net/ipv4/tcp.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the  BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              Implementation of the Transmission Control Protocol(TCP).
   7  *
   8  * Authors:     Ross Biro
   9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
  11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
  12  *              Florian La Roche, <flla@stud.uni-sb.de>
  13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
  15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
  16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
  17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  18  *              Jorge Cwik, <jorge@laser.satlink.net>
  19  *
  20  * Fixes:
  21  *              Alan Cox        :       Numerous verify_area() calls
  22  *              Alan Cox        :       Set the ACK bit on a reset
  23  *              Alan Cox        :       Stopped it crashing if it closed while
  24  *                                      sk->inuse=1 and was trying to connect
  25  *                                      (tcp_err()).
  26  *              Alan Cox        :       All icmp error handling was broken
  27  *                                      pointers passed where wrong and the
  28  *                                      socket was looked up backwards. Nobody
  29  *                                      tested any icmp error code obviously.
  30  *              Alan Cox        :       tcp_err() now handled properly. It
  31  *                                      wakes people on errors. poll
  32  *                                      behaves and the icmp error race
  33  *                                      has gone by moving it into sock.c
  34  *              Alan Cox        :       tcp_send_reset() fixed to work for
  35  *                                      everything not just packets for
  36  *                                      unknown sockets.
  37  *              Alan Cox        :       tcp option processing.
  38  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
  39  *                                      syn rule wrong]
  40  *              Herp Rosmanith  :       More reset fixes
  41  *              Alan Cox        :       No longer acks invalid rst frames.
  42  *                                      Acking any kind of RST is right out.
  43  *              Alan Cox        :       Sets an ignore me flag on an rst
  44  *                                      receive otherwise odd bits of prattle
  45  *                                      escape still
  46  *              Alan Cox        :       Fixed another acking RST frame bug.
  47  *                                      Should stop LAN workplace lockups.
  48  *              Alan Cox        :       Some tidyups using the new skb list
  49  *                                      facilities
  50  *              Alan Cox        :       sk->keepopen now seems to work
  51  *              Alan Cox        :       Pulls options out correctly on accepts
  52  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
  53  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
  54  *                                      bit to skb ops.
  55  *              Alan Cox        :       Tidied tcp_data to avoid a potential
  56  *                                      nasty.
  57  *              Alan Cox        :       Added some better commenting, as the
  58  *                                      tcp is hard to follow
  59  *              Alan Cox        :       Removed incorrect check for 20 * psh
  60  *      Michael O'Reilly        :       ack < copied bug fix.
  61  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
  62  *              Alan Cox        :       FIN with no memory -> CRASH
  63  *              Alan Cox        :       Added socket option proto entries.
  64  *                                      Also added awareness of them to accept.
  65  *              Alan Cox        :       Added TCP options (SOL_TCP)
  66  *              Alan Cox        :       Switched wakeup calls to callbacks,
  67  *                                      so the kernel can layer network
  68  *                                      sockets.
  69  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
  70  *              Alan Cox        :       Handle FIN (more) properly (we hope).
  71  *              Alan Cox        :       RST frames sent on unsynchronised
  72  *                                      state ack error.
  73  *              Alan Cox        :       Put in missing check for SYN bit.
  74  *              Alan Cox        :       Added tcp_select_window() aka NET2E
  75  *                                      window non shrink trick.
  76  *              Alan Cox        :       Added a couple of small NET2E timer
  77  *                                      fixes
  78  *              Charles Hedrick :       TCP fixes
  79  *              Toomas Tamm     :       TCP window fixes
  80  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
  81  *              Charles Hedrick :       Rewrote most of it to actually work
  82  *              Linus           :       Rewrote tcp_read() and URG handling
  83  *                                      completely
  84  *              Gerhard Koerting:       Fixed some missing timer handling
  85  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
  86  *              Gerhard Koerting:       PC/TCP workarounds
  87  *              Adam Caldwell   :       Assorted timer/timing errors
  88  *              Matthew Dillon  :       Fixed another RST bug
  89  *              Alan Cox        :       Move to kernel side addressing changes.
  90  *              Alan Cox        :       Beginning work on TCP fastpathing
  91  *                                      (not yet usable)
  92  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
  93  *              Alan Cox        :       TCP fast path debugging
  94  *              Alan Cox        :       Window clamping
  95  *              Michael Riepe   :       Bug in tcp_check()
  96  *              Matt Dillon     :       More TCP improvements and RST bug fixes
  97  *              Matt Dillon     :       Yet more small nasties remove from the
  98  *                                      TCP code (Be very nice to this man if
  99  *                                      tcp finally works 100%) 8)
 100  *              Alan Cox        :       BSD accept semantics.
 101  *              Alan Cox        :       Reset on closedown bug.
 102  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
 103  *              Michael Pall    :       Handle poll() after URG properly in
 104  *                                      all cases.
 105  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
 106  *                                      (multi URG PUSH broke rlogin).
 107  *              Michael Pall    :       Fix the multi URG PUSH problem in
 108  *                                      tcp_readable(), poll() after URG
 109  *                                      works now.
 110  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
 111  *                                      BSD api.
 112  *              Alan Cox        :       Changed the semantics of sk->socket to
 113  *                                      fix a race and a signal problem with
 114  *                                      accept() and async I/O.
 115  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
 116  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
 117  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
 118  *                                      clients/servers which listen in on
 119  *                                      fixed ports.
 120  *              Alan Cox        :       Cleaned the above up and shrank it to
 121  *                                      a sensible code size.
 122  *              Alan Cox        :       Self connect lockup fix.
 123  *              Alan Cox        :       No connect to multicast.
 124  *              Ross Biro       :       Close unaccepted children on master
 125  *                                      socket close.
 126  *              Alan Cox        :       Reset tracing code.
 127  *              Alan Cox        :       Spurious resets on shutdown.
 128  *              Alan Cox        :       Giant 15 minute/60 second timer error
 129  *              Alan Cox        :       Small whoops in polling before an
 130  *                                      accept.
 131  *              Alan Cox        :       Kept the state trace facility since
 132  *                                      it's handy for debugging.
 133  *              Alan Cox        :       More reset handler fixes.
 134  *              Alan Cox        :       Started rewriting the code based on
 135  *                                      the RFC's for other useful protocol
 136  *                                      references see: Comer, KA9Q NOS, and
 137  *                                      for a reference on the difference
 138  *                                      between specifications and how BSD
 139  *                                      works see the 4.4lite source.
 140  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
 141  *                                      close.
 142  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
 143  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
 144  *              Alan Cox        :       Reimplemented timers as per the RFC
 145  *                                      and using multiple timers for sanity.
 146  *              Alan Cox        :       Small bug fixes, and a lot of new
 147  *                                      comments.
 148  *              Alan Cox        :       Fixed dual reader crash by locking
 149  *                                      the buffers (much like datagram.c)
 150  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
 151  *                                      now gets fed up of retrying without
 152  *                                      (even a no space) answer.
 153  *              Alan Cox        :       Extracted closing code better
 154  *              Alan Cox        :       Fixed the closing state machine to
 155  *                                      resemble the RFC.
 156  *              Alan Cox        :       More 'per spec' fixes.
 157  *              Jorge Cwik      :       Even faster checksumming.
 158  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
 159  *                                      only frames. At least one pc tcp stack
 160  *                                      generates them.
 161  *              Alan Cox        :       Cache last socket.
 162  *              Alan Cox        :       Per route irtt.
 163  *              Matt Day        :       poll()->select() match BSD precisely on error
 164  *              Alan Cox        :       New buffers
 165  *              Marc Tamsky     :       Various sk->prot->retransmits and
 166  *                                      sk->retransmits misupdating fixed.
 167  *                                      Fixed tcp_write_timeout: stuck close,
 168  *                                      and TCP syn retries gets used now.
 169  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
 170  *                                      ack if state is TCP_CLOSED.
 171  *              Alan Cox        :       Look up device on a retransmit - routes may
 172  *                                      change. Doesn't yet cope with MSS shrink right
 173  *                                      but it's a start!
 174  *              Marc Tamsky     :       Closing in closing fixes.
 175  *              Mike Shaver     :       RFC1122 verifications.
 176  *              Alan Cox        :       rcv_saddr errors.
 177  *              Alan Cox        :       Block double connect().
 178  *              Alan Cox        :       Small hooks for enSKIP.
 179  *              Alexey Kuznetsov:       Path MTU discovery.
 180  *              Alan Cox        :       Support soft errors.
 181  *              Alan Cox        :       Fix MTU discovery pathological case
 182  *                                      when the remote claims no mtu!
 183  *              Marc Tamsky     :       TCP_CLOSE fix.
 184  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
 185  *                                      window but wrong (fixes NT lpd problems)
 186  *              Pedro Roque     :       Better TCP window handling, delayed ack.
 187  *              Joerg Reuter    :       No modification of locked buffers in
 188  *                                      tcp_do_retransmit()
 189  *              Eric Schenk     :       Changed receiver side silly window
 190  *                                      avoidance algorithm to BSD style
 191  *                                      algorithm. This doubles throughput
 192  *                                      against machines running Solaris,
 193  *                                      and seems to result in general
 194  *                                      improvement.
 195  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
 196  *      Willy Konynenberg       :       Transparent proxying support.
 197  *      Mike McLagan            :       Routing by source
 198  *              Keith Owens     :       Do proper merging with partial SKB's in
 199  *                                      tcp_do_sendmsg to avoid burstiness.
 200  *              Eric Schenk     :       Fix fast close down bug with
 201  *                                      shutdown() followed by close().
 202  *              Andi Kleen      :       Make poll agree with SIGIO
 203  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
 204  *                                      lingertime == 0 (RFC 793 ABORT Call)
 205  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
 206  *                                      csum_and_copy_from_user() if possible.
 207  *
 208  *              This program is free software; you can redistribute it and/or
 209  *              modify it under the terms of the GNU General Public License
 210  *              as published by the Free Software Foundation; either version
 211  *              2 of the License, or(at your option) any later version.
 212  *
 213  * Description of States:
 214  *
 215  *      TCP_SYN_SENT            sent a connection request, waiting for ack
 216  *
 217  *      TCP_SYN_RECV            received a connection request, sent ack,
 218  *                              waiting for final ack in three-way handshake.
 219  *
 220  *      TCP_ESTABLISHED         connection established
 221  *
 222  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
 223  *                              transmission of remaining buffered data
 224  *
 225  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
 226  *                              to shutdown
 227  *
 228  *      TCP_CLOSING             both sides have shutdown but we still have
 229  *                              data we have to finish sending
 230  *
 231  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
 232  *                              closed, can only be entered from FIN_WAIT2
 233  *                              or CLOSING.  Required because the other end
 234  *                              may not have gotten our last ACK causing it
 235  *                              to retransmit the data packet (which we ignore)
 236  *
 237  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
 238  *                              us to finish writing our data and to shutdown
 239  *                              (we have to close() to move on to LAST_ACK)
 240  *
 241  *      TCP_LAST_ACK            out side has shutdown after remote has
 242  *                              shutdown.  There may still be data in our
 243  *                              buffer that we have to finish sending
 244  *
 245  *      TCP_CLOSE               socket is finished
 246  */
 247
 248 #define pr_fmt(fmt) "TCP: " fmt
 249
 250 #include <linux/kernel.h>
 251 #include <linux/module.h>
 252 #include <linux/types.h>
 253 #include <linux/fcntl.h>
 254 #include <linux/poll.h>
 255 #include <linux/inet_diag.h>
 256 #include <linux/init.h>
 257 #include <linux/fs.h>
 258 #include <linux/skbuff.h>
 259 #include <linux/scatterlist.h>
 260 #include <linux/splice.h>
 261 #include <linux/net.h>
 262 #include <linux/socket.h>
 263 #include <linux/random.h>
 264 #include <linux/bootmem.h>
 265 #include <linux/highmem.h>
 266 #include <linux/swap.h>
 267 #include <linux/cache.h>
 268 #include <linux/err.h>
 269 #include <linux/crypto.h>
 270 #include <linux/time.h>
 271 #include <linux/slab.h>
 272
 273 #include <net/icmp.h>
 274 #include <net/inet_common.h>
 275 #include <net/tcp.h>
 276 #include <net/xfrm.h>
 277 #include <net/ip.h>
 278 #include <net/sock.h>
 279
 280 #include <asm/uaccess.h>
 281 #include <asm/ioctls.h>
 282 #include <net/busy_poll.h>
 283
 284 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
 285
 286 int sysctl_tcp_min_tso_segs __read_mostly = 2;
 287
 288 int sysctl_tcp_autocorking __read_mostly = 1;
 289
 290 struct percpu_counter tcp_orphan_count;
 291 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 292
 293 long sysctl_tcp_mem[3] __read_mostly;
 294 int sysctl_tcp_wmem[3] __read_mostly;
 295 int sysctl_tcp_rmem[3] __read_mostly;
 296
 297 EXPORT_SYMBOL(sysctl_tcp_mem);
 298 EXPORT_SYMBOL(sysctl_tcp_rmem);
 299 EXPORT_SYMBOL(sysctl_tcp_wmem);
 300
 301 atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
 302 EXPORT_SYMBOL(tcp_memory_allocated);
 303
 304 /*
 305  * Current number of TCP sockets.
