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 /* In some cases, both sendpage() and sendmsg() could have added
 939  * an skb to the write queue, but failed adding payload on it.
 940  * We need to remove it to consume less memory, but more
 941  * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
 942  * users.
 943  */
 944 static void tcp_remove_empty_skb(struct sock *sk, struct sk_buff *skb)
 945 {
 946         if (skb && !skb->len) {
 947                 tcp_unlink_write_queue(skb, sk);
 948                 if (tcp_write_queue_empty(sk))
 949                         tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
 950                 sk_wmem_free_skb(sk, skb);
 951         }
 952 }
 953
 954 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
 955                          size_t size, int flags)
 956 {
 957         struct tcp_sock *tp = tcp_sk(sk);
 958         int mss_now, size_goal;
 959         int err;
 960         ssize_t copied;
 961         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 962
 963         if (IS_ENABLED(CONFIG_DEBUG_VM) &&
 964             WARN_ONCE(PageSlab(page), "page must not be a Slab one"))
 965                 return -EINVAL;
 966
 967         /* Wait for a connection to finish. One exception is TCP Fast Open
 968          * (passive side) where data is allowed to be sent before a connection
 969          * is fully established.
 970          */
 971         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
 972             !tcp_passive_fastopen(sk)) {
 973                 err = sk_stream_wait_connect(sk, &timeo);
 974                 if (err != 0)
 975                         goto out_err;
 976         }
 977
 978         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 979
 980         mss_now = tcp_send_mss(sk, &size_goal, flags);
 981         copied = 0;
 982
 983         err = -EPIPE;
 984         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 985                 goto out_err;
 986
 987         while (size > 0) {
 988                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 989                 int copy, i;
 990                 bool can_coalesce;
 991
 992                 if (!skb || (copy = size_goal - skb->len) <= 0 ||
 993                     !tcp_skb_can_collapse_to(skb)) {
 994 new_segment:
 995                         if (!sk_stream_memory_free(sk))
 996                                 goto wait_for_sndbuf;
 997
 998                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
 999                                         tcp_rtx_and_write_queues_empty(sk));
1000                         if (!skb)
1001                                 goto wait_for_memory;
1002
1003 #ifdef CONFIG_TLS_DEVICE
1004                         skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
1005 #endif
1006                         skb_entail(sk, skb);
1007                         copy = size_goal;
1008                 }
1009
1010                 if (copy > size)
1011                         copy = size;
1012
1013                 i = skb_shinfo(skb)->nr_frags;
1014                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
1015                 if (!can_coalesce && i >= sysctl_max_skb_frags) {
1016                         tcp_mark_push(tp, skb);
1017                         goto new_segment;
1018                 }
1019                 if (!sk_wmem_schedule(sk, copy))
1020                         goto wait_for_memory;
1021
1022                 if (can_coalesce) {
1023                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1024                 } else {
1025                         get_page(page);
1026                         skb_fill_page_desc(skb, i, page, offset, copy);
1027                 }
1028
1029                 if (!(flags & MSG_NO_SHARED_FRAGS))
1030                         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1031
1032                 skb->len += copy;
1033                 skb->data_len += copy;
1034                 skb->truesize += copy;
1035                 sk->sk_wmem_queued += copy;
1036                 sk_mem_charge(sk, copy);
1037                 skb->ip_summed = CHECKSUM_PARTIAL;
1038                 tp->write_seq += copy;
1039                 TCP_SKB_CB(skb)->end_seq += copy;
1040                 tcp_skb_pcount_set(skb, 0);
1041
1042                 if (!copied)
1043                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1044
1045                 copied += copy;
1046                 offset += copy;
1047                 size -= copy;
1048                 if (!size)
1049                         goto out;
1050
1051                 if (skb->len < size_goal || (flags & MSG_OOB))
1052                         continue;
1053
1054                 if (forced_push(tp)) {
1055                         tcp_mark_push(tp, skb);
1056                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1057                 } else if (skb == tcp_send_head(sk))
1058                         tcp_push_one(sk, mss_now);
1059                 continue;
1060
1061 wait_for_sndbuf:
1062                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1063 wait_for_memory:
1064                 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1065                          TCP_NAGLE_PUSH, size_goal);
1066
1067                 err = sk_stream_wait_memory(sk, &timeo);
1068                 if (err != 0)
1069                         goto do_error;
1070
1071                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1072         }
1073
1074 out:
1075         if (copied) {
1076                 tcp_tx_timestamp(sk, sk->sk_tsflags);
1077                 if (!(flags & MSG_SENDPAGE_NOTLAST))
1078                         tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1079         }
1080         return copied;
1081
1082 do_error:
1083         tcp_remove_empty_skb(sk, tcp_write_queue_tail(sk));
1084         if (copied)
1085                 goto out;
1086 out_err:
1087         /* make sure we wake any epoll edge trigger waiter */
1088         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1089                      err == -EAGAIN)) {
1090                 sk->sk_write_space(sk);
1091                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1092         }
1093         return sk_stream_error(sk, flags, err);
1094 }
1095 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1096
1097 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1098                         size_t size, int flags)
1099 {
1100         if (!(sk->sk_route_caps & NETIF_F_SG))
1101                 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1102
1103         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1104
1105         return do_tcp_sendpages(sk, page, offset, size, flags);
1106 }
1107 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1108
1109 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1110                  size_t size, int flags)
1111 {
1112         int ret;
1113
1114         lock_sock(sk);
1115         ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1116         release_sock(sk);
1117
1118         return ret;
1119 }
1120 EXPORT_SYMBOL(tcp_sendpage);
1121
1122 void tcp_free_fastopen_req(struct tcp_sock *tp)
1123 {
1124         if (tp->fastopen_req) {
1125                 kfree(tp->fastopen_req);
1126                 tp->fastopen_req = NULL;
1127         }
1128 }
1129
1130 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1131                                 int *copied, size_t size,
1132                                 struct ubuf_info *uarg)
1133 {
1134         struct tcp_sock *tp = tcp_sk(sk);
1135         struct inet_sock *inet = inet_sk(sk);
1136         struct sockaddr *uaddr = msg->msg_name;
1137         int err, flags;
1138
1139         if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1140             (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1141              uaddr->sa_family == AF_UNSPEC))
1142                 return -EOPNOTSUPP;
1143         if (tp->fastopen_req)
1144                 return -EALREADY; /* Another Fast Open is in progress */
1145
1146         tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1147                                    sk->sk_allocation);
1148         if (unlikely(!tp->fastopen_req))
1149                 return -ENOBUFS;
1150         tp->fastopen_req->data = msg;
1151         tp->fastopen_req->size = size;
1152         tp->fastopen_req->uarg = uarg;
1153
1154         if (inet->defer_connect) {
1155                 err = tcp_connect(sk);
1156                 /* Same failure procedure as in tcp_v4/6_connect */
1157                 if (err) {
1158                         tcp_set_state(sk, TCP_CLOSE);
1159                         inet->inet_dport = 0;
1160                         sk->sk_route_caps = 0;
1161                 }
1162         }
1163         flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1164         err = __inet_stream_connect(sk->sk_socket, uaddr,
1165                                     msg->msg_namelen, flags, 1);
1166         /* fastopen_req could already be freed in __inet_stream_connect
1167          * if the connection times out or gets rst
1168          */
1169         if (tp->fastopen_req) {
1170                 *copied = tp->fastopen_req->copied;
1171                 tcp_free_fastopen_req(tp);
1172                 inet->defer_connect = 0;
1173         }
1174         return err;
1175 }
1176
1177 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1178 {
1179         struct tcp_sock *tp = tcp_sk(sk);
1180         struct ubuf_info *uarg = NULL;
1181         struct sk_buff *skb;
1182         struct sockcm_cookie sockc;
1183         int flags, err, copied = 0;
1184         int mss_now = 0, size_goal, copied_syn = 0;
1185         int process_backlog = 0;
1186         bool zc = false;
1187         long timeo;
1188
1189         flags = msg->msg_flags;
1190
1191         if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
1192                 skb = tcp_write_queue_tail(sk);
1193                 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
1194                 if (!uarg) {
1195                         err = -ENOBUFS;
1196                         goto out_err;
1197                 }
1198
1199                 zc = sk->sk_route_caps & NETIF_F_SG;
1200                 if (!zc)
1201                         uarg->zerocopy = 0;
1202         }
1203
1204         if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1205             !tp->repair) {
1206                 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1207                 if (err == -EINPROGRESS && copied_syn > 0)
1208                         goto out;
1209                 else if (err)
1210                         goto out_err;
1211         }
1212
1213         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1214
1215         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1216
1217         /* Wait for a connection to finish. One exception is TCP Fast Open
1218          * (passive side) where data is allowed to be sent before a connection
1219          * is fully established.
1220          */
1221         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1222             !tcp_passive_fastopen(sk)) {
1223                 err = sk_stream_wait_connect(sk, &timeo);
1224                 if (err != 0)
1225                         goto do_error;
1226         }
1227
1228         if (unlikely(tp->repair)) {
1229                 if (tp->repair_queue == TCP_RECV_QUEUE) {
1230                         copied = tcp_send_rcvq(sk, msg, size);
1231                         goto out_nopush;
1232                 }
1233
1234                 err = -EINVAL;
1235                 if (tp->repair_queue == TCP_NO_QUEUE)
1236                         goto out_err;
1237
1238                 /* 'common' sending to sendq */
1239         }
1240
1241         sockcm_init(&sockc, sk);
1242         if (msg->msg_controllen) {
1243                 err = sock_cmsg_send(sk, msg, &sockc);
1244                 if (unlikely(err)) {
1245                         err = -EINVAL;
1246                         goto out_err;
1247                 }
1248         }
1249
1250         /* This should be in poll */
1251         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1252
1253         /* Ok commence sending. */
1254         copied = 0;
1255
1256 restart:
1257         mss_now = tcp_send_mss(sk, &size_goal, flags);
1258
1259         err = -EPIPE;
1260         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1261                 goto do_error;
1262
1263         while (msg_data_left(msg)) {
1264                 int copy = 0;
1265
1266                 skb = tcp_write_queue_tail(sk);
1267                 if (skb)
1268                         copy = size_goal - skb->len;
1269
1270                 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1271                         bool first_skb;
1272
1273 new_segment:
1274                         if (!sk_stream_memory_free(sk))
1275                                 goto wait_for_sndbuf;
1276
1277                         if (unlikely(process_backlog >= 16)) {
1278                                 process_backlog = 0;
1279                                 if (sk_flush_backlog(sk))
1280                                         goto restart;
1281                         }
1282                         first_skb = tcp_rtx_and_write_queues_empty(sk);
1283                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
1284                                                   first_skb);
1285                         if (!skb)
1286                                 goto wait_for_memory;
1287
1288                         process_backlog++;
1289                         skb->ip_summed = CHECKSUM_PARTIAL;
1290
1291                         skb_entail(sk, skb);
1292                         copy = size_goal;
1293
1294                         /* All packets are restored as if they have
1295                          * already been sent. skb_mstamp_ns isn't set to
1296                          * avoid wrong rtt estimation.
