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> 279 280#include <typedefs.h> 281#include <bcmdefs.h> 282 283int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT; 284 285struct percpu_counter tcp_orphan_count; 286EXPORT_SYMBOL_GPL(tcp_orphan_count); 287 288int sysctl_tcp_mem[3] __read_mostly; 289int sysctl_tcp_wmem[3] __read_mostly; 290int sysctl_tcp_rmem[3] __read_mostly; 291 292EXPORT_SYMBOL(sysctl_tcp_mem); 293EXPORT_SYMBOL(sysctl_tcp_rmem); 294EXPORT_SYMBOL(sysctl_tcp_wmem); 295 296atomic_t tcp_memory_allocated; /* Current allocated memory. */ 297EXPORT_SYMBOL(tcp_memory_allocated); 298 299/* 300 * Current number of TCP sockets. 301 */ 302struct percpu_counter tcp_sockets_allocated; 303EXPORT_SYMBOL(tcp_sockets_allocated); 304 305/* 306 * TCP splice context 307 */ 308struct tcp_splice_state { 309 struct pipe_inode_info *pipe; 310 size_t len; 311 unsigned int flags; 312}; 313 314/* 315 * Pressure flag: try to collapse. 316 * Technical note: it is used by multiple contexts non atomically. 317 * All the __sk_mem_schedule() is of this nature: accounting 318 * is strict, actions are advisory and have some latency. 319 */ 320int tcp_memory_pressure __read_mostly; 321EXPORT_SYMBOL(tcp_memory_pressure); 322 323void tcp_enter_memory_pressure(struct sock *sk) 324{ 325 if (!tcp_memory_pressure) { 326 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES); 327 tcp_memory_pressure = 1; 328 } 329} 330EXPORT_SYMBOL(tcp_enter_memory_pressure); 331 332/* Convert seconds to retransmits based on initial and max timeout */ 333static u8 secs_to_retrans(int seconds, int timeout, int rto_max) 334{ 335 u8 res = 0; 336 337 if (seconds > 0) { 338 int period = timeout; 339 340 res = 1; 341 while (seconds > period && res < 255) { 342 res++; 343 timeout <<= 1; 344 if (timeout > rto_max) 345 timeout = rto_max; 346 period += timeout; 347 } 348 } 349 return res; 350} 351 352/* Convert retransmits to seconds based on initial and max timeout */ 353static int retrans_to_secs(u8 retrans, int timeout, int rto_max) 354{ 355 int period = 0; 356 357 if (retrans > 0) { 358 period = timeout; 359 while (--retrans) { 360 timeout <<= 1; 361 if (timeout > rto_max) 362 timeout = rto_max; 363 period += timeout; 364 } 365 } 366 return period; 367} 368 369/* 370 * Wait for a TCP event. 371 * 372 * Note that we don't need to lock the socket, as the upper poll layers 373 * take care of normal races (between the test and the event) and we don't 374 * go look at any of the socket buffers directly. 375 */ 376unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait) 377{ 378 unsigned int mask; 379 struct sock *sk = sock->sk; 380 struct tcp_sock *tp = tcp_sk(sk); 381 382 sock_poll_wait(file, sk_sleep(sk), wait); 383 if (sk->sk_state == TCP_LISTEN) 384 return inet_csk_listen_poll(sk); 385 386 /* Socket is not locked. We are protected from async events 387 * by poll logic and correct handling of state changes 388 * made by other threads is impossible in any case. 389 */ 390 391 mask = 0; 392 393 /* 394 * POLLHUP is certainly not done right. But poll() doesn't 395 * have a notion of HUP in just one direction, and for a 396 * socket the read side is more interesting. 397 * 398 * Some poll() documentation says that POLLHUP is incompatible 399 * with the POLLOUT/POLLWR flags, so somebody should check this 400 * all. But careful, it tends to be safer to return too many 401 * bits than too few, and you can easily break real applications 402 * if you don't tell them that something has hung up! 403 * 404 * Check-me. 405 * 406 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and 407 * our fs/select.c). It means that after we received EOF, 408 * poll always returns immediately, making impossible poll() on write() 409 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP 410 * if and only if shutdown has been made in both directions. 411 * Actually, it is interesting to look how Solaris and DUX 412 * solve this dilemma. I would prefer, if POLLHUP were maskable, 413 * then we could set it on SND_SHUTDOWN. BTW examples given 414 * in Stevens' books assume exactly this behaviour, it explains 415 * why POLLHUP is incompatible with POLLOUT. --ANK 416 * 417 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent 418 * blocking on fresh not-connected or disconnected socket. --ANK 419 */ 420 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE) 421 mask |= POLLHUP; 422 if (sk->sk_shutdown & RCV_SHUTDOWN) 423 mask |= POLLIN | POLLRDNORM | POLLRDHUP; 424 425 /* Connected? */ 426 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) { 427 int target = sock_rcvlowat(sk, 0, INT_MAX); 428 429 if (tp->urg_seq == tp->copied_seq && 430 !sock_flag(sk, SOCK_URGINLINE) && 431 tp->urg_data) 432 target++; 433 434 /* Potential race condition. If read of tp below will 435 * escape above sk->sk_state, we can be illegally awaken 436 * in SYN_* states. */ 437 if (tp->rcv_nxt - tp->copied_seq >= target) 438 mask |= POLLIN | POLLRDNORM; 439 440 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) { 441 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) { 442 mask |= POLLOUT | POLLWRNORM; 443 } else { /* send SIGIO later */ 444 set_bit(SOCK_ASYNC_NOSPACE, 445 &sk->sk_socket->flags); 446 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 447 448 /* Race breaker. If space is freed after 449 * wspace test but before the flags are set, 450 * IO signal will be lost. 451 */ 452 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) 453 mask |= POLLOUT | POLLWRNORM; 454 } 455 } else 456 mask |= POLLOUT | POLLWRNORM; 457 458 if (tp->urg_data & TCP_URG_VALID) 459 mask |= POLLPRI; 460 } 461 /* This barrier is coupled with smp_wmb() in tcp_reset() */ 462 smp_rmb(); 463 if (sk->sk_err) 464 mask |= POLLERR; 465 466 return mask; 467} 468EXPORT_SYMBOL(tcp_poll); 469 470int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg) 471{ 472 struct tcp_sock *tp = tcp_sk(sk); 473 int answ; 474 475 switch (cmd) { 476 case SIOCINQ: 477 if (sk->sk_state == TCP_LISTEN) 478 return -EINVAL; 479 480 lock_sock(sk); 481 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 482 answ = 0; 483 else if (sock_flag(sk, SOCK_URGINLINE) || 484 !tp->urg_data || 485 before(tp->urg_seq, tp->copied_seq) || 486 !before(tp->urg_seq, tp->rcv_nxt)) { 487 struct sk_buff *skb; 488 489 answ = tp->rcv_nxt - tp->copied_seq; 490 491 /* Subtract 1, if FIN is in queue. */ 492 skb = skb_peek_tail(&sk->sk_receive_queue); 493 if (answ && skb) 494 answ -= tcp_hdr(skb)->fin; 495 } else 496 answ = tp->urg_seq - tp->copied_seq; 497 release_sock(sk); 498 break; 499 case SIOCATMARK: 500 answ = tp->urg_data && tp->urg_seq == tp->copied_seq; 501 break; 502 case SIOCOUTQ: 503 if (sk->sk_state == TCP_LISTEN) 504 return -EINVAL; 505 506 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 507 answ = 0; 508 else 509 answ = tp->write_seq - tp->snd_una; 510 break; 511 default: 512 return -ENOIOCTLCMD; 513 } 514 515 return put_user(answ, (int __user *)arg); 516} 517EXPORT_SYMBOL(tcp_ioctl); 518 519static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb) 520{ 521 TCP_SKB_CB(skb)->flags |= TCPHDR_PSH; 522 tp->pushed_seq = tp->write_seq; 523} 524 525static inline int forced_push(struct tcp_sock *tp) 526{ 527 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1)); 528} 529 530static inline void skb_entail(struct sock *sk, struct sk_buff *skb) 531{ 532 struct tcp_sock *tp = tcp_sk(sk); 533 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 534 535 skb->csum = 0; 536 tcb->seq = tcb->end_seq = tp->write_seq; 537 tcb->flags = TCPHDR_ACK; 538 tcb->sacked = 0; 539 skb_header_release(skb); 540 tcp_add_write_queue_tail(sk, skb); 541 sk->sk_wmem_queued += skb->truesize; 542 sk_mem_charge(sk, skb->truesize); 543 if (tp->nonagle & TCP_NAGLE_PUSH) 544 tp->nonagle &= ~TCP_NAGLE_PUSH; 545} 546 547static inline void tcp_mark_urg(struct tcp_sock *tp, int flags) 548{ 549 if (flags & MSG_OOB) 550 tp->snd_up = tp->write_seq; 551} 552 553static inline void tcp_push(struct sock *sk, int flags, int mss_now, 554 int nonagle) 555{ 556 if (tcp_send_head(sk)) { 557 struct tcp_sock *tp = tcp_sk(sk); 558 559 if (!(flags & MSG_MORE) || forced_push(tp)) 560 tcp_mark_push(tp, tcp_write_queue_tail(sk)); 561 562 tcp_mark_urg(tp, flags); 563 __tcp_push_pending_frames(sk, mss_now, 564 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle); 565 } 566} 567 568static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, 569 unsigned int offset, size_t len) 570{ 571 struct tcp_splice_state *tss = rd_desc->arg.data; 572 int ret; 573 574 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len), 575 tss->flags); 576 if (ret > 0) 577 rd_desc->count -= ret; 578 return ret; 579} 580 581static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss) 582{ 583 /* Store TCP splice context information in read_descriptor_t. */ 584 read_descriptor_t rd_desc = { 585 .arg.data = tss, 586 .count = tss->len, 587 }; 588 589 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv); 590} 591 592/** 593 * tcp_splice_read - splice data from TCP socket to a pipe 594 * @sock: socket to splice from 595 * @ppos: position (not valid) 596 * @pipe: pipe to splice to 597 * @len: number of bytes to splice 598 * @flags: splice modifier flags 599 * 600 * Description: 601 * Will read pages from given socket and fill them into a pipe. 602 * 603 **/ 604ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos, 605 struct pipe_inode_info *pipe, size_t len, 606 unsigned int flags) 607{ 608 struct sock *sk = sock->sk; 609 struct tcp_splice_state tss = { 610 .pipe = pipe, 611 .len = len, 612 .flags = flags, 613 }; 614 long timeo; 615 ssize_t spliced; 616 int ret; 617 618 sock_rps_record_flow(sk); 619 /* 620 * We can't seek on a socket input 621 */ 622 if (unlikely(*ppos)) 623 return -ESPIPE; 624 625 ret = spliced = 0; 626 627 lock_sock(sk); 628 629 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK); 630 while (tss.len) { 631 ret = __tcp_splice_read(sk, &tss); 632 if (ret < 0) 633 break; 634 else if (!ret) { 635 if (spliced) 636 break; 637 if (sock_flag(sk, SOCK_DONE)) 638 break; 639 if (sk->sk_err) { 640 ret = sock_error(sk); 641 break; 642 } 643 if (sk->sk_shutdown & RCV_SHUTDOWN) 644 break; 645 if (sk->sk_state == TCP_CLOSE) { 646 /* 647 * This occurs when user tries to read 648 * from never connected socket. 