stat[i - 1];
3157                 if (i == tp->chrono_type)
3158                         stats[i] += tcp_jiffies32 - tp->chrono_start;
3159                 stats[i] *= USEC_PER_SEC / HZ;
3160                 total += stats[i];
3161         }
3162
3163         info->tcpi_busy_time = total;
3164         info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3165         info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3166 }
3167
3168 /* Return information about state of tcp endpoint in API format. */
3169 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3170 {
3171         const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3172         const struct inet_connection_sock *icsk = inet_csk(sk);
3173         unsigned long rate;
3174         u32 now;
3175         u64 rate64;
3176         bool slow;
3177
3178         memset(info, 0, sizeof(*info));
3179         if (sk->sk_type != SOCK_STREAM)
3180                 return;
3181
3182         info->tcpi_state = inet_sk_state_load(sk);
3183
3184         /* Report meaningful fields for all TCP states, including listeners */
3185         rate = READ_ONCE(sk->sk_pacing_rate);
3186         rate64 = (rate != ~0UL) ? rate : ~0ULL;
3187         info->tcpi_pacing_rate = rate64;
3188
3189         rate = READ_ONCE(sk->sk_max_pacing_rate);
3190         rate64 = (rate != ~0UL) ? rate : ~0ULL;
3191         info->tcpi_max_pacing_rate = rate64;
3192
3193         info->tcpi_reordering = tp->reordering;
3194         info->tcpi_snd_cwnd = tp->snd_cwnd;
3195
3196         if (info->tcpi_state == TCP_LISTEN) {
3197                 /* listeners aliased fields :
3198                  * tcpi_unacked -> Number of children ready for accept()
3199                  * tcpi_sacked  -> max backlog
3200                  */
3201                 info->tcpi_unacked = sk->sk_ack_backlog;
3202                 info->tcpi_sacked = sk->sk_max_ack_backlog;
3203                 return;
3204         }
3205
3206         slow = lock_sock_fast(sk);
3207
3208         info->tcpi_ca_state = icsk->icsk_ca_state;
3209         info->tcpi_retransmits = icsk->icsk_retransmits;
3210         info->tcpi_probes = icsk->icsk_probes_out;
3211         info->tcpi_backoff = icsk->icsk_backoff;
3212
3213         if (tp->rx_opt.tstamp_ok)
3214                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3215         if (tcp_is_sack(tp))
3216                 info->tcpi_options |= TCPI_OPT_SACK;
3217         if (tp->rx_opt.wscale_ok) {
3218                 info->tcpi_options |= TCPI_OPT_WSCALE;
3219                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3220                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3221         }
3222
3223         if (tp->ecn_flags & TCP_ECN_OK)
3224                 info->tcpi_options |= TCPI_OPT_ECN;
3225         if (tp->ecn_flags & TCP_ECN_SEEN)
3226                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3227         if (tp->syn_data_acked)
3228                 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3229
3230         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3231         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3232         info->tcpi_snd_mss = tp->mss_cache;
3233         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3234
3235         info->tcpi_unacked = tp->packets_out;
3236         info->tcpi_sacked = tp->sacked_out;
3237
3238         info->tcpi_lost = tp->lost_out;
3239         info->tcpi_retrans = tp->retrans_out;
3240
3241         now = tcp_jiffies32;
3242         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3243         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3244         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3245
3246         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3247         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3248         info->tcpi_rtt = tp->srtt_us >> 3;
3249         info->tcpi_rttvar = tp->mdev_us >> 2;
3250         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3251         info->tcpi_advmss = tp->advmss;
3252
3253         info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3254         info->tcpi_rcv_space = tp->rcvq_space.space;
3255
3256         info->tcpi_total_retrans = tp->total_retrans;
3257
3258         info->tcpi_bytes_acked = tp->bytes_acked;
3259         info->tcpi_bytes_received = tp->bytes_received;
3260         info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3261         tcp_get_info_chrono_stats(tp, info);
3262
3263         info->tcpi_segs_out = tp->segs_out;
3264         info->tcpi_segs_in = tp->segs_in;
3265
3266         info->tcpi_min_rtt = tcp_min_rtt(tp);
3267         info->tcpi_data_segs_in = tp->data_segs_in;
3268         info->tcpi_data_segs_out = tp->data_segs_out;
3269
3270         info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3271         rate64 = tcp_compute_delivery_rate(tp);
3272         if (rate64)
3273                 info->tcpi_delivery_rate = rate64;
3274         info->tcpi_delivered = tp->delivered;
3275         info->tcpi_delivered_ce = tp->delivered_ce;
3276         info->tcpi_bytes_sent = tp->bytes_sent;
3277         info->tcpi_bytes_retrans = tp->bytes_retrans;
3278         info->tcpi_dsack_dups = tp->dsack_dups;
3279         info->tcpi_reord_seen = tp->reord_seen;
3280         unlock_sock_fast(sk, slow);
3281 }
3282 EXPORT_SYMBOL_GPL(tcp_get_info);
3283
3284 static size_t tcp_opt_stats_get_size(void)
3285 {
3286         return
