tcp.c revision 089c34827e52346f0303d1e6a7b744c1f4da3095
1/* 2 * INET An implementation of the TCP/IP protocol suite for the LINUX 3 * operating system. INET is implemented using the BSD Socket 4 * interface as the means of communication with the user level. 5 * 6 * Implementation of the Transmission Control Protocol(TCP). 7 * 8 * Authors: Ross Biro 9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 10 * Mark Evans, <evansmp@uhura.aston.ac.uk> 11 * Corey Minyard <wf-rch!minyard@relay.EU.net> 12 * Florian La Roche, <flla@stud.uni-sb.de> 13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 14 * Linus Torvalds, <torvalds@cs.helsinki.fi> 15 * Alan Cox, <gw4pts@gw4pts.ampr.org> 16 * Matthew Dillon, <dillon@apollo.west.oic.com> 17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 18 * Jorge Cwik, <jorge@laser.satlink.net> 19 * 20 * Fixes: 21 * Alan Cox : Numerous verify_area() calls 22 * Alan Cox : Set the ACK bit on a reset 23 * Alan Cox : Stopped it crashing if it closed while 24 * sk->inuse=1 and was trying to connect 25 * (tcp_err()). 26 * Alan Cox : All icmp error handling was broken 27 * pointers passed where wrong and the 28 * socket was looked up backwards. Nobody 29 * tested any icmp error code obviously. 30 * Alan Cox : tcp_err() now handled properly. It 31 * wakes people on errors. poll 32 * behaves and the icmp error race 33 * has gone by moving it into sock.c 34 * Alan Cox : tcp_send_reset() fixed to work for 35 * everything not just packets for 36 * unknown sockets. 37 * Alan Cox : tcp option processing. 38 * Alan Cox : Reset tweaked (still not 100%) [Had 39 * syn rule wrong] 40 * Herp Rosmanith : More reset fixes 41 * Alan Cox : No longer acks invalid rst frames. 42 * Acking any kind of RST is right out. 43 * Alan Cox : Sets an ignore me flag on an rst 44 * receive otherwise odd bits of prattle 45 * escape still 46 * Alan Cox : Fixed another acking RST frame bug. 47 * Should stop LAN workplace lockups. 48 * Alan Cox : Some tidyups using the new skb list 49 * facilities 50 * Alan Cox : sk->keepopen now seems to work 51 * Alan Cox : Pulls options out correctly on accepts 52 * Alan Cox : Fixed assorted sk->rqueue->next errors 53 * Alan Cox : PSH doesn't end a TCP read. Switched a 54 * bit to skb ops. 55 * Alan Cox : Tidied tcp_data to avoid a potential 56 * nasty. 57 * Alan Cox : Added some better commenting, as the 58 * tcp is hard to follow 59 * Alan Cox : Removed incorrect check for 20 * psh 60 * Michael O'Reilly : ack < copied bug fix. 61 * Johannes Stille : Misc tcp fixes (not all in yet). 62 * Alan Cox : FIN with no memory -> CRASH 63 * Alan Cox : Added socket option proto entries. 64 * Also added awareness of them to accept. 65 * Alan Cox : Added TCP options (SOL_TCP) 66 * Alan Cox : Switched wakeup calls to callbacks, 67 * so the kernel can layer network 68 * sockets. 69 * Alan Cox : Use ip_tos/ip_ttl settings. 70 * Alan Cox : Handle FIN (more) properly (we hope). 71 * Alan Cox : RST frames sent on unsynchronised 72 * state ack error. 73 * Alan Cox : Put in missing check for SYN bit. 74 * Alan Cox : Added tcp_select_window() aka NET2E 75 * window non shrink trick. 76 * Alan Cox : Added a couple of small NET2E timer 77 * fixes 78 * Charles Hedrick : TCP fixes 79 * Toomas Tamm : TCP window fixes 80 * Alan Cox : Small URG fix to rlogin ^C ack fight 81 * Charles Hedrick : Rewrote most of it to actually work 82 * Linus : Rewrote tcp_read() and URG handling 83 * completely 84 * Gerhard Koerting: Fixed some missing timer handling 85 * Matthew Dillon : Reworked TCP machine states as per RFC 86 * Gerhard Koerting: PC/TCP workarounds 87 * Adam Caldwell : Assorted timer/timing errors 88 * Matthew Dillon : Fixed another RST bug 89 * Alan Cox : Move to kernel side addressing changes. 90 * Alan Cox : Beginning work on TCP fastpathing 91 * (not yet usable) 92 * Arnt Gulbrandsen: Turbocharged tcp_check() routine. 93 * Alan Cox : TCP fast path debugging 94 * Alan Cox : Window clamping 95 * Michael Riepe : Bug in tcp_check() 96 * Matt Dillon : More TCP improvements and RST bug fixes 97 * Matt Dillon : Yet more small nasties remove from the 98 * TCP code (Be very nice to this man if 99 * tcp finally works 100%) 8) 100 * Alan Cox : BSD accept semantics. 101 * Alan Cox : Reset on closedown bug. 102 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto(). 103 * Michael Pall : Handle poll() after URG properly in 104 * all cases. 105 * Michael Pall : Undo the last fix in tcp_read_urg() 106 * (multi URG PUSH broke rlogin). 107 * Michael Pall : Fix the multi URG PUSH problem in 108 * tcp_readable(), poll() after URG 109 * works now. 110 * Michael Pall : recv(...,MSG_OOB) never blocks in the 111 * BSD api. 112 * Alan Cox : Changed the semantics of sk->socket to 113 * fix a race and a signal problem with 114 * accept() and async I/O. 115 * Alan Cox : Relaxed the rules on tcp_sendto(). 116 * Yury Shevchuk : Really fixed accept() blocking problem. 117 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for 118 * clients/servers which listen in on 119 * fixed ports. 120 * Alan Cox : Cleaned the above up and shrank it to 121 * a sensible code size. 122 * Alan Cox : Self connect lockup fix. 123 * Alan Cox : No connect to multicast. 124 * Ross Biro : Close unaccepted children on master 125 * socket close. 126 * Alan Cox : Reset tracing code. 127 * Alan Cox : Spurious resets on shutdown. 128 * Alan Cox : Giant 15 minute/60 second timer error 129 * Alan Cox : Small whoops in polling before an 130 * accept. 131 * Alan Cox : Kept the state trace facility since 132 * it's handy for debugging. 133 * Alan Cox : More reset handler fixes. 134 * Alan Cox : Started rewriting the code based on 135 * the RFC's for other useful protocol 136 * references see: Comer, KA9Q NOS, and 137 * for a reference on the difference 138 * between specifications and how BSD 139 * works see the 4.4lite source. 140 * A.N.Kuznetsov : Don't time wait on completion of tidy 141 * close. 142 * Linus Torvalds : Fin/Shutdown & copied_seq changes. 143 * Linus Torvalds : Fixed BSD port reuse to work first syn 144 * Alan Cox : Reimplemented timers as per the RFC 145 * and using multiple timers for sanity. 146 * Alan Cox : Small bug fixes, and a lot of new 147 * comments. 148 * Alan Cox : Fixed dual reader crash by locking 149 * the buffers (much like datagram.c) 150 * Alan Cox : Fixed stuck sockets in probe. A probe 151 * now gets fed up of retrying without 152 * (even a no space) answer. 153 * Alan Cox : Extracted closing code better 154 * Alan Cox : Fixed the closing state machine to 155 * resemble the RFC. 156 * Alan Cox : More 'per spec' fixes. 157 * Jorge Cwik : Even faster checksumming. 158 * Alan Cox : tcp_data() doesn't ack illegal PSH 159 * only frames. At least one pc tcp stack 160 * generates them. 161 * Alan Cox : Cache last socket. 162 * Alan Cox : Per route irtt. 163 * Matt Day : poll()->select() match BSD precisely on error 164 * Alan Cox : New buffers 165 * Marc Tamsky : Various sk->prot->retransmits and 166 * sk->retransmits misupdating fixed. 167 * Fixed tcp_write_timeout: stuck close, 168 * and TCP syn retries gets used now. 169 * Mark Yarvis : In tcp_read_wakeup(), don't send an 170 * ack if state is TCP_CLOSED. 171 * Alan Cox : Look up device on a retransmit - routes may 172 * change. Doesn't yet cope with MSS shrink right 173 * but it's a start! 174 * Marc Tamsky : Closing in closing fixes. 175 * Mike Shaver : RFC1122 verifications. 176 * Alan Cox : rcv_saddr errors. 177 * Alan Cox : Block double connect(). 178 * Alan Cox : Small hooks for enSKIP. 179 * Alexey Kuznetsov: Path MTU discovery. 180 * Alan Cox : Support soft errors. 181 * Alan Cox : Fix MTU discovery pathological case 182 * when the remote claims no mtu! 183 * Marc Tamsky : TCP_CLOSE fix. 184 * Colin (G3TNE) : Send a reset on syn ack replies in 185 * window but wrong (fixes NT lpd problems) 186 * Pedro Roque : Better TCP window handling, delayed ack. 187 * Joerg Reuter : No modification of locked buffers in 188 * tcp_do_retransmit() 189 * Eric Schenk : Changed receiver side silly window 190 * avoidance algorithm to BSD style 191 * algorithm. This doubles throughput 192 * against machines running Solaris, 193 * and seems to result in general 194 * improvement. 195 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD 196 * Willy Konynenberg : Transparent proxying support. 197 * Mike McLagan : Routing by source 198 * Keith Owens : Do proper merging with partial SKB's in 199 * tcp_do_sendmsg to avoid burstiness. 200 * Eric Schenk : Fix fast close down bug with 201 * shutdown() followed by close(). 202 * Andi Kleen : Make poll agree with SIGIO 203 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and 204 * lingertime == 0 (RFC 793 ABORT Call) 205 * Hirokazu Takahashi : Use copy_from_user() instead of 206 * csum_and_copy_from_user() if possible. 207 * 208 * This program is free software; you can redistribute it and/or 209 * modify it under the terms of the GNU General Public License 210 * as published by the Free Software Foundation; either version 211 * 2 of the License, or(at your option) any later version. 212 * 213 * Description of States: 214 * 215 * TCP_SYN_SENT sent a connection request, waiting for ack 216 * 217 * TCP_SYN_RECV received a connection request, sent ack, 218 * waiting for final ack in three-way handshake. 219 * 220 * TCP_ESTABLISHED connection established 221 * 222 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete 223 * transmission of remaining buffered data 224 * 225 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote 226 * to shutdown 227 * 228 * TCP_CLOSING both sides have shutdown but we still have 229 * data we have to finish sending 230 * 231 * TCP_TIME_WAIT timeout to catch resent junk before entering 232 * closed, can only be entered from FIN_WAIT2 233 * or CLOSING. Required because the other end 234 * may not have gotten our last ACK causing it 235 * to retransmit the data packet (which we ignore) 236 * 237 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for 238 * us to finish writing our data and to shutdown 239 * (we have to close() to move on to LAST_ACK) 240 * 241 * TCP_LAST_ACK out side has shutdown after remote has 242 * shutdown. There may still be data in our 243 * buffer that we have to finish sending 244 * 245 * TCP_CLOSE socket is finished 246 */ 247 248#include <linux/kernel.h> 249#include <linux/module.h> 250#include <linux/types.h> 251#include <linux/fcntl.h> 252#include <linux/poll.h> 253#include <linux/init.h> 254#include <linux/fs.h> 255#include <linux/skbuff.h> 256#include <linux/scatterlist.h> 257#include <linux/splice.h> 258#include <linux/net.h> 259#include <linux/socket.h> 260#include <linux/random.h> 261#include <linux/bootmem.h> 262#include <linux/highmem.h> 263#include <linux/swap.h> 264#include <linux/cache.h> 265#include <linux/err.h> 266#include <linux/crypto.h> 267#include <linux/time.h> 268#include <linux/slab.h> 269 270#include <net/icmp.h> 271#include <net/tcp.h> 272#include <net/xfrm.h> 273#include <net/ip.h> 274#include <net/netdma.h> 275#include <net/sock.h> 276 277#include <asm/uaccess.h> 278#include <asm/ioctls.h> 279 280int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT; 281 282struct percpu_counter tcp_orphan_count; 283EXPORT_SYMBOL_GPL(tcp_orphan_count); 284 285long sysctl_tcp_mem[3] __read_mostly; 286int sysctl_tcp_wmem[3] __read_mostly; 287int sysctl_tcp_rmem[3] __read_mostly; 288 289EXPORT_SYMBOL(sysctl_tcp_mem); 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 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 default: 509 return -ENOIOCTLCMD; 510 } 511 512 return put_user(answ, (int __user *)arg); 513} 514EXPORT_SYMBOL(tcp_ioctl); 515 516static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb) 517{ 518 TCP_SKB_CB(skb)->flags |= TCPHDR_PSH; 519 tp->pushed_seq = tp->write_seq; 520} 521 522static inline int forced_push(struct tcp_sock *tp) 523{ 524 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1)); 525} 526 527static inline void skb_entail(struct sock *sk, struct sk_buff *skb) 528{ 529 struct tcp_sock *tp = tcp_sk(sk); 530 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 531 532 skb->csum = 0; 533 tcb->seq = tcb->end_seq = tp->write_seq; 534 tcb->flags = TCPHDR_ACK; 535 tcb->sacked = 0; 536 skb_header_release(skb); 537 tcp_add_write_queue_tail(sk, skb); 538 sk->sk_wmem_queued += skb->truesize; 539 sk_mem_charge(sk, skb->truesize); 540 if (tp->nonagle & TCP_NAGLE_PUSH) 541 tp->nonagle &= ~TCP_NAGLE_PUSH; 542} 543 544static inline void tcp_mark_urg(struct tcp_sock *tp, int flags) 545{ 546 if (flags & MSG_OOB) 547 tp->snd_up = tp->write_seq; 548} 549 550static inline void tcp_push(struct sock *sk, int flags, int mss_now, 551 int nonagle) 552{ 553 if (tcp_send_head(sk)) { 554 struct tcp_sock *tp = tcp_sk(sk); 555 556 if (!