1297                          */
1298                         if (tp->repair)
1299                                 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1300                 }
1301
1302                 /* Try to append data to the end of skb. */
1303                 if (copy > msg_data_left(msg))
1304                         copy = msg_data_left(msg);
1305
1306                 /* Where to copy to? */
1307                 if (skb_availroom(skb) > 0 && !zc) {
1308                         /* We have some space in skb head. Superb! */
1309                         copy = min_t(int, copy, skb_availroom(skb));
1310                         err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1311                         if (err)
1312                                 goto do_fault;
1313                 } else if (!zc) {
1314                         bool merge = true;
1315                         int i = skb_shinfo(skb)->nr_frags;
1316                         struct page_frag *pfrag = sk_page_frag(sk);
1317
1318                         if (!sk_page_frag_refill(sk, pfrag))
1319                                 goto wait_for_memory;
1320
1321                         if (!skb_can_coalesce(skb, i, pfrag->page,
1322                                               pfrag->offset)) {
1323                                 if (i >= sysctl_max_skb_frags) {
1324                                         tcp_mark_push(tp, skb);
1325                                         goto new_segment;
1326                                 }
1327                                 merge = false;
1328                         }
1329
1330                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1331
1332                         if (!sk_wmem_schedule(sk, copy))
1333                                 goto wait_for_memory;
1334
1335                         err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1336                                                        pfrag->page,
1337                                                        pfrag->offset,
1338                                                        copy);
1339                         if (err)
1340                                 goto do_error;
1341
1342                         /* Update the skb. */
1343                         if (merge) {
1344                                 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1345                         } else {
1346                                 skb_fill_page_desc(skb, i, pfrag->page,
1347                                                    pfrag->offset, copy);
1348                                 page_ref_inc(pfrag->page);
1349                         }
1350                         pfrag->offset += copy;
1351                 } else {
1352                         err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1353                         if (err == -EMSGSIZE || err == -EEXIST) {
1354                                 tcp_mark_push(tp, skb);
1355                                 goto new_segment;
1356                         }
1357                         if (err < 0)
1358                                 goto do_error;
1359                         copy = err;
1360                 }
1361
1362                 if (!copied)
1363                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1364
1365                 tp->write_seq += copy;
1366                 TCP_SKB_CB(skb)->end_seq += copy;
1367                 tcp_skb_pcount_set(skb, 0);
1368
1369                 copied += copy;
1370                 if (!msg_data_left(msg)) {
1371                         if (unlikely(flags & MSG_EOR))
1372                                 TCP_SKB_CB(skb)->eor = 1;
1373                         goto out;
1374                 }
1375
1376                 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1377                         continue;
1378
1379                 if (forced_push(tp)) {
1380                         tcp_mark_push(tp, skb);
1381                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1382                 } else if (skb == tcp_send_head(sk))
1383                         tcp_push_one(sk, mss_now);
1384                 continue;
1385
1386 wait_for_sndbuf:
1387                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1388 wait_for_memory:
1389                 if (copied)
1390                         tcp_push(sk, flags & ~MSG_MORE, mss_now,
1391                                  TCP_NAGLE_PUSH, size_goal);
1392
1393                 err = sk_stream_wait_memory(sk, &timeo);
1394                 if (err != 0)
1395                         goto do_error;
1396
1397                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1398         }
1399
1400 out:
1401         if (copied) {
1402                 tcp_tx_timestamp(sk, sockc.tsflags);
1403                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1404         }
1405 out_nopush:
1406         sock_zerocopy_put(uarg);
1407         return copied + copied_syn;
1408
1409 do_error:
1410         skb = tcp_write_queue_tail(sk);
1411 do_fault:
1412         tcp_remove_empty_skb(sk, skb);
1413
1414         if (copied + copied_syn)
1415                 goto out;
1416 out_err:
1417         sock_zerocopy_put_abort(uarg, true);
1418         err = sk_stream_error(sk, flags, err);
1419         /* make sure we wake any epoll edge trigger waiter */
1420         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1421                      err == -EAGAIN)) {
1422                 sk->sk_write_space(sk);
1423                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1424         }
1425         return err;
1426 }
1427 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1428
1429 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1430 {
1431         int ret;
1432
1433         lock_sock(sk);
1434         ret = tcp_sendmsg_locked(sk, msg, size);
1435         release_sock(sk);
1436
1437         return ret;
1438 }
1439 EXPORT_SYMBOL(tcp_sendmsg);
1440
1441 /*
1442  *      Handle reading urgent data. BSD has very simple semantics for
1443  *      this, no blocking and very strange errors 8)
1444  */
1445
1446 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1447 {
1448         struct tcp_sock *tp = tcp_sk(sk);
1449
1450         /* No URG data to read. */
1451         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1452             tp->urg_data == TCP_URG_READ)
1453                 return -EINVAL; /* Yes this is right ! */
1454
1455         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1456                 return -ENOTCONN;
1457
1458         if (tp->urg_data & TCP_URG_VALID) {
1459                 int err = 0;
1460                 char c = tp->urg_data;
1461
1462                 if (!(flags & MSG_PEEK))
1463                         tp->urg_data = TCP_URG_READ;
1464
1465                 /* Read urgent data. */
1466                 msg->msg_flags |= MSG_OOB;
1467
1468                 if (len > 0) {
1469                         if (!(flags & MSG_TRUNC))
1470                                 err = memcpy_to_msg(msg, &c, 1);
1471                         len = 1;
1472                 } else
1473                         msg->msg_flags |= MSG_TRUNC;
1474
1475                 return err ? -EFAULT : len;
1476         }
1477
1478         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1479                 return 0;
1480
1481         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1482          * the available implementations agree in this case:
1483          * this call should never block, independent of the
1484          * blocking state of the socket.
1485          * Mike <pall@rz.uni-karlsruhe.de>
1486          */
1487         return -EAGAIN;
1488 }
1489
1490 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1491 {
1492         struct sk_buff *skb;
1493         int copied = 0, err = 0;
1494
1495         /* XXX -- need to support SO_PEEK_OFF */
1496
1497         skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1498                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1499                 if (err)
1500                         return err;
1501                 copied += skb->len;
1502         }
1503
1504         skb_queue_walk(&sk->sk_write_queue, skb) {
1505                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1506                 if (err)
1507                         break;
1508
1509                 copied += skb->len;
1510         }
1511
1512         return err ?: copied;
1513 }
1514
1515 /* Clean up the receive buffer for full frames taken by the user,
1516  * then send an ACK if necessary.  COPIED is the number of bytes
1517  * tcp_recvmsg has given to the user so far, it speeds up the
1518  * calculation of whether or not we must ACK for the sake of
1519  * a window update.
1520  */
1521 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1522 {
1523         struct tcp_sock *tp = tcp_sk(sk);
1524         bool time_to_ack = false;
1525
1526         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1527
1528         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1529              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1530              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1531
1532         if (inet_csk_ack_scheduled(sk)) {
1533                 const struct inet_connection_sock *icsk = inet_csk(sk);
1534                    /* Delayed ACKs frequently hit locked sockets during bulk
1535                     * receive. */
1536                 if (icsk->icsk_ack.blocked ||
1537                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1538                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1539                     /*
1540                      * If this read emptied read buffer, we send ACK, if
1541                      * connection is not bidirectional, user drained
1542                      * receive buffer and there was a small segment
1543                      * in queue.
1544                      */
1545                     (copied > 0 &&
1546                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1547                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1548                        !inet_csk_in_pingpong_mode(sk))) &&
1549                       !atomic_read(&sk->sk_rmem_alloc)))
1550                         time_to_ack = true;
1551         }
1552
1553         /* We send an ACK if we can now advertise a non-zero window
1554          * which has been raised "significantly".
1555          *
1556          * Even if window raised up to infinity, do not send window open ACK
1557          * in states, where we will not receive more. It is useless.
1558          */
1559         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1560                 __u32 rcv_window_now = tcp_receive_window(tp);
1561
1562                 /* Optimize, __tcp_select_window() is not cheap. */
1563                 if (2*rcv_window_now <= tp->window_clamp) {
1564                         __u32 new_window = __tcp_select_window(sk);
1565
1566                         /* Send ACK now, if this read freed lots of space
1567                          * in our buffer. Certainly, new_window is new window.
1568                          * We can advertise it now, if it is not less than current one.
1569                          * "Lots" means "at least twice" here.
1570                          */
1571                         if (new_window && new_window >= 2 * rcv_window_now)
1572                                 time_to_ack = true;
1573                 }
1574         }
1575         if (time_to_ack)
1576                 tcp_send_ack(sk);
1577 }
1578
1579 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1580 {
1581         struct sk_buff *skb;
1582         u32 offset;
1583
1584         while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1585                 offset = seq - TCP_SKB_CB(skb)->seq;
1586                 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1587                         pr_err_once("%s: found a SYN, please report !\n", __func__);
1588                         offset--;
1589                 }
1590                 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1591                         *off = offset;
1592                         return skb;
1593                 }
1594                 /* This looks weird, but this can happen if TCP collapsing
1595                  * splitted a fat GRO packet, while we released socket lock
1596                  * in skb_splice_bits()
1597                  */
1598                 sk_eat_skb(sk, skb);
1599         }
1600         return NULL;
1601 }
1602
1603 /*
1604  * This routine provides an alternative to tcp_recvmsg() for routines
1605  * that would like to handle copying from skbuffs directly in 'sendfile'
1606  * fashion.
1607  * Note:
1608  *      - It is assumed that the socket was locked by the caller.
1609  *      - The routine does not block.
1610  *      - At present, there is no support for reading OOB data
1611  *        or for 'peeking' the socket using this routine
1612  *        (although both would be easy to implement).
1613  */
1614 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1615                   sk_read_actor_t recv_actor)
1616 {
1617         struct sk_buff *skb;
1618         struct tcp_sock *tp = tcp_sk(sk);
1619         u32 seq = tp->copied_seq;
1620         u32 offset;
1621         int copied = 0;
1622
1623         if (sk->sk_state == TCP_LISTEN)
1624                 return -ENOTCONN;
1625         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1626                 if (offset < skb->len) {
1627                         int used;
1628                         size_t len;
1629
1630                         len = skb->len - offset;
1631                         /* Stop reading if we hit a patch of urgent data */
1632                         if (tp->urg_data) {
1633                                 u32 urg_offset = tp->urg_seq - seq;
1634                                 if (urg_offset < len)
1635                                         len = urg_offset;
1636                                 if (!len)
1637                                         break;
1638                         }
1639                         used = recv_actor(desc, skb, offset, len);
1640                         if (used <= 0) {
1641                                 if (!copied)
1642                                         copied = used;
1643                                 break;
1644                         } else if (used <= len) {
1645                                 seq += used;
1646                                 copied += used;
1647                                 offset += used;
1648                         }
1649                         /* If recv_actor drops the lock (e.g. TCP splice
1650                          * receive) the skb pointer might be invalid when
1651                          * getting here: tcp_collapse might have deleted it
1652                          * while aggregating skbs from the socket queue.
1653                          */
1654                         skb = tcp_recv_skb(sk, seq - 1, &offset);
1655                         if (!skb)
1656                                 break;
1657                         /* TCP coalescing might have appended data to the skb.
