tcp.c revision b139ba4e90dccbf4cd4efb112af96a5c9e0b098c
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/init.h> 256#include <linux/fs.h> 257#include <linux/skbuff.h> 258#include <linux/scatterlist.h> 259#include <linux/splice.h> 260#include <linux/net.h> 261#include <linux/socket.h> 262#include <linux/random.h> 263#include <linux/bootmem.h> 264#include <linux/highmem.h> 265#include <linux/swap.h> 266#include <linux/cache.h> 267#include <linux/err.h> 268#include <linux/crypto.h> 269#include <linux/time.h> 270#include <linux/slab.h> 271 272#include <net/icmp.h> 273#include <net/tcp.h> 274#include <net/xfrm.h> 275#include <net/ip.h> 276#include <net/netdma.h> 277#include <net/sock.h> 278 279#include <asm/uaccess.h> 280#include <asm/ioctls.h> 281 282int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT; 283 284struct percpu_counter tcp_orphan_count; 285EXPORT_SYMBOL_GPL(tcp_orphan_count); 286 287int sysctl_tcp_wmem[3] __read_mostly; 288int sysctl_tcp_rmem[3] __read_mostly; 289 290EXPORT_SYMBOL(sysctl_tcp_rmem); 291EXPORT_SYMBOL(sysctl_tcp_wmem); 292 293atomic_long_t tcp_memory_allocated; /* Current allocated memory. */ 294EXPORT_SYMBOL(tcp_memory_allocated); 295 296/* 297 * Current number of TCP sockets. 298 */ 299struct percpu_counter tcp_sockets_allocated; 300EXPORT_SYMBOL(tcp_sockets_allocated); 301 302/* 303 * TCP splice context 304 */ 305struct tcp_splice_state { 306 struct pipe_inode_info *pipe; 307 size_t len; 308 unsigned int flags; 309}; 310 311/* 312 * Pressure flag: try to collapse. 313 * Technical note: it is used by multiple contexts non atomically. 314 * All the __sk_mem_schedule() is of this nature: accounting 315 * is strict, actions are advisory and have some latency. 316 */ 317int tcp_memory_pressure __read_mostly; 318EXPORT_SYMBOL(tcp_memory_pressure); 319 320void tcp_enter_memory_pressure(struct sock *sk) 321{ 322 if (!tcp_memory_pressure) { 323 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES); 324 tcp_memory_pressure = 1; 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 const struct tcp_sock *tp = tcp_sk(sk); 378 379 sock_poll_wait(file, sk_sleep(sk), 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; 389 390 /* 391 * POLLHUP is certainly not done right. But poll() doesn't 392 * have a notion of HUP in just one direction, and for a 393 * socket the read side is more interesting. 394 * 395 * Some poll() documentation says that POLLHUP is incompatible 396 * with the POLLOUT/POLLWR flags, so somebody should check this 397 * all. But careful, it tends to be safer to return too many 398 * bits than too few, and you can easily break real applications 399 * if you don't tell them that something has hung up! 400 * 401 * Check-me. 402 * 403 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and 404 * our fs/select.c). It means that after we received EOF, 405 * poll always returns immediately, making impossible poll() on write() 406 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP 407 * if and only if shutdown has been made in both directions. 408 * Actually, it is interesting to look how Solaris and DUX 409 * solve this dilemma. I would prefer, if POLLHUP were maskable, 410 * then we could set it on SND_SHUTDOWN. BTW examples given 411 * in Stevens' books assume exactly this behaviour, it explains 412 * why POLLHUP is incompatible with POLLOUT. --ANK 413 * 414 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent 415 * blocking on fresh not-connected or disconnected socket. --ANK 416 */ 417 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE) 418 mask |= POLLHUP; 419 if (sk->sk_shutdown & RCV_SHUTDOWN) 420 mask |= POLLIN | POLLRDNORM | POLLRDHUP; 421 422 /* Connected? */ 423 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) { 424 int target = sock_rcvlowat(sk, 0, INT_MAX); 425 426 if (tp->urg_seq == tp->copied_seq && 427 !sock_flag(sk, SOCK_URGINLINE) && 428 tp->urg_data) 429 target++; 430 431 /* Potential race condition. If read of tp below will 432 * escape above sk->sk_state, we can be illegally awaken 433 * in SYN_* states. */ 434 if (tp->rcv_nxt - tp->copied_seq >= target) 435 mask |= POLLIN | POLLRDNORM; 436 437 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) { 438 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) { 439 mask |= POLLOUT | POLLWRNORM; 440 } else { /* send SIGIO later */ 441 set_bit(SOCK_ASYNC_NOSPACE, 442 &sk->sk_socket->flags); 443 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 444 445 /* Race breaker. If space is freed after 446 * wspace test but before the flags are set, 447 * IO signal will be lost. 448 */ 449 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) 450 mask |= POLLOUT | POLLWRNORM; 451 } 452 } else 453 mask |= POLLOUT | POLLWRNORM; 454 455 if (tp->urg_data & TCP_URG_VALID) 456 mask |= POLLPRI; 457 } 458 /* This barrier is coupled with smp_wmb() in tcp_reset() */ 459 smp_rmb(); 460 if (sk->sk_err) 461 mask |= POLLERR; 462 463 return mask; 464} 465EXPORT_SYMBOL(tcp_poll); 466 467int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg) 468{ 469 struct tcp_sock *tp = tcp_sk(sk); 470 int answ; 471 472 switch (cmd) { 473 case SIOCINQ: 474 if (sk->sk_state == TCP_LISTEN) 475 return -EINVAL; 476 477 lock_sock(sk); 478 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 479 answ = 0; 480 else if (sock_flag(sk, SOCK_URGINLINE) || 481 !tp->urg_data || 482 before(tp->urg_seq, tp->copied_seq) || 483 !before(tp->urg_seq, tp->rcv_nxt)) { 484 struct sk_buff *skb; 485 486 answ = tp->rcv_nxt - tp->copied_seq; 487 488 /* Subtract 1, if FIN is in queue. */ 489 skb = skb_peek_tail(&sk->sk_receive_queue); 490 if (answ && skb) 491 answ -= tcp_hdr(skb)->fin; 492 } else 493 answ = tp->urg_seq - tp->copied_seq; 494 release_sock(sk); 495 break; 496 case SIOCATMARK: 497 answ = tp->urg_data && tp->urg_seq == tp->copied_seq; 498 break; 499 case SIOCOUTQ: 500 if (sk->sk_state == TCP_LISTEN) 501 return -EINVAL; 502 503 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 504 answ = 0; 505 else 506 answ = tp->write_seq - tp->snd_una; 507 break; 508 case SIOCOUTQNSD: 509 if (sk->sk_state == TCP_LISTEN) 510 return -EINVAL; 511 512 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 513 answ = 0; 514 else 515 answ = tp->write_seq - tp->snd_nxt; 516 break; 517 default: 518 return -ENOIOCTLCMD; 519 } 520 521 return put_user(answ, (int __user *)arg); 522} 523EXPORT_SYMBOL(tcp_ioctl); 524 525static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb) 526{ 527 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 528 tp->pushed_seq = tp->write_seq; 529} 530 531static inline int forced_push(const struct tcp_sock *tp) 532{ 533 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1)); 534} 535 536static inline void skb_entail(struct sock *sk, struct sk_buff *skb) 537{ 538 struct tcp_sock *tp = tcp_sk(sk); 539 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 540 541 skb->csum = 0; 542 tcb->seq = tcb->end_seq = tp->write_seq; 543 tcb->tcp_flags = TCPHDR_ACK; 544 tcb->sacked = 0; 545 skb_header_release(skb); 546 tcp_add_write_queue_tail(sk, skb); 547 sk->sk_wmem_queued += skb->truesize; 548 sk_mem_charge(sk, skb->truesize); 549 if (tp->nonagle & TCP_NAGLE_PUSH) 550 tp->nonagle &= ~TCP_NAGLE_PUSH; 551} 552 553static inline void tcp_mark_urg(struct tcp_sock *tp, int flags) 554{ 555 if (flags & MSG_OOB) 556 tp->snd_up = tp->write_seq; 557} 558 559static inline void tcp_push(struct sock *sk, int flags, int mss_now, 560 int nonagle) 561{ 562 if (tcp_send_head(sk)) { 563 struct tcp_sock *tp = tcp_sk(sk); 564 565 if (!(flags & MSG_MORE) || forced_push(tp)) 566 tcp_mark_push(tp, tcp_write_queue_tail(sk)); 567 568 tcp_mark_urg(tp, flags); 569 __tcp_push_pending_frames(sk, mss_now, 570 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle); 571 } 572} 573 574static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, 575 unsigned int offset, size_t len) 576{ 577 struct tcp_splice_state *tss = rd_desc->arg.data; 578 int ret; 579 580 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len), 581 tss->flags); 582 if (ret > 0) 583 rd_desc->count -= ret; 584 return ret; 585} 586 587static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss) 588{ 589 /* Store TCP splice context information in read_descriptor_t. */ 590 read_descriptor_t rd_desc = { 591 .arg.data = tss, 592 .count = tss->len, 593 }; 594 595 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv); 596} 597 598/** 599 * tcp_splice_read - splice data from TCP socket to a pipe 600 * @sock: socket to splice from 601 * @ppos: position (not valid) 602 * @pipe: pipe to splice to 603 * @len: number of bytes to splice 604 * @flags: splice modifier flags 605 * 606 * Description: 607 * Will read pages from given socket and fill them into a pipe. 608 * 609 **/ 610ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos, 611 struct pipe_inode_info *pipe, size_t len, 612 unsigned int flags) 613{ 614 struct sock *sk = sock->sk; 615 struct tcp_splice_state tss = { 616 .pipe = pipe, 617 .len = len, 618 .flags = flags, 619 }; 620 long timeo; 621 ssize_t spliced; 622 int ret; 623 624 sock_rps_record_flow(sk); 625 /* 626 * We can't seek on a socket input 627 */ 628 if (unlikely(*ppos)) 629 return -ESPIPE; 630 631 ret = spliced = 0; 632 633 lock_sock(sk); 634 635 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK); 636 while (tss.len) { 637 ret = __tcp_splice_read(sk, &tss); 638 if (ret < 0) 639 break; 640 else if (!ret) { 641 if (spliced) 642 break; 643 if (sock_flag(sk, SOCK_DONE)) 644 break; 645 if (sk->sk_err) { 646 ret = sock_error(sk); 647 break; 648 } 649 if (sk->sk_shutdown & RCV_SHUTDOWN) 650 break; 651 if (sk->sk_state == TCP_CLOSE) { 652 /* 653 * This occurs when user tries to read 654 * from never connected socket. 