3287                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3288                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3289                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3290                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3291                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3292                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3293                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3294                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3295                 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3296                 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3297                 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3298                 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3299                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3300                 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3301                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3302                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3303                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3304                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3305                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3306                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3307                 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3308                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3309                 0;
3310 }
3311
3312 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk)
3313 {
3314         const struct tcp_sock *tp = tcp_sk(sk);
3315         struct sk_buff *stats;
3316         struct tcp_info info;
3317         unsigned long rate;
3318         u64 rate64;
3319
3320         stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3321         if (!stats)
3322                 return NULL;
3323
3324         tcp_get_info_chrono_stats(tp, &info);
3325         nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3326                           info.tcpi_busy_time, TCP_NLA_PAD);
3327         nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3328                           info.tcpi_rwnd_limited, TCP_NLA_PAD);
3329         nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3330                           info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3331         nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3332                           tp->data_segs_out, TCP_NLA_PAD);
3333         nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3334                           tp->total_retrans, TCP_NLA_PAD);
3335
3336         rate = READ_ONCE(sk->sk_pacing_rate);
3337         rate64 = (rate != ~0UL) ? rate : ~0ULL;
3338         nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3339
3340         rate64 = tcp_compute_delivery_rate(tp);
3341         nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3342
3343         nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3344         nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3345         nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3346
3347         nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3348         nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3349         nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3350         nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3351         nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3352
3353         nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3354         nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3355
3356         nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3357                           TCP_NLA_PAD);
3358         nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3359                           TCP_NLA_PAD);
3360         nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3361         nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3362         nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3363
3364         return stats;
3365 }
3366
3367 static int do_tcp_getsockopt(struct sock *sk, int level,
3368                 int optname, char __user *optval, int __user *optlen)
3369 {
3370         struct inet_connection_sock *icsk = inet_csk(sk);
3371         struct tcp_sock *tp = tcp_sk(sk);
3372         struct net *net = sock_net(sk);
3373         int val, len;
3374
3375         if (get_user(len, optlen))
3376                 return -EFAULT;
3377
3378         len = min_t(unsigned int, len, sizeof(int));
3379
3380         if (len < 0)
3381                 return -EINVAL;
3382
3383         switch (optname) {
3384         case TCP_MAXSEG:
3385                 val = tp->mss_cache;
3386                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3387                         val = tp->rx_opt.user_mss;
3388                 if (tp->repair)
3389                         val = tp->rx_opt.mss_clamp;
3390                 break;
3391         case TCP_NODELAY:
3392                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3393                 break;
3394         case TCP_CORK:
3395                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3396                 break;
3397         case TCP_KEEPIDLE:
3398                 val = keepalive_time_when(tp) / HZ;
3399                 break;
3400         case TCP_KEEPINTVL:
3401                 val = keepalive_intvl_when(tp) / HZ;
3402                 break;
3403         case TCP_KEEPCNT:
3404                 val = keepalive_probes(tp);
3405                 break;
3406         case TCP_SYNCNT:
3407                 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3408                 break;
3409         case TCP_LINGER2:
