1 /* 2 * INET An implementation of the TCP/IP protocol suite for the LINUX 3 * operating system. INET is implemented using the BSD Socket 4 * interface as the means of communication with the user level. 5 * 6 * Implementation of the Transmission Control Protocol(TCP). 7 * 8 * Authors: Ross Biro 9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 10 * Mark Evans, <evansmp@uhura.aston.ac.uk> 11 * Corey Minyard <wf-rch!minyard@relay.EU.net> 12 * Florian La Roche, <flla@stud.uni-sb.de> 13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 14 * Linus Torvalds, <torvalds@cs.helsinki.fi> 15 * Alan Cox, <gw4pts@gw4pts.ampr.org> 16 * Matthew Dillon, <dillon@apollo.west.oic.com> 17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 18 * Jorge Cwik, <jorge@laser.satlink.net> 19 * 20 * Fixes: 21 * Alan Cox : Numerous verify_area() calls 22 * Alan Cox : Set the ACK bit on a reset 23 * Alan Cox : Stopped it crashing if it closed while 24 * sk->inuse=1 and was trying to connect 25 * (tcp_err()). 26 * Alan Cox : All icmp error handling was broken 27 * pointers passed where wrong and the 28 * socket was looked up backwards. Nobody 29 * tested any icmp error code obviously. 30 * Alan Cox : tcp_err() now handled properly. It 31 * wakes people on errors. poll 32 * behaves and the icmp error race 33 * has gone by moving it into sock.c 34 * Alan Cox : tcp_send_reset() fixed to work for 35 * everything not just packets for 36 * unknown sockets. 37 * Alan Cox : tcp option processing. 38 * Alan Cox : Reset tweaked (still not 100%) [Had 39 * syn rule wrong] 40 * Herp Rosmanith : More reset fixes 41 * Alan Cox : No longer acks invalid rst frames. 42 * Acking any kind of RST is right out. 43 * Alan Cox : Sets an ignore me flag on an rst 44 * receive otherwise odd bits of prattle 45 * escape still 46 * Alan Cox : Fixed another acking RST frame bug. 47 * Should stop LAN workplace lockups. 48 * Alan Cox : Some tidyups using the new skb list 49 * facilities 50 * Alan Cox : sk->keepopen now seems to work 51 * Alan Cox : Pulls options out correctly on accepts 52 * Alan Cox : Fixed assorted sk->rqueue->next errors 53 * Alan Cox : PSH doesn't end a TCP read. Switched a 54 * bit to skb ops. 55 * Alan Cox : Tidied tcp_data to avoid a potential 56 * nasty. 57 * Alan Cox : Added some better commenting, as the 58 * tcp is hard to follow 59 * Alan Cox : Removed incorrect check for 20 * psh 60 * Michael O'Reilly : ack < copied bug fix. 61 * Johannes Stille : Misc tcp fixes (not all in yet). 62 * Alan Cox : FIN with no memory -> CRASH 63 * Alan Cox : Added socket option proto entries. 64 * Also added awareness of them to accept. 65 * Alan Cox : Added TCP options (SOL_TCP) 66 * Alan Cox : Switched wakeup calls to callbacks, 67 * so the kernel can layer network 68 * sockets. 69 * Alan Cox : Use ip_tos/ip_ttl settings. 70 * Alan Cox : Handle FIN (more) properly (we hope). 71 * Alan Cox : RST frames sent on unsynchronised 72 * state ack error. 73 * Alan Cox : Put in missing check for SYN bit. 74 * Alan Cox : Added tcp_select_window() aka NET2E 75 * window non shrink trick. 76 * Alan Cox : Added a couple of small NET2E timer 77 * fixes 78 * Charles Hedrick : TCP fixes 79 * Toomas Tamm : TCP window fixes 80 * Alan Cox : Small URG fix to rlogin ^C ack fight 81 * Charles Hedrick : Rewrote most of it to actually work 82 * Linus : Rewrote tcp_read() and URG handling 83 * completely 84 * Gerhard Koerting: Fixed some missing timer handling 85 * Matthew Dillon : Reworked TCP machine states as per RFC 86 * Gerhard Koerting: PC/TCP workarounds 87 * Adam Caldwell : Assorted timer/timing errors 88 * Matthew Dillon : Fixed another RST bug 89 * Alan Cox : Move to kernel side addressing changes. 90 * Alan Cox : Beginning work on TCP fastpathing 91 * (not yet usable) 92 * Arnt Gulbrandsen: Turbocharged tcp_check() routine. 93 * Alan Cox : TCP fast path debugging 94 * Alan Cox : Window clamping 95 * Michael Riepe : Bug in tcp_check() 96 * Matt Dillon : More TCP improvements and RST bug fixes 97 * Matt Dillon : Yet more small nasties remove from the 98 * TCP code (Be very nice to this man if 99 * tcp finally works 100%) 8) 100 * Alan Cox : BSD accept semantics. 101 * Alan Cox : Reset on closedown bug. 102 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto(). 103 * Michael Pall : Handle poll() after URG properly in 104 * all cases. 105 * Michael Pall : Undo the last fix in tcp_read_urg() 106 * (multi URG PUSH broke rlogin). 107 * Michael Pall : Fix the multi URG PUSH problem in 108 * tcp_readable(), poll() after URG 109 * works now. 110 * Michael Pall : recv(...,MSG_OOB) never blocks in the 111 * BSD api. 112 * Alan Cox : Changed the semantics of sk->socket to 113 * fix a race and a signal problem with 114 * accept() and async I/O. 115 * Alan Cox : Relaxed the rules on tcp_sendto(). 116 * Yury Shevchuk : Really fixed accept() blocking problem. 117 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for 118 * clients/servers which listen in on 119 * fixed ports. 120 * Alan Cox : Cleaned the above up and shrank it to 121 * a sensible code size. 122 * Alan Cox : Self connect lockup fix. 123 * Alan Cox : No connect to multicast. 124 * Ross Biro : Close unaccepted children on master 125 * socket close. 126 * Alan Cox : Reset tracing code. 127 * Alan Cox : Spurious resets on shutdown. 128 * Alan Cox : Giant 15 minute/60 second timer error 129 * Alan Cox : Small whoops in polling before an 130 * accept. 131 * Alan Cox : Kept the state trace facility since 132 * it's handy for debugging. 133 * Alan Cox : More reset handler fixes. 134 * Alan Cox : Started rewriting the code based on 135 * the RFC's for other useful protocol 136 * references see: Comer, KA9Q NOS, and 137 * for a reference on the difference 138 * between specifications and how BSD 139 * works see the 4.4lite source. 140 * A.N.Kuznetsov : Don't time wait on completion of tidy 141 * close. 142 * Linus Torvalds : Fin/Shutdown & copied_seq changes. 143 * Linus Torvalds : Fixed BSD port reuse to work first syn 144 * Alan Cox : Reimplemented timers as per the RFC 145 * and using multiple timers for sanity. 146 * Alan Cox : Small bug fixes, and a lot of new 147 * comments. 