655 */ 656 if (!sock_flag(sk, SOCK_DONE)) 657 ret = -ENOTCONN; 658 break; 659 } 660 if (!timeo) { 661 ret = -EAGAIN; 662 break; 663 } 664 sk_wait_data(sk, &timeo); 665 if (signal_pending(current)) { 666 ret = sock_intr_errno(timeo); 667 break; 668 } 669 continue; 670 } 671 tss.len -= ret; 672 spliced += ret; 673 674 if (!timeo) 675 break; 676 release_sock(sk); 677 lock_sock(sk); 678 679 if (sk->sk_err || sk->sk_state == TCP_CLOSE || 680 (sk->sk_shutdown & RCV_SHUTDOWN) || 681 signal_pending(current)) 682 break; 683 } 684 685 release_sock(sk); 686 687 if (spliced) 688 return spliced; 689 690 return ret; 691} 692EXPORT_SYMBOL(tcp_splice_read); 693 694struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp) 695{ 696 struct sk_buff *skb; 697 698 /* The TCP header must be at least 32-bit aligned. */ 699 size = ALIGN(size, 4); 700 701 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp); 702 if (skb) { 703 if (sk_wmem_schedule(sk, skb->truesize)) { 704 skb_reserve(skb, sk->sk_prot->max_header); 705 /* 706 * Make sure that we have exactly size bytes 707 * available to the caller, no more, no less. 708 */ 709 skb->avail_size = size; 710 return skb; 711 } 712 __kfree_skb(skb); 713 } else { 714 sk->sk_prot->enter_memory_pressure(sk); 715 sk_stream_moderate_sndbuf(sk); 716 } 717 return NULL; 718} 719 720static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now, 721 int large_allowed) 722{ 723 struct tcp_sock *tp = tcp_sk(sk); 724 u32 xmit_size_goal, old_size_goal; 725 726 xmit_size_goal = mss_now; 727 728 if (large_allowed && sk_can_gso(sk)) { 729 xmit_size_goal = ((sk->sk_gso_max_size - 1) - 730 inet_csk(sk)->icsk_af_ops->net_header_len - 731 inet_csk(sk)->icsk_ext_hdr_len - 732 tp->tcp_header_len); 733 734 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal); 735 736 /* We try hard to avoid divides here */ 737 old_size_goal = tp->xmit_size_goal_segs * mss_now; 738 739 if (likely(old_size_goal <= xmit_size_goal && 740 old_size_goal + mss_now > xmit_size_goal)) { 741 xmit_size_goal = old_size_goal; 742 } else { 743 tp->xmit_size_goal_segs = xmit_size_goal / mss_now; 744 xmit_size_goal = tp->xmit_size_goal_segs * mss_now; 745 } 746 } 747 748 return max(xmit_size_goal, mss_now); 749} 750 751static int tcp_send_mss(struct sock *sk, int *size_goal, int flags) 752{ 753 int mss_now; 754 755 mss_now = tcp_current_mss(sk); 756 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB)); 757 758 return mss_now; 759} 760 761static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset, 762 size_t psize, int flags) 763{ 764 struct tcp_sock *tp = tcp_sk(sk); 765 int mss_now, size_goal; 766 int err; 767 ssize_t copied; 768 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 769 770 /* Wait for a connection to finish. */ 771 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 772 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0) 773 goto out_err; 774 775 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags); 776 777 mss_now = tcp_send_mss(sk, &size_goal, flags); 778 copied = 0; 779 780 err = -EPIPE; 781 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 782 goto out_err; 783 784 while (psize > 0) { 785 struct sk_buff *skb = tcp_write_queue_tail(sk); 786 struct page *page = pages[poffset / PAGE_SIZE]; 787 int copy, i, can_coalesce; 788 int offset = poffset % PAGE_SIZE; 789 int size = min_t(size_t, psize, PAGE_SIZE - offset); 790 791 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) { 792new_segment: 793 if (!sk_stream_memory_free(sk)) 794 goto wait_for_sndbuf; 795 796 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation); 797 if (!skb) 798 goto wait_for_memory; 799 800 skb_entail(sk, skb); 801 copy = size_goal; 802 } 803 804 if (copy > size) 805 copy = size; 806 807 i = skb_shinfo(skb)->nr_frags; 808 can_coalesce = skb_can_coalesce(skb, i, page, offset); 809 if (!can_coalesce && i >= MAX_SKB_FRAGS) { 810 tcp_mark_push(tp, skb); 811 goto new_segment; 812 } 813 if (!sk_wmem_schedule(sk, copy)) 814 goto wait_for_memory; 815 816 if (can_coalesce) { 817 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 818 } else { 819 get_page(page); 820 skb_fill_page_desc(skb, i, page, offset, copy); 821 } 822 823 skb->len += copy; 824 skb->data_len += copy; 825 skb->truesize += copy; 826 sk->sk_wmem_queued += copy; 827 sk_mem_charge(sk, copy); 828 skb->ip_summed = CHECKSUM_PARTIAL; 829 tp->write_seq += copy; 830 TCP_SKB_CB(skb)->end_seq += copy; 831 skb_shinfo(skb)->gso_segs = 0; 832 833 if (!copied) 834 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 835 836 copied += copy; 837 poffset += copy; 838 if (!(psize -= copy)) 839 goto out; 840 841 if (skb->len < size_goal || (flags & MSG_OOB)) 842 continue; 843 844 if (forced_push(tp)) { 845 tcp_mark_push(tp, skb); 846 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 847 } else if (skb == tcp_send_head(sk)) 848 tcp_push_one(sk, mss_now); 849 continue; 850 851wait_for_sndbuf: 852 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 853wait_for_memory: 854 if (copied) 855 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH); 856 857 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0) 858 goto do_error; 859 860 mss_now = tcp_send_mss(sk, &size_goal, flags); 861 } 862 863out: 864 if (copied && !(flags & MSG_SENDPAGE_NOTLAST)) 865 tcp_push(sk, flags, mss_now, tp->nonagle); 866 return copied; 867 868do_error: 869 if (copied) 870 goto out; 871out_err: 872 return sk_stream_error(sk, flags, err); 873} 874 875int tcp_sendpage(struct sock *sk, struct page *page, int offset, 876 size_t size, int flags) 877{ 878 ssize_t res; 879 880 if (!(sk->sk_route_caps & NETIF_F_SG) || 881 !(sk->sk_route_caps & NETIF_F_ALL_CSUM)) 882 return sock_no_sendpage(sk->sk_socket, page, offset, size, 883 flags); 884 885 lock_sock(sk); 886 res = do_tcp_sendpages(sk, &page, offset, size, flags); 887 release_sock(sk); 888 return res; 889} 890EXPORT_SYMBOL(tcp_sendpage); 891 892static inline int select_size(const struct sock *sk, bool sg) 893{ 894 const struct tcp_sock *tp = tcp_sk(sk); 895 int tmp = tp->mss_cache; 896 897 if (sg) { 898 if (sk_can_gso(sk)) { 899 /* Small frames wont use a full page: 900 * Payload will immediately follow tcp header. 901 */ 902 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER); 903 } else { 904 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER); 905 906 if (tmp >= pgbreak && 907 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE) 908 tmp = pgbreak; 909 } 910 } 911 912 return tmp; 913} 914 915static int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size) 916{ 917 struct sk_buff *skb; 918 struct tcp_skb_cb *cb; 919 struct tcphdr *th; 920 921 skb = alloc_skb(size + sizeof(*th), sk->sk_allocation); 922 if (!skb) 923 goto err; 924 925 th = (struct tcphdr *)skb_put(skb, sizeof(*th)); 926 skb_reset_transport_header(skb); 927 memset(th, 0, sizeof(*th)); 928 929 if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size)) 930 goto err_free; 931 932 cb = TCP_SKB_CB(skb); 933 934 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt; 935 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size; 936 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1; 937 938 tcp_queue_rcv(sk, skb, sizeof(*th)); 939 940 return size; 941 942err_free: 943 kfree_skb(skb); 944err: 945 return -ENOMEM; 946} 947 948int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg, 949 size_t size) 950{ 951 struct iovec *iov; 952 struct tcp_sock *tp = tcp_sk(sk); 953 struct sk_buff *skb; 954 int iovlen, flags, err, copied; 955 int mss_now = 0, size_goal; 956 bool sg; 957 long timeo; 958 959 lock_sock(sk); 960 961 flags = msg->msg_flags; 962 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 963 964 /* Wait for a connection to finish. */ 965 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 966 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0) 967 goto out_err; 968 969 if (unlikely(tp->repair)) { 970 if (tp->repair_queue == TCP_RECV_QUEUE) { 971 copied = tcp_send_rcvq(sk, msg, size); 972 goto out; 973 } 974 975 err = -EINVAL; 976 if (tp->repair_queue == TCP_NO_QUEUE) 977 goto out_err; 978 979 /* 'common' sending to sendq */ 980 } 981 982 /* This should be in poll */ 983 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags); 984 985 mss_now = tcp_send_mss(sk, &size_goal, flags); 986 987 /* Ok commence sending. */ 988 iovlen = msg->msg_iovlen; 989 iov = msg->msg_iov; 990 copied = 0; 991 992 err = -EPIPE; 993 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 994 goto out_err; 995 996 sg = !!(sk->sk_route_caps & NETIF_F_SG); 997 998 while (--iovlen >= 0) { 999 size_t seglen = iov->iov_len; 1000 unsigned char __user *from = iov->iov_base; 1001 1002 iov++; 1003 1004 while (seglen > 0) { 1005 int copy = 0; 1006 int max = size_goal; 1007 1008 skb = tcp_write_queue_tail(sk); 1009 if (tcp_send_head(sk)) { 1010 if (skb->ip_summed == CHECKSUM_NONE) 1011 max = mss_now; 1012 copy = max - skb->len; 1013 } 1014 1015 if (copy <= 0) { 1016new_segment: 1017 /* Allocate new segment. If the interface is SG, 1018 * allocate skb fitting to single page. 1019 */ 1020 if (!sk_stream_memory_free(sk)) 1021 goto wait_for_sndbuf; 1022 1023 skb = sk_stream_alloc_skb(sk, 1024 select_size(sk, sg), 1025 sk->sk_allocation); 1026 if (!skb) 1027 goto wait_for_memory; 1028 1029 /* 1030 * Check whether we can use HW checksum. 1031 */ 1032 if (sk->sk_route_caps & NETIF_F_ALL_CSUM) 1033 skb->ip_summed = CHECKSUM_PARTIAL; 1034 1035 skb_entail(sk, skb); 1036 copy = size_goal; 1037 max = size_goal; 1038 } 1039 1040 /* Try to append data to the end of skb. */ 1041 if (copy > seglen) 1042 copy = seglen; 1043 1044 /* Where to copy to? */ 1045 if (skb_availroom(skb) > 0) { 1046 /* We have some space in skb head. Superb! */ 1047 copy = min_t(int, copy, skb_availroom(skb)); 1048 err = skb_add_data_nocache(sk, skb, from, copy); 1049 if (err) 1050 goto do_fault; 1051 } else { 1052 int merge = 0; 1053 int i = skb_shinfo(skb)->nr_frags; 1054 struct page *page = sk->sk_sndmsg_page; 1055 int off; 1056 1057 if (page && page_count(page) == 1) 1058 sk->sk_sndmsg_off = 0; 1059 1060 off = sk->sk_sndmsg_off; 1061 1062 if (skb_can_coalesce(skb, i, page, off) && 1063 off != PAGE_SIZE) { 1064 /* We can extend the last page 1065 * fragment. */ 1066 merge = 1; 1067 } else if (i == MAX_SKB_FRAGS || !sg) { 1068 /* Need to add new fragment and cannot 1069 * do this because interface is non-SG, 1070 * or because all the page slots are 1071 * busy. */ 1072 tcp_mark_push(tp, skb); 1073 goto new_segment; 1074 } else if (page) { 1075 if (off == PAGE_SIZE) { 1076 put_page(page); 1077 sk->sk_sndmsg_page = page = NULL; 1078 off = 0; 1079 } 1080 } else 1081 off = 0; 1082 1083 if (copy > PAGE_SIZE - off) 1084 copy = PAGE_SIZE - off; 1085 1086 if (!sk_wmem_schedule(sk, copy)) 1087 goto wait_for_memory; 1088 1089 if (!page) { 1090 /* Allocate new cache page. */ 1091 if (!(page = sk_stream_alloc_page(sk))) 1092 goto wait_for_memory; 1093 } 1094 1095 /* Time to copy data. We are close to 1096 * the end! */ 1097 err = skb_copy_to_page_nocache(sk, from, skb, 1098 page, off, copy); 1099 if (err) { 1100 /* If this page was new, give it to the 1101 * socket so it does not get leaked. 1102 */ 1103 if (!sk->sk_sndmsg_page) { 1104 sk->sk_sndmsg_page = page; 1105 sk->sk_sndmsg_off = 0; 1106 } 1107 goto do_error; 1108 } 1109 1110 /* Update the skb. */ 1111 if (merge) { 1112 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 1113 } else { 1114 skb_fill_page_desc(skb, i, page, off, copy); 1115 if (sk->sk_sndmsg_page) { 1116 get_page(page); 1117 } else if (off + copy < PAGE_SIZE) { 1118 get_page(page); 1119 sk->sk_sndmsg_page = page; 1120 } 1121 } 1122 1123 sk->sk_sndmsg_off = off + copy; 1124 } 1125 1126 if (!copied) 1127 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 1128 1129 tp->write_seq += copy; 1130 TCP_SKB_CB(skb)->end_seq += copy; 1131 skb_shinfo(skb)->gso_segs = 0; 1132 1133 from += copy; 1134 copied += copy; 1135 if ((seglen -= copy) == 0 && iovlen == 0) 1136 goto out; 1137 1138 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair)) 1139 continue; 1140 1141 if (forced_push(tp)) { 1142 tcp_mark_push(tp, skb); 1143 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 1144 } else if (skb == tcp_send_head(sk)) 1145 tcp_push_one(sk, mss_now); 1146 continue; 1147 1148wait_for_sndbuf: 1149 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1150wait_for_memory: 1151 if (copied && likely(!tp->repair)) 1152 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH); 1153 1154 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0) 1155 goto do_error; 1156 1157 mss_now = tcp_send_mss(sk, &size_goal, flags); 1158 } 1159 } 1160 1161out: 1162 if (copied && likely(!tp->repair)) 1163 tcp_push(sk, flags, mss_now, tp->nonagle); 1164 release_sock(sk); 1165 return copied; 1166 1167do_fault: 1168 if (!skb->len) { 1169 tcp_unlink_write_queue(skb, sk); 1170 /* It is the one place in all of TCP, except connection 1171 * reset, where we can be unlinking the send_head. 1172 */ 1173 tcp_check_send_head(sk, skb); 1174 sk_wmem_free_skb(sk, skb); 1175 } 1176 1177do_error: 1178 if (copied) 1179 goto out; 1180out_err: 1181 err = sk_stream_error(sk, flags, err); 1182 release_sock(sk); 1183 return err; 1184} 1185EXPORT_SYMBOL(tcp_sendmsg); 1186 1187/* 1188 * Handle reading urgent data. BSD has very simple semantics for 1189 * this, no blocking and very strange errors 8) 1190 */ 1191 1192static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags) 1193{ 1194 struct tcp_sock *tp = tcp_sk(sk); 1195 1196 /* No URG data to read. */ 1197 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data || 1198 tp->urg_data == TCP_URG_READ) 1199 return -EINVAL; /* Yes this is right ! */ 1200 1201 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE)) 1202 return -ENOTCONN; 1203 1204 if (tp->urg_data & TCP_URG_VALID) { 1205 int err = 0; 1206 char c = tp->urg_data; 1207 1208 if (!