tcp.c revision b103cf34382f26ff48a87931b83f13b177b47c1a
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#include <linux/kernel.h> 249#include <linux/module.h> 250#include <linux/types.h> 251#include <linux/fcntl.h> 252#include <linux/poll.h> 253#include <linux/init.h> 254#include <linux/fs.h> 255#include <linux/skbuff.h> 256#include <linux/scatterlist.h> 257#include <linux/splice.h> 258#include <linux/net.h> 259#include <linux/socket.h> 260#include <linux/random.h> 261#include <linux/bootmem.h> 262#include <linux/highmem.h> 263#include <linux/swap.h> 264#include <linux/cache.h> 265#include <linux/err.h> 266#include <linux/crypto.h> 267 268#include <net/icmp.h> 269#include <net/tcp.h> 270#include <net/xfrm.h> 271#include <net/ip.h> 272#include <net/netdma.h> 273#include <net/sock.h> 274 275#include <asm/uaccess.h> 276#include <asm/ioctls.h> 277 278int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT; 279 280struct percpu_counter tcp_orphan_count; 281EXPORT_SYMBOL_GPL(tcp_orphan_count); 282 283int sysctl_tcp_mem[3] __read_mostly; 284int sysctl_tcp_wmem[3] __read_mostly; 285int sysctl_tcp_rmem[3] __read_mostly; 286 287EXPORT_SYMBOL(sysctl_tcp_mem); 288EXPORT_SYMBOL(sysctl_tcp_rmem); 289EXPORT_SYMBOL(sysctl_tcp_wmem); 290 291atomic_t tcp_memory_allocated; /* Current allocated memory. */ 292EXPORT_SYMBOL(tcp_memory_allocated); 293 294/* 295 * Current number of TCP sockets. 296 */ 297struct percpu_counter tcp_sockets_allocated; 298EXPORT_SYMBOL(tcp_sockets_allocated); 299 300/* 301 * TCP splice context 302 */ 303struct tcp_splice_state { 304 struct pipe_inode_info *pipe; 305 size_t len; 306 unsigned int flags; 307}; 308 309/* 310 * Pressure flag: try to collapse. 311 * Technical note: it is used by multiple contexts non atomically. 312 * All the __sk_mem_schedule() is of this nature: accounting 313 * is strict, actions are advisory and have some latency. 314 */ 315int tcp_memory_pressure __read_mostly; 316 317EXPORT_SYMBOL(tcp_memory_pressure); 318 319void tcp_enter_memory_pressure(struct sock *sk) 320{ 321 if (!tcp_memory_pressure) { 322 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES); 323 tcp_memory_pressure = 1; 324 } 325} 326 327EXPORT_SYMBOL(tcp_enter_memory_pressure); 328 329/* Convert seconds to retransmits based on initial and max timeout */ 330static u8 secs_to_retrans(int seconds, int timeout, int rto_max) 331{ 332 u8 res = 0; 333 334 if (seconds > 0) { 335 int period = timeout; 336 337 res = 1; 338 while (seconds > period && res < 255) { 339 res++; 340 timeout <<= 1; 341 if (timeout > rto_max) 342 timeout = rto_max; 343 period += timeout; 344 } 345 } 346 return res; 347} 348 349/* Convert retransmits to seconds based on initial and max timeout */ 350static int retrans_to_secs(u8 retrans, int timeout, int rto_max) 351{ 352 int period = 0; 353 354 if (retrans > 0) { 355 period = timeout; 356 while (--retrans) { 357 timeout <<= 1; 358 if (timeout > rto_max) 359 timeout = rto_max; 360 period += timeout; 361 } 362 } 363 return period; 364} 365 366/* 367 * Wait for a TCP event. 368 * 369 * Note that we don't need to lock the socket, as the upper poll layers 370 * take care of normal races (between the test and the event) and we don't 371 * go look at any of the socket buffers directly. 372 */ 373unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait) 374{ 375 unsigned int mask; 376 struct sock *sk = sock->sk; 377 struct tcp_sock *tp = tcp_sk(sk); 378 379 sock_poll_wait(file, sk->sk_sleep, wait); 380 if (sk->sk_state == TCP_LISTEN) 381 return inet_csk_listen_poll(sk); 382 383 /* Socket is not locked. We are protected from async events 384 * by poll logic and correct handling of state changes 385 * made by other threads is impossible in any case. 386 */ 387 388 mask = 0; 389 if (sk->sk_err) 390 mask = POLLERR; 391 392 /* 393 * POLLHUP is certainly not done right. But poll() doesn't 394 * have a notion of HUP in just one direction, and for a 395 * socket the read side is more interesting. 396 * 397 * Some poll() documentation says that POLLHUP is incompatible 398 * with the POLLOUT/POLLWR flags, so somebody should check this 399 * all. But careful, it tends to be safer to return too many 400 * bits than too few, and you can easily break real applications 401 * if you don't tell them that something has hung up! 402 * 403 * Check-me. 404 * 405 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and 406 * our fs/select.c). It means that after we received EOF, 407 * poll always returns immediately, making impossible poll() on write() 408 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP 409 * if and only if shutdown has been made in both directions. 410 * Actually, it is interesting to look how Solaris and DUX 411 * solve this dilemma. I would prefer, if POLLHUP were maskable, 412 * then we could set it on SND_SHUTDOWN. BTW examples given 413 * in Stevens' books assume exactly this behaviour, it explains 414 * why POLLHUP is incompatible with POLLOUT. --ANK 415 * 416 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent 417 * blocking on fresh not-connected or disconnected socket. --ANK 418 */ 419 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE) 420 mask |= POLLHUP; 421 if (sk->sk_shutdown & RCV_SHUTDOWN) 422 mask |= POLLIN | POLLRDNORM | POLLRDHUP; 423 424 /* Connected? */ 425 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) { 426 int target = sock_rcvlowat(sk, 0, INT_MAX); 427 428 if (tp->urg_seq == tp->copied_seq && 429 !sock_flag(sk, SOCK_URGINLINE) && 430 tp->urg_data) 431 target--; 432 433 /* Potential race condition. If read of tp below will 434 * escape above sk->sk_state, we can be illegally awaken 435 * in SYN_* states. */ 436 if (tp->rcv_nxt - tp->copied_seq >= target) 437 mask |= POLLIN | POLLRDNORM; 438 439 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) { 440 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) { 441 mask |= POLLOUT | POLLWRNORM; 442 } else { /* send SIGIO later */ 443 set_bit(SOCK_ASYNC_NOSPACE, 444 &sk->sk_socket->flags); 445 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 446 447 /* Race breaker. If space is freed after 448 * wspace test but before the flags are set, 449 * IO signal will be lost. 450 */ 451 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) 452 mask |= POLLOUT | POLLWRNORM; 453 } 454 } 455 456 if (tp->urg_data & TCP_URG_VALID) 457 mask |= POLLPRI; 458 } 459 return mask; 460} 461 462int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg) 463{ 464 struct tcp_sock *tp = tcp_sk(sk); 465 int answ; 466 467 switch (cmd) { 468 case SIOCINQ: 469 if (sk->sk_state == TCP_LISTEN) 470 return -EINVAL; 471 472 lock_sock(sk); 473 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 474 answ = 0; 475 else if (sock_flag(sk, SOCK_URGINLINE) || 476 !tp->urg_data || 477 before(tp->urg_seq, tp->copied_seq) || 478 !before(tp->urg_seq, tp->rcv_nxt)) { 479 struct sk_buff *skb; 480 481 answ = tp->rcv_nxt - tp->copied_seq; 482 483 /* Subtract 1, if FIN is in queue. */ 484 skb = skb_peek_tail(&sk->sk_receive_queue); 485 if (answ && skb) 486 answ -= tcp_hdr(skb)->fin; 487 } else 488 answ = tp->urg_seq - tp->copied_seq; 489 release_sock(sk); 490 break; 491 case SIOCATMARK: 492 answ = tp->urg_data && tp->urg_seq == tp->copied_seq; 493 break; 494 case SIOCOUTQ: 495 if (sk->sk_state == TCP_LISTEN) 496 return -EINVAL; 497 498 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 499 answ = 0; 500 else 501 answ = tp->write_seq - tp->snd_una; 502 break; 503 default: 504 return -ENOIOCTLCMD; 505 } 506 507 return put_user(answ, (int __user *)arg); 508} 509 510static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb) 511{ 512 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 513 tp->pushed_seq = tp->write_seq; 514} 515 516static inline int forced_push(struct tcp_sock *tp) 517{ 518 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1)); 519} 520 521static inline void skb_entail(struct sock *sk, struct sk_buff *skb) 522{ 523 struct tcp_sock *tp = tcp_sk(sk); 524 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 525 526 skb->csum = 0; 527 tcb->seq = tcb->end_seq = tp->write_seq; 528 tcb->flags = TCPCB_FLAG_ACK; 529 tcb->sacked = 0; 530 skb_header_release(skb); 531 tcp_add_write_queue_tail(sk, skb); 532 sk->sk_wmem_queued += skb->truesize; 533 sk_mem_charge(sk, skb->truesize); 534 if (tp->nonagle & TCP_NAGLE_PUSH) 535 tp->nonagle &= ~TCP_NAGLE_PUSH; 536} 537 538static inline void tcp_mark_urg(struct tcp_sock *tp, int flags, 539 struct sk_buff *skb) 540{ 541 if (flags & MSG_OOB) 542 tp->snd_up = tp->write_seq; 543} 544 545static inline void tcp_push(struct sock *sk, int flags, int mss_now, 546 int nonagle) 547{ 548 struct tcp_sock *tp = tcp_sk(sk); 549 550 if (tcp_send_head(sk)) { 551 struct sk_buff *skb = tcp_write_queue_tail(sk); 552 if (!