148 * Alan Cox : Fixed dual reader crash by locking 149 * the buffers (much like datagram.c) 150 * Alan Cox : Fixed stuck sockets in probe. A probe 151 * now gets fed up of retrying without 152 * (even a no space) answer. 153 * Alan Cox : Extracted closing code better 154 * Alan Cox : Fixed the closing state machine to 155 * resemble the RFC. 156 * Alan Cox : More 'per spec' fixes. 157 * Jorge Cwik : Even faster checksumming. 158 * Alan Cox : tcp_data() doesn't ack illegal PSH 159 * only frames. At least one pc tcp stack 160 * generates them. 161 * Alan Cox : Cache last socket. 162 * Alan Cox : Per route irtt. 163 * Matt Day : poll()->select() match BSD precisely on error 164 * Alan Cox : New buffers 165 * Marc Tamsky : Various sk->prot->retransmits and 166 * sk->retransmits misupdating fixed. 167 * Fixed tcp_write_timeout: stuck close, 168 * and TCP syn retries gets used now. 169 * Mark Yarvis : In tcp_read_wakeup(), don't send an 170 * ack if state is TCP_CLOSED. 171 * Alan Cox : Look up device on a retransmit - routes may 172 * change. Doesn't yet cope with MSS shrink right 173 * but it's a start! 174 * Marc Tamsky : Closing in closing fixes. 175 * Mike Shaver : RFC1122 verifications. 176 * Alan Cox : rcv_saddr errors. 177 * Alan Cox : Block double connect(). 178 * Alan Cox : Small hooks for enSKIP. 179 * Alexey Kuznetsov: Path MTU discovery. 180 * Alan Cox : Support soft errors. 181 * Alan Cox : Fix MTU discovery pathological case 182 * when the remote claims no mtu! 183 * Marc Tamsky : TCP_CLOSE fix. 184 * Colin (G3TNE) : Send a reset on syn ack replies in 185 * window but wrong (fixes NT lpd problems) 186 * Pedro Roque : Better TCP window handling, delayed ack. 187 * Joerg Reuter : No modification of locked buffers in 188 * tcp_do_retransmit() 189 * Eric Schenk : Changed receiver side silly window 190 * avoidance algorithm to BSD style 191 * algorithm. This doubles throughput 192 * against machines running Solaris, 193 * and seems to result in general 194 * improvement. 195 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD 196 * Willy Konynenberg : Transparent proxying support. 197 * Mike McLagan : Routing by source 198 * Keith Owens : Do proper merging with partial SKB's in 199 * tcp_do_sendmsg to avoid burstiness. 200 * Eric Schenk : Fix fast close down bug with 201 * shutdown() followed by close(). 202 * Andi Kleen : Make poll agree with SIGIO 203 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and 204 * lingertime == 0 (RFC 793 ABORT Call) 205 * Hirokazu Takahashi : Use copy_from_user() instead of 206 * csum_and_copy_from_user() if possible. 207 * 208 * This program is free software; you can redistribute it and/or 209 * modify it under the terms of the GNU General Public License 210 * as published by the Free Software Foundation; either version 211 * 2 of the License, or(at your option) any later version. 212 * 213 * Description of States: 214 * 215 * TCP_SYN_SENT sent a connection request, waiting for ack 216 * 217 * TCP_SYN_RECV received a connection request, sent ack, 218 * waiting for final ack in three-way handshake. 219 * 220 * TCP_ESTABLISHED connection established 221 * 222 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete 223 * transmission of remaining buffered data 224 * 225 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote 226 * to shutdown 227 * 228 * TCP_CLOSING both sides have shutdown but we still have 229 * data we have to finish sending 230 * 231 * TCP_TIME_WAIT timeout to catch resent junk before entering 232 * closed, can only be entered from FIN_WAIT2 233 * or CLOSING. Required because the other end 234 * may not have gotten our last ACK causing it 235 * to retransmit the data packet (which we ignore) 236 * 237 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for 238 * us to finish writing our data and to shutdown 239 * (we have to close() to move on to LAST_ACK) 240 * 241 * TCP_LAST_ACK out side has shutdown after remote has 242 * shutdown. There may still be data in our 243 * buffer that we have to finish sending 244 * 245 * TCP_CLOSE socket is finished 246 */ 247 248 #include <linux/kernel.h> 249 #include <linux/module.h> 250 #include <linux/types.h> 251 #include <linux/fcntl.h> 252 #include <linux/poll.h> 253 #include <linux/init.h> 254 #include <linux/fs.h> 255 #include <linux/skbuff.h> 256 #include <linux/scatterlist.h> 257 #include <linux/splice.h> 258 #include <linux/net.h> 259 #include <linux/socket.h> 260 #include <linux/random.h> 261 #include <linux/bootmem.h> 262 #include <linux/highmem.h> 263 #include <linux/swap.h> 264 #include <linux/cache.h> 265 #include <linux/err.h> 266 #include <linux/crypto.h> 267 #include <linux/time.h> 268 #include <linux/slab.h> 269 270 #include <net/icmp.h> 271 #include <net/tcp.h> 272 #include <net/xfrm.h> 273 #include <net/ip.h> 274 #include <net/netdma.h> 275 #include <net/sock.h> 276 277 #include <asm/uaccess.h> 278 #include <asm/ioctls.h> 279 280 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT; 281 282 struct percpu_counter tcp_orphan_count; 283 EXPORT_SYMBOL_GPL(tcp_orphan_count); 284 285 int sysctl_tcp_mem[3] __read_mostly; 286 int sysctl_tcp_wmem[3] __read_mostly; 287 int sysctl_tcp_rmem[3] __read_mostly; 288 289 EXPORT_SYMBOL(sysctl_tcp_mem); 290 EXPORT_SYMBOL(sysctl_tcp_rmem); 291 EXPORT_SYMBOL(sysctl_tcp_wmem); 292 293 atomic_t tcp_memory_allocated; /* Current allocated memory. */ 294 EXPORT_SYMBOL(tcp_memory_allocated); 295 296 /* 297 * Current number of TCP sockets. 298 */ 299 struct percpu_counter tcp_sockets_allocated; 300 EXPORT_SYMBOL(tcp_sockets_allocated); 301 302 /* 303 * TCP splice context 304 */ 305 struct 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 */ 317 int tcp_memory_pressure __read_mostly; 318 EXPORT_SYMBOL(tcp_memory_pressure); 319 320 void 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 } 327 EXPORT_SYMBOL(tcp_enter_memory_pressure); 328 329 /* Convert seconds to retransmits based on initial and max timeout */ 330 static 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 */ 350 static int retrans_to_secs(u8 retrans, int timeout, int rto_max) 351 { 352 int period = 0; 353 354 if (retrans > 0) { 355 period = timeout; 356 while (--retrans) { 357 timeout <<= 1; 358 if (timeout > rto_max) 359 timeout = rto_max; 360 period += timeout; 361 } 362 } 363 return period; 364 } 365 366 /* 367 * Wait for a TCP event. 