(flags & MSG_PEEK)) 1209 tp->urg_data = TCP_URG_READ; 1210 1211 /* Read urgent data. */ 1212 msg->msg_flags |= MSG_OOB; 1213 1214 if (len > 0) { 1215 if (!(flags & MSG_TRUNC)) 1216 err = memcpy_toiovec(msg->msg_iov, &c, 1); 1217 len = 1; 1218 } else 1219 msg->msg_flags |= MSG_TRUNC; 1220 1221 return err ? -EFAULT : len; 1222 } 1223 1224 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 1225 return 0; 1226 1227 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and 1228 * the available implementations agree in this case: 1229 * this call should never block, independent of the 1230 * blocking state of the socket. 1231 * Mike <pall@rz.uni-karlsruhe.de> 1232 */ 1233 return -EAGAIN; 1234} 1235 1236static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len) 1237{ 1238 struct sk_buff *skb; 1239 int copied = 0, err = 0; 1240 1241 /* XXX -- need to support SO_PEEK_OFF */ 1242 1243 skb_queue_walk(&sk->sk_write_queue, skb) { 1244 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len); 1245 if (err) 1246 break; 1247 1248 copied += skb->len; 1249 } 1250 1251 return err ?: copied; 1252} 1253 1254/* Clean up the receive buffer for full frames taken by the user, 1255 * then send an ACK if necessary. COPIED is the number of bytes 1256 * tcp_recvmsg has given to the user so far, it speeds up the 1257 * calculation of whether or not we must ACK for the sake of 1258 * a window update. 1259 */ 1260void tcp_cleanup_rbuf(struct sock *sk, int copied) 1261{ 1262 struct tcp_sock *tp = tcp_sk(sk); 1263 int time_to_ack = 0; 1264 1265 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 1266 1267 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq), 1268 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n", 1269 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt); 1270 1271 if (inet_csk_ack_scheduled(sk)) { 1272 const struct inet_connection_sock *icsk = inet_csk(sk); 1273 /* Delayed ACKs frequently hit locked sockets during bulk 1274 * receive. */ 1275 if (icsk->icsk_ack.blocked || 1276 /* Once-per-two-segments ACK was not sent by tcp_input.c */ 1277 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss || 1278 /* 1279 * If this read emptied read buffer, we send ACK, if 1280 * connection is not bidirectional, user drained 1281 * receive buffer and there was a small segment 1282 * in queue. 1283 */ 1284 (copied > 0 && 1285 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) || 1286 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) && 1287 !icsk->icsk_ack.pingpong)) && 1288 !atomic_read(&sk->sk_rmem_alloc))) 1289 time_to_ack = 1; 1290 } 1291 1292 /* We send an ACK if we can now advertise a non-zero window 1293 * which has been raised "significantly". 1294 * 1295 * Even if window raised up to infinity, do not send window open ACK 1296 * in states, where we will not receive more. It is useless. 1297 */ 1298 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) { 1299 __u32 rcv_window_now = tcp_receive_window(tp); 1300 1301 /* Optimize, __tcp_select_window() is not cheap. */ 1302 if (2*rcv_window_now <= tp->window_clamp) { 1303 __u32 new_window = __tcp_select_window(sk); 1304 1305 /* Send ACK now, if this read freed lots of space 1306 * in our buffer. Certainly, new_window is new window. 1307 * We can advertise it now, if it is not less than current one. 1308 * "Lots" means "at least twice" here. 1309 */ 1310 if (new_window && new_window >= 2 * rcv_window_now) 1311 time_to_ack = 1; 1312 } 1313 } 1314 if (time_to_ack) 1315 tcp_send_ack(sk); 1316} 1317 1318static void tcp_prequeue_process(struct sock *sk) 1319{ 1320 struct sk_buff *skb; 1321 struct tcp_sock *tp = tcp_sk(sk); 1322 1323 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED); 1324 1325 /* RX process wants to run with disabled BHs, though it is not 1326 * necessary */ 1327 local_bh_disable(); 1328 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL) 1329 sk_backlog_rcv(sk, skb); 1330 local_bh_enable(); 1331 1332 /* Clear memory counter. */ 1333 tp->ucopy.memory = 0; 1334} 1335 1336#ifdef CONFIG_NET_DMA 1337static void tcp_service_net_dma(struct sock *sk, bool wait) 1338{ 1339 dma_cookie_t done, used; 1340 dma_cookie_t last_issued; 1341 struct tcp_sock *tp = tcp_sk(sk); 1342 1343 if (!tp->ucopy.dma_chan) 1344 return; 1345 1346 last_issued = tp->ucopy.dma_cookie; 1347 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1348 1349 do { 1350 if (dma_async_memcpy_complete(tp->ucopy.dma_chan, 1351 last_issued, &done, 1352 &used) == DMA_SUCCESS) { 1353 /* Safe to free early-copied skbs now */ 1354 __skb_queue_purge(&sk->sk_async_wait_queue); 1355 break; 1356 } else { 1357 struct sk_buff *skb; 1358 while ((skb = skb_peek(&sk->sk_async_wait_queue)) && 1359 (dma_async_is_complete(skb->dma_cookie, done, 1360 used) == DMA_SUCCESS)) { 1361 __skb_dequeue(&sk->sk_async_wait_queue); 1362 kfree_skb(skb); 1363 } 1364 } 1365 } while (wait); 1366} 1367#endif 1368 1369static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off) 1370{ 1371 struct sk_buff *skb; 1372 u32 offset; 1373 1374 skb_queue_walk(&sk->sk_receive_queue, skb) { 1375 offset = seq - TCP_SKB_CB(skb)->seq; 1376 if (tcp_hdr(skb)->syn) 1377 offset--; 1378 if (offset < skb->len || tcp_hdr(skb)->fin) { 1379 *off = offset; 1380 return skb; 1381 } 1382 } 1383 return NULL; 1384} 1385 1386/* 1387 * This routine provides an alternative to tcp_recvmsg() for routines 1388 * that would like to handle copying from skbuffs directly in 'sendfile' 1389 * fashion. 1390 * Note: 1391 * - It is assumed that the socket was locked by the caller. 1392 * - The routine does not block. 1393 * - At present, there is no support for reading OOB data 1394 * or for 'peeking' the socket using this routine 1395 * (although both would be easy to implement). 1396 */ 1397int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, 1398 sk_read_actor_t recv_actor) 1399{ 1400 struct sk_buff *skb; 1401 struct tcp_sock *tp = tcp_sk(sk); 1402 u32 seq = tp->copied_seq; 1403 u32 offset; 1404 int copied = 0; 1405 1406 if (sk->sk_state == TCP_LISTEN) 1407 return -ENOTCONN; 1408 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) { 1409 if (offset < skb->len) { 1410 int used; 1411 size_t len; 1412 1413 len = skb->len - offset; 1414 /* Stop reading if we hit a patch of urgent data */ 1415 if (tp->urg_data) { 1416 u32 urg_offset = tp->urg_seq - seq; 1417 if (urg_offset < len) 1418 len = urg_offset; 1419 if (!len) 1420 break; 1421 } 1422 used = recv_actor(desc, skb, offset, len); 1423 if (used < 0) { 1424 if (!copied) 1425 copied = used; 1426 break; 1427 } else if (used <= len) { 1428 seq += used; 1429 copied += used; 1430 offset += used; 1431 } 1432 /* 1433 * If recv_actor drops the lock (e.g. TCP splice 1434 * receive) the skb pointer might be invalid when 1435 * getting here: tcp_collapse might have deleted it 1436 * while aggregating skbs from the socket queue. 1437 */ 1438 skb = tcp_recv_skb(sk, seq-1, &offset); 1439 if (!skb || (offset+1 != skb->len)) 1440 break; 1441 } 1442 if (tcp_hdr(skb)->fin) { 1443 sk_eat_skb(sk, skb, 0); 1444 ++seq; 1445 break; 1446 } 1447 sk_eat_skb(sk, skb, 0); 1448 if (!desc->count) 1449 break; 1450 tp->copied_seq = seq; 1451 } 1452 tp->copied_seq = seq; 1453 1454 tcp_rcv_space_adjust(sk); 1455 1456 /* Clean up data we have read: This will do ACK frames. */ 1457 if (copied > 0) 1458 tcp_cleanup_rbuf(sk, copied); 1459 return copied; 1460} 1461EXPORT_SYMBOL(tcp_read_sock); 1462 1463/* 1464 * This routine copies from a sock struct into the user buffer. 1465 * 1466 * Technical note: in 2.3 we work on _locked_ socket, so that 1467 * tricks with *seq access order and skb->users are not required. 1468 * Probably, code can be easily improved even more. 1469 */ 1470 1471int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg, 1472 size_t len, int nonblock, int flags, int *addr_len) 1473{ 1474 struct tcp_sock *tp = tcp_sk(sk); 1475 int copied = 0; 1476 u32 peek_seq; 1477 u32 *seq; 1478 unsigned long used; 1479 int err; 1480 int target; /* Read at least this many bytes */ 1481 long timeo; 1482 struct task_struct *user_recv = NULL; 1483 int copied_early = 0; 1484 struct sk_buff *skb; 1485 u32 urg_hole = 0; 1486 1487 lock_sock(sk); 1488 1489 err = -ENOTCONN; 1490 if (sk->sk_state == TCP_LISTEN) 1491 goto out; 1492 1493 timeo = sock_rcvtimeo(sk, nonblock); 1494 1495 /* Urgent data needs to be handled specially. */ 1496 if (flags & MSG_OOB) 1497 goto recv_urg; 1498 1499 if (unlikely(tp->repair)) { 1500 err = -EPERM; 1501 if (!(flags & MSG_PEEK)) 1502 goto out; 1503 1504 if (tp->repair_queue == TCP_SEND_QUEUE) 1505 goto recv_sndq; 1506 1507 err = -EINVAL; 1508 if (tp->repair_queue == TCP_NO_QUEUE) 1509 goto out; 1510 1511 /* 'common' recv queue MSG_PEEK-ing */ 1512 } 1513 1514 seq = &tp->copied_seq; 1515 if (flags & MSG_PEEK) { 1516 peek_seq = tp->copied_seq; 1517 seq = &peek_seq; 1518 } 1519 1520 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 1521 1522#ifdef CONFIG_NET_DMA 1523 tp->ucopy.dma_chan = NULL; 1524 preempt_disable(); 1525 skb = skb_peek_tail(&sk->sk_receive_queue); 1526 { 1527 int available = 0; 1528 1529 if (skb) 1530 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq); 1531 if ((available < target) && 1532 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) && 1533 !sysctl_tcp_low_latency && 1534 net_dma_find_channel()) { 1535 preempt_enable_no_resched(); 1536 tp->ucopy.pinned_list = 1537 dma_pin_iovec_pages(msg->msg_iov, len); 1538 } else { 1539 preempt_enable_no_resched(); 1540 } 1541 } 1542#endif 1543 1544 do { 1545 u32 offset; 1546 1547 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */ 1548 if (tp->urg_data && tp->urg_seq == *seq) { 1549 if (copied) 1550 break; 1551 if (signal_pending(current)) { 1552 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN; 1553 break; 1554 } 1555 } 1556 1557 /* Next get a buffer. */ 1558 1559 skb_queue_walk(&sk->sk_receive_queue, skb) { 1560 /* Now that we have two receive queues this 1561 * shouldn't happen. 1562 */ 1563 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq), 1564 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n", 1565 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, 1566 flags)) 1567 break; 1568 1569 offset = *seq - TCP_SKB_CB(skb)->seq; 1570 if (tcp_hdr(skb)->syn) 1571 offset--; 1572 if (offset < skb->len) 1573 goto found_ok_skb; 1574 if (tcp_hdr(skb)->fin) 1575 goto found_fin_ok; 1576 WARN(!