(flags & MSG_MORE) || forced_push(tp)) 553 tcp_mark_push(tp, skb); 554 tcp_mark_urg(tp, flags, skb); 555 __tcp_push_pending_frames(sk, mss_now, 556 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle); 557 } 558} 559 560static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, 561 unsigned int offset, size_t len) 562{ 563 struct tcp_splice_state *tss = rd_desc->arg.data; 564 int ret; 565 566 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len), 567 tss->flags); 568 if (ret > 0) 569 rd_desc->count -= ret; 570 return ret; 571} 572 573static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss) 574{ 575 /* Store TCP splice context information in read_descriptor_t. */ 576 read_descriptor_t rd_desc = { 577 .arg.data = tss, 578 .count = tss->len, 579 }; 580 581 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv); 582} 583 584/** 585 * tcp_splice_read - splice data from TCP socket to a pipe 586 * @sock: socket to splice from 587 * @ppos: position (not valid) 588 * @pipe: pipe to splice to 589 * @len: number of bytes to splice 590 * @flags: splice modifier flags 591 * 592 * Description: 593 * Will read pages from given socket and fill them into a pipe. 594 * 595 **/ 596ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos, 597 struct pipe_inode_info *pipe, size_t len, 598 unsigned int flags) 599{ 600 struct sock *sk = sock->sk; 601 struct tcp_splice_state tss = { 602 .pipe = pipe, 603 .len = len, 604 .flags = flags, 605 }; 606 long timeo; 607 ssize_t spliced; 608 int ret; 609 610 /* 611 * We can't seek on a socket input 612 */ 613 if (unlikely(*ppos)) 614 return -ESPIPE; 615 616 ret = spliced = 0; 617 618 lock_sock(sk); 619 620 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK); 621 while (tss.len) { 622 ret = __tcp_splice_read(sk, &tss); 623 if (ret < 0) 624 break; 625 else if (!ret) { 626 if (spliced) 627 break; 628 if (sock_flag(sk, SOCK_DONE)) 629 break; 630 if (sk->sk_err) { 631 ret = sock_error(sk); 632 break; 633 } 634 if (sk->sk_shutdown & RCV_SHUTDOWN) 635 break; 636 if (sk->sk_state == TCP_CLOSE) { 637 /* 638 * This occurs when user tries to read 639 * from never connected socket. 640 */ 641 if (!sock_flag(sk, SOCK_DONE)) 642 ret = -ENOTCONN; 643 break; 644 } 645 if (!timeo) { 646 ret = -EAGAIN; 647 break; 648 } 649 sk_wait_data(sk, &timeo); 650 if (signal_pending(current)) { 651 ret = sock_intr_errno(timeo); 652 break; 653 } 654 continue; 655 } 656 tss.len -= ret; 657 spliced += ret; 658 659 if (!timeo) 660 break; 661 release_sock(sk); 662 lock_sock(sk); 663 664 if (sk->sk_err || sk->sk_state == TCP_CLOSE || 665 (sk->sk_shutdown & RCV_SHUTDOWN) || 666 signal_pending(current)) 667 break; 668 } 669 670 release_sock(sk); 671 672 if (spliced) 673 return spliced; 674 675 return ret; 676} 677 678struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp) 679{ 680 struct sk_buff *skb; 681 682 /* The TCP header must be at least 32-bit aligned. */ 683 size = ALIGN(size, 4); 684 685 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp); 686 if (skb) { 687 if (sk_wmem_schedule(sk, skb->truesize)) { 688 /* 689 * Make sure that we have exactly size bytes 690 * available to the caller, no more, no less. 691 */ 692 skb_reserve(skb, skb_tailroom(skb) - size); 693 return skb; 694 } 695 __kfree_skb(skb); 696 } else { 697 sk->sk_prot->enter_memory_pressure(sk); 698 sk_stream_moderate_sndbuf(sk); 699 } 700 return NULL; 701} 702 703static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now, 704 int large_allowed) 705{ 706 struct tcp_sock *tp = tcp_sk(sk); 707 u32 xmit_size_goal, old_size_goal; 708 709 xmit_size_goal = mss_now; 710 711 if (large_allowed && sk_can_gso(sk)) { 712 xmit_size_goal = ((sk->sk_gso_max_size - 1) - 713 inet_csk(sk)->icsk_af_ops->net_header_len - 714 inet_csk(sk)->icsk_ext_hdr_len - 715 tp->tcp_header_len); 716 717 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal); 718 719 /* We try hard to avoid divides here */ 720 old_size_goal = tp->xmit_size_goal_segs * mss_now; 721 722 if (likely(old_size_goal <= xmit_size_goal && 723 old_size_goal + mss_now > xmit_size_goal)) { 724 xmit_size_goal = old_size_goal; 725 } else { 726 tp->xmit_size_goal_segs = xmit_size_goal / mss_now; 727 xmit_size_goal = tp->xmit_size_goal_segs * mss_now; 728 } 729 } 730 731 return max(xmit_size_goal, mss_now); 732} 733 734static int tcp_send_mss(struct sock *sk, int *size_goal, int flags) 735{ 736 int mss_now; 737 738 mss_now = tcp_current_mss(sk); 739 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB)); 740 741 return mss_now; 742} 743 744static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset, 745 size_t psize, int flags) 746{ 747 struct tcp_sock *tp = tcp_sk(sk); 748 int mss_now, size_goal; 749 int err; 750 ssize_t copied; 751 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 752 753 /* Wait for a connection to finish. */ 754 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 755 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0) 756 goto out_err; 757 758 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags); 759 760 mss_now = tcp_send_mss(sk, &size_goal, flags); 761 copied = 0; 762 763 err = -EPIPE; 764 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 765 goto out_err; 766 767 while (psize > 0) { 768 struct sk_buff *skb = tcp_write_queue_tail(sk); 769 struct page *page = pages[poffset / PAGE_SIZE]; 770 int copy, i, can_coalesce; 771 int offset = poffset % PAGE_SIZE; 772 int size = min_t(size_t, psize, PAGE_SIZE - offset); 773 774 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) { 775new_segment: 776 if (!sk_stream_memory_free(sk)) 777 goto wait_for_sndbuf; 778 779 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation); 780 if (!skb) 781 goto wait_for_memory; 782 783 skb_entail(sk, skb); 784 copy = size_goal; 785 } 786 787 if (copy > size) 788 copy = size; 789 790 i = skb_shinfo(skb)->nr_frags; 791 can_coalesce = skb_can_coalesce(skb, i, page, offset); 792 if (!can_coalesce && i >= MAX_SKB_FRAGS) { 793 tcp_mark_push(tp, skb); 794 goto new_segment; 795 } 796 if (!sk_wmem_schedule(sk, copy)) 797 goto wait_for_memory; 798 799 if (can_coalesce) { 800 skb_shinfo(skb)->frags[i - 1].size += copy; 801 } else { 802 get_page(page); 803 skb_fill_page_desc(skb, i, page, offset, copy); 804 } 805 806 skb->len += copy; 807 skb->data_len += copy; 808 skb->truesize += copy; 809 sk->sk_wmem_queued += copy; 810 sk_mem_charge(sk, copy); 811 skb->ip_summed = CHECKSUM_PARTIAL; 812 tp->write_seq += copy; 813 TCP_SKB_CB(skb)->end_seq += copy; 814 skb_shinfo(skb)->gso_segs = 0; 815 816 if (!copied) 817 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH; 818 819 copied += copy; 820 poffset += copy; 821 if (!(psize -= copy)) 822 goto out; 823 824 if (skb->len < size_goal || (flags & MSG_OOB)) 825 continue; 826 827 if (forced_push(tp)) { 828 tcp_mark_push(tp, skb); 829 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 830 } else if (skb == tcp_send_head(sk)) 831 tcp_push_one(sk, mss_now); 832 continue; 833 834wait_for_sndbuf: 835 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 836wait_for_memory: 837 if (copied) 838 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH); 839 840 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0) 841 goto do_error; 842 843 mss_now = tcp_send_mss(sk, &size_goal, flags); 844 } 845 846out: 847 if (copied) 848 tcp_push(sk, flags, mss_now, tp->nonagle); 849 return copied; 850 851do_error: 852 if (copied) 853 goto out; 854out_err: 855 return sk_stream_error(sk, flags, err); 856} 857 858ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset, 859 size_t size, int flags) 860{ 861 ssize_t res; 862 struct sock *sk = sock->sk; 863 864 if (!(sk->sk_route_caps & NETIF_F_SG) || 865 !(sk->sk_route_caps & NETIF_F_ALL_CSUM)) 866 return sock_no_sendpage(sock, page, offset, size, flags); 867 868 lock_sock(sk); 869 TCP_CHECK_TIMER(sk); 870 res = do_tcp_sendpages(sk, &page, offset, size, flags); 871 TCP_CHECK_TIMER(sk); 872 release_sock(sk); 873 return res; 874} 875 876#define TCP_PAGE(sk) (sk->sk_sndmsg_page) 877#define TCP_OFF(sk) (sk->sk_sndmsg_off) 878 879static inline int select_size(struct sock *sk) 880{ 881 struct tcp_sock *tp = tcp_sk(sk); 882 int tmp = tp->mss_cache; 883 884 if (sk->sk_route_caps & NETIF_F_SG) { 885 if (sk_can_gso(sk)) 886 tmp = 0; 887 else { 888 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER); 889 890 if (tmp >= pgbreak && 891 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE) 892 tmp = pgbreak; 893 } 894 } 895 896 return tmp; 897} 898 899int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg, 900 size_t size) 901{ 902 struct sock *sk = sock->sk; 903 struct iovec *iov; 904 struct tcp_sock *tp = tcp_sk(sk); 905 struct sk_buff *skb; 906 int iovlen, flags; 907 int mss_now, size_goal; 908 int err, copied; 909 long timeo; 910 911 lock_sock(sk); 912 TCP_CHECK_TIMER(sk); 913 914 flags = msg->msg_flags; 915 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 916 917 /* Wait for a connection to finish. */ 918 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 919 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0) 920 goto out_err; 921 922 /* This should be in poll */ 923 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags); 924 925 mss_now = tcp_send_mss(sk, &size_goal, flags); 926 927 /* Ok commence sending. */ 928 iovlen = msg->msg_iovlen; 929 iov = msg->msg_iov; 930 copied = 0; 931 932 err = -EPIPE; 933 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 934 goto out_err; 935 936 while (--iovlen >= 0) { 937 int seglen = iov->iov_len; 938 unsigned char __user *from = iov->iov_base; 939 940 iov++; 941 942 while (seglen > 0) { 943 int copy = 0; 944 int max = size_goal; 945 946 skb = tcp_write_queue_tail(sk); 947 if (tcp_send_head(sk)) { 948 if (skb->ip_summed == CHECKSUM_NONE) 949 max = mss_now; 950 copy = max - skb->len; 951 } 952 953 if (copy <= 0) { 954new_segment: 955 /* Allocate new segment. If the interface is SG, 956 * allocate skb fitting to single page. 957 */ 958 if (!sk_stream_memory_free(sk)) 959 goto wait_for_sndbuf; 960 961 skb = sk_stream_alloc_skb(sk, select_size(sk), 962 sk->sk_allocation); 963 if (!skb) 964 goto wait_for_memory; 965 966 /* 967 * Check whether we can use HW checksum. 