 306  */
 307 struct percpu_counter tcp_sockets_allocated;
 308 EXPORT_SYMBOL(tcp_sockets_allocated);
 309
 310 /*
 311  * TCP splice context
 312  */
 313 struct tcp_splice_state {
 314         struct pipe_inode_info *pipe;
 315         size_t len;
 316         unsigned int flags;
 317 };
 318
 319 /*
 320  * Pressure flag: try to collapse.
 321  * Technical note: it is used by multiple contexts non atomically.
 322  * All the __sk_mem_schedule() is of this nature: accounting
 323  * is strict, actions are advisory and have some latency.
 324  */
 325 int tcp_memory_pressure __read_mostly;
 326 EXPORT_SYMBOL(tcp_memory_pressure);
 327
 328 void tcp_enter_memory_pressure(struct sock *sk)
 329 {
 330         if (!tcp_memory_pressure) {
 331                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
 332                 tcp_memory_pressure = 1;
 333         }
 334 }
 335 EXPORT_SYMBOL(tcp_enter_memory_pressure);
 336
 337 /* Convert seconds to retransmits based on initial and max timeout */
 338 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
 339 {
 340         u8 res = 0;
 341
 342         if (seconds > 0) {
 343                 int period = timeout;
 344
 345                 res = 1;
 346                 while (seconds > period && res < 255) {
 347                         res++;
 348                         timeout <<= 1;
 349                         if (timeout > rto_max)
 350                                 timeout = rto_max;
 351                         period += timeout;
 352                 }
 353         }
 354         return res;
 355 }
 356
 357 /* Convert retransmits to seconds based on initial and max timeout */
 358 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
 359 {
 360         int period = 0;
 361
 362         if (retrans > 0) {
 363                 period = timeout;
 364                 while (--retrans) {
 365                         timeout <<= 1;
 366                         if (timeout > rto_max)
 367                                 timeout = rto_max;
 368                         period += timeout;
 369                 }
 370         }
 371         return period;
 372 }
 373
 374 /* Address-family independent initialization for a tcp_sock.
 375  *
 376  * NOTE: A lot of things set to zero explicitly by call to
 377  *       sk_alloc() so need not be done here.
 378  */
 379 void tcp_init_sock(struct sock *sk)
 380 {
 381         struct inet_connection_sock *icsk = inet_csk(sk);
 382         struct tcp_sock *tp = tcp_sk(sk);
 383
 384         __skb_queue_head_init(&tp->out_of_order_queue);
 385         tcp_init_xmit_timers(sk);
 386         tcp_prequeue_init(tp);
 387         INIT_LIST_HEAD(&tp->tsq_node);
 388
 389         icsk->icsk_rto = TCP_TIMEOUT_INIT;
 390         tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
 391
 392         /* So many TCP implementations out there (incorrectly) count the
 393          * initial SYN frame in their delayed-ACK and congestion control
 394          * algorithms that we must have the following bandaid to talk
 395          * efficiently to them.  -DaveM
 396          */
 397         tp->snd_cwnd = TCP_INIT_CWND;
 398
 399         /* See draft-stevens-tcpca-spec-01 for discussion of the
 400          * initialization of these values.
 401          */
 402         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 403         tp->snd_cwnd_clamp = ~0;
 404         tp->mss_cache = TCP_MSS_DEFAULT;
 405
 406         tp->reordering = sysctl_tcp_reordering;
 407         tcp_enable_early_retrans(tp);
 408         tcp_assign_congestion_control(sk);
 409
 410         tp->tsoffset = 0;
 411
 412         sk->sk_state = TCP_CLOSE;
 413
 414         sk->sk_write_space = sk_stream_write_space;
 415         sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
 416
 417         icsk->icsk_sync_mss = tcp_sync_mss;
 418
 419         sk->sk_sndbuf = sysctl_tcp_wmem[1];
 420         sk->sk_rcvbuf = sysctl_tcp_rmem[1];
 421
 422         local_bh_disable();
 423         sock_update_memcg(sk);
 424         sk_sockets_allocated_inc(sk);
 425         local_bh_enable();
 426 }
 427 EXPORT_SYMBOL(tcp_init_sock);
 428
 429 static void tcp_tx_timestamp(struct sock *sk, struct sk_buff *skb)
 430 {
 431         if (sk->sk_tsflags) {
 432                 struct skb_shared_info *shinfo = skb_shinfo(skb);
 433
 434                 sock_tx_timestamp(sk, &shinfo->tx_flags);
 435                 if (shinfo->tx_flags & SKBTX_ANY_TSTAMP)
 436                         shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
 437         }
 438 }
 439
 440 /*
 441  *      Wait for a TCP event.
 442  *
 443  *      Note that we don't need to lock the socket, as the upper poll layers
 444  *      take care of normal races (between the test and the event) and we don't
 445  *      go look at any of the socket buffers directly.
 446  */
 447 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 448 {
 449         unsigned int mask;
 450         struct sock *sk = sock->sk;
 451         const struct tcp_sock *tp = tcp_sk(sk);
 452
 453         sock_rps_record_flow(sk);
 454
 455         sock_poll_wait(file, sk_sleep(sk), wait);
 456         if (sk->sk_state == TCP_LISTEN)
 457                 return inet_csk_listen_poll(sk);
 458
 459         /* Socket is not locked. We are protected from async events
 460          * by poll logic and correct handling of state changes
 461          * made by other threads is impossible in any case.
 462          */
 463
 464         mask = 0;
 465
 466         /*
 467          * POLLHUP is certainly not done right. But poll() doesn't
 468          * have a notion of HUP in just one direction, and for a
 469          * socket the read side is more interesting.
 470          *
 471          * Some poll() documentation says that POLLHUP is incompatible
 472          * with the POLLOUT/POLLWR flags, so somebody should check this
 473          * all. But careful, it tends to be safer to return too many
 474          * bits than too few, and you can easily break real applications
 475          * if you don't tell them that something has hung up!
 476          *
 477          * Check-me.
 478          *
 479          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
 480          * our fs/select.c). It means that after we received EOF,
 481          * poll always returns immediately, making impossible poll() on write()
 482          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
 483          * if and only if shutdown has been made in both directions.
 484          * Actually, it is interesting to look how Solaris and DUX
 485          * solve this dilemma. I would prefer, if POLLHUP were maskable,
 486          * then we could set it on SND_SHUTDOWN. BTW examples given
 487          * in Stevens' books assume exactly this behaviour, it explains
 488          * why POLLHUP is incompatible with POLLOUT.    --ANK
 489          *
 490          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 491          * blocking on fresh not-connected or disconnected socket. --ANK
 492          */
 493         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
 494                 mask |= POLLHUP;
 495         if (sk->sk_shutdown & RCV_SHUTDOWN)
 496                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
 497
 498         /* Connected or passive Fast Open socket? */
 499         if (sk->sk_state != TCP_SYN_SENT &&
 500             (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk)) {
 501                 int target = sock_rcvlowat(sk, 0, INT_MAX);
 502
 503                 if (tp->urg_seq == tp->copied_seq &&
 504                     !sock_flag(sk, SOCK_URGINLINE) &&
 505                     tp->urg_data)
 506                         target++;
 507
 508                 /* Potential race condition. If read of tp below will
 509                  * escape above sk->sk_state, we can be illegally awaken
 510                  * in SYN_* states. */
 511                 if (tp->rcv_nxt - tp->copied_seq >= target)
 512                         mask |= POLLIN | POLLRDNORM;
 513
 514                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 515                         if (sk_stream_is_writeable(sk)) {
 516                                 mask |= POLLOUT | POLLWRNORM;
 517                         } else {  /* send SIGIO later */
 518                                 set_bit(SOCK_ASYNC_NOSPACE,
 519                                         &sk->sk_socket->flags);
 520                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 521
 522                                 /* Race breaker. If space is freed after
 523                                  * wspace test but before the flags are set,
 524                                  * IO signal will be lost. Memory barrier
 525                                  * pairs with the input side.
 526                                  */
 527                                 smp_mb__after_atomic();
 528                                 if (sk_stream_is_writeable(sk))
 529                                         mask |= POLLOUT | POLLWRNORM;
 530                         }
 531                 } else
 532                         mask |= POLLOUT | POLLWRNORM;
 533
 534                 if (tp->urg_data & TCP_URG_VALID)
 535                         mask |= POLLPRI;
 536         }
 537         /* This barrier is coupled with smp_wmb() in tcp_reset() */
 538         smp_rmb();
 539         if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
 540                 mask |= POLLERR;
 541
 542         return mask;
 543 }
 544 EXPORT_SYMBOL(tcp_poll);
 545
 546 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 547 {
 548         struct tcp_sock *tp = tcp_sk(sk);
 549         int answ;
 550         bool slow;
 551
 552         switch (cmd) {
 553         case SIOCINQ:
 554                 if (sk->sk_state == TCP_LISTEN)
 555                         return -EINVAL;
 556
 557                 slow = lock_sock_fast(sk);
 558                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 559                         answ = 0;
 560                 else if (sock_flag(sk, SOCK_URGINLINE) ||
 561                          !tp->urg_data ||
 562                          before(tp->urg_seq, tp->copied_seq) ||
 563                          !before(tp->urg_seq, tp->rcv_nxt)) {
 564
 565                         answ = tp->rcv_nxt - tp->copied_seq;
 566
 567                         /* Subtract 1, if FIN was received */
 568                         if (answ && sock_flag(sk, SOCK_DONE))
 569                                 answ--;
 570                 } else
 571                         answ = tp->urg_seq - tp->copied_seq;
 572                 unlock_sock_fast(sk, slow);
 573                 break;
 574         case SIOCATMARK:
 575                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 576                 break;
 577         case SIOCOUTQ:
 578                 if (sk->sk_state == TCP_LISTEN)
 579                         return -EINVAL;
 580
 581                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 582                         answ = 0;
 583                 else
 584                         answ = tp->write_seq - tp->snd_una;
 585                 break;
 586         case SIOCOUTQNSD:
 587                 if (sk->sk_state == TCP_LISTEN)
 588                         return -EINVAL;
 589
 590                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 591                         answ = 0;
 592                 else
 593                         answ = tp->write_seq - tp->snd_nxt;
 594                 break;
 595         default:
 596                 return -ENOIOCTLCMD;
 597         }
 598
 599         return put_user(answ, (int __user *)arg);
 600 }
 601 EXPORT_SYMBOL(tcp_ioctl);
 602
 603 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 604 {
 605         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 606         tp->pushed_seq = tp->write_seq;
 607 }
 608
 609 static inline bool forced_push(const struct tcp_sock *tp)
 610 {
 611         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 612 }
 613
 614 static void skb_entail(struct sock *sk, struct sk_buff *skb)
 615 {
 616         struct tcp_sock *tp = tcp_sk(sk);
 617         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 618
 619         skb->csum    = 0;
 620         tcb->seq     = tcb->end_seq = tp->write_seq;
 621         tcb->tcp_flags = TCPHDR_ACK;
 622         tcb->sacked  = 0;
 623         __skb_header_release(skb);
 624         tcp_add_write_queue_tail(sk, skb);
 625         sk->sk_wmem_queued += skb->truesize;
 626         sk_mem_charge(sk, skb->truesize);
 627         if (tp->nonagle & TCP_NAGLE_PUSH)
 628                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 629 }
 630
 631 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
 632 {
 633         if (flags & MSG_OOB)
 634                 tp->snd_up = tp->write_seq;
 635 }
 636
 637 /* If a not yet filled skb is pushed, do not send it if
 638  * we have data packets in Qdisc or NIC queues :
 639  * Because TX completion will happen shortly, it gives a chance
 640  * to coalesce future sendmsg() payload into this skb, without
 641  * need for a timer, and with no latency trade off.
 642  * As packets containing data payload have a bigger truesize
 643  * than pure acks (dataless) packets, the last checks prevent
 644  * autocorking if we only have an ACK in Qdisc/NIC queues,
 645  * or if TX completion was delayed after we processed ACK packet.
 646  */
 647 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
 648                                 int size_goal)
 649 {
 650         return skb->len < size_goal &&
 651                sysctl_tcp_autocorking &&
 652                skb != tcp_write_queue_head(sk) &&
 653                atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
 654 }
 655
 656 static void tcp_push(struct sock *sk, int flags, int mss_now,
 657                      int nonagle, int size_goal)
 658 {
 659         struct tcp_sock *tp = tcp_sk(sk);
 660         struct sk_buff *skb;
 661
 662         if (!tcp_send_head(sk))
 663                 return;
 664
 665         skb = tcp_write_queue_tail(sk);
 666         if (!(flags & MSG_MORE) || forced_push(tp))
 667                 tcp_mark_push(tp, skb);
 668
 669         tcp_mark_urg(tp, flags);
 670
 671         if (tcp_should_autocork(sk, skb, size_goal)) {
 672
 673                 /* avoid atomic op if TSQ_THROTTLED bit is already set */
 674                 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
 675                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
 676                         set_bit(TSQ_THROTTLED, &tp->tsq_flags);
 677                 }
 678                 /* It is possible TX completion already happened
 679                  * before we set TSQ_THROTTLED.