(flags & MSG_MORE) || forced_push(tp)) 557 tcp_mark_push(tp, tcp_write_queue_tail(sk)); 558 559 tcp_mark_urg(tp, flags); 560 __tcp_push_pending_frames(sk, mss_now, 561 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle); 562 } 563} 564 565static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, 566 unsigned int offset, size_t len) 567{ 568 struct tcp_splice_state *tss = rd_desc->arg.data; 569 int ret; 570 571 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len), 572 tss->flags); 573 if (ret > 0) 574 rd_desc->count -= ret; 575 return ret; 576} 577 578static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss) 579{ 580 /* Store TCP splice context information in read_descriptor_t. */ 581 read_descriptor_t rd_desc = { 582 .arg.data = tss, 583 .count = tss->len, 584 }; 585 586 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv); 587} 588 589/** 590 * tcp_splice_read - splice data from TCP socket to a pipe 591 * @sock: socket to splice from 592 * @ppos: position (not valid) 593 * @pipe: pipe to splice to 594 * @len: number of bytes to splice 595 * @flags: splice modifier flags 596 * 597 * Description: 598 * Will read pages from given socket and fill them into a pipe. 599 * 600 **/ 601ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos, 602 struct pipe_inode_info *pipe, size_t len, 603 unsigned int flags) 604{ 605 struct sock *sk = sock->sk; 606 struct tcp_splice_state tss = { 607 .pipe = pipe, 608 .len = len, 609 .flags = flags, 610 }; 611 long timeo; 612 ssize_t spliced; 613 int ret; 614 615 sock_rps_record_flow(sk); 616 /* 617 * We can't seek on a socket input 618 */ 619 if (unlikely(*ppos)) 620 return -ESPIPE; 621 622 ret = spliced = 0; 623 624 lock_sock(sk); 625 626 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK); 627 while (tss.len) { 628 ret = __tcp_splice_read(sk, &tss); 629 if (ret < 0) 630 break; 631 else if (!ret) { 632 if (spliced) 633 break; 634 if (sock_flag(sk, SOCK_DONE)) 635 break; 636 if (sk->sk_err) { 637 ret = sock_error(sk); 638 break; 639 } 640 if (sk->sk_shutdown & RCV_SHUTDOWN) 641 break; 642 if (sk->sk_state == TCP_CLOSE) { 643 /* 644 * This occurs when user tries to read 645 * from never connected socket. 646 */ 647 if (!sock_flag(sk, SOCK_DONE)) 648 ret = -ENOTCONN; 649 break; 650 } 651 if (!timeo) { 652 ret = -EAGAIN; 653 break; 654 } 655 sk_wait_data(sk, &timeo); 656 if (signal_pending(current)) { 657 ret = sock_intr_errno(timeo); 658 break; 659 } 660 continue; 661 } 662 tss.len -= ret; 663 spliced += ret; 664 665 if (!timeo) 666 break; 667 release_sock(sk); 668 lock_sock(sk); 669 670 if (sk->sk_err || sk->sk_state == TCP_CLOSE || 671 (sk->sk_shutdown & RCV_SHUTDOWN) || 672 signal_pending(current)) 673 break; 674 } 675 676 release_sock(sk); 677 678 if (spliced) 679 return spliced; 680 681 return ret; 682} 683EXPORT_SYMBOL(tcp_splice_read); 684 685struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp) 686{ 687 struct sk_buff *skb; 688 689 /* The TCP header must be at least 32-bit aligned. */ 690 size = ALIGN(size, 4); 691 692 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp); 693 if (skb) { 694 if (sk_wmem_schedule(sk, skb->truesize)) { 695 /* 696 * Make sure that we have exactly size bytes 697 * available to the caller, no more, no less. 698 */ 699 skb_reserve(skb, skb_tailroom(skb) - size); 700 return skb; 701 } 702 __kfree_skb(skb); 703 } else { 704 sk->sk_prot->enter_memory_pressure(sk); 705 sk_stream_moderate_sndbuf(sk); 706 } 707 return NULL; 708} 709 710static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now, 711 int large_allowed) 712{ 713 struct tcp_sock *tp = tcp_sk(sk); 714 u32 xmit_size_goal, old_size_goal; 715 716 xmit_size_goal = mss_now; 717 718 if (large_allowed && sk_can_gso(sk)) { 719 xmit_size_goal = ((sk->sk_gso_max_size - 1) - 720 inet_csk(sk)->icsk_af_ops->net_header_len - 721 inet_csk(sk)->icsk_ext_hdr_len - 722 tp->tcp_header_len); 723 724 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal); 725 726 /* We try hard to avoid divides here */ 727 old_size_goal = tp->xmit_size_goal_segs * mss_now; 728 729 if (likely(old_size_goal <= xmit_size_goal && 730 old_size_goal + mss_now > xmit_size_goal)) { 731 xmit_size_goal = old_size_goal; 732 } else { 733 tp->xmit_size_goal_segs = xmit_size_goal / mss_now; 734 xmit_size_goal = tp->xmit_size_goal_segs * mss_now; 735 } 736 } 737 738 return max(xmit_size_goal, mss_now); 739} 740 741static int tcp_send_mss(struct sock *sk, int *size_goal, int flags) 742{ 743 int mss_now; 744 745 mss_now = tcp_current_mss(sk); 746 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB)); 747 748 return mss_now; 749} 750 751static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset, 752 size_t psize, int flags) 753{ 754 struct tcp_sock *tp = tcp_sk(sk); 755 int mss_now, size_goal; 756 int err; 757 ssize_t copied; 758 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 759 760 /* Wait for a connection to finish. */ 761 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 762 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0) 763 goto out_err; 764 765 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags); 766 767 mss_now = tcp_send_mss(sk, &size_goal, flags); 768 copied = 0; 769 770 err = -EPIPE; 771 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 772 goto out_err; 773 774 while (psize > 0) { 775 struct sk_buff *skb = tcp_write_queue_tail(sk); 776 struct page *page = pages[poffset / PAGE_SIZE]; 777 int copy, i, can_coalesce; 778 int offset = poffset % PAGE_SIZE; 779 int size = min_t(size_t, psize, PAGE_SIZE - offset); 780 781 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) { 782new_segment: 783 if (!sk_stream_memory_free(sk)) 784 goto wait_for_sndbuf; 785 786 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation); 787 if (!skb) 788 goto wait_for_memory; 789 790 skb_entail(sk, skb); 791 copy = size_goal; 792 } 793 794 if (copy > size) 795 copy = size; 796 797 i = skb_shinfo(skb)->nr_frags; 798 can_coalesce = skb_can_coalesce(skb, i, page, offset); 799 if (!can_coalesce && i >= MAX_SKB_FRAGS) { 800 tcp_mark_push(tp, skb); 801 goto new_segment; 802 } 803 if (!sk_wmem_schedule(sk, copy)) 804 goto wait_for_memory; 805 806 if (can_coalesce) { 807 skb_shinfo(skb)->frags[i - 1].size += copy; 808 } else { 809 get_page(page); 810 skb_fill_page_desc(skb, i, page, offset, copy); 811 } 812 813 skb->len += copy; 814 skb->data_len += copy; 815 skb->truesize += copy; 816 sk->sk_wmem_queued += copy; 817 sk_mem_charge(sk, copy); 818 skb->ip_summed = CHECKSUM_PARTIAL; 819 tp->write_seq += copy; 820 TCP_SKB_CB(skb)->end_seq += copy; 821 skb_shinfo(skb)->gso_segs = 0; 822 823 if (!copied) 824 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH; 825 826 copied += copy; 827 poffset += copy; 828 if (!(psize -= copy)) 829 goto out; 830 831 if (skb->len < size_goal || (flags & MSG_OOB)) 832 continue; 833 834 if (forced_push(tp)) { 835 tcp_mark_push(tp, skb); 836 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 837 } else if (skb == tcp_send_head(sk)) 838 tcp_push_one(sk, mss_now); 839 continue; 840 841wait_for_sndbuf: 842 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 843wait_for_memory: 844 if (copied) 845 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH); 846 847 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0) 848 goto do_error; 849 850 mss_now = tcp_send_mss(sk, &size_goal, flags); 851 } 852 853out: 854 if (copied) 855 tcp_push(sk, flags, mss_now, tp->nonagle); 856 return copied; 857 858do_error: 859 if (copied) 860 goto out; 861out_err: 862 return sk_stream_error(sk, flags, err); 863} 864 865int tcp_sendpage(struct sock *sk, struct page *page, int offset, 866 size_t size, int flags) 867{ 868 ssize_t res; 869 870 if (!(sk->sk_route_caps & NETIF_F_SG) || 871 !(sk->sk_route_caps & NETIF_F_ALL_CSUM)) 872 return sock_no_sendpage(sk->sk_socket, page, offset, size, 873 flags); 874 875 lock_sock(sk); 876 res = do_tcp_sendpages(sk, &page, offset, size, flags); 877 release_sock(sk); 878 return res; 879} 880EXPORT_SYMBOL(tcp_sendpage); 881 882#define TCP_PAGE(sk) (sk->sk_sndmsg_page) 883#define TCP_OFF(sk) (sk->sk_sndmsg_off) 884 885static inline int select_size(struct sock *sk, int sg) 886{ 887 struct tcp_sock *tp = tcp_sk(sk); 888 int tmp = tp->mss_cache; 889 890 if (sg) { 891 if (sk_can_gso(sk)) 892 tmp = 0; 893 else { 894 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER); 895 896 if (tmp >= pgbreak && 897 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE) 898 tmp = pgbreak; 899 } 900 } 901 902 return tmp; 903} 904 905int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg, 906 size_t size) 907{ 908 struct iovec *iov; 909 struct tcp_sock *tp = tcp_sk(sk); 910 struct sk_buff *skb; 911 int iovlen, flags; 912 int mss_now, size_goal; 913 int sg, err, copied; 914 long timeo; 915 916 lock_sock(sk); 917 918 flags = msg->msg_flags; 919 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 920 921 /* Wait for a connection to finish. */ 922 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 923 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0) 924 goto out_err; 925 926 /* This should be in poll */ 927 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags); 928 929 mss_now = tcp_send_mss(sk, &size_goal, flags); 930 931 /* Ok commence sending. */ 932 iovlen = msg->msg_iovlen; 933 iov = msg->msg_iov; 934 copied = 0; 935 936 err = -EPIPE; 937 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 938 goto out_err; 939 940 sg = sk->sk_route_caps & NETIF_F_SG; 941 942 while (--iovlen >= 0) { 943 size_t seglen = iov->iov_len; 944 unsigned char __user *from = iov->iov_base; 945 946 iov++; 947 948 while (seglen > 0) { 949 int copy = 0; 950 int max = size_goal; 951 952 skb = tcp_write_queue_tail(sk); 953 if (tcp_send_head(sk)) { 954 if (skb->ip_summed == CHECKSUM_NONE) 955 max = mss_now; 956 copy = max - skb->len; 957 } 958 959 if (copy <= 0) { 960new_segment: 961 /* Allocate new segment. If the interface is SG, 962 * allocate skb fitting to single page. 963 */ 964 if (!sk_stream_memory_free(sk)) 965 goto wait_for_sndbuf; 966 967 skb = sk_stream_alloc_skb(sk, 968 select_size(sk, sg), 969 sk->sk_allocation); 970 if (!skb) 971 goto wait_for_memory; 972 973 /* 974 * Check whether we can use HW checksum. 