tcp.c revision 38ba0a65faf451dd46c7860b4fade84c0b8e444f
1/* 2 * INET An implementation of the TCP/IP protocol suite for the LINUX 3 * operating system. INET is implemented using the BSD Socket 4 * interface as the means of communication with the user level. 5 * 6 * Implementation of the Transmission Control Protocol(TCP). 7 * 8 * Authors: Ross Biro 9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 10 * Mark Evans, <evansmp@uhura.aston.ac.uk> 11 * Corey Minyard <wf-rch!minyard@relay.EU.net> 12 * Florian La Roche, <flla@stud.uni-sb.de> 13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 14 * Linus Torvalds, <torvalds@cs.helsinki.fi> 15 * Alan Cox, <gw4pts@gw4pts.ampr.org> 16 * Matthew Dillon, <dillon@apollo.west.oic.com> 17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 18 * Jorge Cwik, <jorge@laser.satlink.net> 19 * 20 * Fixes: 21 * Alan Cox : Numerous verify_area() calls 22 * Alan Cox : Set the ACK bit on a reset 23 * Alan Cox : Stopped it crashing if it closed while 24 * sk->inuse=1 and was trying to connect 25 * (tcp_err()). 26 * Alan Cox : All icmp error handling was broken 27 * pointers passed where wrong and the 28 * socket was looked up backwards. Nobody 29 * tested any icmp error code obviously. 30 * Alan Cox : tcp_err() now handled properly. It 31 * wakes people on errors. poll 32 * behaves and the icmp error race 33 * has gone by moving it into sock.c 34 * Alan Cox : tcp_send_reset() fixed to work for 35 * everything not just packets for 36 * unknown sockets. 37 * Alan Cox : tcp option processing. 38 * Alan Cox : Reset tweaked (still not 100%) [Had 39 * syn rule wrong] 40 * Herp Rosmanith : More reset fixes 41 * Alan Cox : No longer acks invalid rst frames. 42 * Acking any kind of RST is right out. 43 * Alan Cox : Sets an ignore me flag on an rst 44 * receive otherwise odd bits of prattle 45 * escape still 46 * Alan Cox : Fixed another acking RST frame bug. 47 * Should stop LAN workplace lockups. 48 * Alan Cox : Some tidyups using the new skb list 49 * facilities 50 * Alan Cox : sk->keepopen now seems to work 51 * Alan Cox : Pulls options out correctly on accepts 52 * Alan Cox : Fixed assorted sk->rqueue->next errors 53 * Alan Cox : PSH doesn't end a TCP read. Switched a 54 * bit to skb ops. 55 * Alan Cox : Tidied tcp_data to avoid a potential 56 * nasty. 57 * Alan Cox : Added some better commenting, as the 58 * tcp is hard to follow 59 * Alan Cox : Removed incorrect check for 20 * psh 60 * Michael O'Reilly : ack < copied bug fix. 61 * Johannes Stille : Misc tcp fixes (not all in yet). 62 * Alan Cox : FIN with no memory -> CRASH 63 * Alan Cox : Added socket option proto entries. 64 * Also added awareness of them to accept. 65 * Alan Cox : Added TCP options (SOL_TCP) 66 * Alan Cox : Switched wakeup calls to callbacks, 67 * so the kernel can layer network 68 * sockets. 69 * Alan Cox : Use ip_tos/ip_ttl settings. 70 * Alan Cox : Handle FIN (more) properly (we hope). 71 * Alan Cox : RST frames sent on unsynchronised 72 * state ack error. 73 * Alan Cox : Put in missing check for SYN bit. 74 * Alan Cox : Added tcp_select_window() aka NET2E 75 * window non shrink trick. 76 * Alan Cox : Added a couple of small NET2E timer 77 * fixes 78 * Charles Hedrick : TCP fixes 79 * Toomas Tamm : TCP window fixes 80 * Alan Cox : Small URG fix to rlogin ^C ack fight 81 * Charles Hedrick : Rewrote most of it to actually work 82 * Linus : Rewrote tcp_read() and URG handling 83 * completely 84 * Gerhard Koerting: Fixed some missing timer handling 85 * Matthew Dillon : Reworked TCP machine states as per RFC 86 * Gerhard Koerting: PC/TCP workarounds 87 * Adam Caldwell : Assorted timer/timing errors 88 * Matthew Dillon : Fixed another RST bug 89 * Alan Cox : Move to kernel side addressing changes. 90 * Alan Cox : Beginning work on TCP fastpathing 91 * (not yet usable) 92 * Arnt Gulbrandsen: Turbocharged tcp_check() routine. 93 * Alan Cox : TCP fast path debugging 94 * Alan Cox : Window clamping 95 * Michael Riepe : Bug in tcp_check() 96 * Matt Dillon : More TCP improvements and RST bug fixes 97 * Matt Dillon : Yet more small nasties remove from the 98 * TCP code (Be very nice to this man if 99 * tcp finally works 100%) 8) 100 * Alan Cox : BSD accept semantics. 101 * Alan Cox : Reset on closedown bug. 102 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto(). 103 * Michael Pall : Handle poll() after URG properly in 104 * all cases. 105 * Michael Pall : Undo the last fix in tcp_read_urg() 106 * (multi URG PUSH broke rlogin). 107 * Michael Pall : Fix the multi URG PUSH problem in 108 * tcp_readable(), poll() after URG 109 * works now. 110 * Michael Pall : recv(...,MSG_OOB) never blocks in the 111 * BSD api. 112 * Alan Cox : Changed the semantics of sk->socket to 113 * fix a race and a signal problem with 114 * accept() and async I/O. 115 * Alan Cox : Relaxed the rules on tcp_sendto(). 116 * Yury Shevchuk : Really fixed accept() blocking problem. 117 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for 118 * clients/servers which listen in on 119 * fixed ports. 120 * Alan Cox : Cleaned the above up and shrank it to 121 * a sensible code size. 122 * Alan Cox : Self connect lockup fix. 123 * Alan Cox : No connect to multicast. 124 * Ross Biro : Close unaccepted children on master 125 * socket close. 126 * Alan Cox : Reset tracing code. 127 * Alan Cox : Spurious resets on shutdown. 128 * Alan Cox : Giant 15 minute/60 second timer error 129 * Alan Cox : Small whoops in polling before an 130 * accept. 131 * Alan Cox : Kept the state trace facility since 132 * it's handy for debugging. 133 * Alan Cox : More reset handler fixes. 134 * Alan Cox : Started rewriting the code based on 135 * the RFC's for other useful protocol 136 * references see: Comer, KA9Q NOS, and 137 * for a reference on the difference 138 * between specifications and how BSD 139 * works see the 4.4lite source. 140 * A.N.Kuznetsov : Don't time wait on completion of tidy 141 * close. 142 * Linus Torvalds : Fin/Shutdown & copied_seq changes. 143 * Linus Torvalds : Fixed BSD port reuse to work first syn 144 * Alan Cox : Reimplemented timers as per the RFC 145 * and using multiple timers for sanity. 146 * Alan Cox : Small bug fixes, and a lot of new 147 * comments. 148 * Alan Cox : Fixed dual reader crash by locking 149 * the buffers (much like datagram.c) 150 * Alan Cox : Fixed stuck sockets in probe. A probe 151 * now gets fed up of retrying without 152 * (even a no space) answer. 153 * Alan Cox : Extracted closing code better 154 * Alan Cox : Fixed the closing state machine to 155 * resemble the RFC. 156 * Alan Cox : More 'per spec' fixes. 157 * Jorge Cwik : Even faster checksumming. 158 * Alan Cox : tcp_data() doesn't ack illegal PSH 159 * only frames. At least one pc tcp stack 160 * generates them. 161 * Alan Cox : Cache last socket. 162 * Alan Cox : Per route irtt. 163 * Matt Day : poll()->select() match BSD precisely on error 164 * Alan Cox : New buffers 165 * Marc Tamsky : Various sk->prot->retransmits and 166 * sk->retransmits misupdating fixed. 167 * Fixed tcp_write_timeout: stuck close, 168 * and TCP syn retries gets used now. 169 * Mark Yarvis : In tcp_read_wakeup(), don't send an 170 * ack if state is TCP_CLOSED. 171 * Alan Cox : Look up device on a retransmit - routes may 172 * change. Doesn't yet cope with MSS shrink right 173 * but it's a start! 174 * Marc Tamsky : Closing in closing fixes. 175 * Mike Shaver : RFC1122 verifications. 176 * Alan Cox : rcv_saddr errors. 177 * Alan Cox : Block double connect(). 178 * Alan Cox : Small hooks for enSKIP. 179 * Alexey Kuznetsov: Path MTU discovery. 180 * Alan Cox : Support soft errors. 181 * Alan Cox : Fix MTU discovery pathological case 182 * when the remote claims no mtu! 183 * Marc Tamsky : TCP_CLOSE fix. 184 * Colin (G3TNE) : Send a reset on syn ack replies in 185 * window but wrong (fixes NT lpd problems) 186 * Pedro Roque : Better TCP window handling, delayed ack. 187 * Joerg Reuter : No modification of locked buffers in 188 * tcp_do_retransmit() 189 * Eric Schenk : Changed receiver side silly window 190 * avoidance algorithm to BSD style 191 * algorithm. This doubles throughput 192 * against machines running Solaris, 193 * and seems to result in general 194 * improvement. 195 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD 196 * Willy Konynenberg : Transparent proxying support. 197 * Mike McLagan : Routing by source 198 * Keith Owens : Do proper merging with partial SKB's in 199 * tcp_do_sendmsg to avoid burstiness. 200 * Eric Schenk : Fix fast close down bug with 201 * shutdown() followed by close(). 202 * Andi Kleen : Make poll agree with SIGIO 203 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and 204 * lingertime == 0 (RFC 793 ABORT Call) 205 * Hirokazu Takahashi : Use copy_from_user() instead of 206 * csum_and_copy_from_user() if possible. 207 * 208 * This program is free software; you can redistribute it and/or 209 * modify it under the terms of the GNU General Public License 210 * as published by the Free Software Foundation; either version 211 * 2 of the License, or(at your option) any later version. 212 * 213 * Description of States: 214 * 215 * TCP_SYN_SENT sent a connection request, waiting for ack 216 * 217 * TCP_SYN_RECV received a connection request, sent ack, 218 * waiting for final ack in three-way handshake. 219 * 220 * TCP_ESTABLISHED connection established 221 * 222 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete 223 * transmission of remaining buffered data 224 * 225 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote 226 * to shutdown 227 * 228 * TCP_CLOSING both sides have shutdown but we still have 229 * data we have to finish sending 230 * 231 * TCP_TIME_WAIT timeout to catch resent junk before entering 232 * closed, can only be entered from FIN_WAIT2 233 * or CLOSING. Required because the other end 234 * may not have gotten our last ACK causing it 235 * to retransmit the data packet (which we ignore) 236 * 237 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for 238 * us to finish writing our data and to shutdown 239 * (we have to close() to move on to LAST_ACK) 240 * 241 * TCP_LAST_ACK out side has shutdown after remote has 242 * shutdown. There may still be data in our 243 * buffer that we have to finish sending 244 * 245 * TCP_CLOSE socket is finished 246 */ 247 248#define pr_fmt(fmt) "TCP: " fmt 249 250#include <linux/kernel.h> 251#include <linux/module.h> 252#include <linux/types.h> 253#include <linux/fcntl.h> 254#include <linux/poll.h> 255#include <linux/init.h> 256#include <linux/fs.h> 257#include <linux/skbuff.h> 258#include <linux/scatterlist.h> 259#include <linux/splice.h> 260#include <linux/net.h> 261#include <linux/socket.h> 262#include <linux/random.h> 263#include <linux/bootmem.h> 264#include <linux/highmem.h> 265#include <linux/swap.h> 266#include <linux/cache.h> 267#include <linux/err.h> 268#include <linux/crypto.h> 269#include <linux/time.h> 270#include <linux/slab.h> 271 272#include <net/icmp.h> 273#include <net/tcp.h> 274#include <net/xfrm.h> 275#include <net/ip.h> 276#include <net/netdma.h> 277#include <net/sock.h> 278 279#include <asm/uaccess.h> 280#include <asm/ioctls.h> 281 282int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT; 283 284struct percpu_counter tcp_orphan_count; 285EXPORT_SYMBOL_GPL(tcp_orphan_count); 286 287int sysctl_tcp_wmem[3] __read_mostly; 288int sysctl_tcp_rmem[3] __read_mostly; 289 290EXPORT_SYMBOL(sysctl_tcp_rmem); 291EXPORT_SYMBOL(sysctl_tcp_wmem); 292 293atomic_long_t tcp_memory_allocated; /* Current allocated memory. */ 294EXPORT_SYMBOL(tcp_memory_allocated); 295 296/* 297 * Current number of TCP sockets. 298 */ 299struct percpu_counter tcp_sockets_allocated; 300EXPORT_SYMBOL(tcp_sockets_allocated); 301 302/* 303 * TCP splice context 304 */ 305struct tcp_splice_state { 306 struct pipe_inode_info *pipe; 307 size_t len; 308 unsigned int flags; 309}; 310 311/* 312 * Pressure flag: try to collapse. 313 * Technical note: it is used by multiple contexts non atomically. 314 * All the __sk_mem_schedule() is of this nature: accounting 315 * is strict, actions are advisory and have some latency. 