1658                          * Try to splice more frags
1659                          */
1660                         if (offset + 1 != skb->len)
1661                                 continue;
1662                 }
1663                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1664                         sk_eat_skb(sk, skb);
1665                         ++seq;
1666                         break;
1667                 }
1668                 sk_eat_skb(sk, skb);
1669                 if (!desc->count)
1670                         break;
1671                 tp->copied_seq = seq;
1672         }
1673         tp->copied_seq = seq;
1674
1675         tcp_rcv_space_adjust(sk);
1676
1677         /* Clean up data we have read: This will do ACK frames. */
1678         if (copied > 0) {
1679                 tcp_recv_skb(sk, seq, &offset);
1680                 tcp_cleanup_rbuf(sk, copied);
1681         }
1682         return copied;
1683 }
1684 EXPORT_SYMBOL(tcp_read_sock);
1685
1686 int tcp_peek_len(struct socket *sock)
1687 {
1688         return tcp_inq(sock->sk);
1689 }
1690 EXPORT_SYMBOL(tcp_peek_len);
1691
1692 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1693 int tcp_set_rcvlowat(struct sock *sk, int val)
1694 {
1695         int cap;
1696
1697         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1698                 cap = sk->sk_rcvbuf >> 1;
1699         else
1700                 cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1;
1701         val = min(val, cap);
1702         sk->sk_rcvlowat = val ? : 1;
1703
1704         /* Check if we need to signal EPOLLIN right now */
1705         tcp_data_ready(sk);
1706
1707         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1708                 return 0;
1709
1710         val <<= 1;
1711         if (val > sk->sk_rcvbuf) {
1712                 sk->sk_rcvbuf = val;
1713                 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1714         }
1715         return 0;
1716 }
1717 EXPORT_SYMBOL(tcp_set_rcvlowat);
1718
1719 #ifdef CONFIG_MMU
1720 static const struct vm_operations_struct tcp_vm_ops = {
1721 };
1722
1723 int tcp_mmap(struct file *file, struct socket *sock,
1724              struct vm_area_struct *vma)
1725 {
1726         if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1727                 return -EPERM;
1728         vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1729
1730         /* Instruct vm_insert_page() to not down_read(mmap_sem) */
1731         vma->vm_flags |= VM_MIXEDMAP;
1732
1733         vma->vm_ops = &tcp_vm_ops;
1734         return 0;
1735 }
1736 EXPORT_SYMBOL(tcp_mmap);
1737
1738 static int tcp_zerocopy_receive(struct sock *sk,
1739                                 struct tcp_zerocopy_receive *zc)
1740 {
1741         unsigned long address = (unsigned long)zc->address;
1742         const skb_frag_t *frags = NULL;
1743         u32 length = 0, seq, offset;
1744         struct vm_area_struct *vma;
1745         struct sk_buff *skb = NULL;
1746         struct tcp_sock *tp;
1747         int inq;
1748         int ret;
1749
1750         if (address & (PAGE_SIZE - 1) || address != zc->address)
1751                 return -EINVAL;
1752
1753         if (sk->sk_state == TCP_LISTEN)
1754                 return -ENOTCONN;
1755
1756         sock_rps_record_flow(sk);
1757
1758         down_read(&current->mm->mmap_sem);
1759
1760         ret = -EINVAL;
1761         vma = find_vma(current->mm, address);
1762         if (!vma || vma->vm_start > address || vma->vm_ops != &tcp_vm_ops)
1763                 goto out;
1764         zc->length = min_t(unsigned long, zc->length, vma->vm_end - address);
1765
1766         tp = tcp_sk(sk);
1767         seq = tp->copied_seq;
1768         inq = tcp_inq(sk);
1769         zc->length = min_t(u32, zc->length, inq);
1770         zc->length &= ~(PAGE_SIZE - 1);
1771         if (zc->length) {
1772                 zap_page_range(vma, address, zc->length);
1773                 zc->recv_skip_hint = 0;
1774         } else {
1775                 zc->recv_skip_hint = inq;
1776         }
1777         ret = 0;
1778         while (length + PAGE_SIZE <= zc->length) {
1779                 if (zc->recv_skip_hint < PAGE_SIZE) {
1780                         if (skb) {
1781                                 skb = skb->next;
1782                                 offset = seq - TCP_SKB_CB(skb)->seq;
1783                         } else {
1784                                 skb = tcp_recv_skb(sk, seq, &offset);
1785                         }
1786
1787                         zc->recv_skip_hint = skb->len - offset;
1788                         offset -= skb_headlen(skb);
1789                         if ((int)offset < 0 || skb_has_frag_list(skb))
1790                                 break;
1791                         frags = skb_shinfo(skb)->frags;
1792                         while (offset) {
1793                                 if (skb_frag_size(frags) > offset)
1794                                         goto out;
1795                                 offset -= skb_frag_size(frags);
1796                                 frags++;
1797                         }
1798                 }
1799                 if (skb_frag_size(frags) != PAGE_SIZE || skb_frag_off(frags)) {
1800                         int remaining = zc->recv_skip_hint;
1801                         int size = skb_frag_size(frags);
1802
1803                         while (remaining && (size != PAGE_SIZE ||
1804                                              skb_frag_off(frags))) {
1805                                 remaining -= size;
1806                                 frags++;
1807                                 size = skb_frag_size(frags);
1808                         }
1809                         zc->recv_skip_hint -= remaining;
1810                         break;
1811                 }
1812                 ret = vm_insert_page(vma, address + length,
1813                                      skb_frag_page(frags));
1814                 if (ret)
1815                         break;
1816                 length += PAGE_SIZE;
1817                 seq += PAGE_SIZE;
1818                 zc->recv_skip_hint -= PAGE_SIZE;
1819                 frags++;
1820         }
1821 out:
1822         up_read(&current->mm->mmap_sem);
1823         if (length) {
1824                 tp->copied_seq = seq;
1825                 tcp_rcv_space_adjust(sk);
1826
1827                 /* Clean up data we have read: This will do ACK frames. */
1828                 tcp_recv_skb(sk, seq, &offset);
1829                 tcp_cleanup_rbuf(sk, length);
1830                 ret = 0;
1831                 if (length == zc->length)
1832                         zc->recv_skip_hint = 0;
1833         } else {
1834                 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
1835                         ret = -EIO;
1836         }
1837         zc->length = length;
1838         return ret;
1839 }
1840 #endif
1841
1842 static void tcp_update_recv_tstamps(struct sk_buff *skb,
1843                                     struct scm_timestamping_internal *tss)
1844 {
1845         if (skb->tstamp)
1846                 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1847         else
1848                 tss->ts[0] = (struct timespec64) {0};
1849
1850         if (skb_hwtstamps(skb)->hwtstamp)
1851                 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1852         else
1853                 tss->ts[2] = (struct timespec64) {0};
1854 }
1855
1856 /* Similar to __sock_recv_timestamp, but does not require an skb */
1857 static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
1858                                struct scm_timestamping_internal *tss)
1859 {
1860         int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
1861         bool has_timestamping = false;
1862
1863         if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
1864                 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
1865                         if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
1866                                 if (new_tstamp) {
1867                                         struct __kernel_timespec kts = {tss->ts[0].tv_sec, tss->ts[0].tv_nsec};
1868
1869                                         put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
1870                                                  sizeof(kts), &kts);
1871                                 } else {
1872                                         struct timespec ts_old = timespec64_to_timespec(tss->ts[0]);
1873
1874                                         put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
1875                                                  sizeof(ts_old), &ts_old);
1876                                 }
1877                         } else {
1878                                 if (new_tstamp) {
1879                                         struct __kernel_sock_timeval stv;
1880
1881                                         stv.tv_sec = tss->ts[0].tv_sec;
1882                                         stv.tv_usec = tss->ts[0].tv_nsec / 1000;
1883                                         put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
1884                                                  sizeof(stv), &stv);
1885                                 } else {
1886                                         struct __kernel_old_timeval tv;
1887
1888                                         tv.tv_sec = tss->ts[0].tv_sec;
1889                                         tv.tv_usec = tss->ts[0].tv_nsec / 1000;
1890                                         put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
1891                                                  sizeof(tv), &tv);
1892                                 }
1893                         }
1894                 }
1895
1896                 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
1897                         has_timestamping = true;
1898                 else
1899                         tss->ts[0] = (struct timespec64) {0};
1900         }
1901
1902         if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
1903                 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
1904                         has_timestamping = true;
1905                 else
1906                         tss->ts[2] = (struct timespec64) {0};
1907         }
1908
1909         if (has_timestamping) {
1910                 tss->ts[1] = (struct timespec64) {0};
1911                 if (sock_flag(sk, SOCK_TSTAMP_NEW))
1912                         put_cmsg_scm_timestamping64(msg, tss);
1913                 else
1914                         put_cmsg_scm_timestamping(msg, tss);
1915         }
1916 }
1917
1918 static int tcp_inq_hint(struct sock *sk)
1919 {
1920         const struct tcp_sock *tp = tcp_sk(sk);
1921         u32 copied_seq = READ_ONCE(tp->copied_seq);
1922         u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
1923         int inq;
1924
1925         inq = rcv_nxt - copied_seq;
1926         if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
1927                 lock_sock(sk);
1928                 inq = tp->rcv_nxt - tp->copied_seq;
1929                 release_sock(sk);
1930         }
1931         /* After receiving a FIN, tell the user-space to continue reading
1932          * by returning a non-zero inq.
1933          */
1934         if (inq == 0 && sock_flag(sk, SOCK_DONE))
1935                 inq = 1;
1936         return inq;
1937 }
1938
1939 /*
1940  *      This routine copies from a sock struct into the user buffer.
1941  *
1942  *      Technical note: in 2.3 we work on _locked_ socket, so that
1943  *      tricks with *seq access order and skb->users are not required.
1944  *      Probably, code can be easily improved even more.
1945  */
1946
1947 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1948                 int flags, int *addr_len)
1949 {
1950         struct tcp_sock *tp = tcp_sk(sk);
1951         int copied = 0;
1952         u32 peek_seq;
1953         u32 *seq;
1954         unsigned long used;
1955         int err, inq;
1956         int target;             /* Read at least this many bytes */
1957         long timeo;
1958         struct sk_buff *skb, *last;
1959         u32 urg_hole = 0;
1960         struct scm_timestamping_internal tss;
1961         bool has_tss = false;
1962         bool has_cmsg;
1963
1964         if (unlikely(flags & MSG_ERRQUEUE))
1965                 return inet_recv_error(sk, msg, len, addr_len);
1966
1967         if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1968             (sk->sk_state == TCP_ESTABLISHED))
1969                 sk_busy_loop(sk, nonblock);
1970
1971         lock_sock(sk);
1972
1973         err = -ENOTCONN;
1974         if (sk->sk_state == TCP_LISTEN)
1975                 goto out;
1976
1977         has_cmsg = tp->recvmsg_inq;
1978         timeo = sock_rcvtimeo(sk, nonblock);
1979
1980         /* Urgent data needs to be handled specially. */
1981         if (flags & MSG_OOB)
1982                 goto recv_urg;
1983
1984         if (unlikely(tp->repair)) {
1985                 err = -EPERM;
1986                 if (!(flags & MSG_PEEK))
1987                         goto out;
1988
1989                 if (tp->repair_queue == TCP_SEND_QUEUE)
1990                         goto recv_sndq;
1991
1992                 err = -EINVAL;
1993                 if (tp->repair_queue == TCP_NO_QUEUE)
1994                         goto out;
1995
1996                 /* 'common' recv queue MSG_PEEK-ing */
1997         }
1998
1999         seq = &tp->copied_seq;
2000         if (flags & MSG_PEEK) {
2001                 peek_seq = tp->copied_seq;
2002                 seq = &peek_seq;
2003         }
2004
2005         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2006
2007         do {
2008                 u32 offset;
2009
2010                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2011                 if (tp->urg_data && tp->urg_seq == *seq) {
2012                         if (copied)
2013                                 break;
2014                         if (signal_pending(current)) {
2015                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2016                                 break;
2017                         }
2018                 }
2019
2020                 /* Next get a buffer. */
2021
2022                 last = skb_peek_tail(&sk->sk_receive_queue);
2023                 skb_queue_walk(&sk->sk_receive_queue, skb) {
2024                         last = skb;
2025                         /* Now that we have two receive queues this
2026                          * shouldn't happen.
2027                          */
2028                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2029                                  "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2030                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2031                                  flags))
2032                                 break;
2033
2034                         offset = *seq - TCP_SKB_CB(skb)->seq;
2035                         if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2036                                 pr_err_once("%s: found a SYN, please report !\n", __func__);
2037                                 offset--;
2038                         }
2039                         if (offset < skb->len)
2040                                 goto found_ok_skb;
2041                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2042                                 goto found_fin_ok;
2043                         WARN(!(flags & MSG_PEEK),
2044                              "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2045                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2046                 }
2047
2048                 /* Well, if we have backlog, try to process it now yet. */
2049
2050                 if (copied >= target && !sk->sk_backlog.tail)
2051                         break;
2052
2053                 if (copied) {
2054                         if (sk->sk_err ||
2055                             sk->sk_state == TCP_CLOSE ||
2056                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
2057                             !timeo ||
2058                             signal_pending(current))
2059                                 break;
2060                 } else {
2061                         if (sock_flag(sk, SOCK_DONE))
2062                                 break;
2063
2064                         if (sk->sk_err) {
2065                                 copied = sock_error(sk);
2066                                 break;
2067                         }
2068
2069                         if (sk->sk_shutdown & RCV_SHUTDOWN)
2070                                 break;
2071
2072                         if (sk->sk_state == TCP_CLOSE) {
2073                                 /* This occurs when user tries to read
2074                                  * from never connected socket.