3410                 val = tp->linger2;
3411                 if (val >= 0)
3412                         val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3413                 break;
3414         case TCP_DEFER_ACCEPT:
3415                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3416                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3417                 break;
3418         case TCP_WINDOW_CLAMP:
3419                 val = tp->window_clamp;
3420                 break;
3421         case TCP_INFO: {
3422                 struct tcp_info info;
3423
3424                 if (get_user(len, optlen))
3425                         return -EFAULT;
3426
3427                 tcp_get_info(sk, &info);
3428
3429                 len = min_t(unsigned int, len, sizeof(info));
3430                 if (put_user(len, optlen))
3431                         return -EFAULT;
3432                 if (copy_to_user(optval, &info, len))
3433                         return -EFAULT;
3434                 return 0;
3435         }
3436         case TCP_CC_INFO: {
3437                 const struct tcp_congestion_ops *ca_ops;
3438                 union tcp_cc_info info;
3439                 size_t sz = 0;
3440                 int attr;
3441
3442                 if (get_user(len, optlen))
3443                         return -EFAULT;
3444
3445                 ca_ops = icsk->icsk_ca_ops;
3446                 if (ca_ops && ca_ops->get_info)
3447                         sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3448
3449                 len = min_t(unsigned int, len, sz);
3450                 if (put_user(len, optlen))
3451                         return -EFAULT;
3452                 if (copy_to_user(optval, &info, len))
3453                         return -EFAULT;
3454                 return 0;
3455         }
3456         case TCP_QUICKACK:
3457                 val = !inet_csk_in_pingpong_mode(sk);
3458                 break;
3459
3460         case TCP_CONGESTION:
3461                 if (get_user(len, optlen))
3462                         return -EFAULT;
3463                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
3464                 if (put_user(len, optlen))
3465                         return -EFAULT;
3466                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3467                         return -EFAULT;
3468                 return 0;
3469
3470         case TCP_ULP:
3471                 if (get_user(len, optlen))
3472                         return -EFAULT;
3473                 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
3474                 if (!icsk->icsk_ulp_ops) {
3475                         if (put_user(0, optlen))
3476                                 return -EFAULT;
3477                         return 0;
3478                 }
3479                 if (put_user(len, optlen))
3480                         return -EFAULT;
3481                 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
3482                         return -EFAULT;
3483                 return 0;
3484
3485         case TCP_FASTOPEN_KEY: {
3486                 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3487                 struct tcp_fastopen_context *ctx;
3488                 unsigned int key_len = 0;
3489
3490                 if (get_user(len, optlen))
3491                         return -EFAULT;
3492
3493                 rcu_read_lock();
3494                 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
3495                 if (ctx) {
3496                         key_len = tcp_fastopen_context_len(ctx) *
3497                                         TCP_FASTOPEN_KEY_LENGTH;
3498                         memcpy(&key[0], &ctx->key[0], key_len);
3499                 }
3500                 rcu_read_unlock();
3501
3502                 len = min_t(unsigned int, len, key_len);
3503                 if (put_user(len, optlen))
3504                         return -EFAULT;
3505                 if (copy_to_user(optval, key, len))
3506                         return -EFAULT;
3507                 return 0;
3508         }
3509         case TCP_THIN_LINEAR_TIMEOUTS:
3510                 val = tp->thin_lto;
3511                 break;
3512
3513         case TCP_THIN_DUPACK:
3514                 val = 0;
3515                 break;
3516
3517         case TCP_REPAIR:
3518                 val = tp->repair;
3519                 break;
3520
3521         case TCP_REPAIR_QUEUE:
3522                 if (tp->repair)
3523                         val = tp->repair_queue;
3524                 else
3525                         return -EINVAL;
3526                 break;
3527
3528         case TCP_REPAIR_WINDOW: {
3529                 struct tcp_repair_window opt;
3530
3531                 if (get_user(len, optlen))
3532                         return -EFAULT;
3533
3534                 if (len != sizeof(opt))
3535                         return -EINVAL;
3536
3537                 if (!tp->repair)
3538                         return -EPERM;
3539
3540                 opt.snd_wl1     = tp->snd_wl1;
3541                 opt.snd_wnd     = tp->snd_wnd;
3542                 opt.max_window  = tp->max_window;
3543                 opt.rcv_wnd     = tp->rcv_wnd;
3544                 opt.rcv_wup     = tp->rcv_wup;
3545
3546                 if (copy_to_user(optval, &opt, len))
3547                         return -EFAULT;
3548                 return 0;
3549         }
3550         case TCP_QUEUE_SEQ:
3551                 if (tp->repair_queue == TCP_SEND_QUEUE)
3552                         val = tp->write_seq;
3553                 else if (tp->repair_queue == TCP_RECV_QUEUE)
3554                         val = tp->rcv_nxt;