316 */ 317int tcp_memory_pressure __read_mostly; 318EXPORT_SYMBOL(tcp_memory_pressure); 319 320void tcp_enter_memory_pressure(struct sock *sk) 321{ 322 if (!tcp_memory_pressure) { 323 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES); 324 tcp_memory_pressure = 1; 325 } 326} 327EXPORT_SYMBOL(tcp_enter_memory_pressure); 328 329/* Convert seconds to retransmits based on initial and max timeout */ 330static u8 secs_to_retrans(int seconds, int timeout, int rto_max) 331{ 332 u8 res = 0; 333 334 if (seconds > 0) { 335 int period = timeout; 336 337 res = 1; 338 while (seconds > period && res < 255) { 339 res++; 340 timeout <<= 1; 341 if (timeout > rto_max) 342 timeout = rto_max; 343 period += timeout; 344 } 345 } 346 return res; 347} 348 349/* Convert retransmits to seconds based on initial and max timeout */ 350static int retrans_to_secs(u8 retrans, int timeout, int rto_max) 351{ 352 int period = 0; 353 354 if (retrans > 0) { 355 period = timeout; 356 while (--retrans) { 357 timeout <<= 1; 358 if (timeout > rto_max) 359 timeout = rto_max; 360 period += timeout; 361 } 362 } 363 return period; 364} 365 366/* Address-family independent initialization for a tcp_sock. 367 * 368 * NOTE: A lot of things set to zero explicitly by call to 369 * sk_alloc() so need not be done here. 370 */ 371void tcp_init_sock(struct sock *sk) 372{ 373 struct inet_connection_sock *icsk = inet_csk(sk); 374 struct tcp_sock *tp = tcp_sk(sk); 375 376 skb_queue_head_init(&tp->out_of_order_queue); 377 tcp_init_xmit_timers(sk); 378 tcp_prequeue_init(tp); 379 380 icsk->icsk_rto = TCP_TIMEOUT_INIT; 381 tp->mdev = TCP_TIMEOUT_INIT; 382 383 /* So many TCP implementations out there (incorrectly) count the 384 * initial SYN frame in their delayed-ACK and congestion control 385 * algorithms that we must have the following bandaid to talk 386 * efficiently to them. -DaveM 387 */ 388 tp->snd_cwnd = TCP_INIT_CWND; 389 390 /* See draft-stevens-tcpca-spec-01 for discussion of the 391 * initialization of these values. 392 */ 393 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 394 tp->snd_cwnd_clamp = ~0; 395 tp->mss_cache = TCP_MSS_DEFAULT; 396 397 tp->reordering = sysctl_tcp_reordering; 398 icsk->icsk_ca_ops = &tcp_init_congestion_ops; 399 400 sk->sk_state = TCP_CLOSE; 401 402 sk->sk_write_space = sk_stream_write_space; 403 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE); 404 405 icsk->icsk_sync_mss = tcp_sync_mss; 406 407 /* TCP Cookie Transactions */ 408 if (sysctl_tcp_cookie_size > 0) { 409 /* Default, cookies without s_data_payload. */ 410 tp->cookie_values = 411 kzalloc(sizeof(*tp->cookie_values), 412 sk->sk_allocation); 413 if (tp->cookie_values != NULL) 414 kref_init(&tp->cookie_values->kref); 415 } 416 /* Presumed zeroed, in order of appearance: 417 * cookie_in_always, cookie_out_never, 418 * s_data_constant, s_data_in, s_data_out 419 */ 420 sk->sk_sndbuf = sysctl_tcp_wmem[1]; 421 sk->sk_rcvbuf = sysctl_tcp_rmem[1]; 422 423 local_bh_disable(); 424 sock_update_memcg(sk); 425 sk_sockets_allocated_inc(sk); 426 local_bh_enable(); 427} 428EXPORT_SYMBOL(tcp_init_sock); 429 430/* 431 * Wait for a TCP event. 432 * 433 * Note that we don't need to lock the socket, as the upper poll layers 434 * take care of normal races (between the test and the event) and we don't 435 * go look at any of the socket buffers directly. 436 */ 437unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait) 438{ 439 unsigned int mask; 440 struct sock *sk = sock->sk; 441 const struct tcp_sock *tp = tcp_sk(sk); 442 443 sock_poll_wait(file, sk_sleep(sk), wait); 444 if (sk->sk_state == TCP_LISTEN) 445 return inet_csk_listen_poll(sk); 446 447 /* Socket is not locked. We are protected from async events 448 * by poll logic and correct handling of state changes 449 * made by other threads is impossible in any case. 450 */ 451 452 mask = 0; 453 454 /* 455 * POLLHUP is certainly not done right. But poll() doesn't 456 * have a notion of HUP in just one direction, and for a 457 * socket the read side is more interesting. 458 * 459 * Some poll() documentation says that POLLHUP is incompatible 460 * with the POLLOUT/POLLWR flags, so somebody should check this 461 * all. But careful, it tends to be safer to return too many 462 * bits than too few, and you can easily break real applications 463 * if you don't tell them that something has hung up! 464 * 465 * Check-me. 466 * 467 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and 468 * our fs/select.c). It means that after we received EOF, 469 * poll always returns immediately, making impossible poll() on write() 470 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP 471 * if and only if shutdown has been made in both directions. 472 * Actually, it is interesting to look how Solaris and DUX 473 * solve this dilemma. I would prefer, if POLLHUP were maskable, 474 * then we could set it on SND_SHUTDOWN. BTW examples given 475 * in Stevens' books assume exactly this behaviour, it explains 476 * why POLLHUP is incompatible with POLLOUT. --ANK 477 * 478 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent 479 * blocking on fresh not-connected or disconnected socket. --ANK 480 */ 481 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE) 482 mask |= POLLHUP; 483 if (sk->sk_shutdown & RCV_SHUTDOWN) 484 mask |= POLLIN | POLLRDNORM | POLLRDHUP; 485 486 /* Connected? */ 487 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) { 488 int target = sock_rcvlowat(sk, 0, INT_MAX); 489 490 if (tp->urg_seq == tp->copied_seq && 491 !sock_flag(sk, SOCK_URGINLINE) && 492 tp->urg_data) 493 target++; 494 495 /* Potential race condition. If read of tp below will 496 * escape above sk->sk_state, we can be illegally awaken 497 * in SYN_* states. */ 498 if (tp->rcv_nxt - tp->copied_seq >= target) 499 mask |= POLLIN | POLLRDNORM; 500 501 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) { 502 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) { 503 mask |= POLLOUT | POLLWRNORM; 504 } else { /* send SIGIO later */ 505 set_bit(SOCK_ASYNC_NOSPACE, 506 &sk->sk_socket->flags); 507 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 508 509 /* Race breaker. If space is freed after 510 * wspace test but before the flags are set, 511 * IO signal will be lost. 512 */ 513 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) 514 mask |= POLLOUT | POLLWRNORM; 515 } 516 } else 517 mask |= POLLOUT | POLLWRNORM; 518 519 if (tp->urg_data & TCP_URG_VALID) 520 mask |= POLLPRI; 521 } 522 /* This barrier is coupled with smp_wmb() in tcp_reset() */ 523 smp_rmb(); 524 if (sk->sk_err) 525 mask |= POLLERR; 526 527 return mask; 528} 529EXPORT_SYMBOL(tcp_poll); 530 531int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg) 532{ 533 struct tcp_sock *tp = tcp_sk(sk); 534 int answ; 535 536 switch (cmd) { 537 case SIOCINQ: 538 if (sk->sk_state == TCP_LISTEN) 539 return -EINVAL; 540 541 lock_sock(sk); 542 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 543 answ = 0; 544 else if (sock_flag(sk, SOCK_URGINLINE) || 545 !tp->urg_data || 546 before(tp->urg_seq, tp->copied_seq) || 547 !before(tp->urg_seq, tp->rcv_nxt)) { 548 struct sk_buff *skb; 549 550 answ = tp->rcv_nxt - tp->copied_seq; 551 552 /* Subtract 1, if FIN is in queue. */ 553 skb = skb_peek_tail(&sk->sk_receive_queue); 554 if (answ && skb) 555 answ -= tcp_hdr(skb)->fin; 556 } else 557 answ = tp->urg_seq - tp->copied_seq; 558 release_sock(sk); 559 break; 560 case SIOCATMARK: 561 answ = tp->urg_data && tp->urg_seq == tp->copied_seq; 562 break; 563 case SIOCOUTQ: 564 if (sk->sk_state == TCP_LISTEN) 565 return -EINVAL; 566 567 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 568 answ = 0; 569 else 570 answ = tp->write_seq - tp->snd_una; 571 break; 572 case SIOCOUTQNSD: 573 if (sk->sk_state == TCP_LISTEN) 574 return -EINVAL; 575 576 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 577 answ = 0; 578 else 579 answ = tp->write_seq - tp->snd_nxt; 580 break; 581 default: 582 return -ENOIOCTLCMD; 583 } 584 585 return put_user(answ, (int __user *)arg); 586} 587EXPORT_SYMBOL(tcp_ioctl); 588 589static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb) 590{ 591 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 592 tp->pushed_seq = tp->write_seq; 593} 594 595static inline int forced_push(const struct tcp_sock *tp) 596{ 597 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1)); 598} 599 600static inline void skb_entail(struct sock *sk, struct sk_buff *skb) 601{ 602 struct tcp_sock *tp = tcp_sk(sk); 603 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 604 605 skb->csum = 0; 606 tcb->seq = tcb->end_seq = tp->write_seq; 607 tcb->tcp_flags = TCPHDR_ACK; 608 tcb->sacked = 0; 609 skb_header_release(skb); 610 tcp_add_write_queue_tail(sk, skb); 611 sk->sk_wmem_queued += skb->truesize; 612 sk_mem_charge(sk, skb->truesize); 613 if (tp->nonagle & TCP_NAGLE_PUSH) 614 tp->nonagle &= ~TCP_NAGLE_PUSH; 615} 616 617static inline void tcp_mark_urg(struct tcp_sock *tp, int flags) 618{ 619 if (flags & MSG_OOB) 620 tp->snd_up = tp->write_seq; 621} 622 623static inline void tcp_push(struct sock *sk, int flags, int mss_now, 624 int nonagle) 625{ 626 if (tcp_send_head(sk)) { 627 struct tcp_sock *tp = tcp_sk(sk); 628 629 if (!(flags & MSG_MORE) || forced_push(tp)) 630 tcp_mark_push(tp, tcp_write_queue_tail(sk)); 631 632 tcp_mark_urg(tp, flags); 633 __tcp_push_pending_frames(sk, mss_now, 634 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle); 635 } 636} 637 638static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, 639 unsigned int offset, size_t len) 640{ 641 struct tcp_splice_state *tss = rd_desc->arg.data; 642 int ret; 643 644 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len), 645 tss->flags); 646 if (ret > 0) 647 rd_desc->count -= ret; 648 return ret; 649} 650 651static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss) 652{ 653 /* Store TCP splice context information in read_descriptor_t. */ 654 read_descriptor_t rd_desc = { 655 .arg.data = tss, 656 .count = tss->len, 657 }; 658 659 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv); 660} 661 662/** 663 * tcp_splice_read - splice data from TCP socket to a pipe 664 * @sock: socket to splice from 665 * @ppos: position (not valid) 666 * @pipe: pipe to splice to 667 * @len: number of bytes to splice 668 * @flags: splice modifier flags 669 * 670 * Description: 671 * Will read pages from given socket and fill them into a pipe. 672 * 673 **/ 674ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos, 675 struct pipe_inode_info *pipe, size_t len, 676 unsigned int flags) 677{ 678 struct sock *sk = sock->sk; 679 struct tcp_splice_state tss = { 680 .pipe = pipe, 681 .len = len, 682 .flags = flags, 683 }; 684 long timeo; 685 ssize_t spliced; 686 int ret; 687 688 sock_rps_record_flow(sk); 689 /* 690 * We can't seek on a socket input 691 */ 692 if (unlikely(*ppos)) 693 return -ESPIPE; 694 695 ret = spliced = 0; 696 697 lock_sock(sk); 698 699 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK); 700 while (tss.len) { 701 ret = __tcp_splice_read(sk, &tss); 702 if (ret < 0) 703 break; 704 else if (!ret) { 705 if (spliced) 706 break; 707 if (sock_flag(sk, SOCK_DONE)) 708 break; 709 if (sk->sk_err) { 710 ret = sock_error(sk); 711 break; 712 } 713 if (sk->sk_shutdown & RCV_SHUTDOWN) 714 break; 715 if (sk->sk_state == TCP_CLOSE) { 716 /* 717 * This occurs when user tries to read 718 * from never connected socket. 719 */ 720 if (!sock_flag(sk, SOCK_DONE)) 721 ret = -ENOTCONN; 722 break; 723 } 724 if (!timeo) { 725 ret = -EAGAIN; 726 break; 727 } 728 sk_wait_data(sk, &timeo); 729 if (signal_pending(current)) { 730 ret = sock_intr_errno(timeo); 731 break; 732 } 733 continue; 734 } 735 tss.len -= ret; 736 spliced += ret; 737 738 if (!timeo) 739 break; 740 release_sock(sk); 741 lock_sock(sk); 742 743 if (sk->sk_err || sk->sk_state == TCP_CLOSE || 744 (sk->sk_shutdown & RCV_SHUTDOWN) || 745 signal_pending(current)) 746 break; 747 } 748 749 release_sock(sk); 750 751 if (spliced) 752 return spliced; 753 754 return ret; 755} 756EXPORT_SYMBOL(tcp_splice_read); 757 758struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp) 759{ 760 struct sk_buff *skb; 761 762 /* The TCP header must be at least 32-bit aligned. */ 763 size = ALIGN(size, 4); 764 765 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp); 766 if (skb) { 767 if (sk_wmem_schedule(sk, skb->truesize)) { 768 skb_reserve(skb, sk->sk_prot->max_header); 769 /* 770 * Make sure that we have exactly size bytes 771 * available to the caller, no more, no less. 772 */ 773 skb->avail_size = size; 774 return skb; 775 } 776 __kfree_skb(skb); 777 } else { 778 sk->sk_prot->enter_memory_pressure(sk); 779 sk_stream_moderate_sndbuf(sk); 780 } 781 return NULL; 782} 783 784static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now, 785 int large_allowed) 786{ 787 struct tcp_sock *tp = tcp_sk(sk); 788 u32 xmit_size_goal, old_size_goal; 789 790 xmit_size_goal = mss_now; 791 792 if (large_allowed && sk_can_gso(sk)) { 793 xmit_size_goal = ((sk->sk_gso_max_size - 1) - 794 inet_csk(sk)->icsk_af_ops->net_header_len - 795 inet_csk(sk)->icsk_ext_hdr_len - 796 tp->tcp_header_len); 797 798 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal); 799 800 /* We try hard to avoid divides here */ 801 old_size_goal = tp->xmit_size_goal_segs * mss_now; 802 803 if (likely(old_size_goal <= xmit_size_goal && 804 old_size_goal + mss_now > xmit_size_goal)) { 805 xmit_size_goal = old_size_goal; 806 } else { 807 tp->xmit_size_goal_segs = xmit_size_goal / mss_now; 808 xmit_size_goal = tp->xmit_size_goal_segs * mss_now; 809 } 810 } 811 812 return max(xmit_size_goal, mss_now); 813} 814 815static int tcp_send_mss(struct sock *sk, int *size_goal, int flags) 816{ 817 int mss_now; 818 819 mss_now = tcp_current_mss(sk); 820 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB)); 821 822 return mss_now; 823} 824 