368 * 369 * Note that we don't need to lock the socket, as the upper poll layers 370 * take care of normal races (between the test and the event) and we don't 371 * go look at any of the socket buffers directly. 372 */ 373 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait) 374 { 375 unsigned int mask; 376 struct sock *sk = sock->sk; 377 struct tcp_sock *tp = tcp_sk(sk); 378 379 sock_poll_wait(file, sk_sleep(sk), wait); 380 if (sk->sk_state == TCP_LISTEN) 381 return inet_csk_listen_poll(sk); 382 383 /* Socket is not locked. We are protected from async events 384 * by poll logic and correct handling of state changes 385 * made by other threads is impossible in any case. 386 */ 387 388 mask = 0; 389 if (sk->sk_err) 390 mask = POLLERR; 391 392 /* 393 * POLLHUP is certainly not done right. But poll() doesn't 394 * have a notion of HUP in just one direction, and for a 395 * socket the read side is more interesting. 396 * 397 * Some poll() documentation says that POLLHUP is incompatible 398 * with the POLLOUT/POLLWR flags, so somebody should check this 399 * all. But careful, it tends to be safer to return too many 400 * bits than too few, and you can easily break real applications 401 * if you don't tell them that something has hung up! 402 * 403 * Check-me. 404 * 405 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and 406 * our fs/select.c). It means that after we received EOF, 407 * poll always returns immediately, making impossible poll() on write() 408 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP 409 * if and only if shutdown has been made in both directions. 410 * Actually, it is interesting to look how Solaris and DUX 411 * solve this dilemma. I would prefer, if POLLHUP were maskable, 412 * then we could set it on SND_SHUTDOWN. BTW examples given 413 * in Stevens' books assume exactly this behaviour, it explains 414 * why POLLHUP is incompatible with POLLOUT. --ANK 415 * 416 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent 417 * blocking on fresh not-connected or disconnected socket. --ANK 418 */ 419 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE) 420 mask |= POLLHUP; 421 if (sk->sk_shutdown & RCV_SHUTDOWN) 422 mask |= POLLIN | POLLRDNORM | POLLRDHUP; 423 424 /* Connected? */ 425 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) { 426 int target = sock_rcvlowat(sk, 0, INT_MAX); 427 428 if (tp->urg_seq == tp->copied_seq && 429 !sock_flag(sk, SOCK_URGINLINE) && 430 tp->urg_data) 431 target++; 432 433 /* Potential race condition. If read of tp below will 434 * escape above sk->sk_state, we can be illegally awaken 435 * in SYN_* states. */ 436 if (tp->rcv_nxt - tp->copied_seq >= target) 437 mask |= POLLIN | POLLRDNORM; 438 439 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) { 440 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) { 441 mask |= POLLOUT | POLLWRNORM; 442 } else { /* send SIGIO later */ 443 set_bit(SOCK_ASYNC_NOSPACE, 444 &sk->sk_socket->flags); 445 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 446 447 /* Race breaker. If space is freed after 448 * wspace test but before the flags are set, 449 * IO signal will be lost. 450 */ 451 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) 452 mask |= POLLOUT | POLLWRNORM; 453 } 454 } else 455 mask |= POLLOUT | POLLWRNORM; 456 457 if (tp->urg_data & TCP_URG_VALID) 458 mask |= POLLPRI; 459 } 460 return mask; 461 } 462 EXPORT_SYMBOL(tcp_poll); 463 464 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg) 465 { 466 struct tcp_sock *tp = tcp_sk(sk); 467 int answ; 468 469 switch (cmd) { 470 case SIOCINQ: 471 if (sk->sk_state == TCP_LISTEN) 472 return -EINVAL; 473 474 lock_sock(sk); 475 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 476 answ = 0; 477 else if (sock_flag(sk, SOCK_URGINLINE) || 478 !tp->urg_data || 479 before(tp->urg_seq, tp->copied_seq) || 480 !before(tp->urg_seq, tp->rcv_nxt)) { 481 struct sk_buff *skb; 482 483 answ = tp->rcv_nxt - tp->copied_seq; 484 485 /* Subtract 1, if FIN is in queue. */ 486 skb = skb_peek_tail(&sk->sk_receive_queue); 487 if (answ && skb) 488 answ -= tcp_hdr(skb)->fin; 489 } else 490 answ = tp->urg_seq - tp->copied_seq; 491 release_sock(sk); 492 break; 493 case SIOCATMARK: 494 answ = tp->urg_data && tp->urg_seq == tp->copied_seq; 495 break; 496 case SIOCOUTQ: 497 if (sk->sk_state == TCP_LISTEN) 498 return -EINVAL; 499 500 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 501 answ = 0; 502 else 503 answ = tp->write_seq - tp->snd_una; 504 break; 505 default: 506 return -ENOIOCTLCMD; 507 } 508 509 return put_user(answ, (int __user *)arg); 510 } 511 EXPORT_SYMBOL(tcp_ioctl); 512 513 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb) 514 { 515 TCP_SKB_CB(skb)->flags |= TCPHDR_PSH; 516 tp->pushed_seq = tp->write_seq; 517 } 518 519 static inline int forced_push(struct tcp_sock *tp) 520 { 521 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1)); 522 } 523 524 static inline void skb_entail(struct sock *sk, struct sk_buff *skb) 525 { 526 struct tcp_sock *tp = tcp_sk(sk); 527 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 528 529 skb->csum = 0; 530 tcb->seq = tcb->end_seq = tp->write_seq; 531 tcb->flags = TCPHDR_ACK; 532 tcb->sacked = 0; 533 skb_header_release(skb); 534 tcp_add_write_queue_tail(sk, skb); 535 sk->sk_wmem_queued += skb->truesize; 536 sk_mem_charge(sk, skb->truesize); 537 if (tp->nonagle & TCP_NAGLE_PUSH) 538 tp->nonagle &= ~TCP_NAGLE_PUSH; 539 } 540 541 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags) 542 { 543 if (flags & MSG_OOB) 544 tp->snd_up = tp->write_seq; 545 } 546 547 static inline void tcp_push(struct sock *sk, int flags, int mss_now, 548 int nonagle) 549 { 550 if (tcp_send_head(sk)) { 551 struct tcp_sock *tp = tcp_sk(sk); 552 553 if (!