(flags & MSG_PEEK), 1577 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n", 1578 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags); 1579 } 1580 1581 /* Well, if we have backlog, try to process it now yet. */ 1582 1583 if (copied >= target && !sk->sk_backlog.tail) 1584 break; 1585 1586 if (copied) { 1587 if (sk->sk_err || 1588 sk->sk_state == TCP_CLOSE || 1589 (sk->sk_shutdown & RCV_SHUTDOWN) || 1590 !timeo || 1591 signal_pending(current)) 1592 break; 1593 } else { 1594 if (sock_flag(sk, SOCK_DONE)) 1595 break; 1596 1597 if (sk->sk_err) { 1598 copied = sock_error(sk); 1599 break; 1600 } 1601 1602 if (sk->sk_shutdown & RCV_SHUTDOWN) 1603 break; 1604 1605 if (sk->sk_state == TCP_CLOSE) { 1606 if (!sock_flag(sk, SOCK_DONE)) { 1607 /* This occurs when user tries to read 1608 * from never connected socket. 1609 */ 1610 copied = -ENOTCONN; 1611 break; 1612 } 1613 break; 1614 } 1615 1616 if (!timeo) { 1617 copied = -EAGAIN; 1618 break; 1619 } 1620 1621 if (signal_pending(current)) { 1622 copied = sock_intr_errno(timeo); 1623 break; 1624 } 1625 } 1626 1627 tcp_cleanup_rbuf(sk, copied); 1628 1629 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) { 1630 /* Install new reader */ 1631 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) { 1632 user_recv = current; 1633 tp->ucopy.task = user_recv; 1634 tp->ucopy.iov = msg->msg_iov; 1635 } 1636 1637 tp->ucopy.len = len; 1638 1639 WARN_ON(tp->copied_seq != tp->rcv_nxt && 1640 !(flags & (MSG_PEEK | MSG_TRUNC))); 1641 1642 /* Ugly... If prequeue is not empty, we have to 1643 * process it before releasing socket, otherwise 1644 * order will be broken at second iteration. 1645 * More elegant solution is required!!! 1646 * 1647 * Look: we have the following (pseudo)queues: 1648 * 1649 * 1. packets in flight 1650 * 2. backlog 1651 * 3. prequeue 1652 * 4. receive_queue 1653 * 1654 * Each queue can be processed only if the next ones 1655 * are empty. At this point we have empty receive_queue. 1656 * But prequeue _can_ be not empty after 2nd iteration, 1657 * when we jumped to start of loop because backlog 1658 * processing added something to receive_queue. 1659 * We cannot release_sock(), because backlog contains 1660 * packets arrived _after_ prequeued ones. 1661 * 1662 * Shortly, algorithm is clear --- to process all 1663 * the queues in order. We could make it more directly, 1664 * requeueing packets from backlog to prequeue, if 1665 * is not empty. It is more elegant, but eats cycles, 1666 * unfortunately. 1667 */ 1668 if (!skb_queue_empty(&tp->ucopy.prequeue)) 1669 goto do_prequeue; 1670 1671 /* __ Set realtime policy in scheduler __ */ 1672 } 1673 1674#ifdef CONFIG_NET_DMA 1675 if (tp->ucopy.dma_chan) 1676 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1677#endif 1678 if (copied >= target) { 1679 /* Do not sleep, just process backlog. */ 1680 release_sock(sk); 1681 lock_sock(sk); 1682 } else 1683 sk_wait_data(sk, &timeo); 1684 1685#ifdef CONFIG_NET_DMA 1686 tcp_service_net_dma(sk, false); /* Don't block */ 1687 tp->ucopy.wakeup = 0; 1688#endif 1689 1690 if (user_recv) { 1691 int chunk; 1692 1693 /* __ Restore normal policy in scheduler __ */ 1694 1695 if ((chunk = len - tp->ucopy.len) != 0) { 1696 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk); 1697 len -= chunk; 1698 copied += chunk; 1699 } 1700 1701 if (tp->rcv_nxt == tp->copied_seq && 1702 !skb_queue_empty(&tp->ucopy.prequeue)) { 1703do_prequeue: 1704 tcp_prequeue_process(sk); 1705 1706 if ((chunk = len - tp->ucopy.len) != 0) { 1707 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1708 len -= chunk; 1709 copied += chunk; 1710 } 1711 } 1712 } 1713 if ((flags & MSG_PEEK) && 1714 (peek_seq - copied - urg_hole != tp->copied_seq)) { 1715 if (net_ratelimit()) 1716 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n", 1717 current->comm, task_pid_nr(current)); 1718 peek_seq = tp->copied_seq; 1719 } 1720 continue; 1721 1722 found_ok_skb: 1723 /* Ok so how much can we use? */ 1724 used = skb->len - offset; 1725 if (len < used) 1726 used = len; 1727 1728 /* Do we have urgent data here? */ 1729 if (tp->urg_data) { 1730 u32 urg_offset = tp->urg_seq - *seq; 1731 if (urg_offset < used) { 1732 if (!urg_offset) { 1733 if (!sock_flag(sk, SOCK_URGINLINE)) { 1734 ++*seq; 1735 urg_hole++; 1736 offset++; 1737 used--; 1738 if (!used) 1739 goto skip_copy; 1740 } 1741 } else 1742 used = urg_offset; 1743 } 1744 } 1745 1746 if (!(flags & MSG_TRUNC)) { 1747#ifdef CONFIG_NET_DMA 1748 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 1749 tp->ucopy.dma_chan = net_dma_find_channel(); 1750 1751 if (tp->ucopy.dma_chan) { 1752 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec( 1753 tp->ucopy.dma_chan, skb, offset, 1754 msg->msg_iov, used, 1755 tp->ucopy.pinned_list); 1756 1757 if (tp->ucopy.dma_cookie < 0) { 1758 1759 pr_alert("%s: dma_cookie < 0\n", 1760 __func__); 1761 1762 /* Exception. Bailout! */ 1763 if (!copied) 1764 copied = -EFAULT; 1765 break; 1766 } 1767 1768 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1769 1770 if ((offset + used) == skb->len) 1771 copied_early = 1; 1772 1773 } else 1774#endif 1775 { 1776 err = skb_copy_datagram_iovec(skb, offset, 1777 msg->msg_iov, used); 1778 if (err) { 1779 /* Exception. Bailout! */ 1780 if (!copied) 1781 copied = -EFAULT; 1782 break; 1783 } 1784 } 1785 } 1786 1787 *seq += used; 1788 copied += used; 1789 len -= used; 1790 1791 tcp_rcv_space_adjust(sk); 1792 1793skip_copy: 1794 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) { 1795 tp->urg_data = 0; 1796 tcp_fast_path_check(sk); 1797 } 1798 if (used + offset < skb->len) 1799 continue; 1800 1801 if (tcp_hdr(skb)->fin) 1802 goto found_fin_ok; 1803 if (!(flags & MSG_PEEK)) { 1804 sk_eat_skb(sk, skb, copied_early); 1805 copied_early = 0; 1806 } 1807 continue; 1808 1809 found_fin_ok: 1810 /* Process the FIN. */ 1811 ++*seq; 1812 if (!(flags & MSG_PEEK)) { 1813 sk_eat_skb(sk, skb, copied_early); 1814 copied_early = 0; 1815 } 1816 break; 1817 } while (len > 0); 1818 1819 if (user_recv) { 1820 if (!skb_queue_empty(&tp->ucopy.prequeue)) { 1821 int chunk; 1822 1823 tp->ucopy.len = copied > 0 ? len : 0; 1824 1825 tcp_prequeue_process(sk); 1826 1827 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) { 1828 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1829 len -= chunk; 1830 copied += chunk; 1831 } 1832 } 1833 1834 tp->ucopy.task = NULL; 1835 tp->ucopy.len = 0; 1836 } 1837 1838#ifdef CONFIG_NET_DMA 1839 tcp_service_net_dma(sk, true); /* Wait for queue to drain */ 1840 tp->ucopy.dma_chan = NULL; 1841 1842 if (tp->ucopy.pinned_list) { 1843 dma_unpin_iovec_pages(tp->ucopy.pinned_list); 1844 tp->ucopy.pinned_list = NULL; 1845 } 1846#endif 1847 1848 /* According to UNIX98, msg_name/msg_namelen are ignored 1849 * on connected socket. I was just happy when found this 8) --ANK 1850 */ 1851 1852 /* Clean up data we have read: This will do ACK frames. */ 1853 tcp_cleanup_rbuf(sk, copied); 1854 1855 release_sock(sk); 1856 return copied; 1857 1858out: 1859 release_sock(sk); 1860 return err; 1861 1862recv_urg: 1863 err = tcp_recv_urg(sk, msg, len, flags); 1864 goto out; 1865 1866recv_sndq: 1867 err = tcp_peek_sndq(sk, msg, len); 1868 goto out; 1869} 1870EXPORT_SYMBOL(tcp_recvmsg); 1871 1872void tcp_set_state(struct sock *sk, int state) 1873{ 1874 int oldstate = sk->sk_state; 1875 1876 switch (state) { 1877 case TCP_ESTABLISHED: 1878 if (oldstate != TCP_ESTABLISHED) 1879 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1880 break; 1881 1882 case TCP_CLOSE: 1883 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED) 1884 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS); 1885 1886 sk->sk_prot->unhash(sk); 1887 if (inet_csk(sk)->icsk_bind_hash && 1888 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK)) 1889 inet_put_port(sk); 1890 /* fall through */ 1891 default: 1892 if (oldstate == TCP_ESTABLISHED) 1893 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1894 } 1895 1896 /* Change state AFTER socket is unhashed to avoid closed 1897 * socket sitting in hash tables. 1898 */ 1899 sk->sk_state = state; 1900 1901#ifdef STATE_TRACE 1902 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]); 1903#endif 1904} 1905EXPORT_SYMBOL_GPL(tcp_set_state); 1906 1907/* 1908 * State processing on a close. This implements the state shift for 1909 * sending our FIN frame. Note that we only send a FIN for some 1910 * states. A shutdown() may have already sent the FIN, or we may be 1911 * closed. 1912 */ 1913 1914static const unsigned char new_state[16] = { 1915 /* current state: new state: action: */ 1916 /* (Invalid) */ TCP_CLOSE, 1917 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1918 /* TCP_SYN_SENT */ TCP_CLOSE, 1919 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1920 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1, 1921 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2, 1922 /* TCP_TIME_WAIT */ TCP_CLOSE, 1923 /* TCP_CLOSE */ TCP_CLOSE, 1924 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN, 1925 /* TCP_LAST_ACK */ TCP_LAST_ACK, 1926 /* TCP_LISTEN */ TCP_CLOSE, 1927 /* TCP_CLOSING */ TCP_CLOSING, 1928}; 1929 1930static int tcp_close_state(struct sock *sk) 1931{ 1932 int next = (int)new_state[sk->sk_state]; 1933 int ns = next & TCP_STATE_MASK; 1934 1935 tcp_set_state(sk, ns); 1936 1937 return next & TCP_ACTION_FIN; 1938} 1939 1940/* 1941 * Shutdown the sending side of a connection. Much like close except 1942 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD). 1943 */ 1944 1945void tcp_shutdown(struct sock *sk, int how) 1946{ 1947 /* We need to grab some memory, and put together a FIN, 1948 * and then put it into the queue to be sent. 1949 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92. 1950 */ 1951 if (!(how & SEND_SHUTDOWN)) 1952 return; 1953 1954 /* If we've already sent a FIN, or it's a closed state, skip this. */ 1955 if ((1 << sk->sk_state) & 1956 (TCPF_ESTABLISHED | TCPF_SYN_SENT | 1957 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) { 1958 /* Clear out any half completed packets. FIN if needed. */ 1959 if (tcp_close_state(sk)) 1960 tcp_send_fin(sk); 1961 } 1962} 1963EXPORT_SYMBOL(tcp_shutdown); 1964 1965bool tcp_check_oom(struct sock *sk, int shift) 1966{ 1967 bool too_many_orphans, out_of_socket_memory; 1968 1969 too_many_orphans = tcp_too_many_orphans(sk, shift); 1970 out_of_socket_memory = tcp_out_of_memory(sk); 1971 1972 if (too_many_orphans && net_ratelimit()) 1973 pr_info("too many orphaned sockets\n"); 1974 if (out_of_socket_memory && net_ratelimit()) 1975 pr_info("out of memory -- consider tuning tcp_mem\n"); 1976 return too_many_orphans || out_of_socket_memory; 1977} 1978 1979void tcp_close(struct sock *sk, long timeout) 1980{ 1981 struct sk_buff *skb; 1982 int data_was_unread = 0; 1983 int state; 1984 1985 lock_sock(sk); 1986 sk->sk_shutdown = SHUTDOWN_MASK; 1987 1988 if (sk->sk_state == TCP_LISTEN) { 1989 tcp_set_state(sk, TCP_CLOSE); 1990 1991 /* Special case. */ 1992 inet_csk_listen_stop(sk); 1993 1994 goto adjudge_to_death; 1995 } 1996 1997 /* We need to flush the recv. buffs. We do this only on the 1998 * descriptor close, not protocol-sourced closes, because the 1999 * reader process may not have drained the data yet! 2000 */ 2001 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) { 2002 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq - 2003 tcp_hdr(skb)->fin; 2004 data_was_unread += len; 2005 __kfree_skb(skb); 2006 } 2007 2008 sk_mem_reclaim(sk); 2009 2010 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */ 2011 if (sk->sk_state == TCP_CLOSE) 2012 goto adjudge_to_death; 2013 2014 /* As outlined in RFC 2525, section 2.17, we send a RST here because 2015 * data was lost. To witness the awful effects of the old behavior of 2016 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk 2017 * GET in an FTP client, suspend the process, wait for the client to 2018 * advertise a zero window, then kill -9 the FTP client, wheee... 