825static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset, 826 size_t psize, int flags) 827{ 828 struct tcp_sock *tp = tcp_sk(sk); 829 int mss_now, size_goal; 830 int err; 831 ssize_t copied; 832 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 833 834 /* Wait for a connection to finish. */ 835 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 836 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0) 837 goto out_err; 838 839 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags); 840 841 mss_now = tcp_send_mss(sk, &size_goal, flags); 842 copied = 0; 843 844 err = -EPIPE; 845 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 846 goto out_err; 847 848 while (psize > 0) { 849 struct sk_buff *skb = tcp_write_queue_tail(sk); 850 struct page *page = pages[poffset / PAGE_SIZE]; 851 int copy, i; 852 int offset = poffset % PAGE_SIZE; 853 int size = min_t(size_t, psize, PAGE_SIZE - offset); 854 bool can_coalesce; 855 856 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) { 857new_segment: 858 if (!sk_stream_memory_free(sk)) 859 goto wait_for_sndbuf; 860 861 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation); 862 if (!skb) 863 goto wait_for_memory; 864 865 skb_entail(sk, skb); 866 copy = size_goal; 867 } 868 869 if (copy > size) 870 copy = size; 871 872 i = skb_shinfo(skb)->nr_frags; 873 can_coalesce = skb_can_coalesce(skb, i, page, offset); 874 if (!can_coalesce && i >= MAX_SKB_FRAGS) { 875 tcp_mark_push(tp, skb); 876 goto new_segment; 877 } 878 if (!sk_wmem_schedule(sk, copy)) 879 goto wait_for_memory; 880 881 if (can_coalesce) { 882 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 883 } else { 884 get_page(page); 885 skb_fill_page_desc(skb, i, page, offset, copy); 886 } 887 888 skb->len += copy; 889 skb->data_len += copy; 890 skb->truesize += copy; 891 sk->sk_wmem_queued += copy; 892 sk_mem_charge(sk, copy); 893 skb->ip_summed = CHECKSUM_PARTIAL; 894 tp->write_seq += copy; 895 TCP_SKB_CB(skb)->end_seq += copy; 896 skb_shinfo(skb)->gso_segs = 0; 897 898 if (!copied) 899 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 900 901 copied += copy; 902 poffset += copy; 903 if (!(psize -= copy)) 904 goto out; 905 906 if (skb->len < size_goal || (flags & MSG_OOB)) 907 continue; 908 909 if (forced_push(tp)) { 910 tcp_mark_push(tp, skb); 911 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 912 } else if (skb == tcp_send_head(sk)) 913 tcp_push_one(sk, mss_now); 914 continue; 915 916wait_for_sndbuf: 917 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 918wait_for_memory: 919 if (copied) 920 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH); 921 922 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0) 923 goto do_error; 924 925 mss_now = tcp_send_mss(sk, &size_goal, flags); 926 } 927 928out: 929 if (copied && !(flags & MSG_SENDPAGE_NOTLAST)) 930 tcp_push(sk, flags, mss_now, tp->nonagle); 931 return copied; 932 933do_error: 934 if (copied) 935 goto out; 936out_err: 937 return sk_stream_error(sk, flags, err); 938} 939 940int tcp_sendpage(struct sock *sk, struct page *page, int offset, 941 size_t size, int flags) 942{ 943 ssize_t res; 944 945 if (!(sk->sk_route_caps & NETIF_F_SG) || 946 !(sk->sk_route_caps & NETIF_F_ALL_CSUM)) 947 return sock_no_sendpage(sk->sk_socket, page, offset, size, 948 flags); 949 950 lock_sock(sk); 951 res = do_tcp_sendpages(sk, &page, offset, size, flags); 952 release_sock(sk); 953 return res; 954} 955EXPORT_SYMBOL(tcp_sendpage); 956 957static inline int select_size(const struct sock *sk, bool sg) 958{ 959 const struct tcp_sock *tp = tcp_sk(sk); 960 int tmp = tp->mss_cache; 961 962 if (sg) { 963 if (sk_can_gso(sk)) { 964 /* Small frames wont use a full page: 965 * Payload will immediately follow tcp header. 966 */ 967 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER); 968 } else { 969 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER); 970 971 if (tmp >= pgbreak && 972 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE) 973 tmp = pgbreak; 974 } 975 } 976 977 return tmp; 978} 979 980static int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size) 981{ 982 struct sk_buff *skb; 983 struct tcp_skb_cb *cb; 984 struct tcphdr *th; 985 986 skb = alloc_skb(size + sizeof(*th), sk->sk_allocation); 987 if (!skb) 988 goto err; 989 990 th = (struct tcphdr *)skb_put(skb, sizeof(*th)); 991 skb_reset_transport_header(skb); 992 memset(th, 0, sizeof(*th)); 993 994 if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size)) 995 goto err_free; 996 997 cb = TCP_SKB_CB(skb); 998 999 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt; 1000 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size; 1001 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1; 1002 1003 tcp_queue_rcv(sk, skb, sizeof(*th)); 1004 1005 return size; 1006 1007err_free: 1008 kfree_skb(skb); 1009err: 1010 return -ENOMEM; 1011} 1012 1013int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg, 1014 size_t size) 1015{ 1016 struct iovec *iov; 1017 struct tcp_sock *tp = tcp_sk(sk); 1018 struct sk_buff *skb; 1019 int iovlen, flags, err, copied; 1020 int mss_now = 0, size_goal; 1021 bool sg; 1022 long timeo; 1023 1024 lock_sock(sk); 1025 1026 flags = msg->msg_flags; 1027 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 1028 1029 /* Wait for a connection to finish. */ 1030 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1031 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0) 1032 goto out_err; 1033 1034 if (unlikely(tp->repair)) { 1035 if (tp->repair_queue == TCP_RECV_QUEUE) { 1036 copied = tcp_send_rcvq(sk, msg, size); 1037 goto out; 1038 } 1039 1040 err = -EINVAL; 1041 if (tp->repair_queue == TCP_NO_QUEUE) 1042 goto out_err; 1043 1044 /* 'common' sending to sendq */ 1045 } 1046 1047 /* This should be in poll */ 1048 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags); 1049 1050 mss_now = tcp_send_mss(sk, &size_goal, flags); 1051 1052 /* Ok commence sending. */ 1053 iovlen = msg->msg_iovlen; 1054 iov = msg->msg_iov; 1055 copied = 0; 1056 1057 err = -EPIPE; 1058 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 1059 goto out_err; 1060 1061 sg = !!(sk->sk_route_caps & NETIF_F_SG); 1062 1063 while (--iovlen >= 0) { 1064 size_t seglen = iov->iov_len; 1065 unsigned char __user *from = iov->iov_base; 1066 1067 iov++; 1068 1069 while (seglen > 0) { 1070 int copy = 0; 1071 int max = size_goal; 1072 1073 skb = tcp_write_queue_tail(sk); 1074 if (tcp_send_head(sk)) { 1075 if (skb->ip_summed == CHECKSUM_NONE) 1076 max = mss_now; 1077 copy = max - skb->len; 1078 } 1079 1080 if (copy <= 0) { 1081new_segment: 1082 /* Allocate new segment. If the interface is SG, 1083 * allocate skb fitting to single page. 1084 */ 1085 if (!sk_stream_memory_free(sk)) 1086 goto wait_for_sndbuf; 1087 1088 skb = sk_stream_alloc_skb(sk, 1089 select_size(sk, sg), 1090 sk->sk_allocation); 1091 if (!skb) 1092 goto wait_for_memory; 1093 1094 /* 1095 * Check whether we can use HW checksum. 1096 */ 1097 if (sk->sk_route_caps & NETIF_F_ALL_CSUM) 1098 skb->ip_summed = CHECKSUM_PARTIAL; 1099 1100 skb_entail(sk, skb); 1101 copy = size_goal; 1102 max = size_goal; 1103 } 1104 1105 /* Try to append data to the end of skb. */ 1106 if (copy > seglen) 1107 copy = seglen; 1108 1109 /* Where to copy to? */ 1110 if (skb_availroom(skb) > 0) { 1111 /* We have some space in skb head. Superb! */ 1112 copy = min_t(int, copy, skb_availroom(skb)); 1113 err = skb_add_data_nocache(sk, skb, from, copy); 1114 if (err) 1115 goto do_fault; 1116 } else { 1117 int merge = 0; 1118 int i = skb_shinfo(skb)->nr_frags; 1119 struct page *page = sk->sk_sndmsg_page; 1120 int off; 1121 1122 if (page && page_count(page) == 1) 1123 sk->sk_sndmsg_off = 0; 1124 1125 off = sk->sk_sndmsg_off; 1126 1127 if (skb_can_coalesce(skb, i, page, off) && 1128 off != PAGE_SIZE) { 1129 /* We can extend the last page 1130 * fragment. */ 1131 merge = 1; 1132 } else if (i == MAX_SKB_FRAGS || !sg) { 1133 /* Need to add new fragment and cannot 1134 * do this because interface is non-SG, 1135 * or because all the page slots are 1136 * busy. */ 1137 tcp_mark_push(tp, skb); 1138 goto new_segment; 1139 } else if (page) { 1140 if (off == PAGE_SIZE) { 1141 put_page(page); 1142 sk->sk_sndmsg_page = page = NULL; 1143 off = 0; 1144 } 1145 } else 1146 off = 0; 1147 1148 if (copy > PAGE_SIZE - off) 1149 copy = PAGE_SIZE - off; 1150 1151 if (!sk_wmem_schedule(sk, copy)) 1152 goto wait_for_memory; 1153 1154 if (!page) { 1155 /* Allocate new cache page. */ 1156 if (!(page = sk_stream_alloc_page(sk))) 1157 goto wait_for_memory; 1158 } 1159 1160 /* Time to copy data. We are close to 1161 * the end! */ 1162 err = skb_copy_to_page_nocache(sk, from, skb, 1163 page, off, copy); 1164 if (err) { 1165 /* If this page was new, give it to the 1166 * socket so it does not get leaked. 1167 */ 1168 if (!sk->sk_sndmsg_page) { 1169 sk->sk_sndmsg_page = page; 1170 sk->sk_sndmsg_off = 0; 1171 } 1172 goto do_error; 1173 } 1174 1175 /* Update the skb. */ 1176 if (merge) { 1177 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 1178 } else { 1179 skb_fill_page_desc(skb, i, page, off, copy); 1180 if (sk->sk_sndmsg_page) { 1181 get_page(page); 1182 } else if (off + copy < PAGE_SIZE) { 1183 get_page(page); 1184 sk->sk_sndmsg_page = page; 1185 } 1186 } 1187 1188 sk->sk_sndmsg_off = off + copy; 1189 } 1190 1191 if (!copied) 1192 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 1193 1194 tp->write_seq += copy; 1195 TCP_SKB_CB(skb)->end_seq += copy; 1196 skb_shinfo(skb)->gso_segs = 0; 1197 1198 from += copy; 1199 copied += copy; 1200 if ((seglen -= copy) == 0 && iovlen == 0) 1201 goto out; 1202 1203 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair)) 1204 continue; 1205 1206 if (forced_push(tp)) { 1207 tcp_mark_push(tp, skb); 1208 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 1209 } else if (skb == tcp_send_head(sk)) 1210 tcp_push_one(sk, mss_now); 1211 continue; 1212 1213wait_for_sndbuf: 1214 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1215wait_for_memory: 1216 if (copied && likely(!tp->repair)) 1217 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH); 1218 1219 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0) 1220 goto do_error; 1221 1222 mss_now = tcp_send_mss(sk, &size_goal, flags); 1223 } 1224 } 1225 1226out: 1227 if (copied && likely(!tp->repair)) 1228 tcp_push(sk, flags, mss_now, tp->nonagle); 1229 release_sock(sk); 1230 return copied; 1231 1232do_fault: 1233 if (!skb->len) { 1234 tcp_unlink_write_queue(skb, sk); 1235 /* It is the one place in all of TCP, except connection 1236 * reset, where we can be unlinking the send_head. 1237 */ 1238 tcp_check_send_head(sk, skb); 1239 sk_wmem_free_skb(sk, skb); 1240 } 1241 1242do_error: 1243 if (copied) 1244 goto out; 1245out_err: 1246 err = sk_stream_error(sk, flags, err); 1247 release_sock(sk); 1248 return err; 1249} 1250EXPORT_SYMBOL(tcp_sendmsg); 1251 1252/* 1253 * Handle reading urgent data. BSD has very simple semantics for 1254 * this, no blocking and very strange errors 8) 1255 */ 1256 1257static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags) 1258{ 1259 struct tcp_sock *tp = tcp_sk(sk); 1260 1261 /* No URG data to read. */ 1262 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data || 1263 tp->urg_data == TCP_URG_READ) 1264 return -EINVAL; /* Yes this is right ! */ 1265 1266 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE)) 1267 return -ENOTCONN; 1268 1269 if (tp->urg_data & TCP_URG_VALID) { 1270 int err = 0; 1271 char c = tp->urg_data; 1272 1273 if (!(flags & MSG_PEEK)) 1274 tp->urg_data = TCP_URG_READ; 1275 1276 /* Read urgent data. */ 1277 msg->msg_flags |= MSG_OOB; 1278 1279 if (len > 0) { 1280 if (!(flags & MSG_TRUNC)) 1281 err = memcpy_toiovec(msg->msg_iov, &c, 1); 1282 len = 1; 1283 } else 1284 msg->msg_flags |= MSG_TRUNC; 1285 1286 return err ? -EFAULT : len; 1287 } 1288 1289 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 1290 return 0; 1291 1292 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and 1293 * the available implementations agree in this case: 1294 * this call should never block, independent of the 1295 * blocking state of the socket. 1296 * Mike <pall@rz.uni-karlsruhe.de> 1297 */ 1298 return -EAGAIN; 1299} 1300 1301static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len) 1302{ 1303 struct sk_buff *skb; 1304 int copied = 0, err = 0; 1305 1306 /* XXX -- need to support SO_PEEK_OFF */ 1307 1308 skb_queue_walk(&sk->sk_write_queue, skb) { 1309 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len); 1310 if (err) 1311 break; 1312 1313 copied += skb->len; 1314 } 1315 1316 return err ?: copied; 1317} 1318 1319/* Clean up the receive buffer for full frames taken by the user, 1320 * then send an ACK if necessary. COPIED is the number of bytes 1321 * tcp_recvmsg has given to the user so far, it speeds up the 1322 * calculation of whether or not we must ACK for the sake of 1323 * a window update. 