(flags & MSG_MORE) || forced_push(tp)) 554 tcp_mark_push(tp, tcp_write_queue_tail(sk)); 555 556 tcp_mark_urg(tp, flags); 557 __tcp_push_pending_frames(sk, mss_now, 558 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle); 559 } 560 } 561 562 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, 563 unsigned int offset, size_t len) 564 { 565 struct tcp_splice_state *tss = rd_desc->arg.data; 566 int ret; 567 568 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len), 569 tss->flags); 570 if (ret > 0) 571 rd_desc->count -= ret; 572 return ret; 573 } 574 575 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss) 576 { 577 /* Store TCP splice context information in read_descriptor_t. */ 578 read_descriptor_t rd_desc = { 579 .arg.data = tss, 580 .count = tss->len, 581 }; 582 583 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv); 584 } 585 586 /** 587 * tcp_splice_read - splice data from TCP socket to a pipe 588 * @sock: socket to splice from 589 * @ppos: position (not valid) 590 * @pipe: pipe to splice to 591 * @len: number of bytes to splice 592 * @flags: splice modifier flags 593 * 594 * Description: 595 * Will read pages from given socket and fill them into a pipe. 596 * 597 **/ 598 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos, 599 struct pipe_inode_info *pipe, size_t len, 600 unsigned int flags) 601 { 602 struct sock *sk = sock->sk; 603 struct tcp_splice_state tss = { 604 .pipe = pipe, 605 .len = len, 606 .flags = flags, 607 }; 608 long timeo; 609 ssize_t spliced; 610 int ret; 611 612 sock_rps_record_flow(sk); 613 /* 614 * We can't seek on a socket input 615 */ 616 if (unlikely(*ppos)) 617 return -ESPIPE; 618 619 ret = spliced = 0; 620 621 lock_sock(sk); 622 623 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK); 624 while (tss.len) { 625 ret = __tcp_splice_read(sk, &tss); 626 if (ret < 0) 627 break; 628 else if (!ret) { 629 if (spliced) 630 break; 631 if (sock_flag(sk, SOCK_DONE)) 632 break; 633 if (sk->sk_err) { 634 ret = sock_error(sk); 635 break; 636 } 637 if (sk->sk_shutdown & RCV_SHUTDOWN) 638 break; 639 if (sk->sk_state == TCP_CLOSE) { 640 /* 641 * This occurs when user tries to read 642 * from never connected socket. 643 */ 644 if (!sock_flag(sk, SOCK_DONE)) 645 ret = -ENOTCONN; 646 break; 647 } 648 if (!timeo) { 649 ret = -EAGAIN; 650 break; 651 } 652 sk_wait_data(sk, &timeo); 653 if (signal_pending(current)) { 654 ret = sock_intr_errno(timeo); 655 break; 656 } 657 continue; 658 } 659 tss.len -= ret; 660 spliced += ret; 661 662 if (!timeo) 663 break; 664 release_sock(sk); 665 lock_sock(sk); 666 667 if (sk->sk_err || sk->sk_state == TCP_CLOSE || 668 (sk->sk_shutdown & RCV_SHUTDOWN) || 669 signal_pending(current)) 670 break; 671 } 672 673 release_sock(sk); 674 675 if (spliced) 676 return spliced; 677 678 return ret; 679 } 680 EXPORT_SYMBOL(tcp_splice_read); 681 682 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp) 683 { 684 struct sk_buff *skb; 685 686 /* The TCP header must be at least 32-bit aligned. */ 687 size = ALIGN(size, 4); 688 689 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp); 690 if (skb) { 691 if (sk_wmem_schedule(sk, skb->truesize)) { 692 /* 693 * Make sure that we have exactly size bytes 694 * available to the caller, no more, no less. 695 */ 696 skb_reserve(skb, skb_tailroom(skb) - size); 697 return skb; 698 } 699 __kfree_skb(skb); 700 } else { 701 sk->sk_prot->enter_memory_pressure(sk); 702 sk_stream_moderate_sndbuf(sk); 703 } 704 return NULL; 705 } 706 707 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now, 708 int large_allowed) 709 { 710 struct tcp_sock *tp = tcp_sk(sk); 711 u32 xmit_size_goal, old_size_goal; 712 713 xmit_size_goal = mss_now; 714 715 if (large_allowed && sk_can_gso(sk)) { 716 xmit_size_goal = ((sk->sk_gso_max_size - 1) - 717 inet_csk(sk)->icsk_af_ops->net_header_len - 718 inet_csk(sk)->icsk_ext_hdr_len - 719 tp->tcp_header_len); 720 721 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal); 722 723 /* We try hard to avoid divides here */ 724 old_size_goal = tp->xmit_size_goal_segs * mss_now; 725 726 if (likely(old_size_goal <= xmit_size_goal && 727 old_size_goal + mss_now > xmit_size_goal)) { 728 xmit_size_goal = old_size_goal; 729 } else { 730 tp->xmit_size_goal_segs = xmit_size_goal / mss_now; 731 xmit_size_goal = tp->xmit_size_goal_segs * mss_now; 732 } 733 } 734 735 return max(xmit_size_goal, mss_now); 736 } 737 738 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags) 739 { 740 int mss_now; 741 742 mss_now = tcp_current_mss(sk); 743 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB)); 744 745 return mss_now; 746 } 747 748 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset, 749 size_t psize, int flags) 750 { 751 struct tcp_sock *tp = tcp_sk(sk); 752 int mss_now, size_goal; 753 int err; 754 ssize_t copied; 755 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 756 757 /* Wait for a connection to finish. */ 758 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 759 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0) 760 goto out_err; 761 762 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags); 763 764 mss_now = tcp_send_mss(sk, &size_goal, flags); 765 copied = 0; 766 767 err = -EPIPE; 768 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 769 goto out_err; 770 771 while (psize > 0) { 772 struct sk_buff *skb = tcp_write_queue_tail(sk); 773 struct page *page = pages[poffset / PAGE_SIZE]; 774 int copy, i, can_coalesce; 775 int offset = poffset % PAGE_SIZE; 776 int size = min_t(size_t, psize, PAGE_SIZE - offset); 777 778 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) { 779 new_segment: 780 if (!sk_stream_memory_free(sk)) 781 goto wait_for_sndbuf; 782 783 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation); 784 if (!skb) 785 goto wait_for_memory; 786 787 skb_entail(sk, skb); 788 copy = size_goal; 789 } 790 791 if (copy > size) 792 copy = size; 793 794 i = skb_shinfo(skb)->nr_frags; 795 can_coalesce = skb_can_coalesce(skb, i, page, offset); 796 if (!can_coalesce && i >= MAX_SKB_FRAGS) { 797 tcp_mark_push(tp, skb); 798 goto new_segment; 799 } 800 if (!sk_wmem_schedule(sk, copy)) 801 goto wait_for_memory; 802 803 if (can_coalesce) { 804 skb_shinfo(skb)->frags[i - 1].size += copy; 805 } else { 806 get_page(page); 807 skb_fill_page_desc(skb, i, page, offset, copy); 808 } 809 810 skb->len += copy; 811 skb->data_len += copy; 812 skb->truesize += copy; 813 sk->sk_wmem_queued += copy; 814 sk_mem_charge(sk, copy); 815 skb->ip_summed = CHECKSUM_PARTIAL; 816 tp->write_seq += copy; 817 TCP_SKB_CB(skb)->end_seq += copy; 818 skb_shinfo(skb)->gso_segs = 0; 819 820 if (!copied) 821 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH; 822 823 copied += copy; 824 poffset += copy; 825 if (!