968 */ 969 if (sk->sk_route_caps & NETIF_F_ALL_CSUM) 970 skb->ip_summed = CHECKSUM_PARTIAL; 971 972 skb_entail(sk, skb); 973 copy = size_goal; 974 max = size_goal; 975 } 976 977 /* Try to append data to the end of skb. */ 978 if (copy > seglen) 979 copy = seglen; 980 981 /* Where to copy to? */ 982 if (skb_tailroom(skb) > 0) { 983 /* We have some space in skb head. Superb! */ 984 if (copy > skb_tailroom(skb)) 985 copy = skb_tailroom(skb); 986 if ((err = skb_add_data(skb, from, copy)) != 0) 987 goto do_fault; 988 } else { 989 int merge = 0; 990 int i = skb_shinfo(skb)->nr_frags; 991 struct page *page = TCP_PAGE(sk); 992 int off = TCP_OFF(sk); 993 994 if (skb_can_coalesce(skb, i, page, off) && 995 off != PAGE_SIZE) { 996 /* We can extend the last page 997 * fragment. */ 998 merge = 1; 999 } else if (i == MAX_SKB_FRAGS || 1000 (!i && 1001 !(sk->sk_route_caps & NETIF_F_SG))) { 1002 /* Need to add new fragment and cannot 1003 * do this because interface is non-SG, 1004 * or because all the page slots are 1005 * busy. */ 1006 tcp_mark_push(tp, skb); 1007 goto new_segment; 1008 } else if (page) { 1009 if (off == PAGE_SIZE) { 1010 put_page(page); 1011 TCP_PAGE(sk) = page = NULL; 1012 off = 0; 1013 } 1014 } else 1015 off = 0; 1016 1017 if (copy > PAGE_SIZE - off) 1018 copy = PAGE_SIZE - off; 1019 1020 if (!sk_wmem_schedule(sk, copy)) 1021 goto wait_for_memory; 1022 1023 if (!page) { 1024 /* Allocate new cache page. */ 1025 if (!(page = sk_stream_alloc_page(sk))) 1026 goto wait_for_memory; 1027 } 1028 1029 /* Time to copy data. We are close to 1030 * the end! */ 1031 err = skb_copy_to_page(sk, from, skb, page, 1032 off, copy); 1033 if (err) { 1034 /* If this page was new, give it to the 1035 * socket so it does not get leaked. 1036 */ 1037 if (!TCP_PAGE(sk)) { 1038 TCP_PAGE(sk) = page; 1039 TCP_OFF(sk) = 0; 1040 } 1041 goto do_error; 1042 } 1043 1044 /* Update the skb. */ 1045 if (merge) { 1046 skb_shinfo(skb)->frags[i - 1].size += 1047 copy; 1048 } else { 1049 skb_fill_page_desc(skb, i, page, off, copy); 1050 if (TCP_PAGE(sk)) { 1051 get_page(page); 1052 } else if (off + copy < PAGE_SIZE) { 1053 get_page(page); 1054 TCP_PAGE(sk) = page; 1055 } 1056 } 1057 1058 TCP_OFF(sk) = off + copy; 1059 } 1060 1061 if (!copied) 1062 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH; 1063 1064 tp->write_seq += copy; 1065 TCP_SKB_CB(skb)->end_seq += copy; 1066 skb_shinfo(skb)->gso_segs = 0; 1067 1068 from += copy; 1069 copied += copy; 1070 if ((seglen -= copy) == 0 && iovlen == 0) 1071 goto out; 1072 1073 if (skb->len < max || (flags & MSG_OOB)) 1074 continue; 1075 1076 if (forced_push(tp)) { 1077 tcp_mark_push(tp, skb); 1078 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 1079 } else if (skb == tcp_send_head(sk)) 1080 tcp_push_one(sk, mss_now); 1081 continue; 1082 1083wait_for_sndbuf: 1084 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1085wait_for_memory: 1086 if (copied) 1087 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH); 1088 1089 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0) 1090 goto do_error; 1091 1092 mss_now = tcp_send_mss(sk, &size_goal, flags); 1093 } 1094 } 1095 1096out: 1097 if (copied) 1098 tcp_push(sk, flags, mss_now, tp->nonagle); 1099 TCP_CHECK_TIMER(sk); 1100 release_sock(sk); 1101 return copied; 1102 1103do_fault: 1104 if (!skb->len) { 1105 tcp_unlink_write_queue(skb, sk); 1106 /* It is the one place in all of TCP, except connection 1107 * reset, where we can be unlinking the send_head. 1108 */ 1109 tcp_check_send_head(sk, skb); 1110 sk_wmem_free_skb(sk, skb); 1111 } 1112 1113do_error: 1114 if (copied) 1115 goto out; 1116out_err: 1117 err = sk_stream_error(sk, flags, err); 1118 TCP_CHECK_TIMER(sk); 1119 release_sock(sk); 1120 return err; 1121} 1122 1123/* 1124 * Handle reading urgent data. BSD has very simple semantics for 1125 * this, no blocking and very strange errors 8) 1126 */ 1127 1128static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags) 1129{ 1130 struct tcp_sock *tp = tcp_sk(sk); 1131 1132 /* No URG data to read. */ 1133 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data || 1134 tp->urg_data == TCP_URG_READ) 1135 return -EINVAL; /* Yes this is right ! */ 1136 1137 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE)) 1138 return -ENOTCONN; 1139 1140 if (tp->urg_data & TCP_URG_VALID) { 1141 int err = 0; 1142 char c = tp->urg_data; 1143 1144 if (!(flags & MSG_PEEK)) 1145 tp->urg_data = TCP_URG_READ; 1146 1147 /* Read urgent data. */ 1148 msg->msg_flags |= MSG_OOB; 1149 1150 if (len > 0) { 1151 if (!(flags & MSG_TRUNC)) 1152 err = memcpy_toiovec(msg->msg_iov, &c, 1); 1153 len = 1; 1154 } else 1155 msg->msg_flags |= MSG_TRUNC; 1156 1157 return err ? -EFAULT : len; 1158 } 1159 1160 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 1161 return 0; 1162 1163 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and 1164 * the available implementations agree in this case: 1165 * this call should never block, independent of the 1166 * blocking state of the socket. 1167 * Mike <pall@rz.uni-karlsruhe.de> 1168 */ 1169 return -EAGAIN; 1170} 1171 1172/* Clean up the receive buffer for full frames taken by the user, 1173 * then send an ACK if necessary. COPIED is the number of bytes 1174 * tcp_recvmsg has given to the user so far, it speeds up the 1175 * calculation of whether or not we must ACK for the sake of 1176 * a window update. 1177 */ 1178void tcp_cleanup_rbuf(struct sock *sk, int copied) 1179{ 1180 struct tcp_sock *tp = tcp_sk(sk); 1181 int time_to_ack = 0; 1182 1183#if TCP_DEBUG 1184 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 1185 1186 WARN_ON(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)); 1187#endif 1188 1189 if (inet_csk_ack_scheduled(sk)) { 1190 const struct inet_connection_sock *icsk = inet_csk(sk); 1191 /* Delayed ACKs frequently hit locked sockets during bulk 1192 * receive. */ 1193 if (icsk->icsk_ack.blocked || 1194 /* Once-per-two-segments ACK was not sent by tcp_input.c */ 1195 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss || 1196 /* 1197 * If this read emptied read buffer, we send ACK, if 1198 * connection is not bidirectional, user drained 1199 * receive buffer and there was a small segment 1200 * in queue. 1201 */ 1202 (copied > 0 && 1203 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) || 1204 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) && 1205 !icsk->icsk_ack.pingpong)) && 1206 !atomic_read(&sk->sk_rmem_alloc))) 1207 time_to_ack = 1; 1208 } 1209 1210 /* We send an ACK if we can now advertise a non-zero window 1211 * which has been raised "significantly". 1212 * 1213 * Even if window raised up to infinity, do not send window open ACK 1214 * in states, where we will not receive more. It is useless. 1215 */ 1216 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) { 1217 __u32 rcv_window_now = tcp_receive_window(tp); 1218 1219 /* Optimize, __tcp_select_window() is not cheap. */ 1220 if (2*rcv_window_now <= tp->window_clamp) { 1221 __u32 new_window = __tcp_select_window(sk); 1222 1223 /* Send ACK now, if this read freed lots of space 1224 * in our buffer. Certainly, new_window is new window. 1225 * We can advertise it now, if it is not less than current one. 1226 * "Lots" means "at least twice" here. 1227 */ 1228 if (new_window && new_window >= 2 * rcv_window_now) 1229 time_to_ack = 1; 1230 } 1231 } 1232 if (time_to_ack) 1233 tcp_send_ack(sk); 1234} 1235 1236static void tcp_prequeue_process(struct sock *sk) 1237{ 1238 struct sk_buff *skb; 1239 struct tcp_sock *tp = tcp_sk(sk); 1240 1241 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED); 1242 1243 /* RX process wants to run with disabled BHs, though it is not 1244 * necessary */ 1245 local_bh_disable(); 1246 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL) 1247 sk_backlog_rcv(sk, skb); 1248 local_bh_enable(); 1249 1250 /* Clear memory counter. */ 1251 tp->ucopy.memory = 0; 1252} 1253 1254static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off) 1255{ 1256 struct sk_buff *skb; 1257 u32 offset; 1258 1259 skb_queue_walk(&sk->sk_receive_queue, skb) { 1260 offset = seq - TCP_SKB_CB(skb)->seq; 1261 if (tcp_hdr(skb)->syn) 1262 offset--; 1263 if (offset < skb->len || tcp_hdr(skb)->fin) { 1264 *off = offset; 1265 return skb; 1266 } 1267 } 1268 return NULL; 1269} 1270 1271/* 1272 * This routine provides an alternative to tcp_recvmsg() for routines 1273 * that would like to handle copying from skbuffs directly in 'sendfile' 1274 * fashion. 1275 * Note: 1276 * - It is assumed that the socket was locked by the caller. 1277 * - The routine does not block. 1278 * - At present, there is no support for reading OOB data 1279 * or for 'peeking' the socket using this routine 1280 * (although both would be easy to implement). 