649 */ 650 if (!sock_flag(sk, SOCK_DONE)) 651 ret = -ENOTCONN; 652 break; 653 } 654 if (!timeo) { 655 ret = -EAGAIN; 656 break; 657 } 658 sk_wait_data(sk, &timeo); 659 if (signal_pending(current)) { 660 ret = sock_intr_errno(timeo); 661 break; 662 } 663 continue; 664 } 665 tss.len -= ret; 666 spliced += ret; 667 668 if (!timeo) 669 break; 670 release_sock(sk); 671 lock_sock(sk); 672 673 if (sk->sk_err || sk->sk_state == TCP_CLOSE || 674 (sk->sk_shutdown & RCV_SHUTDOWN) || 675 signal_pending(current)) 676 break; 677 } 678 679 release_sock(sk); 680 681 if (spliced) 682 return spliced; 683 684 return ret; 685} 686EXPORT_SYMBOL(tcp_splice_read); 687 688struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp) 689{ 690 struct sk_buff *skb; 691 692 /* The TCP header must be at least 32-bit aligned. */ 693 size = ALIGN(size, 4); 694 695 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp); 696 if (skb) { 697 if (sk_wmem_schedule(sk, skb->truesize)) { 698 /* 699 * Make sure that we have exactly size bytes 700 * available to the caller, no more, no less. 701 */ 702 skb_reserve(skb, skb_tailroom(skb) - size); 703 return skb; 704 } 705 __kfree_skb(skb); 706 } else { 707 sk->sk_prot->enter_memory_pressure(sk); 708 sk_stream_moderate_sndbuf(sk); 709 } 710 return NULL; 711} 712 713static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now, 714 int large_allowed) 715{ 716 struct tcp_sock *tp = tcp_sk(sk); 717 u32 xmit_size_goal, old_size_goal; 718 719 xmit_size_goal = mss_now; 720 721 if (large_allowed && sk_can_gso(sk)) { 722 xmit_size_goal = ((sk->sk_gso_max_size - 1) - 723 inet_csk(sk)->icsk_af_ops->net_header_len - 724 inet_csk(sk)->icsk_ext_hdr_len - 725 tp->tcp_header_len); 726 727 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal); 728 729 /* We try hard to avoid divides here */ 730 old_size_goal = tp->xmit_size_goal_segs * mss_now; 731 732 if (likely(old_size_goal <= xmit_size_goal && 733 old_size_goal + mss_now > xmit_size_goal)) { 734 xmit_size_goal = old_size_goal; 735 } else { 736 tp->xmit_size_goal_segs = xmit_size_goal / mss_now; 737 xmit_size_goal = tp->xmit_size_goal_segs * mss_now; 738 } 739 } 740 741 return max(xmit_size_goal, mss_now); 742} 743 744static int tcp_send_mss(struct sock *sk, int *size_goal, int flags) 745{ 746 int mss_now; 747 748 mss_now = tcp_current_mss(sk); 749 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB)); 750 751 return mss_now; 752} 753 754static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset, 755 size_t psize, int flags) 756{ 757 struct tcp_sock *tp = tcp_sk(sk); 758 int mss_now, size_goal; 759 int err; 760 ssize_t copied; 761 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 762 763 /* Wait for a connection to finish. */ 764 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 765 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0) 766 goto out_err; 767 768 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags); 769 770 mss_now = tcp_send_mss(sk, &size_goal, flags); 771 copied = 0; 772 773 err = -EPIPE; 774 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 775 goto out_err; 776 777 while (psize > 0) { 778 struct sk_buff *skb = tcp_write_queue_tail(sk); 779 struct page *page = pages[poffset / PAGE_SIZE]; 780 int copy, i, can_coalesce; 781 int offset = poffset % PAGE_SIZE; 782 int size = min_t(size_t, psize, PAGE_SIZE - offset); 783 784 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) { 785new_segment: 786 if (!sk_stream_memory_free(sk)) 787 goto wait_for_sndbuf; 788 789 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation); 790 if (!skb) 791 goto wait_for_memory; 792 793 skb_entail(sk, skb); 794 copy = size_goal; 795 } 796 797 if (copy > size) 798 copy = size; 799 800 i = skb_shinfo(skb)->nr_frags; 801 can_coalesce = skb_can_coalesce(skb, i, page, offset); 802 if (!can_coalesce && i >= MAX_SKB_FRAGS) { 803 tcp_mark_push(tp, skb); 804 goto new_segment; 805 } 806 if (!sk_wmem_schedule(sk, copy)) 807 goto wait_for_memory; 808 809 if (can_coalesce) { 810 skb_shinfo(skb)->frags[i - 1].size += copy; 811 } else { 812 get_page(page); 813 skb_fill_page_desc(skb, i, page, offset, copy); 814 } 815 816 skb->len += copy; 817 skb->data_len += copy; 818 skb->truesize += copy; 819 sk->sk_wmem_queued += copy; 820 sk_mem_charge(sk, copy); 821 skb->ip_summed = CHECKSUM_PARTIAL; 822 tp->write_seq += copy; 823 TCP_SKB_CB(skb)->end_seq += copy; 824 skb_shinfo(skb)->gso_segs = 0; 825 826 if (!copied) 827 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH; 828 829 copied += copy; 830 poffset += copy; 831 if (!(psize -= copy)) 832 goto out; 833 834 if (skb->len < size_goal || (flags & MSG_OOB)) 835 continue; 836 837 if (forced_push(tp)) { 838 tcp_mark_push(tp, skb); 839 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 840 } else if (skb == tcp_send_head(sk)) 841 tcp_push_one(sk, mss_now); 842 continue; 843 844wait_for_sndbuf: 845 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 846wait_for_memory: 847 if (copied) 848 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH); 849 850 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0) 851 goto do_error; 852 853 mss_now = tcp_send_mss(sk, &size_goal, flags); 854 } 855 856out: 857#ifdef CONFIG_BCM47XX 858 if (copied && !(flags & MSG_MORE)) 859#else 860 if (copied) 861#endif 862 tcp_push(sk, flags, mss_now, tp->nonagle); 863 return copied; 864 865do_error: 866 if (copied) 867 goto out; 868out_err: 869 return sk_stream_error(sk, flags, err); 870} 871 872int BCMFASTPATH_HOST tcp_sendpage(struct sock *sk, struct page *page, int offset, 873 size_t size, int flags) 874{ 875 ssize_t res; 876 877 if (!(sk->sk_route_caps & NETIF_F_SG) || 878 !(sk->sk_route_caps & NETIF_F_ALL_CSUM)) 879 return sock_no_sendpage(sk->sk_socket, page, offset, size, 880 flags); 881 882 lock_sock(sk); 883 TCP_CHECK_TIMER(sk); 884 res = do_tcp_sendpages(sk, &page, offset, size, flags); 885 TCP_CHECK_TIMER(sk); 886 release_sock(sk); 887 return res; 888} 889EXPORT_SYMBOL(tcp_sendpage); 890 891#define TCP_PAGE(sk) (sk->sk_sndmsg_page) 892#define TCP_OFF(sk) (sk->sk_sndmsg_off) 893 894static inline int select_size(struct sock *sk, int sg) 895{ 896 struct tcp_sock *tp = tcp_sk(sk); 897 int tmp = tp->mss_cache; 898 899 if (sg) { 900 if (sk_can_gso(sk)) 901 tmp = 0; 902 else { 903 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER); 904 905 if (tmp >= pgbreak && 906 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE) 907 tmp = pgbreak; 908 } 909 } 910 911 return tmp; 912} 913 914int BCMFASTPATH_HOST tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg, 915 size_t size) 916{ 917 struct iovec *iov; 918 struct tcp_sock *tp = tcp_sk(sk); 919 struct sk_buff *skb; 920 int iovlen, flags; 921 int mss_now, size_goal; 922 int sg, err, copied; 923 long timeo; 924 925 lock_sock(sk); 926 TCP_CHECK_TIMER(sk); 927 928 flags = msg->msg_flags; 929 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 930 931 /* Wait for a connection to finish. */ 932 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 933 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0) 934 goto out_err; 935 936 /* This should be in poll */ 937 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags); 938 939 mss_now = tcp_send_mss(sk, &size_goal, flags); 940 941 /* Ok commence sending. */ 942 iovlen = msg->msg_iovlen; 943 iov = msg->msg_iov; 944 copied = 0; 945 946 err = -EPIPE; 947 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 948 goto out_err; 949 950 sg = sk->sk_route_caps & NETIF_F_SG; 951 952 while (--iovlen >= 0) { 953 size_t seglen = iov->iov_len; 954 unsigned char __user *from = iov->iov_base; 955 956 iov++; 957 958 while (seglen > 0) { 959 int copy = 0; 960 int max = size_goal; 961 962 skb = tcp_write_queue_tail(sk); 963 if (tcp_send_head(sk)) { 964 if (skb->ip_summed == CHECKSUM_NONE) 965 max = mss_now; 966 copy = max - skb->len; 967 } 968 969 if (copy <= 0) { 970new_segment: 971 /* Allocate new segment. If the interface is SG, 972 * allocate skb fitting to single page. 973 */ 974 if (!sk_stream_memory_free(sk)) 975 goto wait_for_sndbuf; 976 977 skb = sk_stream_alloc_skb(sk, 978 select_size(sk, sg), 979 sk->sk_allocation); 980 if (!skb) 981 goto wait_for_memory; 982 983 /* 984 * Check whether we can use HW checksum. 985 */ 986 if (sk->sk_route_caps & NETIF_F_ALL_CSUM) 987 skb->ip_summed = CHECKSUM_PARTIAL; 988 989 skb_entail(sk, skb); 990 copy = size_goal; 991 max = size_goal; 992 } 993 994 /* Try to append data to the end of skb. */ 995 if (copy > seglen) 996 copy = seglen; 997 998 /* Where to copy to? */ 999 if (skb_tailroom(skb) > 0) { 1000 /* We have some space in skb head. Superb! */ 1001 if (copy > skb_tailroom(skb)) 1002 copy = skb_tailroom(skb); 1003 if ((err = skb_add_data(skb, from, copy)) != 0) 1004 goto do_fault; 1005 } else { 1006 int merge = 0; 1007 int i = skb_shinfo(skb)->nr_frags; 1008 struct page *page = TCP_PAGE(sk); 1009 int off = TCP_OFF(sk); 1010 1011 if (skb_can_coalesce(skb, i, page, off) && 1012 off != PAGE_SIZE) { 1013 /* We can extend the last page 1014 * fragment. */ 1015 merge = 1; 1016 } else if (i == MAX_SKB_FRAGS || !sg) { 1017 /* Need to add new fragment and cannot 1018 * do this because interface is non-SG, 1019 * or because all the page slots are 1020 * busy. */ 1021 tcp_mark_push(tp, skb); 1022 goto new_segment; 1023 } else if (page) { 1024 if (off == PAGE_SIZE) { 1025 put_page(page); 1026 TCP_PAGE(sk) = page = NULL; 1027 off = 0; 1028 } 1029 } else 1030 off = 0; 1031 1032 if (copy > PAGE_SIZE - off) 1033 copy = PAGE_SIZE - off; 1034 1035 if (!sk_wmem_schedule(sk, copy)) 1036 goto wait_for_memory; 1037 1038 if (!page) { 1039 /* Allocate new cache page. */ 1040 if (!(page = sk_stream_alloc_page(sk))) 1041 goto wait_for_memory; 1042 } 1043 1044 /* Time to copy data. We are close to 1045 * the end! */ 1046 err = skb_copy_to_page(sk, from, skb, page, 1047 off, copy); 1048 if (err) { 1049 /* If this page was new, give it to the 1050 * socket so it does not get leaked. 