 680                  */
 681                 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
 682                         return;
 683         }
 684
 685         if (flags & MSG_MORE)
 686                 nonagle = TCP_NAGLE_CORK;
 687
 688         __tcp_push_pending_frames(sk, mss_now, nonagle);
 689 }
 690
 691 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 692                                 unsigned int offset, size_t len)
 693 {
 694         struct tcp_splice_state *tss = rd_desc->arg.data;
 695         int ret;
 696
 697         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
 698                               tss->flags);
 699         if (ret > 0)
 700                 rd_desc->count -= ret;
 701         return ret;
 702 }
 703
 704 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 705 {
 706         /* Store TCP splice context information in read_descriptor_t. */
 707         read_descriptor_t rd_desc = {
 708                 .arg.data = tss,
 709                 .count    = tss->len,
 710         };
 711
 712         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 713 }
 714
 715 /**
 716  *  tcp_splice_read - splice data from TCP socket to a pipe
 717  * @sock:       socket to splice from
 718  * @ppos:       position (not valid)
 719  * @pipe:       pipe to splice to
 720  * @len:        number of bytes to splice
 721  * @flags:      splice modifier flags
 722  *
 723  * Description:
 724  *    Will read pages from given socket and fill them into a pipe.
 725  *
 726  **/
 727 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 728                         struct pipe_inode_info *pipe, size_t len,
 729                         unsigned int flags)
 730 {
 731         struct sock *sk = sock->sk;
 732         struct tcp_splice_state tss = {
 733                 .pipe = pipe,
 734                 .len = len,
 735                 .flags = flags,
 736         };
 737         long timeo;
 738         ssize_t spliced;
 739         int ret;
 740
 741         sock_rps_record_flow(sk);
 742         /*
 743          * We can't seek on a socket input
 744          */
 745         if (unlikely(*ppos))
 746                 return -ESPIPE;
 747
 748         ret = spliced = 0;
 749
 750         lock_sock(sk);
 751
 752         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 753         while (tss.len) {
 754                 ret = __tcp_splice_read(sk, &tss);
 755                 if (ret < 0)
 756                         break;
 757                 else if (!ret) {
 758                         if (spliced)
 759                                 break;
 760                         if (sock_flag(sk, SOCK_DONE))
 761                                 break;
 762                         if (sk->sk_err) {
 763                                 ret = sock_error(sk);
 764                                 break;
 765                         }
 766                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 767                                 break;
 768                         if (sk->sk_state == TCP_CLOSE) {
 769                                 /*
 770                                  * This occurs when user tries to read
 771                                  * from never connected socket.
 772                                  */
 773                                 if (!sock_flag(sk, SOCK_DONE))
 774                                         ret = -ENOTCONN;
 775                                 break;
 776                         }
 777                         if (!timeo) {
 778                                 ret = -EAGAIN;
 779                                 break;
 780                         }
 781                         sk_wait_data(sk, &timeo);
 782                         if (signal_pending(current)) {
 783                                 ret = sock_intr_errno(timeo);
 784                                 break;
 785                         }
 786                         continue;
 787                 }
 788                 tss.len -= ret;
 789                 spliced += ret;
 790
 791                 if (!timeo)
 792                         break;
 793                 release_sock(sk);
 794                 lock_sock(sk);
 795
 796                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 797                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 798                     signal_pending(current))
 799                         break;
 800         }
 801
 802         release_sock(sk);
 803
 804         if (spliced)
 805                 return spliced;
 806
 807         return ret;
 808 }
 809 EXPORT_SYMBOL(tcp_splice_read);
 810
 811 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
 812                                     bool force_schedule)
 813 {
 814         struct sk_buff *skb;
 815
 816         /* The TCP header must be at least 32-bit aligned.  */
 817         size = ALIGN(size, 4);
 818
 819         if (unlikely(tcp_under_memory_pressure(sk)))
 820                 sk_mem_reclaim_partial(sk);
 821
 822         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 823         if (likely(skb)) {
 824                 bool mem_scheduled;
 825
 826                 if (force_schedule) {
 827                         mem_scheduled = true;
 828                         sk_forced_mem_schedule(sk, skb->truesize);
 829                 } else {
 830                         mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
 831                 }
 832                 if (likely(mem_scheduled)) {
 833                         skb_reserve(skb, sk->sk_prot->max_header);
 834                         /*
 835                          * Make sure that we have exactly size bytes
 836                          * available to the caller, no more, no less.
 837                          */
 838                         skb->reserved_tailroom = skb->end - skb->tail - size;
 839                         return skb;
 840                 }
 841                 __kfree_skb(skb);
 842         } else {
 843                 sk->sk_prot->enter_memory_pressure(sk);
 844                 sk_stream_moderate_sndbuf(sk);
 845         }
 846         return NULL;
 847 }
 848
 849 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 850                                        int large_allowed)
 851 {
 852         struct tcp_sock *tp = tcp_sk(sk);
 853         u32 new_size_goal, size_goal;
 854
 855         if (!large_allowed || !sk_can_gso(sk))
 856                 return mss_now;
 857
 858         /* Note : tcp_tso_autosize() will eventually split this later */
 859         new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
 860         new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
 861
 862         /* We try hard to avoid divides here */
 863         size_goal = tp->gso_segs * mss_now;
 864         if (unlikely(new_size_goal < size_goal ||
 865                      new_size_goal >= size_goal + mss_now)) {
 866                 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
 867                                      sk->sk_gso_max_segs);
 868                 size_goal = tp->gso_segs * mss_now;
 869         }
 870
 871         return max(size_goal, mss_now);
 872 }
 873
 874 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 875 {
 876         int mss_now;
 877
 878         mss_now = tcp_current_mss(sk);
 879         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 880
 881         return mss_now;
 882 }
 883
 884 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
 885                                 size_t size, int flags)
 886 {
 887         struct tcp_sock *tp = tcp_sk(sk);
 888         int mss_now, size_goal;
 889         int err;
 890         ssize_t copied;
 891         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 892
 893         /* Wait for a connection to finish. One exception is TCP Fast Open
 894          * (passive side) where data is allowed to be sent before a connection
 895          * is fully established.
 896          */
 897         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
 898             !tcp_passive_fastopen(sk)) {
 899                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 900                         goto out_err;
 901         }
 902
 903         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 904
 905         mss_now = tcp_send_mss(sk, &size_goal, flags);
 906         copied = 0;
 907
 908         err = -EPIPE;
 909         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 910                 goto out_err;
 911
 912         while (size > 0) {
 913                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 914                 int copy, i;
 915                 bool can_coalesce;
 916
 917                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
 918 new_segment:
 919                         if (!sk_stream_memory_free(sk))
 920                                 goto wait_for_sndbuf;
 921
 922                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
 923                                                   skb_queue_empty(&sk->sk_write_queue));
 924                         if (!skb)
 925                                 goto wait_for_memory;
 926
 927                         skb_entail(sk, skb);
 928                         copy = size_goal;
 929                 }
 930
 931                 if (copy > size)
 932                         copy = size;
 933
 934                 i = skb_shinfo(skb)->nr_frags;
 935                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 936                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
 937                         tcp_mark_push(tp, skb);
 938                         goto new_segment;
 939                 }
 940                 if (!sk_wmem_schedule(sk, copy))
 941                         goto wait_for_memory;
 942
 943                 if (can_coalesce) {
 944                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
 945                 } else {
 946                         get_page(page);
 947                         skb_fill_page_desc(skb, i, page, offset, copy);
 948                 }
 949                 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
 950
 951                 skb->len += copy;
 952                 skb->data_len += copy;
 953                 skb->truesize += copy;
 954                 sk->sk_wmem_queued += copy;
 955                 sk_mem_charge(sk, copy);
 956                 skb->ip_summed = CHECKSUM_PARTIAL;
 957                 tp->write_seq += copy;
 958                 TCP_SKB_CB(skb)->end_seq += copy;
 959                 tcp_skb_pcount_set(skb, 0);
 960
 961                 if (!copied)
 962                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
 963
 964                 copied += copy;
 965                 offset += copy;
 966                 if (!(size -= copy)) {
 967                         tcp_tx_timestamp(sk, skb);
 968                         goto out;
 969                 }
 970
 971                 if (skb->len < size_goal || (flags & MSG_OOB))
 972                         continue;
 973
 974                 if (forced_push(tp)) {
 975                         tcp_mark_push(tp, skb);
 976                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
 977                 } else if (skb == tcp_send_head(sk))
 978                         tcp_push_one(sk, mss_now);
 979                 continue;
 980
 981 wait_for_sndbuf:
 982                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 983 wait_for_memory:
 984                 tcp_push(sk, flags & ~MSG_MORE, mss_now,
 985                          TCP_NAGLE_PUSH, size_goal);
 986
 987                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 988                         goto do_error;
 989
 990                 mss_now = tcp_send_mss(sk, &size_goal, flags);
 991         }
 992
 993 out:
 994         if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
 995                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
 996         return copied;
 997
 998 do_error:
 999         if (copied)
1000                 goto out;
1001 out_err:
1002         return sk_stream_error(sk, flags, err);
1003 }
1004
1005 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1006                  size_t size, int flags)
1007 {
1008         ssize_t res;
1009
1010         if (!(sk->sk_route_caps & NETIF_F_SG) ||
1011             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
1012                 return sock_no_sendpage(sk->sk_socket, page, offset, size,
1013                                         flags);
1014
1015         lock_sock(sk);
1016         res = do_tcp_sendpages(sk, page, offset, size, flags);
1017         release_sock(sk);
1018         return res;
1019 }
1020 EXPORT_SYMBOL(tcp_sendpage);
1021
1022 static inline int select_size(const struct sock *sk, bool sg)
1023 {
1024         const struct tcp_sock *tp = tcp_sk(sk);
1025         int tmp = tp->mss_cache;
1026
1027         if (sg) {
1028                 if (sk_can_gso(sk)) {
1029                         /* Small frames wont use a full page:
1030                          * Payload will immediately follow tcp header.
1031                          */
1032                         tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1033                 } else {
1034                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1035
1036                         if (tmp >= pgbreak &&
1037                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1038                                 tmp = pgbreak;
1039                 }
1040         }
1041
1042         return tmp;
1043 }
1044
1045 void tcp_free_fastopen_req(struct tcp_sock *tp)
1046 {
1047         if (tp->fastopen_req) {
1048                 kfree(tp->fastopen_req);
1049                 tp->fastopen_req = NULL;
1050         }
1051 }
1052
1053 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1054                                 int *copied, size_t size)
1055 {
1056         struct tcp_sock *tp = tcp_sk(sk);
1057         int err, flags;
1058
1059         if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1060                 return -EOPNOTSUPP;
1061         if (tp->fastopen_req)
1062                 return -EALREADY; /* Another Fast Open is in progress */
1063
1064         tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1065                                    sk->sk_allocation);
1066         if (unlikely(!tp->fastopen_req))
1067                 return -ENOBUFS;
1068         tp->fastopen_req->data = msg;
1069         tp->fastopen_req->size = size;
1070
1071         flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1072         err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1073                                     msg->msg_namelen, flags);
1074         *copied = tp->fastopen_req->copied;
1075         tcp_free_fastopen_req(tp);
1076         return err;
1077 }
1078
1079 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1080 {
1081         struct tcp_sock *tp = tcp_sk(sk);
1082         struct sk_buff *skb;
1083         int flags, err, copied = 0;
1084         int mss_now = 0, size_goal, copied_syn = 0;
1085         bool sg;
1086         long timeo;
1087
1088         lock_sock(sk);
1089
1090         flags = msg->msg_flags;
1091         if (flags & MSG_FASTOPEN) {
1092                 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1093                 if (err == -EINPROGRESS && copied_syn > 0)
1094                         goto out;
1095                 else if (err)
1096                         goto out_err;
1097         }
1098
1099         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1100
1101         /* Wait for a connection to finish. One exception is TCP Fast Open
1102          * (passive side) where data is allowed to be sent before a connection
1103          * is fully established.
1104          */
1105         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1106             !tcp_passive_fastopen(sk)) {
1107                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1108                         goto do_error;
1109         }
1110
1111         if (unlikely(tp->repair)) {
1112                 if (tp->repair_queue == TCP_RECV_QUEUE) {
1113                         copied = tcp_send_rcvq(sk, msg, size);
1114                         goto out_nopush;
1115                 }
1116
1117                 err = -EINVAL;
1118                 if (tp->repair_queue == TCP_NO_QUEUE)
1119                         goto out_err;
1120
1121                 /* 'common' sending to sendq */
1122         }
1123
1124         /* This should be in poll */
1125         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1126
1127         mss_now = tcp_send_mss(sk, &size_goal, flags);
1128
1129         /* Ok commence sending. */
1130         copied = 0;
1131
1132         err = -EPIPE;
1133         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1134                 goto out_err;
1135
1136         sg = !!(sk->sk_route_caps & NETIF_F_SG);
1137
1138         while (msg_data_left(msg)) {
1139                 int copy = 0;
1140                 int max = size_goal;
1141
1142                 skb = tcp_write_queue_tail(sk);
1143                 if (tcp_send_head(sk)) {
1144                         if (skb->ip_summed == CHECKSUM_NONE)
1145                                 max = mss_now;
1146                         copy = max - skb->len;
1147                 }
1148
1149                 if (copy <= 0) {
1150 new_segment:
1151                         /* Allocate new segment. If the interface is SG,
1152                          * allocate skb fitting to single page.