1281 */ 1282int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, 1283 sk_read_actor_t recv_actor) 1284{ 1285 struct sk_buff *skb; 1286 struct tcp_sock *tp = tcp_sk(sk); 1287 u32 seq = tp->copied_seq; 1288 u32 offset; 1289 int copied = 0; 1290 1291 if (sk->sk_state == TCP_LISTEN) 1292 return -ENOTCONN; 1293 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) { 1294 if (offset < skb->len) { 1295 int used; 1296 size_t len; 1297 1298 len = skb->len - offset; 1299 /* Stop reading if we hit a patch of urgent data */ 1300 if (tp->urg_data) { 1301 u32 urg_offset = tp->urg_seq - seq; 1302 if (urg_offset < len) 1303 len = urg_offset; 1304 if (!len) 1305 break; 1306 } 1307 used = recv_actor(desc, skb, offset, len); 1308 if (used < 0) { 1309 if (!copied) 1310 copied = used; 1311 break; 1312 } else if (used <= len) { 1313 seq += used; 1314 copied += used; 1315 offset += used; 1316 } 1317 /* 1318 * If recv_actor drops the lock (e.g. TCP splice 1319 * receive) the skb pointer might be invalid when 1320 * getting here: tcp_collapse might have deleted it 1321 * while aggregating skbs from the socket queue. 1322 */ 1323 skb = tcp_recv_skb(sk, seq-1, &offset); 1324 if (!skb || (offset+1 != skb->len)) 1325 break; 1326 } 1327 if (tcp_hdr(skb)->fin) { 1328 sk_eat_skb(sk, skb, 0); 1329 ++seq; 1330 break; 1331 } 1332 sk_eat_skb(sk, skb, 0); 1333 if (!desc->count) 1334 break; 1335 } 1336 tp->copied_seq = seq; 1337 1338 tcp_rcv_space_adjust(sk); 1339 1340 /* Clean up data we have read: This will do ACK frames. */ 1341 if (copied > 0) 1342 tcp_cleanup_rbuf(sk, copied); 1343 return copied; 1344} 1345 1346/* 1347 * This routine copies from a sock struct into the user buffer. 1348 * 1349 * Technical note: in 2.3 we work on _locked_ socket, so that 1350 * tricks with *seq access order and skb->users are not required. 1351 * Probably, code can be easily improved even more. 1352 */ 1353 1354int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg, 1355 size_t len, int nonblock, int flags, int *addr_len) 1356{ 1357 struct tcp_sock *tp = tcp_sk(sk); 1358 int copied = 0; 1359 u32 peek_seq; 1360 u32 *seq; 1361 unsigned long used; 1362 int err; 1363 int target; /* Read at least this many bytes */ 1364 long timeo; 1365 struct task_struct *user_recv = NULL; 1366 int copied_early = 0; 1367 struct sk_buff *skb; 1368 u32 urg_hole = 0; 1369 1370 lock_sock(sk); 1371 1372 TCP_CHECK_TIMER(sk); 1373 1374 err = -ENOTCONN; 1375 if (sk->sk_state == TCP_LISTEN) 1376 goto out; 1377 1378 timeo = sock_rcvtimeo(sk, nonblock); 1379 1380 /* Urgent data needs to be handled specially. */ 1381 if (flags & MSG_OOB) 1382 goto recv_urg; 1383 1384 seq = &tp->copied_seq; 1385 if (flags & MSG_PEEK) { 1386 peek_seq = tp->copied_seq; 1387 seq = &peek_seq; 1388 } 1389 1390 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 1391 1392#ifdef CONFIG_NET_DMA 1393 tp->ucopy.dma_chan = NULL; 1394 preempt_disable(); 1395 skb = skb_peek_tail(&sk->sk_receive_queue); 1396 { 1397 int available = 0; 1398 1399 if (skb) 1400 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq); 1401 if ((available < target) && 1402 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) && 1403 !sysctl_tcp_low_latency && 1404 dma_find_channel(DMA_MEMCPY)) { 1405 preempt_enable_no_resched(); 1406 tp->ucopy.pinned_list = 1407 dma_pin_iovec_pages(msg->msg_iov, len); 1408 } else { 1409 preempt_enable_no_resched(); 1410 } 1411 } 1412#endif 1413 1414 do { 1415 u32 offset; 1416 1417 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */ 1418 if (tp->urg_data && tp->urg_seq == *seq) { 1419 if (copied) 1420 break; 1421 if (signal_pending(current)) { 1422 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN; 1423 break; 1424 } 1425 } 1426 1427 /* Next get a buffer. */ 1428 1429 skb_queue_walk(&sk->sk_receive_queue, skb) { 1430 /* Now that we have two receive queues this 1431 * shouldn't happen. 1432 */ 1433 if (before(*seq, TCP_SKB_CB(skb)->seq)) { 1434 printk(KERN_INFO "recvmsg bug: copied %X " 1435 "seq %X\n", *seq, TCP_SKB_CB(skb)->seq); 1436 break; 1437 } 1438 offset = *seq - TCP_SKB_CB(skb)->seq; 1439 if (tcp_hdr(skb)->syn) 1440 offset--; 1441 if (offset < skb->len) 1442 goto found_ok_skb; 1443 if (tcp_hdr(skb)->fin) 1444 goto found_fin_ok; 1445 WARN_ON(!(flags & MSG_PEEK)); 1446 } 1447 1448 /* Well, if we have backlog, try to process it now yet. */ 1449 1450 if (copied >= target && !sk->sk_backlog.tail) 1451 break; 1452 1453 if (copied) { 1454 if (sk->sk_err || 1455 sk->sk_state == TCP_CLOSE || 1456 (sk->sk_shutdown & RCV_SHUTDOWN) || 1457 !timeo || 1458 signal_pending(current)) 1459 break; 1460 } else { 1461 if (sock_flag(sk, SOCK_DONE)) 1462 break; 1463 1464 if (sk->sk_err) { 1465 copied = sock_error(sk); 1466 break; 1467 } 1468 1469 if (sk->sk_shutdown & RCV_SHUTDOWN) 1470 break; 1471 1472 if (sk->sk_state == TCP_CLOSE) { 1473 if (!sock_flag(sk, SOCK_DONE)) { 1474 /* This occurs when user tries to read 1475 * from never connected socket. 1476 */ 1477 copied = -ENOTCONN; 1478 break; 1479 } 1480 break; 1481 } 1482 1483 if (!timeo) { 1484 copied = -EAGAIN; 1485 break; 1486 } 1487 1488 if (signal_pending(current)) { 1489 copied = sock_intr_errno(timeo); 1490 break; 1491 } 1492 } 1493 1494 tcp_cleanup_rbuf(sk, copied); 1495 1496 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) { 1497 /* Install new reader */ 1498 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) { 1499 user_recv = current; 1500 tp->ucopy.task = user_recv; 1501 tp->ucopy.iov = msg->msg_iov; 1502 } 1503 1504 tp->ucopy.len = len; 1505 1506 WARN_ON(tp->copied_seq != tp->rcv_nxt && 1507 !(flags & (MSG_PEEK | MSG_TRUNC))); 1508 1509 /* Ugly... If prequeue is not empty, we have to 1510 * process it before releasing socket, otherwise 1511 * order will be broken at second iteration. 1512 * More elegant solution is required!!! 1513 * 1514 * Look: we have the following (pseudo)queues: 1515 * 1516 * 1. packets in flight 1517 * 2. backlog 1518 * 3. prequeue 1519 * 4. receive_queue 1520 * 1521 * Each queue can be processed only if the next ones 1522 * are empty. At this point we have empty receive_queue. 1523 * But prequeue _can_ be not empty after 2nd iteration, 1524 * when we jumped to start of loop because backlog 1525 * processing added something to receive_queue. 1526 * We cannot release_sock(), because backlog contains 1527 * packets arrived _after_ prequeued ones. 1528 * 1529 * Shortly, algorithm is clear --- to process all 1530 * the queues in order. We could make it more directly, 1531 * requeueing packets from backlog to prequeue, if 1532 * is not empty. It is more elegant, but eats cycles, 1533 * unfortunately. 1534 */ 1535 if (!skb_queue_empty(&tp->ucopy.prequeue)) 1536 goto do_prequeue; 1537 1538 /* __ Set realtime policy in scheduler __ */ 1539 } 1540 1541 if (copied >= target) { 1542 /* Do not sleep, just process backlog. */ 1543 release_sock(sk); 1544 lock_sock(sk); 1545 } else 1546 sk_wait_data(sk, &timeo); 1547 1548#ifdef CONFIG_NET_DMA 1549 tp->ucopy.wakeup = 0; 1550#endif 1551 1552 if (user_recv) { 1553 int chunk; 1554 1555 /* __ Restore normal policy in scheduler __ */ 1556 1557 if ((chunk = len - tp->ucopy.len) != 0) { 1558 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk); 1559 len -= chunk; 1560 copied += chunk; 1561 } 1562 1563 if (tp->rcv_nxt == tp->copied_seq && 1564 !skb_queue_empty(&tp->ucopy.prequeue)) { 1565do_prequeue: 1566 tcp_prequeue_process(sk); 1567 1568 if ((chunk = len - tp->ucopy.len) != 0) { 1569 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1570 len -= chunk; 1571 copied += chunk; 1572 } 1573 } 1574 } 1575 if ((flags & MSG_PEEK) && 1576 (peek_seq - copied - urg_hole != tp->copied_seq)) { 1577 if (net_ratelimit()) 1578 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n", 1579 current->comm, task_pid_nr(current)); 1580 peek_seq = tp->copied_seq; 1581 } 1582 continue; 1583 1584 found_ok_skb: 1585 /* Ok so how much can we use? */ 1586 used = skb->len - offset; 1587 if (len < used) 1588 used = len; 1589 1590 /* Do we have urgent data here? */ 1591 if (tp->urg_data) { 1592 u32 urg_offset = tp->urg_seq - *seq; 1593 if (urg_offset < used) { 1594 if (!urg_offset) { 1595 if (!sock_flag(sk, SOCK_URGINLINE)) { 1596 ++*seq; 1597 urg_hole++; 1598 offset++; 1599 used--; 1600 if (!used) 1601 goto skip_copy; 1602 } 1603 } else 1604 used = urg_offset; 1605 } 1606 } 1607 1608 if (!(flags & MSG_TRUNC)) { 1609#ifdef CONFIG_NET_DMA 1610 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 1611 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY); 1612 1613 if (tp->ucopy.dma_chan) { 1614 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec( 1615 tp->ucopy.dma_chan, skb, offset, 1616 msg->msg_iov, used, 1617 tp->ucopy.pinned_list); 1618 1619 if (tp->ucopy.dma_cookie < 0) { 1620 1621 printk(KERN_ALERT "dma_cookie < 0\n"); 1622 1623 /* Exception. Bailout! */ 1624 if (!copied) 1625 copied = -EFAULT; 1626 break; 1627 } 1628 if ((offset + used) == skb->len) 1629 copied_early = 1; 1630 1631 } else 1632#endif 1633 { 1634 err = skb_copy_datagram_iovec(skb, offset, 1635 msg->msg_iov, used); 1636 if (err) { 1637 /* Exception. Bailout! */ 1638 if (!copied) 1639 copied = -EFAULT; 1640 break; 1641 } 1642 } 1643 } 1644 1645 *seq += used; 1646 copied += used; 1647 len -= used; 1648 1649 tcp_rcv_space_adjust(sk); 1650 1651skip_copy: 1652 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) { 1653 tp->urg_data = 0; 1654 tcp_fast_path_check(sk); 1655 } 1656 if (used + offset < skb->len) 1657 continue; 1658 1659 if (tcp_hdr(skb)->fin) 1660 goto found_fin_ok; 1661 if (!(flags & MSG_PEEK)) { 1662 sk_eat_skb(sk, skb, copied_early); 1663 copied_early = 0; 1664 } 1665 continue; 1666 1667 found_fin_ok: 1668 /* Process the FIN. */ 1669 ++*seq; 1670 if (!