3555                 else
3556                         return -EINVAL;
3557                 break;
3558
3559         case TCP_USER_TIMEOUT:
3560                 val = icsk->icsk_user_timeout;
3561                 break;
3562
3563         case TCP_FASTOPEN:
3564                 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3565                 break;
3566
3567         case TCP_FASTOPEN_CONNECT:
3568                 val = tp->fastopen_connect;
3569                 break;
3570
3571         case TCP_FASTOPEN_NO_COOKIE:
3572                 val = tp->fastopen_no_cookie;
3573                 break;
3574
3575         case TCP_TX_DELAY:
3576                 val = tp->tcp_tx_delay;
3577                 break;
3578
3579         case TCP_TIMESTAMP:
3580                 val = tcp_time_stamp_raw() + tp->tsoffset;
3581                 break;
3582         case TCP_NOTSENT_LOWAT:
3583                 val = tp->notsent_lowat;
3584                 break;
3585         case TCP_INQ:
3586                 val = tp->recvmsg_inq;
3587                 break;
3588         case TCP_SAVE_SYN:
3589                 val = tp->save_syn;
3590                 break;
3591         case TCP_SAVED_SYN: {
3592                 if (get_user(len, optlen))
3593                         return -EFAULT;
3594
3595                 lock_sock(sk);
3596                 if (tp->saved_syn) {
3597                         if (len < tp->saved_syn[0]) {
3598                                 if (put_user(tp->saved_syn[0], optlen)) {
3599                                         release_sock(sk);
3600                                         return -EFAULT;
3601                                 }
3602                                 release_sock(sk);
3603                                 return -EINVAL;
3604                         }
3605                         len = tp->saved_syn[0];
3606                         if (put_user(len, optlen)) {
3607                                 release_sock(sk);
3608                                 return -EFAULT;
3609                         }
3610                         if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3611                                 release_sock(sk);
3612                                 return -EFAULT;
3613                         }
3614                         tcp_saved_syn_free(tp);
3615                         release_sock(sk);
3616                 } else {
3617                         release_sock(sk);
3618                         len = 0;
3619                         if (put_user(len, optlen))
3620                                 return -EFAULT;
3621                 }
3622                 return 0;
3623         }
3624 #ifdef CONFIG_MMU
3625         case TCP_ZEROCOPY_RECEIVE: {
3626                 struct tcp_zerocopy_receive zc;
3627                 int err;
3628
3629                 if (get_user(len, optlen))
3630                         return -EFAULT;
3631                 if (len != sizeof(zc))
3632                         return -EINVAL;
3633                 if (copy_from_user(&zc, optval, len))
3634                         return -EFAULT;
3635                 lock_sock(sk);
3636                 err = tcp_zerocopy_receive(sk, &zc);
3637                 release_sock(sk);
3638                 if (!err && copy_to_user(optval, &zc, len))
3639                         err = -EFAULT;
3640                 return err;
3641         }
3642 #endif
3643         default:
3644                 return -ENOPROTOOPT;
3645         }
3646
3647         if (put_user(len, optlen))
3648                 return -EFAULT;
3649         if (copy_to_user(optval, &val, len))
3650                 return -EFAULT;
3651         return 0;
3652 }
3653
3654 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3655                    int __user *optlen)
3656 {
3657         struct inet_connection_sock *icsk = inet_csk(sk);
3658
3659         if (level != SOL_TCP)
3660                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3661                                                      optval, optlen);
3662         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3663 }
3664 EXPORT_SYMBOL(tcp_getsockopt);
3665
3666 #ifdef CONFIG_COMPAT
3667 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3668                           char __user *optval, int __user *optlen)
3669 {
3670         if (level != SOL_TCP)
3671                 return inet_csk_compat_getsockopt(sk, level, optname,
3672                                                   optval, optlen);
3673         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3674 }
3675 EXPORT_SYMBOL(compat_tcp_getsockopt);
3676 #endif
3677
3678 #ifdef CONFIG_TCP_MD5SIG
3679 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3680 static DEFINE_MUTEX(tcp_md5sig_mutex);
3681 static bool tcp_md5sig_pool_populated = false;
3682
3683 static void __tcp_alloc_md5sig_pool(void)
3684 {
3685         struct crypto_ahash *hash;
3686         int cpu;
3687
3688         hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3689         if (IS_ERR(hash))
3690                 return;
3691
3692         for_each_possible_cpu(cpu) {
3693                 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3694                 struct ahash_request *req;
3695
3696                 if (!scratch) {
3697                         scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3698                                                sizeof(struct tcphdr),
3699                                                GFP_KERNEL,
3700                                                cpu_to_node(cpu));
3701                         if (!scratch)
3702                                 return;