975 */ 976 if (sk->sk_route_caps & NETIF_F_ALL_CSUM) 977 skb->ip_summed = CHECKSUM_PARTIAL; 978 979 skb_entail(sk, skb); 980 copy = size_goal; 981 max = size_goal; 982 } 983 984 /* Try to append data to the end of skb. */ 985 if (copy > seglen) 986 copy = seglen; 987 988 /* Where to copy to? */ 989 if (skb_tailroom(skb) > 0) { 990 /* We have some space in skb head. Superb! */ 991 if (copy > skb_tailroom(skb)) 992 copy = skb_tailroom(skb); 993 if ((err = skb_add_data(skb, from, copy)) != 0) 994 goto do_fault; 995 } else { 996 int merge = 0; 997 int i = skb_shinfo(skb)->nr_frags; 998 struct page *page = TCP_PAGE(sk); 999 int off = TCP_OFF(sk); 1000 1001 if (skb_can_coalesce(skb, i, page, off) && 1002 off != PAGE_SIZE) { 1003 /* We can extend the last page 1004 * fragment. */ 1005 merge = 1; 1006 } else if (i == MAX_SKB_FRAGS || !sg) { 1007 /* Need to add new fragment and cannot 1008 * do this because interface is non-SG, 1009 * or because all the page slots are 1010 * busy. */ 1011 tcp_mark_push(tp, skb); 1012 goto new_segment; 1013 } else if (page) { 1014 if (off == PAGE_SIZE) { 1015 put_page(page); 1016 TCP_PAGE(sk) = page = NULL; 1017 off = 0; 1018 } 1019 } else 1020 off = 0; 1021 1022 if (copy > PAGE_SIZE - off) 1023 copy = PAGE_SIZE - off; 1024 1025 if (!sk_wmem_schedule(sk, copy)) 1026 goto wait_for_memory; 1027 1028 if (!page) { 1029 /* Allocate new cache page. */ 1030 if (!(page = sk_stream_alloc_page(sk))) 1031 goto wait_for_memory; 1032 } 1033 1034 /* Time to copy data. We are close to 1035 * the end! */ 1036 err = skb_copy_to_page(sk, from, skb, page, 1037 off, copy); 1038 if (err) { 1039 /* If this page was new, give it to the 1040 * socket so it does not get leaked. 1041 */ 1042 if (!TCP_PAGE(sk)) { 1043 TCP_PAGE(sk) = page; 1044 TCP_OFF(sk) = 0; 1045 } 1046 goto do_error; 1047 } 1048 1049 /* Update the skb. */ 1050 if (merge) { 1051 skb_shinfo(skb)->frags[i - 1].size += 1052 copy; 1053 } else { 1054 skb_fill_page_desc(skb, i, page, off, copy); 1055 if (TCP_PAGE(sk)) { 1056 get_page(page); 1057 } else if (off + copy < PAGE_SIZE) { 1058 get_page(page); 1059 TCP_PAGE(sk) = page; 1060 } 1061 } 1062 1063 TCP_OFF(sk) = off + copy; 1064 } 1065 1066 if (!copied) 1067 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH; 1068 1069 tp->write_seq += copy; 1070 TCP_SKB_CB(skb)->end_seq += copy; 1071 skb_shinfo(skb)->gso_segs = 0; 1072 1073 from += copy; 1074 copied += copy; 1075 if ((seglen -= copy) == 0 && iovlen == 0) 1076 goto out; 1077 1078 if (skb->len < max || (flags & MSG_OOB)) 1079 continue; 1080 1081 if (forced_push(tp)) { 1082 tcp_mark_push(tp, skb); 1083 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 1084 } else if (skb == tcp_send_head(sk)) 1085 tcp_push_one(sk, mss_now); 1086 continue; 1087 1088wait_for_sndbuf: 1089 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1090wait_for_memory: 1091 if (copied) 1092 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH); 1093 1094 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0) 1095 goto do_error; 1096 1097 mss_now = tcp_send_mss(sk, &size_goal, flags); 1098 } 1099 } 1100 1101out: 1102 if (copied) 1103 tcp_push(sk, flags, mss_now, tp->nonagle); 1104 release_sock(sk); 1105 return copied; 1106 1107do_fault: 1108 if (!skb->len) { 1109 tcp_unlink_write_queue(skb, sk); 1110 /* It is the one place in all of TCP, except connection 1111 * reset, where we can be unlinking the send_head. 1112 */ 1113 tcp_check_send_head(sk, skb); 1114 sk_wmem_free_skb(sk, skb); 1115 } 1116 1117do_error: 1118 if (copied) 1119 goto out; 1120out_err: 1121 err = sk_stream_error(sk, flags, err); 1122 release_sock(sk); 1123 return err; 1124} 1125EXPORT_SYMBOL(tcp_sendmsg); 1126 1127/* 1128 * Handle reading urgent data. BSD has very simple semantics for 1129 * this, no blocking and very strange errors 8) 1130 */ 1131 1132static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags) 1133{ 1134 struct tcp_sock *tp = tcp_sk(sk); 1135 1136 /* No URG data to read. */ 1137 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data || 1138 tp->urg_data == TCP_URG_READ) 1139 return -EINVAL; /* Yes this is right ! */ 1140 1141 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE)) 1142 return -ENOTCONN; 1143 1144 if (tp->urg_data & TCP_URG_VALID) { 1145 int err = 0; 1146 char c = tp->urg_data; 1147 1148 if (!(flags & MSG_PEEK)) 1149 tp->urg_data = TCP_URG_READ; 1150 1151 /* Read urgent data. */ 1152 msg->msg_flags |= MSG_OOB; 1153 1154 if (len > 0) { 1155 if (!(flags & MSG_TRUNC)) 1156 err = memcpy_toiovec(msg->msg_iov, &c, 1); 1157 len = 1; 1158 } else 1159 msg->msg_flags |= MSG_TRUNC; 1160 1161 return err ? -EFAULT : len; 1162 } 1163 1164 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 1165 return 0; 1166 1167 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and 1168 * the available implementations agree in this case: 1169 * this call should never block, independent of the 1170 * blocking state of the socket. 1171 * Mike <pall@rz.uni-karlsruhe.de> 1172 */ 1173 return -EAGAIN; 1174} 1175 1176/* Clean up the receive buffer for full frames taken by the user, 1177 * then send an ACK if necessary. COPIED is the number of bytes 1178 * tcp_recvmsg has given to the user so far, it speeds up the 1179 * calculation of whether or not we must ACK for the sake of 1180 * a window update. 1181 */ 1182void tcp_cleanup_rbuf(struct sock *sk, int copied) 1183{ 1184 struct tcp_sock *tp = tcp_sk(sk); 1185 int time_to_ack = 0; 1186 1187#if TCP_DEBUG 1188 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 1189 1190 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq), 1191 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n", 1192 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt); 1193#endif 1194 1195 if (inet_csk_ack_scheduled(sk)) { 1196 const struct inet_connection_sock *icsk = inet_csk(sk); 1197 /* Delayed ACKs frequently hit locked sockets during bulk 1198 * receive. */ 1199 if (icsk->icsk_ack.blocked || 1200 /* Once-per-two-segments ACK was not sent by tcp_input.c */ 1201 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss || 1202 /* 1203 * If this read emptied read buffer, we send ACK, if 1204 * connection is not bidirectional, user drained 1205 * receive buffer and there was a small segment 1206 * in queue. 1207 */ 1208 (copied > 0 && 1209 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) || 1210 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) && 1211 !icsk->icsk_ack.pingpong)) && 1212 !atomic_read(&sk->sk_rmem_alloc))) 1213 time_to_ack = 1; 1214 } 1215 1216 /* We send an ACK if we can now advertise a non-zero window 1217 * which has been raised "significantly". 1218 * 1219 * Even if window raised up to infinity, do not send window open ACK 1220 * in states, where we will not receive more. It is useless. 1221 */ 1222 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) { 1223 __u32 rcv_window_now = tcp_receive_window(tp); 1224 1225 /* Optimize, __tcp_select_window() is not cheap. */ 1226 if (2*rcv_window_now <= tp->window_clamp) { 1227 __u32 new_window = __tcp_select_window(sk); 1228 1229 /* Send ACK now, if this read freed lots of space 1230 * in our buffer. Certainly, new_window is new window. 1231 * We can advertise it now, if it is not less than current one. 1232 * "Lots" means "at least twice" here. 1233 */ 1234 if (new_window && new_window >= 2 * rcv_window_now) 1235 time_to_ack = 1; 1236 } 1237 } 1238 if (time_to_ack) 1239 tcp_send_ack(sk); 1240} 1241 1242static void tcp_prequeue_process(struct sock *sk) 1243{ 1244 struct sk_buff *skb; 1245 struct tcp_sock *tp = tcp_sk(sk); 1246 1247 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED); 1248 1249 /* RX process wants to run with disabled BHs, though it is not 1250 * necessary */ 1251 local_bh_disable(); 1252 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL) 1253 sk_backlog_rcv(sk, skb); 1254 local_bh_enable(); 1255 1256 /* Clear memory counter. */ 1257 tp->ucopy.memory = 0; 1258} 1259 1260#ifdef CONFIG_NET_DMA 1261static void tcp_service_net_dma(struct sock *sk, bool wait) 1262{ 1263 dma_cookie_t done, used; 1264 dma_cookie_t last_issued; 1265 struct tcp_sock *tp = tcp_sk(sk); 1266 1267 if (!tp->ucopy.dma_chan) 1268 return; 1269 1270 last_issued = tp->ucopy.dma_cookie; 1271 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1272 1273 do { 1274 if (dma_async_memcpy_complete(tp->ucopy.dma_chan, 1275 last_issued, &done, 1276 &used) == DMA_SUCCESS) { 1277 /* Safe to free early-copied skbs now */ 1278 __skb_queue_purge(&sk->sk_async_wait_queue); 1279 break; 1280 } else { 1281 struct sk_buff *skb; 1282 while ((skb = skb_peek(&sk->sk_async_wait_queue)) && 1283 (dma_async_is_complete(skb->dma_cookie, done, 1284 used) == DMA_SUCCESS)) { 1285 __skb_dequeue(&sk->sk_async_wait_queue); 1286 kfree_skb(skb); 1287 } 1288 } 1289 } while (wait); 1290} 1291#endif 1292 1293static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off) 1294{ 1295 struct sk_buff *skb; 1296 u32 offset; 1297 1298 skb_queue_walk(&sk->sk_receive_queue, skb) { 1299 offset = seq - TCP_SKB_CB(skb)->seq; 1300 if (tcp_hdr(skb)->syn) 1301 offset--; 1302 if (offset < skb->len || tcp_hdr(skb)->fin) { 1303 *off = offset; 1304 return skb; 1305 } 1306 } 1307 return NULL; 1308} 1309 1310/* 1311 * This routine provides an alternative to tcp_recvmsg() for routines 1312 * that would like to handle copying from skbuffs directly in 'sendfile' 1313 * fashion. 1314 * Note: 1315 * - It is assumed that the socket was locked by the caller. 1316 * - The routine does not block. 1317 * - At present, there is no support for reading OOB data 1318 * or for 'peeking' the socket using this routine 1319 * (although both would be easy to implement). 1320 */ 1321int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, 1322 sk_read_actor_t recv_actor) 1323{ 1324 struct sk_buff *skb; 1325 struct tcp_sock *tp = tcp_sk(sk); 1326 u32 seq = tp->copied_seq; 1327 u32 offset; 1328 int copied = 0; 1329 1330 if (sk->sk_state == TCP_LISTEN) 1331 return -ENOTCONN; 1332 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) { 1333 if (offset < skb->len) { 1334 int used; 1335 size_t len; 1336 1337 len = skb->len - offset; 1338 /* Stop reading if we hit a patch of urgent data */ 1339 if (tp->urg_data) { 1340 u32 urg_offset = tp->urg_seq - seq; 1341 if (urg_offset < len) 1342 len = urg_offset; 1343 if (!len) 1344 break; 1345 } 1346 used = recv_actor(desc, skb, offset, len); 1347 if (used < 0) { 1348 if (!copied) 1349 copied = used; 1350 break; 1351 } else if (used <= len) { 1352 seq += used; 1353 copied += used; 1354 offset += used; 1355 } 1356 /* 1357 * If recv_actor drops the lock (e.g. TCP splice 1358 * receive) the skb pointer might be invalid when 1359 * getting here: tcp_collapse might have deleted it 1360 * while aggregating skbs from the socket queue. 