(flags & MSG_PEEK)) { 1671 sk_eat_skb(sk, skb, copied_early); 1672 copied_early = 0; 1673 } 1674 break; 1675 } while (len > 0); 1676 1677 if (user_recv) { 1678 if (!skb_queue_empty(&tp->ucopy.prequeue)) { 1679 int chunk; 1680 1681 tp->ucopy.len = copied > 0 ? len : 0; 1682 1683 tcp_prequeue_process(sk); 1684 1685 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) { 1686 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1687 len -= chunk; 1688 copied += chunk; 1689 } 1690 } 1691 1692 tp->ucopy.task = NULL; 1693 tp->ucopy.len = 0; 1694 } 1695 1696#ifdef CONFIG_NET_DMA 1697 if (tp->ucopy.dma_chan) { 1698 dma_cookie_t done, used; 1699 1700 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1701 1702 while (dma_async_memcpy_complete(tp->ucopy.dma_chan, 1703 tp->ucopy.dma_cookie, &done, 1704 &used) == DMA_IN_PROGRESS) { 1705 /* do partial cleanup of sk_async_wait_queue */ 1706 while ((skb = skb_peek(&sk->sk_async_wait_queue)) && 1707 (dma_async_is_complete(skb->dma_cookie, done, 1708 used) == DMA_SUCCESS)) { 1709 __skb_dequeue(&sk->sk_async_wait_queue); 1710 kfree_skb(skb); 1711 } 1712 } 1713 1714 /* Safe to free early-copied skbs now */ 1715 __skb_queue_purge(&sk->sk_async_wait_queue); 1716 tp->ucopy.dma_chan = NULL; 1717 } 1718 if (tp->ucopy.pinned_list) { 1719 dma_unpin_iovec_pages(tp->ucopy.pinned_list); 1720 tp->ucopy.pinned_list = NULL; 1721 } 1722#endif 1723 1724 /* According to UNIX98, msg_name/msg_namelen are ignored 1725 * on connected socket. I was just happy when found this 8) --ANK 1726 */ 1727 1728 /* Clean up data we have read: This will do ACK frames. */ 1729 tcp_cleanup_rbuf(sk, copied); 1730 1731 TCP_CHECK_TIMER(sk); 1732 release_sock(sk); 1733 return copied; 1734 1735out: 1736 TCP_CHECK_TIMER(sk); 1737 release_sock(sk); 1738 return err; 1739 1740recv_urg: 1741 err = tcp_recv_urg(sk, msg, len, flags); 1742 goto out; 1743} 1744 1745void tcp_set_state(struct sock *sk, int state) 1746{ 1747 int oldstate = sk->sk_state; 1748 1749 switch (state) { 1750 case TCP_ESTABLISHED: 1751 if (oldstate != TCP_ESTABLISHED) 1752 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1753 break; 1754 1755 case TCP_CLOSE: 1756 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED) 1757 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS); 1758 1759 sk->sk_prot->unhash(sk); 1760 if (inet_csk(sk)->icsk_bind_hash && 1761 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK)) 1762 inet_put_port(sk); 1763 /* fall through */ 1764 default: 1765 if (oldstate == TCP_ESTABLISHED) 1766 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1767 } 1768 1769 /* Change state AFTER socket is unhashed to avoid closed 1770 * socket sitting in hash tables. 1771 */ 1772 sk->sk_state = state; 1773 1774#ifdef STATE_TRACE 1775 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]); 1776#endif 1777} 1778EXPORT_SYMBOL_GPL(tcp_set_state); 1779 1780/* 1781 * State processing on a close. This implements the state shift for 1782 * sending our FIN frame. Note that we only send a FIN for some 1783 * states. A shutdown() may have already sent the FIN, or we may be 1784 * closed. 1785 */ 1786 1787static const unsigned char new_state[16] = { 1788 /* current state: new state: action: */ 1789 /* (Invalid) */ TCP_CLOSE, 1790 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1791 /* TCP_SYN_SENT */ TCP_CLOSE, 1792 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1793 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1, 1794 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2, 1795 /* TCP_TIME_WAIT */ TCP_CLOSE, 1796 /* TCP_CLOSE */ TCP_CLOSE, 1797 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN, 1798 /* TCP_LAST_ACK */ TCP_LAST_ACK, 1799 /* TCP_LISTEN */ TCP_CLOSE, 1800 /* TCP_CLOSING */ TCP_CLOSING, 1801}; 1802 1803static int tcp_close_state(struct sock *sk) 1804{ 1805 int next = (int)new_state[sk->sk_state]; 1806 int ns = next & TCP_STATE_MASK; 1807 1808 tcp_set_state(sk, ns); 1809 1810 return next & TCP_ACTION_FIN; 1811} 1812 1813/* 1814 * Shutdown the sending side of a connection. Much like close except 1815 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD). 1816 */ 1817 1818void tcp_shutdown(struct sock *sk, int how) 1819{ 1820 /* We need to grab some memory, and put together a FIN, 1821 * and then put it into the queue to be sent. 1822 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92. 1823 */ 1824 if (!(how & SEND_SHUTDOWN)) 1825 return; 1826 1827 /* If we've already sent a FIN, or it's a closed state, skip this. */ 1828 if ((1 << sk->sk_state) & 1829 (TCPF_ESTABLISHED | TCPF_SYN_SENT | 1830 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) { 1831 /* Clear out any half completed packets. FIN if needed. */ 1832 if (tcp_close_state(sk)) 1833 tcp_send_fin(sk); 1834 } 1835} 1836 1837void tcp_close(struct sock *sk, long timeout) 1838{ 1839 struct sk_buff *skb; 1840 int data_was_unread = 0; 1841 int state; 1842 1843 lock_sock(sk); 1844 sk->sk_shutdown = SHUTDOWN_MASK; 1845 1846 if (sk->sk_state == TCP_LISTEN) { 1847 tcp_set_state(sk, TCP_CLOSE); 1848 1849 /* Special case. */ 1850 inet_csk_listen_stop(sk); 1851 1852 goto adjudge_to_death; 1853 } 1854 1855 /* We need to flush the recv. buffs. We do this only on the 1856 * descriptor close, not protocol-sourced closes, because the 1857 * reader process may not have drained the data yet! 1858 */ 1859 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) { 1860 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq - 1861 tcp_hdr(skb)->fin; 1862 data_was_unread += len; 1863 __kfree_skb(skb); 1864 } 1865 1866 sk_mem_reclaim(sk); 1867 1868 /* As outlined in RFC 2525, section 2.17, we send a RST here because 1869 * data was lost. To witness the awful effects of the old behavior of 1870 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk 1871 * GET in an FTP client, suspend the process, wait for the client to 1872 * advertise a zero window, then kill -9 the FTP client, wheee... 1873 * Note: timeout is always zero in such a case. 1874 */ 1875 if (data_was_unread) { 1876 /* Unread data was tossed, zap the connection. */ 1877 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE); 1878 tcp_set_state(sk, TCP_CLOSE); 1879 tcp_send_active_reset(sk, sk->sk_allocation); 1880 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) { 1881 /* Check zero linger _after_ checking for unread data. */ 1882 sk->sk_prot->disconnect(sk, 0); 1883 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 1884 } else if (tcp_close_state(sk)) { 1885 /* We FIN if the application ate all the data before 1886 * zapping the connection. 1887 */ 1888 1889 /* RED-PEN. Formally speaking, we have broken TCP state 1890 * machine. State transitions: 1891 * 1892 * TCP_ESTABLISHED -> TCP_FIN_WAIT1 1893 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible) 1894 * TCP_CLOSE_WAIT -> TCP_LAST_ACK 1895 * 1896 * are legal only when FIN has been sent (i.e. in window), 1897 * rather than queued out of window. Purists blame. 1898 * 1899 * F.e. "RFC state" is ESTABLISHED, 1900 * if Linux state is FIN-WAIT-1, but FIN is still not sent. 1901 * 1902 * The visible declinations are that sometimes 1903 * we enter time-wait state, when it is not required really 1904 * (harmless), do not send active resets, when they are 1905 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when 1906 * they look as CLOSING or LAST_ACK for Linux) 1907 * Probably, I missed some more holelets. 1908 * --ANK 1909 */ 1910 tcp_send_fin(sk); 1911 } 1912 1913 sk_stream_wait_close(sk, timeout); 1914 1915adjudge_to_death: 1916 state = sk->sk_state; 1917 sock_hold(sk); 1918 sock_orphan(sk); 1919 1920 /* It is the last release_sock in its life. It will remove backlog. */ 1921 release_sock(sk); 1922 1923 1924 /* Now socket is owned by kernel and we acquire BH lock 1925 to finish close. No need to check for user refs. 1926 */ 1927 local_bh_disable(); 1928 bh_lock_sock(sk); 1929 WARN_ON(sock_owned_by_user(sk)); 1930 1931 percpu_counter_inc(sk->sk_prot->orphan_count); 1932 1933 /* Have we already been destroyed by a softirq or backlog? */ 1934 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE) 1935 goto out; 1936 1937 /* This is a (useful) BSD violating of the RFC. There is a 1938 * problem with TCP as specified in that the other end could 1939 * keep a socket open forever with no application left this end. 1940 * We use a 3 minute timeout (about the same as BSD) then kill 1941 * our end. If they send after that then tough - BUT: long enough 1942 * that we won't make the old 4*rto = almost no time - whoops 1943 * reset mistake. 