1051 */ 1052 if (!TCP_PAGE(sk)) { 1053 TCP_PAGE(sk) = page; 1054 TCP_OFF(sk) = 0; 1055 } 1056 goto do_error; 1057 } 1058 1059 /* Update the skb. */ 1060 if (merge) { 1061 skb_shinfo(skb)->frags[i - 1].size += 1062 copy; 1063 } else { 1064 skb_fill_page_desc(skb, i, page, off, copy); 1065 if (TCP_PAGE(sk)) { 1066 get_page(page); 1067 } else if (off + copy < PAGE_SIZE) { 1068 get_page(page); 1069 TCP_PAGE(sk) = page; 1070 } 1071 } 1072 1073 TCP_OFF(sk) = off + copy; 1074 } 1075 1076 if (!copied) 1077 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH; 1078 1079 tp->write_seq += copy; 1080 TCP_SKB_CB(skb)->end_seq += copy; 1081 skb_shinfo(skb)->gso_segs = 0; 1082 1083 from += copy; 1084 copied += copy; 1085 if ((seglen -= copy) == 0 && iovlen == 0) 1086 goto out; 1087 1088 if (skb->len < max || (flags & MSG_OOB)) 1089 continue; 1090 1091 if (forced_push(tp)) { 1092 tcp_mark_push(tp, skb); 1093 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 1094 } else if (skb == tcp_send_head(sk)) 1095 tcp_push_one(sk, mss_now); 1096 continue; 1097 1098wait_for_sndbuf: 1099 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1100wait_for_memory: 1101 if (copied) 1102 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH); 1103 1104 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0) 1105 goto do_error; 1106 1107 mss_now = tcp_send_mss(sk, &size_goal, flags); 1108 } 1109 } 1110 1111out: 1112 if (copied) 1113 tcp_push(sk, flags, mss_now, tp->nonagle); 1114 TCP_CHECK_TIMER(sk); 1115 release_sock(sk); 1116 return copied; 1117 1118do_fault: 1119 if (!skb->len) { 1120 tcp_unlink_write_queue(skb, sk); 1121 /* It is the one place in all of TCP, except connection 1122 * reset, where we can be unlinking the send_head. 1123 */ 1124 tcp_check_send_head(sk, skb); 1125 sk_wmem_free_skb(sk, skb); 1126 } 1127 1128do_error: 1129 if (copied) 1130 goto out; 1131out_err: 1132 err = sk_stream_error(sk, flags, err); 1133 TCP_CHECK_TIMER(sk); 1134 release_sock(sk); 1135 return err; 1136} 1137EXPORT_SYMBOL(tcp_sendmsg); 1138 1139/* 1140 * Handle reading urgent data. BSD has very simple semantics for 1141 * this, no blocking and very strange errors 8) 1142 */ 1143 1144static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags) 1145{ 1146 struct tcp_sock *tp = tcp_sk(sk); 1147 1148 /* No URG data to read. */ 1149 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data || 1150 tp->urg_data == TCP_URG_READ) 1151 return -EINVAL; /* Yes this is right ! */ 1152 1153 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE)) 1154 return -ENOTCONN; 1155 1156 if (tp->urg_data & TCP_URG_VALID) { 1157 int err = 0; 1158 char c = tp->urg_data; 1159 1160 if (!(flags & MSG_PEEK)) 1161 tp->urg_data = TCP_URG_READ; 1162 1163 /* Read urgent data. */ 1164 msg->msg_flags |= MSG_OOB; 1165 1166 if (len > 0) { 1167 if (!(flags & MSG_TRUNC)) 1168 err = memcpy_toiovec(msg->msg_iov, &c, 1); 1169 len = 1; 1170 } else 1171 msg->msg_flags |= MSG_TRUNC; 1172 1173 return err ? -EFAULT : len; 1174 } 1175 1176 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 1177 return 0; 1178 1179 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and 1180 * the available implementations agree in this case: 1181 * this call should never block, independent of the 1182 * blocking state of the socket. 1183 * Mike <pall@rz.uni-karlsruhe.de> 1184 */ 1185 return -EAGAIN; 1186} 1187 1188/* Clean up the receive buffer for full frames taken by the user, 1189 * then send an ACK if necessary. COPIED is the number of bytes 1190 * tcp_recvmsg has given to the user so far, it speeds up the 1191 * calculation of whether or not we must ACK for the sake of 1192 * a window update. 1193 */ 1194void tcp_cleanup_rbuf(struct sock *sk, int copied) 1195{ 1196 struct tcp_sock *tp = tcp_sk(sk); 1197 int time_to_ack = 0; 1198 1199#if TCP_DEBUG 1200 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 1201 1202 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq), 1203 KERN_INFO "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n", 1204 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt); 1205#endif 1206 1207 if (inet_csk_ack_scheduled(sk)) { 1208 const struct inet_connection_sock *icsk = inet_csk(sk); 1209 /* Delayed ACKs frequently hit locked sockets during bulk 1210 * receive. */ 1211 if (icsk->icsk_ack.blocked || 1212 /* Once-per-two-segments ACK was not sent by tcp_input.c */ 1213 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss || 1214 /* 1215 * If this read emptied read buffer, we send ACK, if 1216 * connection is not bidirectional, user drained 1217 * receive buffer and there was a small segment 1218 * in queue. 1219 */ 1220 (copied > 0 && 1221 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) || 1222 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) && 1223 !icsk->icsk_ack.pingpong)) && 1224 !atomic_read(&sk->sk_rmem_alloc))) 1225 time_to_ack = 1; 1226 } 1227 1228 /* We send an ACK if we can now advertise a non-zero window 1229 * which has been raised "significantly". 1230 * 1231 * Even if window raised up to infinity, do not send window open ACK 1232 * in states, where we will not receive more. It is useless. 1233 */ 1234 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) { 1235 __u32 rcv_window_now = tcp_receive_window(tp); 1236 1237 /* Optimize, __tcp_select_window() is not cheap. */ 1238 if (2*rcv_window_now <= tp->window_clamp) { 1239 __u32 new_window = __tcp_select_window(sk); 1240 1241 /* Send ACK now, if this read freed lots of space 1242 * in our buffer. Certainly, new_window is new window. 1243 * We can advertise it now, if it is not less than current one. 1244 * "Lots" means "at least twice" here. 1245 */ 1246 if (new_window && new_window >= 2 * rcv_window_now) 1247 time_to_ack = 1; 1248 } 1249 } 1250 if (time_to_ack) 1251 tcp_send_ack(sk); 1252} 1253 1254static void tcp_prequeue_process(struct sock *sk) 1255{ 1256 struct sk_buff *skb; 1257 struct tcp_sock *tp = tcp_sk(sk); 1258 1259 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED); 1260 1261 /* RX process wants to run with disabled BHs, though it is not 1262 * necessary */ 1263 local_bh_disable(); 1264 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL) 1265 sk_backlog_rcv(sk, skb); 1266 local_bh_enable(); 1267 1268 /* Clear memory counter. */ 1269 tp->ucopy.memory = 0; 1270} 1271 1272#ifdef CONFIG_NET_DMA 1273static void tcp_service_net_dma(struct sock *sk, bool wait) 1274{ 1275 dma_cookie_t done, used; 1276 dma_cookie_t last_issued; 1277 struct tcp_sock *tp = tcp_sk(sk); 1278 1279 if (!tp->ucopy.dma_chan) 1280 return; 1281 1282 last_issued = tp->ucopy.dma_cookie; 1283 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1284 1285 do { 1286 if (dma_async_memcpy_complete(tp->ucopy.dma_chan, 1287 last_issued, &done, 1288 &used) == DMA_SUCCESS) { 1289 /* Safe to free early-copied skbs now */ 1290 __skb_queue_purge(&sk->sk_async_wait_queue); 1291 break; 1292 } else { 1293 struct sk_buff *skb; 1294 while ((skb = skb_peek(&sk->sk_async_wait_queue)) && 1295 (dma_async_is_complete(skb->dma_cookie, done, 1296 used) == DMA_SUCCESS)) { 1297 __skb_dequeue(&sk->sk_async_wait_queue); 1298 kfree_skb(skb); 1299 } 1300 } 1301 } while (wait); 1302} 1303#endif 1304 1305static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off) 1306{ 1307 struct sk_buff *skb; 1308 u32 offset; 1309 1310 skb_queue_walk(&sk->sk_receive_queue, skb) { 1311 offset = seq - TCP_SKB_CB(skb)->seq; 1312 if (tcp_hdr(skb)->syn) 1313 offset--; 1314 if (offset < skb->len || tcp_hdr(skb)->fin) { 1315 *off = offset; 1316 return skb; 1317 } 1318 } 1319 return NULL; 1320} 1321 1322/* 1323 * This routine provides an alternative to tcp_recvmsg() for routines 1324 * that would like to handle copying from skbuffs directly in 'sendfile' 1325 * fashion. 1326 * Note: 1327 * - It is assumed that the socket was locked by the caller. 1328 * - The routine does not block. 1329 * - At present, there is no support for reading OOB data 1330 * or for 'peeking' the socket using this routine 1331 * (although both would be easy to implement). 1332 */ 1333int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, 1334 sk_read_actor_t recv_actor) 1335{ 1336 struct sk_buff *skb; 1337 struct tcp_sock *tp = tcp_sk(sk); 1338 u32 seq = tp->copied_seq; 1339 u32 offset; 1340 int copied = 0; 1341 1342 if (sk->sk_state == TCP_LISTEN) 1343 return -ENOTCONN; 1344 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) { 1345 if (offset < skb->len) { 1346 int used; 1347 size_t len; 1348 1349 len = skb->len - offset; 1350 /* Stop reading if we hit a patch of urgent data */ 1351 if (tp->urg_data) { 1352 u32 urg_offset = tp->urg_seq - seq; 1353 if (urg_offset < len) 1354 len = urg_offset; 1355 if (!len) 1356 break; 1357 } 1358 used = recv_actor(desc, skb, offset, len); 1359 if (used < 0) { 1360 if (!copied) 1361 copied = used; 1362 break; 1363 } else if (used <= len) { 1364 seq += used; 1365 copied += used; 1366 offset += used; 1367 } 1368 /* 1369 * If recv_actor drops the lock (e.g. TCP splice 1370 * receive) the skb pointer might be invalid when 1371 * getting here: tcp_collapse might have deleted it 1372 * while aggregating skbs from the socket queue. 1373 */ 1374 skb = tcp_recv_skb(sk, seq-1, &offset); 1375 if (!skb || (offset+1 != skb->len)) 1376 break; 1377 } 1378 if (tcp_hdr(skb)->fin) { 1379 sk_eat_skb(sk, skb, 0); 1380 ++seq; 1381 break; 1382 } 1383 sk_eat_skb(sk, skb, 0); 1384 if (!desc->count) 1385 break; 1386 tp->copied_seq = seq; 1387 } 1388 tp->copied_seq = seq; 1389 1390 tcp_rcv_space_adjust(sk); 1391 1392 /* Clean up data we have read: This will do ACK frames. */ 1393 if (copied > 0) 1394 tcp_cleanup_rbuf(sk, copied); 1395 return copied; 1396} 1397EXPORT_SYMBOL(tcp_read_sock); 1398 1399/* 1400 * This routine copies from a sock struct into the user buffer. 1401 * 1402 * Technical note: in 2.3 we work on _locked_ socket, so that 1403 * tricks with *seq access order and skb->users are not required. 1404 * Probably, code can be easily improved even more. 1405 */ 1406 1407int BCMFASTPATH_HOST tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg, 1408 size_t len, int nonblock, int flags, int *addr_len) 1409{ 1410 struct tcp_sock *tp = tcp_sk(sk); 1411 int copied = 0; 1412 u32 peek_seq; 1413 u32 *seq; 1414 unsigned long used; 1415 int err; 1416 int target; /* Read at least this many bytes */ 1417 long timeo; 1418 struct task_struct *user_recv = NULL; 1419 int copied_early = 0; 1420 struct sk_buff *skb; 1421 u32 urg_hole = 0; 1422 1423 lock_sock(sk); 1424 1425 TCP_CHECK_TIMER(sk); 1426 1427 err = -ENOTCONN; 1428 if (sk->sk_state == TCP_LISTEN) 1429 goto out; 1430 1431 timeo = sock_rcvtimeo(sk, nonblock); 1432 1433 /* Urgent data needs to be handled specially. */ 1434 if (flags & MSG_OOB) 1435 goto recv_urg; 1436 1437 seq = &tp->copied_seq; 1438 if (flags & MSG_PEEK) { 1439 peek_seq = tp->copied_seq; 1440 seq = &peek_seq; 1441 } 1442 1443 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 1444 1445#ifdef CONFIG_NET_DMA 1446 tp->ucopy.dma_chan = NULL; 1447 preempt_disable(); 1448 skb = skb_peek_tail(&sk->sk_receive_queue); 1449 { 1450 int available = 0; 1451 1452 if (skb) 1453 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq); 1454 if ((available < target) && 1455 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) && 1456 !sysctl_tcp_low_latency && 1457 dma_find_channel(DMA_MEMCPY)) { 1458 preempt_enable_no_resched(); 1459 tp->ucopy.pinned_list = 1460 dma_pin_iovec_pages(msg->msg_iov, len); 1461 } else { 1462 preempt_enable_no_resched(); 1463 } 1464 } 1465#endif 1466 1467 do { 1468 u32 offset; 1469 1470 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */ 1471 if (tp->urg_data && tp->urg_seq == *seq) { 1472 if (copied) 1473 break; 1474 if (signal_pending(current)) { 1475 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN; 1476 break; 1477 } 1478 } 1479 1480 /* Next get a buffer. */ 1481 1482 skb_queue_walk(&sk->sk_receive_queue, skb) { 1483 /* Now that we have two receive queues this 1484 * shouldn't happen. 