(psize -= copy)) 826 goto out; 827 828 if (skb->len < size_goal || (flags & MSG_OOB)) 829 continue; 830 831 if (forced_push(tp)) { 832 tcp_mark_push(tp, skb); 833 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 834 } else if (skb == tcp_send_head(sk)) 835 tcp_push_one(sk, mss_now); 836 continue; 837 838 wait_for_sndbuf: 839 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 840 wait_for_memory: 841 if (copied) 842 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH); 843 844 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0) 845 goto do_error; 846 847 mss_now = tcp_send_mss(sk, &size_goal, flags); 848 } 849 850 out: 851 if (copied) 852 tcp_push(sk, flags, mss_now, tp->nonagle); 853 return copied; 854 855 do_error: 856 if (copied) 857 goto out; 858 out_err: 859 return sk_stream_error(sk, flags, err); 860 } 861 862 int tcp_sendpage(struct sock *sk, struct page *page, int offset, 863 size_t size, int flags) 864 { 865 ssize_t res; 866 867 if (!(sk->sk_route_caps & NETIF_F_SG) || 868 !(sk->sk_route_caps & NETIF_F_ALL_CSUM)) 869 return sock_no_sendpage(sk->sk_socket, page, offset, size, 870 flags); 871 872 lock_sock(sk); 873 TCP_CHECK_TIMER(sk); 874 res = do_tcp_sendpages(sk, &page, offset, size, flags); 875 TCP_CHECK_TIMER(sk); 876 release_sock(sk); 877 return res; 878 } 879 EXPORT_SYMBOL(tcp_sendpage); 880 881 #define TCP_PAGE(sk) (sk->sk_sndmsg_page) 882 #define TCP_OFF(sk) (sk->sk_sndmsg_off) 883 884 static inline int select_size(struct sock *sk, int sg) 885 { 886 struct tcp_sock *tp = tcp_sk(sk); 887 int tmp = tp->mss_cache; 888 889 if (sg) { 890 if (sk_can_gso(sk)) 891 tmp = 0; 892 else { 893 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER); 894 895 if (tmp >= pgbreak && 896 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE) 897 tmp = pgbreak; 898 } 899 } 900 901 return tmp; 902 } 903 904 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg, 905 size_t size) 906 { 907 struct iovec *iov; 908 struct tcp_sock *tp = tcp_sk(sk); 909 struct sk_buff *skb; 910 int iovlen, flags; 911 int mss_now, size_goal; 912 int sg, err, copied; 913 long timeo; 914 915 lock_sock(sk); 916 TCP_CHECK_TIMER(sk); 917 918 flags = msg->msg_flags; 919 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 920 921 /* Wait for a connection to finish. */ 922 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 923 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0) 924 goto out_err; 925 926 /* This should be in poll */ 927 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags); 928 929 mss_now = tcp_send_mss(sk, &size_goal, flags); 930 931 /* Ok commence sending. */ 932 iovlen = msg->msg_iovlen; 933 iov = msg->msg_iov; 934 copied = 0; 935 936 err = -EPIPE; 937 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 938 goto out_err; 939 940 sg = sk->sk_route_caps & NETIF_F_SG; 941 942 while (--iovlen >= 0) { 943 int seglen = iov->iov_len; 944 unsigned char __user *from = iov->iov_base; 945 946 iov++; 947 948 while (seglen > 0) { 949 int copy = 0; 950 int max = size_goal; 951 952 skb = tcp_write_queue_tail(sk); 953 if (tcp_send_head(sk)) { 954 if (skb->ip_summed == CHECKSUM_NONE) 955 max = mss_now; 956 copy = max - skb->len; 957 } 958 959 if (copy <= 0) { 960 new_segment: 961 /* Allocate new segment. If the interface is SG, 962 * allocate skb fitting to single page. 963 */ 964 if (!sk_stream_memory_free(sk)) 965 goto wait_for_sndbuf; 966 967 skb = sk_stream_alloc_skb(sk, 968 select_size(sk, sg), 969 sk->sk_allocation); 970 if (!skb) 971 goto wait_for_memory; 972 973 /* 974 * Check whether we can use HW checksum. 975 */ 976 if (sk->sk_route_caps & NETIF_F_ALL_CSUM) 977 skb->ip_summed = CHECKSUM_PARTIAL; 978 979 skb_entail(sk, skb); 980 copy = size_goal; 981 max = size_goal; 982 } 983 984 /* Try to append data to the end of skb. */ 985 if (copy > seglen) 986 copy = seglen; 987 988 /* Where to copy to? */ 989 if (skb_tailroom(skb) > 0) { 990 /* We have some space in skb head. Superb! */ 991 if (copy > skb_tailroom(skb)) 992 copy = skb_tailroom(skb); 993 if ((err = skb_add_data(skb, from, copy)) != 0) 994 goto do_fault; 995 } else { 996 int merge = 0; 997 int i = skb_shinfo(skb)->nr_frags; 998 struct page *page = TCP_PAGE(sk); 999 int off = TCP_OFF(sk); 1000 1001 if (skb_can_coalesce(skb, i, page, off) && 1002 off != PAGE_SIZE) { 1003 /* We can extend the last page 1004 * fragment. */ 1005 merge = 1; 1006 } else if (i == MAX_SKB_FRAGS || !sg) { 1007 /* Need to add new fragment and cannot 1008 * do this because interface is non-SG, 1009 * or because all the page slots are 1010 * busy. */ 1011 tcp_mark_push(tp, skb); 1012 goto new_segment; 1013 } else if (page) { 1014 if (off == PAGE_SIZE) { 1015 put_page(page); 1016 TCP_PAGE(sk) = page = NULL; 1017 off = 0; 1018 } 1019 } else 1020 off = 0; 1021 1022 if (copy > PAGE_SIZE - off) 1023 copy = PAGE_SIZE - off; 1024 1025 if (!sk_wmem_schedule(sk, copy)) 1026 goto wait_for_memory; 1027 1028 if (!page) { 1029 /* Allocate new cache page. */ 1030 if (!(page = sk_stream_alloc_page(sk))) 1031 goto wait_for_memory; 1032 } 1033 1034 /* Time to copy data. We are close to 1035 * the end! */ 1036 err = skb_copy_to_page(sk, from, skb, page, 1037 off, copy); 1038 if (err) { 1039 /* If this page was new, give it to the 1040 * socket so it does not get leaked. 