1944 * 1945 * Nope, it was not mistake. It is really desired behaviour 1946 * f.e. on http servers, when such sockets are useless, but 1947 * consume significant resources. Let's do it with special 1948 * linger2 option. --ANK 1949 */ 1950 1951 if (sk->sk_state == TCP_FIN_WAIT2) { 1952 struct tcp_sock *tp = tcp_sk(sk); 1953 if (tp->linger2 < 0) { 1954 tcp_set_state(sk, TCP_CLOSE); 1955 tcp_send_active_reset(sk, GFP_ATOMIC); 1956 NET_INC_STATS_BH(sock_net(sk), 1957 LINUX_MIB_TCPABORTONLINGER); 1958 } else { 1959 const int tmo = tcp_fin_time(sk); 1960 1961 if (tmo > TCP_TIMEWAIT_LEN) { 1962 inet_csk_reset_keepalive_timer(sk, 1963 tmo - TCP_TIMEWAIT_LEN); 1964 } else { 1965 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 1966 goto out; 1967 } 1968 } 1969 } 1970 if (sk->sk_state != TCP_CLOSE) { 1971 int orphan_count = percpu_counter_read_positive( 1972 sk->sk_prot->orphan_count); 1973 1974 sk_mem_reclaim(sk); 1975 if (tcp_too_many_orphans(sk, orphan_count)) { 1976 if (net_ratelimit()) 1977 printk(KERN_INFO "TCP: too many of orphaned " 1978 "sockets\n"); 1979 tcp_set_state(sk, TCP_CLOSE); 1980 tcp_send_active_reset(sk, GFP_ATOMIC); 1981 NET_INC_STATS_BH(sock_net(sk), 1982 LINUX_MIB_TCPABORTONMEMORY); 1983 } 1984 } 1985 1986 if (sk->sk_state == TCP_CLOSE) 1987 inet_csk_destroy_sock(sk); 1988 /* Otherwise, socket is reprieved until protocol close. */ 1989 1990out: 1991 bh_unlock_sock(sk); 1992 local_bh_enable(); 1993 sock_put(sk); 1994} 1995 1996/* These states need RST on ABORT according to RFC793 */ 1997 1998static inline int tcp_need_reset(int state) 1999{ 2000 return (1 << state) & 2001 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 | 2002 TCPF_FIN_WAIT2 | TCPF_SYN_RECV); 2003} 2004 2005int tcp_disconnect(struct sock *sk, int flags) 2006{ 2007 struct inet_sock *inet = inet_sk(sk); 2008 struct inet_connection_sock *icsk = inet_csk(sk); 2009 struct tcp_sock *tp = tcp_sk(sk); 2010 int err = 0; 2011 int old_state = sk->sk_state; 2012 2013 if (old_state != TCP_CLOSE) 2014 tcp_set_state(sk, TCP_CLOSE); 2015 2016 /* ABORT function of RFC793 */ 2017 if (old_state == TCP_LISTEN) { 2018 inet_csk_listen_stop(sk); 2019 } else if (tcp_need_reset(old_state) || 2020 (tp->snd_nxt != tp->write_seq && 2021 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) { 2022 /* The last check adjusts for discrepancy of Linux wrt. RFC 2023 * states 2024 */ 2025 tcp_send_active_reset(sk, gfp_any()); 2026 sk->sk_err = ECONNRESET; 2027 } else if (old_state == TCP_SYN_SENT) 2028 sk->sk_err = ECONNRESET; 2029 2030 tcp_clear_xmit_timers(sk); 2031 __skb_queue_purge(&sk->sk_receive_queue); 2032 tcp_write_queue_purge(sk); 2033 __skb_queue_purge(&tp->out_of_order_queue); 2034#ifdef CONFIG_NET_DMA 2035 __skb_queue_purge(&sk->sk_async_wait_queue); 2036#endif 2037 2038 inet->dport = 0; 2039 2040 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK)) 2041 inet_reset_saddr(sk); 2042 2043 sk->sk_shutdown = 0; 2044 sock_reset_flag(sk, SOCK_DONE); 2045 tp->srtt = 0; 2046 if ((tp->write_seq += tp->max_window + 2) == 0) 2047 tp->write_seq = 1; 2048 icsk->icsk_backoff = 0; 2049 tp->snd_cwnd = 2; 2050 icsk->icsk_probes_out = 0; 2051 tp->packets_out = 0; 2052 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 2053 tp->snd_cwnd_cnt = 0; 2054 tp->bytes_acked = 0; 2055 tcp_set_ca_state(sk, TCP_CA_Open); 2056 tcp_clear_retrans(tp); 2057 inet_csk_delack_init(sk); 2058 tcp_init_send_head(sk); 2059 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt)); 2060 __sk_dst_reset(sk); 2061 2062 WARN_ON(inet->num && !icsk->icsk_bind_hash); 2063 2064 sk->sk_error_report(sk); 2065 return err; 2066} 2067 2068/* 2069 * Socket option code for TCP. 2070 */ 2071static int do_tcp_setsockopt(struct sock *sk, int level, 2072 int optname, char __user *optval, unsigned int optlen) 2073{ 2074 struct tcp_sock *tp = tcp_sk(sk); 2075 struct inet_connection_sock *icsk = inet_csk(sk); 2076 int val; 2077 int err = 0; 2078 2079 /* This is a string value all the others are int's */ 2080 if (optname == TCP_CONGESTION) { 2081 char name[TCP_CA_NAME_MAX]; 2082 2083 if (optlen < 1) 2084 return -EINVAL; 2085 2086 val = strncpy_from_user(name, optval, 2087 min_t(long, TCP_CA_NAME_MAX-1, optlen)); 2088 if (val < 0) 2089 return -EFAULT; 2090 name[val] = 0; 2091 2092 lock_sock(sk); 2093 err = tcp_set_congestion_control(sk, name); 2094 release_sock(sk); 2095 return err; 2096 } 2097 2098 if (optlen < sizeof(int)) 2099 return -EINVAL; 2100 2101 if (get_user(val, (int __user *)optval)) 2102 return -EFAULT; 2103 2104 lock_sock(sk); 2105 2106 switch (optname) { 2107 case TCP_MAXSEG: 2108 /* Values greater than interface MTU won't take effect. However 2109 * at the point when this call is done we typically don't yet 2110 * know which interface is going to be used */ 2111 if (val < 8 || val > MAX_TCP_WINDOW) { 2112 err = -EINVAL; 2113 break; 2114 } 2115 tp->rx_opt.user_mss = val; 2116 break; 2117 2118 case TCP_NODELAY: 2119 if (val) { 2120 /* TCP_NODELAY is weaker than TCP_CORK, so that 2121 * this option on corked socket is remembered, but 2122 * it is not activated until cork is cleared. 2123 * 2124 * However, when TCP_NODELAY is set we make 2125 * an explicit push, which overrides even TCP_CORK 2126 * for currently queued segments. 2127 */ 2128 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH; 2129 tcp_push_pending_frames(sk); 2130 } else { 2131 tp->nonagle &= ~TCP_NAGLE_OFF; 2132 } 2133 break; 2134 2135 case TCP_CORK: 2136 /* When set indicates to always queue non-full frames. 2137 * Later the user clears this option and we transmit 2138 * any pending partial frames in the queue. This is 2139 * meant to be used alongside sendfile() to get properly 2140 * filled frames when the user (for example) must write 2141 * out headers with a write() call first and then use 2142 * sendfile to send out the data parts. 2143 * 2144 * TCP_CORK can be set together with TCP_NODELAY and it is 2145 * stronger than TCP_NODELAY. 2146 */ 2147 if (val) { 2148 tp->nonagle |= TCP_NAGLE_CORK; 2149 } else { 2150 tp->nonagle &= ~TCP_NAGLE_CORK; 2151 if (tp->nonagle&TCP_NAGLE_OFF) 2152 tp->nonagle |= TCP_NAGLE_PUSH; 2153 tcp_push_pending_frames(sk); 2154 } 2155 break; 2156 2157 case TCP_KEEPIDLE: 2158 if (val < 1 || val > MAX_TCP_KEEPIDLE) 2159 err = -EINVAL; 2160 else { 2161 tp->keepalive_time = val * HZ; 2162 if (sock_flag(sk, SOCK_KEEPOPEN) && 2163 !((1 << sk->sk_state) & 2164 (TCPF_CLOSE | TCPF_LISTEN))) { 2165 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp; 2166 if (tp->keepalive_time > elapsed) 2167 elapsed = tp->keepalive_time - elapsed; 2168 else 2169 elapsed = 0; 2170 inet_csk_reset_keepalive_timer(sk, elapsed); 2171 } 2172 } 2173 break; 2174 case TCP_KEEPINTVL: 2175 if (val < 1 || val > MAX_TCP_KEEPINTVL) 2176 err = -EINVAL; 2177 else 2178 tp->keepalive_intvl = val * HZ; 2179 break; 2180 case TCP_KEEPCNT: 2181 if (val < 1 || val > MAX_TCP_KEEPCNT) 2182 err = -EINVAL; 2183 else 2184 tp->keepalive_probes = val; 2185 break; 2186 case TCP_SYNCNT: 2187 if (val < 1 || val > MAX_TCP_SYNCNT) 2188 err = -EINVAL; 2189 else 2190 icsk->icsk_syn_retries = val; 2191 break; 2192 2193 case TCP_LINGER2: 2194 if (val < 0) 2195 tp->linger2 = -1; 2196 else if (val > sysctl_tcp_fin_timeout / HZ) 2197 tp->linger2 = 0; 2198 else 2199 tp->linger2 = val * HZ; 2200 break; 2201 2202 case TCP_DEFER_ACCEPT: 2203 /* Translate value in seconds to number of retransmits */ 2204 icsk->icsk_accept_queue.rskq_defer_accept = 2205 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ, 2206 TCP_RTO_MAX / HZ); 2207 break; 2208 2209 case TCP_WINDOW_CLAMP: 2210 if (!val) { 2211 if (sk->sk_state != TCP_CLOSE) { 2212 err = -EINVAL; 2213 break; 2214 } 2215 tp->window_clamp = 0; 2216 } else 2217 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ? 2218 SOCK_MIN_RCVBUF / 2 : val; 2219 break; 2220 2221 case TCP_QUICKACK: 2222 if (!val) { 2223 icsk->icsk_ack.pingpong = 1; 2224 } else { 2225 icsk->icsk_ack.pingpong = 0; 2226 if ((1 << sk->sk_state) & 2227 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) && 2228 inet_csk_ack_scheduled(sk)) { 2229 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 2230 tcp_cleanup_rbuf(sk, 1); 2231 if (!(val & 1)) 2232 icsk->icsk_ack.pingpong = 1; 2233 } 2234 } 2235 break; 2236 2237#ifdef CONFIG_TCP_MD5SIG 2238 case TCP_MD5SIG: 2239 /* Read the IP->Key mappings from userspace */ 2240 err = tp->af_specific->md5_parse(sk, optval, optlen); 2241 break; 2242#endif 2243 2244 default: 2245 err = -ENOPROTOOPT; 2246 break; 2247 } 2248 2249 release_sock(sk); 2250 return err; 2251} 2252 2253int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval, 2254 unsigned int optlen) 2255{ 2256 struct inet_connection_sock *icsk = inet_csk(sk); 2257 2258 if (level != SOL_TCP) 2259 return icsk->icsk_af_ops->setsockopt(sk, level, optname, 2260 optval, optlen); 2261 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2262} 2263 2264#ifdef CONFIG_COMPAT 2265int compat_tcp_setsockopt(struct sock *sk, int level, int optname, 2266 char __user *optval, unsigned int optlen) 2267{ 2268 if (level != SOL_TCP) 2269 return inet_csk_compat_setsockopt(sk, level, optname, 2270 optval, optlen); 2271 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2272} 2273 2274EXPORT_SYMBOL(compat_tcp_setsockopt); 2275#endif 2276 2277/* Return information about state of tcp endpoint in API format. */ 2278void tcp_get_info(struct sock *sk, struct tcp_info *info) 2279{ 2280 struct tcp_sock *tp = tcp_sk(sk); 2281 const struct inet_connection_sock *icsk = inet_csk(sk); 2282 u32 now = tcp_time_stamp; 2283 2284 memset(info, 0, sizeof(*info)); 2285 2286 info->tcpi_state = sk->sk_state; 2287 info->tcpi_ca_state = icsk->icsk_ca_state; 2288 info->tcpi_retransmits = icsk->icsk_retransmits; 2289 info->tcpi_probes = icsk->icsk_probes_out; 2290 info->tcpi_backoff = icsk->icsk_backoff; 2291 2292 if (tp->rx_opt.tstamp_ok) 2293 info->tcpi_options |= TCPI_OPT_TIMESTAMPS; 2294 if (tcp_is_sack(tp)) 2295 info->tcpi_options |= TCPI_OPT_SACK; 2296 if (tp->rx_opt.wscale_ok) { 2297 info->tcpi_options |= TCPI_OPT_WSCALE; 2298 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale; 2299 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale; 2300 } 2301 2302 if (tp->ecn_flags&TCP_ECN_OK) 2303 info->tcpi_options |= TCPI_OPT_ECN; 2304 2305 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto); 2306 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato); 2307 info->tcpi_snd_mss = tp->mss_cache; 2308 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss; 