1485 */ 1486 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq), 1487 KERN_INFO "recvmsg bug: copied %X " 1488 "seq %X rcvnxt %X fl %X\n", *seq, 1489 TCP_SKB_CB(skb)->seq, tp->rcv_nxt, 1490 flags)) 1491 break; 1492 1493 offset = *seq - TCP_SKB_CB(skb)->seq; 1494 if (tcp_hdr(skb)->syn) 1495 offset--; 1496 if (offset < skb->len) 1497 goto found_ok_skb; 1498 if (tcp_hdr(skb)->fin) 1499 goto found_fin_ok; 1500 WARN(!(flags & MSG_PEEK), KERN_INFO "recvmsg bug 2: " 1501 "copied %X seq %X rcvnxt %X fl %X\n", 1502 *seq, TCP_SKB_CB(skb)->seq, 1503 tp->rcv_nxt, flags); 1504 } 1505 1506 /* Well, if we have backlog, try to process it now yet. */ 1507 1508 if (copied >= target && !sk->sk_backlog.tail) 1509 break; 1510 1511 if (copied) { 1512 if (sk->sk_err || 1513 sk->sk_state == TCP_CLOSE || 1514 (sk->sk_shutdown & RCV_SHUTDOWN) || 1515 !timeo || 1516 signal_pending(current)) 1517 break; 1518 } else { 1519 if (sock_flag(sk, SOCK_DONE)) 1520 break; 1521 1522 if (sk->sk_err) { 1523 copied = sock_error(sk); 1524 break; 1525 } 1526 1527 if (sk->sk_shutdown & RCV_SHUTDOWN) 1528 break; 1529 1530 if (sk->sk_state == TCP_CLOSE) { 1531 if (!sock_flag(sk, SOCK_DONE)) { 1532 /* This occurs when user tries to read 1533 * from never connected socket. 1534 */ 1535 copied = -ENOTCONN; 1536 break; 1537 } 1538 break; 1539 } 1540 1541 if (!timeo) { 1542 copied = -EAGAIN; 1543 break; 1544 } 1545 1546 if (signal_pending(current)) { 1547 copied = sock_intr_errno(timeo); 1548 break; 1549 } 1550 } 1551 1552 tcp_cleanup_rbuf(sk, copied); 1553 1554 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) { 1555 /* Install new reader */ 1556 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) { 1557 user_recv = current; 1558 tp->ucopy.task = user_recv; 1559 tp->ucopy.iov = msg->msg_iov; 1560 } 1561 1562 tp->ucopy.len = len; 1563 1564 WARN_ON(tp->copied_seq != tp->rcv_nxt && 1565 !(flags & (MSG_PEEK | MSG_TRUNC))); 1566 1567 /* Ugly... If prequeue is not empty, we have to 1568 * process it before releasing socket, otherwise 1569 * order will be broken at second iteration. 1570 * More elegant solution is required!!! 1571 * 1572 * Look: we have the following (pseudo)queues: 1573 * 1574 * 1. packets in flight 1575 * 2. backlog 1576 * 3. prequeue 1577 * 4. receive_queue 1578 * 1579 * Each queue can be processed only if the next ones 1580 * are empty. At this point we have empty receive_queue. 1581 * But prequeue _can_ be not empty after 2nd iteration, 1582 * when we jumped to start of loop because backlog 1583 * processing added something to receive_queue. 1584 * We cannot release_sock(), because backlog contains 1585 * packets arrived _after_ prequeued ones. 1586 * 1587 * Shortly, algorithm is clear --- to process all 1588 * the queues in order. We could make it more directly, 1589 * requeueing packets from backlog to prequeue, if 1590 * is not empty. It is more elegant, but eats cycles, 1591 * unfortunately. 1592 */ 1593 if (!skb_queue_empty(&tp->ucopy.prequeue)) 1594 goto do_prequeue; 1595 1596 /* __ Set realtime policy in scheduler __ */ 1597 } 1598 1599#ifdef CONFIG_NET_DMA 1600 if (tp->ucopy.dma_chan) 1601 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1602#endif 1603 if (copied >= target) { 1604 /* Do not sleep, just process backlog. */ 1605 release_sock(sk); 1606 lock_sock(sk); 1607 } else 1608 sk_wait_data(sk, &timeo); 1609 1610#ifdef CONFIG_NET_DMA 1611 tcp_service_net_dma(sk, false); /* Don't block */ 1612 tp->ucopy.wakeup = 0; 1613#endif 1614 1615 if (user_recv) { 1616 int chunk; 1617 1618 /* __ Restore normal policy in scheduler __ */ 1619 1620 if ((chunk = len - tp->ucopy.len) != 0) { 1621 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk); 1622 len -= chunk; 1623 copied += chunk; 1624 } 1625 1626 if (tp->rcv_nxt == tp->copied_seq && 1627 !skb_queue_empty(&tp->ucopy.prequeue)) { 1628do_prequeue: 1629 tcp_prequeue_process(sk); 1630 1631 if ((chunk = len - tp->ucopy.len) != 0) { 1632 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1633 len -= chunk; 1634 copied += chunk; 1635 } 1636 } 1637 } 1638 if ((flags & MSG_PEEK) && 1639 (peek_seq - copied - urg_hole != tp->copied_seq)) { 1640 if (net_ratelimit()) 1641 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n", 1642 current->comm, task_pid_nr(current)); 1643 peek_seq = tp->copied_seq; 1644 } 1645 continue; 1646 1647 found_ok_skb: 1648 /* Ok so how much can we use? */ 1649 used = skb->len - offset; 1650 if (len < used) 1651 used = len; 1652 1653 /* Do we have urgent data here? */ 1654 if (tp->urg_data) { 1655 u32 urg_offset = tp->urg_seq - *seq; 1656 if (urg_offset < used) { 1657 if (!urg_offset) { 1658 if (!sock_flag(sk, SOCK_URGINLINE)) { 1659 ++*seq; 1660 urg_hole++; 1661 offset++; 1662 used--; 1663 if (!used) 1664 goto skip_copy; 1665 } 1666 } else 1667 used = urg_offset; 1668 } 1669 } 1670 1671 if (!(flags & MSG_TRUNC)) { 1672#ifdef CONFIG_NET_DMA 1673 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 1674 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY); 1675 1676 if (tp->ucopy.dma_chan) { 1677 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec( 1678 tp->ucopy.dma_chan, skb, offset, 1679 msg->msg_iov, used, 1680 tp->ucopy.pinned_list); 1681 1682 if (tp->ucopy.dma_cookie < 0) { 1683 1684 printk(KERN_ALERT "dma_cookie < 0\n"); 1685 1686 /* Exception. Bailout! */ 1687 if (!copied) 1688 copied = -EFAULT; 1689 break; 1690 } 1691 1692 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1693 1694 if ((offset + used) == skb->len) 1695 copied_early = 1; 1696 1697 } else 1698#endif 1699 { 1700 err = skb_copy_datagram_iovec(skb, offset, 1701 msg->msg_iov, used); 1702 if (err) { 1703 /* Exception. Bailout! */ 1704 if (!copied) 1705 copied = -EFAULT; 1706 break; 1707 } 1708 } 1709 } 1710 1711 *seq += used; 1712 copied += used; 1713 len -= used; 1714 1715 tcp_rcv_space_adjust(sk); 1716 1717skip_copy: 1718 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) { 1719 tp->urg_data = 0; 1720 tcp_fast_path_check(sk); 1721 } 1722 if (used + offset < skb->len) 1723 continue; 1724 1725 if (tcp_hdr(skb)->fin) 1726 goto found_fin_ok; 1727 if (!(flags & MSG_PEEK)) { 1728 sk_eat_skb(sk, skb, copied_early); 1729 copied_early = 0; 1730 } 1731 continue; 1732 1733 found_fin_ok: 1734 /* Process the FIN. */ 1735 ++*seq; 1736 if (!(flags & MSG_PEEK)) { 1737 sk_eat_skb(sk, skb, copied_early); 1738 copied_early = 0; 1739 } 1740 break; 1741 } while (len > 0); 1742 1743 if (user_recv) { 1744 if (!skb_queue_empty(&tp->ucopy.prequeue)) { 1745 int chunk; 1746 1747 tp->ucopy.len = copied > 0 ? len : 0; 1748 1749 tcp_prequeue_process(sk); 1750 1751 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) { 1752 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1753 len -= chunk; 1754 copied += chunk; 1755 } 1756 } 1757 1758 tp->ucopy.task = NULL; 1759 tp->ucopy.len = 0; 1760 } 1761 1762#ifdef CONFIG_NET_DMA 1763 tcp_service_net_dma(sk, true); /* Wait for queue to drain */ 1764 tp->ucopy.dma_chan = NULL; 1765 1766 if (tp->ucopy.pinned_list) { 1767 dma_unpin_iovec_pages(tp->ucopy.pinned_list); 1768 tp->ucopy.pinned_list = NULL; 1769 } 1770#endif 1771 1772 /* According to UNIX98, msg_name/msg_namelen are ignored 1773 * on connected socket. I was just happy when found this 8) --ANK 1774 */ 1775 1776 /* Clean up data we have read: This will do ACK frames. */ 1777 tcp_cleanup_rbuf(sk, copied); 1778 1779 TCP_CHECK_TIMER(sk); 1780 release_sock(sk); 1781 return copied; 1782 1783out: 1784 TCP_CHECK_TIMER(sk); 1785 release_sock(sk); 1786 return err; 1787 1788recv_urg: 1789 err = tcp_recv_urg(sk, msg, len, flags); 1790 goto out; 1791} 1792EXPORT_SYMBOL(tcp_recvmsg); 1793 1794void tcp_set_state(struct sock *sk, int state) 1795{ 1796 int oldstate = sk->sk_state; 1797 1798 switch (state) { 1799 case TCP_ESTABLISHED: 1800 if (oldstate != TCP_ESTABLISHED) 1801 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1802 break; 1803 1804 case TCP_CLOSE: 1805 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED) 1806 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS); 1807 1808 sk->sk_prot->unhash(sk); 1809 if (inet_csk(sk)->icsk_bind_hash && 1810 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK)) 1811 inet_put_port(sk); 1812 /* fall through */ 1813 default: 1814 if (oldstate == TCP_ESTABLISHED) 1815 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1816 } 1817 1818 /* Change state AFTER socket is unhashed to avoid closed 1819 * socket sitting in hash tables. 1820 */ 1821 sk->sk_state = state; 1822 1823#ifdef STATE_TRACE 1824 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]); 1825#endif 1826} 1827EXPORT_SYMBOL_GPL(tcp_set_state); 1828 1829/* 1830 * State processing on a close. This implements the state shift for 1831 * sending our FIN frame. Note that we only send a FIN for some 1832 * states. A shutdown() may have already sent the FIN, or we may be 1833 * closed. 1834 */ 1835 1836static const unsigned char new_state[16] = { 1837 /* current state: new state: action: */ 1838 /* (Invalid) */ TCP_CLOSE, 1839 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1840 /* TCP_SYN_SENT */ TCP_CLOSE, 1841 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1842 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1, 1843 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2, 1844 /* TCP_TIME_WAIT */ TCP_CLOSE, 1845 /* TCP_CLOSE */ TCP_CLOSE, 1846 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN, 1847 /* TCP_LAST_ACK */ TCP_LAST_ACK, 1848 /* TCP_LISTEN */ TCP_CLOSE, 1849 /* TCP_CLOSING */ TCP_CLOSING, 1850}; 1851 1852static int tcp_close_state(struct sock *sk) 1853{ 1854 int next = (int)new_state[sk->sk_state]; 1855 int ns = next & TCP_STATE_MASK; 1856 1857 tcp_set_state(sk, ns); 1858 1859 return next & TCP_ACTION_FIN; 1860} 1861 1862/* 1863 * Shutdown the sending side of a connection. Much like close except 1864 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD). 1865 */ 1866 1867void tcp_shutdown(struct sock *sk, int how) 1868{ 1869 /* We need to grab some memory, and put together a FIN, 1870 * and then put it into the queue to be sent. 1871 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92. 1872 */ 1873 if (!