1324 */ 1325void tcp_cleanup_rbuf(struct sock *sk, int copied) 1326{ 1327 struct tcp_sock *tp = tcp_sk(sk); 1328 int time_to_ack = 0; 1329 1330 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 1331 1332 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq), 1333 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n", 1334 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt); 1335 1336 if (inet_csk_ack_scheduled(sk)) { 1337 const struct inet_connection_sock *icsk = inet_csk(sk); 1338 /* Delayed ACKs frequently hit locked sockets during bulk 1339 * receive. */ 1340 if (icsk->icsk_ack.blocked || 1341 /* Once-per-two-segments ACK was not sent by tcp_input.c */ 1342 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss || 1343 /* 1344 * If this read emptied read buffer, we send ACK, if 1345 * connection is not bidirectional, user drained 1346 * receive buffer and there was a small segment 1347 * in queue. 1348 */ 1349 (copied > 0 && 1350 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) || 1351 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) && 1352 !icsk->icsk_ack.pingpong)) && 1353 !atomic_read(&sk->sk_rmem_alloc))) 1354 time_to_ack = 1; 1355 } 1356 1357 /* We send an ACK if we can now advertise a non-zero window 1358 * which has been raised "significantly". 1359 * 1360 * Even if window raised up to infinity, do not send window open ACK 1361 * in states, where we will not receive more. It is useless. 1362 */ 1363 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) { 1364 __u32 rcv_window_now = tcp_receive_window(tp); 1365 1366 /* Optimize, __tcp_select_window() is not cheap. */ 1367 if (2*rcv_window_now <= tp->window_clamp) { 1368 __u32 new_window = __tcp_select_window(sk); 1369 1370 /* Send ACK now, if this read freed lots of space 1371 * in our buffer. Certainly, new_window is new window. 1372 * We can advertise it now, if it is not less than current one. 1373 * "Lots" means "at least twice" here. 1374 */ 1375 if (new_window && new_window >= 2 * rcv_window_now) 1376 time_to_ack = 1; 1377 } 1378 } 1379 if (time_to_ack) 1380 tcp_send_ack(sk); 1381} 1382 1383static void tcp_prequeue_process(struct sock *sk) 1384{ 1385 struct sk_buff *skb; 1386 struct tcp_sock *tp = tcp_sk(sk); 1387 1388 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED); 1389 1390 /* RX process wants to run with disabled BHs, though it is not 1391 * necessary */ 1392 local_bh_disable(); 1393 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL) 1394 sk_backlog_rcv(sk, skb); 1395 local_bh_enable(); 1396 1397 /* Clear memory counter. */ 1398 tp->ucopy.memory = 0; 1399} 1400 1401#ifdef CONFIG_NET_DMA 1402static void tcp_service_net_dma(struct sock *sk, bool wait) 1403{ 1404 dma_cookie_t done, used; 1405 dma_cookie_t last_issued; 1406 struct tcp_sock *tp = tcp_sk(sk); 1407 1408 if (!tp->ucopy.dma_chan) 1409 return; 1410 1411 last_issued = tp->ucopy.dma_cookie; 1412 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1413 1414 do { 1415 if (dma_async_memcpy_complete(tp->ucopy.dma_chan, 1416 last_issued, &done, 1417 &used) == DMA_SUCCESS) { 1418 /* Safe to free early-copied skbs now */ 1419 __skb_queue_purge(&sk->sk_async_wait_queue); 1420 break; 1421 } else { 1422 struct sk_buff *skb; 1423 while ((skb = skb_peek(&sk->sk_async_wait_queue)) && 1424 (dma_async_is_complete(skb->dma_cookie, done, 1425 used) == DMA_SUCCESS)) { 1426 __skb_dequeue(&sk->sk_async_wait_queue); 1427 kfree_skb(skb); 1428 } 1429 } 1430 } while (wait); 1431} 1432#endif 1433 1434static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off) 1435{ 1436 struct sk_buff *skb; 1437 u32 offset; 1438 1439 skb_queue_walk(&sk->sk_receive_queue, skb) { 1440 offset = seq - TCP_SKB_CB(skb)->seq; 1441 if (tcp_hdr(skb)->syn) 1442 offset--; 1443 if (offset < skb->len || tcp_hdr(skb)->fin) { 1444 *off = offset; 1445 return skb; 1446 } 1447 } 1448 return NULL; 1449} 1450 1451/* 1452 * This routine provides an alternative to tcp_recvmsg() for routines 1453 * that would like to handle copying from skbuffs directly in 'sendfile' 1454 * fashion. 1455 * Note: 1456 * - It is assumed that the socket was locked by the caller. 1457 * - The routine does not block. 1458 * - At present, there is no support for reading OOB data 1459 * or for 'peeking' the socket using this routine 1460 * (although both would be easy to implement). 1461 */ 1462int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, 1463 sk_read_actor_t recv_actor) 1464{ 1465 struct sk_buff *skb; 1466 struct tcp_sock *tp = tcp_sk(sk); 1467 u32 seq = tp->copied_seq; 1468 u32 offset; 1469 int copied = 0; 1470 1471 if (sk->sk_state == TCP_LISTEN) 1472 return -ENOTCONN; 1473 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) { 1474 if (offset < skb->len) { 1475 int used; 1476 size_t len; 1477 1478 len = skb->len - offset; 1479 /* Stop reading if we hit a patch of urgent data */ 1480 if (tp->urg_data) { 1481 u32 urg_offset = tp->urg_seq - seq; 1482 if (urg_offset < len) 1483 len = urg_offset; 1484 if (!len) 1485 break; 1486 } 1487 used = recv_actor(desc, skb, offset, len); 1488 if (used < 0) { 1489 if (!copied) 1490 copied = used; 1491 break; 1492 } else if (used <= len) { 1493 seq += used; 1494 copied += used; 1495 offset += used; 1496 } 1497 /* 1498 * If recv_actor drops the lock (e.g. TCP splice 1499 * receive) the skb pointer might be invalid when 1500 * getting here: tcp_collapse might have deleted it 1501 * while aggregating skbs from the socket queue. 1502 */ 1503 skb = tcp_recv_skb(sk, seq-1, &offset); 1504 if (!skb || (offset+1 != skb->len)) 1505 break; 1506 } 1507 if (tcp_hdr(skb)->fin) { 1508 sk_eat_skb(sk, skb, 0); 1509 ++seq; 1510 break; 1511 } 1512 sk_eat_skb(sk, skb, 0); 1513 if (!desc->count) 1514 break; 1515 tp->copied_seq = seq; 1516 } 1517 tp->copied_seq = seq; 1518 1519 tcp_rcv_space_adjust(sk); 1520 1521 /* Clean up data we have read: This will do ACK frames. */ 1522 if (copied > 0) 1523 tcp_cleanup_rbuf(sk, copied); 1524 return copied; 1525} 1526EXPORT_SYMBOL(tcp_read_sock); 1527 1528/* 1529 * This routine copies from a sock struct into the user buffer. 1530 * 1531 * Technical note: in 2.3 we work on _locked_ socket, so that 1532 * tricks with *seq access order and skb->users are not required. 1533 * Probably, code can be easily improved even more. 1534 */ 1535 1536int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg, 1537 size_t len, int nonblock, int flags, int *addr_len) 1538{ 1539 struct tcp_sock *tp = tcp_sk(sk); 1540 int copied = 0; 1541 u32 peek_seq; 1542 u32 *seq; 1543 unsigned long used; 1544 int err; 1545 int target; /* Read at least this many bytes */ 1546 long timeo; 1547 struct task_struct *user_recv = NULL; 1548 int copied_early = 0; 1549 struct sk_buff *skb; 1550 u32 urg_hole = 0; 1551 1552 lock_sock(sk); 1553 1554 err = -ENOTCONN; 1555 if (sk->sk_state == TCP_LISTEN) 1556 goto out; 1557 1558 timeo = sock_rcvtimeo(sk, nonblock); 1559 1560 /* Urgent data needs to be handled specially. */ 1561 if (flags & MSG_OOB) 1562 goto recv_urg; 1563 1564 if (unlikely(tp->repair)) { 1565 err = -EPERM; 1566 if (!(flags & MSG_PEEK)) 1567 goto out; 1568 1569 if (tp->repair_queue == TCP_SEND_QUEUE) 1570 goto recv_sndq; 1571 1572 err = -EINVAL; 1573 if (tp->repair_queue == TCP_NO_QUEUE) 1574 goto out; 1575 1576 /* 'common' recv queue MSG_PEEK-ing */ 1577 } 1578 1579 seq = &tp->copied_seq; 1580 if (flags & MSG_PEEK) { 1581 peek_seq = tp->copied_seq; 1582 seq = &peek_seq; 1583 } 1584 1585 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 1586 1587#ifdef CONFIG_NET_DMA 1588 tp->ucopy.dma_chan = NULL; 1589 preempt_disable(); 1590 skb = skb_peek_tail(&sk->sk_receive_queue); 1591 { 1592 int available = 0; 1593 1594 if (skb) 1595 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq); 1596 if ((available < target) && 1597 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) && 1598 !sysctl_tcp_low_latency && 1599 net_dma_find_channel()) { 1600 preempt_enable_no_resched(); 1601 tp->ucopy.pinned_list = 1602 dma_pin_iovec_pages(msg->msg_iov, len); 1603 } else { 1604 preempt_enable_no_resched(); 1605 } 1606 } 1607#endif 1608 1609 do { 1610 u32 offset; 1611 1612 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */ 1613 if (tp->urg_data && tp->urg_seq == *seq) { 1614 if (copied) 1615 break; 1616 if (signal_pending(current)) { 1617 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN; 1618 break; 1619 } 1620 } 1621 1622 /* Next get a buffer. */ 1623 1624 skb_queue_walk(&sk->sk_receive_queue, skb) { 1625 /* Now that we have two receive queues this 1626 * shouldn't happen. 1627 */ 1628 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq), 1629 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n", 1630 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, 1631 flags)) 1632 break; 1633 1634 offset = *seq - TCP_SKB_CB(skb)->seq; 1635 if (tcp_hdr(skb)->syn) 1636 offset--; 1637 if (offset < skb->len) 1638 goto found_ok_skb; 1639 if (tcp_hdr(skb)->fin) 1640 goto found_fin_ok; 1641 WARN(!(flags & MSG_PEEK), 1642 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n", 1643 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags); 1644 } 1645 1646 /* Well, if we have backlog, try to process it now yet. */ 1647 1648 if (copied >= target && !sk->sk_backlog.tail) 1649 break; 1650 1651 if (copied) { 1652 if (sk->sk_err || 1653 sk->sk_state == TCP_CLOSE || 1654 (sk->sk_shutdown & RCV_SHUTDOWN) || 1655 !timeo || 1656 signal_pending(current)) 1657 break; 1658 } else { 1659 if (sock_flag(sk, SOCK_DONE)) 1660 break; 1661 1662 if (sk->sk_err) { 1663 copied = sock_error(sk); 1664 break; 1665 } 1666 1667 if (sk->sk_shutdown & RCV_SHUTDOWN) 1668 break; 1669 1670 if (sk->sk_state == TCP_CLOSE) { 1671 if (!sock_flag(sk, SOCK_DONE)) { 1672 /* This occurs when user tries to read 1673 * from never connected socket. 1674 */ 1675 copied = -ENOTCONN; 1676 break; 1677 } 1678 break; 1679 } 1680 1681 if (!timeo) { 1682 copied = -EAGAIN; 1683 break; 1684 } 1685 1686 if (signal_pending(current)) { 1687 copied = sock_intr_errno(timeo); 1688 break; 1689 } 1690 } 1691 1692 tcp_cleanup_rbuf(sk, copied); 1693 1694 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) { 1695 /* Install new reader */ 1696 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) { 1697 user_recv = current; 1698 tp->ucopy.task = user_recv; 1699 tp->ucopy.iov = msg->msg_iov; 1700 } 1701 1702 tp->ucopy.len = len; 1703 1704 WARN_ON(tp->copied_seq != tp->rcv_nxt && 1705 !(flags & (MSG_PEEK | MSG_TRUNC))); 1706 1707 /* Ugly... If prequeue is not empty, we have to 1708 * process it before releasing socket, otherwise 1709 * order will be broken at second iteration. 1710 * More elegant solution is required!!! 1711 * 1712 * Look: we have the following (pseudo)queues: 1713 * 1714 * 1. packets in flight 1715 * 2. backlog 1716 * 3. prequeue 1717 * 4. receive_queue 1718 * 1719 * Each queue can be processed only if the next ones 1720 * are empty. At this point we have empty receive_queue. 1721 * But prequeue _can_ be not empty after 2nd iteration, 1722 * when we jumped to start of loop because backlog 1723 * processing added something to receive_queue. 1724 * We cannot release_sock(), because backlog contains 1725 * packets arrived _after_ prequeued ones. 1726 * 1727 * Shortly, algorithm is clear --- to process all 1728 * the queues in order. We could make it more directly, 1729 * requeueing packets from backlog to prequeue, if 1730 * is not empty. It is more elegant, but eats cycles, 1731 * unfortunately. 1732 */ 1733 if (!skb_queue_empty(&tp->ucopy.prequeue)) 1734 goto do_prequeue; 1735 1736 /* __ Set realtime policy in scheduler __ */ 1737 } 1738 1739#ifdef CONFIG_NET_DMA 1740 if (tp->ucopy.dma_chan) 1741 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1742#endif 1743 if (copied >= target) { 1744 /* Do not sleep, just process backlog. */ 1745 release_sock(sk); 1746 lock_sock(sk); 1747 } else 1748 sk_wait_data(sk, &timeo); 1749 1750#ifdef CONFIG_NET_DMA 1751 tcp_service_net_dma(sk, false); /* Don't block */ 1752 tp->ucopy.wakeup = 0; 1753#endif 1754 1755 if (user_recv) { 1756 int chunk; 1757 1758 /* __ Restore normal policy in scheduler __ */ 1759 1760 if ((chunk = len - tp->ucopy.len) != 0) { 1761 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk); 1762 len -= chunk; 1763 copied += chunk; 1764 } 1765 1766 if (tp->rcv_nxt == tp->copied_seq && 1767 !skb_queue_empty(&tp->ucopy.prequeue)) { 1768do_prequeue: 1769 tcp_prequeue_process(sk); 1770 1771 if ((chunk = len - tp->ucopy.len) != 0) { 1772 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1773 len -= chunk; 1774 copied += chunk; 1775 } 1776 } 1777 } 1778 if ((flags & MSG_PEEK) && 1779 (peek_seq - copied - urg_hole != tp->copied_seq)) { 1780 if (net_ratelimit()) 1781 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n", 1782 current->comm, task_pid_nr(current)); 1783 peek_seq = tp->copied_seq; 1784 } 1785 continue; 1786 1787 found_ok_skb: 1788 /* Ok so how much can we use? */ 1789 used = skb->len - offset; 1790 if (len < used) 1791 used = len; 1792 1793 /* Do we have urgent data here? */ 1794 if (tp->urg_data) { 1795 u32 urg_offset = tp->urg_seq - *seq; 1796 if (urg_offset < used) { 1797 if (!urg_offset) { 1798 if (!sock_flag(sk, SOCK_URGINLINE)) { 1799 ++*seq; 1800 urg_hole++; 1801 offset++; 1802 used--; 1803 if (!used) 1804 goto skip_copy; 1805 } 1806 } else 1807 used = urg_offset; 1808 } 1809 } 1810 1811 if (!