1361 */ 1362 skb = tcp_recv_skb(sk, seq-1, &offset); 1363 if (!skb || (offset+1 != skb->len)) 1364 break; 1365 } 1366 if (tcp_hdr(skb)->fin) { 1367 sk_eat_skb(sk, skb, 0); 1368 ++seq; 1369 break; 1370 } 1371 sk_eat_skb(sk, skb, 0); 1372 if (!desc->count) 1373 break; 1374 tp->copied_seq = seq; 1375 } 1376 tp->copied_seq = seq; 1377 1378 tcp_rcv_space_adjust(sk); 1379 1380 /* Clean up data we have read: This will do ACK frames. */ 1381 if (copied > 0) 1382 tcp_cleanup_rbuf(sk, copied); 1383 return copied; 1384} 1385EXPORT_SYMBOL(tcp_read_sock); 1386 1387/* 1388 * This routine copies from a sock struct into the user buffer. 1389 * 1390 * Technical note: in 2.3 we work on _locked_ socket, so that 1391 * tricks with *seq access order and skb->users are not required. 1392 * Probably, code can be easily improved even more. 1393 */ 1394 1395int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg, 1396 size_t len, int nonblock, int flags, int *addr_len) 1397{ 1398 struct tcp_sock *tp = tcp_sk(sk); 1399 int copied = 0; 1400 u32 peek_seq; 1401 u32 *seq; 1402 unsigned long used; 1403 int err; 1404 int target; /* Read at least this many bytes */ 1405 long timeo; 1406 struct task_struct *user_recv = NULL; 1407 int copied_early = 0; 1408 struct sk_buff *skb; 1409 u32 urg_hole = 0; 1410 1411 lock_sock(sk); 1412 1413 err = -ENOTCONN; 1414 if (sk->sk_state == TCP_LISTEN) 1415 goto out; 1416 1417 timeo = sock_rcvtimeo(sk, nonblock); 1418 1419 /* Urgent data needs to be handled specially. */ 1420 if (flags & MSG_OOB) 1421 goto recv_urg; 1422 1423 seq = &tp->copied_seq; 1424 if (flags & MSG_PEEK) { 1425 peek_seq = tp->copied_seq; 1426 seq = &peek_seq; 1427 } 1428 1429 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 1430 1431#ifdef CONFIG_NET_DMA 1432 tp->ucopy.dma_chan = NULL; 1433 preempt_disable(); 1434 skb = skb_peek_tail(&sk->sk_receive_queue); 1435 { 1436 int available = 0; 1437 1438 if (skb) 1439 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq); 1440 if ((available < target) && 1441 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) && 1442 !sysctl_tcp_low_latency && 1443 dma_find_channel(DMA_MEMCPY)) { 1444 preempt_enable_no_resched(); 1445 tp->ucopy.pinned_list = 1446 dma_pin_iovec_pages(msg->msg_iov, len); 1447 } else { 1448 preempt_enable_no_resched(); 1449 } 1450 } 1451#endif 1452 1453 do { 1454 u32 offset; 1455 1456 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */ 1457 if (tp->urg_data && tp->urg_seq == *seq) { 1458 if (copied) 1459 break; 1460 if (signal_pending(current)) { 1461 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN; 1462 break; 1463 } 1464 } 1465 1466 /* Next get a buffer. */ 1467 1468 skb_queue_walk(&sk->sk_receive_queue, skb) { 1469 /* Now that we have two receive queues this 1470 * shouldn't happen. 1471 */ 1472 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq), 1473 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n", 1474 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, 1475 flags)) 1476 break; 1477 1478 offset = *seq - TCP_SKB_CB(skb)->seq; 1479 if (tcp_hdr(skb)->syn) 1480 offset--; 1481 if (offset < skb->len) 1482 goto found_ok_skb; 1483 if (tcp_hdr(skb)->fin) 1484 goto found_fin_ok; 1485 WARN(!(flags & MSG_PEEK), 1486 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n", 1487 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags); 1488 } 1489 1490 /* Well, if we have backlog, try to process it now yet. */ 1491 1492 if (copied >= target && !sk->sk_backlog.tail) 1493 break; 1494 1495 if (copied) { 1496 if (sk->sk_err || 1497 sk->sk_state == TCP_CLOSE || 1498 (sk->sk_shutdown & RCV_SHUTDOWN) || 1499 !timeo || 1500 signal_pending(current)) 1501 break; 1502 } else { 1503 if (sock_flag(sk, SOCK_DONE)) 1504 break; 1505 1506 if (sk->sk_err) { 1507 copied = sock_error(sk); 1508 break; 1509 } 1510 1511 if (sk->sk_shutdown & RCV_SHUTDOWN) 1512 break; 1513 1514 if (sk->sk_state == TCP_CLOSE) { 1515 if (!sock_flag(sk, SOCK_DONE)) { 1516 /* This occurs when user tries to read 1517 * from never connected socket. 1518 */ 1519 copied = -ENOTCONN; 1520 break; 1521 } 1522 break; 1523 } 1524 1525 if (!timeo) { 1526 copied = -EAGAIN; 1527 break; 1528 } 1529 1530 if (signal_pending(current)) { 1531 copied = sock_intr_errno(timeo); 1532 break; 1533 } 1534 } 1535 1536 tcp_cleanup_rbuf(sk, copied); 1537 1538 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) { 1539 /* Install new reader */ 1540 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) { 1541 user_recv = current; 1542 tp->ucopy.task = user_recv; 1543 tp->ucopy.iov = msg->msg_iov; 1544 } 1545 1546 tp->ucopy.len = len; 1547 1548 WARN_ON(tp->copied_seq != tp->rcv_nxt && 1549 !(flags & (MSG_PEEK | MSG_TRUNC))); 1550 1551 /* Ugly... If prequeue is not empty, we have to 1552 * process it before releasing socket, otherwise 1553 * order will be broken at second iteration. 1554 * More elegant solution is required!!! 1555 * 1556 * Look: we have the following (pseudo)queues: 1557 * 1558 * 1. packets in flight 1559 * 2. backlog 1560 * 3. prequeue 1561 * 4. receive_queue 1562 * 1563 * Each queue can be processed only if the next ones 1564 * are empty. At this point we have empty receive_queue. 1565 * But prequeue _can_ be not empty after 2nd iteration, 1566 * when we jumped to start of loop because backlog 1567 * processing added something to receive_queue. 1568 * We cannot release_sock(), because backlog contains 1569 * packets arrived _after_ prequeued ones. 1570 * 1571 * Shortly, algorithm is clear --- to process all 1572 * the queues in order. We could make it more directly, 1573 * requeueing packets from backlog to prequeue, if 1574 * is not empty. It is more elegant, but eats cycles, 1575 * unfortunately. 1576 */ 1577 if (!skb_queue_empty(&tp->ucopy.prequeue)) 1578 goto do_prequeue; 1579 1580 /* __ Set realtime policy in scheduler __ */ 1581 } 1582 1583#ifdef CONFIG_NET_DMA 1584 if (tp->ucopy.dma_chan) 1585 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1586#endif 1587 if (copied >= target) { 1588 /* Do not sleep, just process backlog. */ 1589 release_sock(sk); 1590 lock_sock(sk); 1591 } else 1592 sk_wait_data(sk, &timeo); 1593 1594#ifdef CONFIG_NET_DMA 1595 tcp_service_net_dma(sk, false); /* Don't block */ 1596 tp->ucopy.wakeup = 0; 1597#endif 1598 1599 if (user_recv) { 1600 int chunk; 1601 1602 /* __ Restore normal policy in scheduler __ */ 1603 1604 if ((chunk = len - tp->ucopy.len) != 0) { 1605 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk); 1606 len -= chunk; 1607 copied += chunk; 1608 } 1609 1610 if (tp->rcv_nxt == tp->copied_seq && 1611 !skb_queue_empty(&tp->ucopy.prequeue)) { 1612do_prequeue: 1613 tcp_prequeue_process(sk); 1614 1615 if ((chunk = len - tp->ucopy.len) != 0) { 1616 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1617 len -= chunk; 1618 copied += chunk; 1619 } 1620 } 1621 } 1622 if ((flags & MSG_PEEK) && 1623 (peek_seq - copied - urg_hole != tp->copied_seq)) { 1624 if (net_ratelimit()) 1625 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n", 1626 current->comm, task_pid_nr(current)); 1627 peek_seq = tp->copied_seq; 1628 } 1629 continue; 1630 1631 found_ok_skb: 1632 /* Ok so how much can we use? */ 1633 used = skb->len - offset; 1634 if (len < used) 1635 used = len; 1636 1637 /* Do we have urgent data here? */ 1638 if (tp->urg_data) { 1639 u32 urg_offset = tp->urg_seq - *seq; 1640 if (urg_offset < used) { 1641 if (!urg_offset) { 1642 if (!sock_flag(sk, SOCK_URGINLINE)) { 1643 ++*seq; 1644 urg_hole++; 1645 offset++; 1646 used--; 1647 if (!used) 1648 goto skip_copy; 1649 } 1650 } else 1651 used = urg_offset; 1652 } 1653 } 1654 1655 if (!(flags & MSG_TRUNC)) { 1656#ifdef CONFIG_NET_DMA 1657 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 1658 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY); 1659 1660 if (tp->ucopy.dma_chan) { 1661 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec( 1662 tp->ucopy.dma_chan, skb, offset, 1663 msg->msg_iov, used, 1664 tp->ucopy.pinned_list); 1665 1666 if (tp->ucopy.dma_cookie < 0) { 1667 1668 printk(KERN_ALERT "dma_cookie < 0\n"); 1669 1670 /* Exception. Bailout! */ 1671 if (!copied) 1672 copied = -EFAULT; 1673 break; 1674 } 1675 1676 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1677 1678 if ((offset + used) == skb->len) 1679 copied_early = 1; 1680 1681 } else 1682#endif 1683 { 1684 err = skb_copy_datagram_iovec(skb, offset, 1685 msg->msg_iov, used); 1686 if (err) { 1687 /* Exception. Bailout! */ 1688 if (!copied) 1689 copied = -EFAULT; 1690 break; 1691 } 1692 } 1693 } 1694 1695 *seq += used; 1696 copied += used; 1697 len -= used; 1698 1699 tcp_rcv_space_adjust(sk); 1700 1701skip_copy: 1702 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) { 1703 tp->urg_data = 0; 1704 tcp_fast_path_check(sk); 1705 } 1706 if (used + offset < skb->len) 1707 continue; 1708 1709 if (tcp_hdr(skb)->fin) 1710 goto found_fin_ok; 1711 if (!(flags & MSG_PEEK)) { 1712 sk_eat_skb(sk, skb, copied_early); 1713 copied_early = 0; 1714 } 1715 continue; 1716 1717 found_fin_ok: 1718 /* Process the FIN. */ 1719 ++*seq; 1720 if (!(flags & MSG_PEEK)) { 1721 sk_eat_skb(sk, skb, copied_early); 1722 copied_early = 0; 1723 } 1724 break; 1725 } while (len > 0); 1726 1727 if (user_recv) { 1728 if (!skb_queue_empty(&tp->ucopy.prequeue)) { 1729 int chunk; 1730 1731 tp->ucopy.len = copied > 0 ? len : 0; 1732 1733 tcp_prequeue_process(sk); 1734 1735 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) { 1736 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1737 len -= chunk; 1738 copied += chunk; 1739 } 1740 } 1741 1742 tp->ucopy.task = NULL; 1743 tp->ucopy.len = 0; 1744 } 1745 1746#ifdef CONFIG_NET_DMA 1747 tcp_service_net_dma(sk, true); /* Wait for queue to drain */ 1748 tp->ucopy.dma_chan = NULL; 1749 1750 if (tp->ucopy.pinned_list) { 1751 dma_unpin_iovec_pages(tp->ucopy.pinned_list); 1752 tp->ucopy.pinned_list = NULL; 1753 } 1754#endif 1755 1756 /* According to UNIX98, msg_name/msg_namelen are ignored 1757 * on connected socket. I was just happy when found this 8) --ANK 1758 */ 1759 1760 /* Clean up data we have read: This will do ACK frames. */ 1761 tcp_cleanup_rbuf(sk, copied); 1762 1763 release_sock(sk); 1764 return copied; 1765 1766out: 1767 release_sock(sk); 1768 return err; 1769 1770recv_urg: 1771 err = tcp_recv_urg(sk, msg, len, flags); 1772 goto out; 1773} 1774EXPORT_SYMBOL(tcp_recvmsg); 1775 1776void tcp_set_state(struct sock *sk, int state) 1777{ 1778 int oldstate = sk->sk_state; 1779 1780 switch (state) { 1781 case TCP_ESTABLISHED: 1782 if (oldstate != TCP_ESTABLISHED) 1783 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1784 break; 1785 1786 case TCP_CLOSE: 1787 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED) 1788 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS); 1789 1790 sk->sk_prot->unhash(sk); 1791 if (inet_csk(sk)->icsk_bind_hash && 1792 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK)) 1793 inet_put_port(sk); 1794 /* fall through */ 1795 default: 1796 if (oldstate == TCP_ESTABLISHED) 1797 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1798 } 1799 1800 /* Change state AFTER socket is unhashed to avoid closed 1801 * socket sitting in hash tables. 