(how & SEND_SHUTDOWN)) 1874 return; 1875 1876 /* If we've already sent a FIN, or it's a closed state, skip this. */ 1877 if ((1 << sk->sk_state) & 1878 (TCPF_ESTABLISHED | TCPF_SYN_SENT | 1879 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) { 1880 /* Clear out any half completed packets. FIN if needed. */ 1881 if (tcp_close_state(sk)) 1882 tcp_send_fin(sk); 1883 } 1884} 1885EXPORT_SYMBOL(tcp_shutdown); 1886 1887void tcp_close(struct sock *sk, long timeout) 1888{ 1889 struct sk_buff *skb; 1890 int data_was_unread = 0; 1891 int state; 1892 1893 lock_sock(sk); 1894 sk->sk_shutdown = SHUTDOWN_MASK; 1895 1896 if (sk->sk_state == TCP_LISTEN) { 1897 tcp_set_state(sk, TCP_CLOSE); 1898 1899 /* Special case. */ 1900 inet_csk_listen_stop(sk); 1901 1902 goto adjudge_to_death; 1903 } 1904 1905 /* We need to flush the recv. buffs. We do this only on the 1906 * descriptor close, not protocol-sourced closes, because the 1907 * reader process may not have drained the data yet! 1908 */ 1909 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) { 1910 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq - 1911 tcp_hdr(skb)->fin; 1912 data_was_unread += len; 1913 __kfree_skb(skb); 1914 } 1915 1916 sk_mem_reclaim(sk); 1917 1918 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */ 1919 if (sk->sk_state == TCP_CLOSE) 1920 goto adjudge_to_death; 1921 1922 /* As outlined in RFC 2525, section 2.17, we send a RST here because 1923 * data was lost. To witness the awful effects of the old behavior of 1924 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk 1925 * GET in an FTP client, suspend the process, wait for the client to 1926 * advertise a zero window, then kill -9 the FTP client, wheee... 1927 * Note: timeout is always zero in such a case. 1928 */ 1929 if (data_was_unread) { 1930 /* Unread data was tossed, zap the connection. */ 1931 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE); 1932 tcp_set_state(sk, TCP_CLOSE); 1933 tcp_send_active_reset(sk, sk->sk_allocation); 1934 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) { 1935 /* Check zero linger _after_ checking for unread data. */ 1936 sk->sk_prot->disconnect(sk, 0); 1937 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 1938 } else if (tcp_close_state(sk)) { 1939 /* We FIN if the application ate all the data before 1940 * zapping the connection. 1941 */ 1942 1943 /* RED-PEN. Formally speaking, we have broken TCP state 1944 * machine. State transitions: 1945 * 1946 * TCP_ESTABLISHED -> TCP_FIN_WAIT1 1947 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible) 1948 * TCP_CLOSE_WAIT -> TCP_LAST_ACK 1949 * 1950 * are legal only when FIN has been sent (i.e. in window), 1951 * rather than queued out of window. Purists blame. 1952 * 1953 * F.e. "RFC state" is ESTABLISHED, 1954 * if Linux state is FIN-WAIT-1, but FIN is still not sent. 1955 * 1956 * The visible declinations are that sometimes 1957 * we enter time-wait state, when it is not required really 1958 * (harmless), do not send active resets, when they are 1959 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when 1960 * they look as CLOSING or LAST_ACK for Linux) 1961 * Probably, I missed some more holelets. 1962 * --ANK 1963 */ 1964 tcp_send_fin(sk); 1965 } 1966 1967 sk_stream_wait_close(sk, timeout); 1968 1969adjudge_to_death: 1970 state = sk->sk_state; 1971 sock_hold(sk); 1972 sock_orphan(sk); 1973 1974 /* It is the last release_sock in its life. It will remove backlog. */ 1975 release_sock(sk); 1976 1977 1978 /* Now socket is owned by kernel and we acquire BH lock 1979 to finish close. No need to check for user refs. 1980 */ 1981 local_bh_disable(); 1982 bh_lock_sock(sk); 1983 WARN_ON(sock_owned_by_user(sk)); 1984 1985 percpu_counter_inc(sk->sk_prot->orphan_count); 1986 1987 /* Have we already been destroyed by a softirq or backlog? */ 1988 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE) 1989 goto out; 1990 1991 /* This is a (useful) BSD violating of the RFC. There is a 1992 * problem with TCP as specified in that the other end could 1993 * keep a socket open forever with no application left this end. 1994 * We use a 3 minute timeout (about the same as BSD) then kill 1995 * our end. If they send after that then tough - BUT: long enough 1996 * that we won't make the old 4*rto = almost no time - whoops 1997 * reset mistake. 1998 * 1999 * Nope, it was not mistake. It is really desired behaviour 2000 * f.e. on http servers, when such sockets are useless, but 2001 * consume significant resources. Let's do it with special 2002 * linger2 option. --ANK 2003 */ 2004 2005 if (sk->sk_state == TCP_FIN_WAIT2) { 2006 struct tcp_sock *tp = tcp_sk(sk); 2007 if (tp->linger2 < 0) { 2008 tcp_set_state(sk, TCP_CLOSE); 2009 tcp_send_active_reset(sk, GFP_ATOMIC); 2010 NET_INC_STATS_BH(sock_net(sk), 2011 LINUX_MIB_TCPABORTONLINGER); 2012 } else { 2013 const int tmo = tcp_fin_time(sk); 2014 2015 if (tmo > TCP_TIMEWAIT_LEN) { 2016 inet_csk_reset_keepalive_timer(sk, 2017 tmo - TCP_TIMEWAIT_LEN); 2018 } else { 2019 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 2020 goto out; 2021 } 2022 } 2023 } 2024 if (sk->sk_state != TCP_CLOSE) { 2025 sk_mem_reclaim(sk); 2026 if (tcp_too_many_orphans(sk, 0)) { 2027 if (net_ratelimit()) 2028 printk(KERN_INFO "TCP: too many of orphaned " 2029 "sockets\n"); 2030 tcp_set_state(sk, TCP_CLOSE); 2031 tcp_send_active_reset(sk, GFP_ATOMIC); 2032 NET_INC_STATS_BH(sock_net(sk), 2033 LINUX_MIB_TCPABORTONMEMORY); 2034 } 2035 } 2036 2037 if (sk->sk_state == TCP_CLOSE) 2038 inet_csk_destroy_sock(sk); 2039 /* Otherwise, socket is reprieved until protocol close. */ 2040 2041out: 2042 bh_unlock_sock(sk); 2043 local_bh_enable(); 2044 sock_put(sk); 2045} 2046EXPORT_SYMBOL(tcp_close); 2047 2048/* These states need RST on ABORT according to RFC793 */ 2049 2050static inline int tcp_need_reset(int state) 2051{ 2052 return (1 << state) & 2053 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 | 2054 TCPF_FIN_WAIT2 | TCPF_SYN_RECV); 2055} 2056 2057int tcp_disconnect(struct sock *sk, int flags) 2058{ 2059 struct inet_sock *inet = inet_sk(sk); 2060 struct inet_connection_sock *icsk = inet_csk(sk); 2061 struct tcp_sock *tp = tcp_sk(sk); 2062 int err = 0; 2063 int old_state = sk->sk_state; 2064 2065 if (old_state != TCP_CLOSE) 2066 tcp_set_state(sk, TCP_CLOSE); 2067 2068 /* ABORT function of RFC793 */ 2069 if (old_state == TCP_LISTEN) { 2070 inet_csk_listen_stop(sk); 2071 } else if (tcp_need_reset(old_state) || 2072 (tp->snd_nxt != tp->write_seq && 2073 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) { 2074 /* The last check adjusts for discrepancy of Linux wrt. RFC 2075 * states 2076 */ 2077 tcp_send_active_reset(sk, gfp_any()); 2078 sk->sk_err = ECONNRESET; 2079 } else if (old_state == TCP_SYN_SENT) 2080 sk->sk_err = ECONNRESET; 2081 2082 tcp_clear_xmit_timers(sk); 2083 __skb_queue_purge(&sk->sk_receive_queue); 2084 tcp_write_queue_purge(sk); 2085 __skb_queue_purge(&tp->out_of_order_queue); 2086#ifdef CONFIG_NET_DMA 2087 __skb_queue_purge(&sk->sk_async_wait_queue); 2088#endif 2089 2090 inet->inet_dport = 0; 2091 2092 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK)) 2093 inet_reset_saddr(sk); 2094 2095 sk->sk_shutdown = 0; 2096 sock_reset_flag(sk, SOCK_DONE); 2097 tp->srtt = 0; 2098 if ((tp->write_seq += tp->max_window + 2) == 0) 2099 tp->write_seq = 1; 2100 icsk->icsk_backoff = 0; 2101 tp->snd_cwnd = 2; 2102 icsk->icsk_probes_out = 0; 2103 tp->packets_out = 0; 2104 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 2105 tp->snd_cwnd_cnt = 0; 2106 tp->bytes_acked = 0; 2107 tp->window_clamp = 0; 2108 tcp_set_ca_state(sk, TCP_CA_Open); 2109 tcp_clear_retrans(tp); 2110 inet_csk_delack_init(sk); 2111 tcp_init_send_head(sk); 2112 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt)); 2113 __sk_dst_reset(sk); 2114 2115 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash); 2116 2117 sk->sk_error_report(sk); 2118 return err; 2119} 2120EXPORT_SYMBOL(tcp_disconnect); 2121 2122/* 2123 * Socket option code for TCP. 2124 */ 2125static int do_tcp_setsockopt(struct sock *sk, int level, 2126 int optname, char __user *optval, unsigned int optlen) 2127{ 2128 struct tcp_sock *tp = tcp_sk(sk); 2129 struct inet_connection_sock *icsk = inet_csk(sk); 2130 int val; 2131 int err = 0; 2132 2133 /* These are data/string values, all the others are ints */ 2134 switch (optname) { 2135 case TCP_CONGESTION: { 2136 char name[TCP_CA_NAME_MAX]; 2137 2138 if (optlen < 1) 2139 return -EINVAL; 2140 2141 val = strncpy_from_user(name, optval, 2142 min_t(long, TCP_CA_NAME_MAX-1, optlen)); 2143 if (val < 0) 2144 return -EFAULT; 2145 name[val] = 0; 2146 2147 lock_sock(sk); 2148 err = tcp_set_congestion_control(sk, name); 2149 release_sock(sk); 2150 return err; 2151 } 2152 case TCP_COOKIE_TRANSACTIONS: { 2153 struct tcp_cookie_transactions ctd; 2154 struct tcp_cookie_values *cvp = NULL; 2155 2156 if (sizeof(ctd) > optlen) 2157 return -EINVAL; 2158 if (copy_from_user(&ctd, optval, sizeof(ctd))) 2159 return -EFAULT; 2160 2161 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) || 2162 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED) 2163 return -EINVAL; 2164 2165 if (ctd.tcpct_cookie_desired == 0) { 2166 /* default to global value */ 2167 } else if ((0x1 & ctd.tcpct_cookie_desired) || 2168 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX || 2169 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) { 2170 return -EINVAL; 2171 } 2172 2173 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) { 2174 /* Supercedes all other values */ 2175 lock_sock(sk); 2176 if (tp->cookie_values != NULL) { 2177 kref_put(&tp->cookie_values->kref, 2178 tcp_cookie_values_release); 2179 tp->cookie_values = NULL; 2180 } 2181 tp->rx_opt.cookie_in_always = 0; /* false */ 2182 tp->rx_opt.cookie_out_never = 1; /* true */ 2183 release_sock(sk); 2184 return err; 2185 } 2186 2187 /* Allocate ancillary memory before locking. 2188 */ 2189 if (ctd.tcpct_used > 0 || 2190 (tp->cookie_values == NULL && 2191 (sysctl_tcp_cookie_size > 0 || 2192 ctd.tcpct_cookie_desired > 0 || 2193 ctd.tcpct_s_data_desired > 0))) { 2194 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used, 2195 GFP_KERNEL); 2196 if (cvp == NULL) 2197 return -ENOMEM; 2198 2199 kref_init(&cvp->kref); 2200 } 2201 lock_sock(sk); 2202 tp->rx_opt.cookie_in_always = 2203 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags); 2204 tp->rx_opt.cookie_out_never = 0; /* false */ 2205 2206 if (tp->cookie_values != NULL) { 2207 if (cvp != NULL) { 2208 /* Changed values are recorded by a changed 2209 * pointer, ensuring the cookie will differ, 2210 * without separately hashing each value later. 