2019 * Note: timeout is always zero in such a case. 2020 */ 2021 if (unlikely(tcp_sk(sk)->repair)) { 2022 sk->sk_prot->disconnect(sk, 0); 2023 } else if (data_was_unread) { 2024 /* Unread data was tossed, zap the connection. */ 2025 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE); 2026 tcp_set_state(sk, TCP_CLOSE); 2027 tcp_send_active_reset(sk, sk->sk_allocation); 2028 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) { 2029 /* Check zero linger _after_ checking for unread data. */ 2030 sk->sk_prot->disconnect(sk, 0); 2031 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 2032 } else if (tcp_close_state(sk)) { 2033 /* We FIN if the application ate all the data before 2034 * zapping the connection. 2035 */ 2036 2037 /* RED-PEN. Formally speaking, we have broken TCP state 2038 * machine. State transitions: 2039 * 2040 * TCP_ESTABLISHED -> TCP_FIN_WAIT1 2041 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible) 2042 * TCP_CLOSE_WAIT -> TCP_LAST_ACK 2043 * 2044 * are legal only when FIN has been sent (i.e. in window), 2045 * rather than queued out of window. Purists blame. 2046 * 2047 * F.e. "RFC state" is ESTABLISHED, 2048 * if Linux state is FIN-WAIT-1, but FIN is still not sent. 2049 * 2050 * The visible declinations are that sometimes 2051 * we enter time-wait state, when it is not required really 2052 * (harmless), do not send active resets, when they are 2053 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when 2054 * they look as CLOSING or LAST_ACK for Linux) 2055 * Probably, I missed some more holelets. 2056 * --ANK 2057 */ 2058 tcp_send_fin(sk); 2059 } 2060 2061 sk_stream_wait_close(sk, timeout); 2062 2063adjudge_to_death: 2064 state = sk->sk_state; 2065 sock_hold(sk); 2066 sock_orphan(sk); 2067 2068 /* It is the last release_sock in its life. It will remove backlog. */ 2069 release_sock(sk); 2070 2071 2072 /* Now socket is owned by kernel and we acquire BH lock 2073 to finish close. No need to check for user refs. 2074 */ 2075 local_bh_disable(); 2076 bh_lock_sock(sk); 2077 WARN_ON(sock_owned_by_user(sk)); 2078 2079 percpu_counter_inc(sk->sk_prot->orphan_count); 2080 2081 /* Have we already been destroyed by a softirq or backlog? */ 2082 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE) 2083 goto out; 2084 2085 /* This is a (useful) BSD violating of the RFC. There is a 2086 * problem with TCP as specified in that the other end could 2087 * keep a socket open forever with no application left this end. 2088 * We use a 3 minute timeout (about the same as BSD) then kill 2089 * our end. If they send after that then tough - BUT: long enough 2090 * that we won't make the old 4*rto = almost no time - whoops 2091 * reset mistake. 2092 * 2093 * Nope, it was not mistake. It is really desired behaviour 2094 * f.e. on http servers, when such sockets are useless, but 2095 * consume significant resources. Let's do it with special 2096 * linger2 option. --ANK 2097 */ 2098 2099 if (sk->sk_state == TCP_FIN_WAIT2) { 2100 struct tcp_sock *tp = tcp_sk(sk); 2101 if (tp->linger2 < 0) { 2102 tcp_set_state(sk, TCP_CLOSE); 2103 tcp_send_active_reset(sk, GFP_ATOMIC); 2104 NET_INC_STATS_BH(sock_net(sk), 2105 LINUX_MIB_TCPABORTONLINGER); 2106 } else { 2107 const int tmo = tcp_fin_time(sk); 2108 2109 if (tmo > TCP_TIMEWAIT_LEN) { 2110 inet_csk_reset_keepalive_timer(sk, 2111 tmo - TCP_TIMEWAIT_LEN); 2112 } else { 2113 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 2114 goto out; 2115 } 2116 } 2117 } 2118 if (sk->sk_state != TCP_CLOSE) { 2119 sk_mem_reclaim(sk); 2120 if (tcp_check_oom(sk, 0)) { 2121 tcp_set_state(sk, TCP_CLOSE); 2122 tcp_send_active_reset(sk, GFP_ATOMIC); 2123 NET_INC_STATS_BH(sock_net(sk), 2124 LINUX_MIB_TCPABORTONMEMORY); 2125 } 2126 } 2127 2128 if (sk->sk_state == TCP_CLOSE) 2129 inet_csk_destroy_sock(sk); 2130 /* Otherwise, socket is reprieved until protocol close. */ 2131 2132out: 2133 bh_unlock_sock(sk); 2134 local_bh_enable(); 2135 sock_put(sk); 2136} 2137EXPORT_SYMBOL(tcp_close); 2138 2139/* These states need RST on ABORT according to RFC793 */ 2140 2141static inline int tcp_need_reset(int state) 2142{ 2143 return (1 << state) & 2144 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 | 2145 TCPF_FIN_WAIT2 | TCPF_SYN_RECV); 2146} 2147 2148int tcp_disconnect(struct sock *sk, int flags) 2149{ 2150 struct inet_sock *inet = inet_sk(sk); 2151 struct inet_connection_sock *icsk = inet_csk(sk); 2152 struct tcp_sock *tp = tcp_sk(sk); 2153 int err = 0; 2154 int old_state = sk->sk_state; 2155 2156 if (old_state != TCP_CLOSE) 2157 tcp_set_state(sk, TCP_CLOSE); 2158 2159 /* ABORT function of RFC793 */ 2160 if (old_state == TCP_LISTEN) { 2161 inet_csk_listen_stop(sk); 2162 } else if (unlikely(tp->repair)) { 2163 sk->sk_err = ECONNABORTED; 2164 } else if (tcp_need_reset(old_state) || 2165 (tp->snd_nxt != tp->write_seq && 2166 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) { 2167 /* The last check adjusts for discrepancy of Linux wrt. RFC 2168 * states 2169 */ 2170 tcp_send_active_reset(sk, gfp_any()); 2171 sk->sk_err = ECONNRESET; 2172 } else if (old_state == TCP_SYN_SENT) 2173 sk->sk_err = ECONNRESET; 2174 2175 tcp_clear_xmit_timers(sk); 2176 __skb_queue_purge(&sk->sk_receive_queue); 2177 tcp_write_queue_purge(sk); 2178 __skb_queue_purge(&tp->out_of_order_queue); 2179#ifdef CONFIG_NET_DMA 2180 __skb_queue_purge(&sk->sk_async_wait_queue); 2181#endif 2182 2183 inet->inet_dport = 0; 2184 2185 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK)) 2186 inet_reset_saddr(sk); 2187 2188 sk->sk_shutdown = 0; 2189 sock_reset_flag(sk, SOCK_DONE); 2190 tp->srtt = 0; 2191 if ((tp->write_seq += tp->max_window + 2) == 0) 2192 tp->write_seq = 1; 2193 icsk->icsk_backoff = 0; 2194 tp->snd_cwnd = 2; 2195 icsk->icsk_probes_out = 0; 2196 tp->packets_out = 0; 2197 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 2198 tp->snd_cwnd_cnt = 0; 2199 tp->bytes_acked = 0; 2200 tp->window_clamp = 0; 2201 tcp_set_ca_state(sk, TCP_CA_Open); 2202 tcp_clear_retrans(tp); 2203 inet_csk_delack_init(sk); 2204 tcp_init_send_head(sk); 2205 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt)); 2206 __sk_dst_reset(sk); 2207 2208 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash); 2209 2210 sk->sk_error_report(sk); 2211 return err; 2212} 2213EXPORT_SYMBOL(tcp_disconnect); 2214 2215static inline int tcp_can_repair_sock(struct sock *sk) 2216{ 2217 return capable(CAP_NET_ADMIN) && 2218 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED)); 2219} 2220 2221static int tcp_repair_options_est(struct tcp_sock *tp, char __user *optbuf, unsigned int len) 2222{ 2223 /* 2224 * Options are stored in CODE:VALUE form where CODE is 8bit and VALUE 2225 * fits the respective TCPOLEN_ size 2226 */ 2227 2228 while (len > 0) { 2229 u8 opcode; 2230 2231 if (get_user(opcode, optbuf)) 2232 return -EFAULT; 2233 2234 optbuf++; 2235 len--; 2236 2237 switch (opcode) { 2238 case TCPOPT_MSS: { 2239 u16 in_mss; 2240 2241 if (len < sizeof(in_mss)) 2242 return -ENODATA; 2243 if (get_user(in_mss, optbuf)) 2244 return -EFAULT; 2245 2246 tp->rx_opt.mss_clamp = in_mss; 2247 2248 optbuf += sizeof(in_mss); 2249 len -= sizeof(in_mss); 2250 break; 2251 } 2252 case TCPOPT_WINDOW: { 2253 u8 wscale; 2254 2255 if (len < sizeof(wscale)) 2256 return -ENODATA; 2257 if (get_user(wscale, optbuf)) 2258 return -EFAULT; 2259 2260 if (wscale > 14) 2261 return -EFBIG; 2262 2263 tp->rx_opt.snd_wscale = wscale; 2264 2265 optbuf += sizeof(wscale); 2266 len -= sizeof(wscale); 2267 break; 2268 } 2269 case TCPOPT_SACK_PERM: 2270 tp->rx_opt.sack_ok |= TCP_SACK_SEEN; 2271 if (sysctl_tcp_fack) 2272 tcp_enable_fack(tp); 2273 break; 2274 case TCPOPT_TIMESTAMP: 2275 tp->rx_opt.tstamp_ok = 1; 2276 break; 2277 } 2278 } 2279 2280 return 0; 2281} 2282 2283/* 2284 * Socket option code for TCP. 2285 */ 2286static int do_tcp_setsockopt(struct sock *sk, int level, 2287 int optname, char __user *optval, unsigned int optlen) 2288{ 2289 struct tcp_sock *tp = tcp_sk(sk); 2290 struct inet_connection_sock *icsk = inet_csk(sk); 2291 int val; 2292 int err = 0; 2293 2294 /* These are data/string values, all the others are ints */ 2295 switch (optname) { 2296 case TCP_CONGESTION: { 2297 char name[TCP_CA_NAME_MAX]; 2298 2299 if (optlen < 1) 2300 return -EINVAL; 2301 2302 val = strncpy_from_user(name, optval, 2303 min_t(long, TCP_CA_NAME_MAX-1, optlen)); 2304 if (val < 0) 2305 return -EFAULT; 2306 name[val] = 0; 2307 2308 lock_sock(sk); 2309 err = tcp_set_congestion_control(sk, name); 2310 release_sock(sk); 2311 return err; 2312 } 2313 case TCP_COOKIE_TRANSACTIONS: { 2314 struct tcp_cookie_transactions ctd; 2315 struct tcp_cookie_values *cvp = NULL; 2316 2317 if (sizeof(ctd) > optlen) 2318 return -EINVAL; 2319 if (copy_from_user(&ctd, optval, sizeof(ctd))) 2320 return -EFAULT; 2321 2322 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) || 2323 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED) 2324 return -EINVAL; 2325 2326 if (ctd.tcpct_cookie_desired == 0) { 2327 /* default to global value */ 2328 } else if ((0x1 & ctd.tcpct_cookie_desired) || 2329 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX || 2330 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) { 2331 return -EINVAL; 2332 } 2333 2334 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) { 2335 /* Supercedes all other values */ 2336 lock_sock(sk); 2337 if (tp->cookie_values != NULL) { 2338 kref_put(&tp->cookie_values->kref, 2339 tcp_cookie_values_release); 2340 tp->cookie_values = NULL; 2341 } 2342 tp->rx_opt.cookie_in_always = 0; /* false */ 2343 tp->rx_opt.cookie_out_never = 1; /* true */ 2344 release_sock(sk); 2345 return err; 2346 } 2347 2348 /* Allocate ancillary memory before locking. 2349 */ 2350 if (ctd.tcpct_used > 0 || 2351 (tp->cookie_values == NULL && 2352 (sysctl_tcp_cookie_size > 0 || 2353 ctd.tcpct_cookie_desired > 0 || 2354 ctd.tcpct_s_data_desired > 0))) { 2355 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used, 2356 GFP_KERNEL); 2357 if (cvp == NULL) 2358 return -ENOMEM; 2359 2360 kref_init(&cvp->kref); 2361 } 2362 lock_sock(sk); 2363 tp->rx_opt.cookie_in_always = 2364 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags); 2365 tp->rx_opt.cookie_out_never = 0; /* false */ 2366 2367 if (tp->cookie_values != NULL) { 2368 if (cvp != NULL) { 2369 /* Changed values are recorded by a changed 2370 * pointer, ensuring the cookie will differ, 2371 * without separately hashing each value later. 