(flags & MSG_TRUNC)) { 1812#ifdef CONFIG_NET_DMA 1813 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 1814 tp->ucopy.dma_chan = net_dma_find_channel(); 1815 1816 if (tp->ucopy.dma_chan) { 1817 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec( 1818 tp->ucopy.dma_chan, skb, offset, 1819 msg->msg_iov, used, 1820 tp->ucopy.pinned_list); 1821 1822 if (tp->ucopy.dma_cookie < 0) { 1823 1824 pr_alert("%s: dma_cookie < 0\n", 1825 __func__); 1826 1827 /* Exception. Bailout! */ 1828 if (!copied) 1829 copied = -EFAULT; 1830 break; 1831 } 1832 1833 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1834 1835 if ((offset + used) == skb->len) 1836 copied_early = 1; 1837 1838 } else 1839#endif 1840 { 1841 err = skb_copy_datagram_iovec(skb, offset, 1842 msg->msg_iov, used); 1843 if (err) { 1844 /* Exception. Bailout! */ 1845 if (!copied) 1846 copied = -EFAULT; 1847 break; 1848 } 1849 } 1850 } 1851 1852 *seq += used; 1853 copied += used; 1854 len -= used; 1855 1856 tcp_rcv_space_adjust(sk); 1857 1858skip_copy: 1859 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) { 1860 tp->urg_data = 0; 1861 tcp_fast_path_check(sk); 1862 } 1863 if (used + offset < skb->len) 1864 continue; 1865 1866 if (tcp_hdr(skb)->fin) 1867 goto found_fin_ok; 1868 if (!(flags & MSG_PEEK)) { 1869 sk_eat_skb(sk, skb, copied_early); 1870 copied_early = 0; 1871 } 1872 continue; 1873 1874 found_fin_ok: 1875 /* Process the FIN. */ 1876 ++*seq; 1877 if (!(flags & MSG_PEEK)) { 1878 sk_eat_skb(sk, skb, copied_early); 1879 copied_early = 0; 1880 } 1881 break; 1882 } while (len > 0); 1883 1884 if (user_recv) { 1885 if (!skb_queue_empty(&tp->ucopy.prequeue)) { 1886 int chunk; 1887 1888 tp->ucopy.len = copied > 0 ? len : 0; 1889 1890 tcp_prequeue_process(sk); 1891 1892 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) { 1893 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1894 len -= chunk; 1895 copied += chunk; 1896 } 1897 } 1898 1899 tp->ucopy.task = NULL; 1900 tp->ucopy.len = 0; 1901 } 1902 1903#ifdef CONFIG_NET_DMA 1904 tcp_service_net_dma(sk, true); /* Wait for queue to drain */ 1905 tp->ucopy.dma_chan = NULL; 1906 1907 if (tp->ucopy.pinned_list) { 1908 dma_unpin_iovec_pages(tp->ucopy.pinned_list); 1909 tp->ucopy.pinned_list = NULL; 1910 } 1911#endif 1912 1913 /* According to UNIX98, msg_name/msg_namelen are ignored 1914 * on connected socket. I was just happy when found this 8) --ANK 1915 */ 1916 1917 /* Clean up data we have read: This will do ACK frames. */ 1918 tcp_cleanup_rbuf(sk, copied); 1919 1920 release_sock(sk); 1921 return copied; 1922 1923out: 1924 release_sock(sk); 1925 return err; 1926 1927recv_urg: 1928 err = tcp_recv_urg(sk, msg, len, flags); 1929 goto out; 1930 1931recv_sndq: 1932 err = tcp_peek_sndq(sk, msg, len); 1933 goto out; 1934} 1935EXPORT_SYMBOL(tcp_recvmsg); 1936 1937void tcp_set_state(struct sock *sk, int state) 1938{ 1939 int oldstate = sk->sk_state; 1940 1941 switch (state) { 1942 case TCP_ESTABLISHED: 1943 if (oldstate != TCP_ESTABLISHED) 1944 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1945 break; 1946 1947 case TCP_CLOSE: 1948 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED) 1949 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS); 1950 1951 sk->sk_prot->unhash(sk); 1952 if (inet_csk(sk)->icsk_bind_hash && 1953 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK)) 1954 inet_put_port(sk); 1955 /* fall through */ 1956 default: 1957 if (oldstate == TCP_ESTABLISHED) 1958 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1959 } 1960 1961 /* Change state AFTER socket is unhashed to avoid closed 1962 * socket sitting in hash tables. 1963 */ 1964 sk->sk_state = state; 1965 1966#ifdef STATE_TRACE 1967 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]); 1968#endif 1969} 1970EXPORT_SYMBOL_GPL(tcp_set_state); 1971 1972/* 1973 * State processing on a close. This implements the state shift for 1974 * sending our FIN frame. Note that we only send a FIN for some 1975 * states. A shutdown() may have already sent the FIN, or we may be 1976 * closed. 1977 */ 1978 1979static const unsigned char new_state[16] = { 1980 /* current state: new state: action: */ 1981 /* (Invalid) */ TCP_CLOSE, 1982 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1983 /* TCP_SYN_SENT */ TCP_CLOSE, 1984 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1985 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1, 1986 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2, 1987 /* TCP_TIME_WAIT */ TCP_CLOSE, 1988 /* TCP_CLOSE */ TCP_CLOSE, 1989 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN, 1990 /* TCP_LAST_ACK */ TCP_LAST_ACK, 1991 /* TCP_LISTEN */ TCP_CLOSE, 1992 /* TCP_CLOSING */ TCP_CLOSING, 1993}; 1994 1995static int tcp_close_state(struct sock *sk) 1996{ 1997 int next = (int)new_state[sk->sk_state]; 1998 int ns = next & TCP_STATE_MASK; 1999 2000 tcp_set_state(sk, ns); 2001 2002 return next & TCP_ACTION_FIN; 2003} 2004 2005/* 2006 * Shutdown the sending side of a connection. Much like close except 2007 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD). 2008 */ 2009 2010void tcp_shutdown(struct sock *sk, int how) 2011{ 2012 /* We need to grab some memory, and put together a FIN, 2013 * and then put it into the queue to be sent. 2014 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92. 2015 */ 2016 if (!(how & SEND_SHUTDOWN)) 2017 return; 2018 2019 /* If we've already sent a FIN, or it's a closed state, skip this. */ 2020 if ((1 << sk->sk_state) & 2021 (TCPF_ESTABLISHED | TCPF_SYN_SENT | 2022 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) { 2023 /* Clear out any half completed packets. FIN if needed. */ 2024 if (tcp_close_state(sk)) 2025 tcp_send_fin(sk); 2026 } 2027} 2028EXPORT_SYMBOL(tcp_shutdown); 2029 2030bool tcp_check_oom(struct sock *sk, int shift) 2031{ 2032 bool too_many_orphans, out_of_socket_memory; 2033 2034 too_many_orphans = tcp_too_many_orphans(sk, shift); 2035 out_of_socket_memory = tcp_out_of_memory(sk); 2036 2037 if (too_many_orphans && net_ratelimit()) 2038 pr_info("too many orphaned sockets\n"); 2039 if (out_of_socket_memory && net_ratelimit()) 2040 pr_info("out of memory -- consider tuning tcp_mem\n"); 2041 return too_many_orphans || out_of_socket_memory; 2042} 2043 2044void tcp_close(struct sock *sk, long timeout) 2045{ 2046 struct sk_buff *skb; 2047 int data_was_unread = 0; 2048 int state; 2049 2050 lock_sock(sk); 2051 sk->sk_shutdown = SHUTDOWN_MASK; 2052 2053 if (sk->sk_state == TCP_LISTEN) { 2054 tcp_set_state(sk, TCP_CLOSE); 2055 2056 /* Special case. */ 2057 inet_csk_listen_stop(sk); 2058 2059 goto adjudge_to_death; 2060 } 2061 2062 /* We need to flush the recv. buffs. We do this only on the 2063 * descriptor close, not protocol-sourced closes, because the 2064 * reader process may not have drained the data yet! 2065 */ 2066 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) { 2067 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq - 2068 tcp_hdr(skb)->fin; 2069 data_was_unread += len; 2070 __kfree_skb(skb); 2071 } 2072 2073 sk_mem_reclaim(sk); 2074 2075 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */ 2076 if (sk->sk_state == TCP_CLOSE) 2077 goto adjudge_to_death; 2078 2079 /* As outlined in RFC 2525, section 2.17, we send a RST here because 2080 * data was lost. To witness the awful effects of the old behavior of 2081 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk 2082 * GET in an FTP client, suspend the process, wait for the client to 2083 * advertise a zero window, then kill -9 the FTP client, wheee... 2084 * Note: timeout is always zero in such a case. 2085 */ 2086 if (unlikely(tcp_sk(sk)->repair)) { 2087 sk->sk_prot->disconnect(sk, 0); 2088 } else if (data_was_unread) { 2089 /* Unread data was tossed, zap the connection. */ 2090 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE); 2091 tcp_set_state(sk, TCP_CLOSE); 2092 tcp_send_active_reset(sk, sk->sk_allocation); 2093 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) { 2094 /* Check zero linger _after_ checking for unread data. */ 2095 sk->sk_prot->disconnect(sk, 0); 2096 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 2097 } else if (tcp_close_state(sk)) { 2098 /* We FIN if the application ate all the data before 2099 * zapping the connection. 2100 */ 2101 2102 /* RED-PEN. Formally speaking, we have broken TCP state 2103 * machine. State transitions: 2104 * 2105 * TCP_ESTABLISHED -> TCP_FIN_WAIT1 2106 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible) 2107 * TCP_CLOSE_WAIT -> TCP_LAST_ACK 2108 * 2109 * are legal only when FIN has been sent (i.e. in window), 2110 * rather than queued out of window. Purists blame. 2111 * 2112 * F.e. "RFC state" is ESTABLISHED, 2113 * if Linux state is FIN-WAIT-1, but FIN is still not sent. 2114 * 2115 * The visible declinations are that sometimes 2116 * we enter time-wait state, when it is not required really 2117 * (harmless), do not send active resets, when they are 2118 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when 2119 * they look as CLOSING or LAST_ACK for Linux) 2120 * Probably, I missed some more holelets. 2121 * --ANK 2122 */ 2123 tcp_send_fin(sk); 2124 } 2125 2126 sk_stream_wait_close(sk, timeout); 2127 2128adjudge_to_death: 2129 state = sk->sk_state; 2130 sock_hold(sk); 2131 sock_orphan(sk); 2132 2133 /* It is the last release_sock in its life. It will remove backlog. */ 2134 release_sock(sk); 2135 2136 2137 /* Now socket is owned by kernel and we acquire BH lock 2138 to finish close. No need to check for user refs. 2139 */ 2140 local_bh_disable(); 2141 bh_lock_sock(sk); 2142 WARN_ON(sock_owned_by_user(sk)); 2143 2144 percpu_counter_inc(sk->sk_prot->orphan_count); 2145 2146 /* Have we already been destroyed by a softirq or backlog? */ 2147 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE) 2148 goto out; 2149 2150 /* This is a (useful) BSD violating of the RFC. There is a 2151 * problem with TCP as specified in that the other end could 2152 * keep a socket open forever with no application left this end. 2153 * We use a 3 minute timeout (about the same as BSD) then kill 2154 * our end. If they send after that then tough - BUT: long enough 2155 * that we won't make the old 4*rto = almost no time - whoops 2156 * reset mistake. 2157 * 2158 * Nope, it was not mistake. It is really desired behaviour 2159 * f.e. on http servers, when such sockets are useless, but 2160 * consume significant resources. Let's do it with special 2161 * linger2 option. --ANK 2162 */ 2163 2164 if (sk->sk_state == TCP_FIN_WAIT2) { 2165 struct tcp_sock *tp = tcp_sk(sk); 2166 if (tp->linger2 < 0) { 2167 tcp_set_state(sk, TCP_CLOSE); 2168 tcp_send_active_reset(sk, GFP_ATOMIC); 2169 NET_INC_STATS_BH(sock_net(sk), 2170 LINUX_MIB_TCPABORTONLINGER); 2171 } else { 2172 const int tmo = tcp_fin_time(sk); 2173 2174 if (tmo > TCP_TIMEWAIT_LEN) { 2175 inet_csk_reset_keepalive_timer(sk, 2176 tmo - TCP_TIMEWAIT_LEN); 2177 } else { 2178 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 2179 goto out; 2180 } 2181 } 2182 } 2183 if (sk->sk_state != TCP_CLOSE) { 2184 sk_mem_reclaim(sk); 2185 if (tcp_check_oom(sk, 0)) { 2186 tcp_set_state(sk, TCP_CLOSE); 2187 tcp_send_active_reset(sk, GFP_ATOMIC); 2188 NET_INC_STATS_BH(sock_net(sk), 2189 LINUX_MIB_TCPABORTONMEMORY); 2190 } 2191 } 2192 2193 if (sk->sk_state == TCP_CLOSE) 2194 inet_csk_destroy_sock(sk); 2195 /* Otherwise, socket is reprieved until protocol close. */ 2196 2197out: 2198 bh_unlock_sock(sk); 2199 local_bh_enable(); 2200 sock_put(sk); 2201} 2202EXPORT_SYMBOL(tcp_close); 2203 2204/* These states need RST on ABORT according to RFC793 */ 2205 2206static inline int tcp_need_reset(int state) 2207{ 2208 return (1 << state) & 2209 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 | 2210 TCPF_FIN_WAIT2 | TCPF_SYN_RECV); 2211} 2212 2213int tcp_disconnect(struct sock *sk, int flags) 2214{ 2215 struct inet_sock *inet = inet_sk(sk); 2216 struct inet_connection_sock *icsk = inet_csk(sk); 2217 struct tcp_sock *tp = tcp_sk(sk); 2218 int err = 0; 2219 int old_state = sk->sk_state; 2220 2221 if (old_state != TCP_CLOSE) 2222 tcp_set_state(sk, TCP_CLOSE); 2223 2224 /* ABORT function of RFC793 */ 2225 if (old_state == TCP_LISTEN) { 2226 inet_csk_listen_stop(sk); 2227 } else if (unlikely(tp->repair)) { 2228 sk->sk_err = ECONNABORTED; 2229 } else if (tcp_need_reset(old_state) || 2230 (tp->snd_nxt != tp->write_seq && 2231 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) { 2232 /* The last check adjusts for discrepancy of Linux wrt. RFC 2233 * states 2234 */ 2235 tcp_send_active_reset(sk, gfp_any()); 2236 sk->sk_err = ECONNRESET; 2237 } else if (old_state == TCP_SYN_SENT) 2238 sk->sk_err = ECONNRESET; 2239 2240 tcp_clear_xmit_timers(sk); 2241 __skb_queue_purge(&sk->sk_receive_queue); 