2211 */ 2212 kref_put(&tp->cookie_values->kref, 2213 tcp_cookie_values_release); 2214 } else { 2215 cvp = tp->cookie_values; 2216 } 2217 } 2218 2219 if (cvp != NULL) { 2220 cvp->cookie_desired = ctd.tcpct_cookie_desired; 2221 2222 if (ctd.tcpct_used > 0) { 2223 memcpy(cvp->s_data_payload, ctd.tcpct_value, 2224 ctd.tcpct_used); 2225 cvp->s_data_desired = ctd.tcpct_used; 2226 cvp->s_data_constant = 1; /* true */ 2227 } else { 2228 /* No constant payload data. */ 2229 cvp->s_data_desired = ctd.tcpct_s_data_desired; 2230 cvp->s_data_constant = 0; /* false */ 2231 } 2232 2233 tp->cookie_values = cvp; 2234 } 2235 release_sock(sk); 2236 return err; 2237 } 2238 default: 2239 /* fallthru */ 2240 break; 2241 } 2242 2243 if (optlen < sizeof(int)) 2244 return -EINVAL; 2245 2246 if (get_user(val, (int __user *)optval)) 2247 return -EFAULT; 2248 2249 lock_sock(sk); 2250 2251 switch (optname) { 2252 case TCP_MAXSEG: 2253 /* Values greater than interface MTU won't take effect. However 2254 * at the point when this call is done we typically don't yet 2255 * know which interface is going to be used */ 2256 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) { 2257 err = -EINVAL; 2258 break; 2259 } 2260 tp->rx_opt.user_mss = val; 2261 break; 2262 2263 case TCP_NODELAY: 2264 if (val) { 2265 /* TCP_NODELAY is weaker than TCP_CORK, so that 2266 * this option on corked socket is remembered, but 2267 * it is not activated until cork is cleared. 2268 * 2269 * However, when TCP_NODELAY is set we make 2270 * an explicit push, which overrides even TCP_CORK 2271 * for currently queued segments. 2272 */ 2273 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH; 2274 tcp_push_pending_frames(sk); 2275 } else { 2276 tp->nonagle &= ~TCP_NAGLE_OFF; 2277 } 2278 break; 2279 2280 case TCP_THIN_LINEAR_TIMEOUTS: 2281 if (val < 0 || val > 1) 2282 err = -EINVAL; 2283 else 2284 tp->thin_lto = val; 2285 break; 2286 2287 case TCP_THIN_DUPACK: 2288 if (val < 0 || val > 1) 2289 err = -EINVAL; 2290 else 2291 tp->thin_dupack = val; 2292 break; 2293 2294 case TCP_CORK: 2295 /* When set indicates to always queue non-full frames. 2296 * Later the user clears this option and we transmit 2297 * any pending partial frames in the queue. This is 2298 * meant to be used alongside sendfile() to get properly 2299 * filled frames when the user (for example) must write 2300 * out headers with a write() call first and then use 2301 * sendfile to send out the data parts. 2302 * 2303 * TCP_CORK can be set together with TCP_NODELAY and it is 2304 * stronger than TCP_NODELAY. 2305 */ 2306 if (val) { 2307 tp->nonagle |= TCP_NAGLE_CORK; 2308 } else { 2309 tp->nonagle &= ~TCP_NAGLE_CORK; 2310 if (tp->nonagle&TCP_NAGLE_OFF) 2311 tp->nonagle |= TCP_NAGLE_PUSH; 2312 tcp_push_pending_frames(sk); 2313 } 2314 break; 2315 2316 case TCP_KEEPIDLE: 2317 if (val < 1 || val > MAX_TCP_KEEPIDLE) 2318 err = -EINVAL; 2319 else { 2320 tp->keepalive_time = val * HZ; 2321 if (sock_flag(sk, SOCK_KEEPOPEN) && 2322 !((1 << sk->sk_state) & 2323 (TCPF_CLOSE | TCPF_LISTEN))) { 2324 u32 elapsed = keepalive_time_elapsed(tp); 2325 if (tp->keepalive_time > elapsed) 2326 elapsed = tp->keepalive_time - elapsed; 2327 else 2328 elapsed = 0; 2329 inet_csk_reset_keepalive_timer(sk, elapsed); 2330 } 2331 } 2332 break; 2333 case TCP_KEEPINTVL: 2334 if (val < 1 || val > MAX_TCP_KEEPINTVL) 2335 err = -EINVAL; 2336 else 2337 tp->keepalive_intvl = val * HZ; 2338 break; 2339 case TCP_KEEPCNT: 2340 if (val < 1 || val > MAX_TCP_KEEPCNT) 2341 err = -EINVAL; 2342 else 2343 tp->keepalive_probes = val; 2344 break; 2345 case TCP_SYNCNT: 2346 if (val < 1 || val > MAX_TCP_SYNCNT) 2347 err = -EINVAL; 2348 else 2349 icsk->icsk_syn_retries = val; 2350 break; 2351 2352 case TCP_LINGER2: 2353 if (val < 0) 2354 tp->linger2 = -1; 2355 else if (val > sysctl_tcp_fin_timeout / HZ) 2356 tp->linger2 = 0; 2357 else 2358 tp->linger2 = val * HZ; 2359 break; 2360 2361 case TCP_DEFER_ACCEPT: 2362 /* Translate value in seconds to number of retransmits */ 2363 icsk->icsk_accept_queue.rskq_defer_accept = 2364 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ, 2365 TCP_RTO_MAX / HZ); 2366 break; 2367 2368 case TCP_WINDOW_CLAMP: 2369 if (!val) { 2370 if (sk->sk_state != TCP_CLOSE) { 2371 err = -EINVAL; 2372 break; 2373 } 2374 tp->window_clamp = 0; 2375 } else 2376 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ? 2377 SOCK_MIN_RCVBUF / 2 : val; 2378 break; 2379 2380 case TCP_QUICKACK: 2381 if (!val) { 2382 icsk->icsk_ack.pingpong = 1; 2383 } else { 2384 icsk->icsk_ack.pingpong = 0; 2385 if ((1 << sk->sk_state) & 2386 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) && 2387 inet_csk_ack_scheduled(sk)) { 2388 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 2389 tcp_cleanup_rbuf(sk, 1); 2390 if (!(val & 1)) 2391 icsk->icsk_ack.pingpong = 1; 2392 } 2393 } 2394 break; 2395 2396#ifdef CONFIG_TCP_MD5SIG 2397 case TCP_MD5SIG: 2398 /* Read the IP->Key mappings from userspace */ 2399 err = tp->af_specific->md5_parse(sk, optval, optlen); 2400 break; 2401#endif 2402 2403 default: 2404 err = -ENOPROTOOPT; 2405 break; 2406 } 2407 2408 release_sock(sk); 2409 return err; 2410} 2411 2412int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval, 2413 unsigned int optlen) 2414{ 2415 struct inet_connection_sock *icsk = inet_csk(sk); 2416 2417 if (level != SOL_TCP) 2418 return icsk->icsk_af_ops->setsockopt(sk, level, optname, 2419 optval, optlen); 2420 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2421} 2422EXPORT_SYMBOL(tcp_setsockopt); 2423 2424#ifdef CONFIG_COMPAT 2425int compat_tcp_setsockopt(struct sock *sk, int level, int optname, 2426 char __user *optval, unsigned int optlen) 2427{ 2428 if (level != SOL_TCP) 2429 return inet_csk_compat_setsockopt(sk, level, optname, 2430 optval, optlen); 2431 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2432} 2433EXPORT_SYMBOL(compat_tcp_setsockopt); 2434#endif 2435 2436/* Return information about state of tcp endpoint in API format. */ 2437void tcp_get_info(struct sock *sk, struct tcp_info *info) 2438{ 2439 struct tcp_sock *tp = tcp_sk(sk); 2440 const struct inet_connection_sock *icsk = inet_csk(sk); 2441 u32 now = tcp_time_stamp; 2442 2443 memset(info, 0, sizeof(*info)); 2444 2445 info->tcpi_state = sk->sk_state; 2446 info->tcpi_ca_state = icsk->icsk_ca_state; 2447 info->tcpi_retransmits = icsk->icsk_retransmits; 2448 info->tcpi_probes = icsk->icsk_probes_out; 2449 info->tcpi_backoff = icsk->icsk_backoff; 2450 2451 if (tp->rx_opt.tstamp_ok) 2452 info->tcpi_options |= TCPI_OPT_TIMESTAMPS; 2453 if (tcp_is_sack(tp)) 2454 info->tcpi_options |= TCPI_OPT_SACK; 2455 if (tp->rx_opt.wscale_ok) { 2456 info->tcpi_options |= TCPI_OPT_WSCALE; 2457 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale; 2458 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale; 2459 } 2460 2461 if (tp->ecn_flags&TCP_ECN_OK) 2462 info->tcpi_options |= TCPI_OPT_ECN; 2463 2464 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto); 2465 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato); 2466 info->tcpi_snd_mss = tp->mss_cache; 2467 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss; 2468 2469 if (sk->sk_state == TCP_LISTEN) { 2470 info->tcpi_unacked = sk->sk_ack_backlog; 2471 info->tcpi_sacked = sk->sk_max_ack_backlog; 2472 } else { 2473 info->tcpi_unacked = tp->packets_out; 2474 info->tcpi_sacked = tp->sacked_out; 2475 } 2476 info->tcpi_lost = tp->lost_out; 2477 info->tcpi_retrans = tp->retrans_out; 2478 info->tcpi_fackets = tp->fackets_out; 2479 2480 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime); 2481 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime); 2482 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp); 2483 2484 info->tcpi_pmtu = icsk->icsk_pmtu_cookie; 2485 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh; 2486 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3; 2487 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2; 2488 info->tcpi_snd_ssthresh = tp->snd_ssthresh; 2489 info->tcpi_snd_cwnd = tp->snd_cwnd; 2490 info->tcpi_advmss = tp->advmss; 2491 info->tcpi_reordering = tp->reordering; 2492 2493 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3; 2494 info->tcpi_rcv_space = tp->rcvq_space.space; 2495 2496 info->tcpi_total_retrans = tp->total_retrans; 2497} 2498EXPORT_SYMBOL_GPL(tcp_get_info); 2499 2500static int do_tcp_getsockopt(struct sock *sk, int level, 2501 int optname, char __user *optval, int __user *optlen) 2502{ 2503 struct inet_connection_sock *icsk = inet_csk(sk); 2504 struct tcp_sock *tp = tcp_sk(sk); 2505 int val, len; 2506 2507 if (get_user(len, optlen)) 2508 return -EFAULT; 2509 2510 len = min_t(unsigned int, len, sizeof(int)); 2511 2512 if (len < 0) 2513 return -EINVAL; 2514 2515 switch (optname) { 2516 case TCP_MAXSEG: 2517 val = tp->mss_cache; 2518 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 2519 val = tp->rx_opt.user_mss; 2520 break; 2521 case TCP_NODELAY: 2522 val = !!(tp->nonagle&TCP_NAGLE_OFF); 2523 break; 2524 case TCP_CORK: 2525 val = !!(tp->nonagle&TCP_NAGLE_CORK); 2526 break; 2527 case TCP_KEEPIDLE: 2528 val = keepalive_time_when(tp) / HZ; 2529 break; 2530 case TCP_KEEPINTVL: 2531 val = keepalive_intvl_when(tp) / HZ; 2532 break; 2533 case TCP_KEEPCNT: 2534 val = keepalive_probes(tp); 2535 break; 2536 case TCP_SYNCNT: 2537 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries; 2538 break; 2539 case TCP_LINGER2: 2540 val = tp->linger2; 2541 if (val >= 0) 2542 val = (val ? : sysctl_tcp_fin_timeout) / HZ; 2543 break; 2544 case TCP_DEFER_ACCEPT: 2545 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept, 2546 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ); 2547 break; 2548 case TCP_WINDOW_CLAMP: 2549 val = tp->window_clamp; 2550 break; 2551 case TCP_INFO: { 2552 struct tcp_info info; 2553 2554 if (get_user(len, optlen)) 2555 return -EFAULT; 2556 2557 tcp_get_info(sk, &info); 2558 2559 len = min_t(unsigned int, len, sizeof(info)); 2560 if (put_user(len, optlen)) 2561 return -EFAULT; 2562 if (copy_to_user(optval, &info, len)) 2563 return -EFAULT; 2564 return 0; 2565 } 2566 case TCP_QUICKACK: 2567 val = !icsk->icsk_ack.pingpong; 2568 break; 2569 2570 case TCP_CONGESTION: 2571 if (get_user(len, optlen)) 2572 return -EFAULT; 2573 len = min_t(unsigned int, len, TCP_CA_NAME_MAX); 2574 if (put_user(len, optlen)) 2575 return -EFAULT; 2576 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len)) 2577 return -EFAULT; 2578 return 0; 2579 2580 case TCP_COOKIE_TRANSACTIONS: { 2581 struct tcp_cookie_transactions ctd; 2582 struct tcp_cookie_values *cvp = tp->cookie_values; 2583 2584 if (get_user(len, optlen)) 2585 return -EFAULT; 2586 if (len < sizeof(ctd)) 2587 return -EINVAL; 2588 2589 memset(&ctd, 0, sizeof(ctd)); 2590 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ? 2591 TCP_COOKIE_IN_ALWAYS : 0) 2592 | (tp->rx_opt.cookie_out_never ? 2593 TCP_COOKIE_OUT_NEVER : 0); 2594 2595 if (cvp != NULL) { 2596 ctd.tcpct_flags |= (cvp->s_data_in ? 2597 TCP_S_DATA_IN : 0) 2598 | (cvp->s_data_out ? 