1802 */ 1803 sk->sk_state = state; 1804 1805#ifdef STATE_TRACE 1806 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]); 1807#endif 1808} 1809EXPORT_SYMBOL_GPL(tcp_set_state); 1810 1811/* 1812 * State processing on a close. This implements the state shift for 1813 * sending our FIN frame. Note that we only send a FIN for some 1814 * states. A shutdown() may have already sent the FIN, or we may be 1815 * closed. 1816 */ 1817 1818static const unsigned char new_state[16] = { 1819 /* current state: new state: action: */ 1820 /* (Invalid) */ TCP_CLOSE, 1821 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1822 /* TCP_SYN_SENT */ TCP_CLOSE, 1823 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1824 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1, 1825 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2, 1826 /* TCP_TIME_WAIT */ TCP_CLOSE, 1827 /* TCP_CLOSE */ TCP_CLOSE, 1828 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN, 1829 /* TCP_LAST_ACK */ TCP_LAST_ACK, 1830 /* TCP_LISTEN */ TCP_CLOSE, 1831 /* TCP_CLOSING */ TCP_CLOSING, 1832}; 1833 1834static int tcp_close_state(struct sock *sk) 1835{ 1836 int next = (int)new_state[sk->sk_state]; 1837 int ns = next & TCP_STATE_MASK; 1838 1839 tcp_set_state(sk, ns); 1840 1841 return next & TCP_ACTION_FIN; 1842} 1843 1844/* 1845 * Shutdown the sending side of a connection. Much like close except 1846 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD). 1847 */ 1848 1849void tcp_shutdown(struct sock *sk, int how) 1850{ 1851 /* We need to grab some memory, and put together a FIN, 1852 * and then put it into the queue to be sent. 1853 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92. 1854 */ 1855 if (!(how & SEND_SHUTDOWN)) 1856 return; 1857 1858 /* If we've already sent a FIN, or it's a closed state, skip this. */ 1859 if ((1 << sk->sk_state) & 1860 (TCPF_ESTABLISHED | TCPF_SYN_SENT | 1861 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) { 1862 /* Clear out any half completed packets. FIN if needed. */ 1863 if (tcp_close_state(sk)) 1864 tcp_send_fin(sk); 1865 } 1866} 1867EXPORT_SYMBOL(tcp_shutdown); 1868 1869void tcp_close(struct sock *sk, long timeout) 1870{ 1871 struct sk_buff *skb; 1872 int data_was_unread = 0; 1873 int state; 1874 1875 lock_sock(sk); 1876 sk->sk_shutdown = SHUTDOWN_MASK; 1877 1878 if (sk->sk_state == TCP_LISTEN) { 1879 tcp_set_state(sk, TCP_CLOSE); 1880 1881 /* Special case. */ 1882 inet_csk_listen_stop(sk); 1883 1884 goto adjudge_to_death; 1885 } 1886 1887 /* We need to flush the recv. buffs. We do this only on the 1888 * descriptor close, not protocol-sourced closes, because the 1889 * reader process may not have drained the data yet! 1890 */ 1891 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) { 1892 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq - 1893 tcp_hdr(skb)->fin; 1894 data_was_unread += len; 1895 __kfree_skb(skb); 1896 } 1897 1898 sk_mem_reclaim(sk); 1899 1900 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */ 1901 if (sk->sk_state == TCP_CLOSE) 1902 goto adjudge_to_death; 1903 1904 /* As outlined in RFC 2525, section 2.17, we send a RST here because 1905 * data was lost. To witness the awful effects of the old behavior of 1906 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk 1907 * GET in an FTP client, suspend the process, wait for the client to 1908 * advertise a zero window, then kill -9 the FTP client, wheee... 1909 * Note: timeout is always zero in such a case. 1910 */ 1911 if (data_was_unread) { 1912 /* Unread data was tossed, zap the connection. */ 1913 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE); 1914 tcp_set_state(sk, TCP_CLOSE); 1915 tcp_send_active_reset(sk, sk->sk_allocation); 1916 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) { 1917 /* Check zero linger _after_ checking for unread data. */ 1918 sk->sk_prot->disconnect(sk, 0); 1919 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 1920 } else if (tcp_close_state(sk)) { 1921 /* We FIN if the application ate all the data before 1922 * zapping the connection. 1923 */ 1924 1925 /* RED-PEN. Formally speaking, we have broken TCP state 1926 * machine. State transitions: 1927 * 1928 * TCP_ESTABLISHED -> TCP_FIN_WAIT1 1929 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible) 1930 * TCP_CLOSE_WAIT -> TCP_LAST_ACK 1931 * 1932 * are legal only when FIN has been sent (i.e. in window), 1933 * rather than queued out of window. Purists blame. 1934 * 1935 * F.e. "RFC state" is ESTABLISHED, 1936 * if Linux state is FIN-WAIT-1, but FIN is still not sent. 1937 * 1938 * The visible declinations are that sometimes 1939 * we enter time-wait state, when it is not required really 1940 * (harmless), do not send active resets, when they are 1941 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when 1942 * they look as CLOSING or LAST_ACK for Linux) 1943 * Probably, I missed some more holelets. 1944 * --ANK 1945 */ 1946 tcp_send_fin(sk); 1947 } 1948 1949 sk_stream_wait_close(sk, timeout); 1950 1951adjudge_to_death: 1952 state = sk->sk_state; 1953 sock_hold(sk); 1954 sock_orphan(sk); 1955 1956 /* It is the last release_sock in its life. It will remove backlog. */ 1957 release_sock(sk); 1958 1959 1960 /* Now socket is owned by kernel and we acquire BH lock 1961 to finish close. No need to check for user refs. 1962 */ 1963 local_bh_disable(); 1964 bh_lock_sock(sk); 1965 WARN_ON(sock_owned_by_user(sk)); 1966 1967 percpu_counter_inc(sk->sk_prot->orphan_count); 1968 1969 /* Have we already been destroyed by a softirq or backlog? */ 1970 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE) 1971 goto out; 1972 1973 /* This is a (useful) BSD violating of the RFC. There is a 1974 * problem with TCP as specified in that the other end could 1975 * keep a socket open forever with no application left this end. 1976 * We use a 3 minute timeout (about the same as BSD) then kill 1977 * our end. If they send after that then tough - BUT: long enough 1978 * that we won't make the old 4*rto = almost no time - whoops 1979 * reset mistake. 1980 * 1981 * Nope, it was not mistake. It is really desired behaviour 1982 * f.e. on http servers, when such sockets are useless, but 1983 * consume significant resources. Let's do it with special 1984 * linger2 option. --ANK 1985 */ 1986 1987 if (sk->sk_state == TCP_FIN_WAIT2) { 1988 struct tcp_sock *tp = tcp_sk(sk); 1989 if (tp->linger2 < 0) { 1990 tcp_set_state(sk, TCP_CLOSE); 1991 tcp_send_active_reset(sk, GFP_ATOMIC); 1992 NET_INC_STATS_BH(sock_net(sk), 1993 LINUX_MIB_TCPABORTONLINGER); 1994 } else { 1995 const int tmo = tcp_fin_time(sk); 1996 1997 if (tmo > TCP_TIMEWAIT_LEN) { 1998 inet_csk_reset_keepalive_timer(sk, 1999 tmo - TCP_TIMEWAIT_LEN); 2000 } else { 2001 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 2002 goto out; 2003 } 2004 } 2005 } 2006 if (sk->sk_state != TCP_CLOSE) { 2007 sk_mem_reclaim(sk); 2008 if (tcp_too_many_orphans(sk, 0)) { 2009 if (net_ratelimit()) 2010 printk(KERN_INFO "TCP: too many of orphaned " 2011 "sockets\n"); 2012 tcp_set_state(sk, TCP_CLOSE); 2013 tcp_send_active_reset(sk, GFP_ATOMIC); 2014 NET_INC_STATS_BH(sock_net(sk), 2015 LINUX_MIB_TCPABORTONMEMORY); 2016 } 2017 } 2018 2019 if (sk->sk_state == TCP_CLOSE) 2020 inet_csk_destroy_sock(sk); 2021 /* Otherwise, socket is reprieved until protocol close. */ 2022 2023out: 2024 bh_unlock_sock(sk); 2025 local_bh_enable(); 2026 sock_put(sk); 2027} 2028EXPORT_SYMBOL(tcp_close); 2029 2030/* These states need RST on ABORT according to RFC793 */ 2031 2032static inline int tcp_need_reset(int state) 2033{ 2034 return (1 << state) & 2035 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 | 2036 TCPF_FIN_WAIT2 | TCPF_SYN_RECV); 2037} 2038 2039int tcp_disconnect(struct sock *sk, int flags) 2040{ 2041 struct inet_sock *inet = inet_sk(sk); 2042 struct inet_connection_sock *icsk = inet_csk(sk); 2043 struct tcp_sock *tp = tcp_sk(sk); 2044 int err = 0; 2045 int old_state = sk->sk_state; 2046 2047 if (old_state != TCP_CLOSE) 2048 tcp_set_state(sk, TCP_CLOSE); 2049 2050 /* ABORT function of RFC793 */ 2051 if (old_state == TCP_LISTEN) { 2052 inet_csk_listen_stop(sk); 2053 } else if (tcp_need_reset(old_state) || 2054 (tp->snd_nxt != tp->write_seq && 2055 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) { 2056 /* The last check adjusts for discrepancy of Linux wrt. RFC 2057 * states 2058 */ 2059 tcp_send_active_reset(sk, gfp_any()); 2060 sk->sk_err = ECONNRESET; 2061 } else if (old_state == TCP_SYN_SENT) 2062 sk->sk_err = ECONNRESET; 2063 2064 tcp_clear_xmit_timers(sk); 2065 __skb_queue_purge(&sk->sk_receive_queue); 2066 tcp_write_queue_purge(sk); 2067 __skb_queue_purge(&tp->out_of_order_queue); 2068#ifdef CONFIG_NET_DMA 2069 __skb_queue_purge(&sk->sk_async_wait_queue); 2070#endif 2071 2072 inet->inet_dport = 0; 2073 2074 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK)) 2075 inet_reset_saddr(sk); 2076 2077 sk->sk_shutdown = 0; 2078 sock_reset_flag(sk, SOCK_DONE); 2079 tp->srtt = 0; 2080 if ((tp->write_seq += tp->max_window + 2) == 0) 2081 tp->write_seq = 1; 2082 icsk->icsk_backoff = 0; 2083 tp->snd_cwnd = 2; 2084 icsk->icsk_probes_out = 0; 2085 tp->packets_out = 0; 2086 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 2087 tp->snd_cwnd_cnt = 0; 2088 tp->bytes_acked = 0; 2089 tp->window_clamp = 0; 2090 tcp_set_ca_state(sk, TCP_CA_Open); 2091 tcp_clear_retrans(tp); 2092 inet_csk_delack_init(sk); 2093 tcp_init_send_head(sk); 2094 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt)); 2095 __sk_dst_reset(sk); 2096 2097 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash); 2098 2099 sk->sk_error_report(sk); 2100 return err; 2101} 2102EXPORT_SYMBOL(tcp_disconnect); 2103 2104/* 2105 * Socket option code for TCP. 2106 */ 2107static int do_tcp_setsockopt(struct sock *sk, int level, 2108 int optname, char __user *optval, unsigned int optlen) 2109{ 2110 struct tcp_sock *tp = tcp_sk(sk); 2111 struct inet_connection_sock *icsk = inet_csk(sk); 2112 int val; 2113 int err = 0; 2114 2115 /* These are data/string values, all the others are ints */ 2116 switch (optname) { 2117 case TCP_CONGESTION: { 2118 char name[TCP_CA_NAME_MAX]; 2119 2120 if (optlen < 1) 2121 return -EINVAL; 2122 2123 val = strncpy_from_user(name, optval, 2124 min_t(long, TCP_CA_NAME_MAX-1, optlen)); 2125 if (val < 0) 2126 return -EFAULT; 2127 name[val] = 0; 2128 2129 lock_sock(sk); 2130 err = tcp_set_congestion_control(sk, name); 2131 release_sock(sk); 2132 return err; 2133 } 2134 case TCP_COOKIE_TRANSACTIONS: { 2135 struct tcp_cookie_transactions ctd; 2136 struct tcp_cookie_values *cvp = NULL; 2137 2138 if (sizeof(ctd) > optlen) 2139 return -EINVAL; 2140 if (copy_from_user(&ctd, optval, sizeof(ctd))) 2141 return -EFAULT; 2142 2143 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) || 2144 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED) 2145 return -EINVAL; 2146 2147 if (ctd.tcpct_cookie_desired == 0) { 2148 /* default to global value */ 2149 } else if ((0x1 & ctd.tcpct_cookie_desired) || 2150 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX || 2151 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) { 2152 return -EINVAL; 2153 } 2154 2155 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) { 2156 /* Supercedes all other values */ 2157 lock_sock(sk); 2158 if (tp->cookie_values != NULL) { 2159 kref_put(&tp->cookie_values->kref, 2160 tcp_cookie_values_release); 2161 tp->cookie_values = NULL; 2162 } 2163 tp->rx_opt.cookie_in_always = 0; /* false */ 2164 tp->rx_opt.cookie_out_never = 1; /* true */ 2165 release_sock(sk); 2166 return err; 2167 } 2168 2169 /* Allocate ancillary memory before locking. 