2372 */ 2373 kref_put(&tp->cookie_values->kref, 2374 tcp_cookie_values_release); 2375 } else { 2376 cvp = tp->cookie_values; 2377 } 2378 } 2379 2380 if (cvp != NULL) { 2381 cvp->cookie_desired = ctd.tcpct_cookie_desired; 2382 2383 if (ctd.tcpct_used > 0) { 2384 memcpy(cvp->s_data_payload, ctd.tcpct_value, 2385 ctd.tcpct_used); 2386 cvp->s_data_desired = ctd.tcpct_used; 2387 cvp->s_data_constant = 1; /* true */ 2388 } else { 2389 /* No constant payload data. */ 2390 cvp->s_data_desired = ctd.tcpct_s_data_desired; 2391 cvp->s_data_constant = 0; /* false */ 2392 } 2393 2394 tp->cookie_values = cvp; 2395 } 2396 release_sock(sk); 2397 return err; 2398 } 2399 default: 2400 /* fallthru */ 2401 break; 2402 } 2403 2404 if (optlen < sizeof(int)) 2405 return -EINVAL; 2406 2407 if (get_user(val, (int __user *)optval)) 2408 return -EFAULT; 2409 2410 lock_sock(sk); 2411 2412 switch (optname) { 2413 case TCP_MAXSEG: 2414 /* Values greater than interface MTU won't take effect. However 2415 * at the point when this call is done we typically don't yet 2416 * know which interface is going to be used */ 2417 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) { 2418 err = -EINVAL; 2419 break; 2420 } 2421 tp->rx_opt.user_mss = val; 2422 break; 2423 2424 case TCP_NODELAY: 2425 if (val) { 2426 /* TCP_NODELAY is weaker than TCP_CORK, so that 2427 * this option on corked socket is remembered, but 2428 * it is not activated until cork is cleared. 2429 * 2430 * However, when TCP_NODELAY is set we make 2431 * an explicit push, which overrides even TCP_CORK 2432 * for currently queued segments. 2433 */ 2434 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH; 2435 tcp_push_pending_frames(sk); 2436 } else { 2437 tp->nonagle &= ~TCP_NAGLE_OFF; 2438 } 2439 break; 2440 2441 case TCP_THIN_LINEAR_TIMEOUTS: 2442 if (val < 0 || val > 1) 2443 err = -EINVAL; 2444 else 2445 tp->thin_lto = val; 2446 break; 2447 2448 case TCP_THIN_DUPACK: 2449 if (val < 0 || val > 1) 2450 err = -EINVAL; 2451 else 2452 tp->thin_dupack = val; 2453 break; 2454 2455 case TCP_REPAIR: 2456 if (!tcp_can_repair_sock(sk)) 2457 err = -EPERM; 2458 else if (val == 1) { 2459 tp->repair = 1; 2460 sk->sk_reuse = SK_FORCE_REUSE; 2461 tp->repair_queue = TCP_NO_QUEUE; 2462 } else if (val == 0) { 2463 tp->repair = 0; 2464 sk->sk_reuse = SK_NO_REUSE; 2465 tcp_send_window_probe(sk); 2466 } else 2467 err = -EINVAL; 2468 2469 break; 2470 2471 case TCP_REPAIR_QUEUE: 2472 if (!tp->repair) 2473 err = -EPERM; 2474 else if (val < TCP_QUEUES_NR) 2475 tp->repair_queue = val; 2476 else 2477 err = -EINVAL; 2478 break; 2479 2480 case TCP_QUEUE_SEQ: 2481 if (sk->sk_state != TCP_CLOSE) 2482 err = -EPERM; 2483 else if (tp->repair_queue == TCP_SEND_QUEUE) 2484 tp->write_seq = val; 2485 else if (tp->repair_queue == TCP_RECV_QUEUE) 2486 tp->rcv_nxt = val; 2487 else 2488 err = -EINVAL; 2489 break; 2490 2491 case TCP_REPAIR_OPTIONS: 2492 if (!tp->repair) 2493 err = -EINVAL; 2494 else if (sk->sk_state == TCP_ESTABLISHED) 2495 err = tcp_repair_options_est(tp, optval, optlen); 2496 else 2497 err = -EPERM; 2498 break; 2499 2500 case TCP_CORK: 2501 /* When set indicates to always queue non-full frames. 2502 * Later the user clears this option and we transmit 2503 * any pending partial frames in the queue. This is 2504 * meant to be used alongside sendfile() to get properly 2505 * filled frames when the user (for example) must write 2506 * out headers with a write() call first and then use 2507 * sendfile to send out the data parts. 2508 * 2509 * TCP_CORK can be set together with TCP_NODELAY and it is 2510 * stronger than TCP_NODELAY. 2511 */ 2512 if (val) { 2513 tp->nonagle |= TCP_NAGLE_CORK; 2514 } else { 2515 tp->nonagle &= ~TCP_NAGLE_CORK; 2516 if (tp->nonagle&TCP_NAGLE_OFF) 2517 tp->nonagle |= TCP_NAGLE_PUSH; 2518 tcp_push_pending_frames(sk); 2519 } 2520 break; 2521 2522 case TCP_KEEPIDLE: 2523 if (val < 1 || val > MAX_TCP_KEEPIDLE) 2524 err = -EINVAL; 2525 else { 2526 tp->keepalive_time = val * HZ; 2527 if (sock_flag(sk, SOCK_KEEPOPEN) && 2528 !((1 << sk->sk_state) & 2529 (TCPF_CLOSE | TCPF_LISTEN))) { 2530 u32 elapsed = keepalive_time_elapsed(tp); 2531 if (tp->keepalive_time > elapsed) 2532 elapsed = tp->keepalive_time - elapsed; 2533 else 2534 elapsed = 0; 2535 inet_csk_reset_keepalive_timer(sk, elapsed); 2536 } 2537 } 2538 break; 2539 case TCP_KEEPINTVL: 2540 if (val < 1 || val > MAX_TCP_KEEPINTVL) 2541 err = -EINVAL; 2542 else 2543 tp->keepalive_intvl = val * HZ; 2544 break; 2545 case TCP_KEEPCNT: 2546 if (val < 1 || val > MAX_TCP_KEEPCNT) 2547 err = -EINVAL; 2548 else 2549 tp->keepalive_probes = val; 2550 break; 2551 case TCP_SYNCNT: 2552 if (val < 1 || val > MAX_TCP_SYNCNT) 2553 err = -EINVAL; 2554 else 2555 icsk->icsk_syn_retries = val; 2556 break; 2557 2558 case TCP_LINGER2: 2559 if (val < 0) 2560 tp->linger2 = -1; 2561 else if (val > sysctl_tcp_fin_timeout / HZ) 2562 tp->linger2 = 0; 2563 else 2564 tp->linger2 = val * HZ; 2565 break; 2566 2567 case TCP_DEFER_ACCEPT: 2568 /* Translate value in seconds to number of retransmits */ 2569 icsk->icsk_accept_queue.rskq_defer_accept = 2570 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ, 2571 TCP_RTO_MAX / HZ); 2572 break; 2573 2574 case TCP_WINDOW_CLAMP: 2575 if (!val) { 2576 if (sk->sk_state != TCP_CLOSE) { 2577 err = -EINVAL; 2578 break; 2579 } 2580 tp->window_clamp = 0; 2581 } else 2582 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ? 2583 SOCK_MIN_RCVBUF / 2 : val; 2584 break; 2585 2586 case TCP_QUICKACK: 2587 if (!val) { 2588 icsk->icsk_ack.pingpong = 1; 2589 } else { 2590 icsk->icsk_ack.pingpong = 0; 2591 if ((1 << sk->sk_state) & 2592 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) && 2593 inet_csk_ack_scheduled(sk)) { 2594 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 2595 tcp_cleanup_rbuf(sk, 1); 2596 if (!(val & 1)) 2597 icsk->icsk_ack.pingpong = 1; 2598 } 2599 } 2600 break; 2601 2602#ifdef CONFIG_TCP_MD5SIG 2603 case TCP_MD5SIG: 2604 /* Read the IP->Key mappings from userspace */ 2605 err = tp->af_specific->md5_parse(sk, optval, optlen); 2606 break; 2607#endif 2608 case TCP_USER_TIMEOUT: 2609 /* Cap the max timeout in ms TCP will retry/retrans 2610 * before giving up and aborting (ETIMEDOUT) a connection. 2611 */ 2612 icsk->icsk_user_timeout = msecs_to_jiffies(val); 2613 break; 2614 default: 2615 err = -ENOPROTOOPT; 2616 break; 2617 } 2618 2619 release_sock(sk); 2620 return err; 2621} 2622 2623int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval, 2624 unsigned int optlen) 2625{ 2626 const struct inet_connection_sock *icsk = inet_csk(sk); 2627 2628 if (level != SOL_TCP) 2629 return icsk->icsk_af_ops->setsockopt(sk, level, optname, 2630 optval, optlen); 2631 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2632} 2633EXPORT_SYMBOL(tcp_setsockopt); 2634 2635#ifdef CONFIG_COMPAT 2636int compat_tcp_setsockopt(struct sock *sk, int level, int optname, 2637 char __user *optval, unsigned int optlen) 2638{ 2639 if (level != SOL_TCP) 2640 return inet_csk_compat_setsockopt(sk, level, optname, 2641 optval, optlen); 2642 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2643} 2644EXPORT_SYMBOL(compat_tcp_setsockopt); 2645#endif 2646 2647/* Return information about state of tcp endpoint in API format. */ 2648void tcp_get_info(const struct sock *sk, struct tcp_info *info) 2649{ 2650 const struct tcp_sock *tp = tcp_sk(sk); 2651 const struct inet_connection_sock *icsk = inet_csk(sk); 2652 u32 now = tcp_time_stamp; 2653 2654 memset(info, 0, sizeof(*info)); 2655 2656 info->tcpi_state = sk->sk_state; 2657 info->tcpi_ca_state = icsk->icsk_ca_state; 2658 info->tcpi_retransmits = icsk->icsk_retransmits; 2659 info->tcpi_probes = icsk->icsk_probes_out; 2660 info->tcpi_backoff = icsk->icsk_backoff; 2661 2662 if (tp->rx_opt.tstamp_ok) 2663 info->tcpi_options |= TCPI_OPT_TIMESTAMPS; 2664 if (tcp_is_sack(tp)) 2665 info->tcpi_options |= TCPI_OPT_SACK; 2666 if (tp->rx_opt.wscale_ok) { 2667 info->tcpi_options |= TCPI_OPT_WSCALE; 2668 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale; 2669 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale; 2670 } 2671 2672 if (tp->ecn_flags & TCP_ECN_OK) 2673 info->tcpi_options |= TCPI_OPT_ECN; 2674 if (tp->ecn_flags & TCP_ECN_SEEN) 2675 info->tcpi_options |= TCPI_OPT_ECN_SEEN; 2676 2677 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto); 2678 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato); 2679 info->tcpi_snd_mss = tp->mss_cache; 2680 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss; 2681 2682 if (sk->sk_state == TCP_LISTEN) { 2683 info->tcpi_unacked = sk->sk_ack_backlog; 2684 info->tcpi_sacked = sk->sk_max_ack_backlog; 2685 } else { 2686 info->tcpi_unacked = tp->packets_out; 2687 info->tcpi_sacked = tp->sacked_out; 2688 } 2689 info->tcpi_lost = tp->lost_out; 2690 info->tcpi_retrans = tp->retrans_out; 2691 info->tcpi_fackets = tp->fackets_out; 2692 2693 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime); 2694 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime); 2695 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp); 2696 2697 info->tcpi_pmtu = icsk->icsk_pmtu_cookie; 2698 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh; 2699 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3; 2700 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2; 2701 info->tcpi_snd_ssthresh = tp->snd_ssthresh; 2702 info->tcpi_snd_cwnd = tp->snd_cwnd; 2703 info->tcpi_advmss = tp->advmss; 2704 info->tcpi_reordering = tp->reordering; 2705 2706 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3; 2707 info->tcpi_rcv_space = tp->rcvq_space.space; 2708 2709 info->tcpi_total_retrans = tp->total_retrans; 2710} 2711EXPORT_SYMBOL_GPL(tcp_get_info); 2712 2713static int do_tcp_getsockopt(struct sock *sk, int level, 2714 int optname, char __user *optval, int __user *optlen) 2715{ 2716 struct inet_connection_sock *icsk = inet_csk(sk); 2717 struct tcp_sock *tp = tcp_sk(sk); 2718 int val, len; 2719 2720 if (get_user(len, optlen)) 2721 return -EFAULT; 2722 2723 len = min_t(unsigned int, len, sizeof(int)); 2724 2725 if (len < 0) 2726 return -EINVAL; 2727 2728 switch (optname) { 2729 case TCP_MAXSEG: 2730 val = tp->mss_cache; 2731 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 2732 val = tp->rx_opt.user_mss; 2733 if (tp->repair) 2734 val = tp->rx_opt.mss_clamp; 2735 break; 2736 case TCP_NODELAY: 2737 val = !!(tp->nonagle&TCP_NAGLE_OFF); 2738 break; 2739 case TCP_CORK: 2740 val = !!