1153                          */
1154                         if (!sk_stream_memory_free(sk))
1155                                 goto wait_for_sndbuf;
1156
1157                         skb = sk_stream_alloc_skb(sk,
1158                                                   select_size(sk, sg),
1159                                                   sk->sk_allocation,
1160                                                   skb_queue_empty(&sk->sk_write_queue));
1161                         if (!skb)
1162                                 goto wait_for_memory;
1163
1164                         /*
1165                          * Check whether we can use HW checksum.
1166                          */
1167                         if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1168                                 skb->ip_summed = CHECKSUM_PARTIAL;
1169
1170                         skb_entail(sk, skb);
1171                         copy = size_goal;
1172                         max = size_goal;
1173
1174                         /* All packets are restored as if they have
1175                          * already been sent. skb_mstamp isn't set to
1176                          * avoid wrong rtt estimation.
1177                          */
1178                         if (tp->repair)
1179                                 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1180                 }
1181
1182                 /* Try to append data to the end of skb. */
1183                 if (copy > msg_data_left(msg))
1184                         copy = msg_data_left(msg);
1185
1186                 /* Where to copy to? */
1187                 if (skb_availroom(skb) > 0) {
1188                         /* We have some space in skb head. Superb! */
1189                         copy = min_t(int, copy, skb_availroom(skb));
1190                         err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1191                         if (err)
1192                                 goto do_fault;
1193                 } else {
1194                         bool merge = true;
1195                         int i = skb_shinfo(skb)->nr_frags;
1196                         struct page_frag *pfrag = sk_page_frag(sk);
1197
1198                         if (!sk_page_frag_refill(sk, pfrag))
1199                                 goto wait_for_memory;
1200
1201                         if (!skb_can_coalesce(skb, i, pfrag->page,
1202                                               pfrag->offset)) {
1203                                 if (i == MAX_SKB_FRAGS || !sg) {
1204                                         tcp_mark_push(tp, skb);
1205                                         goto new_segment;
1206                                 }
1207                                 merge = false;
1208                         }
1209
1210                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1211
1212                         if (!sk_wmem_schedule(sk, copy))
1213                                 goto wait_for_memory;
1214
1215                         err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1216                                                        pfrag->page,
1217                                                        pfrag->offset,
1218                                                        copy);
1219                         if (err)
1220                                 goto do_error;
1221
1222                         /* Update the skb. */
1223                         if (merge) {
1224                                 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1225                         } else {
1226                                 skb_fill_page_desc(skb, i, pfrag->page,
1227                                                    pfrag->offset, copy);
1228                                 get_page(pfrag->page);
1229                         }
1230                         pfrag->offset += copy;
1231                 }
1232
1233                 if (!copied)
1234                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1235
1236                 tp->write_seq += copy;
1237                 TCP_SKB_CB(skb)->end_seq += copy;
1238                 tcp_skb_pcount_set(skb, 0);
1239
1240                 copied += copy;
1241                 if (!msg_data_left(msg)) {
1242                         tcp_tx_timestamp(sk, skb);
1243                         goto out;
1244                 }
1245
1246                 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1247                         continue;
1248
1249                 if (forced_push(tp)) {
1250                         tcp_mark_push(tp, skb);
1251                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1252                 } else if (skb == tcp_send_head(sk))
1253                         tcp_push_one(sk, mss_now);
1254                 continue;
1255
1256 wait_for_sndbuf:
1257                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1258 wait_for_memory:
1259                 if (copied)
1260                         tcp_push(sk, flags & ~MSG_MORE, mss_now,
1261                                  TCP_NAGLE_PUSH, size_goal);
1262
1263                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1264                         goto do_error;
1265
1266                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1267         }
1268
1269 out:
1270         if (copied)
1271                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1272 out_nopush:
1273         release_sock(sk);
1274         return copied + copied_syn;
1275
1276 do_fault:
1277         if (!skb->len) {
1278                 tcp_unlink_write_queue(skb, sk);
1279                 /* It is the one place in all of TCP, except connection
1280                  * reset, where we can be unlinking the send_head.
1281                  */
1282                 tcp_check_send_head(sk, skb);
1283                 sk_wmem_free_skb(sk, skb);
1284         }
1285
1286 do_error:
1287         if (copied + copied_syn)
1288                 goto out;
1289 out_err:
1290         err = sk_stream_error(sk, flags, err);
1291         release_sock(sk);
1292         return err;
1293 }
1294 EXPORT_SYMBOL(tcp_sendmsg);
1295
1296 /*
1297  *      Handle reading urgent data. BSD has very simple semantics for
1298  *      this, no blocking and very strange errors 8)
1299  */
1300
1301 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1302 {
1303         struct tcp_sock *tp = tcp_sk(sk);
1304
1305         /* No URG data to read. */
1306         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1307             tp->urg_data == TCP_URG_READ)
1308                 return -EINVAL; /* Yes this is right ! */
1309
1310         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1311                 return -ENOTCONN;
1312
1313         if (tp->urg_data & TCP_URG_VALID) {
1314                 int err = 0;
1315                 char c = tp->urg_data;
1316
1317                 if (!(flags & MSG_PEEK))
1318                         tp->urg_data = TCP_URG_READ;
1319
1320                 /* Read urgent data. */
1321                 msg->msg_flags |= MSG_OOB;
1322
1323                 if (len > 0) {
1324                         if (!(flags & MSG_TRUNC))
1325                                 err = memcpy_to_msg(msg, &c, 1);
1326                         len = 1;
1327                 } else
1328                         msg->msg_flags |= MSG_TRUNC;
1329
1330                 return err ? -EFAULT : len;
1331         }
1332
1333         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1334                 return 0;
1335
1336         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1337          * the available implementations agree in this case:
1338          * this call should never block, independent of the
1339          * blocking state of the socket.
1340          * Mike <pall@rz.uni-karlsruhe.de>
1341          */
1342         return -EAGAIN;
1343 }
1344
1345 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1346 {
1347         struct sk_buff *skb;
1348         int copied = 0, err = 0;
1349
1350         /* XXX -- need to support SO_PEEK_OFF */
1351
1352         skb_queue_walk(&sk->sk_write_queue, skb) {
1353                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1354                 if (err)
1355                         break;
1356
1357                 copied += skb->len;
1358         }
1359
1360         return err ?: copied;
1361 }
1362
1363 /* Clean up the receive buffer for full frames taken by the user,
1364  * then send an ACK if necessary.  COPIED is the number of bytes
1365  * tcp_recvmsg has given to the user so far, it speeds up the
1366  * calculation of whether or not we must ACK for the sake of
1367  * a window update.
1368  */
1369 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1370 {
1371         struct tcp_sock *tp = tcp_sk(sk);
1372         bool time_to_ack = false;
1373
1374         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1375
1376         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1377              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1378              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1379
1380         if (inet_csk_ack_scheduled(sk)) {
1381                 const struct inet_connection_sock *icsk = inet_csk(sk);
1382                    /* Delayed ACKs frequently hit locked sockets during bulk
1383                     * receive. */
1384                 if (icsk->icsk_ack.blocked ||
1385                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1386                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1387                     /*
1388                      * If this read emptied read buffer, we send ACK, if
1389                      * connection is not bidirectional, user drained
1390                      * receive buffer and there was a small segment
1391                      * in queue.
1392                      */
1393                     (copied > 0 &&
1394                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1395                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1396                        !icsk->icsk_ack.pingpong)) &&
1397                       !atomic_read(&sk->sk_rmem_alloc)))
1398                         time_to_ack = true;
1399         }
1400
1401         /* We send an ACK if we can now advertise a non-zero window
1402          * which has been raised "significantly".
1403          *
1404          * Even if window raised up to infinity, do not send window open ACK
1405          * in states, where we will not receive more. It is useless.
1406          */
1407         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1408                 __u32 rcv_window_now = tcp_receive_window(tp);
1409
1410                 /* Optimize, __tcp_select_window() is not cheap. */
1411                 if (2*rcv_window_now <= tp->window_clamp) {
1412                         __u32 new_window = __tcp_select_window(sk);
1413
1414                         /* Send ACK now, if this read freed lots of space
1415                          * in our buffer. Certainly, new_window is new window.
1416                          * We can advertise it now, if it is not less than current one.
1417                          * "Lots" means "at least twice" here.
1418                          */
1419                         if (new_window && new_window >= 2 * rcv_window_now)
1420                                 time_to_ack = true;
1421                 }
1422         }
1423         if (time_to_ack)
1424                 tcp_send_ack(sk);
1425 }
1426
1427 static void tcp_prequeue_process(struct sock *sk)
1428 {
1429         struct sk_buff *skb;
1430         struct tcp_sock *tp = tcp_sk(sk);
1431
1432         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1433
1434         /* RX process wants to run with disabled BHs, though it is not
1435          * necessary */
1436         local_bh_disable();
1437         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1438                 sk_backlog_rcv(sk, skb);
1439         local_bh_enable();
1440
1441         /* Clear memory counter. */
1442         tp->ucopy.memory = 0;
1443 }
1444
1445 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1446 {
1447         struct sk_buff *skb;
1448         u32 offset;
1449
1450         while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1451                 offset = seq - TCP_SKB_CB(skb)->seq;
1452                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1453                         offset--;
1454                 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1455                         *off = offset;
1456                         return skb;
1457                 }
1458                 /* This looks weird, but this can happen if TCP collapsing
1459                  * splitted a fat GRO packet, while we released socket lock
1460                  * in skb_splice_bits()
1461                  */
1462                 sk_eat_skb(sk, skb);
1463         }
1464         return NULL;
1465 }
1466
1467 /*
1468  * This routine provides an alternative to tcp_recvmsg() for routines
1469  * that would like to handle copying from skbuffs directly in 'sendfile'
1470  * fashion.
1471  * Note:
1472  *      - It is assumed that the socket was locked by the caller.
1473  *      - The routine does not block.
1474  *      - At present, there is no support for reading OOB data
1475  *        or for 'peeking' the socket using this routine
1476  *        (although both would be easy to implement).
1477  */
1478 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1479                   sk_read_actor_t recv_actor)
1480 {
1481         struct sk_buff *skb;
1482         struct tcp_sock *tp = tcp_sk(sk);
1483         u32 seq = tp->copied_seq;
1484         u32 offset;
1485         int copied = 0;
1486
1487         if (sk->sk_state == TCP_LISTEN)
1488                 return -ENOTCONN;
1489         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1490                 if (offset < skb->len) {
1491                         int used;
1492                         size_t len;
1493
1494                         len = skb->len - offset;
1495                         /* Stop reading if we hit a patch of urgent data */
1496                         if (tp->urg_data) {
1497                                 u32 urg_offset = tp->urg_seq - seq;
1498                                 if (urg_offset < len)
1499                                         len = urg_offset;
1500                                 if (!len)
1501                                         break;
1502                         }
1503                         used = recv_actor(desc, skb, offset, len);
1504                         if (used <= 0) {
1505                                 if (!copied)
1506                                         copied = used;
1507                                 break;
1508                         } else if (used <= len) {
1509                                 seq += used;
1510                                 copied += used;
1511                                 offset += used;
1512                         }
1513                         /* If recv_actor drops the lock (e.g. TCP splice
1514                          * receive) the skb pointer might be invalid when
1515                          * getting here: tcp_collapse might have deleted it
1516                          * while aggregating skbs from the socket queue.
1517                          */
1518                         skb = tcp_recv_skb(sk, seq - 1, &offset);
1519                         if (!skb)
1520                                 break;
1521                         /* TCP coalescing might have appended data to the skb.
1522                          * Try to splice more frags
1523                          */
1524                         if (offset + 1 != skb->len)
1525                                 continue;
1526                 }
1527                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1528                         sk_eat_skb(sk, skb);
1529                         ++seq;
1530                         break;
1531                 }
1532                 sk_eat_skb(sk, skb);
1533                 if (!desc->count)
1534                         break;
1535                 tp->copied_seq = seq;
1536         }
1537         tp->copied_seq = seq;
1538
1539         tcp_rcv_space_adjust(sk);
1540
1541         /* Clean up data we have read: This will do ACK frames. */
1542         if (copied > 0) {
1543                 tcp_recv_skb(sk, seq, &offset);
1544                 tcp_cleanup_rbuf(sk, copied);
1545         }
1546         return copied;
1547 }
1548 EXPORT_SYMBOL(tcp_read_sock);
1549
1550 /*
1551  *      This routine copies from a sock struct into the user buffer.
1552  *
1553  *      Technical note: in 2.3 we work on _locked_ socket, so that
1554  *      tricks with *seq access order and skb->users are not required.
1555  *      Probably, code can be easily improved even more.