2170 */ 2171 if (ctd.tcpct_used > 0 || 2172 (tp->cookie_values == NULL && 2173 (sysctl_tcp_cookie_size > 0 || 2174 ctd.tcpct_cookie_desired > 0 || 2175 ctd.tcpct_s_data_desired > 0))) { 2176 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used, 2177 GFP_KERNEL); 2178 if (cvp == NULL) 2179 return -ENOMEM; 2180 2181 kref_init(&cvp->kref); 2182 } 2183 lock_sock(sk); 2184 tp->rx_opt.cookie_in_always = 2185 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags); 2186 tp->rx_opt.cookie_out_never = 0; /* false */ 2187 2188 if (tp->cookie_values != NULL) { 2189 if (cvp != NULL) { 2190 /* Changed values are recorded by a changed 2191 * pointer, ensuring the cookie will differ, 2192 * without separately hashing each value later. 2193 */ 2194 kref_put(&tp->cookie_values->kref, 2195 tcp_cookie_values_release); 2196 } else { 2197 cvp = tp->cookie_values; 2198 } 2199 } 2200 2201 if (cvp != NULL) { 2202 cvp->cookie_desired = ctd.tcpct_cookie_desired; 2203 2204 if (ctd.tcpct_used > 0) { 2205 memcpy(cvp->s_data_payload, ctd.tcpct_value, 2206 ctd.tcpct_used); 2207 cvp->s_data_desired = ctd.tcpct_used; 2208 cvp->s_data_constant = 1; /* true */ 2209 } else { 2210 /* No constant payload data. */ 2211 cvp->s_data_desired = ctd.tcpct_s_data_desired; 2212 cvp->s_data_constant = 0; /* false */ 2213 } 2214 2215 tp->cookie_values = cvp; 2216 } 2217 release_sock(sk); 2218 return err; 2219 } 2220 default: 2221 /* fallthru */ 2222 break; 2223 } 2224 2225 if (optlen < sizeof(int)) 2226 return -EINVAL; 2227 2228 if (get_user(val, (int __user *)optval)) 2229 return -EFAULT; 2230 2231 lock_sock(sk); 2232 2233 switch (optname) { 2234 case TCP_MAXSEG: 2235 /* Values greater than interface MTU won't take effect. However 2236 * at the point when this call is done we typically don't yet 2237 * know which interface is going to be used */ 2238 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) { 2239 err = -EINVAL; 2240 break; 2241 } 2242 tp->rx_opt.user_mss = val; 2243 break; 2244 2245 case TCP_NODELAY: 2246 if (val) { 2247 /* TCP_NODELAY is weaker than TCP_CORK, so that 2248 * this option on corked socket is remembered, but 2249 * it is not activated until cork is cleared. 2250 * 2251 * However, when TCP_NODELAY is set we make 2252 * an explicit push, which overrides even TCP_CORK 2253 * for currently queued segments. 2254 */ 2255 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH; 2256 tcp_push_pending_frames(sk); 2257 } else { 2258 tp->nonagle &= ~TCP_NAGLE_OFF; 2259 } 2260 break; 2261 2262 case TCP_THIN_LINEAR_TIMEOUTS: 2263 if (val < 0 || val > 1) 2264 err = -EINVAL; 2265 else 2266 tp->thin_lto = val; 2267 break; 2268 2269 case TCP_THIN_DUPACK: 2270 if (val < 0 || val > 1) 2271 err = -EINVAL; 2272 else 2273 tp->thin_dupack = val; 2274 break; 2275 2276 case TCP_CORK: 2277 /* When set indicates to always queue non-full frames. 2278 * Later the user clears this option and we transmit 2279 * any pending partial frames in the queue. This is 2280 * meant to be used alongside sendfile() to get properly 2281 * filled frames when the user (for example) must write 2282 * out headers with a write() call first and then use 2283 * sendfile to send out the data parts. 2284 * 2285 * TCP_CORK can be set together with TCP_NODELAY and it is 2286 * stronger than TCP_NODELAY. 2287 */ 2288 if (val) { 2289 tp->nonagle |= TCP_NAGLE_CORK; 2290 } else { 2291 tp->nonagle &= ~TCP_NAGLE_CORK; 2292 if (tp->nonagle&TCP_NAGLE_OFF) 2293 tp->nonagle |= TCP_NAGLE_PUSH; 2294 tcp_push_pending_frames(sk); 2295 } 2296 break; 2297 2298 case TCP_KEEPIDLE: 2299 if (val < 1 || val > MAX_TCP_KEEPIDLE) 2300 err = -EINVAL; 2301 else { 2302 tp->keepalive_time = val * HZ; 2303 if (sock_flag(sk, SOCK_KEEPOPEN) && 2304 !((1 << sk->sk_state) & 2305 (TCPF_CLOSE | TCPF_LISTEN))) { 2306 u32 elapsed = keepalive_time_elapsed(tp); 2307 if (tp->keepalive_time > elapsed) 2308 elapsed = tp->keepalive_time - elapsed; 2309 else 2310 elapsed = 0; 2311 inet_csk_reset_keepalive_timer(sk, elapsed); 2312 } 2313 } 2314 break; 2315 case TCP_KEEPINTVL: 2316 if (val < 1 || val > MAX_TCP_KEEPINTVL) 2317 err = -EINVAL; 2318 else 2319 tp->keepalive_intvl = val * HZ; 2320 break; 2321 case TCP_KEEPCNT: 2322 if (val < 1 || val > MAX_TCP_KEEPCNT) 2323 err = -EINVAL; 2324 else 2325 tp->keepalive_probes = val; 2326 break; 2327 case TCP_SYNCNT: 2328 if (val < 1 || val > MAX_TCP_SYNCNT) 2329 err = -EINVAL; 2330 else 2331 icsk->icsk_syn_retries = val; 2332 break; 2333 2334 case TCP_LINGER2: 2335 if (val < 0) 2336 tp->linger2 = -1; 2337 else if (val > sysctl_tcp_fin_timeout / HZ) 2338 tp->linger2 = 0; 2339 else 2340 tp->linger2 = val * HZ; 2341 break; 2342 2343 case TCP_DEFER_ACCEPT: 2344 /* Translate value in seconds to number of retransmits */ 2345 icsk->icsk_accept_queue.rskq_defer_accept = 2346 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ, 2347 TCP_RTO_MAX / HZ); 2348 break; 2349 2350 case TCP_WINDOW_CLAMP: 2351 if (!val) { 2352 if (sk->sk_state != TCP_CLOSE) { 2353 err = -EINVAL; 2354 break; 2355 } 2356 tp->window_clamp = 0; 2357 } else 2358 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ? 2359 SOCK_MIN_RCVBUF / 2 : val; 2360 break; 2361 2362 case TCP_QUICKACK: 2363 if (!val) { 2364 icsk->icsk_ack.pingpong = 1; 2365 } else { 2366 icsk->icsk_ack.pingpong = 0; 2367 if ((1 << sk->sk_state) & 2368 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) && 2369 inet_csk_ack_scheduled(sk)) { 2370 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 2371 tcp_cleanup_rbuf(sk, 1); 2372 if (!(val & 1)) 2373 icsk->icsk_ack.pingpong = 1; 2374 } 2375 } 2376 break; 2377 2378#ifdef CONFIG_TCP_MD5SIG 2379 case TCP_MD5SIG: 2380 /* Read the IP->Key mappings from userspace */ 2381 err = tp->af_specific->md5_parse(sk, optval, optlen); 2382 break; 2383#endif 2384 case TCP_USER_TIMEOUT: 2385 /* Cap the max timeout in ms TCP will retry/retrans 2386 * before giving up and aborting (ETIMEDOUT) a connection. 2387 */ 2388 icsk->icsk_user_timeout = msecs_to_jiffies(val); 2389 break; 2390 default: 2391 err = -ENOPROTOOPT; 2392 break; 2393 } 2394 2395 release_sock(sk); 2396 return err; 2397} 2398 2399int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval, 2400 unsigned int optlen) 2401{ 2402 struct inet_connection_sock *icsk = inet_csk(sk); 2403 2404 if (level != SOL_TCP) 2405 return icsk->icsk_af_ops->setsockopt(sk, level, optname, 2406 optval, optlen); 2407 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2408} 2409EXPORT_SYMBOL(tcp_setsockopt); 2410 2411#ifdef CONFIG_COMPAT 2412int compat_tcp_setsockopt(struct sock *sk, int level, int optname, 2413 char __user *optval, unsigned int optlen) 2414{ 2415 if (level != SOL_TCP) 2416 return inet_csk_compat_setsockopt(sk, level, optname, 2417 optval, optlen); 2418 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2419} 2420EXPORT_SYMBOL(compat_tcp_setsockopt); 2421#endif 2422 2423/* Return information about state of tcp endpoint in API format. */ 2424void tcp_get_info(struct sock *sk, struct tcp_info *info) 2425{ 2426 struct tcp_sock *tp = tcp_sk(sk); 2427 const struct inet_connection_sock *icsk = inet_csk(sk); 2428 u32 now = tcp_time_stamp; 2429 2430 memset(info, 0, sizeof(*info)); 2431 2432 info->tcpi_state = sk->sk_state; 2433 info->tcpi_ca_state = icsk->icsk_ca_state; 2434 info->tcpi_retransmits = icsk->icsk_retransmits; 2435 info->tcpi_probes = icsk->icsk_probes_out; 2436 info->tcpi_backoff = icsk->icsk_backoff; 2437 2438 if (tp->rx_opt.tstamp_ok) 2439 info->tcpi_options |= TCPI_OPT_TIMESTAMPS; 2440 if (tcp_is_sack(tp)) 2441 info->tcpi_options |= TCPI_OPT_SACK; 2442 if (tp->rx_opt.wscale_ok) { 2443 info->tcpi_options |= TCPI_OPT_WSCALE; 2444 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale; 2445 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale; 2446 } 2447 2448 if (tp->ecn_flags&TCP_ECN_OK) 2449 info->tcpi_options |= TCPI_OPT_ECN; 2450 2451 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto); 2452 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato); 2453 info->tcpi_snd_mss = tp->mss_cache; 2454 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss; 2455 2456 if (sk->sk_state == TCP_LISTEN) { 2457 info->tcpi_unacked = sk->sk_ack_backlog; 2458 info->tcpi_sacked = sk->sk_max_ack_backlog; 2459 } else { 2460 info->tcpi_unacked = tp->packets_out; 2461 info->tcpi_sacked = tp->sacked_out; 2462 } 2463 info->tcpi_lost = tp->lost_out; 2464 info->tcpi_retrans = tp->retrans_out; 2465 info->tcpi_fackets = tp->fackets_out; 2466 2467 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime); 2468 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime); 2469 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp); 2470 2471 info->tcpi_pmtu = icsk->icsk_pmtu_cookie; 2472 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh; 2473 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3; 2474 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2; 2475 info->tcpi_snd_ssthresh = tp->snd_ssthresh; 2476 info->tcpi_snd_cwnd = tp->snd_cwnd; 2477 info->tcpi_advmss = tp->advmss; 2478 info->tcpi_reordering = tp->reordering; 2479 2480 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3; 2481 info->tcpi_rcv_space = tp->rcvq_space.space; 2482 2483 info->tcpi_total_retrans = tp->total_retrans; 2484} 2485EXPORT_SYMBOL_GPL(tcp_get_info); 2486 2487static int do_tcp_getsockopt(struct sock *sk, int level, 2488 int optname, char __user *optval, int __user *optlen) 2489{ 2490 struct inet_connection_sock *icsk = inet_csk(sk); 2491 struct tcp_sock *tp = tcp_sk(sk); 2492 int val, len; 2493 2494 if (get_user(len, optlen)) 2495 return -EFAULT; 2496 2497 len = min_t(unsigned int, len, sizeof(int)); 2498 2499 if (len < 0) 2500 return -EINVAL; 2501 2502 switch (optname) { 2503 case TCP_MAXSEG: 2504 val = tp->mss_cache; 2505 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 2506 val = tp->rx_opt.user_mss; 2507 break; 2508 case TCP_NODELAY: 2509 val = !!(tp->nonagle&TCP_NAGLE_OFF); 2510 break; 2511 case TCP_CORK: 2512 val = !!