2075                                  */
2076                                 copied = -ENOTCONN;
2077                                 break;
2078                         }
2079
2080                         if (!timeo) {
2081                                 copied = -EAGAIN;
2082                                 break;
2083                         }
2084
2085                         if (signal_pending(current)) {
2086                                 copied = sock_intr_errno(timeo);
2087                                 break;
2088                         }
2089                 }
2090
2091                 tcp_cleanup_rbuf(sk, copied);
2092
2093                 if (copied >= target) {
2094                         /* Do not sleep, just process backlog. */
2095                         release_sock(sk);
2096                         lock_sock(sk);
2097                 } else {
2098                         sk_wait_data(sk, &timeo, last);
2099                 }
2100
2101                 if ((flags & MSG_PEEK) &&
2102                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
2103                         net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2104                                             current->comm,
2105                                             task_pid_nr(current));
2106                         peek_seq = tp->copied_seq;
2107                 }
2108                 continue;
2109
2110 found_ok_skb:
2111                 /* Ok so how much can we use? */
2112                 used = skb->len - offset;
2113                 if (len < used)
2114                         used = len;
2115
2116                 /* Do we have urgent data here? */
2117                 if (tp->urg_data) {
2118                         u32 urg_offset = tp->urg_seq - *seq;
2119                         if (urg_offset < used) {
2120                                 if (!urg_offset) {
2121                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
2122                                                 ++*seq;
2123                                                 urg_hole++;
2124                                                 offset++;
2125                                                 used--;
2126                                                 if (!used)
2127                                                         goto skip_copy;
2128                                         }
2129                                 } else
2130                                         used = urg_offset;
2131                         }
2132                 }
2133
2134                 if (!(flags & MSG_TRUNC)) {
2135                         err = skb_copy_datagram_msg(skb, offset, msg, used);
2136                         if (err) {
2137                                 /* Exception. Bailout! */
2138                                 if (!copied)
2139                                         copied = -EFAULT;
2140                                 break;
2141                         }
2142                 }
2143
2144                 *seq += used;
2145                 copied += used;
2146                 len -= used;
2147
2148                 tcp_rcv_space_adjust(sk);
2149
2150 skip_copy:
2151                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
2152                         tp->urg_data = 0;
2153                         tcp_fast_path_check(sk);
2154                 }
2155                 if (used + offset < skb->len)
2156                         continue;
2157
2158                 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2159                         tcp_update_recv_tstamps(skb, &tss);
2160                         has_tss = true;
2161                         has_cmsg = true;
2162                 }
2163                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2164                         goto found_fin_ok;
2165                 if (!(flags & MSG_PEEK))
2166                         sk_eat_skb(sk, skb);
2167                 continue;
2168
2169 found_fin_ok:
2170                 /* Process the FIN. */
2171                 ++*seq;
2172                 if (!(flags & MSG_PEEK))
2173                         sk_eat_skb(sk, skb);
2174                 break;
2175         } while (len > 0);
2176
2177         /* According to UNIX98, msg_name/msg_namelen are ignored
2178          * on connected socket. I was just happy when found this 8) --ANK
2179          */
2180
2181         /* Clean up data we have read: This will do ACK frames. */
2182         tcp_cleanup_rbuf(sk, copied);
2183
2184         release_sock(sk);
2185
2186         if (has_cmsg) {
2187                 if (has_tss)
2188                         tcp_recv_timestamp(msg, sk, &tss);
2189                 if (tp->recvmsg_inq) {
2190                         inq = tcp_inq_hint(sk);
2191                         put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2192                 }
2193         }
2194
2195         return copied;
2196
2197 out:
2198         release_sock(sk);
2199         return err;
2200
2201 recv_urg:
2202         err = tcp_recv_urg(sk, msg, len, flags);
2203         goto out;
2204
2205 recv_sndq:
2206         err = tcp_peek_sndq(sk, msg, len);
2207         goto out;
2208 }
2209 EXPORT_SYMBOL(tcp_recvmsg);
2210
2211 void tcp_set_state(struct sock *sk, int state)
2212 {
2213         int oldstate = sk->sk_state;
2214
2215         /* We defined a new enum for TCP states that are exported in BPF
2216          * so as not force the internal TCP states to be frozen. The
2217          * following checks will detect if an internal state value ever
2218          * differs from the BPF value. If this ever happens, then we will
2219          * need to remap the internal value to the BPF value before calling
2220          * tcp_call_bpf_2arg.
2221          */
2222         BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2223         BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2224         BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2225         BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2226         BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2227         BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2228         BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2229         BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2230         BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2231         BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2232         BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2233         BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2234         BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2235
2236         if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2237                 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2238
2239         switch (state) {
2240         case TCP_ESTABLISHED:
2241                 if (oldstate != TCP_ESTABLISHED)
2242                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2243                 break;
2244
2245         case TCP_CLOSE:
2246                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2247                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2248
2249                 sk->sk_prot->unhash(sk);
2250                 if (inet_csk(sk)->icsk_bind_hash &&
2251                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2252                         inet_put_port(sk);
2253                 /* fall through */
2254         default:
2255                 if (oldstate == TCP_ESTABLISHED)
2256                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2257         }
2258
2259         /* Change state AFTER socket is unhashed to avoid closed
2260          * socket sitting in hash tables.
2261          */
2262         inet_sk_state_store(sk, state);
2263 }
2264 EXPORT_SYMBOL_GPL(tcp_set_state);
2265
2266 /*
2267  *      State processing on a close. This implements the state shift for
2268  *      sending our FIN frame. Note that we only send a FIN for some
2269  *      states. A shutdown() may have already sent the FIN, or we may be
2270  *      closed.
2271  */
2272
2273 static const unsigned char new_state[16] = {
2274   /* current state:        new state:      action:      */
2275   [0 /* (Invalid) */]   = TCP_CLOSE,
2276   [TCP_ESTABLISHED]     = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2277   [TCP_SYN_SENT]        = TCP_CLOSE,
2278   [TCP_SYN_RECV]        = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2279   [TCP_FIN_WAIT1]       = TCP_FIN_WAIT1,
2280   [TCP_FIN_WAIT2]       = TCP_FIN_WAIT2,
2281   [TCP_TIME_WAIT]       = TCP_CLOSE,
2282   [TCP_CLOSE]           = TCP_CLOSE,
2283   [TCP_CLOSE_WAIT]      = TCP_LAST_ACK  | TCP_ACTION_FIN,
2284   [TCP_LAST_ACK]        = TCP_LAST_ACK,
2285   [TCP_LISTEN]          = TCP_CLOSE,
2286   [TCP_CLOSING]         = TCP_CLOSING,
2287   [TCP_NEW_SYN_RECV]    = TCP_CLOSE,    /* should not happen ! */
2288 };
2289
2290 static int tcp_close_state(struct sock *sk)
2291 {
2292         int next = (int)new_state[sk->sk_state];
2293         int ns = next & TCP_STATE_MASK;
2294
2295         tcp_set_state(sk, ns);
2296
2297         return next & TCP_ACTION_FIN;
2298 }
2299
2300 /*
2301  *      Shutdown the sending side of a connection. Much like close except
2302  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2303  */
2304
2305 void tcp_shutdown(struct sock *sk, int how)
2306 {
2307         /*      We need to grab some memory, and put together a FIN,
2308          *      and then put it into the queue to be sent.
2309          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2310          */
2311         if (!(how & SEND_SHUTDOWN))
2312                 return;
2313
2314         /* If we've already sent a FIN, or it's a closed state, skip this. */
2315         if ((1 << sk->sk_state) &
2316             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2317              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2318                 /* Clear out any half completed packets.  FIN if needed. */
2319                 if (tcp_close_state(sk))
2320                         tcp_send_fin(sk);
2321         }
2322 }
2323 EXPORT_SYMBOL(tcp_shutdown);
2324
2325 bool tcp_check_oom(struct sock *sk, int shift)
2326 {
2327         bool too_many_orphans, out_of_socket_memory;
2328
2329         too_many_orphans = tcp_too_many_orphans(sk, shift);
2330         out_of_socket_memory = tcp_out_of_memory(sk);
2331
2332         if (too_many_orphans)
2333                 net_info_ratelimited("too many orphaned sockets\n");
2334         if (out_of_socket_memory)
2335                 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2336         return too_many_orphans || out_of_socket_memory;
2337 }
2338
2339 void tcp_close(struct sock *sk, long timeout)
2340 {
2341         struct sk_buff *skb;
2342         int data_was_unread = 0;
2343         int state;
2344
2345         lock_sock(sk);
2346         sk->sk_shutdown = SHUTDOWN_MASK;
2347
2348         if (sk->sk_state == TCP_LISTEN) {
2349                 tcp_set_state(sk, TCP_CLOSE);
2350
2351                 /* Special case. */
2352                 inet_csk_listen_stop(sk);
2353
2354                 goto adjudge_to_death;
2355         }
2356
2357         /*  We need to flush the recv. buffs.  We do this only on the
2358          *  descriptor close, not protocol-sourced closes, because the
2359          *  reader process may not have drained the data yet!
2360          */
2361         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2362                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2363
2364                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2365                         len--;
2366                 data_was_unread += len;
2367                 __kfree_skb(skb);
2368         }
2369
2370         sk_mem_reclaim(sk);
2371
2372         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2373         if (sk->sk_state == TCP_CLOSE)
2374                 goto adjudge_to_death;
2375
2376         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2377          * data was lost. To witness the awful effects of the old behavior of
2378          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2379          * GET in an FTP client, suspend the process, wait for the client to
2380          * advertise a zero window, then kill -9 the FTP client, wheee...
2381          * Note: timeout is always zero in such a case.
2382          */
2383         if (unlikely(tcp_sk(sk)->repair)) {
2384                 sk->sk_prot->disconnect(sk, 0);
2385         } else if (data_was_unread) {
2386                 /* Unread data was tossed, zap the connection. */
2387                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2388                 tcp_set_state(sk, TCP_CLOSE);
2389                 tcp_send_active_reset(sk, sk->sk_allocation);
2390         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2391                 /* Check zero linger _after_ checking for unread data. */
2392                 sk->sk_prot->disconnect(sk, 0);
2393                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2394         } else if (tcp_close_state(sk)) {
2395                 /* We FIN if the application ate all the data before
2396                  * zapping the connection.
2397                  */
2398
2399                 /* RED-PEN. Formally speaking, we have broken TCP state
2400                  * machine. State transitions:
2401                  *
2402                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2403                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2404                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2405                  *
2406                  * are legal only when FIN has been sent (i.e. in window),
2407                  * rather than queued out of window. Purists blame.
2408                  *
2409                  * F.e. "RFC state" is ESTABLISHED,
2410                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2411                  *
2412                  * The visible declinations are that sometimes
2413                  * we enter time-wait state, when it is not required really
2414                  * (harmless), do not send active resets, when they are
2415                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2416                  * they look as CLOSING or LAST_ACK for Linux)
2417                  * Probably, I missed some more holelets.
2418                  *                                              --ANK
2419                  * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2420                  * in a single packet! (May consider it later but will
2421                  * probably need API support or TCP_CORK SYN-ACK until
2422                  * data is written and socket is closed.)
2423                  */
2424                 tcp_send_fin(sk);
2425         }
2426
2427         sk_stream_wait_close(sk, timeout);
2428
2429 adjudge_to_death:
2430         state = sk->sk_state;
2431         sock_hold(sk);
2432         sock_orphan(sk);
2433
2434         local_bh_disable();
2435         bh_lock_sock(sk);
2436         /* remove backlog if any, without releasing ownership. */
2437         __release_sock(sk);
2438
2439         percpu_counter_inc(sk->sk_prot->orphan_count);
2440
2441         /* Have we already been destroyed by a softirq or backlog? */
2442         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2443                 goto out;
2444
2445         /*      This is a (useful) BSD violating of the RFC. There is a
2446          *      problem with TCP as specified in that the other end could
2447          *      keep a socket open forever with no application left this end.
2448          *      We use a 1 minute timeout (about the same as BSD) then kill
2449          *      our end. If they send after that then tough - BUT: long enough
2450          *      that we won't make the old 4*rto = almost no time - whoops
2451          *      reset mistake.
2452          *
2453          *      Nope, it was not mistake. It is really desired behaviour
2454          *      f.e. on http servers, when such sockets are useless, but
2455          *      consume significant resources. Let's do it with special
2456          *      linger2 option.                                 --ANK
2457          */
2458
2459         if (sk->sk_state == TCP_FIN_WAIT2) {
2460                 struct tcp_sock *tp = tcp_sk(sk);
2461                 if (tp->linger2 < 0) {
2462                         tcp_set_state(sk, TCP_CLOSE);
2463                         tcp_send_active_reset(sk, GFP_ATOMIC);
2464                         __NET_INC_STATS(sock_net(sk),
2465                                         LINUX_MIB_TCPABORTONLINGER);
2466                 } else {
2467                         const int tmo = tcp_fin_time(sk);
2468
2469                         if (tmo > TCP_TIMEWAIT_LEN) {
2470                                 inet_csk_reset_keepalive_timer(sk,
2471                                                 tmo - TCP_TIMEWAIT_LEN);
2472                         } else {
2473                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2474                                 goto out;
2475                         }
2476                 }
2477         }
2478         if (sk->sk_state != TCP_CLOSE) {
2479                 sk_mem_reclaim(sk);
2480                 if (tcp_check_oom(sk, 0)) {
2481                         tcp_set_state(sk, TCP_CLOSE);
2482                         tcp_send_active_reset(sk, GFP_ATOMIC);
2483                         __NET_INC_STATS(sock_net(sk),
2484                                         LINUX_MIB_TCPABORTONMEMORY);
2485                 } else if (!check_net(sock_net(sk))) {
2486                         /* Not possible to send reset; just close */
2487                         tcp_set_state(sk, TCP_CLOSE);
2488                 }
2489         }
2490
2491         if (sk->sk_state == TCP_CLOSE) {
2492                 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2493                 /* We could get here with a non-NULL req if the socket is
2494                  * aborted (e.g., closed with unread data) before 3WHS
2495                  * finishes.