1556  */
1557
1558 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1559                 int flags, int *addr_len)
1560 {
1561         struct tcp_sock *tp = tcp_sk(sk);
1562         int copied = 0;
1563         u32 peek_seq;
1564         u32 *seq;
1565         unsigned long used;
1566         int err;
1567         int target;             /* Read at least this many bytes */
1568         long timeo;
1569         struct task_struct *user_recv = NULL;
1570         struct sk_buff *skb;
1571         u32 urg_hole = 0;
1572
1573         if (unlikely(flags & MSG_ERRQUEUE))
1574                 return inet_recv_error(sk, msg, len, addr_len);
1575
1576         if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1577             (sk->sk_state == TCP_ESTABLISHED))
1578                 sk_busy_loop(sk, nonblock);
1579
1580         lock_sock(sk);
1581
1582         err = -ENOTCONN;
1583         if (sk->sk_state == TCP_LISTEN)
1584                 goto out;
1585
1586         timeo = sock_rcvtimeo(sk, nonblock);
1587
1588         /* Urgent data needs to be handled specially. */
1589         if (flags & MSG_OOB)
1590                 goto recv_urg;
1591
1592         if (unlikely(tp->repair)) {
1593                 err = -EPERM;
1594                 if (!(flags & MSG_PEEK))
1595                         goto out;
1596
1597                 if (tp->repair_queue == TCP_SEND_QUEUE)
1598                         goto recv_sndq;
1599
1600                 err = -EINVAL;
1601                 if (tp->repair_queue == TCP_NO_QUEUE)
1602                         goto out;
1603
1604                 /* 'common' recv queue MSG_PEEK-ing */
1605         }
1606
1607         seq = &tp->copied_seq;
1608         if (flags & MSG_PEEK) {
1609                 peek_seq = tp->copied_seq;
1610                 seq = &peek_seq;
1611         }
1612
1613         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1614
1615         do {
1616                 u32 offset;
1617
1618                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1619                 if (tp->urg_data && tp->urg_seq == *seq) {
1620                         if (copied)
1621                                 break;
1622                         if (signal_pending(current)) {
1623                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1624                                 break;
1625                         }
1626                 }
1627
1628                 /* Next get a buffer. */
1629
1630                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1631                         /* Now that we have two receive queues this
1632                          * shouldn't happen.
1633                          */
1634                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1635                                  "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1636                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1637                                  flags))
1638                                 break;
1639
1640                         offset = *seq - TCP_SKB_CB(skb)->seq;
1641                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1642                                 offset--;
1643                         if (offset < skb->len)
1644                                 goto found_ok_skb;
1645                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1646                                 goto found_fin_ok;
1647                         WARN(!(flags & MSG_PEEK),
1648                              "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1649                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1650                 }
1651
1652                 /* Well, if we have backlog, try to process it now yet. */
1653
1654                 if (copied >= target && !sk->sk_backlog.tail)
1655                         break;
1656
1657                 if (copied) {
1658                         if (sk->sk_err ||
1659                             sk->sk_state == TCP_CLOSE ||
1660                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1661                             !timeo ||
1662                             signal_pending(current))
1663                                 break;
1664                 } else {
1665                         if (sock_flag(sk, SOCK_DONE))
1666                                 break;
1667
1668                         if (sk->sk_err) {
1669                                 copied = sock_error(sk);
1670                                 break;
1671                         }
1672
1673                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1674                                 break;
1675
1676                         if (sk->sk_state == TCP_CLOSE) {
1677                                 if (!sock_flag(sk, SOCK_DONE)) {
1678                                         /* This occurs when user tries to read
1679                                          * from never connected socket.
1680                                          */
1681                                         copied = -ENOTCONN;
1682                                         break;
1683                                 }
1684                                 break;
1685                         }
1686
1687                         if (!timeo) {
1688                                 copied = -EAGAIN;
1689                                 break;
1690                         }
1691
1692                         if (signal_pending(current)) {
1693                                 copied = sock_intr_errno(timeo);
1694                                 break;
1695                         }
1696                 }
1697
1698                 tcp_cleanup_rbuf(sk, copied);
1699
1700                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1701                         /* Install new reader */
1702                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1703                                 user_recv = current;
1704                                 tp->ucopy.task = user_recv;
1705                                 tp->ucopy.msg = msg;
1706                         }
1707
1708                         tp->ucopy.len = len;
1709
1710                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1711                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1712
1713                         /* Ugly... If prequeue is not empty, we have to
1714                          * process it before releasing socket, otherwise
1715                          * order will be broken at second iteration.
1716                          * More elegant solution is required!!!
1717                          *
1718                          * Look: we have the following (pseudo)queues:
1719                          *
1720                          * 1. packets in flight
1721                          * 2. backlog
1722                          * 3. prequeue
1723                          * 4. receive_queue
1724                          *
1725                          * Each queue can be processed only if the next ones
1726                          * are empty. At this point we have empty receive_queue.
1727                          * But prequeue _can_ be not empty after 2nd iteration,
1728                          * when we jumped to start of loop because backlog
1729                          * processing added something to receive_queue.
1730                          * We cannot release_sock(), because backlog contains
1731                          * packets arrived _after_ prequeued ones.
1732                          *
1733                          * Shortly, algorithm is clear --- to process all
1734                          * the queues in order. We could make it more directly,
1735                          * requeueing packets from backlog to prequeue, if
1736                          * is not empty. It is more elegant, but eats cycles,
1737                          * unfortunately.
1738                          */
1739                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1740                                 goto do_prequeue;
1741
1742                         /* __ Set realtime policy in scheduler __ */
1743                 }
1744
1745                 if (copied >= target) {
1746                         /* Do not sleep, just process backlog. */
1747                         release_sock(sk);
1748                         lock_sock(sk);
1749                 } else
1750                         sk_wait_data(sk, &timeo);
1751
1752                 if (user_recv) {
1753                         int chunk;
1754
1755                         /* __ Restore normal policy in scheduler __ */
1756
1757                         if ((chunk = len - tp->ucopy.len) != 0) {
1758                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1759                                 len -= chunk;
1760                                 copied += chunk;
1761                         }
1762
1763                         if (tp->rcv_nxt == tp->copied_seq &&
1764                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1765 do_prequeue:
1766                                 tcp_prequeue_process(sk);
1767
1768                                 if ((chunk = len - tp->ucopy.len) != 0) {
1769                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1770                                         len -= chunk;
1771                                         copied += chunk;
1772                                 }
1773                         }
1774                 }
1775                 if ((flags & MSG_PEEK) &&
1776                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1777                         net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1778                                             current->comm,
1779                                             task_pid_nr(current));
1780                         peek_seq = tp->copied_seq;
1781                 }
1782                 continue;
1783
1784         found_ok_skb:
1785                 /* Ok so how much can we use? */
1786                 used = skb->len - offset;
1787                 if (len < used)
1788                         used = len;
1789
1790                 /* Do we have urgent data here? */
1791                 if (tp->urg_data) {
1792                         u32 urg_offset = tp->urg_seq - *seq;
1793                         if (urg_offset < used) {
1794                                 if (!urg_offset) {
1795                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1796                                                 ++*seq;
1797                                                 urg_hole++;
1798                                                 offset++;
1799                                                 used--;
1800                                                 if (!used)
1801                                                         goto skip_copy;
1802                                         }
1803                                 } else
1804                                         used = urg_offset;
1805                         }
1806                 }
1807
1808                 if (!(flags & MSG_TRUNC)) {
1809                         err = skb_copy_datagram_msg(skb, offset, msg, used);
1810                         if (err) {
1811                                 /* Exception. Bailout! */
1812                                 if (!copied)
1813                                         copied = -EFAULT;
1814                                 break;
1815                         }
1816                 }
1817
1818                 *seq += used;
1819                 copied += used;
1820                 len -= used;
1821
1822                 tcp_rcv_space_adjust(sk);
1823
1824 skip_copy:
1825                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1826                         tp->urg_data = 0;
1827                         tcp_fast_path_check(sk);
1828                 }
1829                 if (used + offset < skb->len)
1830                         continue;
1831
1832                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1833                         goto found_fin_ok;
1834                 if (!(flags & MSG_PEEK))
1835                         sk_eat_skb(sk, skb);
1836                 continue;
1837
1838         found_fin_ok:
1839                 /* Process the FIN. */
1840                 ++*seq;
1841                 if (!(flags & MSG_PEEK))
1842                         sk_eat_skb(sk, skb);
1843                 break;
1844         } while (len > 0);
1845
1846         if (user_recv) {
1847                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1848                         int chunk;
1849
1850                         tp->ucopy.len = copied > 0 ? len : 0;
1851
1852                         tcp_prequeue_process(sk);
1853
1854                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1855                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1856                                 len -= chunk;
1857                                 copied += chunk;
1858                         }
1859                 }
1860
1861                 tp->ucopy.task = NULL;
1862                 tp->ucopy.len = 0;
1863         }
1864
1865         /* According to UNIX98, msg_name/msg_namelen are ignored
1866          * on connected socket. I was just happy when found this 8) --ANK
1867          */
1868
1869         /* Clean up data we have read: This will do ACK frames. */
1870         tcp_cleanup_rbuf(sk, copied);
1871
1872         release_sock(sk);
1873         return copied;
1874
1875 out:
1876         release_sock(sk);
1877         return err;
1878
1879 recv_urg:
1880         err = tcp_recv_urg(sk, msg, len, flags);
1881         goto out;
1882
1883 recv_sndq:
1884         err = tcp_peek_sndq(sk, msg, len);
1885         goto out;
1886 }
1887 EXPORT_SYMBOL(tcp_recvmsg);
1888
1889 void tcp_set_state(struct sock *sk, int state)
1890 {
1891         int oldstate = sk->sk_state;
1892
1893         switch (state) {
1894         case TCP_ESTABLISHED:
1895                 if (oldstate != TCP_ESTABLISHED)
1896                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1897                 break;
1898
1899         case TCP_CLOSE:
1900                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1901                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1902
1903                 sk->sk_prot->unhash(sk);
1904                 if (inet_csk(sk)->icsk_bind_hash &&
1905                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1906                         inet_put_port(sk);
1907                 /* fall through */
1908         default:
1909                 if (oldstate == TCP_ESTABLISHED)
1910                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1911         }
1912
1913         /* Change state AFTER socket is unhashed to avoid closed
1914          * socket sitting in hash tables.
1915          */
1916         sk->sk_state = state;
1917
1918 #ifdef STATE_TRACE
1919         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1920 #endif
1921 }
1922 EXPORT_SYMBOL_GPL(tcp_set_state);
1923
1924 /*
1925  *      State processing on a close. This implements the state shift for
1926  *      sending our FIN frame. Note that we only send a FIN for some
1927  *      states. A shutdown() may have already sent the FIN, or we may be
1928  *      closed.
1929  */
1930
1931 static const unsigned char new_state[16] = {
1932   /* current state:        new state:      action:      */
1933   [0 /* (Invalid) */]   = TCP_CLOSE,
1934   [TCP_ESTABLISHED]     = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1935   [TCP_SYN_SENT]        = TCP_CLOSE,
1936   [TCP_SYN_RECV]        = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1937   [TCP_FIN_WAIT1]       = TCP_FIN_WAIT1,
1938   [TCP_FIN_WAIT2]       = TCP_FIN_WAIT2,
1939   [TCP_TIME_WAIT]       = TCP_CLOSE,
1940   [TCP_CLOSE]           = TCP_CLOSE,
1941   [TCP_CLOSE_WAIT]      = TCP_LAST_ACK  | TCP_ACTION_FIN,
1942   [TCP_LAST_ACK]        = TCP_LAST_ACK,
1943   [TCP_LISTEN]          = TCP_CLOSE,
1944   [TCP_CLOSING]         = TCP_CLOSING,
1945   [TCP_NEW_SYN_RECV]    = TCP_CLOSE,    /* should not happen ! */
1946 };
1947
1948 static int tcp_close_state(struct sock *sk)
1949 {
1950         int next = (int)new_state[sk->sk_state];
1951         int ns = next & TCP_STATE_MASK;
1952
1953         tcp_set_state(sk, ns);
1954
1955         return next & TCP_ACTION_FIN;
1956 }
1957
1958 /*
1959  *      Shutdown the sending side of a connection. Much like close except
1960  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1961  */
1962
1963 void tcp_shutdown(struct sock *sk, int how)
1964 {
1965         /*      We need to grab some memory, and put together a FIN,
1966          *      and then put it into the queue to be sent.