1041 */ 1042 if (!TCP_PAGE(sk)) { 1043 TCP_PAGE(sk) = page; 1044 TCP_OFF(sk) = 0; 1045 } 1046 goto do_error; 1047 } 1048 1049 /* Update the skb. */ 1050 if (merge) { 1051 skb_shinfo(skb)->frags[i - 1].size += 1052 copy; 1053 } else { 1054 skb_fill_page_desc(skb, i, page, off, copy); 1055 if (TCP_PAGE(sk)) { 1056 get_page(page); 1057 } else if (off + copy < PAGE_SIZE) { 1058 get_page(page); 1059 TCP_PAGE(sk) = page; 1060 } 1061 } 1062 1063 TCP_OFF(sk) = off + copy; 1064 } 1065 1066 if (!copied) 1067 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH; 1068 1069 tp->write_seq += copy; 1070 TCP_SKB_CB(skb)->end_seq += copy; 1071 skb_shinfo(skb)->gso_segs = 0; 1072 1073 from += copy; 1074 copied += copy; 1075 if ((seglen -= copy) == 0 && iovlen == 0) 1076 goto out; 1077 1078 if (skb->len < max || (flags & MSG_OOB)) 1079 continue; 1080 1081 if (forced_push(tp)) { 1082 tcp_mark_push(tp, skb); 1083 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 1084 } else if (skb == tcp_send_head(sk)) 1085 tcp_push_one(sk, mss_now); 1086 continue; 1087 1088 wait_for_sndbuf: 1089 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1090 wait_for_memory: 1091 if (copied) 1092 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH); 1093 1094 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0) 1095 goto do_error; 1096 1097 mss_now = tcp_send_mss(sk, &size_goal, flags); 1098 } 1099 } 1100 1101 out: 1102 if (copied) 1103 tcp_push(sk, flags, mss_now, tp->nonagle); 1104 TCP_CHECK_TIMER(sk); 1105 release_sock(sk); 1106 return copied; 1107 1108 do_fault: 1109 if (!skb->len) { 1110 tcp_unlink_write_queue(skb, sk); 1111 /* It is the one place in all of TCP, except connection 1112 * reset, where we can be unlinking the send_head. 1113 */ 1114 tcp_check_send_head(sk, skb); 1115 sk_wmem_free_skb(sk, skb); 1116 } 1117 1118 do_error: 1119 if (copied) 1120 goto out; 1121 out_err: 1122 err = sk_stream_error(sk, flags, err); 1123 TCP_CHECK_TIMER(sk); 1124 release_sock(sk); 1125 return err; 1126 } 1127 EXPORT_SYMBOL(tcp_sendmsg); 1128 1129 /* 1130 * Handle reading urgent data. BSD has very simple semantics for 1131 * this, no blocking and very strange errors 8) 1132 */ 1133 1134 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags) 1135 { 1136 struct tcp_sock *tp = tcp_sk(sk); 1137 1138 /* No URG data to read. */ 1139 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data || 1140 tp->urg_data == TCP_URG_READ) 1141 return -EINVAL; /* Yes this is right ! */ 1142 1143 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE)) 1144 return -ENOTCONN; 1145 1146 if (tp->urg_data & TCP_URG_VALID) { 1147 int err = 0; 1148 char c = tp->urg_data; 1149 1150 if (!(flags & MSG_PEEK)) 1151 tp->urg_data = TCP_URG_READ; 1152 1153 /* Read urgent data. */ 1154 msg->msg_flags |= MSG_OOB; 1155 1156 if (len > 0) { 1157 if (!(flags & MSG_TRUNC)) 1158 err = memcpy_toiovec(msg->msg_iov, &c, 1); 1159 len = 1; 1160 } else 1161 msg->msg_flags |= MSG_TRUNC; 1162 1163 return err ? -EFAULT : len; 1164 } 1165 1166 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 1167 return 0; 1168 1169 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and 1170 * the available implementations agree in this case: 1171 * this call should never block, independent of the 1172 * blocking state of the socket. 1173 * Mike <pall@rz.uni-karlsruhe.de> 1174 */ 1175 return -EAGAIN; 1176 } 1177 1178 /* Clean up the receive buffer for full frames taken by the user, 1179 * then send an ACK if necessary. COPIED is the number of bytes 1180 * tcp_recvmsg has given to the user so far, it speeds up the 1181 * calculation of whether or not we must ACK for the sake of 1182 * a window update. 1183 */ 1184 void tcp_cleanup_rbuf(struct sock *sk, int copied) 1185 { 1186 struct tcp_sock *tp = tcp_sk(sk); 1187 int time_to_ack = 0; 1188 1189 #if TCP_DEBUG 1190 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 1191 1192 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq), 1193 KERN_INFO "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n", 1194 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt); 1195 #endif 1196 1197 if (inet_csk_ack_scheduled(sk)) { 1198 const struct inet_connection_sock *icsk = inet_csk(sk); 1199 /* Delayed ACKs frequently hit locked sockets during bulk 1200 * receive. */ 1201 if (icsk->icsk_ack.blocked || 1202 /* Once-per-two-segments ACK was not sent by tcp_input.c */ 1203 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss || 1204 /* 1205 * If this read emptied read buffer, we send ACK, if 1206 * connection is not bidirectional, user drained 1207 * receive buffer and there was a small segment 1208 * in queue. 1209 */ 1210 (copied > 0 && 1211 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) || 1212 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) && 1213 !icsk->icsk_ack.pingpong)) && 1214 !atomic_read(&sk->sk_rmem_alloc))) 1215 time_to_ack = 1; 1216 } 1217 1218 /* We send an ACK if we can now advertise a non-zero window 1219 * which has been raised "significantly". 1220 * 1221 * Even if window raised up to infinity, do not send window open ACK 1222 * in states, where we will not receive more. It is useless. 1223 */ 1224 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) { 1225 __u32 rcv_window_now = tcp_receive_window(tp); 1226 1227 /* Optimize, __tcp_select_window() is not cheap. */ 1228 if (2*rcv_window_now <= tp->window_clamp) { 1229 __u32 new_window = __tcp_select_window(sk); 1230 1231 /* Send ACK now, if this read freed lots of space 1232 * in our buffer. Certainly, new_window is new window. 1233 * We can advertise it now, if it is not less than current one. 1234 * "Lots" means "at least twice" here. 1235 */ 1236 if (new_window && new_window >= 2 * rcv_window_now) 1237 time_to_ack = 1; 1238 } 1239 } 1240 if (time_to_ack) 1241 tcp_send_ack(sk); 1242 } 1243 1244 static void tcp_prequeue_process(struct sock *sk) 1245 { 1246 struct sk_buff *skb; 1247 struct tcp_sock *tp = tcp_sk(sk); 1248 1249 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED); 1250 1251 /* RX process wants to run with disabled BHs, though it is not 1252 * necessary */ 1253 local_bh_disable(); 1254 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL) 1255 sk_backlog_rcv(sk, skb); 1256 local_bh_enable(); 1257 1258 /* Clear memory counter. */ 1259 tp->ucopy.memory = 0; 1260 } 1261 1262 #ifdef CONFIG_NET_DMA 1263 static void tcp_service_net_dma(struct sock *sk, bool wait) 1264 { 1265 dma_cookie_t done, used; 1266 dma_cookie_t last_issued; 1267 struct tcp_sock *tp = tcp_sk(sk); 1268 1269 if (!tp->ucopy.dma_chan) 1270 return; 1271 1272 last_issued = tp->ucopy.dma_cookie; 1273 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1274 1275 do { 1276 if (dma_async_memcpy_complete(tp->ucopy.dma_chan, 1277 last_issued, &done, 1278 &used) == DMA_SUCCESS) { 1279 /* Safe to free early-copied skbs now */ 1280 __skb_queue_purge(&sk->sk_async_wait_queue); 1281 break; 1282 } else { 1283 struct sk_buff *skb; 1284 while ((skb = skb_peek(&sk->sk_async_wait_queue)) && 1285 (dma_async_is_complete(skb->dma_cookie, done, 1286 used) == DMA_SUCCESS)) { 1287 __skb_dequeue(&sk->sk_async_wait_queue); 1288 kfree_skb(skb); 1289 } 1290 } 1291 } while (wait); 1292 } 1293 #endif 1294 1295 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off) 1296 { 1297 struct sk_buff *skb; 1298 u32 offset; 1299 1300 skb_queue_walk(&sk->sk_receive_queue, skb) { 1301 offset = seq - TCP_SKB_CB(skb)->seq; 1302 if (tcp_hdr(skb)->syn) 1303 offset--; 1304 if (offset < skb->len || tcp_hdr(skb)->fin) { 1305 *off = offset; 1306 return skb; 1307 } 1308 } 1309 return NULL; 1310 } 1311 1312 /* 1313 * This routine provides an alternative to tcp_recvmsg() for routines 1314 * that would like to handle copying from skbuffs directly in 'sendfile' 1315 * fashion. 