(tp->nonagle&TCP_NAGLE_CORK); 2741 break; 2742 case TCP_KEEPIDLE: 2743 val = keepalive_time_when(tp) / HZ; 2744 break; 2745 case TCP_KEEPINTVL: 2746 val = keepalive_intvl_when(tp) / HZ; 2747 break; 2748 case TCP_KEEPCNT: 2749 val = keepalive_probes(tp); 2750 break; 2751 case TCP_SYNCNT: 2752 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries; 2753 break; 2754 case TCP_LINGER2: 2755 val = tp->linger2; 2756 if (val >= 0) 2757 val = (val ? : sysctl_tcp_fin_timeout) / HZ; 2758 break; 2759 case TCP_DEFER_ACCEPT: 2760 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept, 2761 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ); 2762 break; 2763 case TCP_WINDOW_CLAMP: 2764 val = tp->window_clamp; 2765 break; 2766 case TCP_INFO: { 2767 struct tcp_info info; 2768 2769 if (get_user(len, optlen)) 2770 return -EFAULT; 2771 2772 tcp_get_info(sk, &info); 2773 2774 len = min_t(unsigned int, len, sizeof(info)); 2775 if (put_user(len, optlen)) 2776 return -EFAULT; 2777 if (copy_to_user(optval, &info, len)) 2778 return -EFAULT; 2779 return 0; 2780 } 2781 case TCP_QUICKACK: 2782 val = !icsk->icsk_ack.pingpong; 2783 break; 2784 2785 case TCP_CONGESTION: 2786 if (get_user(len, optlen)) 2787 return -EFAULT; 2788 len = min_t(unsigned int, len, TCP_CA_NAME_MAX); 2789 if (put_user(len, optlen)) 2790 return -EFAULT; 2791 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len)) 2792 return -EFAULT; 2793 return 0; 2794 2795 case TCP_COOKIE_TRANSACTIONS: { 2796 struct tcp_cookie_transactions ctd; 2797 struct tcp_cookie_values *cvp = tp->cookie_values; 2798 2799 if (get_user(len, optlen)) 2800 return -EFAULT; 2801 if (len < sizeof(ctd)) 2802 return -EINVAL; 2803 2804 memset(&ctd, 0, sizeof(ctd)); 2805 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ? 2806 TCP_COOKIE_IN_ALWAYS : 0) 2807 | (tp->rx_opt.cookie_out_never ? 2808 TCP_COOKIE_OUT_NEVER : 0); 2809 2810 if (cvp != NULL) { 2811 ctd.tcpct_flags |= (cvp->s_data_in ? 2812 TCP_S_DATA_IN : 0) 2813 | (cvp->s_data_out ? 2814 TCP_S_DATA_OUT : 0); 2815 2816 ctd.tcpct_cookie_desired = cvp->cookie_desired; 2817 ctd.tcpct_s_data_desired = cvp->s_data_desired; 2818 2819 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0], 2820 cvp->cookie_pair_size); 2821 ctd.tcpct_used = cvp->cookie_pair_size; 2822 } 2823 2824 if (put_user(sizeof(ctd), optlen)) 2825 return -EFAULT; 2826 if (copy_to_user(optval, &ctd, sizeof(ctd))) 2827 return -EFAULT; 2828 return 0; 2829 } 2830 case TCP_THIN_LINEAR_TIMEOUTS: 2831 val = tp->thin_lto; 2832 break; 2833 case TCP_THIN_DUPACK: 2834 val = tp->thin_dupack; 2835 break; 2836 2837 case TCP_REPAIR: 2838 val = tp->repair; 2839 break; 2840 2841 case TCP_REPAIR_QUEUE: 2842 if (tp->repair) 2843 val = tp->repair_queue; 2844 else 2845 return -EINVAL; 2846 break; 2847 2848 case TCP_QUEUE_SEQ: 2849 if (tp->repair_queue == TCP_SEND_QUEUE) 2850 val = tp->write_seq; 2851 else if (tp->repair_queue == TCP_RECV_QUEUE) 2852 val = tp->rcv_nxt; 2853 else 2854 return -EINVAL; 2855 break; 2856 2857 case TCP_USER_TIMEOUT: 2858 val = jiffies_to_msecs(icsk->icsk_user_timeout); 2859 break; 2860 default: 2861 return -ENOPROTOOPT; 2862 } 2863 2864 if (put_user(len, optlen)) 2865 return -EFAULT; 2866 if (copy_to_user(optval, &val, len)) 2867 return -EFAULT; 2868 return 0; 2869} 2870 2871int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval, 2872 int __user *optlen) 2873{ 2874 struct inet_connection_sock *icsk = inet_csk(sk); 2875 2876 if (level != SOL_TCP) 2877 return icsk->icsk_af_ops->getsockopt(sk, level, optname, 2878 optval, optlen); 2879 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2880} 2881EXPORT_SYMBOL(tcp_getsockopt); 2882 2883#ifdef CONFIG_COMPAT 2884int compat_tcp_getsockopt(struct sock *sk, int level, int optname, 2885 char __user *optval, int __user *optlen) 2886{ 2887 if (level != SOL_TCP) 2888 return inet_csk_compat_getsockopt(sk, level, optname, 2889 optval, optlen); 2890 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2891} 2892EXPORT_SYMBOL(compat_tcp_getsockopt); 2893#endif 2894 2895struct sk_buff *tcp_tso_segment(struct sk_buff *skb, 2896 netdev_features_t features) 2897{ 2898 struct sk_buff *segs = ERR_PTR(-EINVAL); 2899 struct tcphdr *th; 2900 unsigned int thlen; 2901 unsigned int seq; 2902 __be32 delta; 2903 unsigned int oldlen; 2904 unsigned int mss; 2905 2906 if (!pskb_may_pull(skb, sizeof(*th))) 2907 goto out; 2908 2909 th = tcp_hdr(skb); 2910 thlen = th->doff * 4; 2911 if (thlen < sizeof(*th)) 2912 goto out; 2913 2914 if (!pskb_may_pull(skb, thlen)) 2915 goto out; 2916 2917 oldlen = (u16)~skb->len; 2918 __skb_pull(skb, thlen); 2919 2920 mss = skb_shinfo(skb)->gso_size; 2921 if (unlikely(skb->len <= mss)) 2922 goto out; 2923 2924 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) { 2925 /* Packet is from an untrusted source, reset gso_segs. */ 2926 int type = skb_shinfo(skb)->gso_type; 2927 2928 if (unlikely(type & 2929 ~(SKB_GSO_TCPV4 | 2930 SKB_GSO_DODGY | 2931 SKB_GSO_TCP_ECN | 2932 SKB_GSO_TCPV6 | 2933 0) || 2934 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))) 2935 goto out; 2936 2937 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss); 2938 2939 segs = NULL; 2940 goto out; 2941 } 2942 2943 segs = skb_segment(skb, features); 2944 if (IS_ERR(segs)) 2945 goto out; 2946 2947 delta = htonl(oldlen + (thlen + mss)); 2948 2949 skb = segs; 2950 th = tcp_hdr(skb); 2951 seq = ntohl(th->seq); 2952 2953 do { 2954 th->fin = th->psh = 0; 2955 2956 th->check = ~csum_fold((__force __wsum)((__force u32)th->check + 2957 (__force u32)delta)); 2958 if (skb->ip_summed != CHECKSUM_PARTIAL) 2959 th->check = 2960 csum_fold(csum_partial(skb_transport_header(skb), 2961 thlen, skb->csum)); 2962 2963 seq += mss; 2964 skb = skb->next; 2965 th = tcp_hdr(skb); 2966 2967 th->seq = htonl(seq); 2968 th->cwr = 0; 2969 } while (skb->next); 2970 2971 delta = htonl(oldlen + (skb->tail - skb->transport_header) + 2972 skb->data_len); 2973 th->check = ~csum_fold((__force __wsum)((__force u32)th->check + 2974 (__force u32)delta)); 2975 if (skb->ip_summed != CHECKSUM_PARTIAL) 2976 th->check = csum_fold(csum_partial(skb_transport_header(skb), 2977 thlen, skb->csum)); 2978 2979out: 2980 return segs; 2981} 2982EXPORT_SYMBOL(tcp_tso_segment); 2983 2984struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb) 2985{ 2986 struct sk_buff **pp = NULL; 2987 struct sk_buff *p; 2988 struct tcphdr *th; 2989 struct tcphdr *th2; 2990 unsigned int len; 2991 unsigned int thlen; 2992 __be32 flags; 2993 unsigned int mss = 1; 2994 unsigned int hlen; 2995 unsigned int off; 2996 int flush = 1; 2997 int i; 2998 2999 off = skb_gro_offset(skb); 3000 hlen = off + sizeof(*th); 3001 th = skb_gro_header_fast(skb, off); 3002 if (skb_gro_header_hard(skb, hlen)) { 3003 th = skb_gro_header_slow(skb, hlen, off); 3004 if (unlikely(!th)) 3005 goto out; 3006 } 3007 3008 thlen = th->doff * 4; 3009 if (thlen < sizeof(*th)) 3010 goto out; 3011 3012 hlen = off + thlen; 3013 if (skb_gro_header_hard(skb, hlen)) { 3014 th = skb_gro_header_slow(skb, hlen, off); 3015 if (unlikely(!th)) 3016 goto out; 3017 } 3018 3019 skb_gro_pull(skb, thlen); 3020 3021 len = skb_gro_len(skb); 3022 flags = tcp_flag_word(th); 3023 3024 for (; (p = *head); head = &p->next) { 3025 if (!NAPI_GRO_CB(p)->same_flow) 3026 continue; 3027 3028 th2 = tcp_hdr(p); 3029 3030 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) { 3031 NAPI_GRO_CB(p)->same_flow = 0; 3032 continue; 3033 } 3034 3035 goto found; 3036 } 3037 3038 goto out_check_final; 3039 3040found: 3041 flush = NAPI_GRO_CB(p)->flush; 3042 flush |= (__force int)(flags & TCP_FLAG_CWR); 3043 flush |= (__force int)((flags ^ tcp_flag_word(th2)) & 3044 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH)); 3045 flush |= (__force int)(th->ack_seq ^ th2->ack_seq); 3046 for (i = sizeof(*th); i < thlen; i += 4) 3047 flush |= *(u32 *)((u8 *)th + i) ^ 3048 *(u32 *)((u8 *)th2 + i); 3049 3050 mss = skb_shinfo(p)->gso_size; 3051 3052 flush |= (len - 1) >= mss; 3053 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq); 3054 3055 if (flush || skb_gro_receive(head, skb)) { 3056 mss = 1; 3057 goto out_check_final; 3058 } 3059 3060 p = *head; 3061 th2 = tcp_hdr(p); 3062 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH); 3063 3064out_check_final: 3065 flush = len < mss; 3066 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH | 3067 TCP_FLAG_RST | TCP_FLAG_SYN | 3068 TCP_FLAG_FIN)); 3069 3070 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush)) 3071 pp = head; 3072 3073out: 3074 NAPI_GRO_CB(skb)->flush |= flush; 3075 3076 return pp; 3077} 3078EXPORT_SYMBOL(tcp_gro_receive); 3079 3080int tcp_gro_complete(struct sk_buff *skb) 3081{ 3082 struct tcphdr *th = tcp_hdr(skb); 3083 3084 skb->csum_start = skb_transport_header(skb) - skb->head; 3085 skb->csum_offset = offsetof(struct tcphdr, check); 3086 skb->ip_summed = CHECKSUM_PARTIAL; 3087 3088 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count; 3089 3090 if (th->cwr) 3091 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 3092 3093 return 0; 3094} 3095EXPORT_SYMBOL(tcp_gro_complete); 3096 3097#ifdef CONFIG_TCP_MD5SIG 3098static unsigned long tcp_md5sig_users; 3099static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool; 3100static DEFINE_SPINLOCK(tcp_md5sig_pool_lock); 3101 3102static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool) 3103{ 3104 int cpu; 3105 3106 for_each_possible_cpu(cpu) { 3107 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu); 3108 3109 if (p->md5_desc.tfm) 3110 crypto_free_hash(p->md5_desc.tfm); 3111 } 3112 free_percpu(pool); 3113} 3114 3115void tcp_free_md5sig_pool(void) 3116{ 3117 struct tcp_md5sig_pool __percpu *pool = NULL; 3118 3119 spin_lock_bh(&tcp_md5sig_pool_lock); 3120 if (--tcp_md5sig_users == 0) { 3121 pool = tcp_md5sig_pool; 3122 tcp_md5sig_pool = NULL; 3123 } 3124 spin_unlock_bh(&tcp_md5sig_pool_lock); 3125 if (pool) 3126 __tcp_free_md5sig_pool(pool); 3127} 3128EXPORT_SYMBOL(tcp_free_md5sig_pool); 3129 3130static struct tcp_md5sig_pool __percpu * 3131__tcp_alloc_md5sig_pool(struct sock *sk) 3132{ 3133 int cpu; 3134 struct tcp_md5sig_pool __percpu *pool; 3135 3136 pool = alloc_percpu(struct tcp_md5sig_pool); 3137 if (!pool) 3138 return NULL; 3139 3140 for_each_possible_cpu(cpu) { 3141 struct crypto_hash *hash; 3142 3143 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC); 3144 if (!hash || IS_ERR(hash)) 3145 goto out_free; 3146 3147 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash; 3148 } 3149 return pool; 3150out_free: 3151 __tcp_free_md5sig_pool(pool); 3152 return NULL; 3153} 3154 3155struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk) 3156{ 3157 struct tcp_md5sig_pool __percpu *pool; 3158 int alloc = 0; 3159 3160retry: 3161 spin_lock_bh(&tcp_md5sig_pool_lock); 3162 pool = tcp_md5sig_pool; 3163 if (tcp_md5sig_users++ == 0) { 3164 alloc = 1; 3165 spin_unlock_bh(&tcp_md5sig_pool_lock); 3166 } else if (!pool) { 3167 tcp_md5sig_users--; 3168 spin_unlock_bh(&tcp_md5sig_pool_lock); 3169 cpu_relax(); 3170 goto retry; 3171 } else 3172 spin_unlock_bh(&tcp_md5sig_pool_lock); 3173 3174 if (alloc) { 3175 /* we cannot hold spinlock here because this may sleep. */ 3176 struct tcp_md5sig_pool __percpu *p; 3177 3178 p = __tcp_alloc_md5sig_pool(sk); 3179 spin_lock_bh(&tcp_md5sig_pool_lock); 3180 if (!p) { 3181 tcp_md5sig_users--; 3182 spin_unlock_bh(&tcp_md5sig_pool_lock); 3183 return NULL; 3184 } 3185 pool = tcp_md5sig_pool; 3186 if (pool) { 3187 /* oops, it has already been assigned. */ 3188 spin_unlock_bh(&tcp_md5sig_pool_lock); 3189 __tcp_free_md5sig_pool(p); 3190 } else { 3191 tcp_md5sig_pool = pool = p; 3192 spin_unlock_bh(&tcp_md5sig_pool_lock); 3193 } 3194 } 3195 return pool; 3196} 3197EXPORT_SYMBOL(tcp_alloc_md5sig_pool); 3198 3199 3200/** 3201 * tcp_get_md5sig_pool - get md5sig_pool for this user 3202 * 3203 * We use percpu structure, so if we succeed, we exit with preemption 3204 * and BH disabled, to make sure another thread or softirq handling 3205 * wont try to get same context. 