2599 TCP_S_DATA_OUT : 0); 2600 2601 ctd.tcpct_cookie_desired = cvp->cookie_desired; 2602 ctd.tcpct_s_data_desired = cvp->s_data_desired; 2603 2604 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0], 2605 cvp->cookie_pair_size); 2606 ctd.tcpct_used = cvp->cookie_pair_size; 2607 } 2608 2609 if (put_user(sizeof(ctd), optlen)) 2610 return -EFAULT; 2611 if (copy_to_user(optval, &ctd, sizeof(ctd))) 2612 return -EFAULT; 2613 return 0; 2614 } 2615 case TCP_THIN_LINEAR_TIMEOUTS: 2616 val = tp->thin_lto; 2617 break; 2618 case TCP_THIN_DUPACK: 2619 val = tp->thin_dupack; 2620 break; 2621 default: 2622 return -ENOPROTOOPT; 2623 } 2624 2625 if (put_user(len, optlen)) 2626 return -EFAULT; 2627 if (copy_to_user(optval, &val, len)) 2628 return -EFAULT; 2629 return 0; 2630} 2631 2632int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval, 2633 int __user *optlen) 2634{ 2635 struct inet_connection_sock *icsk = inet_csk(sk); 2636 2637 if (level != SOL_TCP) 2638 return icsk->icsk_af_ops->getsockopt(sk, level, optname, 2639 optval, optlen); 2640 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2641} 2642EXPORT_SYMBOL(tcp_getsockopt); 2643 2644#ifdef CONFIG_COMPAT 2645int compat_tcp_getsockopt(struct sock *sk, int level, int optname, 2646 char __user *optval, int __user *optlen) 2647{ 2648 if (level != SOL_TCP) 2649 return inet_csk_compat_getsockopt(sk, level, optname, 2650 optval, optlen); 2651 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2652} 2653EXPORT_SYMBOL(compat_tcp_getsockopt); 2654#endif 2655 2656struct sk_buff BCMFASTPATH_HOST *tcp_tso_segment(struct sk_buff *skb, int features) 2657{ 2658 struct sk_buff *segs = ERR_PTR(-EINVAL); 2659 struct tcphdr *th; 2660 unsigned thlen; 2661 unsigned int seq; 2662 __be32 delta; 2663 unsigned int oldlen; 2664 unsigned int mss; 2665 2666 if (!pskb_may_pull(skb, sizeof(*th))) 2667 goto out; 2668 2669 th = tcp_hdr(skb); 2670 thlen = th->doff * 4; 2671 if (thlen < sizeof(*th)) 2672 goto out; 2673 2674 if (!pskb_may_pull(skb, thlen)) 2675 goto out; 2676 2677 oldlen = (u16)~skb->len; 2678 __skb_pull(skb, thlen); 2679 2680 mss = skb_shinfo(skb)->gso_size; 2681 if (unlikely(skb->len <= mss)) 2682 goto out; 2683 2684 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) { 2685 /* Packet is from an untrusted source, reset gso_segs. */ 2686 int type = skb_shinfo(skb)->gso_type; 2687 2688 if (unlikely(type & 2689 ~(SKB_GSO_TCPV4 | 2690 SKB_GSO_DODGY | 2691 SKB_GSO_TCP_ECN | 2692 SKB_GSO_TCPV6 | 2693 0) || 2694 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))) 2695 goto out; 2696 2697 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss); 2698 2699 segs = NULL; 2700 goto out; 2701 } 2702 2703 /* 2704 * For normal optimized packet handling, it calls skb_tcp_segmented(). 2705 * However, packets marked for Netfilter needs to be segmented using 2706 * the old method since the packets are passed up to the application 2707 * layer. 2708 */ 2709 if (!skb->tcpf_nf && (skb->protocol != htons(ETH_P_IPV6))) { 2710 return skb_tcp_segment(skb, features, oldlen, thlen); 2711 } 2712 2713 /* Old method */ 2714 skb->tcpf_nf = 0; 2715 segs = skb_segment(skb, features); 2716 if (IS_ERR(segs)) 2717 goto out; 2718 2719 delta = htonl(oldlen + (thlen + mss)); 2720 2721 skb = segs; 2722 th = tcp_hdr(skb); 2723 seq = ntohl(th->seq); 2724 2725 do { 2726 th->fin = th->psh = 0; 2727 2728 th->check = ~csum_fold((__force __wsum)((__force u32)th->check + 2729 (__force u32)delta)); 2730 if (skb->ip_summed != CHECKSUM_PARTIAL) 2731 th->check = 2732 csum_fold(csum_partial(skb_transport_header(skb), 2733 thlen, skb->csum)); 2734 2735 seq += mss; 2736 skb = skb->next; 2737 th = tcp_hdr(skb); 2738 2739 th->seq = htonl(seq); 2740 th->cwr = 0; 2741 } while (skb->next); 2742 2743 delta = htonl(oldlen + (skb->tail - skb->transport_header) + 2744 skb->data_len); 2745 th->check = ~csum_fold((__force __wsum)((__force u32)th->check + 2746 (__force u32)delta)); 2747 if (skb->ip_summed != CHECKSUM_PARTIAL) 2748 th->check = csum_fold(csum_partial(skb_transport_header(skb), 2749 thlen, skb->csum)); 2750 2751out: 2752 return segs; 2753} 2754EXPORT_SYMBOL(tcp_tso_segment); 2755 2756struct sk_buff ** BCMFASTPATH_HOST tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb) 2757{ 2758 struct sk_buff **pp = NULL; 2759 struct sk_buff *p; 2760 struct tcphdr *th; 2761 struct tcphdr *th2; 2762 unsigned int len; 2763 unsigned int thlen; 2764 __be32 flags; 2765 unsigned int mss = 1; 2766 unsigned int hlen; 2767 unsigned int off; 2768 int flush = 1; 2769 int i; 2770 2771 off = skb_gro_offset(skb); 2772 hlen = off + sizeof(*th); 2773 th = skb_gro_header_fast(skb, off); 2774 if (skb_gro_header_hard(skb, hlen)) { 2775 th = skb_gro_header_slow(skb, hlen, off); 2776 if (unlikely(!th)) 2777 goto out; 2778 } 2779 2780 thlen = th->doff * 4; 2781 if (thlen < sizeof(*th)) 2782 goto out; 2783 2784 hlen = off + thlen; 2785 if (skb_gro_header_hard(skb, hlen)) { 2786 th = skb_gro_header_slow(skb, hlen, off); 2787 if (unlikely(!th)) 2788 goto out; 2789 } 2790 2791 skb_gro_pull(skb, thlen); 2792 2793 len = skb_gro_len(skb); 2794 flags = tcp_flag_word(th); 2795 2796 for (; (p = *head); head = &p->next) { 2797 if (!NAPI_GRO_CB(p)->same_flow) 2798 continue; 2799 2800 th2 = tcp_hdr(p); 2801 2802 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) { 2803 NAPI_GRO_CB(p)->same_flow = 0; 2804 continue; 2805 } 2806 2807 goto found; 2808 } 2809 2810 goto out_check_final; 2811 2812found: 2813 flush = NAPI_GRO_CB(p)->flush; 2814 flush |= (__force int)(flags & TCP_FLAG_CWR); 2815 flush |= (__force int)((flags ^ tcp_flag_word(th2)) & 2816 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH)); 2817 flush |= (__force int)(th->ack_seq ^ th2->ack_seq); 2818 for (i = sizeof(*th); i < thlen; i += 4) 2819 flush |= *(u32 *)((u8 *)th + i) ^ 2820 *(u32 *)((u8 *)th2 + i); 2821 2822 mss = skb_shinfo(p)->gso_size; 2823 2824 flush |= (len - 1) >= mss; 2825 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq); 2826 2827 if (flush || skb_gro_receive(head, skb)) { 2828 mss = 1; 2829 goto out_check_final; 2830 } 2831 2832 p = *head; 2833 th2 = tcp_hdr(p); 2834 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH); 2835 2836out_check_final: 2837 flush = len < mss; 2838 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH | 2839 TCP_FLAG_RST | TCP_FLAG_SYN | 2840 TCP_FLAG_FIN)); 2841 2842 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush)) 2843 pp = head; 2844 2845out: 2846 NAPI_GRO_CB(skb)->flush |= flush; 2847 2848 return pp; 2849} 2850EXPORT_SYMBOL(tcp_gro_receive); 2851 2852int BCMFASTPATH_HOST tcp_gro_complete(struct sk_buff *skb) 2853{ 2854 struct tcphdr *th = tcp_hdr(skb); 2855 2856 skb->csum_start = skb_transport_header(skb) - skb->head; 2857 skb->csum_offset = offsetof(struct tcphdr, check); 2858 skb->ip_summed = CHECKSUM_PARTIAL; 2859 2860 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count; 2861 2862 if (th->cwr) 2863 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 2864 2865 return 0; 2866} 2867EXPORT_SYMBOL(tcp_gro_complete); 2868 2869#ifdef CONFIG_TCP_MD5SIG 2870static unsigned long tcp_md5sig_users; 2871static struct tcp_md5sig_pool * __percpu *tcp_md5sig_pool; 2872static DEFINE_SPINLOCK(tcp_md5sig_pool_lock); 2873 2874static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool * __percpu *pool) 2875{ 2876 int cpu; 2877 for_each_possible_cpu(cpu) { 2878 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu); 2879 if (p) { 2880 if (p->md5_desc.tfm) 2881 crypto_free_hash(p->md5_desc.tfm); 2882 kfree(p); 2883 } 2884 } 2885 free_percpu(pool); 2886} 2887 2888void tcp_free_md5sig_pool(void) 2889{ 2890 struct tcp_md5sig_pool * __percpu *pool = NULL; 2891 2892 spin_lock_bh(&tcp_md5sig_pool_lock); 2893 if (--tcp_md5sig_users == 0) { 2894 pool = tcp_md5sig_pool; 2895 tcp_md5sig_pool = NULL; 2896 } 2897 spin_unlock_bh(&tcp_md5sig_pool_lock); 2898 if (pool) 2899 __tcp_free_md5sig_pool(pool); 2900} 2901EXPORT_SYMBOL(tcp_free_md5sig_pool); 2902 2903static struct tcp_md5sig_pool * __percpu * 2904__tcp_alloc_md5sig_pool(struct sock *sk) 2905{ 2906 int cpu; 2907 struct tcp_md5sig_pool * __percpu *pool; 2908 2909 pool = alloc_percpu(struct tcp_md5sig_pool *); 2910 if (!pool) 2911 return NULL; 2912 2913 for_each_possible_cpu(cpu) { 2914 struct tcp_md5sig_pool *p; 2915 struct crypto_hash *hash; 2916 2917 p = kzalloc(sizeof(*p), sk->sk_allocation); 2918 if (!p) 2919 goto out_free; 2920 *per_cpu_ptr(pool, cpu) = p; 2921 2922 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC); 2923 if (!hash || IS_ERR(hash)) 2924 goto out_free; 2925 2926 p->md5_desc.tfm = hash; 2927 } 2928 return pool; 2929out_free: 2930 __tcp_free_md5sig_pool(pool); 2931 return NULL; 2932} 2933 2934struct tcp_md5sig_pool * __percpu *tcp_alloc_md5sig_pool(struct sock *sk) 2935{ 2936 struct tcp_md5sig_pool * __percpu *pool; 2937 int alloc = 0; 2938 2939retry: 2940 spin_lock_bh(&tcp_md5sig_pool_lock); 2941 pool = tcp_md5sig_pool; 2942 if (tcp_md5sig_users++ == 0) { 2943 alloc = 1; 2944 spin_unlock_bh(&tcp_md5sig_pool_lock); 2945 } else if (!pool) { 2946 tcp_md5sig_users--; 2947 spin_unlock_bh(&tcp_md5sig_pool_lock); 2948 cpu_relax(); 2949 goto retry; 2950 } else 2951 spin_unlock_bh(&tcp_md5sig_pool_lock); 2952 2953 if (alloc) { 2954 /* we cannot hold spinlock here because this may sleep. */ 2955 struct tcp_md5sig_pool * __percpu *p; 2956 2957 p = __tcp_alloc_md5sig_pool(sk); 2958 spin_lock_bh(&tcp_md5sig_pool_lock); 2959 if (!p) { 2960 tcp_md5sig_users--; 2961 spin_unlock_bh(&tcp_md5sig_pool_lock); 2962 return NULL; 2963 } 2964 pool = tcp_md5sig_pool; 2965 if (pool) { 2966 /* oops, it has already been assigned. */ 2967 spin_unlock_bh(&tcp_md5sig_pool_lock); 2968 __tcp_free_md5sig_pool(p); 2969 } else { 2970 tcp_md5sig_pool = pool = p; 2971 spin_unlock_bh(&tcp_md5sig_pool_lock); 2972 } 2973 } 2974 return pool; 2975} 2976EXPORT_SYMBOL(tcp_alloc_md5sig_pool); 2977 2978 2979/** 2980 * tcp_get_md5sig_pool - get md5sig_pool for this user 2981 * 2982 * We use percpu structure, so if we succeed, we exit with preemption 2983 * and BH disabled, to make sure another thread or softirq handling 2984 * wont try to get same context. 2985 */ 2986struct tcp_md5sig_pool *tcp_get_md5sig_pool(void) 2987{ 2988 struct tcp_md5sig_pool * __percpu *p; 2989 2990 local_bh_disable(); 2991 2992 spin_lock(&tcp_md5sig_pool_lock); 2993 p = tcp_md5sig_pool; 2994 if (p) 2995 tcp_md5sig_users++; 2996 spin_unlock(&tcp_md5sig_pool_lock); 2997 2998 if (p) 2999 return *this_cpu_ptr(p); 3000 3001 local_bh_enable(); 3002 return NULL; 3003} 3004EXPORT_SYMBOL(tcp_get_md5sig_pool); 3005 3006void tcp_put_md5sig_pool(void) 3007{ 3008 local_bh_enable(); 3009 tcp_free_md5sig_pool(); 3010} 3011EXPORT_SYMBOL(tcp_put_md5sig_pool); 3012 3013int tcp_md5_hash_header(struct tcp_md5sig_pool *hp, 3014 struct tcphdr *th) 3015{ 3016 struct scatterlist sg; 3017 int err; 3018 3019 __sum16 old_checksum = th->check; 3020 th->check = 0; 3021 /* options aren't included in the hash */ 3022 sg_init_one(&sg, th, sizeof(struct tcphdr)); 3023 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr)); 3024 th->check = old_checksum; 3025 return err; 3026} 3027EXPORT_SYMBOL(tcp_md5_hash_header); 3028 3029int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp, 3030 struct sk_buff *skb, unsigned header_len) 3031{ 3032 struct scatterlist sg; 3033 const struct tcphdr *tp = tcp_hdr(skb); 3034 struct hash_desc *desc = &hp->md5_desc; 3035 unsigned i; 3036 const unsigned head_data_len = skb_headlen(skb) > header_len ? 