1316 * Note: 1317 * - It is assumed that the socket was locked by the caller. 1318 * - The routine does not block. 1319 * - At present, there is no support for reading OOB data 1320 * or for 'peeking' the socket using this routine 1321 * (although both would be easy to implement). 1322 */ 1323 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, 1324 sk_read_actor_t recv_actor) 1325 { 1326 struct sk_buff *skb; 1327 struct tcp_sock *tp = tcp_sk(sk); 1328 u32 seq = tp->copied_seq; 1329 u32 offset; 1330 int copied = 0; 1331 1332 if (sk->sk_state == TCP_LISTEN) 1333 return -ENOTCONN; 1334 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) { 1335 if (offset < skb->len) { 1336 int used; 1337 size_t len; 1338 1339 len = skb->len - offset; 1340 /* Stop reading if we hit a patch of urgent data */ 1341 if (tp->urg_data) { 1342 u32 urg_offset = tp->urg_seq - seq; 1343 if (urg_offset < len) 1344 len = urg_offset; 1345 if (!len) 1346 break; 1347 } 1348 used = recv_actor(desc, skb, offset, len); 1349 if (used < 0) { 1350 if (!copied) 1351 copied = used; 1352 break; 1353 } else if (used <= len) { 1354 seq += used; 1355 copied += used; 1356 offset += used; 1357 } 1358 /* 1359 * If recv_actor drops the lock (e.g. TCP splice 1360 * receive) the skb pointer might be invalid when 1361 * getting here: tcp_collapse might have deleted it 1362 * while aggregating skbs from the socket queue. 1363 */ 1364 skb = tcp_recv_skb(sk, seq-1, &offset); 1365 if (!skb || (offset+1 != skb->len)) 1366 break; 1367 } 1368 if (tcp_hdr(skb)->fin) { 1369 sk_eat_skb(sk, skb, 0); 1370 ++seq; 1371 break; 1372 } 1373 sk_eat_skb(sk, skb, 0); 1374 if (!desc->count) 1375 break; 1376 tp->copied_seq = seq; 1377 } 1378 tp->copied_seq = seq; 1379 1380 tcp_rcv_space_adjust(sk); 1381 1382 /* Clean up data we have read: This will do ACK frames. */ 1383 if (copied > 0) 1384 tcp_cleanup_rbuf(sk, copied); 1385 return copied; 1386 } 1387 EXPORT_SYMBOL(tcp_read_sock); 1388 1389 /* 1390 * This routine copies from a sock struct into the user buffer. 1391 * 1392 * Technical note: in 2.3 we work on _locked_ socket, so that 1393 * tricks with *seq access order and skb->users are not required. 1394 * Probably, code can be easily improved even more. 1395 */ 1396 1397 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg, 1398 size_t len, int nonblock, int flags, int *addr_len) 1399 { 1400 struct tcp_sock *tp = tcp_sk(sk); 1401 int copied = 0; 1402 u32 peek_seq; 1403 u32 *seq; 1404 unsigned long used; 1405 int err; 1406 int target; /* Read at least this many bytes */ 1407 long timeo; 1408 struct task_struct *user_recv = NULL; 1409 int copied_early = 0; 1410 struct sk_buff *skb; 1411 u32 urg_hole = 0; 1412 1413 lock_sock(sk); 1414 1415 TCP_CHECK_TIMER(sk); 1416 1417 err = -ENOTCONN; 1418 if (sk->sk_state == TCP_LISTEN) 1419 goto out; 1420 1421 timeo = sock_rcvtimeo(sk, nonblock); 1422 1423 /* Urgent data needs to be handled specially. */ 1424 if (flags & MSG_OOB) 1425 goto recv_urg; 1426 1427 seq = &tp->copied_seq; 1428 if (flags & MSG_PEEK) { 1429 peek_seq = tp->copied_seq; 1430 seq = &peek_seq; 1431 } 1432 1433 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 1434 1435 #ifdef CONFIG_NET_DMA 1436 tp->ucopy.dma_chan = NULL; 1437 preempt_disable(); 1438 skb = skb_peek_tail(&sk->sk_receive_queue); 1439 { 1440 int available = 0; 1441 1442 if (skb) 1443 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq); 1444 if ((available < target) && 1445 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) && 1446 !sysctl_tcp_low_latency && 1447 dma_find_channel(DMA_MEMCPY)) { 1448 preempt_enable_no_resched(); 1449 tp->ucopy.pinned_list = 1450 dma_pin_iovec_pages(msg->msg_iov, len); 1451 } else { 1452 preempt_enable_no_resched(); 1453 } 1454 } 1455 #endif 1456 1457 do { 1458 u32 offset; 1459 1460 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */ 1461 if (tp->urg_data && tp->urg_seq == *seq) { 1462 if (copied) 1463 break; 1464 if (signal_pending(current)) { 1465 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN; 1466 break; 1467 } 1468 } 1469 1470 /* Next get a buffer. */ 1471 1472 skb_queue_walk(&sk->sk_receive_queue, skb) { 1473 /* Now that we have two receive queues this 1474 * shouldn't happen. 1475 */ 1476 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq), 1477 KERN_INFO "recvmsg bug: copied %X " 1478 "seq %X rcvnxt %X fl %X\n", *seq, 1479 TCP_SKB_CB(skb)->seq, tp->rcv_nxt, 1480 flags)) 1481 break; 1482 1483 offset = *seq - TCP_SKB_CB(skb)->seq; 1484 if (tcp_hdr(skb)->syn) 1485 offset--; 1486 if (offset < skb->len) 1487 goto found_ok_skb; 1488 if (tcp_hdr(skb)->fin) 1489 goto found_fin_ok; 1490 WARN(!(flags & MSG_PEEK), KERN_INFO "recvmsg bug 2: " 1491 "copied %X seq %X rcvnxt %X fl %X\n", 1492 *seq, TCP_SKB_CB(skb)->seq, 1493 tp->rcv_nxt, flags); 1494 } 1495 1496 /* Well, if we have backlog, try to process it now yet. */ 1497 1498 if (copied >= target && !sk->sk_backlog.tail) 1499 break; 1500 1501 if (copied) { 1502 if (sk->sk_err || 1503 sk->sk_state == TCP_CLOSE || 1504 (sk->sk_shutdown & RCV_SHUTDOWN) || 1505 !timeo || 1506 signal_pending(current)) 1507 break; 1508 } else { 1509 if (sock_flag(sk, SOCK_DONE)) 1510 break; 1511 1512 if (sk->sk_err) { 1513 copied = sock_error(sk); 1514 break; 1515 } 1516 1517 if (sk->sk_shutdown & RCV_SHUTDOWN) 1518 break; 1519 1520 if (sk->sk_state == TCP_CLOSE) { 1521 if (!sock_flag(sk, SOCK_DONE)) { 1522 /* This occurs when user tries to read 1523 * from never connected socket. 