3206 */ 3207struct tcp_md5sig_pool *tcp_get_md5sig_pool(void) 3208{ 3209 struct tcp_md5sig_pool __percpu *p; 3210 3211 local_bh_disable(); 3212 3213 spin_lock(&tcp_md5sig_pool_lock); 3214 p = tcp_md5sig_pool; 3215 if (p) 3216 tcp_md5sig_users++; 3217 spin_unlock(&tcp_md5sig_pool_lock); 3218 3219 if (p) 3220 return this_cpu_ptr(p); 3221 3222 local_bh_enable(); 3223 return NULL; 3224} 3225EXPORT_SYMBOL(tcp_get_md5sig_pool); 3226 3227void tcp_put_md5sig_pool(void) 3228{ 3229 local_bh_enable(); 3230 tcp_free_md5sig_pool(); 3231} 3232EXPORT_SYMBOL(tcp_put_md5sig_pool); 3233 3234int tcp_md5_hash_header(struct tcp_md5sig_pool *hp, 3235 const struct tcphdr *th) 3236{ 3237 struct scatterlist sg; 3238 struct tcphdr hdr; 3239 int err; 3240 3241 /* We are not allowed to change tcphdr, make a local copy */ 3242 memcpy(&hdr, th, sizeof(hdr)); 3243 hdr.check = 0; 3244 3245 /* options aren't included in the hash */ 3246 sg_init_one(&sg, &hdr, sizeof(hdr)); 3247 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr)); 3248 return err; 3249} 3250EXPORT_SYMBOL(tcp_md5_hash_header); 3251 3252int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp, 3253 const struct sk_buff *skb, unsigned int header_len) 3254{ 3255 struct scatterlist sg; 3256 const struct tcphdr *tp = tcp_hdr(skb); 3257 struct hash_desc *desc = &hp->md5_desc; 3258 unsigned int i; 3259 const unsigned int head_data_len = skb_headlen(skb) > header_len ? 3260 skb_headlen(skb) - header_len : 0; 3261 const struct skb_shared_info *shi = skb_shinfo(skb); 3262 struct sk_buff *frag_iter; 3263 3264 sg_init_table(&sg, 1); 3265 3266 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len); 3267 if (crypto_hash_update(desc, &sg, head_data_len)) 3268 return 1; 3269 3270 for (i = 0; i < shi->nr_frags; ++i) { 3271 const struct skb_frag_struct *f = &shi->frags[i]; 3272 struct page *page = skb_frag_page(f); 3273 sg_set_page(&sg, page, skb_frag_size(f), f->page_offset); 3274 if (crypto_hash_update(desc, &sg, skb_frag_size(f))) 3275 return 1; 3276 } 3277 3278 skb_walk_frags(skb, frag_iter) 3279 if (tcp_md5_hash_skb_data(hp, frag_iter, 0)) 3280 return 1; 3281 3282 return 0; 3283} 3284EXPORT_SYMBOL(tcp_md5_hash_skb_data); 3285 3286int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key) 3287{ 3288 struct scatterlist sg; 3289 3290 sg_init_one(&sg, key->key, key->keylen); 3291 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen); 3292} 3293EXPORT_SYMBOL(tcp_md5_hash_key); 3294 3295#endif 3296 3297/** 3298 * Each Responder maintains up to two secret values concurrently for 3299 * efficient secret rollover. Each secret value has 4 states: 3300 * 3301 * Generating. (tcp_secret_generating != tcp_secret_primary) 3302 * Generates new Responder-Cookies, but not yet used for primary 3303 * verification. This is a short-term state, typically lasting only 3304 * one round trip time (RTT). 3305 * 3306 * Primary. (tcp_secret_generating == tcp_secret_primary) 3307 * Used both for generation and primary verification. 3308 * 3309 * Retiring. (tcp_secret_retiring != tcp_secret_secondary) 3310 * Used for verification, until the first failure that can be 3311 * verified by the newer Generating secret. At that time, this 3312 * cookie's state is changed to Secondary, and the Generating 3313 * cookie's state is changed to Primary. This is a short-term state, 3314 * typically lasting only one round trip time (RTT). 3315 * 3316 * Secondary. (tcp_secret_retiring == tcp_secret_secondary) 3317 * Used for secondary verification, after primary verification 3318 * failures. This state lasts no more than twice the Maximum Segment 3319 * Lifetime (2MSL). Then, the secret is discarded. 3320 */ 3321struct tcp_cookie_secret { 3322 /* The secret is divided into two parts. The digest part is the 3323 * equivalent of previously hashing a secret and saving the state, 3324 * and serves as an initialization vector (IV). The message part 3325 * serves as the trailing secret. 3326 */ 3327 u32 secrets[COOKIE_WORKSPACE_WORDS]; 3328 unsigned long expires; 3329}; 3330 3331#define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL) 3332#define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2) 3333#define TCP_SECRET_LIFE (HZ * 600) 3334 3335static struct tcp_cookie_secret tcp_secret_one; 3336static struct tcp_cookie_secret tcp_secret_two; 3337 3338/* Essentially a circular list, without dynamic allocation. */ 3339static struct tcp_cookie_secret *tcp_secret_generating; 3340static struct tcp_cookie_secret *tcp_secret_primary; 3341static struct tcp_cookie_secret *tcp_secret_retiring; 3342static struct tcp_cookie_secret *tcp_secret_secondary; 3343 3344static DEFINE_SPINLOCK(tcp_secret_locker); 3345 3346/* Select a pseudo-random word in the cookie workspace. 3347 */ 3348static inline u32 tcp_cookie_work(const u32 *ws, const int n) 3349{ 3350 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])]; 3351} 3352 3353/* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed. 3354 * Called in softirq context. 3355 * Returns: 0 for success. 3356 */ 3357int tcp_cookie_generator(u32 *bakery) 3358{ 3359 unsigned long jiffy = jiffies; 3360 3361 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) { 3362 spin_lock_bh(&tcp_secret_locker); 3363 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) { 3364 /* refreshed by another */ 3365 memcpy(bakery, 3366 &tcp_secret_generating->secrets[0], 3367 COOKIE_WORKSPACE_WORDS); 3368 } else { 3369 /* still needs refreshing */ 3370 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS); 3371 3372 /* The first time, paranoia assumes that the 3373 * randomization function isn't as strong. But, 3374 * this secret initialization is delayed until 3375 * the last possible moment (packet arrival). 3376 * Although that time is observable, it is 3377 * unpredictably variable. Mash in the most 3378 * volatile clock bits available, and expire the 3379 * secret extra quickly. 3380 */ 3381 if (unlikely(tcp_secret_primary->expires == 3382 tcp_secret_secondary->expires)) { 3383 struct timespec tv; 3384 3385 getnstimeofday(&tv); 3386 bakery[COOKIE_DIGEST_WORDS+0] ^= 3387 (u32)tv.tv_nsec; 3388 3389 tcp_secret_secondary->expires = jiffy 3390 + TCP_SECRET_1MSL 3391 + (0x0f & tcp_cookie_work(bakery, 0)); 3392 } else { 3393 tcp_secret_secondary->expires = jiffy 3394 + TCP_SECRET_LIFE 3395 + (0xff & tcp_cookie_work(bakery, 1)); 3396 tcp_secret_primary->expires = jiffy 3397 + TCP_SECRET_2MSL 3398 + (0x1f & tcp_cookie_work(bakery, 2)); 3399 } 3400 memcpy(&tcp_secret_secondary->secrets[0], 3401 bakery, COOKIE_WORKSPACE_WORDS); 3402 3403 rcu_assign_pointer(tcp_secret_generating, 3404 tcp_secret_secondary); 3405 rcu_assign_pointer(tcp_secret_retiring, 3406 tcp_secret_primary); 3407 /* 3408 * Neither call_rcu() nor synchronize_rcu() needed. 3409 * Retiring data is not freed. It is replaced after 3410 * further (locked) pointer updates, and a quiet time 3411 * (minimum 1MSL, maximum LIFE - 2MSL). 3412 */ 3413 } 3414 spin_unlock_bh(&tcp_secret_locker); 3415 } else { 3416 rcu_read_lock_bh(); 3417 memcpy(bakery, 3418 &rcu_dereference(tcp_secret_generating)->secrets[0], 3419 COOKIE_WORKSPACE_WORDS); 3420 rcu_read_unlock_bh(); 3421 } 3422 return 0; 3423} 3424EXPORT_SYMBOL(tcp_cookie_generator); 3425 3426void tcp_done(struct sock *sk) 3427{ 3428 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV) 3429 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 3430 3431 tcp_set_state(sk, TCP_CLOSE); 3432 tcp_clear_xmit_timers(sk); 3433 3434 sk->sk_shutdown = SHUTDOWN_MASK; 3435 3436 if (!sock_flag(sk, SOCK_DEAD)) 3437 sk->sk_state_change(sk); 3438 else 3439 inet_csk_destroy_sock(sk); 3440} 3441EXPORT_SYMBOL_GPL(tcp_done); 3442 3443extern struct tcp_congestion_ops tcp_reno; 3444 3445static __initdata unsigned long thash_entries; 3446static int __init set_thash_entries(char *str) 3447{ 3448 if (!str) 3449 return 0; 3450 thash_entries = simple_strtoul(str, &str, 0); 3451 return 1; 3452} 3453__setup("thash_entries=", set_thash_entries); 3454 3455void tcp_init_mem(struct net *net) 3456{ 3457 unsigned long limit = nr_free_buffer_pages() / 8; 3458 limit = max(limit, 128UL); 3459 net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3; 3460 net->ipv4.sysctl_tcp_mem[1] = limit; 3461 net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2; 3462} 3463 3464void __init tcp_init(void) 3465{ 3466 struct sk_buff *skb = NULL; 3467 unsigned long limit; 3468 int max_share, cnt; 3469 unsigned int i; 3470 unsigned long jiffy = jiffies; 3471 3472 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb)); 3473 3474 percpu_counter_init(&tcp_sockets_allocated, 0); 3475 percpu_counter_init(&tcp_orphan_count, 0); 3476 tcp_hashinfo.bind_bucket_cachep = 3477 kmem_cache_create("tcp_bind_bucket", 3478 sizeof(struct inet_bind_bucket), 0, 3479 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL); 3480 3481 /* Size and allocate the main established and bind bucket 3482 * hash tables. 3483 * 3484 * The methodology is similar to that of the buffer cache. 3485 */ 3486 tcp_hashinfo.ehash = 3487 alloc_large_system_hash("TCP established", 3488 sizeof(struct inet_ehash_bucket), 3489 thash_entries, 3490 (totalram_pages >= 128 * 1024) ? 3491 13 : 15, 3492 0, 3493 NULL, 3494 &tcp_hashinfo.ehash_mask, 3495 thash_entries ? 0 : 512 * 1024); 3496 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) { 3497 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i); 3498 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i); 3499 } 3500 if (inet_ehash_locks_alloc(&tcp_hashinfo)) 3501 panic("TCP: failed to alloc ehash_locks"); 3502 tcp_hashinfo.bhash = 3503 alloc_large_system_hash("TCP bind", 3504 sizeof(struct inet_bind_hashbucket), 3505 tcp_hashinfo.ehash_mask + 1, 3506 (totalram_pages >= 128 * 1024) ? 3507 13 : 15, 3508 0, 3509 &tcp_hashinfo.bhash_size, 3510 NULL, 3511 64 * 1024); 3512 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size; 3513 for (i = 0; i < tcp_hashinfo.bhash_size; i++) { 3514 spin_lock_init(&tcp_hashinfo.bhash[i].lock); 3515 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain); 3516 } 3517 3518 3519 cnt = tcp_hashinfo.ehash_mask + 1; 3520 3521 tcp_death_row.sysctl_max_tw_buckets = cnt / 2; 3522 sysctl_tcp_max_orphans = cnt / 2; 3523 sysctl_max_syn_backlog = max(128, cnt / 256); 3524 3525 tcp_init_mem(&init_net); 3526 /* Set per-socket limits to no more than 1/128 the pressure threshold */ 3527 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7); 3528 max_share = min(4UL*1024*1024, limit); 3529 3530 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM; 3531 sysctl_tcp_wmem[1] = 16*1024; 3532 sysctl_tcp_wmem[2] = max(64*1024, max_share); 3533 3534 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM; 3535 sysctl_tcp_rmem[1] = 87380; 3536 sysctl_tcp_rmem[2] = max(87380, max_share); 3537 3538 pr_info("Hash tables configured (established %u bind %u)\n", 3539 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size); 3540 3541 tcp_register_congestion_control(&tcp_reno); 3542 3543 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets)); 3544 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets)); 3545 tcp_secret_one.expires = jiffy; /* past due */ 3546 tcp_secret_two.expires = jiffy; /* past due */ 3547 tcp_secret_generating = &tcp_secret_one; 3548 tcp_secret_primary = &tcp_secret_one; 3549 tcp_secret_retiring = &tcp_secret_two; 3550 tcp_secret_secondary = &tcp_secret_two; 3551} 3552