2242 tcp_write_queue_purge(sk); 2243 __skb_queue_purge(&tp->out_of_order_queue); 2244#ifdef CONFIG_NET_DMA 2245 __skb_queue_purge(&sk->sk_async_wait_queue); 2246#endif 2247 2248 inet->inet_dport = 0; 2249 2250 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK)) 2251 inet_reset_saddr(sk); 2252 2253 sk->sk_shutdown = 0; 2254 sock_reset_flag(sk, SOCK_DONE); 2255 tp->srtt = 0; 2256 if ((tp->write_seq += tp->max_window + 2) == 0) 2257 tp->write_seq = 1; 2258 icsk->icsk_backoff = 0; 2259 tp->snd_cwnd = 2; 2260 icsk->icsk_probes_out = 0; 2261 tp->packets_out = 0; 2262 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 2263 tp->snd_cwnd_cnt = 0; 2264 tp->bytes_acked = 0; 2265 tp->window_clamp = 0; 2266 tcp_set_ca_state(sk, TCP_CA_Open); 2267 tcp_clear_retrans(tp); 2268 inet_csk_delack_init(sk); 2269 tcp_init_send_head(sk); 2270 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt)); 2271 __sk_dst_reset(sk); 2272 2273 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash); 2274 2275 sk->sk_error_report(sk); 2276 return err; 2277} 2278EXPORT_SYMBOL(tcp_disconnect); 2279 2280static inline int tcp_can_repair_sock(struct sock *sk) 2281{ 2282 return capable(CAP_NET_ADMIN) && 2283 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED)); 2284} 2285 2286static int tcp_repair_options_est(struct tcp_sock *tp, char __user *optbuf, unsigned int len) 2287{ 2288 /* 2289 * Options are stored in CODE:VALUE form where CODE is 8bit and VALUE 2290 * fits the respective TCPOLEN_ size 2291 */ 2292 2293 while (len > 0) { 2294 u8 opcode; 2295 2296 if (get_user(opcode, optbuf)) 2297 return -EFAULT; 2298 2299 optbuf++; 2300 len--; 2301 2302 switch (opcode) { 2303 case TCPOPT_MSS: { 2304 u16 in_mss; 2305 2306 if (len < sizeof(in_mss)) 2307 return -ENODATA; 2308 if (get_user(in_mss, optbuf)) 2309 return -EFAULT; 2310 2311 tp->rx_opt.mss_clamp = in_mss; 2312 2313 optbuf += sizeof(in_mss); 2314 len -= sizeof(in_mss); 2315 break; 2316 } 2317 case TCPOPT_WINDOW: { 2318 u8 wscale; 2319 2320 if (len < sizeof(wscale)) 2321 return -ENODATA; 2322 if (get_user(wscale, optbuf)) 2323 return -EFAULT; 2324 2325 if (wscale > 14) 2326 return -EFBIG; 2327 2328 tp->rx_opt.snd_wscale = wscale; 2329 2330 optbuf += sizeof(wscale); 2331 len -= sizeof(wscale); 2332 break; 2333 } 2334 case TCPOPT_SACK_PERM: 2335 tp->rx_opt.sack_ok |= TCP_SACK_SEEN; 2336 if (sysctl_tcp_fack) 2337 tcp_enable_fack(tp); 2338 break; 2339 case TCPOPT_TIMESTAMP: 2340 tp->rx_opt.tstamp_ok = 1; 2341 break; 2342 } 2343 } 2344 2345 return 0; 2346} 2347 2348/* 2349 * Socket option code for TCP. 2350 */ 2351static int do_tcp_setsockopt(struct sock *sk, int level, 2352 int optname, char __user *optval, unsigned int optlen) 2353{ 2354 struct tcp_sock *tp = tcp_sk(sk); 2355 struct inet_connection_sock *icsk = inet_csk(sk); 2356 int val; 2357 int err = 0; 2358 2359 /* These are data/string values, all the others are ints */ 2360 switch (optname) { 2361 case TCP_CONGESTION: { 2362 char name[TCP_CA_NAME_MAX]; 2363 2364 if (optlen < 1) 2365 return -EINVAL; 2366 2367 val = strncpy_from_user(name, optval, 2368 min_t(long, TCP_CA_NAME_MAX-1, optlen)); 2369 if (val < 0) 2370 return -EFAULT; 2371 name[val] = 0; 2372 2373 lock_sock(sk); 2374 err = tcp_set_congestion_control(sk, name); 2375 release_sock(sk); 2376 return err; 2377 } 2378 case TCP_COOKIE_TRANSACTIONS: { 2379 struct tcp_cookie_transactions ctd; 2380 struct tcp_cookie_values *cvp = NULL; 2381 2382 if (sizeof(ctd) > optlen) 2383 return -EINVAL; 2384 if (copy_from_user(&ctd, optval, sizeof(ctd))) 2385 return -EFAULT; 2386 2387 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) || 2388 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED) 2389 return -EINVAL; 2390 2391 if (ctd.tcpct_cookie_desired == 0) { 2392 /* default to global value */ 2393 } else if ((0x1 & ctd.tcpct_cookie_desired) || 2394 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX || 2395 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) { 2396 return -EINVAL; 2397 } 2398 2399 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) { 2400 /* Supercedes all other values */ 2401 lock_sock(sk); 2402 if (tp->cookie_values != NULL) { 2403 kref_put(&tp->cookie_values->kref, 2404 tcp_cookie_values_release); 2405 tp->cookie_values = NULL; 2406 } 2407 tp->rx_opt.cookie_in_always = 0; /* false */ 2408 tp->rx_opt.cookie_out_never = 1; /* true */ 2409 release_sock(sk); 2410 return err; 2411 } 2412 2413 /* Allocate ancillary memory before locking. 2414 */ 2415 if (ctd.tcpct_used > 0 || 2416 (tp->cookie_values == NULL && 2417 (sysctl_tcp_cookie_size > 0 || 2418 ctd.tcpct_cookie_desired > 0 || 2419 ctd.tcpct_s_data_desired > 0))) { 2420 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used, 2421 GFP_KERNEL); 2422 if (cvp == NULL) 2423 return -ENOMEM; 2424 2425 kref_init(&cvp->kref); 2426 } 2427 lock_sock(sk); 2428 tp->rx_opt.cookie_in_always = 2429 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags); 2430 tp->rx_opt.cookie_out_never = 0; /* false */ 2431 2432 if (tp->cookie_values != NULL) { 2433 if (cvp != NULL) { 2434 /* Changed values are recorded by a changed 2435 * pointer, ensuring the cookie will differ, 2436 * without separately hashing each value later. 2437 */ 2438 kref_put(&tp->cookie_values->kref, 2439 tcp_cookie_values_release); 2440 } else { 2441 cvp = tp->cookie_values; 2442 } 2443 } 2444 2445 if (cvp != NULL) { 2446 cvp->cookie_desired = ctd.tcpct_cookie_desired; 2447 2448 if (ctd.tcpct_used > 0) { 2449 memcpy(cvp->s_data_payload, ctd.tcpct_value, 2450 ctd.tcpct_used); 2451 cvp->s_data_desired = ctd.tcpct_used; 2452 cvp->s_data_constant = 1; /* true */ 2453 } else { 2454 /* No constant payload data. */ 2455 cvp->s_data_desired = ctd.tcpct_s_data_desired; 2456 cvp->s_data_constant = 0; /* false */ 2457 } 2458 2459 tp->cookie_values = cvp; 2460 } 2461 release_sock(sk); 2462 return err; 2463 } 2464 default: 2465 /* fallthru */ 2466 break; 2467 } 2468 2469 if (optlen < sizeof(int)) 2470 return -EINVAL; 2471 2472 if (get_user(val, (int __user *)optval)) 2473 return -EFAULT; 2474 2475 lock_sock(sk); 2476 2477 switch (optname) { 2478 case TCP_MAXSEG: 2479 /* Values greater than interface MTU won't take effect. However 2480 * at the point when this call is done we typically don't yet 2481 * know which interface is going to be used */ 2482 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) { 2483 err = -EINVAL; 2484 break; 2485 } 2486 tp->rx_opt.user_mss = val; 2487 break; 2488 2489 case TCP_NODELAY: 2490 if (val) { 2491 /* TCP_NODELAY is weaker than TCP_CORK, so that 2492 * this option on corked socket is remembered, but 2493 * it is not activated until cork is cleared. 2494 * 2495 * However, when TCP_NODELAY is set we make 2496 * an explicit push, which overrides even TCP_CORK 2497 * for currently queued segments. 2498 */ 2499 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH; 2500 tcp_push_pending_frames(sk); 2501 } else { 2502 tp->nonagle &= ~TCP_NAGLE_OFF; 2503 } 2504 break; 2505 2506 case TCP_THIN_LINEAR_TIMEOUTS: 2507 if (val < 0 || val > 1) 2508 err = -EINVAL; 2509 else 2510 tp->thin_lto = val; 2511 break; 2512 2513 case TCP_THIN_DUPACK: 2514 if (val < 0 || val > 1) 2515 err = -EINVAL; 2516 else 2517 tp->thin_dupack = val; 2518 break; 2519 2520 case TCP_REPAIR: 2521 if (!tcp_can_repair_sock(sk)) 2522 err = -EPERM; 2523 else if (val == 1) { 2524 tp->repair = 1; 2525 sk->sk_reuse = SK_FORCE_REUSE; 2526 tp->repair_queue = TCP_NO_QUEUE; 2527 } else if (val == 0) { 2528 tp->repair = 0; 2529 sk->sk_reuse = SK_NO_REUSE; 2530 tcp_send_window_probe(sk); 2531 } else 2532 err = -EINVAL; 2533 2534 break; 2535 2536 case TCP_REPAIR_QUEUE: 2537 if (!tp->repair) 2538 err = -EPERM; 2539 else if (val < TCP_QUEUES_NR) 2540 tp->repair_queue = val; 2541 else 2542 err = -EINVAL; 2543 break; 2544 2545 case TCP_QUEUE_SEQ: 2546 if (sk->sk_state != TCP_CLOSE) 2547 err = -EPERM; 2548 else if (tp->repair_queue == TCP_SEND_QUEUE) 2549 tp->write_seq = val; 2550 else if (tp->repair_queue == TCP_RECV_QUEUE) 2551 tp->rcv_nxt = val; 2552 else 2553 err = -EINVAL; 2554 break; 2555 2556 case TCP_REPAIR_OPTIONS: 2557 if (!tp->repair) 2558 err = -EINVAL; 2559 else if (sk->sk_state == TCP_ESTABLISHED) 2560 err = tcp_repair_options_est(tp, optval, optlen); 2561 else 2562 err = -EPERM; 2563 break; 2564 2565 case TCP_CORK: 2566 /* When set indicates to always queue non-full frames. 2567 * Later the user clears this option and we transmit 2568 * any pending partial frames in the queue. This is 2569 * meant to be used alongside sendfile() to get properly 2570 * filled frames when the user (for example) must write 2571 * out headers with a write() call first and then use 2572 * sendfile to send out the data parts. 2573 * 2574 * TCP_CORK can be set together with TCP_NODELAY and it is 2575 * stronger than TCP_NODELAY. 2576 */ 2577 if (val) { 2578 tp->nonagle |= TCP_NAGLE_CORK; 2579 } else { 2580 tp->nonagle &= ~TCP_NAGLE_CORK; 2581 if (tp->nonagle&TCP_NAGLE_OFF) 2582 tp->nonagle |= TCP_NAGLE_PUSH; 2583 tcp_push_pending_frames(sk); 2584 } 2585 break; 2586 2587 case TCP_KEEPIDLE: 2588 if (val < 1 || val > MAX_TCP_KEEPIDLE) 2589 err = -EINVAL; 2590 else { 2591 tp->keepalive_time = val * HZ; 2592 if (sock_flag(sk, SOCK_KEEPOPEN) && 2593 !((1 << sk->sk_state) & 2594 (TCPF_CLOSE | TCPF_LISTEN))) { 2595 u32 elapsed = keepalive_time_elapsed(tp); 2596 if (tp->keepalive_time > elapsed) 2597 elapsed = tp->keepalive_time - elapsed; 2598 else 2599 elapsed = 0; 2600 inet_csk_reset_keepalive_timer(sk, elapsed); 2601 } 2602 } 2603 break; 2604 case TCP_KEEPINTVL: 2605 if (val < 1 || val > MAX_TCP_KEEPINTVL) 2606 err = -EINVAL; 2607 else 2608 tp->keepalive_intvl = val * HZ; 2609 break; 2610 case TCP_KEEPCNT: 2611 if (val < 1 || val > MAX_TCP_KEEPCNT) 2612 err = -EINVAL; 2613 else 2614 tp->keepalive_probes = val; 2615 break; 2616 case TCP_SYNCNT: 2617 if (val < 1 || val > MAX_TCP_SYNCNT) 2618 err = -EINVAL; 2619 else 2620 icsk->icsk_syn_retries = val; 2621 break; 2622 2623 case TCP_LINGER2: 2624 if (val < 0) 2625 tp->linger2 = -1; 2626 else if (val > sysctl_tcp_fin_timeout / HZ) 2627 tp->linger2 = 0; 2628 else 2629 tp->linger2 = val * HZ; 2630 break; 2631 2632 case TCP_DEFER_ACCEPT: 2633 /* Translate value in seconds to number of retransmits */ 2634 icsk->icsk_accept_queue.rskq_defer_accept = 2635 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ, 2636 TCP_RTO_MAX / HZ); 2637 break; 2638 2639 case TCP_WINDOW_CLAMP: 2640 if (!val) { 2641 if (sk->sk_state != TCP_CLOSE) { 2642 err = -EINVAL; 2643 break; 2644 } 2645 tp->window_clamp = 0; 2646 } else 2647 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ? 2648 SOCK_MIN_RCVBUF / 2 : val; 2649 break; 2650 2651 case TCP_QUICKACK: 2652 if (!val) { 2653 icsk->icsk_ack.pingpong = 1; 2654 } else { 2655 icsk->icsk_ack.pingpong = 0; 2656 if ((1 << sk->sk_state) & 2657 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) && 2658 inet_csk_ack_scheduled(sk)) { 2659 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 2660 tcp_cleanup_rbuf(sk, 1); 2661 if (!(val & 1)) 2662 icsk->icsk_ack.pingpong = 1; 2663 } 2664 } 2665 break; 2666 2667#ifdef CONFIG_TCP_MD5SIG 2668 case TCP_MD5SIG: 2669 /* Read the IP->Key mappings from userspace */ 2670 err = tp->af_specific->md5_parse(sk, optval, optlen); 2671 break; 2672#endif 2673 case TCP_USER_TIMEOUT: 2674 /* Cap the max timeout in ms TCP will retry/retrans 2675 * before giving up and aborting (ETIMEDOUT) a connection. 