2309 2310 if (sk->sk_state == TCP_LISTEN) { 2311 info->tcpi_unacked = sk->sk_ack_backlog; 2312 info->tcpi_sacked = sk->sk_max_ack_backlog; 2313 } else { 2314 info->tcpi_unacked = tp->packets_out; 2315 info->tcpi_sacked = tp->sacked_out; 2316 } 2317 info->tcpi_lost = tp->lost_out; 2318 info->tcpi_retrans = tp->retrans_out; 2319 info->tcpi_fackets = tp->fackets_out; 2320 2321 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime); 2322 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime); 2323 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp); 2324 2325 info->tcpi_pmtu = icsk->icsk_pmtu_cookie; 2326 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh; 2327 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3; 2328 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2; 2329 info->tcpi_snd_ssthresh = tp->snd_ssthresh; 2330 info->tcpi_snd_cwnd = tp->snd_cwnd; 2331 info->tcpi_advmss = tp->advmss; 2332 info->tcpi_reordering = tp->reordering; 2333 2334 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3; 2335 info->tcpi_rcv_space = tp->rcvq_space.space; 2336 2337 info->tcpi_total_retrans = tp->total_retrans; 2338} 2339 2340EXPORT_SYMBOL_GPL(tcp_get_info); 2341 2342static int do_tcp_getsockopt(struct sock *sk, int level, 2343 int optname, char __user *optval, int __user *optlen) 2344{ 2345 struct inet_connection_sock *icsk = inet_csk(sk); 2346 struct tcp_sock *tp = tcp_sk(sk); 2347 int val, len; 2348 2349 if (get_user(len, optlen)) 2350 return -EFAULT; 2351 2352 len = min_t(unsigned int, len, sizeof(int)); 2353 2354 if (len < 0) 2355 return -EINVAL; 2356 2357 switch (optname) { 2358 case TCP_MAXSEG: 2359 val = tp->mss_cache; 2360 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 2361 val = tp->rx_opt.user_mss; 2362 break; 2363 case TCP_NODELAY: 2364 val = !!(tp->nonagle&TCP_NAGLE_OFF); 2365 break; 2366 case TCP_CORK: 2367 val = !!(tp->nonagle&TCP_NAGLE_CORK); 2368 break; 2369 case TCP_KEEPIDLE: 2370 val = keepalive_time_when(tp) / HZ; 2371 break; 2372 case TCP_KEEPINTVL: 2373 val = keepalive_intvl_when(tp) / HZ; 2374 break; 2375 case TCP_KEEPCNT: 2376 val = keepalive_probes(tp); 2377 break; 2378 case TCP_SYNCNT: 2379 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries; 2380 break; 2381 case TCP_LINGER2: 2382 val = tp->linger2; 2383 if (val >= 0) 2384 val = (val ? : sysctl_tcp_fin_timeout) / HZ; 2385 break; 2386 case TCP_DEFER_ACCEPT: 2387 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept, 2388 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ); 2389 break; 2390 case TCP_WINDOW_CLAMP: 2391 val = tp->window_clamp; 2392 break; 2393 case TCP_INFO: { 2394 struct tcp_info info; 2395 2396 if (get_user(len, optlen)) 2397 return -EFAULT; 2398 2399 tcp_get_info(sk, &info); 2400 2401 len = min_t(unsigned int, len, sizeof(info)); 2402 if (put_user(len, optlen)) 2403 return -EFAULT; 2404 if (copy_to_user(optval, &info, len)) 2405 return -EFAULT; 2406 return 0; 2407 } 2408 case TCP_QUICKACK: 2409 val = !icsk->icsk_ack.pingpong; 2410 break; 2411 2412 case TCP_CONGESTION: 2413 if (get_user(len, optlen)) 2414 return -EFAULT; 2415 len = min_t(unsigned int, len, TCP_CA_NAME_MAX); 2416 if (put_user(len, optlen)) 2417 return -EFAULT; 2418 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len)) 2419 return -EFAULT; 2420 return 0; 2421 default: 2422 return -ENOPROTOOPT; 2423 } 2424 2425 if (put_user(len, optlen)) 2426 return -EFAULT; 2427 if (copy_to_user(optval, &val, len)) 2428 return -EFAULT; 2429 return 0; 2430} 2431 2432int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval, 2433 int __user *optlen) 2434{ 2435 struct inet_connection_sock *icsk = inet_csk(sk); 2436 2437 if (level != SOL_TCP) 2438 return icsk->icsk_af_ops->getsockopt(sk, level, optname, 2439 optval, optlen); 2440 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2441} 2442 2443#ifdef CONFIG_COMPAT 2444int compat_tcp_getsockopt(struct sock *sk, int level, int optname, 2445 char __user *optval, int __user *optlen) 2446{ 2447 if (level != SOL_TCP) 2448 return inet_csk_compat_getsockopt(sk, level, optname, 2449 optval, optlen); 2450 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2451} 2452 2453EXPORT_SYMBOL(compat_tcp_getsockopt); 2454#endif 2455 2456struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features) 2457{ 2458 struct sk_buff *segs = ERR_PTR(-EINVAL); 2459 struct tcphdr *th; 2460 unsigned thlen; 2461 unsigned int seq; 2462 __be32 delta; 2463 unsigned int oldlen; 2464 unsigned int mss; 2465 2466 if (!pskb_may_pull(skb, sizeof(*th))) 2467 goto out; 2468 2469 th = tcp_hdr(skb); 2470 thlen = th->doff * 4; 2471 if (thlen < sizeof(*th)) 2472 goto out; 2473 2474 if (!pskb_may_pull(skb, thlen)) 2475 goto out; 2476 2477 oldlen = (u16)~skb->len; 2478 __skb_pull(skb, thlen); 2479 2480 mss = skb_shinfo(skb)->gso_size; 2481 if (unlikely(skb->len <= mss)) 2482 goto out; 2483 2484 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) { 2485 /* Packet is from an untrusted source, reset gso_segs. */ 2486 int type = skb_shinfo(skb)->gso_type; 2487 2488 if (unlikely(type & 2489 ~(SKB_GSO_TCPV4 | 2490 SKB_GSO_DODGY | 2491 SKB_GSO_TCP_ECN | 2492 SKB_GSO_TCPV6 | 2493 0) || 2494 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))) 2495 goto out; 2496 2497 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss); 2498 2499 segs = NULL; 2500 goto out; 2501 } 2502 2503 segs = skb_segment(skb, features); 2504 if (IS_ERR(segs)) 2505 goto out; 2506 2507 delta = htonl(oldlen + (thlen + mss)); 2508 2509 skb = segs; 2510 th = tcp_hdr(skb); 2511 seq = ntohl(th->seq); 2512 2513 do { 2514 th->fin = th->psh = 0; 2515 2516 th->check = ~csum_fold((__force __wsum)((__force u32)th->check + 2517 (__force u32)delta)); 2518 if (skb->ip_summed != CHECKSUM_PARTIAL) 2519 th->check = 2520 csum_fold(csum_partial(skb_transport_header(skb), 2521 thlen, skb->csum)); 2522 2523 seq += mss; 2524 skb = skb->next; 2525 th = tcp_hdr(skb); 2526 2527 th->seq = htonl(seq); 2528 th->cwr = 0; 2529 } while (skb->next); 2530 2531 delta = htonl(oldlen + (skb->tail - skb->transport_header) + 2532 skb->data_len); 2533 th->check = ~csum_fold((__force __wsum)((__force u32)th->check + 2534 (__force u32)delta)); 2535 if (skb->ip_summed != CHECKSUM_PARTIAL) 2536 th->check = csum_fold(csum_partial(skb_transport_header(skb), 2537 thlen, skb->csum)); 2538 2539out: 2540 return segs; 2541} 2542EXPORT_SYMBOL(tcp_tso_segment); 2543 2544struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb) 2545{ 2546 struct sk_buff **pp = NULL; 2547 struct sk_buff *p; 2548 struct tcphdr *th; 2549 struct tcphdr *th2; 2550 unsigned int len; 2551 unsigned int thlen; 2552 unsigned int flags; 2553 unsigned int mss = 1; 2554 unsigned int hlen; 2555 unsigned int off; 2556 int flush = 1; 2557 int i; 2558 2559 off = skb_gro_offset(skb); 2560 hlen = off + sizeof(*th); 2561 th = skb_gro_header_fast(skb, off); 2562 if (skb_gro_header_hard(skb, hlen)) { 2563 th = skb_gro_header_slow(skb, hlen, off); 2564 if (unlikely(!th)) 2565 goto out; 2566 } 2567 2568 thlen = th->doff * 4; 2569 if (thlen < sizeof(*th)) 2570 goto out; 2571 2572 hlen = off + thlen; 2573 if (skb_gro_header_hard(skb, hlen)) { 2574 th = skb_gro_header_slow(skb, hlen, off); 2575 if (unlikely(!th)) 2576 goto out; 2577 } 2578 2579 skb_gro_pull(skb, thlen); 2580 2581 len = skb_gro_len(skb); 2582 flags = tcp_flag_word(th); 2583 2584 for (; (p = *head); head = &p->next) { 2585 if (!NAPI_GRO_CB(p)->same_flow) 2586 continue; 2587 2588 th2 = tcp_hdr(p); 2589 2590 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) { 2591 NAPI_GRO_CB(p)->same_flow = 0; 2592 continue; 2593 } 2594 2595 goto found; 2596 } 2597 2598 goto out_check_final; 2599 2600found: 2601 flush = NAPI_GRO_CB(p)->flush; 2602 flush |= flags & TCP_FLAG_CWR; 2603 flush |= (flags ^ tcp_flag_word(th2)) & 2604 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH); 2605 flush |= th->ack_seq ^ th2->ack_seq; 2606 for (i = sizeof(*th); i < thlen; i += 4) 2607 flush |= *(u32 *)((u8 *)th + i) ^ 2608 *(u32 *)((u8 *)th2 + i); 2609 2610 mss = skb_shinfo(p)->gso_size; 2611 2612 flush |= (len - 1) >= mss; 2613 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq); 2614 2615 if (flush || skb_gro_receive(head, skb)) { 2616 mss = 1; 2617 goto out_check_final; 2618 } 2619 2620 p = *head; 2621 th2 = tcp_hdr(p); 2622 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH); 2623 2624out_check_final: 2625 flush = len < mss; 2626 flush |= flags & (TCP_FLAG_URG | TCP_FLAG_PSH | TCP_FLAG_RST | 2627 TCP_FLAG_SYN | TCP_FLAG_FIN); 2628 2629 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush)) 2630 pp = head; 2631 2632out: 2633 NAPI_GRO_CB(skb)->flush |= flush; 2634 2635 return pp; 2636} 2637EXPORT_SYMBOL(tcp_gro_receive); 2638 2639int tcp_gro_complete(struct sk_buff *skb) 2640{ 2641 struct tcphdr *th = tcp_hdr(skb); 2642 2643 skb->csum_start = skb_transport_header(skb) - skb->head; 2644 skb->csum_offset = offsetof(struct tcphdr, check); 2645 skb->ip_summed = CHECKSUM_PARTIAL; 2646 2647 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count; 2648 2649 if (th->cwr) 2650 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 2651 2652 return 0; 2653} 2654EXPORT_SYMBOL(tcp_gro_complete); 2655 2656#ifdef CONFIG_TCP_MD5SIG 2657static unsigned long tcp_md5sig_users; 2658static struct tcp_md5sig_pool **tcp_md5sig_pool; 2659static DEFINE_SPINLOCK(tcp_md5sig_pool_lock); 2660 2661static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool) 