2496                  */
2497                 if (req)
2498                         reqsk_fastopen_remove(sk, req, false);
2499                 inet_csk_destroy_sock(sk);
2500         }
2501         /* Otherwise, socket is reprieved until protocol close. */
2502
2503 out:
2504         bh_unlock_sock(sk);
2505         local_bh_enable();
2506         release_sock(sk);
2507         sock_put(sk);
2508 }
2509 EXPORT_SYMBOL(tcp_close);
2510
2511 /* These states need RST on ABORT according to RFC793 */
2512
2513 static inline bool tcp_need_reset(int state)
2514 {
2515         return (1 << state) &
2516                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2517                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2518 }
2519
2520 static void tcp_rtx_queue_purge(struct sock *sk)
2521 {
2522         struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2523
2524         while (p) {
2525                 struct sk_buff *skb = rb_to_skb(p);
2526
2527                 p = rb_next(p);
2528                 /* Since we are deleting whole queue, no need to
2529                  * list_del(&skb->tcp_tsorted_anchor)
2530                  */
2531                 tcp_rtx_queue_unlink(skb, sk);
2532                 sk_wmem_free_skb(sk, skb);
2533         }
2534 }
2535
2536 void tcp_write_queue_purge(struct sock *sk)
2537 {
2538         struct sk_buff *skb;
2539
2540         tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2541         while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2542                 tcp_skb_tsorted_anchor_cleanup(skb);
2543                 sk_wmem_free_skb(sk, skb);
2544         }
2545         tcp_rtx_queue_purge(sk);
2546         skb = sk->sk_tx_skb_cache;
2547         if (skb) {
2548                 __kfree_skb(skb);
2549                 sk->sk_tx_skb_cache = NULL;
2550         }
2551         INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2552         sk_mem_reclaim(sk);
2553         tcp_clear_all_retrans_hints(tcp_sk(sk));
2554         tcp_sk(sk)->packets_out = 0;
2555         inet_csk(sk)->icsk_backoff = 0;
2556 }
2557
2558 int tcp_disconnect(struct sock *sk, int flags)
2559 {
2560         struct inet_sock *inet = inet_sk(sk);
2561         struct inet_connection_sock *icsk = inet_csk(sk);
2562         struct tcp_sock *tp = tcp_sk(sk);
2563         int old_state = sk->sk_state;
2564
2565         if (old_state != TCP_CLOSE)
2566                 tcp_set_state(sk, TCP_CLOSE);
2567
2568         /* ABORT function of RFC793 */
2569         if (old_state == TCP_LISTEN) {
2570                 inet_csk_listen_stop(sk);
2571         } else if (unlikely(tp->repair)) {
2572                 sk->sk_err = ECONNABORTED;
2573         } else if (tcp_need_reset(old_state) ||
2574                    (tp->snd_nxt != tp->write_seq &&
2575                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2576                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2577                  * states
2578                  */
2579                 tcp_send_active_reset(sk, gfp_any());
2580                 sk->sk_err = ECONNRESET;
2581         } else if (old_state == TCP_SYN_SENT)
2582                 sk->sk_err = ECONNRESET;
2583
2584         tcp_clear_xmit_timers(sk);
2585         __skb_queue_purge(&sk->sk_receive_queue);
2586         if (sk->sk_rx_skb_cache) {
2587                 __kfree_skb(sk->sk_rx_skb_cache);
2588                 sk->sk_rx_skb_cache = NULL;
2589         }
2590         tp->copied_seq = tp->rcv_nxt;
2591         tp->urg_data = 0;
2592         tcp_write_queue_purge(sk);
2593         tcp_fastopen_active_disable_ofo_check(sk);
2594         skb_rbtree_purge(&tp->out_of_order_queue);
2595
2596         inet->inet_dport = 0;
2597
2598         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2599                 inet_reset_saddr(sk);
2600
2601         sk->sk_shutdown = 0;
2602         sock_reset_flag(sk, SOCK_DONE);
2603         tp->srtt_us = 0;
2604         tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
2605         tp->rcv_rtt_last_tsecr = 0;
2606         tp->write_seq += tp->max_window + 2;
2607         if (tp->write_seq == 0)
2608                 tp->write_seq = 1;
2609         icsk->icsk_backoff = 0;
2610         tp->snd_cwnd = 2;
2611         icsk->icsk_probes_out = 0;
2612         icsk->icsk_rto = TCP_TIMEOUT_INIT;
2613         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2614         tp->snd_cwnd = TCP_INIT_CWND;
2615         tp->snd_cwnd_cnt = 0;
2616         tp->window_clamp = 0;
2617         tp->delivered_ce = 0;
2618         tcp_set_ca_state(sk, TCP_CA_Open);
2619         tp->is_sack_reneg = 0;
2620         tcp_clear_retrans(tp);
2621         inet_csk_delack_init(sk);
2622         /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2623          * issue in __tcp_select_window()
2624          */
2625         icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2626         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2627         __sk_dst_reset(sk);
2628         dst_release(sk->sk_rx_dst);
2629         sk->sk_rx_dst = NULL;
2630         tcp_saved_syn_free(tp);
2631         tp->compressed_ack = 0;
2632         tp->bytes_sent = 0;
2633         tp->bytes_acked = 0;
2634         tp->bytes_received = 0;
2635         tp->bytes_retrans = 0;
2636         tp->duplicate_sack[0].start_seq = 0;
2637         tp->duplicate_sack[0].end_seq = 0;
2638         tp->dsack_dups = 0;
2639         tp->reord_seen = 0;
2640         tp->retrans_out = 0;
2641         tp->sacked_out = 0;
2642         tp->tlp_high_seq = 0;
2643         tp->last_oow_ack_time = 0;
2644         /* There's a bubble in the pipe until at least the first ACK. */
2645         tp->app_limited = ~0U;
2646         tp->rack.mstamp = 0;
2647         tp->rack.advanced = 0;
2648         tp->rack.reo_wnd_steps = 1;
2649         tp->rack.last_delivered = 0;
2650         tp->rack.reo_wnd_persist = 0;
2651         tp->rack.dsack_seen = 0;
2652         tp->syn_data_acked = 0;
2653         tp->rx_opt.saw_tstamp = 0;
2654         tp->rx_opt.dsack = 0;
2655         tp->rx_opt.num_sacks = 0;
2656         tp->rcv_ooopack = 0;
2657
2658
2659         /* Clean up fastopen related fields */
2660         tcp_free_fastopen_req(tp);
2661         inet->defer_connect = 0;
2662
2663         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2664
2665         if (sk->sk_frag.page) {
2666                 put_page(sk->sk_frag.page);
2667                 sk->sk_frag.page = NULL;
2668                 sk->sk_frag.offset = 0;
2669         }
2670
2671         sk->sk_error_report(sk);
2672         return 0;
2673 }
2674 EXPORT_SYMBOL(tcp_disconnect);
2675
2676 static inline bool tcp_can_repair_sock(const struct sock *sk)
2677 {
2678         return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2679                 (sk->sk_state != TCP_LISTEN);
2680 }
2681
2682 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2683 {
2684         struct tcp_repair_window opt;
2685
2686         if (!tp->repair)
2687                 return -EPERM;
2688
2689         if (len != sizeof(opt))
2690                 return -EINVAL;
2691
2692         if (copy_from_user(&opt, optbuf, sizeof(opt)))
2693                 return -EFAULT;
2694
2695         if (opt.max_window < opt.snd_wnd)
2696                 return -EINVAL;
2697
2698         if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2699                 return -EINVAL;
2700
2701         if (after(opt.rcv_wup, tp->rcv_nxt))
2702                 return -EINVAL;
2703
2704         tp->snd_wl1     = opt.snd_wl1;
2705         tp->snd_wnd     = opt.snd_wnd;
2706         tp->max_window  = opt.max_window;
2707
2708         tp->rcv_wnd     = opt.rcv_wnd;
2709         tp->rcv_wup     = opt.rcv_wup;
2710
2711         return 0;
2712 }
2713
2714 static int tcp_repair_options_est(struct sock *sk,
2715                 struct tcp_repair_opt __user *optbuf, unsigned int len)
2716 {
2717         struct tcp_sock *tp = tcp_sk(sk);
2718         struct tcp_repair_opt opt;
2719
2720         while (len >= sizeof(opt)) {
2721                 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2722                         return -EFAULT;
2723
2724                 optbuf++;
2725                 len -= sizeof(opt);
2726
2727                 switch (opt.opt_code) {
2728                 case TCPOPT_MSS:
2729                         tp->rx_opt.mss_clamp = opt.opt_val;
2730                         tcp_mtup_init(sk);
2731                         break;
2732                 case TCPOPT_WINDOW:
2733                         {
2734                                 u16 snd_wscale = opt.opt_val & 0xFFFF;
2735                                 u16 rcv_wscale = opt.opt_val >> 16;
2736
2737                                 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
2738                                         return -EFBIG;
2739
2740                                 tp->rx_opt.snd_wscale = snd_wscale;
2741                                 tp->rx_opt.rcv_wscale = rcv_wscale;
2742                                 tp->rx_opt.wscale_ok = 1;
2743                         }
2744                         break;
2745                 case TCPOPT_SACK_PERM:
2746                         if (opt.opt_val != 0)
2747                                 return -EINVAL;
2748
2749                         tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2750                         break;
2751                 case TCPOPT_TIMESTAMP:
2752                         if (opt.opt_val != 0)
2753                                 return -EINVAL;
2754
2755                         tp->rx_opt.tstamp_ok = 1;
2756                         break;
2757                 }
2758         }
2759
2760         return 0;
2761 }
2762
2763 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
2764 EXPORT_SYMBOL(tcp_tx_delay_enabled);
2765
2766 static void tcp_enable_tx_delay(void)
2767 {
2768         if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
2769                 static int __tcp_tx_delay_enabled = 0;
2770
2771                 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
2772                         static_branch_enable(&tcp_tx_delay_enabled);
2773                         pr_info("TCP_TX_DELAY enabled\n");
2774                 }
2775         }
2776 }
2777
2778 /*
2779  *      Socket option code for TCP.
2780  */
2781 static int do_tcp_setsockopt(struct sock *sk, int level,
2782                 int optname, char __user *optval, unsigned int optlen)
2783 {
2784         struct tcp_sock *tp = tcp_sk(sk);
2785         struct inet_connection_sock *icsk = inet_csk(sk);
2786         struct net *net = sock_net(sk);
2787         int val;
2788         int err = 0;
2789
2790         /* These are data/string values, all the others are ints */
2791         switch (optname) {
2792         case TCP_CONGESTION: {
2793                 char name[TCP_CA_NAME_MAX];
2794
2795                 if (optlen < 1)
2796                         return -EINVAL;
2797
2798                 val = strncpy_from_user(name, optval,
2799                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2800                 if (val < 0)
2801                         return -EFAULT;
2802                 name[val] = 0;
2803
2804                 lock_sock(sk);
2805                 err = tcp_set_congestion_control(sk, name, true, true,
2806                                                  ns_capable(sock_net(sk)->user_ns,
2807                                                             CAP_NET_ADMIN));
2808                 release_sock(sk);
2809                 return err;
2810         }
2811         case TCP_ULP: {
2812                 char name[TCP_ULP_NAME_MAX];
2813
2814                 if (optlen < 1)
2815                         return -EINVAL;
2816
2817                 val = strncpy_from_user(name, optval,
2818                                         min_t(long, TCP_ULP_NAME_MAX - 1,
2819                                               optlen));
2820                 if (val < 0)
2821                         return -EFAULT;
2822                 name[val] = 0;
2823
2824                 lock_sock(sk);
2825                 err = tcp_set_ulp(sk, name);
2826                 release_sock(sk);
2827                 return err;
2828         }
2829         case TCP_FASTOPEN_KEY: {
2830                 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
2831                 __u8 *backup_key = NULL;
2832
2833                 /* Allow a backup key as well to facilitate key rotation
2834                  * First key is the active one.