3037 skb_headlen(skb) - header_len : 0; 3038 const struct skb_shared_info *shi = skb_shinfo(skb); 3039 struct sk_buff *frag_iter; 3040 3041 sg_init_table(&sg, 1); 3042 3043 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len); 3044 if (crypto_hash_update(desc, &sg, head_data_len)) 3045 return 1; 3046 3047 for (i = 0; i < shi->nr_frags; ++i) { 3048 const struct skb_frag_struct *f = &shi->frags[i]; 3049 sg_set_page(&sg, f->page, f->size, f->page_offset); 3050 if (crypto_hash_update(desc, &sg, f->size)) 3051 return 1; 3052 } 3053 3054 skb_walk_frags(skb, frag_iter) 3055 if (tcp_md5_hash_skb_data(hp, frag_iter, 0)) 3056 return 1; 3057 3058 return 0; 3059} 3060EXPORT_SYMBOL(tcp_md5_hash_skb_data); 3061 3062int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key) 3063{ 3064 struct scatterlist sg; 3065 3066 sg_init_one(&sg, key->key, key->keylen); 3067 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen); 3068} 3069EXPORT_SYMBOL(tcp_md5_hash_key); 3070 3071#endif 3072 3073/** 3074 * Each Responder maintains up to two secret values concurrently for 3075 * efficient secret rollover. Each secret value has 4 states: 3076 * 3077 * Generating. (tcp_secret_generating != tcp_secret_primary) 3078 * Generates new Responder-Cookies, but not yet used for primary 3079 * verification. This is a short-term state, typically lasting only 3080 * one round trip time (RTT). 3081 * 3082 * Primary. (tcp_secret_generating == tcp_secret_primary) 3083 * Used both for generation and primary verification. 3084 * 3085 * Retiring. (tcp_secret_retiring != tcp_secret_secondary) 3086 * Used for verification, until the first failure that can be 3087 * verified by the newer Generating secret. At that time, this 3088 * cookie's state is changed to Secondary, and the Generating 3089 * cookie's state is changed to Primary. This is a short-term state, 3090 * typically lasting only one round trip time (RTT). 3091 * 3092 * Secondary. (tcp_secret_retiring == tcp_secret_secondary) 3093 * Used for secondary verification, after primary verification 3094 * failures. This state lasts no more than twice the Maximum Segment 3095 * Lifetime (2MSL). Then, the secret is discarded. 3096 */ 3097struct tcp_cookie_secret { 3098 /* The secret is divided into two parts. The digest part is the 3099 * equivalent of previously hashing a secret and saving the state, 3100 * and serves as an initialization vector (IV). The message part 3101 * serves as the trailing secret. 3102 */ 3103 u32 secrets[COOKIE_WORKSPACE_WORDS]; 3104 unsigned long expires; 3105}; 3106 3107#define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL) 3108#define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2) 3109#define TCP_SECRET_LIFE (HZ * 600) 3110 3111static struct tcp_cookie_secret tcp_secret_one; 3112static struct tcp_cookie_secret tcp_secret_two; 3113 3114/* Essentially a circular list, without dynamic allocation. */ 3115static struct tcp_cookie_secret *tcp_secret_generating; 3116static struct tcp_cookie_secret *tcp_secret_primary; 3117static struct tcp_cookie_secret *tcp_secret_retiring; 3118static struct tcp_cookie_secret *tcp_secret_secondary; 3119 3120static DEFINE_SPINLOCK(tcp_secret_locker); 3121 3122/* Select a pseudo-random word in the cookie workspace. 3123 */ 3124static inline u32 tcp_cookie_work(const u32 *ws, const int n) 3125{ 3126 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])]; 3127} 3128 3129/* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed. 3130 * Called in softirq context. 3131 * Returns: 0 for success. 3132 */ 3133int tcp_cookie_generator(u32 *bakery) 3134{ 3135 unsigned long jiffy = jiffies; 3136 3137 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) { 3138 spin_lock_bh(&tcp_secret_locker); 3139 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) { 3140 /* refreshed by another */ 3141 memcpy(bakery, 3142 &tcp_secret_generating->secrets[0], 3143 COOKIE_WORKSPACE_WORDS); 3144 } else { 3145 /* still needs refreshing */ 3146 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS); 3147 3148 /* The first time, paranoia assumes that the 3149 * randomization function isn't as strong. But, 3150 * this secret initialization is delayed until 3151 * the last possible moment (packet arrival). 3152 * Although that time is observable, it is 3153 * unpredictably variable. Mash in the most 3154 * volatile clock bits available, and expire the 3155 * secret extra quickly. 3156 */ 3157 if (unlikely(tcp_secret_primary->expires == 3158 tcp_secret_secondary->expires)) { 3159 struct timespec tv; 3160 3161 getnstimeofday(&tv); 3162 bakery[COOKIE_DIGEST_WORDS+0] ^= 3163 (u32)tv.tv_nsec; 3164 3165 tcp_secret_secondary->expires = jiffy 3166 + TCP_SECRET_1MSL 3167 + (0x0f & tcp_cookie_work(bakery, 0)); 3168 } else { 3169 tcp_secret_secondary->expires = jiffy 3170 + TCP_SECRET_LIFE 3171 + (0xff & tcp_cookie_work(bakery, 1)); 3172 tcp_secret_primary->expires = jiffy 3173 + TCP_SECRET_2MSL 3174 + (0x1f & tcp_cookie_work(bakery, 2)); 3175 } 3176 memcpy(&tcp_secret_secondary->secrets[0], 3177 bakery, COOKIE_WORKSPACE_WORDS); 3178 3179 rcu_assign_pointer(tcp_secret_generating, 3180 tcp_secret_secondary); 3181 rcu_assign_pointer(tcp_secret_retiring, 3182 tcp_secret_primary); 3183 /* 3184 * Neither call_rcu() nor synchronize_rcu() needed. 3185 * Retiring data is not freed. It is replaced after 3186 * further (locked) pointer updates, and a quiet time 3187 * (minimum 1MSL, maximum LIFE - 2MSL). 3188 */ 3189 } 3190 spin_unlock_bh(&tcp_secret_locker); 3191 } else { 3192 rcu_read_lock_bh(); 3193 memcpy(bakery, 3194 &rcu_dereference(tcp_secret_generating)->secrets[0], 3195 COOKIE_WORKSPACE_WORDS); 3196 rcu_read_unlock_bh(); 3197 } 3198 return 0; 3199} 3200EXPORT_SYMBOL(tcp_cookie_generator); 3201 3202void tcp_done(struct sock *sk) 3203{ 3204 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV) 3205 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 3206 3207 tcp_set_state(sk, TCP_CLOSE); 3208 tcp_clear_xmit_timers(sk); 3209 3210 sk->sk_shutdown = SHUTDOWN_MASK; 3211 3212 if (!sock_flag(sk, SOCK_DEAD)) 3213 sk->sk_state_change(sk); 3214 else 3215 inet_csk_destroy_sock(sk); 3216} 3217EXPORT_SYMBOL_GPL(tcp_done); 3218 3219extern struct tcp_congestion_ops tcp_reno; 3220 3221static __initdata unsigned long thash_entries; 3222static int __init set_thash_entries(char *str) 3223{ 3224 if (!str) 3225 return 0; 3226 thash_entries = simple_strtoul(str, &str, 0); 3227 return 1; 3228} 3229__setup("thash_entries=", set_thash_entries); 3230 3231void __init tcp_init(void) 3232{ 3233 struct sk_buff *skb = NULL; 3234 unsigned long nr_pages, limit; 3235 int i, max_share, cnt; 3236 unsigned long jiffy = jiffies; 3237 3238 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb)); 3239 3240 percpu_counter_init(&tcp_sockets_allocated, 0); 3241 percpu_counter_init(&tcp_orphan_count, 0); 3242 tcp_hashinfo.bind_bucket_cachep = 3243 kmem_cache_create("tcp_bind_bucket", 3244 sizeof(struct inet_bind_bucket), 0, 3245 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL); 3246 3247 /* Size and allocate the main established and bind bucket 3248 * hash tables. 3249 * 3250 * The methodology is similar to that of the buffer cache. 3251 */ 3252 tcp_hashinfo.ehash = 3253 alloc_large_system_hash("TCP established", 3254 sizeof(struct inet_ehash_bucket), 3255 thash_entries, 3256 (totalram_pages >= 128 * 1024) ? 3257 13 : 15, 3258 0, 3259 NULL, 3260 &tcp_hashinfo.ehash_mask, 3261 thash_entries ? 0 : 512 * 1024); 3262 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) { 3263 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i); 3264 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i); 3265 } 3266 if (inet_ehash_locks_alloc(&tcp_hashinfo)) 3267 panic("TCP: failed to alloc ehash_locks"); 3268 tcp_hashinfo.bhash = 3269 alloc_large_system_hash("TCP bind", 3270 sizeof(struct inet_bind_hashbucket), 3271 tcp_hashinfo.ehash_mask + 1, 3272 (totalram_pages >= 128 * 1024) ? 3273 13 : 15, 3274 0, 3275 &tcp_hashinfo.bhash_size, 3276 NULL, 3277 64 * 1024); 3278 tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size; 3279 for (i = 0; i < tcp_hashinfo.bhash_size; i++) { 3280 spin_lock_init(&tcp_hashinfo.bhash[i].lock); 3281 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain); 3282 } 3283 3284 3285 cnt = tcp_hashinfo.ehash_mask + 1; 3286 3287 tcp_death_row.sysctl_max_tw_buckets = cnt / 2; 3288 sysctl_tcp_max_orphans = cnt / 2; 3289 sysctl_max_syn_backlog = max(128, cnt / 256); 3290 3291 /* Set the pressure threshold to be a fraction of global memory that 3292 * is up to 1/2 at 256 MB, decreasing toward zero with the amount of 3293 * memory, with a floor of 128 pages. 3294 */ 3295 nr_pages = totalram_pages - totalhigh_pages; 3296 limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT); 3297 limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11); 3298 limit = max(limit, 128UL); 3299 sysctl_tcp_mem[0] = limit / 4 * 3; 3300 sysctl_tcp_mem[1] = limit; 3301 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; 3302 3303 /* Set per-socket limits to no more than 1/128 the pressure threshold */ 3304 limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7); 3305 max_share = min(4UL*1024*1024, limit); 3306 3307 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM; 3308 sysctl_tcp_wmem[1] = 16*1024; 3309 sysctl_tcp_wmem[2] = max(64*1024, max_share); 3310 3311 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM; 3312 sysctl_tcp_rmem[1] = 87380; 3313 sysctl_tcp_rmem[2] = max(87380, max_share); 3314 3315 printk(KERN_INFO "TCP: Hash tables configured " 3316 "(established %u bind %u)\n", 3317 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size); 3318 3319 tcp_register_congestion_control(&tcp_reno); 3320 3321 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets)); 3322 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets)); 3323 tcp_secret_one.expires = jiffy; /* past due */ 3324 tcp_secret_two.expires = jiffy; /* past due */ 3325 tcp_secret_generating = &tcp_secret_one; 3326 tcp_secret_primary = &tcp_secret_one; 3327 tcp_secret_retiring = &tcp_secret_two; 3328 tcp_secret_secondary = &tcp_secret_two; 3329} 3330