1967          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1968          */
1969         if (!(how & SEND_SHUTDOWN))
1970                 return;
1971
1972         /* If we've already sent a FIN, or it's a closed state, skip this. */
1973         if ((1 << sk->sk_state) &
1974             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1975              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1976                 /* Clear out any half completed packets.  FIN if needed. */
1977                 if (tcp_close_state(sk))
1978                         tcp_send_fin(sk);
1979         }
1980 }
1981 EXPORT_SYMBOL(tcp_shutdown);
1982
1983 bool tcp_check_oom(struct sock *sk, int shift)
1984 {
1985         bool too_many_orphans, out_of_socket_memory;
1986
1987         too_many_orphans = tcp_too_many_orphans(sk, shift);
1988         out_of_socket_memory = tcp_out_of_memory(sk);
1989
1990         if (too_many_orphans)
1991                 net_info_ratelimited("too many orphaned sockets\n");
1992         if (out_of_socket_memory)
1993                 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
1994         return too_many_orphans || out_of_socket_memory;
1995 }
1996
1997 void tcp_close(struct sock *sk, long timeout)
1998 {
1999         struct sk_buff *skb;
2000         int data_was_unread = 0;
2001         int state;
2002
2003         lock_sock(sk);
2004         sk->sk_shutdown = SHUTDOWN_MASK;
2005
2006         if (sk->sk_state == TCP_LISTEN) {
2007                 tcp_set_state(sk, TCP_CLOSE);
2008
2009                 /* Special case. */
2010                 inet_csk_listen_stop(sk);
2011
2012                 goto adjudge_to_death;
2013         }
2014
2015         /*  We need to flush the recv. buffs.  We do this only on the
2016          *  descriptor close, not protocol-sourced closes, because the
2017          *  reader process may not have drained the data yet!
2018          */
2019         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2020                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2021
2022                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2023                         len--;
2024                 data_was_unread += len;
2025                 __kfree_skb(skb);
2026         }
2027
2028         sk_mem_reclaim(sk);
2029
2030         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2031         if (sk->sk_state == TCP_CLOSE)
2032                 goto adjudge_to_death;
2033
2034         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2035          * data was lost. To witness the awful effects of the old behavior of
2036          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2037          * GET in an FTP client, suspend the process, wait for the client to
2038          * advertise a zero window, then kill -9 the FTP client, wheee...
2039          * Note: timeout is always zero in such a case.
2040          */
2041         if (unlikely(tcp_sk(sk)->repair)) {
2042                 sk->sk_prot->disconnect(sk, 0);
2043         } else if (data_was_unread) {
2044                 /* Unread data was tossed, zap the connection. */
2045                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2046                 tcp_set_state(sk, TCP_CLOSE);
2047                 tcp_send_active_reset(sk, sk->sk_allocation);
2048         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2049                 /* Check zero linger _after_ checking for unread data. */
2050                 sk->sk_prot->disconnect(sk, 0);
2051                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2052         } else if (tcp_close_state(sk)) {
2053                 /* We FIN if the application ate all the data before
2054                  * zapping the connection.
2055                  */
2056
2057                 /* RED-PEN. Formally speaking, we have broken TCP state
2058                  * machine. State transitions:
2059                  *
2060                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2061                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2062                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2063                  *
2064                  * are legal only when FIN has been sent (i.e. in window),
2065                  * rather than queued out of window. Purists blame.
2066                  *
2067                  * F.e. "RFC state" is ESTABLISHED,
2068                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2069                  *
2070                  * The visible declinations are that sometimes
2071                  * we enter time-wait state, when it is not required really
2072                  * (harmless), do not send active resets, when they are
2073                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2074                  * they look as CLOSING or LAST_ACK for Linux)
2075                  * Probably, I missed some more holelets.
2076                  *                                              --ANK
2077                  * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2078                  * in a single packet! (May consider it later but will
2079                  * probably need API support or TCP_CORK SYN-ACK until
2080                  * data is written and socket is closed.)
2081                  */
2082                 tcp_send_fin(sk);
2083         }
2084
2085         sk_stream_wait_close(sk, timeout);
2086
2087 adjudge_to_death:
2088         state = sk->sk_state;
2089         sock_hold(sk);
2090         sock_orphan(sk);
2091
2092         /* It is the last release_sock in its life. It will remove backlog. */
2093         release_sock(sk);
2094
2095
2096         /* Now socket is owned by kernel and we acquire BH lock
2097            to finish close. No need to check for user refs.
2098          */
2099         local_bh_disable();
2100         bh_lock_sock(sk);
2101         WARN_ON(sock_owned_by_user(sk));
2102
2103         percpu_counter_inc(sk->sk_prot->orphan_count);
2104
2105         /* Have we already been destroyed by a softirq or backlog? */
2106         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2107                 goto out;
2108
2109         /*      This is a (useful) BSD violating of the RFC. There is a
2110          *      problem with TCP as specified in that the other end could
2111          *      keep a socket open forever with no application left this end.
2112          *      We use a 1 minute timeout (about the same as BSD) then kill
2113          *      our end. If they send after that then tough - BUT: long enough
2114          *      that we won't make the old 4*rto = almost no time - whoops
2115          *      reset mistake.
2116          *
2117          *      Nope, it was not mistake. It is really desired behaviour
2118          *      f.e. on http servers, when such sockets are useless, but
2119          *      consume significant resources. Let's do it with special
2120          *      linger2 option.                                 --ANK
2121          */
2122
2123         if (sk->sk_state == TCP_FIN_WAIT2) {
2124                 struct tcp_sock *tp = tcp_sk(sk);
2125                 if (tp->linger2 < 0) {
2126                         tcp_set_state(sk, TCP_CLOSE);
2127                         tcp_send_active_reset(sk, GFP_ATOMIC);
2128                         NET_INC_STATS_BH(sock_net(sk),
2129                                         LINUX_MIB_TCPABORTONLINGER);
2130                 } else {
2131                         const int tmo = tcp_fin_time(sk);
2132
2133                         if (tmo > TCP_TIMEWAIT_LEN) {
2134                                 inet_csk_reset_keepalive_timer(sk,
2135                                                 tmo - TCP_TIMEWAIT_LEN);
2136                         } else {
2137                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2138                                 goto out;
2139                         }
2140                 }
2141         }
2142         if (sk->sk_state != TCP_CLOSE) {
2143                 sk_mem_reclaim(sk);
2144                 if (tcp_check_oom(sk, 0)) {
2145                         tcp_set_state(sk, TCP_CLOSE);
2146                         tcp_send_active_reset(sk, GFP_ATOMIC);
2147                         NET_INC_STATS_BH(sock_net(sk),
2148                                         LINUX_MIB_TCPABORTONMEMORY);
2149                 }
2150         }
2151
2152         if (sk->sk_state == TCP_CLOSE) {
2153                 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2154                 /* We could get here with a non-NULL req if the socket is
2155                  * aborted (e.g., closed with unread data) before 3WHS
2156                  * finishes.
2157                  */
2158                 if (req)
2159                         reqsk_fastopen_remove(sk, req, false);
2160                 inet_csk_destroy_sock(sk);
2161         }
2162         /* Otherwise, socket is reprieved until protocol close. */
2163
2164 out:
2165         bh_unlock_sock(sk);
2166         local_bh_enable();
2167         sock_put(sk);
2168 }
2169 EXPORT_SYMBOL(tcp_close);
2170
2171 /* These states need RST on ABORT according to RFC793 */
2172
2173 static inline bool tcp_need_reset(int state)
2174 {
2175         return (1 << state) &
2176                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2177                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2178 }
2179
2180 int tcp_disconnect(struct sock *sk, int flags)
2181 {
2182         struct inet_sock *inet = inet_sk(sk);
2183         struct inet_connection_sock *icsk = inet_csk(sk);
2184         struct tcp_sock *tp = tcp_sk(sk);
2185         int err = 0;
2186         int old_state = sk->sk_state;
2187
2188         if (old_state != TCP_CLOSE)
2189                 tcp_set_state(sk, TCP_CLOSE);
2190
2191         /* ABORT function of RFC793 */
2192         if (old_state == TCP_LISTEN) {
2193                 inet_csk_listen_stop(sk);
2194         } else if (unlikely(tp->repair)) {
2195                 sk->sk_err = ECONNABORTED;
2196         } else if (tcp_need_reset(old_state) ||
2197                    (tp->snd_nxt != tp->write_seq &&
2198                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2199                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2200                  * states
2201                  */
2202                 tcp_send_active_reset(sk, gfp_any());
2203                 sk->sk_err = ECONNRESET;
2204         } else if (old_state == TCP_SYN_SENT)
2205                 sk->sk_err = ECONNRESET;
2206
2207         tcp_clear_xmit_timers(sk);
2208         __skb_queue_purge(&sk->sk_receive_queue);
2209         tcp_write_queue_purge(sk);
2210         __skb_queue_purge(&tp->out_of_order_queue);
2211
2212         inet->inet_dport = 0;
2213
2214         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2215                 inet_reset_saddr(sk);
2216
2217         sk->sk_shutdown = 0;
2218         sock_reset_flag(sk, SOCK_DONE);
2219         tp->srtt_us = 0;
2220         if ((tp->write_seq += tp->max_window + 2) == 0)
2221                 tp->write_seq = 1;
2222         icsk->icsk_backoff = 0;
2223         tp->snd_cwnd = 2;
2224         icsk->icsk_probes_out = 0;
2225         tp->packets_out = 0;
2226         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2227         tp->snd_cwnd_cnt = 0;
2228         tp->window_clamp = 0;
2229         tcp_set_ca_state(sk, TCP_CA_Open);
2230         tcp_clear_retrans(tp);
2231         inet_csk_delack_init(sk);
2232         tcp_init_send_head(sk);
2233         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2234         __sk_dst_reset(sk);
2235
2236         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2237
2238         sk->sk_error_report(sk);
2239         return err;
2240 }
2241 EXPORT_SYMBOL(tcp_disconnect);
2242
2243 void tcp_sock_destruct(struct sock *sk)
2244 {
2245         inet_sock_destruct(sk);
2246
2247         kfree(inet_csk(sk)->icsk_accept_queue.fastopenq);
2248 }
2249
2250 static inline bool tcp_can_repair_sock(const struct sock *sk)
2251 {
2252         return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2253                 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2254 }
2255
2256 static int tcp_repair_options_est(struct tcp_sock *tp,
2257                 struct tcp_repair_opt __user *optbuf, unsigned int len)
2258 {
2259         struct tcp_repair_opt opt;
2260
2261         while (len >= sizeof(opt)) {
2262                 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2263                         return -EFAULT;
2264
2265                 optbuf++;
2266                 len -= sizeof(opt);
2267
2268                 switch (opt.opt_code) {
2269                 case TCPOPT_MSS:
2270                         tp->rx_opt.mss_clamp = opt.opt_val;
2271                         break;
2272                 case TCPOPT_WINDOW:
2273                         {
2274                                 u16 snd_wscale = opt.opt_val & 0xFFFF;
2275                                 u16 rcv_wscale = opt.opt_val >> 16;
2276
2277                                 if (snd_wscale > 14 || rcv_wscale > 14)
2278                                         return -EFBIG;
2279
2280                                 tp->rx_opt.snd_wscale = snd_wscale;
2281                                 tp->rx_opt.rcv_wscale = rcv_wscale;
2282                                 tp->rx_opt.wscale_ok = 1;
2283                         }
2284                         break;
2285                 case TCPOPT_SACK_PERM:
2286                         if (opt.opt_val != 0)
2287                                 return -EINVAL;
2288
2289                         tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2290                         if (sysctl_tcp_fack)
2291                                 tcp_enable_fack(tp);
2292                         break;
2293                 case TCPOPT_TIMESTAMP:
2294                         if (opt.opt_val != 0)
2295                                 return -EINVAL;
2296
2297                         tp->rx_opt.tstamp_ok = 1;
2298                         break;
2299                 }
2300         }
2301
2302         return 0;
2303 }
2304
2305 /*
2306  *      Socket option code for TCP.
2307  */
2308 static int do_tcp_setsockopt(struct sock *sk, int level,
2309                 int optname, char __user *optval, unsigned int optlen)
2310 {
2311         struct tcp_sock *tp = tcp_sk(sk);
2312         struct inet_connection_sock *icsk = inet_csk(sk);
2313         int val;
2314         int err = 0;
2315
2316         /* These are data/string values, all the others are ints */
2317         switch (optname) {
2318         case TCP_CONGESTION: {
2319                 char name[TCP_CA_NAME_MAX];
2320
2321                 if (optlen < 1)
2322                         return -EINVAL;
2323
2324                 val = strncpy_from_user(name, optval,
2325                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2326                 if (val < 0)
2327                         return -EFAULT;
2328                 name[val] = 0;
2329
2330                 lock_sock(sk);
2331                 err = tcp_set_congestion_control(sk, name);
2332                 release_sock(sk);
2333                 return err;
2334         }
2335         default:
2336                 /* fallthru */
2337                 break;
2338         }
2339
2340         if (optlen < sizeof(int))
2341                 return -EINVAL;
2342
2343         if (get_user(val, (int __user *)optval))
2344                 return -EFAULT;
2345
2346         lock_sock(sk);
2347
2348         switch (optname) {
2349         case TCP_MAXSEG:
2350                 /* Values greater than interface MTU won't take effect. However
2351                  * at the point when this call is done we typically don't yet
2352                  * know which interface is going to be used */
2353                 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2354                         err = -EINVAL;
2355                         break;
2356                 }
2357                 tp->rx_opt.user_mss = val;
2358                 break;
2359
2360         case TCP_NODELAY:
2361                 if (val) {
2362                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2363                          * this option on corked socket is remembered, but
2364                          * it is not activated until cork is cleared.
2365                          *
2366                          * However, when TCP_NODELAY is set we make
2367                          * an explicit push, which overrides even TCP_CORK
2368                          * for currently queued segments.