2835                  */
2836                 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
2837                     optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
2838                         return -EINVAL;
2839
2840                 if (copy_from_user(key, optval, optlen))
2841                         return -EFAULT;
2842
2843                 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
2844                         backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
2845
2846                 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
2847         }
2848         default:
2849                 /* fallthru */
2850                 break;
2851         }
2852
2853         if (optlen < sizeof(int))
2854                 return -EINVAL;
2855
2856         if (get_user(val, (int __user *)optval))
2857                 return -EFAULT;
2858
2859         lock_sock(sk);
2860
2861         switch (optname) {
2862         case TCP_MAXSEG:
2863                 /* Values greater than interface MTU won't take effect. However
2864                  * at the point when this call is done we typically don't yet
2865                  * know which interface is going to be used
2866                  */
2867                 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
2868                         err = -EINVAL;
2869                         break;
2870                 }
2871                 tp->rx_opt.user_mss = val;
2872                 break;
2873
2874         case TCP_NODELAY:
2875                 if (val) {
2876                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2877                          * this option on corked socket is remembered, but
2878                          * it is not activated until cork is cleared.
2879                          *
2880                          * However, when TCP_NODELAY is set we make
2881                          * an explicit push, which overrides even TCP_CORK
2882                          * for currently queued segments.
2883                          */
2884                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2885                         tcp_push_pending_frames(sk);
2886                 } else {
2887                         tp->nonagle &= ~TCP_NAGLE_OFF;
2888                 }
2889                 break;
2890
2891         case TCP_THIN_LINEAR_TIMEOUTS:
2892                 if (val < 0 || val > 1)
2893                         err = -EINVAL;
2894                 else
2895                         tp->thin_lto = val;
2896                 break;
2897
2898         case TCP_THIN_DUPACK:
2899                 if (val < 0 || val > 1)
2900                         err = -EINVAL;
2901                 break;
2902
2903         case TCP_REPAIR:
2904                 if (!tcp_can_repair_sock(sk))
2905                         err = -EPERM;
2906                 else if (val == TCP_REPAIR_ON) {
2907                         tp->repair = 1;
2908                         sk->sk_reuse = SK_FORCE_REUSE;
2909                         tp->repair_queue = TCP_NO_QUEUE;
2910                 } else if (val == TCP_REPAIR_OFF) {
2911                         tp->repair = 0;
2912                         sk->sk_reuse = SK_NO_REUSE;
2913                         tcp_send_window_probe(sk);
2914                 } else if (val == TCP_REPAIR_OFF_NO_WP) {
2915                         tp->repair = 0;
2916                         sk->sk_reuse = SK_NO_REUSE;
2917                 } else
2918                         err = -EINVAL;
2919
2920                 break;
2921
2922         case TCP_REPAIR_QUEUE:
2923                 if (!tp->repair)
2924                         err = -EPERM;
2925                 else if ((unsigned int)val < TCP_QUEUES_NR)
2926                         tp->repair_queue = val;
2927                 else
2928                         err = -EINVAL;
2929                 break;
2930
2931         case TCP_QUEUE_SEQ:
2932                 if (sk->sk_state != TCP_CLOSE)
2933                         err = -EPERM;
2934                 else if (tp->repair_queue == TCP_SEND_QUEUE)
2935                         tp->write_seq = val;
2936                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2937                         tp->rcv_nxt = val;
2938                 else
2939                         err = -EINVAL;
2940                 break;
2941
2942         case TCP_REPAIR_OPTIONS:
2943                 if (!tp->repair)
2944                         err = -EINVAL;
2945                 else if (sk->sk_state == TCP_ESTABLISHED)
2946                         err = tcp_repair_options_est(sk,
2947                                         (struct tcp_repair_opt __user *)optval,
2948                                         optlen);
2949                 else
2950                         err = -EPERM;
2951                 break;
2952
2953         case TCP_CORK:
2954                 /* When set indicates to always queue non-full frames.
2955                  * Later the user clears this option and we transmit
2956                  * any pending partial frames in the queue.  This is
2957                  * meant to be used alongside sendfile() to get properly
2958                  * filled frames when the user (for example) must write
2959                  * out headers with a write() call first and then use
2960                  * sendfile to send out the data parts.
2961                  *
2962                  * TCP_CORK can be set together with TCP_NODELAY and it is
2963                  * stronger than TCP_NODELAY.
2964                  */
2965                 if (val) {
2966                         tp->nonagle |= TCP_NAGLE_CORK;
2967                 } else {
2968                         tp->nonagle &= ~TCP_NAGLE_CORK;
2969                         if (tp->nonagle&TCP_NAGLE_OFF)
2970                                 tp->nonagle |= TCP_NAGLE_PUSH;
2971                         tcp_push_pending_frames(sk);
2972                 }
2973                 break;
2974
2975         case TCP_KEEPIDLE:
2976                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2977                         err = -EINVAL;
2978                 else {
2979                         tp->keepalive_time = val * HZ;
2980                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2981                             !((1 << sk->sk_state) &
2982                               (TCPF_CLOSE | TCPF_LISTEN))) {
2983                                 u32 elapsed = keepalive_time_elapsed(tp);
2984                                 if (tp->keepalive_time > elapsed)
2985                                         elapsed = tp->keepalive_time - elapsed;
2986                                 else
2987                                         elapsed = 0;
2988                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2989                         }
2990                 }
2991                 break;
2992         case TCP_KEEPINTVL:
2993                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2994                         err = -EINVAL;
2995                 else
2996                         tp->keepalive_intvl = val * HZ;
2997                 break;
2998         case TCP_KEEPCNT:
2999                 if (val < 1 || val > MAX_TCP_KEEPCNT)
3000                         err = -EINVAL;
3001                 else
3002                         tp->keepalive_probes = val;
3003                 break;
3004         case TCP_SYNCNT:
3005                 if (val < 1 || val > MAX_TCP_SYNCNT)
3006                         err = -EINVAL;
3007                 else
3008                         icsk->icsk_syn_retries = val;
3009                 break;
3010
3011         case TCP_SAVE_SYN:
3012                 if (val < 0 || val > 1)
3013                         err = -EINVAL;
3014                 else
3015                         tp->save_syn = val;
3016                 break;
3017
3018         case TCP_LINGER2:
3019                 if (val < 0)
3020                         tp->linger2 = -1;
3021                 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
3022                         tp->linger2 = 0;
3023                 else
3024                         tp->linger2 = val * HZ;
3025                 break;
3026
3027         case TCP_DEFER_ACCEPT:
3028                 /* Translate value in seconds to number of retransmits */
3029                 icsk->icsk_accept_queue.rskq_defer_accept =
3030                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3031                                         TCP_RTO_MAX / HZ);
3032                 break;
3033
3034         case TCP_WINDOW_CLAMP:
3035                 if (!val) {
3036                         if (sk->sk_state != TCP_CLOSE) {
3037                                 err = -EINVAL;
3038                                 break;
3039                         }
3040                         tp->window_clamp = 0;
3041                 } else
3042                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3043                                                 SOCK_MIN_RCVBUF / 2 : val;
3044                 break;
3045
3046         case TCP_QUICKACK:
3047                 if (!val) {
3048                         inet_csk_enter_pingpong_mode(sk);
3049                 } else {
3050                         inet_csk_exit_pingpong_mode(sk);
3051                         if ((1 << sk->sk_state) &
3052                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3053                             inet_csk_ack_scheduled(sk)) {
3054                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
3055                                 tcp_cleanup_rbuf(sk, 1);
3056                                 if (!(val & 1))
3057                                         inet_csk_enter_pingpong_mode(sk);
3058                         }
3059                 }
3060                 break;
3061
3062 #ifdef CONFIG_TCP_MD5SIG
3063         case TCP_MD5SIG:
3064         case TCP_MD5SIG_EXT:
3065                 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
3066                         err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3067                 else
3068                         err = -EINVAL;
3069                 break;
3070 #endif
3071         case TCP_USER_TIMEOUT:
3072                 /* Cap the max time in ms TCP will retry or probe the window
3073                  * before giving up and aborting (ETIMEDOUT) a connection.
3074                  */
3075                 if (val < 0)
3076                         err = -EINVAL;
3077                 else
3078                         icsk->icsk_user_timeout = val;
3079                 break;
3080
3081         case TCP_FASTOPEN:
3082                 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3083                     TCPF_LISTEN))) {
3084                         tcp_fastopen_init_key_once(net);
3085
3086                         fastopen_queue_tune(sk, val);
3087                 } else {
3088                         err = -EINVAL;
3089                 }
3090                 break;
3091         case TCP_FASTOPEN_CONNECT:
3092                 if (val > 1 || val < 0) {
3093                         err = -EINVAL;
3094                 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
3095                         if (sk->sk_state == TCP_CLOSE)
3096                                 tp->fastopen_connect = val;
3097                         else
3098                                 err = -EINVAL;
3099                 } else {
3100                         err = -EOPNOTSUPP;
3101                 }
3102                 break;
3103         case TCP_FASTOPEN_NO_COOKIE:
3104                 if (val > 1 || val < 0)
3105                         err = -EINVAL;
3106                 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3107                         err = -EINVAL;
3108                 else
3109                         tp->fastopen_no_cookie = val;
3110                 break;
3111         case TCP_TIMESTAMP:
3112                 if (!tp->repair)
3113                         err = -EPERM;
3114                 else
3115                         tp->tsoffset = val - tcp_time_stamp_raw();
3116                 break;
3117         case TCP_REPAIR_WINDOW:
3118                 err = tcp_repair_set_window(tp, optval, optlen);
3119                 break;
3120         case TCP_NOTSENT_LOWAT:
3121                 tp->notsent_lowat = val;
3122                 sk->sk_write_space(sk);
3123                 break;
3124         case TCP_INQ:
3125                 if (val > 1 || val < 0)
3126                         err = -EINVAL;
3127                 else
3128                         tp->recvmsg_inq = val;
3129                 break;
3130         case TCP_TX_DELAY:
3131                 if (val)
3132                         tcp_enable_tx_delay();
3133                 tp->tcp_tx_delay = val;
3134                 break;
3135         default:
3136                 err = -ENOPROTOOPT;
3137                 break;
3138         }
3139
3140         release_sock(sk);
3141         return err;
3142 }
3143
3144 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
3145                    unsigned int optlen)
3146 {
3147         const struct inet_connection_sock *icsk = inet_csk(sk);
3148
3149         if (level != SOL_TCP)
3150                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3151                                                      optval, optlen);
3152         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3153 }
3154 EXPORT_SYMBOL(tcp_setsockopt);
3155
3156 #ifdef CONFIG_COMPAT
3157 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
3158                           char __user *optval, unsigned int optlen)
3159 {
3160         if (level != SOL_TCP)
3161                 return inet_csk_compat_setsockopt(sk, level, optname,
3162                                                   optval, optlen);
3163         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3164 }
3165 EXPORT_SYMBOL(compat_tcp_setsockopt);
3166 #endif
3167
3168 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3169                                       struct tcp_info *info)
3170 {
3171         u64 stats[__TCP_CHRONO_MAX], total = 0;
3172         enum tcp_chrono i;
3173
3174         for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3175                 stats[i] = tp->chrono_stat[i - 1];
3176                 if (i == tp->chrono_type)
3177                         stats[i] += tcp_jiffies32 - tp->chrono_start;
3178                 stats[i] *= USEC_PER_SEC / HZ;
3179                 total += stats[i];
3180         }
3181
3182         info->tcpi_busy_time = total;
3183         info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3184         info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3185 }
3186
3187 /* Return information about state of tcp endpoint in API format. */
3188 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3189 {
3190         const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3191         const struct inet_connection_sock *icsk = inet_csk(sk);
3192         unsigned long rate;
3193         u32 now;
3194         u64 rate64;
3195         bool slow;
3196
3197         memset(info, 0, sizeof(*info));
3198         if (sk->sk_type != SOCK_STREAM)
3199                 return;
3200
3201         info->tcpi_state = inet_sk_state_load(sk);
3202
3203         /* Report meaningful fields for all TCP states, including listeners */
3204         rate = READ_ONCE(sk->sk_pacing_rate);
3205         rate64 = (rate != ~0UL) ? rate : ~0ULL;
3206         info->tcpi_pacing_rate = rate64;
3207
3208         rate = READ_ONCE(sk->sk_max_pacing_rate);
3209         rate64 = (rate != ~0UL) ? rate : ~0ULL;
3210         info->tcpi_max_pacing_rate = rate64;
3211
3212         info->tcpi_reordering = tp->reordering;
3213         info->tcpi_snd_cwnd = tp->snd_cwnd;
3214
3215         if (info->tcpi_state == TCP_LISTEN) {
3216                 /* listeners aliased fields :
3217                  * tcpi_unacked -> Number of children ready for accept()
3218                  * tcpi_sacked  -> max backlog
3219                  */
3220                 info->tcpi_unacked = sk->sk_ack_backlog;