2676 */ 2677 icsk->icsk_user_timeout = msecs_to_jiffies(val); 2678 break; 2679 default: 2680 err = -ENOPROTOOPT; 2681 break; 2682 } 2683 2684 release_sock(sk); 2685 return err; 2686} 2687 2688int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval, 2689 unsigned int optlen) 2690{ 2691 const struct inet_connection_sock *icsk = inet_csk(sk); 2692 2693 if (level != SOL_TCP) 2694 return icsk->icsk_af_ops->setsockopt(sk, level, optname, 2695 optval, optlen); 2696 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2697} 2698EXPORT_SYMBOL(tcp_setsockopt); 2699 2700#ifdef CONFIG_COMPAT 2701int compat_tcp_setsockopt(struct sock *sk, int level, int optname, 2702 char __user *optval, unsigned int optlen) 2703{ 2704 if (level != SOL_TCP) 2705 return inet_csk_compat_setsockopt(sk, level, optname, 2706 optval, optlen); 2707 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2708} 2709EXPORT_SYMBOL(compat_tcp_setsockopt); 2710#endif 2711 2712/* Return information about state of tcp endpoint in API format. */ 2713void tcp_get_info(const struct sock *sk, struct tcp_info *info) 2714{ 2715 const struct tcp_sock *tp = tcp_sk(sk); 2716 const struct inet_connection_sock *icsk = inet_csk(sk); 2717 u32 now = tcp_time_stamp; 2718 2719 memset(info, 0, sizeof(*info)); 2720 2721 info->tcpi_state = sk->sk_state; 2722 info->tcpi_ca_state = icsk->icsk_ca_state; 2723 info->tcpi_retransmits = icsk->icsk_retransmits; 2724 info->tcpi_probes = icsk->icsk_probes_out; 2725 info->tcpi_backoff = icsk->icsk_backoff; 2726 2727 if (tp->rx_opt.tstamp_ok) 2728 info->tcpi_options |= TCPI_OPT_TIMESTAMPS; 2729 if (tcp_is_sack(tp)) 2730 info->tcpi_options |= TCPI_OPT_SACK; 2731 if (tp->rx_opt.wscale_ok) { 2732 info->tcpi_options |= TCPI_OPT_WSCALE; 2733 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale; 2734 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale; 2735 } 2736 2737 if (tp->ecn_flags & TCP_ECN_OK) 2738 info->tcpi_options |= TCPI_OPT_ECN; 2739 if (tp->ecn_flags & TCP_ECN_SEEN) 2740 info->tcpi_options |= TCPI_OPT_ECN_SEEN; 2741 2742 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto); 2743 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato); 2744 info->tcpi_snd_mss = tp->mss_cache; 2745 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss; 2746 2747 if (sk->sk_state == TCP_LISTEN) { 2748 info->tcpi_unacked = sk->sk_ack_backlog; 2749 info->tcpi_sacked = sk->sk_max_ack_backlog; 2750 } else { 2751 info->tcpi_unacked = tp->packets_out; 2752 info->tcpi_sacked = tp->sacked_out; 2753 } 2754 info->tcpi_lost = tp->lost_out; 2755 info->tcpi_retrans = tp->retrans_out; 2756 info->tcpi_fackets = tp->fackets_out; 2757 2758 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime); 2759 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime); 2760 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp); 2761 2762 info->tcpi_pmtu = icsk->icsk_pmtu_cookie; 2763 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh; 2764 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3; 2765 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2; 2766 info->tcpi_snd_ssthresh = tp->snd_ssthresh; 2767 info->tcpi_snd_cwnd = tp->snd_cwnd; 2768 info->tcpi_advmss = tp->advmss; 2769 info->tcpi_reordering = tp->reordering; 2770 2771 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3; 2772 info->tcpi_rcv_space = tp->rcvq_space.space; 2773 2774 info->tcpi_total_retrans = tp->total_retrans; 2775} 2776EXPORT_SYMBOL_GPL(tcp_get_info); 2777 2778static int do_tcp_getsockopt(struct sock *sk, int level, 2779 int optname, char __user *optval, int __user *optlen) 2780{ 2781 struct inet_connection_sock *icsk = inet_csk(sk); 2782 struct tcp_sock *tp = tcp_sk(sk); 2783 int val, len; 2784 2785 if (get_user(len, optlen)) 2786 return -EFAULT; 2787 2788 len = min_t(unsigned int, len, sizeof(int)); 2789 2790 if (len < 0) 2791 return -EINVAL; 2792 2793 switch (optname) { 2794 case TCP_MAXSEG: 2795 val = tp->mss_cache; 2796 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 2797 val = tp->rx_opt.user_mss; 2798 if (tp->repair) 2799 val = tp->rx_opt.mss_clamp; 2800 break; 2801 case TCP_NODELAY: 2802 val = !!(tp->nonagle&TCP_NAGLE_OFF); 2803 break; 2804 case TCP_CORK: 2805 val = !!(tp->nonagle&TCP_NAGLE_CORK); 2806 break; 2807 case TCP_KEEPIDLE: 2808 val = keepalive_time_when(tp) / HZ; 2809 break; 2810 case TCP_KEEPINTVL: 2811 val = keepalive_intvl_when(tp) / HZ; 2812 break; 2813 case TCP_KEEPCNT: 2814 val = keepalive_probes(tp); 2815 break; 2816 case TCP_SYNCNT: 2817 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries; 2818 break; 2819 case TCP_LINGER2: 2820 val = tp->linger2; 2821 if (val >= 0) 2822 val = (val ? : sysctl_tcp_fin_timeout) / HZ; 2823 break; 2824 case TCP_DEFER_ACCEPT: 2825 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept, 2826 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ); 2827 break; 2828 case TCP_WINDOW_CLAMP: 2829 val = tp->window_clamp; 2830 break; 2831 case TCP_INFO: { 2832 struct tcp_info info; 2833 2834 if (get_user(len, optlen)) 2835 return -EFAULT; 2836 2837 tcp_get_info(sk, &info); 2838 2839 len = min_t(unsigned int, len, sizeof(info)); 2840 if (put_user(len, optlen)) 2841 return -EFAULT; 2842 if (copy_to_user(optval, &info, len)) 2843 return -EFAULT; 2844 return 0; 2845 } 2846 case TCP_QUICKACK: 2847 val = !icsk->icsk_ack.pingpong; 2848 break; 2849 2850 case TCP_CONGESTION: 2851 if (get_user(len, optlen)) 2852 return -EFAULT; 2853 len = min_t(unsigned int, len, TCP_CA_NAME_MAX); 2854 if (put_user(len, optlen)) 2855 return -EFAULT; 2856 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len)) 2857 return -EFAULT; 2858 return 0; 2859 2860 case TCP_COOKIE_TRANSACTIONS: { 2861 struct tcp_cookie_transactions ctd; 2862 struct tcp_cookie_values *cvp = tp->cookie_values; 2863 2864 if (get_user(len, optlen)) 2865 return -EFAULT; 2866 if (len < sizeof(ctd)) 2867 return -EINVAL; 2868 2869 memset(&ctd, 0, sizeof(ctd)); 2870 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ? 2871 TCP_COOKIE_IN_ALWAYS : 0) 2872 | (tp->rx_opt.cookie_out_never ? 2873 TCP_COOKIE_OUT_NEVER : 0); 2874 2875 if (cvp != NULL) { 2876 ctd.tcpct_flags |= (cvp->s_data_in ? 2877 TCP_S_DATA_IN : 0) 2878 | (cvp->s_data_out ? 2879 TCP_S_DATA_OUT : 0); 2880 2881 ctd.tcpct_cookie_desired = cvp->cookie_desired; 2882 ctd.tcpct_s_data_desired = cvp->s_data_desired; 2883 2884 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0], 2885 cvp->cookie_pair_size); 2886 ctd.tcpct_used = cvp->cookie_pair_size; 2887 } 2888 2889 if (put_user(sizeof(ctd), optlen)) 2890 return -EFAULT; 2891 if (copy_to_user(optval, &ctd, sizeof(ctd))) 2892 return -EFAULT; 2893 return 0; 2894 } 2895 case TCP_THIN_LINEAR_TIMEOUTS: 2896 val = tp->thin_lto; 2897 break; 2898 case TCP_THIN_DUPACK: 2899 val = tp->thin_dupack; 2900 break; 2901 2902 case TCP_REPAIR: 2903 val = tp->repair; 2904 break; 2905 2906 case TCP_REPAIR_QUEUE: 2907 if (tp->repair) 2908 val = tp->repair_queue; 2909 else 2910 return -EINVAL; 2911 break; 2912 2913 case TCP_QUEUE_SEQ: 2914 if (tp->repair_queue == TCP_SEND_QUEUE) 2915 val = tp->write_seq; 2916 else if (tp->repair_queue == TCP_RECV_QUEUE) 2917 val = tp->rcv_nxt; 2918 else 2919 return -EINVAL; 2920 break; 2921 2922 case TCP_USER_TIMEOUT: 2923 val = jiffies_to_msecs(icsk->icsk_user_timeout); 2924 break; 2925 default: 2926 return -ENOPROTOOPT; 2927 } 2928 2929 if (put_user(len, optlen)) 2930 return -EFAULT; 2931 if (copy_to_user(optval, &val, len)) 2932 return -EFAULT; 2933 return 0; 2934} 2935 2936int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval, 2937 int __user *optlen) 2938{ 2939 struct inet_connection_sock *icsk = inet_csk(sk); 2940 2941 if (level != SOL_TCP) 2942 return icsk->icsk_af_ops->getsockopt(sk, level, optname, 2943 optval, optlen); 2944 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2945} 2946EXPORT_SYMBOL(tcp_getsockopt); 2947 2948#ifdef CONFIG_COMPAT 2949int compat_tcp_getsockopt(struct sock *sk, int level, int optname, 2950 char __user *optval, int __user *optlen) 2951{ 2952 if (level != SOL_TCP) 2953 return inet_csk_compat_getsockopt(sk, level, optname, 2954 optval, optlen); 2955 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2956} 2957EXPORT_SYMBOL(compat_tcp_getsockopt); 2958#endif 2959 2960struct sk_buff *tcp_tso_segment(struct sk_buff *skb, 2961 netdev_features_t features) 2962{ 2963 struct sk_buff *segs = ERR_PTR(-EINVAL); 2964 struct tcphdr *th; 2965 unsigned int thlen; 2966 unsigned int seq; 2967 __be32 delta; 2968 unsigned int oldlen; 2969 unsigned int mss; 2970 2971 if (!pskb_may_pull(skb, sizeof(*th))) 2972 goto out; 2973 2974 th = tcp_hdr(skb); 2975 thlen = th->doff * 4; 2976 if (thlen < sizeof(*th)) 2977 goto out; 2978 2979 if (!pskb_may_pull(skb, thlen)) 2980 goto out; 2981 2982 oldlen = (u16)~skb->len; 2983 __skb_pull(skb, thlen); 2984 2985 mss = skb_shinfo(skb)->gso_size; 2986 if (unlikely(skb->len <= mss)) 2987 goto out; 2988 2989 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) { 2990 /* Packet is from an untrusted source, reset gso_segs. */ 2991 int type = skb_shinfo(skb)->gso_type; 2992 2993 if (unlikely(type & 2994 ~(SKB_GSO_TCPV4 | 2995 SKB_GSO_DODGY | 2996 SKB_GSO_TCP_ECN | 2997 SKB_GSO_TCPV6 | 2998 0) || 2999 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))) 3000 goto out; 3001 3002 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss); 3003 3004 segs = NULL; 3005 goto out; 3006 } 3007 3008 segs = skb_segment(skb, features); 3009 if (IS_ERR(segs)) 3010 goto out; 3011 3012 delta = htonl(oldlen + (thlen + mss)); 3013 3014 skb = segs; 3015 th = tcp_hdr(skb); 3016 seq = ntohl(th->seq); 3017 3018 do { 3019 th->fin = th->psh = 0; 3020 3021 th->check = ~csum_fold((__force __wsum)((__force u32)th->check + 3022 (__force u32)delta)); 3023 if (skb->ip_summed != CHECKSUM_PARTIAL) 3024 th->check = 3025 csum_fold(csum_partial(skb_transport_header(skb), 3026 thlen, skb->csum)); 3027 3028 seq += mss; 3029 skb = skb->next; 3030 th = tcp_hdr(skb); 3031 3032 th->seq = htonl(seq); 3033 th->cwr = 0; 3034 } while (skb->next); 3035 3036 delta = htonl(oldlen + (skb->tail - skb->transport_header) + 3037 skb->data_len); 3038 th->check = ~csum_fold((__force __wsum)((__force u32)th->check + 3039 (__force u32)delta)); 3040 if (skb->ip_summed != CHECKSUM_PARTIAL) 3041 th->check = csum_fold(csum_partial(skb_transport_header(skb), 3042 thlen, skb->csum)); 3043 3044out: 3045 return segs; 3046} 3047EXPORT_SYMBOL(tcp_tso_segment); 3048 3049struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb) 3050{ 3051 struct sk_buff **pp = NULL; 3052 struct sk_buff *p; 3053 struct tcphdr *th; 3054 struct tcphdr *th2; 3055 unsigned int len; 3056 unsigned int thlen; 3057 __be32 flags; 3058 unsigned int mss = 1; 3059 unsigned int hlen; 3060 unsigned int off; 3061 int flush = 1; 3062 int i; 3063 3064 off = skb_gro_offset(skb); 3065 hlen = off + sizeof(*th); 3066 th = skb_gro_header_fast(skb, off); 3067 if (skb_gro_header_hard(skb, hlen)) { 3068 th = skb_gro_header_slow(skb, hlen, off); 3069 if (unlikely(!th)) 3070 goto out; 3071 } 3072 3073 thlen = th->doff * 4; 3074 if (thlen < sizeof(*th)) 3075 goto out; 3076 3077 hlen = off + thlen; 3078 if (skb_gro_header_hard(skb, hlen)) { 3079 th = skb_gro_header_slow(skb, hlen, off); 3080 if (unlikely(!th)) 3081 goto out; 3082 } 3083 3084 skb_gro_pull(skb, thlen); 3085 3086 len = skb_gro_len(skb); 3087 flags = tcp_flag_word(th); 3088 3089 for (; (p = *head); head = &p->next) { 3090 if (!NAPI_GRO_CB(p)->same_flow) 3091 continue; 3092 3093 th2 = tcp_hdr(p); 3094 3095 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) { 3096 NAPI_GRO_CB(p)->same_flow = 0; 3097 continue; 3098 } 3099 3100 goto found; 3101 } 3102 3103 goto out_check_final; 3104 3105found: 3106 flush = NAPI_GRO_CB(p)->flush; 3107 flush |= (__force int)(flags & TCP_FLAG_CWR); 3108 flush |= (__force int)((flags ^ tcp_flag_word(th2)) & 3109 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH)); 3110 flush |= (__force int)(th->ack_seq ^ th2->ack_seq); 3111 for (i = sizeof(*th); i < thlen; i += 4) 3112 flush |= *(u32 *)((u8 *)th + i) ^ 3113 *(u32 *)((u8 *)th2 + i); 3114 3115 mss = skb_shinfo(p)->gso_size; 3116 3117 flush |= (len - 1) >= mss; 3118 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq); 3119 3120 if (flush || skb_gro_receive(head, skb)) { 3121 mss = 1; 3122 goto out_check_final; 3123 } 3124 3125 p = *head; 3126 th2 = tcp_hdr(p); 3127 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH); 3128 3129out_check_final: 3130 flush = len < mss; 3131 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH | 3132 TCP_FLAG_RST | TCP_FLAG_SYN | 3133 TCP_FLAG_FIN)); 3134 3135 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush)) 3136 pp = head; 3137 3138out: 3139 NAPI_GRO_CB(skb)->flush |= flush; 3140 3141 return pp; 3142} 3143EXPORT_SYMBOL(tcp_gro_receive); 3144 3145int tcp_gro_complete(struct sk_buff *skb) 3146{ 3147 struct tcphdr *th = tcp_hdr(skb); 3148 3149 skb->csum_start = skb_transport_header(skb) - skb->head; 3150 skb->csum_offset = offsetof(struct tcphdr, check); 3151 skb->ip_summed = CHECKSUM_PARTIAL; 3152 3153 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count; 3154 3155 if (th->cwr) 3156 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 3157 3158 return 0; 3159} 3160EXPORT_SYMBOL(tcp_gro_complete); 3161 3162#ifdef CONFIG_TCP_MD5SIG 3163static unsigned long tcp_md5sig_users; 3164static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool; 3165static DEFINE_SPINLOCK(tcp_md5sig_pool_lock); 3166 3167static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool) 3168{ 3169 int cpu; 3170 3171 for_each_possible_cpu(cpu) { 3172 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu); 3173 3174 if (p->md5_desc.tfm) 3175 crypto_free_hash(p->md5_desc.tfm); 3176 } 3177 free_percpu(pool); 3178} 3179 3180void tcp_free_md5sig_pool(void) 3181{ 3182 struct tcp_md5sig_pool __percpu *pool = NULL; 3183 3184 spin_lock_bh(&tcp_md5sig_pool_lock); 3185 if (--tcp_md5sig_users == 0) { 3186 pool = tcp_md5sig_pool; 3187 tcp_md5sig_pool = NULL; 3188 } 3189 spin_unlock_bh(&tcp_md5sig_pool_lock); 3190 if (pool) 3191 __tcp_free_md5sig_pool(pool); 3192} 3193EXPORT_SYMBOL(tcp_free_md5sig_pool); 3194 3195static struct tcp_md5sig_pool __percpu * 3196__tcp_alloc_md5sig_pool(struct sock *sk) 3197{ 3198 int cpu; 3199 struct tcp_md5sig_pool __percpu *pool; 3200 3201 pool = alloc_percpu(struct tcp_md5sig_pool); 3202 if (!pool) 3203 return NULL; 3204 3205 for_each_possible_cpu(cpu) { 3206 struct crypto_hash *hash; 3207 3208 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC); 3209 if (!hash || IS_ERR(hash)) 3210 goto out_free; 3211 3212 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash; 3213 } 3214 return pool; 3215out_free: 3216 __tcp_free_md5sig_pool(pool); 3217 return NULL; 3218} 3219 3220struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk) 3221{ 3222 struct tcp_md5sig_pool __percpu *pool; 3223 int alloc = 0; 3224 3225retry: 3226 spin_lock_bh(&tcp_md5sig_pool_lock); 3227 pool = tcp_md5sig_pool; 3228 if (tcp_md5sig_users++ == 0) { 3229 alloc = 1; 3230 spin_unlock_bh(&tcp_md5sig_pool_lock); 3231 } else if (!pool) { 3232 tcp_md5sig_users--; 3233 spin_unlock_bh(&tcp_md5sig_pool_lock); 3234 cpu_relax(); 3235 goto retry; 3236 } else 3237 spin_unlock_bh(&tcp_md5sig_pool_lock); 3238 3239 if (alloc) { 3240 /* we cannot hold spinlock here because this may sleep. */ 3241 struct tcp_md5sig_pool __percpu *p; 3242 3243 p = __tcp_alloc_md5sig_pool(sk); 3244 spin_lock_bh(&tcp_md5sig_pool_lock); 3245 if (!p) { 3246 tcp_md5sig_users--; 3247 spin_unlock_bh(&tcp_md5sig_pool_lock); 3248 return NULL; 3249 } 3250 pool = tcp_md5sig_pool; 3251 if (pool) { 3252 /* oops, it has already been assigned. */ 3253 spin_unlock_bh(&tcp_md5sig_pool_lock); 3254 __tcp_free_md5sig_pool(p); 3255 } else { 3256 tcp_md5sig_pool = pool = p; 3257 spin_unlock_bh(&tcp_md5sig_pool_lock); 3258 } 3259 } 3260 return pool; 3261} 3262EXPORT_SYMBOL(tcp_alloc_md5sig_pool); 3263 3264 3265/** 3266 * tcp_get_md5sig_pool - get md5sig_pool for this user 3267 * 3268 * We use percpu structure, so if we succeed, we exit with preemption 3269 * and BH disabled, to make sure another thread or softirq handling 3270 * wont try to get same context. 