3703                         per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3704                 }
3705                 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3706                         continue;
3707
3708                 req = ahash_request_alloc(hash, GFP_KERNEL);
3709                 if (!req)
3710                         return;
3711
3712                 ahash_request_set_callback(req, 0, NULL, NULL);
3713
3714                 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3715         }
3716         /* before setting tcp_md5sig_pool_populated, we must commit all writes
3717          * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3718          */
3719         smp_wmb();
3720         tcp_md5sig_pool_populated = true;
3721 }
3722
3723 bool tcp_alloc_md5sig_pool(void)
3724 {
3725         if (unlikely(!tcp_md5sig_pool_populated)) {
3726                 mutex_lock(&tcp_md5sig_mutex);
3727
3728                 if (!tcp_md5sig_pool_populated) {
3729                         __tcp_alloc_md5sig_pool();
3730                         if (tcp_md5sig_pool_populated)
3731                                 static_branch_inc(&tcp_md5_needed);
3732                 }
3733
3734                 mutex_unlock(&tcp_md5sig_mutex);
3735         }
3736         return tcp_md5sig_pool_populated;
3737 }
3738 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3739
3740
3741 /**
3742  *      tcp_get_md5sig_pool - get md5sig_pool for this user
3743  *
3744  *      We use percpu structure, so if we succeed, we exit with preemption
3745  *      and BH disabled, to make sure another thread or softirq handling
3746  *      wont try to get same context.
3747  */
3748 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3749 {
3750         local_bh_disable();
3751
3752         if (tcp_md5sig_pool_populated) {
3753                 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3754                 smp_rmb();
3755                 return this_cpu_ptr(&tcp_md5sig_pool);
3756         }
3757         local_bh_enable();
3758         return NULL;
3759 }
3760 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3761
3762 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3763                           const struct sk_buff *skb, unsigned int header_len)
3764 {
3765         struct scatterlist sg;
3766         const struct tcphdr *tp = tcp_hdr(skb);
3767         struct ahash_request *req = hp->md5_req;
3768         unsigned int i;
3769         const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3770                                            skb_headlen(skb) - header_len : 0;
3771         const struct skb_shared_info *shi = skb_shinfo(skb);
3772         struct sk_buff *frag_iter;
3773
3774         sg_init_table(&sg, 1);
3775
3776         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3777         ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3778         if (crypto_ahash_update(req))
3779                 return 1;
3780
3781         for (i = 0; i < shi->nr_frags; ++i) {
3782                 const struct skb_frag_struct *f = &shi->frags[i];
3783                 unsigned int offset = f->page_offset;
3784                 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3785
3786                 sg_set_page(&sg, page, skb_frag_size(f),
3787                             offset_in_page(offset));
3788                 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3789                 if (crypto_ahash_update(req))
3790                         return 1;
3791         }
3792
3793         skb_walk_frags(skb, frag_iter)
3794                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3795                         return 1;
3796
3797         return 0;
3798 }
3799 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3800
3801 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3802 {
3803         struct scatterlist sg;
3804
3805         sg_init_one(&sg, key->key, key->keylen);
3806         ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen);
3807         return crypto_ahash_update(hp->md5_req);
3808 }
3809 EXPORT_SYMBOL(tcp_md5_hash_key);
3810
3811 #endif
3812
3813 void tcp_done(struct sock *sk)
3814 {
3815         struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3816
3817         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3818                 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3819
3820         tcp_set_state(sk, TCP_CLOSE);
3821         tcp_clear_xmit_timers(sk);
3822         if (req)
3823                 reqsk_fastopen_remove(sk, req, false);
3824
3825         sk->sk_shutdown = SHUTDOWN_MASK;
3826
3827         if (!sock_flag(sk, SOCK_DEAD))
3828                 sk->sk_state_change(sk);
3829         else
3830                 inet_csk_destroy_sock(sk);
3831 }
3832 EXPORT_SYMBOL_GPL(tcp_done);
3833
3834 int tcp_abort(struct sock *sk, int err)
3835 {
3836         if (!sk_fullsock(sk)) {
3837                 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3838                         struct request_sock *req = inet_reqsk(sk);
3839
3840                         local_bh_disable();
3841                         inet_csk_reqsk_queue_drop(req->rsk_listener, req);
3842                         local_bh_enable();
3843                         return 0;
3844                 }
3845                 return -EOPNOTSUPP;
3846         }
3847
3848         /* Don't race with userspace socket closes such as tcp_close. */
3849         lock_sock(sk);
3850
3851         if (sk->sk_state == TCP_LISTEN) {
3852                 tcp_set_state(sk, TCP_CLOSE);
3853                 inet_csk_listen_stop(sk);
3854         }
3855