(tp->nonagle&TCP_NAGLE_CORK); 2513 break; 2514 case TCP_KEEPIDLE: 2515 val = keepalive_time_when(tp) / HZ; 2516 break; 2517 case TCP_KEEPINTVL: 2518 val = keepalive_intvl_when(tp) / HZ; 2519 break; 2520 case TCP_KEEPCNT: 2521 val = keepalive_probes(tp); 2522 break; 2523 case TCP_SYNCNT: 2524 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries; 2525 break; 2526 case TCP_LINGER2: 2527 val = tp->linger2; 2528 if (val >= 0) 2529 val = (val ? : sysctl_tcp_fin_timeout) / HZ; 2530 break; 2531 case TCP_DEFER_ACCEPT: 2532 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept, 2533 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ); 2534 break; 2535 case TCP_WINDOW_CLAMP: 2536 val = tp->window_clamp; 2537 break; 2538 case TCP_INFO: { 2539 struct tcp_info info; 2540 2541 if (get_user(len, optlen)) 2542 return -EFAULT; 2543 2544 tcp_get_info(sk, &info); 2545 2546 len = min_t(unsigned int, len, sizeof(info)); 2547 if (put_user(len, optlen)) 2548 return -EFAULT; 2549 if (copy_to_user(optval, &info, len)) 2550 return -EFAULT; 2551 return 0; 2552 } 2553 case TCP_QUICKACK: 2554 val = !icsk->icsk_ack.pingpong; 2555 break; 2556 2557 case TCP_CONGESTION: 2558 if (get_user(len, optlen)) 2559 return -EFAULT; 2560 len = min_t(unsigned int, len, TCP_CA_NAME_MAX); 2561 if (put_user(len, optlen)) 2562 return -EFAULT; 2563 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len)) 2564 return -EFAULT; 2565 return 0; 2566 2567 case TCP_COOKIE_TRANSACTIONS: { 2568 struct tcp_cookie_transactions ctd; 2569 struct tcp_cookie_values *cvp = tp->cookie_values; 2570 2571 if (get_user(len, optlen)) 2572 return -EFAULT; 2573 if (len < sizeof(ctd)) 2574 return -EINVAL; 2575 2576 memset(&ctd, 0, sizeof(ctd)); 2577 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ? 2578 TCP_COOKIE_IN_ALWAYS : 0) 2579 | (tp->rx_opt.cookie_out_never ? 2580 TCP_COOKIE_OUT_NEVER : 0); 2581 2582 if (cvp != NULL) { 2583 ctd.tcpct_flags |= (cvp->s_data_in ? 2584 TCP_S_DATA_IN : 0) 2585 | (cvp->s_data_out ? 2586 TCP_S_DATA_OUT : 0); 2587 2588 ctd.tcpct_cookie_desired = cvp->cookie_desired; 2589 ctd.tcpct_s_data_desired = cvp->s_data_desired; 2590 2591 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0], 2592 cvp->cookie_pair_size); 2593 ctd.tcpct_used = cvp->cookie_pair_size; 2594 } 2595 2596 if (put_user(sizeof(ctd), optlen)) 2597 return -EFAULT; 2598 if (copy_to_user(optval, &ctd, sizeof(ctd))) 2599 return -EFAULT; 2600 return 0; 2601 } 2602 case TCP_THIN_LINEAR_TIMEOUTS: 2603 val = tp->thin_lto; 2604 break; 2605 case TCP_THIN_DUPACK: 2606 val = tp->thin_dupack; 2607 break; 2608 2609 case TCP_USER_TIMEOUT: 2610 val = jiffies_to_msecs(icsk->icsk_user_timeout); 2611 break; 2612 default: 2613 return -ENOPROTOOPT; 2614 } 2615 2616 if (put_user(len, optlen)) 2617 return -EFAULT; 2618 if (copy_to_user(optval, &val, len)) 2619 return -EFAULT; 2620 return 0; 2621} 2622 2623int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval, 2624 int __user *optlen) 2625{ 2626 struct inet_connection_sock *icsk = inet_csk(sk); 2627 2628 if (level != SOL_TCP) 2629 return icsk->icsk_af_ops->getsockopt(sk, level, optname, 2630 optval, optlen); 2631 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2632} 2633EXPORT_SYMBOL(tcp_getsockopt); 2634 2635#ifdef CONFIG_COMPAT 2636int compat_tcp_getsockopt(struct sock *sk, int level, int optname, 2637 char __user *optval, int __user *optlen) 2638{ 2639 if (level != SOL_TCP) 2640 return inet_csk_compat_getsockopt(sk, level, optname, 2641 optval, optlen); 2642 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2643} 2644EXPORT_SYMBOL(compat_tcp_getsockopt); 2645#endif 2646 2647struct sk_buff *tcp_tso_segment(struct sk_buff *skb, u32 features) 2648{ 2649 struct sk_buff *segs = ERR_PTR(-EINVAL); 2650 struct tcphdr *th; 2651 unsigned thlen; 2652 unsigned int seq; 2653 __be32 delta; 2654 unsigned int oldlen; 2655 unsigned int mss; 2656 2657 if (!pskb_may_pull(skb, sizeof(*th))) 2658 goto out; 2659 2660 th = tcp_hdr(skb); 2661 thlen = th->doff * 4; 2662 if (thlen < sizeof(*th)) 2663 goto out; 2664 2665 if (!pskb_may_pull(skb, thlen)) 2666 goto out; 2667 2668 oldlen = (u16)~skb->len; 2669 __skb_pull(skb, thlen); 2670 2671 mss = skb_shinfo(skb)->gso_size; 2672 if (unlikely(skb->len <= mss)) 2673 goto out; 2674 2675 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) { 2676 /* Packet is from an untrusted source, reset gso_segs. */ 2677 int type = skb_shinfo(skb)->gso_type; 2678 2679 if (unlikely(type & 2680 ~(SKB_GSO_TCPV4 | 2681 SKB_GSO_DODGY | 2682 SKB_GSO_TCP_ECN | 2683 SKB_GSO_TCPV6 | 2684 0) || 2685 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))) 2686 goto out; 2687 2688 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss); 2689 2690 segs = NULL; 2691 goto out; 2692 } 2693 2694 segs = skb_segment(skb, features); 2695 if (IS_ERR(segs)) 2696 goto out; 2697 2698 delta = htonl(oldlen + (thlen + mss)); 2699 2700 skb = segs; 2701 th = tcp_hdr(skb); 2702 seq = ntohl(th->seq); 2703 2704 do { 2705 th->fin = th->psh = 0; 2706 2707 th->check = ~csum_fold((__force __wsum)((__force u32)th->check + 2708 (__force u32)delta)); 2709 if (skb->ip_summed != CHECKSUM_PARTIAL) 2710 th->check = 2711 csum_fold(csum_partial(skb_transport_header(skb), 2712 thlen, skb->csum)); 2713 2714 seq += mss; 2715 skb = skb->next; 2716 th = tcp_hdr(skb); 2717 2718 th->seq = htonl(seq); 2719 th->cwr = 0; 2720 } while (skb->next); 2721 2722 delta = htonl(oldlen + (skb->tail - skb->transport_header) + 2723 skb->data_len); 2724 th->check = ~csum_fold((__force __wsum)((__force u32)th->check + 2725 (__force u32)delta)); 2726 if (skb->ip_summed != CHECKSUM_PARTIAL) 2727 th->check = csum_fold(csum_partial(skb_transport_header(skb), 2728 thlen, skb->csum)); 2729 2730out: 2731 return segs; 2732} 2733EXPORT_SYMBOL(tcp_tso_segment); 2734 2735struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb) 2736{ 2737 struct sk_buff **pp = NULL; 2738 struct sk_buff *p; 2739 struct tcphdr *th; 2740 struct tcphdr *th2; 2741 unsigned int len; 2742 unsigned int thlen; 2743 __be32 flags; 2744 unsigned int mss = 1; 2745 unsigned int hlen; 2746 unsigned int off; 2747 int flush = 1; 2748 int i; 2749 2750 off = skb_gro_offset(skb); 2751 hlen = off + sizeof(*th); 2752 th = skb_gro_header_fast(skb, off); 2753 if (skb_gro_header_hard(skb, hlen)) { 2754 th = skb_gro_header_slow(skb, hlen, off); 2755 if (unlikely(!th)) 2756 goto out; 2757 } 2758 2759 thlen = th->doff * 4; 2760 if (thlen < sizeof(*th)) 2761 goto out; 2762 2763 hlen = off + thlen; 2764 if (skb_gro_header_hard(skb, hlen)) { 2765 th = skb_gro_header_slow(skb, hlen, off); 2766 if (unlikely(!th)) 2767 goto out; 2768 } 2769 2770 skb_gro_pull(skb, thlen); 2771 2772 len = skb_gro_len(skb); 2773 flags = tcp_flag_word(th); 2774 2775 for (; (p = *head); head = &p->next) { 2776 if (!NAPI_GRO_CB(p)->same_flow) 2777 continue; 2778 2779 th2 = tcp_hdr(p); 2780 2781 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) { 2782 NAPI_GRO_CB(p)->same_flow = 0; 2783 continue; 2784 } 2785 2786 goto found; 2787 } 2788 2789 goto out_check_final; 2790 2791found: 2792 flush = NAPI_GRO_CB(p)->flush; 2793 flush |= (__force int)(flags & TCP_FLAG_CWR); 2794 flush |= (__force int)((flags ^ tcp_flag_word(th2)) & 2795 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH)); 2796 flush |= (__force int)(th->ack_seq ^ th2->ack_seq); 2797 for (i = sizeof(*th); i < thlen; i += 4) 2798 flush |= *(u32 *)((u8 *)th + i) ^ 2799 *(u32 *)((u8 *)th2 + i); 2800 2801 mss = skb_shinfo(p)->gso_size; 2802 2803 flush |= (len - 1) >= mss; 2804 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq); 2805 2806 if (flush || skb_gro_receive(head, skb)) { 2807 mss = 1; 2808 goto out_check_final; 2809 } 2810 2811 p = *head; 2812 th2 = tcp_hdr(p); 2813 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH); 2814 2815out_check_final: 2816 flush = len < mss; 2817 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH | 2818 TCP_FLAG_RST | TCP_FLAG_SYN | 2819 TCP_FLAG_FIN)); 2820 2821 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush)) 2822 pp = head; 2823 2824out: 2825 NAPI_GRO_CB(skb)->flush |= flush; 2826 2827 return pp; 2828} 2829EXPORT_SYMBOL(tcp_gro_receive); 2830 2831int tcp_gro_complete(struct sk_buff *skb) 2832{ 2833 struct tcphdr *th = tcp_hdr(skb); 2834 2835 skb->csum_start = skb_transport_header(skb) - skb->head; 2836 skb->csum_offset = offsetof(struct tcphdr, check); 2837 skb->ip_summed = CHECKSUM_PARTIAL; 2838 2839 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count; 2840 2841 if (th->cwr) 2842 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 2843 2844 return 0; 2845} 2846EXPORT_SYMBOL(tcp_gro_complete); 2847 2848#ifdef CONFIG_TCP_MD5SIG 2849static unsigned long tcp_md5sig_users; 2850static struct tcp_md5sig_pool * __percpu *tcp_md5sig_pool; 2851static DEFINE_SPINLOCK(tcp_md5sig_pool_lock); 2852 2853static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool * __percpu *pool) 2854{ 2855 int cpu; 2856 for_each_possible_cpu(cpu) { 2857 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu); 2858 if (p) { 2859 if (p->md5_desc.tfm) 2860 crypto_free_hash(p->md5_desc.tfm); 2861 kfree(p); 2862 } 2863 } 2864 free_percpu(pool); 2865} 2866 2867void tcp_free_md5sig_pool(void) 2868{ 2869 struct tcp_md5sig_pool * __percpu *pool = NULL; 2870 2871 spin_lock_bh(&tcp_md5sig_pool_lock); 2872 if (--tcp_md5sig_users == 0) { 2873 pool = tcp_md5sig_pool; 2874 tcp_md5sig_pool = NULL; 2875 } 2876 spin_unlock_bh(&tcp_md5sig_pool_lock); 2877 if (pool) 2878 __tcp_free_md5sig_pool(pool); 2879} 2880EXPORT_SYMBOL(tcp_free_md5sig_pool); 2881 2882static struct tcp_md5sig_pool * __percpu * 2883__tcp_alloc_md5sig_pool(struct sock *sk) 2884{ 2885 int cpu; 2886 struct tcp_md5sig_pool * __percpu *pool; 2887 2888 pool = alloc_percpu(struct tcp_md5sig_pool *); 2889 if (!pool) 2890 return NULL; 2891 2892 for_each_possible_cpu(cpu) { 2893 struct tcp_md5sig_pool *p; 2894 struct crypto_hash *hash; 2895 2896 p = kzalloc(sizeof(*p), sk->sk_allocation); 2897 if (!p) 2898 goto out_free; 2899 *per_cpu_ptr(pool, cpu) = p; 2900 2901 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC); 2902 if (!hash || IS_ERR(hash)) 2903 goto out_free; 2904 2905 p->md5_desc.tfm = hash; 2906 } 2907 return pool; 2908out_free: 2909 __tcp_free_md5sig_pool(pool); 2910 return NULL; 2911} 2912 2913struct tcp_md5sig_pool * __percpu *tcp_alloc_md5sig_pool(struct sock *sk) 2914{ 2915 struct tcp_md5sig_pool * __percpu *pool; 2916 int alloc = 0; 2917 2918retry: 2919 spin_lock_bh(&tcp_md5sig_pool_lock); 2920 pool = tcp_md5sig_pool; 2921 if (tcp_md5sig_users++ == 0) { 2922 alloc = 1; 2923 spin_unlock_bh(&tcp_md5sig_pool_lock); 2924 } else if (!pool) { 2925 tcp_md5sig_users--; 2926 spin_unlock_bh(&tcp_md5sig_pool_lock); 2927 cpu_relax(); 2928 goto retry; 2929 } else 2930 spin_unlock_bh(&tcp_md5sig_pool_lock); 2931 2932 if (alloc) { 2933 /* we cannot hold spinlock here because this may sleep. */ 2934 struct tcp_md5sig_pool * __percpu *p; 2935 2936 p = __tcp_alloc_md5sig_pool(sk); 2937 spin_lock_bh(&tcp_md5sig_pool_lock); 2938 if (!p) { 2939 tcp_md5sig_users--; 2940 spin_unlock_bh(&tcp_md5sig_pool_lock); 2941 return NULL; 2942 } 2943 pool = tcp_md5sig_pool; 2944 if (pool) { 2945 /* oops, it has already been assigned. */ 2946 spin_unlock_bh(&tcp_md5sig_pool_lock); 2947 __tcp_free_md5sig_pool(p); 2948 } else { 2949 tcp_md5sig_pool = pool = p; 2950 spin_unlock_bh(&tcp_md5sig_pool_lock); 2951 } 2952 } 2953 return pool; 2954} 2955EXPORT_SYMBOL(tcp_alloc_md5sig_pool); 2956 2957 2958/** 2959 * tcp_get_md5sig_pool - get md5sig_pool for this user 2960 * 2961 * We use percpu structure, so if we succeed, we exit with preemption 2962 * and BH disabled, to make sure another thread or softirq handling 2963 * wont try to get same context. 