2662{ 2663 int cpu; 2664 for_each_possible_cpu(cpu) { 2665 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu); 2666 if (p) { 2667 if (p->md5_desc.tfm) 2668 crypto_free_hash(p->md5_desc.tfm); 2669 kfree(p); 2670 p = NULL; 2671 } 2672 } 2673 free_percpu(pool); 2674} 2675 2676void tcp_free_md5sig_pool(void) 2677{ 2678 struct tcp_md5sig_pool **pool = NULL; 2679 2680 spin_lock_bh(&tcp_md5sig_pool_lock); 2681 if (--tcp_md5sig_users == 0) { 2682 pool = tcp_md5sig_pool; 2683 tcp_md5sig_pool = NULL; 2684 } 2685 spin_unlock_bh(&tcp_md5sig_pool_lock); 2686 if (pool) 2687 __tcp_free_md5sig_pool(pool); 2688} 2689 2690EXPORT_SYMBOL(tcp_free_md5sig_pool); 2691 2692static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(struct sock *sk) 2693{ 2694 int cpu; 2695 struct tcp_md5sig_pool **pool; 2696 2697 pool = alloc_percpu(struct tcp_md5sig_pool *); 2698 if (!pool) 2699 return NULL; 2700 2701 for_each_possible_cpu(cpu) { 2702 struct tcp_md5sig_pool *p; 2703 struct crypto_hash *hash; 2704 2705 p = kzalloc(sizeof(*p), sk->sk_allocation); 2706 if (!p) 2707 goto out_free; 2708 *per_cpu_ptr(pool, cpu) = p; 2709 2710 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC); 2711 if (!hash || IS_ERR(hash)) 2712 goto out_free; 2713 2714 p->md5_desc.tfm = hash; 2715 } 2716 return pool; 2717out_free: 2718 __tcp_free_md5sig_pool(pool); 2719 return NULL; 2720} 2721 2722struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(struct sock *sk) 2723{ 2724 struct tcp_md5sig_pool **pool; 2725 int alloc = 0; 2726 2727retry: 2728 spin_lock_bh(&tcp_md5sig_pool_lock); 2729 pool = tcp_md5sig_pool; 2730 if (tcp_md5sig_users++ == 0) { 2731 alloc = 1; 2732 spin_unlock_bh(&tcp_md5sig_pool_lock); 2733 } else if (!pool) { 2734 tcp_md5sig_users--; 2735 spin_unlock_bh(&tcp_md5sig_pool_lock); 2736 cpu_relax(); 2737 goto retry; 2738 } else 2739 spin_unlock_bh(&tcp_md5sig_pool_lock); 2740 2741 if (alloc) { 2742 /* we cannot hold spinlock here because this may sleep. */ 2743 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool(sk); 2744 spin_lock_bh(&tcp_md5sig_pool_lock); 2745 if (!p) { 2746 tcp_md5sig_users--; 2747 spin_unlock_bh(&tcp_md5sig_pool_lock); 2748 return NULL; 2749 } 2750 pool = tcp_md5sig_pool; 2751 if (pool) { 2752 /* oops, it has already been assigned. */ 2753 spin_unlock_bh(&tcp_md5sig_pool_lock); 2754 __tcp_free_md5sig_pool(p); 2755 } else { 2756 tcp_md5sig_pool = pool = p; 2757 spin_unlock_bh(&tcp_md5sig_pool_lock); 2758 } 2759 } 2760 return pool; 2761} 2762 2763EXPORT_SYMBOL(tcp_alloc_md5sig_pool); 2764 2765struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu) 2766{ 2767 struct tcp_md5sig_pool **p; 2768 spin_lock_bh(&tcp_md5sig_pool_lock); 2769 p = tcp_md5sig_pool; 2770 if (p) 2771 tcp_md5sig_users++; 2772 spin_unlock_bh(&tcp_md5sig_pool_lock); 2773 return (p ? *per_cpu_ptr(p, cpu) : NULL); 2774} 2775 2776EXPORT_SYMBOL(__tcp_get_md5sig_pool); 2777 2778void __tcp_put_md5sig_pool(void) 2779{ 2780 tcp_free_md5sig_pool(); 2781} 2782 2783EXPORT_SYMBOL(__tcp_put_md5sig_pool); 2784 2785int tcp_md5_hash_header(struct tcp_md5sig_pool *hp, 2786 struct tcphdr *th) 2787{ 2788 struct scatterlist sg; 2789 int err; 2790 2791 __sum16 old_checksum = th->check; 2792 th->check = 0; 2793 /* options aren't included in the hash */ 2794 sg_init_one(&sg, th, sizeof(struct tcphdr)); 2795 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr)); 2796 th->check = old_checksum; 2797 return err; 2798} 2799 2800EXPORT_SYMBOL(tcp_md5_hash_header); 2801 2802int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp, 2803 struct sk_buff *skb, unsigned header_len) 2804{ 2805 struct scatterlist sg; 2806 const struct tcphdr *tp = tcp_hdr(skb); 2807 struct hash_desc *desc = &hp->md5_desc; 2808 unsigned i; 2809 const unsigned head_data_len = skb_headlen(skb) > header_len ? 2810 skb_headlen(skb) - header_len : 0; 2811 const struct skb_shared_info *shi = skb_shinfo(skb); 2812 2813 sg_init_table(&sg, 1); 2814 2815 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len); 2816 if (crypto_hash_update(desc, &sg, head_data_len)) 2817 return 1; 2818 2819 for (i = 0; i < shi->nr_frags; ++i) { 2820 const struct skb_frag_struct *f = &shi->frags[i]; 2821 sg_set_page(&sg, f->page, f->size, f->page_offset); 2822 if (crypto_hash_update(desc, &sg, f->size)) 2823 return 1; 2824 } 2825 2826 return 0; 2827} 2828 2829EXPORT_SYMBOL(tcp_md5_hash_skb_data); 2830 2831int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key) 2832{ 2833 struct scatterlist sg; 2834 2835 sg_init_one(&sg, key->key, key->keylen); 2836 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen); 2837} 2838 2839EXPORT_SYMBOL(tcp_md5_hash_key); 2840 2841#endif 2842 2843void tcp_done(struct sock *sk) 2844{ 2845 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV) 2846 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 2847 2848 tcp_set_state(sk, TCP_CLOSE); 2849 tcp_clear_xmit_timers(sk); 2850 2851 sk->sk_shutdown = SHUTDOWN_MASK; 2852 2853 if (!sock_flag(sk, SOCK_DEAD)) 2854 sk->sk_state_change(sk); 2855 else 2856 inet_csk_destroy_sock(sk); 2857} 2858EXPORT_SYMBOL_GPL(tcp_done); 2859 2860extern struct tcp_congestion_ops tcp_reno; 2861 2862static __initdata unsigned long thash_entries; 2863static int __init set_thash_entries(char *str) 2864{ 2865 if (!str) 2866 return 0; 2867 thash_entries = simple_strtoul(str, &str, 0); 2868 return 1; 2869} 2870__setup("thash_entries=", set_thash_entries); 2871 2872void __init tcp_init(void) 2873{ 2874 struct sk_buff *skb = NULL; 2875 unsigned long nr_pages, limit; 2876 int order, i, max_share; 2877 2878 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb)); 2879 2880 percpu_counter_init(&tcp_sockets_allocated, 0); 2881 percpu_counter_init(&tcp_orphan_count, 0); 2882 tcp_hashinfo.bind_bucket_cachep = 2883 kmem_cache_create("tcp_bind_bucket", 2884 sizeof(struct inet_bind_bucket), 0, 2885 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL); 2886 2887 /* Size and allocate the main established and bind bucket 2888 * hash tables. 2889 * 2890 * The methodology is similar to that of the buffer cache. 2891 */ 2892 tcp_hashinfo.ehash = 2893 alloc_large_system_hash("TCP established", 2894 sizeof(struct inet_ehash_bucket), 2895 thash_entries, 2896 (totalram_pages >= 128 * 1024) ? 2897 13 : 15, 2898 0, 2899 &tcp_hashinfo.ehash_size, 2900 NULL, 2901 thash_entries ? 0 : 512 * 1024); 2902 tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size; 2903 for (i = 0; i < tcp_hashinfo.ehash_size; i++) { 2904 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i); 2905 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i); 2906 } 2907 if (inet_ehash_locks_alloc(&tcp_hashinfo)) 2908 panic("TCP: failed to alloc ehash_locks"); 2909 tcp_hashinfo.bhash = 2910 alloc_large_system_hash("TCP bind", 2911 sizeof(struct inet_bind_hashbucket), 2912 tcp_hashinfo.ehash_size, 2913 (totalram_pages >= 128 * 1024) ? 2914 13 : 15, 2915 0, 2916 &tcp_hashinfo.bhash_size, 2917 NULL, 2918 64 * 1024); 2919 tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size; 2920 for (i = 0; i < tcp_hashinfo.bhash_size; i++) { 2921 spin_lock_init(&tcp_hashinfo.bhash[i].lock); 2922 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain); 2923 } 2924 2925 /* Try to be a bit smarter and adjust defaults depending 2926 * on available memory. 2927 */ 2928 for (order = 0; ((1 << order) << PAGE_SHIFT) < 2929 (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket)); 2930 order++) 2931 ; 2932 if (order >= 4) { 2933 tcp_death_row.sysctl_max_tw_buckets = 180000; 2934 sysctl_tcp_max_orphans = 4096 << (order - 4); 2935 sysctl_max_syn_backlog = 1024; 2936 } else if (order < 3) { 2937 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order); 2938 sysctl_tcp_max_orphans >>= (3 - order); 2939 sysctl_max_syn_backlog = 128; 2940 } 2941 2942 /* Set the pressure threshold to be a fraction of global memory that 2943 * is up to 1/2 at 256 MB, decreasing toward zero with the amount of 2944 * memory, with a floor of 128 pages. 2945 */ 2946 nr_pages = totalram_pages - totalhigh_pages; 2947 limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT); 2948 limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11); 2949 limit = max(limit, 128UL); 2950 sysctl_tcp_mem[0] = limit / 4 * 3; 2951 sysctl_tcp_mem[1] = limit; 2952 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; 2953 2954 /* Set per-socket limits to no more than 1/128 the pressure threshold */ 2955 limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7); 2956 max_share = min(4UL*1024*1024, limit); 2957 2958 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM; 2959 sysctl_tcp_wmem[1] = 16*1024; 2960 sysctl_tcp_wmem[2] = max(64*1024, max_share); 2961 2962 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM; 2963 sysctl_tcp_rmem[1] = 87380; 2964 sysctl_tcp_rmem[2] = max(87380, max_share); 2965 2966 printk(KERN_INFO "TCP: Hash tables configured " 2967 "(established %d bind %d)\n", 2968 tcp_hashinfo.ehash_size, tcp_hashinfo.bhash_size); 2969 2970 tcp_register_congestion_control(&tcp_reno); 2971} 2972 2973EXPORT_SYMBOL(tcp_close); 2974EXPORT_SYMBOL(tcp_disconnect); 2975EXPORT_SYMBOL(tcp_getsockopt); 2976EXPORT_SYMBOL(tcp_ioctl); 2977EXPORT_SYMBOL(tcp_poll); 2978EXPORT_SYMBOL(tcp_read_sock); 2979EXPORT_SYMBOL(tcp_recvmsg); 2980EXPORT_SYMBOL(tcp_sendmsg); 2981EXPORT_SYMBOL(tcp_splice_read); 2982EXPORT_SYMBOL(tcp_sendpage); 2983EXPORT_SYMBOL(tcp_setsockopt); 2984EXPORT_SYMBOL(tcp_shutdown); 2985