3271 */ 3272struct tcp_md5sig_pool *tcp_get_md5sig_pool(void) 3273{ 3274 struct tcp_md5sig_pool __percpu *p; 3275 3276 local_bh_disable(); 3277 3278 spin_lock(&tcp_md5sig_pool_lock); 3279 p = tcp_md5sig_pool; 3280 if (p) 3281 tcp_md5sig_users++; 3282 spin_unlock(&tcp_md5sig_pool_lock); 3283 3284 if (p) 3285 return this_cpu_ptr(p); 3286 3287 local_bh_enable(); 3288 return NULL; 3289} 3290EXPORT_SYMBOL(tcp_get_md5sig_pool); 3291 3292void tcp_put_md5sig_pool(void) 3293{ 3294 local_bh_enable(); 3295 tcp_free_md5sig_pool(); 3296} 3297EXPORT_SYMBOL(tcp_put_md5sig_pool); 3298 3299int tcp_md5_hash_header(struct tcp_md5sig_pool *hp, 3300 const struct tcphdr *th) 3301{ 3302 struct scatterlist sg; 3303 struct tcphdr hdr; 3304 int err; 3305 3306 /* We are not allowed to change tcphdr, make a local copy */ 3307 memcpy(&hdr, th, sizeof(hdr)); 3308 hdr.check = 0; 3309 3310 /* options aren't included in the hash */ 3311 sg_init_one(&sg, &hdr, sizeof(hdr)); 3312 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr)); 3313 return err; 3314} 3315EXPORT_SYMBOL(tcp_md5_hash_header); 3316 3317int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp, 3318 const struct sk_buff *skb, unsigned int header_len) 3319{ 3320 struct scatterlist sg; 3321 const struct tcphdr *tp = tcp_hdr(skb); 3322 struct hash_desc *desc = &hp->md5_desc; 3323 unsigned int i; 3324 const unsigned int head_data_len = skb_headlen(skb) > header_len ? 3325 skb_headlen(skb) - header_len : 0; 3326 const struct skb_shared_info *shi = skb_shinfo(skb); 3327 struct sk_buff *frag_iter; 3328 3329 sg_init_table(&sg, 1); 3330 3331 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len); 3332 if (crypto_hash_update(desc, &sg, head_data_len)) 3333 return 1; 3334 3335 for (i = 0; i < shi->nr_frags; ++i) { 3336 const struct skb_frag_struct *f = &shi->frags[i]; 3337 struct page *page = skb_frag_page(f); 3338 sg_set_page(&sg, page, skb_frag_size(f), f->page_offset); 3339 if (crypto_hash_update(desc, &sg, skb_frag_size(f))) 3340 return 1; 3341 } 3342 3343 skb_walk_frags(skb, frag_iter) 3344 if (tcp_md5_hash_skb_data(hp, frag_iter, 0)) 3345 return 1; 3346 3347 return 0; 3348} 3349EXPORT_SYMBOL(tcp_md5_hash_skb_data); 3350 3351int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key) 3352{ 3353 struct scatterlist sg; 3354 3355 sg_init_one(&sg, key->key, key->keylen); 3356 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen); 3357} 3358EXPORT_SYMBOL(tcp_md5_hash_key); 3359 3360#endif 3361 3362/** 3363 * Each Responder maintains up to two secret values concurrently for 3364 * efficient secret rollover. Each secret value has 4 states: 3365 * 3366 * Generating. (tcp_secret_generating != tcp_secret_primary) 3367 * Generates new Responder-Cookies, but not yet used for primary 3368 * verification. This is a short-term state, typically lasting only 3369 * one round trip time (RTT). 3370 * 3371 * Primary. (tcp_secret_generating == tcp_secret_primary) 3372 * Used both for generation and primary verification. 3373 * 3374 * Retiring. (tcp_secret_retiring != tcp_secret_secondary) 3375 * Used for verification, until the first failure that can be 3376 * verified by the newer Generating secret. At that time, this 3377 * cookie's state is changed to Secondary, and the Generating 3378 * cookie's state is changed to Primary. This is a short-term state, 3379 * typically lasting only one round trip time (RTT). 3380 * 3381 * Secondary. (tcp_secret_retiring == tcp_secret_secondary) 3382 * Used for secondary verification, after primary verification 3383 * failures. This state lasts no more than twice the Maximum Segment 3384 * Lifetime (2MSL). Then, the secret is discarded. 3385 */ 3386struct tcp_cookie_secret { 3387 /* The secret is divided into two parts. The digest part is the 3388 * equivalent of previously hashing a secret and saving the state, 3389 * and serves as an initialization vector (IV). The message part 3390 * serves as the trailing secret. 3391 */ 3392 u32 secrets[COOKIE_WORKSPACE_WORDS]; 3393 unsigned long expires; 3394}; 3395 3396#define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL) 3397#define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2) 3398#define TCP_SECRET_LIFE (HZ * 600) 3399 3400static struct tcp_cookie_secret tcp_secret_one; 3401static struct tcp_cookie_secret tcp_secret_two; 3402 3403/* Essentially a circular list, without dynamic allocation. */ 3404static struct tcp_cookie_secret *tcp_secret_generating; 3405static struct tcp_cookie_secret *tcp_secret_primary; 3406static struct tcp_cookie_secret *tcp_secret_retiring; 3407static struct tcp_cookie_secret *tcp_secret_secondary; 3408 3409static DEFINE_SPINLOCK(tcp_secret_locker); 3410 3411/* Select a pseudo-random word in the cookie workspace. 3412 */ 3413static inline u32 tcp_cookie_work(const u32 *ws, const int n) 3414{ 3415 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])]; 3416} 3417 3418/* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed. 3419 * Called in softirq context. 3420 * Returns: 0 for success. 3421 */ 3422int tcp_cookie_generator(u32 *bakery) 3423{ 3424 unsigned long jiffy = jiffies; 3425 3426 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) { 3427 spin_lock_bh(&tcp_secret_locker); 3428 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) { 3429 /* refreshed by another */ 3430 memcpy(bakery, 3431 &tcp_secret_generating->secrets[0], 3432 COOKIE_WORKSPACE_WORDS); 3433 } else { 3434 /* still needs refreshing */ 3435 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS); 3436 3437 /* The first time, paranoia assumes that the 3438 * randomization function isn't as strong. But, 3439 * this secret initialization is delayed until 3440 * the last possible moment (packet arrival). 3441 * Although that time is observable, it is 3442 * unpredictably variable. Mash in the most 3443 * volatile clock bits available, and expire the 3444 * secret extra quickly. 3445 */ 3446 if (unlikely(tcp_secret_primary->expires == 3447 tcp_secret_secondary->expires)) { 3448 struct timespec tv; 3449 3450 getnstimeofday(&tv); 3451 bakery[COOKIE_DIGEST_WORDS+0] ^= 3452 (u32)tv.tv_nsec; 3453 3454 tcp_secret_secondary->expires = jiffy 3455 + TCP_SECRET_1MSL 3456 + (0x0f & tcp_cookie_work(bakery, 0)); 3457 } else { 3458 tcp_secret_secondary->expires = jiffy 3459 + TCP_SECRET_LIFE 3460 + (0xff & tcp_cookie_work(bakery, 1)); 3461 tcp_secret_primary->expires = jiffy 3462 + TCP_SECRET_2MSL 3463 + (0x1f & tcp_cookie_work(bakery, 2)); 3464 } 3465 memcpy(&tcp_secret_secondary->secrets[0], 3466 bakery, COOKIE_WORKSPACE_WORDS); 3467 3468 rcu_assign_pointer(tcp_secret_generating, 3469 tcp_secret_secondary); 3470 rcu_assign_pointer(tcp_secret_retiring, 3471 tcp_secret_primary); 3472 /* 3473 * Neither call_rcu() nor synchronize_rcu() needed. 3474 * Retiring data is not freed. It is replaced after 3475 * further (locked) pointer updates, and a quiet time 3476 * (minimum 1MSL, maximum LIFE - 2MSL). 3477 */ 3478 } 3479 spin_unlock_bh(&tcp_secret_locker); 3480 } else { 3481 rcu_read_lock_bh(); 3482 memcpy(bakery, 3483 &rcu_dereference(tcp_secret_generating)->secrets[0], 3484 COOKIE_WORKSPACE_WORDS); 3485 rcu_read_unlock_bh(); 3486 } 3487 return 0; 3488} 3489EXPORT_SYMBOL(tcp_cookie_generator); 3490 3491void tcp_done(struct sock *sk) 3492{ 3493 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV) 3494 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 3495 3496 tcp_set_state(sk, TCP_CLOSE); 3497 tcp_clear_xmit_timers(sk); 3498 3499 sk->sk_shutdown = SHUTDOWN_MASK; 3500 3501 if (!sock_flag(sk, SOCK_DEAD)) 3502 sk->sk_state_change(sk); 3503 else 3504 inet_csk_destroy_sock(sk); 3505} 3506EXPORT_SYMBOL_GPL(tcp_done); 3507 3508extern struct tcp_congestion_ops tcp_reno; 3509 3510static __initdata unsigned long thash_entries; 3511static int __init set_thash_entries(char *str) 3512{ 3513 if (!str) 3514 return 0; 3515 thash_entries = simple_strtoul(str, &str, 0); 3516 return 1; 3517} 3518__setup("thash_entries=", set_thash_entries); 3519 3520void tcp_init_mem(struct net *net) 3521{ 3522 unsigned long limit = nr_free_buffer_pages() / 8; 3523 limit = max(limit, 128UL); 3524 net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3; 3525 net->ipv4.sysctl_tcp_mem[1] = limit; 3526 net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2; 3527} 3528 3529void __init tcp_init(void) 3530{ 3531 struct sk_buff *skb = NULL; 3532 unsigned long limit; 3533 int max_share, cnt; 3534 unsigned int i; 3535 unsigned long jiffy = jiffies; 3536 3537 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb)); 3538 3539 percpu_counter_init(&tcp_sockets_allocated, 0); 3540 percpu_counter_init(&tcp_orphan_count, 0); 3541 tcp_hashinfo.bind_bucket_cachep = 3542 kmem_cache_create("tcp_bind_bucket", 3543 sizeof(struct inet_bind_bucket), 0, 3544 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL); 3545 3546 /* Size and allocate the main established and bind bucket 3547 * hash tables. 3548 * 3549 * The methodology is similar to that of the buffer cache. 3550 */ 3551 tcp_hashinfo.ehash = 3552 alloc_large_system_hash("TCP established", 3553 sizeof(struct inet_ehash_bucket), 3554 thash_entries, 3555 (totalram_pages >= 128 * 1024) ? 3556 13 : 15, 3557 0, 3558 NULL, 3559 &tcp_hashinfo.ehash_mask, 3560 thash_entries ? 0 : 512 * 1024); 3561 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) { 3562 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i); 3563 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i); 3564 } 3565 if (inet_ehash_locks_alloc(&tcp_hashinfo)) 3566 panic("TCP: failed to alloc ehash_locks"); 3567 tcp_hashinfo.bhash = 3568 alloc_large_system_hash("TCP bind", 3569 sizeof(struct inet_bind_hashbucket), 3570 tcp_hashinfo.ehash_mask + 1, 3571 (totalram_pages >= 128 * 1024) ? 3572 13 : 15, 3573 0, 3574 &tcp_hashinfo.bhash_size, 3575 NULL, 3576 64 * 1024); 3577 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size; 3578 for (i = 0; i < tcp_hashinfo.bhash_size; i++) { 3579 spin_lock_init(&tcp_hashinfo.bhash[i].lock); 3580 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain); 3581 } 3582 3583 3584 cnt = tcp_hashinfo.ehash_mask + 1; 3585 3586 tcp_death_row.sysctl_max_tw_buckets = cnt / 2; 3587 sysctl_tcp_max_orphans = cnt / 2; 3588 sysctl_max_syn_backlog = max(128, cnt / 256); 3589 3590 tcp_init_mem(&init_net); 3591 /* Set per-socket limits to no more than 1/128 the pressure threshold */ 3592 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7); 3593 max_share = min(4UL*1024*1024, limit); 3594 3595 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM; 3596 sysctl_tcp_wmem[1] = 16*1024; 3597 sysctl_tcp_wmem[2] = max(64*1024, max_share); 3598 3599 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM; 3600 sysctl_tcp_rmem[1] = 87380; 3601 sysctl_tcp_rmem[2] = max(87380, max_share); 3602 3603 pr_info("Hash tables configured (established %u bind %u)\n", 3604 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size); 3605 3606 tcp_register_congestion_control(&tcp_reno); 3607 3608 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets)); 3609 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets)); 3610 tcp_secret_one.expires = jiffy; /* past due */ 3611 tcp_secret_two.expires = jiffy; /* past due */ 3612 tcp_secret_generating = &tcp_secret_one; 3613 tcp_secret_primary = &tcp_secret_one; 3614 tcp_secret_retiring = &tcp_secret_two; 3615 tcp_secret_secondary = &tcp_secret_two; 3616} 3617