2369                          */
2370                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2371                         tcp_push_pending_frames(sk);
2372                 } else {
2373                         tp->nonagle &= ~TCP_NAGLE_OFF;
2374                 }
2375                 break;
2376
2377         case TCP_THIN_LINEAR_TIMEOUTS:
2378                 if (val < 0 || val > 1)
2379                         err = -EINVAL;
2380                 else
2381                         tp->thin_lto = val;
2382                 break;
2383
2384         case TCP_THIN_DUPACK:
2385                 if (val < 0 || val > 1)
2386                         err = -EINVAL;
2387                 else {
2388                         tp->thin_dupack = val;
2389                         if (tp->thin_dupack)
2390                                 tcp_disable_early_retrans(tp);
2391                 }
2392                 break;
2393
2394         case TCP_REPAIR:
2395                 if (!tcp_can_repair_sock(sk))
2396                         err = -EPERM;
2397                 else if (val == 1) {
2398                         tp->repair = 1;
2399                         sk->sk_reuse = SK_FORCE_REUSE;
2400                         tp->repair_queue = TCP_NO_QUEUE;
2401                 } else if (val == 0) {
2402                         tp->repair = 0;
2403                         sk->sk_reuse = SK_NO_REUSE;
2404                         tcp_send_window_probe(sk);
2405                 } else
2406                         err = -EINVAL;
2407
2408                 break;
2409
2410         case TCP_REPAIR_QUEUE:
2411                 if (!tp->repair)
2412                         err = -EPERM;
2413                 else if (val < TCP_QUEUES_NR)
2414                         tp->repair_queue = val;
2415                 else
2416                         err = -EINVAL;
2417                 break;
2418
2419         case TCP_QUEUE_SEQ:
2420                 if (sk->sk_state != TCP_CLOSE)
2421                         err = -EPERM;
2422                 else if (tp->repair_queue == TCP_SEND_QUEUE)
2423                         tp->write_seq = val;
2424                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2425                         tp->rcv_nxt = val;
2426                 else
2427                         err = -EINVAL;
2428                 break;
2429
2430         case TCP_REPAIR_OPTIONS:
2431                 if (!tp->repair)
2432                         err = -EINVAL;
2433                 else if (sk->sk_state == TCP_ESTABLISHED)
2434                         err = tcp_repair_options_est(tp,
2435                                         (struct tcp_repair_opt __user *)optval,
2436                                         optlen);
2437                 else
2438                         err = -EPERM;
2439                 break;
2440
2441         case TCP_CORK:
2442                 /* When set indicates to always queue non-full frames.
2443                  * Later the user clears this option and we transmit
2444                  * any pending partial frames in the queue.  This is
2445                  * meant to be used alongside sendfile() to get properly
2446                  * filled frames when the user (for example) must write
2447                  * out headers with a write() call first and then use
2448                  * sendfile to send out the data parts.
2449                  *
2450                  * TCP_CORK can be set together with TCP_NODELAY and it is
2451                  * stronger than TCP_NODELAY.
2452                  */
2453                 if (val) {
2454                         tp->nonagle |= TCP_NAGLE_CORK;
2455                 } else {
2456                         tp->nonagle &= ~TCP_NAGLE_CORK;
2457                         if (tp->nonagle&TCP_NAGLE_OFF)
2458                                 tp->nonagle |= TCP_NAGLE_PUSH;
2459                         tcp_push_pending_frames(sk);
2460                 }
2461                 break;
2462
2463         case TCP_KEEPIDLE:
2464                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2465                         err = -EINVAL;
2466                 else {
2467                         tp->keepalive_time = val * HZ;
2468                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2469                             !((1 << sk->sk_state) &
2470                               (TCPF_CLOSE | TCPF_LISTEN))) {
2471                                 u32 elapsed = keepalive_time_elapsed(tp);
2472                                 if (tp->keepalive_time > elapsed)
2473                                         elapsed = tp->keepalive_time - elapsed;
2474                                 else
2475                                         elapsed = 0;
2476                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2477                         }
2478                 }
2479                 break;
2480         case TCP_KEEPINTVL:
2481                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2482                         err = -EINVAL;
2483                 else
2484                         tp->keepalive_intvl = val * HZ;
2485                 break;
2486         case TCP_KEEPCNT:
2487                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2488                         err = -EINVAL;
2489                 else
2490                         tp->keepalive_probes = val;
2491                 break;
2492         case TCP_SYNCNT:
2493                 if (val < 1 || val > MAX_TCP_SYNCNT)
2494                         err = -EINVAL;
2495                 else
2496                         icsk->icsk_syn_retries = val;
2497                 break;
2498
2499         case TCP_SAVE_SYN:
2500                 if (val < 0 || val > 1)
2501                         err = -EINVAL;
2502                 else
2503                         tp->save_syn = val;
2504                 break;
2505
2506         case TCP_LINGER2:
2507                 if (val < 0)
2508                         tp->linger2 = -1;
2509                 else if (val > sysctl_tcp_fin_timeout / HZ)
2510                         tp->linger2 = 0;
2511                 else
2512                         tp->linger2 = val * HZ;
2513                 break;
2514
2515         case TCP_DEFER_ACCEPT:
2516                 /* Translate value in seconds to number of retransmits */
2517                 icsk->icsk_accept_queue.rskq_defer_accept =
2518                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2519                                         TCP_RTO_MAX / HZ);
2520                 break;
2521
2522         case TCP_WINDOW_CLAMP:
2523                 if (!val) {
2524                         if (sk->sk_state != TCP_CLOSE) {
2525                                 err = -EINVAL;
2526                                 break;
2527                         }
2528                         tp->window_clamp = 0;
2529                 } else
2530                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2531                                                 SOCK_MIN_RCVBUF / 2 : val;
2532                 break;
2533
2534         case TCP_QUICKACK:
2535                 if (!val) {
2536                         icsk->icsk_ack.pingpong = 1;
2537                 } else {
2538                         icsk->icsk_ack.pingpong = 0;
2539                         if ((1 << sk->sk_state) &
2540                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2541                             inet_csk_ack_scheduled(sk)) {
2542                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2543                                 tcp_cleanup_rbuf(sk, 1);
2544                                 if (!(val & 1))
2545                                         icsk->icsk_ack.pingpong = 1;
2546                         }
2547                 }
2548                 break;
2549
2550 #ifdef CONFIG_TCP_MD5SIG
2551         case TCP_MD5SIG:
2552                 /* Read the IP->Key mappings from userspace */
2553                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2554                 break;
2555 #endif
2556         case TCP_USER_TIMEOUT:
2557                 /* Cap the max time in ms TCP will retry or probe the window
2558                  * before giving up and aborting (ETIMEDOUT) a connection.
2559                  */
2560                 if (val < 0)
2561                         err = -EINVAL;
2562                 else
2563                         icsk->icsk_user_timeout = msecs_to_jiffies(val);
2564                 break;
2565
2566         case TCP_FASTOPEN:
2567                 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2568                     TCPF_LISTEN)))
2569                         err = fastopen_init_queue(sk, val);
2570                 else
2571                         err = -EINVAL;
2572                 break;
2573         case TCP_TIMESTAMP:
2574                 if (!tp->repair)
2575                         err = -EPERM;
2576                 else
2577                         tp->tsoffset = val - tcp_time_stamp;
2578                 break;
2579         case TCP_NOTSENT_LOWAT:
2580                 tp->notsent_lowat = val;
2581                 sk->sk_write_space(sk);
2582                 break;
2583         default:
2584                 err = -ENOPROTOOPT;
2585                 break;
2586         }
2587
2588         release_sock(sk);
2589         return err;
2590 }
2591
2592 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2593                    unsigned int optlen)
2594 {
2595         const struct inet_connection_sock *icsk = inet_csk(sk);
2596
2597         if (level != SOL_TCP)
2598                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2599                                                      optval, optlen);
2600         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2601 }
2602 EXPORT_SYMBOL(tcp_setsockopt);
2603
2604 #ifdef CONFIG_COMPAT
2605 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2606                           char __user *optval, unsigned int optlen)
2607 {
2608         if (level != SOL_TCP)
2609                 return inet_csk_compat_setsockopt(sk, level, optname,
2610                                                   optval, optlen);
2611         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2612 }
2613 EXPORT_SYMBOL(compat_tcp_setsockopt);
2614 #endif
2615
2616 /* Return information about state of tcp endpoint in API format. */
2617 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2618 {
2619         const struct tcp_sock *tp = tcp_sk(sk);
2620         const struct inet_connection_sock *icsk = inet_csk(sk);
2621         u32 now = tcp_time_stamp;
2622         u32 rate;
2623
2624         memset(info, 0, sizeof(*info));
2625
2626         info->tcpi_state = sk->sk_state;
2627         info->tcpi_ca_state = icsk->icsk_ca_state;
2628         info->tcpi_retransmits = icsk->icsk_retransmits;
2629         info->tcpi_probes = icsk->icsk_probes_out;
2630         info->tcpi_backoff = icsk->icsk_backoff;
2631
2632         if (tp->rx_opt.tstamp_ok)
2633                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2634         if (tcp_is_sack(tp))
2635                 info->tcpi_options |= TCPI_OPT_SACK;
2636         if (tp->rx_opt.wscale_ok) {
2637                 info->tcpi_options |= TCPI_OPT_WSCALE;
2638                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2639                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2640         }
2641
2642         if (tp->ecn_flags & TCP_ECN_OK)
2643                 info->tcpi_options |= TCPI_OPT_ECN;
2644         if (tp->ecn_flags & TCP_ECN_SEEN)
2645                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2646         if (tp->syn_data_acked)
2647                 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2648
2649         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2650         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2651         info->tcpi_snd_mss = tp->mss_cache;
2652         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2653
2654         if (sk->sk_state == TCP_LISTEN) {
2655                 info->tcpi_unacked = sk->sk_ack_backlog;
2656                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2657         } else {
2658                 info->tcpi_unacked = tp->packets_out;
2659                 info->tcpi_sacked = tp->sacked_out;
2660         }
2661         info->tcpi_lost = tp->lost_out;
2662         info->tcpi_retrans = tp->retrans_out;
2663         info->tcpi_fackets = tp->fackets_out;
2664
2665         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2666         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2667         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2668
2669         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2670         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2671         info->tcpi_rtt = tp->srtt_us >> 3;
2672         info->tcpi_rttvar = tp->mdev_us >> 2;
2673         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2674         info->tcpi_snd_cwnd = tp->snd_cwnd;
2675         info->tcpi_advmss = tp->advmss;
2676         info->tcpi_reordering = tp->reordering;
2677
2678         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2679         info->tcpi_rcv_space = tp->rcvq_space.space;
2680
2681         info->tcpi_total_retrans = tp->total_retrans;
2682
2683         rate = READ_ONCE(sk->sk_pacing_rate);
2684         info->tcpi_pacing_rate = rate != ~0U ? rate : ~0ULL;
2685
2686         rate = READ_ONCE(sk->sk_max_pacing_rate);
2687         info->tcpi_max_pacing_rate = rate != ~0U ? rate : ~0ULL;
2688
2689         spin_lock_bh(&sk->sk_lock.slock);
2690         info->tcpi_bytes_acked = tp->bytes_acked;
2691         info->tcpi_bytes_received = tp->bytes_received;
2692         spin_unlock_bh(&sk->sk_lock.slock);
2693 }
2694 EXPORT_SYMBOL_GPL(tcp_get_info);
2695
2696 static int do_tcp_getsockopt(struct sock *sk, int level,
2697                 int optname, char __user *optval, int __user *optlen)
2698 {
2699         struct inet_connection_sock *icsk = inet_csk(sk);
2700         struct tcp_sock *tp = tcp_sk(sk);
2701         int val, len;
2702
2703         if (get_user(len, optlen))
2704                 return -EFAULT;
2705
2706         len = min_t(unsigned int, len, sizeof(int));
2707
2708         if (len < 0)
2709                 return -EINVAL;
2710
2711         switch (optname) {
2712         case TCP_MAXSEG:
2713                 val = tp->mss_cache;
2714                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2715                         val = tp->rx_opt.user_mss;
2716                 if (tp->repair)
2717                         val = tp->rx_opt.mss_clamp;
2718                 break;
2719         case TCP_NODELAY:
2720                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2721                 break;
2722         case TCP_CORK:
2723                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2724                 break;
2725         case TCP_KEEPIDLE:
2726                 val = keepalive_time_when(tp) / HZ;
2727                 break;
2728         case TCP_KEEPINTVL:
2729                 val = keepalive_intvl_when(tp) / HZ;
2730                 break;
2731         case TCP_KEEPCNT:
2732                 val = keepalive_probes(tp);
2733                 break;
2734         case TCP_SYNCNT:
2735                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2736                 break;
2737         case TCP_LINGER2:
2738                 val = tp->linger2;
2739                 if (val >= 0)
2740                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2741                 break;
2742         case TCP_DEFER_ACCEPT:
2743                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2744                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2745                 break;
2746         case TCP_WINDOW_CLAMP:
2747                 val = tp->window_clamp;
2748                 break;
2749         case TCP_INFO: {
2750                 struct tcp_info info;
2751
2752                 if (get_user(len, optlen))