3221                 info->tcpi_sacked = sk->sk_max_ack_backlog;
3222                 return;
3223         }
3224
3225         slow = lock_sock_fast(sk);
3226
3227         info->tcpi_ca_state = icsk->icsk_ca_state;
3228         info->tcpi_retransmits = icsk->icsk_retransmits;
3229         info->tcpi_probes = icsk->icsk_probes_out;
3230         info->tcpi_backoff = icsk->icsk_backoff;
3231
3232         if (tp->rx_opt.tstamp_ok)
3233                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3234         if (tcp_is_sack(tp))
3235                 info->tcpi_options |= TCPI_OPT_SACK;
3236         if (tp->rx_opt.wscale_ok) {
3237                 info->tcpi_options |= TCPI_OPT_WSCALE;
3238                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3239                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3240         }
3241
3242         if (tp->ecn_flags & TCP_ECN_OK)
3243                 info->tcpi_options |= TCPI_OPT_ECN;
3244         if (tp->ecn_flags & TCP_ECN_SEEN)
3245                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3246         if (tp->syn_data_acked)
3247                 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3248
3249         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3250         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3251         info->tcpi_snd_mss = tp->mss_cache;
3252         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3253
3254         info->tcpi_unacked = tp->packets_out;
3255         info->tcpi_sacked = tp->sacked_out;
3256
3257         info->tcpi_lost = tp->lost_out;
3258         info->tcpi_retrans = tp->retrans_out;
3259
3260         now = tcp_jiffies32;
3261         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3262         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3263         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3264
3265         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3266         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3267         info->tcpi_rtt = tp->srtt_us >> 3;
3268         info->tcpi_rttvar = tp->mdev_us >> 2;
3269         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3270         info->tcpi_advmss = tp->advmss;
3271
3272         info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3273         info->tcpi_rcv_space = tp->rcvq_space.space;
3274
3275         info->tcpi_total_retrans = tp->total_retrans;
3276
3277         info->tcpi_bytes_acked = tp->bytes_acked;
3278         info->tcpi_bytes_received = tp->bytes_received;
3279         info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3280         tcp_get_info_chrono_stats(tp, info);
3281
3282         info->tcpi_segs_out = tp->segs_out;
3283         info->tcpi_segs_in = tp->segs_in;
3284
3285         info->tcpi_min_rtt = tcp_min_rtt(tp);
3286         info->tcpi_data_segs_in = tp->data_segs_in;
3287         info->tcpi_data_segs_out = tp->data_segs_out;
3288
3289         info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3290         rate64 = tcp_compute_delivery_rate(tp);
3291         if (rate64)
3292                 info->tcpi_delivery_rate = rate64;
3293         info->tcpi_delivered = tp->delivered;
3294         info->tcpi_delivered_ce = tp->delivered_ce;
3295         info->tcpi_bytes_sent = tp->bytes_sent;
3296         info->tcpi_bytes_retrans = tp->bytes_retrans;
3297         info->tcpi_dsack_dups = tp->dsack_dups;
3298         info->tcpi_reord_seen = tp->reord_seen;
3299         info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3300         info->tcpi_snd_wnd = tp->snd_wnd;
3301         unlock_sock_fast(sk, slow);
3302 }
3303 EXPORT_SYMBOL_GPL(tcp_get_info);
3304
3305 static size_t tcp_opt_stats_get_size(void)
3306 {
3307         return
3308                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3309                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3310                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3311                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3312                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3313                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3314                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3315                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3316                 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3317                 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3318                 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3319                 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3320                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3321                 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3322                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3323                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3324                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3325                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3326                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3327                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3328                 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3329                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3330                 0;
3331 }
3332
3333 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk)
3334 {
3335         const struct tcp_sock *tp = tcp_sk(sk);
3336         struct sk_buff *stats;
3337         struct tcp_info info;
3338         unsigned long rate;
3339         u64 rate64;
3340
3341         stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3342         if (!stats)
3343                 return NULL;
3344
3345         tcp_get_info_chrono_stats(tp, &info);
3346         nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3347                           info.tcpi_busy_time, TCP_NLA_PAD);
3348         nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3349                           info.tcpi_rwnd_limited, TCP_NLA_PAD);
3350         nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3351                           info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3352         nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3353                           tp->data_segs_out, TCP_NLA_PAD);
3354         nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3355                           tp->total_retrans, TCP_NLA_PAD);
3356
3357         rate = READ_ONCE(sk->sk_pacing_rate);
3358         rate64 = (rate != ~0UL) ? rate : ~0ULL;
3359         nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3360
3361         rate64 = tcp_compute_delivery_rate(tp);
3362         nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3363
3364         nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3365         nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3366         nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3367
3368         nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3369         nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3370         nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3371         nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3372         nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3373
3374         nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3375         nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3376
3377         nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3378                           TCP_NLA_PAD);
3379         nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3380                           TCP_NLA_PAD);
3381         nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3382         nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3383         nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3384
3385         return stats;
3386 }
3387
3388 static int do_tcp_getsockopt(struct sock *sk, int level,
3389                 int optname, char __user *optval, int __user *optlen)
3390 {
3391         struct inet_connection_sock *icsk = inet_csk(sk);
3392         struct tcp_sock *tp = tcp_sk(sk);
3393         struct net *net = sock_net(sk);
3394         int val, len;
3395
3396         if (get_user(len, optlen))
3397                 return -EFAULT;
3398
3399         len = min_t(unsigned int, len, sizeof(int));
3400
3401         if (len < 0)
3402                 return -EINVAL;
3403
3404         switch (optname) {
3405         case TCP_MAXSEG:
3406                 val = tp->mss_cache;
3407                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3408                         val = tp->rx_opt.user_mss;
3409                 if (tp->repair)
3410                         val = tp->rx_opt.mss_clamp;
3411                 break;
3412         case TCP_NODELAY:
3413                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3414                 break;
3415         case TCP_CORK:
3416                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3417                 break;
3418         case TCP_KEEPIDLE:
3419                 val = keepalive_time_when(tp) / HZ;
3420                 break;
3421         case TCP_KEEPINTVL:
3422                 val = keepalive_intvl_when(tp) / HZ;
3423                 break;
3424         case TCP_KEEPCNT:
3425                 val = keepalive_probes(tp);
3426                 break;
3427         case TCP_SYNCNT:
3428                 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3429                 break;
3430         case TCP_LINGER2:
3431                 val = tp->linger2;
3432                 if (val >= 0)
3433                         val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3434                 break;
3435         case TCP_DEFER_ACCEPT:
3436                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3437                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3438                 break;
3439         case TCP_WINDOW_CLAMP:
3440                 val = tp->window_clamp;
3441                 break;
3442         case TCP_INFO: {
3443                 struct tcp_info info;
3444
3445                 if (get_user(len, optlen))
3446                         return -EFAULT;
3447
3448                 tcp_get_info(sk, &info);
3449
3450                 len = min_t(unsigned int, len, sizeof(info));
3451                 if (put_user(len, optlen))
3452                         return -EFAULT;
3453                 if (copy_to_user(optval, &info, len))
3454                         return -EFAULT;
3455                 return 0;
3456         }
3457         case TCP_CC_INFO: {
3458                 const struct tcp_congestion_ops *ca_ops;
3459                 union tcp_cc_info info;
3460                 size_t sz = 0;
3461                 int attr;
3462
3463                 if (get_user(len, optlen))
3464                         return -EFAULT;
3465
3466                 ca_ops = icsk->icsk_ca_ops;
3467                 if (ca_ops && ca_ops->get_info)
3468                         sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3469
3470                 len = min_t(unsigned int, len, sz);
3471                 if (put_user(len, optlen))
3472                         return -EFAULT;
3473                 if (copy_to_user(optval, &info, len))
3474                         return -EFAULT;
3475                 return 0;
3476         }
3477         case TCP_QUICKACK:
3478                 val = !inet_csk_in_pingpong_mode(sk);
3479                 break;
3480
3481         case TCP_CONGESTION:
3482                 if (get_user(len, optlen))
3483                         return -EFAULT;
3484                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
3485                 if (put_user(len, optlen))
3486                         return -EFAULT;
3487                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3488                         return -EFAULT;
3489                 return 0;
3490
3491         case TCP_ULP:
3492                 if (get_user(len, optlen))
3493                         return -EFAULT;
3494                 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
3495                 if (!icsk->icsk_ulp_ops) {
3496                         if (put_user(0, optlen))
3497                                 return -EFAULT;
3498                         return 0;
3499                 }
3500                 if (put_user(len, optlen))
3501                         return -EFAULT;
3502                 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
3503                         return -EFAULT;
3504                 return 0;
3505
3506         case TCP_FASTOPEN_KEY: {
3507                 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3508                 struct tcp_fastopen_context *ctx;
3509                 unsigned int key_len = 0;
3510
3511                 if (get_user(len, optlen))
3512                         return -EFAULT;
3513
3514                 rcu_read_lock();
3515                 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
3516                 if (ctx) {
3517                         key_len = tcp_fastopen_context_len(ctx) *
3518                                         TCP_FASTOPEN_KEY_LENGTH;
3519                         memcpy(&key[0], &ctx->key[0], key_len);
3520                 }
3521                 rcu_read_unlock();
3522
3523                 len = min_t(unsigned int, len, key_len);
3524                 if (put_user(len, optlen))
3525                         return -EFAULT;
3526                 if (copy_to_user(optval, key, len))
3527                         return -EFAULT;
3528                 return 0;
3529         }
3530         case TCP_THIN_LINEAR_TIMEOUTS:
3531                 val = tp->thin_lto;
3532                 break;
3533
3534         case TCP_THIN_DUPACK:
3535                 val = 0;
3536                 break;
3537
3538         case TCP_REPAIR:
3539                 val = tp->repair;
3540                 break;
3541
3542         case TCP_REPAIR_QUEUE:
3543                 if (tp->repair)
3544                         val = tp->repair_queue;
3545                 else
3546                         return -EINVAL;
3547                 break;
3548
3549         case TCP_REPAIR_WINDOW: {
3550                 struct tcp_repair_window opt;
3551
3552                 if (get_user(len, optlen))
3553                         return -EFAULT;
3554
3555                 if (len != sizeof(opt))
3556                         return -EINVAL;
3557
3558                 if (!tp->repair)
3559                         return -EPERM;
3560
3561                 opt.snd_wl1     = tp->snd_wl1;
3562                 opt.snd_wnd     = tp->snd_wnd;
3563                 opt.max_window  = tp->max_window;
3564                 opt.rcv_wnd     = tp->rcv_wnd;
3565                 opt.rcv_wup     = tp->rcv_wup;
3566
3567                 if (copy_to_user(optval, &opt, len))
3568                         return -EFAULT;
3569                 return 0;
3570         }
3571         case TCP_QUEUE_SEQ:
3572                 if (tp->repair_queue == TCP_SEND_QUEUE)
3573                         val = tp->write_seq;
3574                 else if (tp->repair_queue == TCP_RECV_QUEUE)
3575                         val = tp->rcv_nxt;
3576                 else
3577                         return -EINVAL;
3578                 break;
3579
3580         case TCP_USER_TIMEOUT:
3581                 val = icsk->icsk_user_timeout;
3582                 break;
3583
3584         case TCP_FASTOPEN:
3585                 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3586                 break;
3587
3588         case TCP_FASTOPEN_CONNECT:
3589                 val = tp->fastopen_connect;
3590                 break;
3591
3592         case TCP_FASTOPEN_NO_COOKIE:
3593                 val = tp->fastopen_no_cookie;
3594                 break;
3595
3596         case TCP_TX_DELAY:
3597                 val = tp->tcp_tx_delay;
3598                 break;
3599
3600         case TCP_TIMESTAMP:
3601                 val = tcp_time_stamp_raw() + tp->tsoffset;