3856         /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3857         local_bh_disable();
3858         bh_lock_sock(sk);
3859
3860         if (!sock_flag(sk, SOCK_DEAD)) {
3861                 sk->sk_err = err;
3862                 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3863                 smp_wmb();
3864                 sk->sk_error_report(sk);
3865                 if (tcp_need_reset(sk->sk_state))
3866                         tcp_send_active_reset(sk, GFP_ATOMIC);
3867                 tcp_done(sk);
3868         }
3869
3870         bh_unlock_sock(sk);
3871         local_bh_enable();
3872         tcp_write_queue_purge(sk);
3873         release_sock(sk);
3874         return 0;
3875 }
3876 EXPORT_SYMBOL_GPL(tcp_abort);
3877
3878 extern struct tcp_congestion_ops tcp_reno;
3879
3880 static __initdata unsigned long thash_entries;
3881 static int __init set_thash_entries(char *str)
3882 {
3883         ssize_t ret;
3884
3885         if (!str)
3886                 return 0;
3887
3888         ret = kstrtoul(str, 0, &thash_entries);
3889         if (ret)
3890                 return 0;
3891
3892         return 1;
3893 }
3894 __setup("thash_entries=", set_thash_entries);
3895
3896 static void __init tcp_init_mem(void)
3897 {
3898         unsigned long limit = nr_free_buffer_pages() / 16;
3899
3900         limit = max(limit, 128UL);
3901         sysctl_tcp_mem[0] = limit / 4 * 3;              /* 4.68 % */
3902         sysctl_tcp_mem[1] = limit;                      /* 6.25 % */
3903         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;      /* 9.37 % */
3904 }
3905
3906 void __init tcp_init(void)
3907 {
3908         int max_rshare, max_wshare, cnt;
3909         unsigned long limit;
3910         unsigned int i;
3911
3912         BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
3913         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3914                      FIELD_SIZEOF(struct sk_buff, cb));
3915
3916         percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3917         percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3918         inet_hashinfo_init(&tcp_hashinfo);
3919         inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
3920                             thash_entries, 21,  /* one slot per 2 MB*/
3921                             0, 64 * 1024);
3922         tcp_hashinfo.bind_bucket_cachep =
3923                 kmem_cache_create("tcp_bind_bucket",
3924                                   sizeof(struct inet_bind_bucket), 0,
3925                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3926
3927         /* Size and allocate the main established and bind bucket
3928          * hash tables.
3929          *
3930          * The methodology is similar to that of the buffer cache.
3931          */
3932         tcp_hashinfo.ehash =
3933                 alloc_large_system_hash("TCP established",
3934                                         sizeof(struct inet_ehash_bucket),
3935                                         thash_entries,
3936                                         17, /* one slot per 128 KB of memory */
3937                                         0,
3938                                         NULL,
3939                                         &tcp_hashinfo.ehash_mask,
3940                                         0,
3941                                         thash_entries ? 0 : 512 * 1024);
3942         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3943                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3944
3945         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3946                 panic("TCP: failed to alloc ehash_locks");
3947         tcp_hashinfo.bhash =
3948                 alloc_large_system_hash("TCP bind",
3949                                         sizeof(struct inet_bind_hashbucket),
3950                                         tcp_hashinfo.ehash_mask + 1,
3951                                         17, /* one slot per 128 KB of memory */
3952                                         0,
3953                                         &tcp_hashinfo.bhash_size,
3954                                         NULL,
3955                                         0,
3956                                         64 * 1024);
3957         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3958         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3959                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3960                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3961         }
3962
3963
3964         cnt = tcp_hashinfo.ehash_mask + 1;
3965         sysctl_tcp_max_orphans = cnt / 2;
3966
3967         tcp_init_mem();
3968         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3969         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3970         max_wshare = min(4UL*1024*1024, limit);
3971         max_rshare = min(6UL*1024*1024, limit);
3972
3973         init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3974         init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
3975         init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3976
3977         init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3978         init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
3979         init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
3980
3981         pr_info("Hash tables configured (established %u bind %u)\n",
3982                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3983
3984         tcp_v4_init();
3985         tcp_metrics_init();
3986         BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3987         tcp_tasklet_init();
3988 }