2753                         return -EFAULT;
2754
2755                 tcp_get_info(sk, &info);
2756
2757                 len = min_t(unsigned int, len, sizeof(info));
2758                 if (put_user(len, optlen))
2759                         return -EFAULT;
2760                 if (copy_to_user(optval, &info, len))
2761                         return -EFAULT;
2762                 return 0;
2763         }
2764         case TCP_CC_INFO: {
2765                 const struct tcp_congestion_ops *ca_ops;
2766                 union tcp_cc_info info;
2767                 size_t sz = 0;
2768                 int attr;
2769
2770                 if (get_user(len, optlen))
2771                         return -EFAULT;
2772
2773                 ca_ops = icsk->icsk_ca_ops;
2774                 if (ca_ops && ca_ops->get_info)
2775                         sz = ca_ops->get_info(sk, ~0U, &attr, &info);
2776
2777                 len = min_t(unsigned int, len, sz);
2778                 if (put_user(len, optlen))
2779                         return -EFAULT;
2780                 if (copy_to_user(optval, &info, len))
2781                         return -EFAULT;
2782                 return 0;
2783         }
2784         case TCP_QUICKACK:
2785                 val = !icsk->icsk_ack.pingpong;
2786                 break;
2787
2788         case TCP_CONGESTION:
2789                 if (get_user(len, optlen))
2790                         return -EFAULT;
2791                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2792                 if (put_user(len, optlen))
2793                         return -EFAULT;
2794                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2795                         return -EFAULT;
2796                 return 0;
2797
2798         case TCP_THIN_LINEAR_TIMEOUTS:
2799                 val = tp->thin_lto;
2800                 break;
2801         case TCP_THIN_DUPACK:
2802                 val = tp->thin_dupack;
2803                 break;
2804
2805         case TCP_REPAIR:
2806                 val = tp->repair;
2807                 break;
2808
2809         case TCP_REPAIR_QUEUE:
2810                 if (tp->repair)
2811                         val = tp->repair_queue;
2812                 else
2813                         return -EINVAL;
2814                 break;
2815
2816         case TCP_QUEUE_SEQ:
2817                 if (tp->repair_queue == TCP_SEND_QUEUE)
2818                         val = tp->write_seq;
2819                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2820                         val = tp->rcv_nxt;
2821                 else
2822                         return -EINVAL;
2823                 break;
2824
2825         case TCP_USER_TIMEOUT:
2826                 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2827                 break;
2828
2829         case TCP_FASTOPEN:
2830                 if (icsk->icsk_accept_queue.fastopenq)
2831                         val = icsk->icsk_accept_queue.fastopenq->max_qlen;
2832                 else
2833                         val = 0;
2834                 break;
2835
2836         case TCP_TIMESTAMP:
2837                 val = tcp_time_stamp + tp->tsoffset;
2838                 break;
2839         case TCP_NOTSENT_LOWAT:
2840                 val = tp->notsent_lowat;
2841                 break;
2842         case TCP_SAVE_SYN:
2843                 val = tp->save_syn;
2844                 break;
2845         case TCP_SAVED_SYN: {
2846                 if (get_user(len, optlen))
2847                         return -EFAULT;
2848
2849                 lock_sock(sk);
2850                 if (tp->saved_syn) {
2851                         if (len < tp->saved_syn[0]) {
2852                                 if (put_user(tp->saved_syn[0], optlen)) {
2853                                         release_sock(sk);
2854                                         return -EFAULT;
2855                                 }
2856                                 release_sock(sk);
2857                                 return -EINVAL;
2858                         }
2859                         len = tp->saved_syn[0];
2860                         if (put_user(len, optlen)) {
2861                                 release_sock(sk);
2862                                 return -EFAULT;
2863                         }
2864                         if (copy_to_user(optval, tp->saved_syn + 1, len)) {
2865                                 release_sock(sk);
2866                                 return -EFAULT;
2867                         }
2868                         tcp_saved_syn_free(tp);
2869                         release_sock(sk);
2870                 } else {
2871                         release_sock(sk);
2872                         len = 0;
2873                         if (put_user(len, optlen))
2874                                 return -EFAULT;
2875                 }
2876                 return 0;
2877         }
2878         default:
2879                 return -ENOPROTOOPT;
2880         }
2881
2882         if (put_user(len, optlen))
2883                 return -EFAULT;
2884         if (copy_to_user(optval, &val, len))
2885                 return -EFAULT;
2886         return 0;
2887 }
2888
2889 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2890                    int __user *optlen)
2891 {
2892         struct inet_connection_sock *icsk = inet_csk(sk);
2893
2894         if (level != SOL_TCP)
2895                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2896                                                      optval, optlen);
2897         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2898 }
2899 EXPORT_SYMBOL(tcp_getsockopt);
2900
2901 #ifdef CONFIG_COMPAT
2902 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2903                           char __user *optval, int __user *optlen)
2904 {
2905         if (level != SOL_TCP)
2906                 return inet_csk_compat_getsockopt(sk, level, optname,
2907                                                   optval, optlen);
2908         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2909 }
2910 EXPORT_SYMBOL(compat_tcp_getsockopt);
2911 #endif
2912
2913 #ifdef CONFIG_TCP_MD5SIG
2914 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
2915 static DEFINE_MUTEX(tcp_md5sig_mutex);
2916 static bool tcp_md5sig_pool_populated = false;
2917
2918 static void __tcp_alloc_md5sig_pool(void)
2919 {
2920         int cpu;
2921
2922         for_each_possible_cpu(cpu) {
2923                 if (!per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm) {
2924                         struct crypto_hash *hash;
2925
2926                         hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2927                         if (IS_ERR_OR_NULL(hash))
2928                                 return;
2929                         per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm = hash;
2930                 }
2931         }
2932         /* before setting tcp_md5sig_pool_populated, we must commit all writes
2933          * to memory. See smp_rmb() in tcp_get_md5sig_pool()
2934          */
2935         smp_wmb();
2936         tcp_md5sig_pool_populated = true;
2937 }
2938
2939 bool tcp_alloc_md5sig_pool(void)
2940 {
2941         if (unlikely(!tcp_md5sig_pool_populated)) {
2942                 mutex_lock(&tcp_md5sig_mutex);
2943
2944                 if (!tcp_md5sig_pool_populated)
2945                         __tcp_alloc_md5sig_pool();
2946
2947                 mutex_unlock(&tcp_md5sig_mutex);
2948         }
2949         return tcp_md5sig_pool_populated;
2950 }
2951 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2952
2953
2954 /**
2955  *      tcp_get_md5sig_pool - get md5sig_pool for this user
2956  *
2957  *      We use percpu structure, so if we succeed, we exit with preemption
2958  *      and BH disabled, to make sure another thread or softirq handling
2959  *      wont try to get same context.
2960  */
2961 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2962 {
2963         local_bh_disable();
2964
2965         if (tcp_md5sig_pool_populated) {
2966                 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
2967                 smp_rmb();
2968                 return this_cpu_ptr(&tcp_md5sig_pool);
2969         }
2970         local_bh_enable();
2971         return NULL;
2972 }
2973 EXPORT_SYMBOL(tcp_get_md5sig_pool);
2974
2975 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2976                         const struct tcphdr *th)
2977 {
2978         struct scatterlist sg;
2979         struct tcphdr hdr;
2980         int err;
2981
2982         /* We are not allowed to change tcphdr, make a local copy */
2983         memcpy(&hdr, th, sizeof(hdr));
2984         hdr.check = 0;
2985
2986         /* options aren't included in the hash */
2987         sg_init_one(&sg, &hdr, sizeof(hdr));
2988         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
2989         return err;
2990 }
2991 EXPORT_SYMBOL(tcp_md5_hash_header);
2992
2993 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2994                           const struct sk_buff *skb, unsigned int header_len)
2995 {
2996         struct scatterlist sg;
2997         const struct tcphdr *tp = tcp_hdr(skb);
2998         struct hash_desc *desc = &hp->md5_desc;
2999         unsigned int i;
3000         const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3001                                            skb_headlen(skb) - header_len : 0;
3002         const struct skb_shared_info *shi = skb_shinfo(skb);
3003         struct sk_buff *frag_iter;
3004
3005         sg_init_table(&sg, 1);
3006
3007         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3008         if (crypto_hash_update(desc, &sg, head_data_len))
3009                 return 1;
3010
3011         for (i = 0; i < shi->nr_frags; ++i) {
3012                 const struct skb_frag_struct *f = &shi->frags[i];
3013                 unsigned int offset = f->page_offset;
3014                 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3015
3016                 sg_set_page(&sg, page, skb_frag_size(f),
3017                             offset_in_page(offset));
3018                 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3019                         return 1;
3020         }
3021
3022         skb_walk_frags(skb, frag_iter)
3023                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3024                         return 1;
3025
3026         return 0;
3027 }
3028 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3029
3030 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3031 {
3032         struct scatterlist sg;
3033
3034         sg_init_one(&sg, key->key, key->keylen);
3035         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3036 }
3037 EXPORT_SYMBOL(tcp_md5_hash_key);
3038
3039 #endif
3040
3041 void tcp_done(struct sock *sk)
3042 {
3043         struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3044
3045         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3046                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3047
3048         tcp_set_state(sk, TCP_CLOSE);
3049         tcp_clear_xmit_timers(sk);
3050         if (req)
3051                 reqsk_fastopen_remove(sk, req, false);
3052
3053         sk->sk_shutdown = SHUTDOWN_MASK;
3054
3055         if (!sock_flag(sk, SOCK_DEAD))
3056                 sk->sk_state_change(sk);
3057         else
3058                 inet_csk_destroy_sock(sk);
3059 }
3060 EXPORT_SYMBOL_GPL(tcp_done);
3061
3062 extern struct tcp_congestion_ops tcp_reno;
3063
3064 static __initdata unsigned long thash_entries;
3065 static int __init set_thash_entries(char *str)
3066 {
3067         ssize_t ret;
3068
3069         if (!str)
3070                 return 0;
3071
3072         ret = kstrtoul(str, 0, &thash_entries);
3073         if (ret)
3074                 return 0;
3075
3076         return 1;
3077 }
3078 __setup("thash_entries=", set_thash_entries);
3079
3080 static void __init tcp_init_mem(void)
3081 {
3082         unsigned long limit = nr_free_buffer_pages() / 16;
3083
3084         limit = max(limit, 128UL);
3085         sysctl_tcp_mem[0] = limit / 4 * 3;              /* 4.68 % */
3086         sysctl_tcp_mem[1] = limit;                      /* 6.25 % */
3087         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;      /* 9.37 % */
3088 }
3089
3090 void __init tcp_init(void)
3091 {
3092         unsigned long limit;
3093         int max_rshare, max_wshare, cnt;
3094         unsigned int i;
3095
3096         sock_skb_cb_check_size(sizeof(struct tcp_skb_cb));
3097
3098         percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3099         percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3100         tcp_hashinfo.bind_bucket_cachep =
3101                 kmem_cache_create("tcp_bind_bucket",
3102                                   sizeof(struct inet_bind_bucket), 0,
3103                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3104
3105         /* Size and allocate the main established and bind bucket
3106          * hash tables.
3107          *
3108          * The methodology is similar to that of the buffer cache.
3109          */
3110         tcp_hashinfo.ehash =
3111                 alloc_large_system_hash("TCP established",
3112                                         sizeof(struct inet_ehash_bucket),
3113                                         thash_entries,
3114                                         17, /* one slot per 128 KB of memory */
3115                                         0,
3116                                         NULL,
3117                                         &tcp_hashinfo.ehash_mask,
3118                                         0,
3119                                         thash_entries ? 0 : 512 * 1024);
3120         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3121                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3122
3123         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3124                 panic("TCP: failed to alloc ehash_locks");
3125         tcp_hashinfo.bhash =
3126                 alloc_large_system_hash("TCP bind",
3127                                         sizeof(struct inet_bind_hashbucket),
3128                                         tcp_hashinfo.ehash_mask + 1,
3129                                         17, /* one slot per 128 KB of memory */
3130                                         0,
3131                                         &tcp_hashinfo.bhash_size,
3132                                         NULL,
3133                                         0,
3134                                         64 * 1024);
3135         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3136         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3137                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3138                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3139         }
3140
3141
3142         cnt = tcp_hashinfo.ehash_mask + 1;
3143
3144         tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3145         sysctl_tcp_max_orphans = cnt / 2;
3146         sysctl_max_syn_backlog = max(128, cnt / 256);
3147
3148         tcp_init_mem();
3149         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3150         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3151         max_wshare = min(4UL*1024*1024, limit);
3152         max_rshare = min(6UL*1024*1024, limit);
3153
3154         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3155         sysctl_tcp_wmem[1] = 16*1024;
3156         sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3157
3158         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3159         sysctl_tcp_rmem[1] = 87380;
3160         sysctl_tcp_rmem[2] = max(87380, max_rshare);
3161
3162         pr_info("Hash tables configured (established %u bind %u)\n",
3163                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3164
3165         tcp_metrics_init();
3166         BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3167         tcp_tasklet_init();
3168 }