3602                 break;
3603         case TCP_NOTSENT_LOWAT:
3604                 val = tp->notsent_lowat;
3605                 break;
3606         case TCP_INQ:
3607                 val = tp->recvmsg_inq;
3608                 break;
3609         case TCP_SAVE_SYN:
3610                 val = tp->save_syn;
3611                 break;
3612         case TCP_SAVED_SYN: {
3613                 if (get_user(len, optlen))
3614                         return -EFAULT;
3615
3616                 lock_sock(sk);
3617                 if (tp->saved_syn) {
3618                         if (len < tp->saved_syn[0]) {
3619                                 if (put_user(tp->saved_syn[0], optlen)) {
3620                                         release_sock(sk);
3621                                         return -EFAULT;
3622                                 }
3623                                 release_sock(sk);
3624                                 return -EINVAL;
3625                         }
3626                         len = tp->saved_syn[0];
3627                         if (put_user(len, optlen)) {
3628                                 release_sock(sk);
3629                                 return -EFAULT;
3630                         }
3631                         if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3632                                 release_sock(sk);
3633                                 return -EFAULT;
3634                         }
3635                         tcp_saved_syn_free(tp);
3636                         release_sock(sk);
3637                 } else {
3638                         release_sock(sk);
3639                         len = 0;
3640                         if (put_user(len, optlen))
3641                                 return -EFAULT;
3642                 }
3643                 return 0;
3644         }
3645 #ifdef CONFIG_MMU
3646         case TCP_ZEROCOPY_RECEIVE: {
3647                 struct tcp_zerocopy_receive zc;
3648                 int err;
3649
3650                 if (get_user(len, optlen))
3651                         return -EFAULT;
3652                 if (len != sizeof(zc))
3653                         return -EINVAL;
3654                 if (copy_from_user(&zc, optval, len))
3655                         return -EFAULT;
3656                 lock_sock(sk);
3657                 err = tcp_zerocopy_receive(sk, &zc);
3658                 release_sock(sk);
3659                 if (!err && copy_to_user(optval, &zc, len))
3660                         err = -EFAULT;
3661                 return err;
3662         }
3663 #endif
3664         default:
3665                 return -ENOPROTOOPT;
3666         }
3667
3668         if (put_user(len, optlen))
3669                 return -EFAULT;
3670         if (copy_to_user(optval, &val, len))
3671                 return -EFAULT;
3672         return 0;
3673 }
3674
3675 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3676                    int __user *optlen)
3677 {
3678         struct inet_connection_sock *icsk = inet_csk(sk);
3679
3680         if (level != SOL_TCP)
3681                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3682                                                      optval, optlen);
3683         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3684 }
3685 EXPORT_SYMBOL(tcp_getsockopt);
3686
3687 #ifdef CONFIG_COMPAT
3688 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3689                           char __user *optval, int __user *optlen)
3690 {
3691         if (level != SOL_TCP)
3692                 return inet_csk_compat_getsockopt(sk, level, optname,
3693                                                   optval, optlen);
3694         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3695 }
3696 EXPORT_SYMBOL(compat_tcp_getsockopt);
3697 #endif
3698
3699 #ifdef CONFIG_TCP_MD5SIG
3700 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3701 static DEFINE_MUTEX(tcp_md5sig_mutex);
3702 static bool tcp_md5sig_pool_populated = false;
3703
3704 static void __tcp_alloc_md5sig_pool(void)
3705 {
3706         struct crypto_ahash *hash;
3707         int cpu;
3708
3709         hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3710         if (IS_ERR(hash))
3711                 return;
3712
3713         for_each_possible_cpu(cpu) {
3714                 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3715                 struct ahash_request *req;
3716
3717                 if (!scratch) {
3718                         scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3719                                                sizeof(struct tcphdr),
3720                                                GFP_KERNEL,
3721                                                cpu_to_node(cpu));
3722                         if (!scratch)
3723                                 return;
3724                         per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3725                 }
3726                 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3727                         continue;
3728
3729                 req = ahash_request_alloc(hash, GFP_KERNEL);
3730                 if (!req)
3731                         return;
3732
3733                 ahash_request_set_callback(req, 0, NULL, NULL);
3734
3735                 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3736         }
3737         /* before setting tcp_md5sig_pool_populated, we must commit all writes
3738          * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3739          */
3740         smp_wmb();
3741         tcp_md5sig_pool_populated = true;
3742 }
3743
3744 bool tcp_alloc_md5sig_pool(void)
3745 {
3746         if (unlikely(!tcp_md5sig_pool_populated)) {
3747                 mutex_lock(&tcp_md5sig_mutex);
3748
3749                 if (!tcp_md5sig_pool_populated) {
3750                         __tcp_alloc_md5sig_pool();
3751                         if (tcp_md5sig_pool_populated)
3752                                 static_branch_inc(&tcp_md5_needed);
3753                 }
3754
3755                 mutex_unlock(&tcp_md5sig_mutex);
3756         }
3757         return tcp_md5sig_pool_populated;
3758 }
3759 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3760
3761
3762 /**
3763  *      tcp_get_md5sig_pool - get md5sig_pool for this user
3764  *
3765  *      We use percpu structure, so if we succeed, we exit with preemption
3766  *      and BH disabled, to make sure another thread or softirq handling
3767  *      wont try to get same context.
3768  */
3769 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3770 {
3771         local_bh_disable();
3772
3773         if (tcp_md5sig_pool_populated) {
3774                 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3775                 smp_rmb();
3776                 return this_cpu_ptr(&tcp_md5sig_pool);
3777         }
3778         local_bh_enable();
3779         return NULL;
3780 }
3781 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3782
3783 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3784                           const struct sk_buff *skb, unsigned int header_len)
3785 {
3786         struct scatterlist sg;
3787         const struct tcphdr *tp = tcp_hdr(skb);
3788         struct ahash_request *req = hp->md5_req;
3789         unsigned int i;
3790         const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3791                                            skb_headlen(skb) - header_len : 0;
3792         const struct skb_shared_info *shi = skb_shinfo(skb);
3793         struct sk_buff *frag_iter;
3794
3795         sg_init_table(&sg, 1);
3796
3797         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3798         ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3799         if (crypto_ahash_update(req))
3800                 return 1;
3801
3802         for (i = 0; i < shi->nr_frags; ++i) {
3803                 const skb_frag_t *f = &shi->frags[i];
3804                 unsigned int offset = skb_frag_off(f);
3805                 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3806
3807                 sg_set_page(&sg, page, skb_frag_size(f),
3808                             offset_in_page(offset));
3809                 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3810                 if (crypto_ahash_update(req))
3811                         return 1;
3812         }
3813
3814         skb_walk_frags(skb, frag_iter)
3815                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3816                         return 1;
3817
3818         return 0;
3819 }
3820 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3821
3822 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3823 {
3824         struct scatterlist sg;
3825
3826         sg_init_one(&sg, key->key, key->keylen);
3827         ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen);
3828         return crypto_ahash_update(hp->md5_req);
3829 }
3830 EXPORT_SYMBOL(tcp_md5_hash_key);
3831
3832 #endif
3833
3834 void tcp_done(struct sock *sk)
3835 {
3836         struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3837
3838         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3839                 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3840
3841         tcp_set_state(sk, TCP_CLOSE);
3842         tcp_clear_xmit_timers(sk);
3843         if (req)
3844                 reqsk_fastopen_remove(sk, req, false);
3845
3846         sk->sk_shutdown = SHUTDOWN_MASK;
3847
3848         if (!sock_flag(sk, SOCK_DEAD))
3849                 sk->sk_state_change(sk);
3850         else
3851                 inet_csk_destroy_sock(sk);
3852 }
3853 EXPORT_SYMBOL_GPL(tcp_done);
3854
3855 int tcp_abort(struct sock *sk, int err)
3856 {
3857         if (!sk_fullsock(sk)) {
3858                 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3859                         struct request_sock *req = inet_reqsk(sk);
3860
3861                         local_bh_disable();
3862                         inet_csk_reqsk_queue_drop(req->rsk_listener, req);
3863                         local_bh_enable();
3864                         return 0;
3865                 }
3866                 return -EOPNOTSUPP;
3867         }
3868
3869         /* Don't race with userspace socket closes such as tcp_close. */
3870         lock_sock(sk);
3871
3872         if (sk->sk_state == TCP_LISTEN) {
3873                 tcp_set_state(sk, TCP_CLOSE);
3874                 inet_csk_listen_stop(sk);
3875         }
3876
3877         /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3878         local_bh_disable();
3879         bh_lock_sock(sk);
3880
3881         if (!sock_flag(sk, SOCK_DEAD)) {
3882                 sk->sk_err = err;
3883                 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3884                 smp_wmb();
3885                 sk->sk_error_report(sk);
3886                 if (tcp_need_reset(sk->sk_state))
3887                         tcp_send_active_reset(sk, GFP_ATOMIC);
3888                 tcp_done(sk);
3889         }
3890
3891         bh_unlock_sock(sk);
3892         local_bh_enable();
3893         tcp_write_queue_purge(sk);
3894         release_sock(sk);
3895         return 0;
3896 }
3897 EXPORT_SYMBOL_GPL(tcp_abort);
3898
3899 extern struct tcp_congestion_ops tcp_reno;
3900
3901 static __initdata unsigned long thash_entries;
3902 static int __init set_thash_entries(char *str)
3903 {
3904         ssize_t ret;
3905
3906         if (!str)
3907                 return 0;
3908
3909         ret = kstrtoul(str, 0, &thash_entries);
3910         if (ret)
3911                 return 0;
3912
3913         return 1;
3914 }
3915 __setup("thash_entries=", set_thash_entries);
3916
3917 static void __init tcp_init_mem(void)
3918 {
3919         unsigned long limit = nr_free_buffer_pages() / 16;
3920
3921         limit = max(limit, 128UL);
3922         sysctl_tcp_mem[0] = limit / 4 * 3;              /* 4.68 % */
3923         sysctl_tcp_mem[1] = limit;                      /* 6.25 % */
3924         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;      /* 9.37 % */
3925 }
3926
3927 void __init tcp_init(void)
3928 {
3929         int max_rshare, max_wshare, cnt;
3930         unsigned long limit;
3931         unsigned int i;
3932
3933         BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
3934         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3935                      FIELD_SIZEOF(struct sk_buff, cb));
3936
3937         percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3938         percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3939         inet_hashinfo_init(&tcp_hashinfo);
3940         inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
3941                             thash_entries, 21,  /* one slot per 2 MB*/
3942                             0, 64 * 1024);
3943         tcp_hashinfo.bind_bucket_cachep =
3944                 kmem_cache_create("tcp_bind_bucket",
3945                                   sizeof(struct inet_bind_bucket), 0,
3946                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3947
3948         /* Size and allocate the main established and bind bucket
3949          * hash tables.
3950          *
3951          * The methodology is similar to that of the buffer cache.
3952          */
3953         tcp_hashinfo.ehash =
3954                 alloc_large_system_hash("TCP established",
3955                                         sizeof(struct inet_ehash_bucket),
3956                                         thash_entries,
3957                                         17, /* one slot per 128 KB of memory */
3958                                         0,
3959                                         NULL,
3960                                         &tcp_hashinfo.ehash_mask,
3961                                         0,
3962                                         thash_entries ? 0 : 512 * 1024);
3963         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3964                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3965
3966         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3967                 panic("TCP: failed to alloc ehash_locks");
3968         tcp_hashinfo.bhash =
3969                 alloc_large_system_hash("TCP bind",
3970                                         sizeof(struct inet_bind_hashbucket),
3971                                         tcp_hashinfo.ehash_mask + 1,
3972                                         17, /* one slot per 128 KB of memory */
3973                                         0,
3974                                         &tcp_hashinfo.bhash_size,
3975                                         NULL,
3976                                         0,
3977                                         64 * 1024);
3978         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3979         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3980                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3981                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3982         }
3983
3984
3985         cnt = tcp_hashinfo.ehash_mask + 1;
3986         sysctl_tcp_max_orphans = cnt / 2;
3987
3988         tcp_init_mem();
3989         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3990         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3991         max_wshare = min(4UL*1024*1024, limit);
3992         max_rshare = min(6UL*1024*1024, limit);
3993
3994         init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3995         init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
3996         init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3997
3998         init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3999         init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4000         init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4001
4002         pr_info("Hash tables configured (established %u bind %u)\n",
4003                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4004
4005         tcp_v4_init();
4006         tcp_metrics_init();
4007         BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
4008         tcp_tasklet_init();
4009 }