2964 */ 2965struct tcp_md5sig_pool *tcp_get_md5sig_pool(void) 2966{ 2967 struct tcp_md5sig_pool * __percpu *p; 2968 2969 local_bh_disable(); 2970 2971 spin_lock(&tcp_md5sig_pool_lock); 2972 p = tcp_md5sig_pool; 2973 if (p) 2974 tcp_md5sig_users++; 2975 spin_unlock(&tcp_md5sig_pool_lock); 2976 2977 if (p) 2978 return *this_cpu_ptr(p); 2979 2980 local_bh_enable(); 2981 return NULL; 2982} 2983EXPORT_SYMBOL(tcp_get_md5sig_pool); 2984 2985void tcp_put_md5sig_pool(void) 2986{ 2987 local_bh_enable(); 2988 tcp_free_md5sig_pool(); 2989} 2990EXPORT_SYMBOL(tcp_put_md5sig_pool); 2991 2992int tcp_md5_hash_header(struct tcp_md5sig_pool *hp, 2993 struct tcphdr *th) 2994{ 2995 struct scatterlist sg; 2996 int err; 2997 2998 __sum16 old_checksum = th->check; 2999 th->check = 0; 3000 /* options aren't included in the hash */ 3001 sg_init_one(&sg, th, sizeof(struct tcphdr)); 3002 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr)); 3003 th->check = old_checksum; 3004 return err; 3005} 3006EXPORT_SYMBOL(tcp_md5_hash_header); 3007 3008int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp, 3009 struct sk_buff *skb, unsigned header_len) 3010{ 3011 struct scatterlist sg; 3012 const struct tcphdr *tp = tcp_hdr(skb); 3013 struct hash_desc *desc = &hp->md5_desc; 3014 unsigned i; 3015 const unsigned head_data_len = skb_headlen(skb) > header_len ? 3016 skb_headlen(skb) - header_len : 0; 3017 const struct skb_shared_info *shi = skb_shinfo(skb); 3018 struct sk_buff *frag_iter; 3019 3020 sg_init_table(&sg, 1); 3021 3022 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len); 3023 if (crypto_hash_update(desc, &sg, head_data_len)) 3024 return 1; 3025 3026 for (i = 0; i < shi->nr_frags; ++i) { 3027 const struct skb_frag_struct *f = &shi->frags[i]; 3028 sg_set_page(&sg, f->page, f->size, f->page_offset); 3029 if (crypto_hash_update(desc, &sg, f->size)) 3030 return 1; 3031 } 3032 3033 skb_walk_frags(skb, frag_iter) 3034 if (tcp_md5_hash_skb_data(hp, frag_iter, 0)) 3035 return 1; 3036 3037 return 0; 3038} 3039EXPORT_SYMBOL(tcp_md5_hash_skb_data); 3040 3041int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key) 3042{ 3043 struct scatterlist sg; 3044 3045 sg_init_one(&sg, key->key, key->keylen); 3046 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen); 3047} 3048EXPORT_SYMBOL(tcp_md5_hash_key); 3049 3050#endif 3051 3052/** 3053 * Each Responder maintains up to two secret values concurrently for 3054 * efficient secret rollover. Each secret value has 4 states: 3055 * 3056 * Generating. (tcp_secret_generating != tcp_secret_primary) 3057 * Generates new Responder-Cookies, but not yet used for primary 3058 * verification. This is a short-term state, typically lasting only 3059 * one round trip time (RTT). 3060 * 3061 * Primary. (tcp_secret_generating == tcp_secret_primary) 3062 * Used both for generation and primary verification. 3063 * 3064 * Retiring. (tcp_secret_retiring != tcp_secret_secondary) 3065 * Used for verification, until the first failure that can be 3066 * verified by the newer Generating secret. At that time, this 3067 * cookie's state is changed to Secondary, and the Generating 3068 * cookie's state is changed to Primary. This is a short-term state, 3069 * typically lasting only one round trip time (RTT). 3070 * 3071 * Secondary. (tcp_secret_retiring == tcp_secret_secondary) 3072 * Used for secondary verification, after primary verification 3073 * failures. This state lasts no more than twice the Maximum Segment 3074 * Lifetime (2MSL). Then, the secret is discarded. 3075 */ 3076struct tcp_cookie_secret { 3077 /* The secret is divided into two parts. The digest part is the 3078 * equivalent of previously hashing a secret and saving the state, 3079 * and serves as an initialization vector (IV). The message part 3080 * serves as the trailing secret. 3081 */ 3082 u32 secrets[COOKIE_WORKSPACE_WORDS]; 3083 unsigned long expires; 3084}; 3085 3086#define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL) 3087#define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2) 3088#define TCP_SECRET_LIFE (HZ * 600) 3089 3090static struct tcp_cookie_secret tcp_secret_one; 3091static struct tcp_cookie_secret tcp_secret_two; 3092 3093/* Essentially a circular list, without dynamic allocation. */ 3094static struct tcp_cookie_secret *tcp_secret_generating; 3095static struct tcp_cookie_secret *tcp_secret_primary; 3096static struct tcp_cookie_secret *tcp_secret_retiring; 3097static struct tcp_cookie_secret *tcp_secret_secondary; 3098 3099static DEFINE_SPINLOCK(tcp_secret_locker); 3100 3101/* Select a pseudo-random word in the cookie workspace. 3102 */ 3103static inline u32 tcp_cookie_work(const u32 *ws, const int n) 3104{ 3105 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])]; 3106} 3107 3108/* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed. 3109 * Called in softirq context. 3110 * Returns: 0 for success. 3111 */ 3112int tcp_cookie_generator(u32 *bakery) 3113{ 3114 unsigned long jiffy = jiffies; 3115 3116 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) { 3117 spin_lock_bh(&tcp_secret_locker); 3118 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) { 3119 /* refreshed by another */ 3120 memcpy(bakery, 3121 &tcp_secret_generating->secrets[0], 3122 COOKIE_WORKSPACE_WORDS); 3123 } else { 3124 /* still needs refreshing */ 3125 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS); 3126 3127 /* The first time, paranoia assumes that the 3128 * randomization function isn't as strong. But, 3129 * this secret initialization is delayed until 3130 * the last possible moment (packet arrival). 3131 * Although that time is observable, it is 3132 * unpredictably variable. Mash in the most 3133 * volatile clock bits available, and expire the 3134 * secret extra quickly. 3135 */ 3136 if (unlikely(tcp_secret_primary->expires == 3137 tcp_secret_secondary->expires)) { 3138 struct timespec tv; 3139 3140 getnstimeofday(&tv); 3141 bakery[COOKIE_DIGEST_WORDS+0] ^= 3142 (u32)tv.tv_nsec; 3143 3144 tcp_secret_secondary->expires = jiffy 3145 + TCP_SECRET_1MSL 3146 + (0x0f & tcp_cookie_work(bakery, 0)); 3147 } else { 3148 tcp_secret_secondary->expires = jiffy 3149 + TCP_SECRET_LIFE 3150 + (0xff & tcp_cookie_work(bakery, 1)); 3151 tcp_secret_primary->expires = jiffy 3152 + TCP_SECRET_2MSL 3153 + (0x1f & tcp_cookie_work(bakery, 2)); 3154 } 3155 memcpy(&tcp_secret_secondary->secrets[0], 3156 bakery, COOKIE_WORKSPACE_WORDS); 3157 3158 rcu_assign_pointer(tcp_secret_generating, 3159 tcp_secret_secondary); 3160 rcu_assign_pointer(tcp_secret_retiring, 3161 tcp_secret_primary); 3162 /* 3163 * Neither call_rcu() nor synchronize_rcu() needed. 3164 * Retiring data is not freed. It is replaced after 3165 * further (locked) pointer updates, and a quiet time 3166 * (minimum 1MSL, maximum LIFE - 2MSL). 3167 */ 3168 } 3169 spin_unlock_bh(&tcp_secret_locker); 3170 } else { 3171 rcu_read_lock_bh(); 3172 memcpy(bakery, 3173 &rcu_dereference(tcp_secret_generating)->secrets[0], 3174 COOKIE_WORKSPACE_WORDS); 3175 rcu_read_unlock_bh(); 3176 } 3177 return 0; 3178} 3179EXPORT_SYMBOL(tcp_cookie_generator); 3180 3181void tcp_done(struct sock *sk) 3182{ 3183 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV) 3184 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 3185 3186 tcp_set_state(sk, TCP_CLOSE); 3187 tcp_clear_xmit_timers(sk); 3188 3189 sk->sk_shutdown = SHUTDOWN_MASK; 3190 3191 if (!sock_flag(sk, SOCK_DEAD)) 3192 sk->sk_state_change(sk); 3193 else 3194 inet_csk_destroy_sock(sk); 3195} 3196EXPORT_SYMBOL_GPL(tcp_done); 3197 3198extern struct tcp_congestion_ops tcp_reno; 3199 3200static __initdata unsigned long thash_entries; 3201static int __init set_thash_entries(char *str) 3202{ 3203 if (!str) 3204 return 0; 3205 thash_entries = simple_strtoul(str, &str, 0); 3206 return 1; 3207} 3208__setup("thash_entries=", set_thash_entries); 3209 3210void __init tcp_init(void) 3211{ 3212 struct sk_buff *skb = NULL; 3213 unsigned long nr_pages, limit; 3214 int i, max_share, cnt; 3215 unsigned long jiffy = jiffies; 3216 3217 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb)); 3218 3219 percpu_counter_init(&tcp_sockets_allocated, 0); 3220 percpu_counter_init(&tcp_orphan_count, 0); 3221 tcp_hashinfo.bind_bucket_cachep = 3222 kmem_cache_create("tcp_bind_bucket", 3223 sizeof(struct inet_bind_bucket), 0, 3224 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL); 3225 3226 /* Size and allocate the main established and bind bucket 3227 * hash tables. 3228 * 3229 * The methodology is similar to that of the buffer cache. 3230 */ 3231 tcp_hashinfo.ehash = 3232 alloc_large_system_hash("TCP established", 3233 sizeof(struct inet_ehash_bucket), 3234 thash_entries, 3235 (totalram_pages >= 128 * 1024) ? 3236 13 : 15, 3237 0, 3238 NULL, 3239 &tcp_hashinfo.ehash_mask, 3240 thash_entries ? 0 : 512 * 1024); 3241 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) { 3242 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i); 3243 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i); 3244 } 3245 if (inet_ehash_locks_alloc(&tcp_hashinfo)) 3246 panic("TCP: failed to alloc ehash_locks"); 3247 tcp_hashinfo.bhash = 3248 alloc_large_system_hash("TCP bind", 3249 sizeof(struct inet_bind_hashbucket), 3250 tcp_hashinfo.ehash_mask + 1, 3251 (totalram_pages >= 128 * 1024) ? 3252 13 : 15, 3253 0, 3254 &tcp_hashinfo.bhash_size, 3255 NULL, 3256 64 * 1024); 3257 tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size; 3258 for (i = 0; i < tcp_hashinfo.bhash_size; i++) { 3259 spin_lock_init(&tcp_hashinfo.bhash[i].lock); 3260 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain); 3261 } 3262 3263 3264 cnt = tcp_hashinfo.ehash_mask + 1; 3265 3266 tcp_death_row.sysctl_max_tw_buckets = cnt / 2; 3267 sysctl_tcp_max_orphans = cnt / 2; 3268 sysctl_max_syn_backlog = max(128, cnt / 256); 3269 3270 /* Set the pressure threshold to be a fraction of global memory that 3271 * is up to 1/2 at 256 MB, decreasing toward zero with the amount of 3272 * memory, with a floor of 128 pages. 3273 */ 3274 nr_pages = totalram_pages - totalhigh_pages; 3275 limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT); 3276 limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11); 3277 limit = max(limit, 128UL); 3278 sysctl_tcp_mem[0] = limit / 4 * 3; 3279 sysctl_tcp_mem[1] = limit; 3280 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; 3281 3282 /* Set per-socket limits to no more than 1/128 the pressure threshold */ 3283 limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7); 3284 max_share = min(4UL*1024*1024, limit); 3285 3286 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM; 3287 sysctl_tcp_wmem[1] = 16*1024; 3288 sysctl_tcp_wmem[2] = max(64*1024, max_share); 3289 3290 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM; 3291 sysctl_tcp_rmem[1] = 87380; 3292 sysctl_tcp_rmem[2] = max(87380, max_share); 3293 3294 printk(KERN_INFO "TCP: Hash tables configured " 3295 "(established %u bind %u)\n", 3296 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size); 3297 3298 tcp_register_congestion_control(&tcp_reno); 3299 3300 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets)); 3301 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets)); 3302 tcp_secret_one.expires = jiffy; /* past due */ 3303 tcp_secret_two.expires = jiffy; /* past due */ 3304 tcp_secret_generating = &tcp_secret_one; 3305 tcp_secret_primary = &tcp_secret_one; 3306 tcp_secret_retiring = &tcp_secret_two; 3307 tcp_secret_secondary = &tcp_secret_two; 3308} 3309