1524 */ 1525 copied = -ENOTCONN; 1526 break; 1527 } 1528 break; 1529 } 1530 1531 if (!timeo) { 1532 copied = -EAGAIN; 1533 break; 1534 } 1535 1536 if (signal_pending(current)) { 1537 copied = sock_intr_errno(timeo); 1538 break; 1539 } 1540 } 1541 1542 tcp_cleanup_rbuf(sk, copied); 1543 1544 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) { 1545 /* Install new reader */ 1546 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) { 1547 user_recv = current; 1548 tp->ucopy.task = user_recv; 1549 tp->ucopy.iov = msg->msg_iov; 1550 } 1551 1552 tp->ucopy.len = len; 1553 1554 WARN_ON(tp->copied_seq != tp->rcv_nxt && 1555 !(flags & (MSG_PEEK | MSG_TRUNC))); 1556 1557 /* Ugly... If prequeue is not empty, we have to 1558 * process it before releasing socket, otherwise 1559 * order will be broken at second iteration. 1560 * More elegant solution is required!!! 1561 * 1562 * Look: we have the following (pseudo)queues: 1563 * 1564 * 1. packets in flight 1565 * 2. backlog 1566 * 3. prequeue 1567 * 4. receive_queue 1568 * 1569 * Each queue can be processed only if the next ones 1570 * are empty. At this point we have empty receive_queue. 1571 * But prequeue _can_ be not empty after 2nd iteration, 1572 * when we jumped to start of loop because backlog 1573 * processing added something to receive_queue. 1574 * We cannot release_sock(), because backlog contains 1575 * packets arrived _after_ prequeued ones. 1576 * 1577 * Shortly, algorithm is clear --- to process all 1578 * the queues in order. We could make it more directly, 1579 * requeueing packets from backlog to prequeue, if 1580 * is not empty. It is more elegant, but eats cycles, 1581 * unfortunately. 1582 */ 1583 if (!skb_queue_empty(&tp->ucopy.prequeue)) 1584 goto do_prequeue; 1585 1586 /* __ Set realtime policy in scheduler __ */ 1587 } 1588 1589 #ifdef CONFIG_NET_DMA 1590 if (tp->ucopy.dma_chan) 1591 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1592 #endif 1593 if (copied >= target) { 1594 /* Do not sleep, just process backlog. */ 1595 release_sock(sk); 1596 lock_sock(sk); 1597 } else 1598 sk_wait_data(sk, &timeo); 1599 1600 #ifdef CONFIG_NET_DMA 1601 tcp_service_net_dma(sk, false); /* Don't block */ 1602 tp->ucopy.wakeup = 0; 1603 #endif 1604 1605 if (user_recv) { 1606 int chunk; 1607 1608 /* __ Restore normal policy in scheduler __ */ 1609 1610 if ((chunk = len - tp->ucopy.len) != 0) { 1611 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk); 1612 len -= chunk; 1613 copied += chunk; 1614 } 1615 1616 if (tp->rcv_nxt == tp->copied_seq && 1617 !skb_queue_empty(&tp->ucopy.prequeue)) { 1618 do_prequeue: 1619 tcp_prequeue_process(sk); 1620 1621 if ((chunk = len - tp->ucopy.len) != 0) { 1622 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1623 len -= chunk; 1624 copied += chunk; 1625 } 1626 } 1627 } 1628 if ((flags & MSG_PEEK) && 1629 (peek_seq - copied - urg_hole != tp->copied_seq)) { 1630 if (net_ratelimit()) 1631 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n", 1632 current->comm, task_pid_nr(current)); 1633 peek_seq = tp->copied_seq; 1634 } 1635 continue; 1636 1637 found_ok_skb: 1638 /* Ok so how much can we use? */ 1639 used = skb->len - offset; 1640 if (len < used) 1641 used = len; 1642 1643 /* Do we have urgent data here? */ 1644 if (tp->urg_data) { 1645 u32 urg_offset = tp->urg_seq - *seq; 1646 if (urg_offset < used) { 1647 if (!urg_offset) { 1648 if (!sock_flag(sk, SOCK_URGINLINE)) { 1649 ++*seq; 1650 urg_hole++; 1651 offset++; 1652 used--; 1653 if (!used) 1654 goto skip_copy; 1655 } 1656 } else 1657 used = urg_offset; 1658 } 1659 } 1660 1661 if (!(flags & MSG_TRUNC)) { 1662 #ifdef CONFIG_NET_DMA 1663 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 1664 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY); 1665 1666 if (tp->ucopy.dma_chan) { 1667 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec( 1668 tp->ucopy.dma_chan, skb, offset, 1669 msg->msg_iov, used, 1670 tp->ucopy.pinned_list); 1671 1672 if (tp->ucopy.dma_cookie < 0) { 1673 1674 printk(KERN_ALERT "dma_cookie < 0\n"); 1675 1676 /* Exception. Bailout! */ 1677 if (!copied) 1678 copied = -EFAULT; 1679 break; 1680 } 1681 1682 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1683 1684 if ((offset + used) == skb->len) 1685 copied_early = 1; 1686 1687 } else 1688 #endif 1689 { 1690 err = skb_copy_datagram_iovec(skb, offset, 1691 msg->msg_iov, used); 1692 if (err) { 1693 /* Exception. Bailout! */ 1694 if (!copied) 1695 copied = -EFAULT; 1696 break; 1697 } 1698 } 1699 } 1700 1701 *seq += used; 1702 copied += used; 1703 len -= used; 1704 1705 tcp_rcv_space_adjust(sk); 1706 1707 skip_copy: 1708 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) { 1709 tp->urg_data = 0; 1710 tcp_fast_path_check(sk); 1711 } 1712 if (used + offset < skb->len) 1713 continue; 1714 1715 if (tcp_hdr(skb)->fin) 1716 goto found_fin_ok; 1717 if (!(flags & MSG_PEEK)) { 1718 sk_eat_skb(sk, skb, copied_early); 1719 copied_early = 0; 1720 } 1721 continue; 1722 1723 found_fin_ok: 1724 /* Process the FIN. */ 1725 ++*seq; 1726 if (!(flags & MSG_PEEK)) { 1727 sk_eat_skb(sk, skb, copied_early); 1728 copied_early = 0; 1729 } 1730 break; 1731 } while (len > 0); 1732 1733 if (user_recv) { 1734 if (!skb_queue_empty(&tp->ucopy.prequeue)) { 1735 int chunk; 1736 1737 tp->ucopy.len = copied > 0 ? len : 0; 1738 1739 tcp_prequeue_process(sk); 1740 1741 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) { 1742 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1743 len -= chunk; 1744 copied += chunk; 1745 } 1746 } 1747 1748 tp->ucopy.task = NULL; 1749 tp->ucopy.len = 0; 1750 } 1751 1752 #ifdef CONFIG_NET_DMA 1753 tcp_service_net_dma(sk, true); /* Wait for queue to drain */ 1754 tp->ucopy.dma_chan = NULL; 1755 1756 if (tp->ucopy.pinned_list) { 1757 dma_unpin_iovec_pages(tp->ucopy.pinned_list); 1758 tp->ucopy.pinned_list = NULL; 1759 } 1760 #endif 1761 1762 /* According to UNIX98, msg_name/msg_namelen are ignored 1763 * on connected socket. I was just happy when found this 8) --ANK 1764 */ 1765 1766 /* Clean up data we have read: This will do ACK frames. */ 1767 tcp_cleanup_rbuf(sk, copied); 1768 1769 TCP_CHECK_TIMER(sk); 1770 release_sock(sk); 1771 return copied; 1772 1773 out: 1774 TCP_CHECK_TIMER(sk); 1775 release_sock(sk); 1776 return err; 1777 1778 recv_urg: 1779 err = tcp_recv_urg(sk, msg, len, flags); 1780 goto out; 1781 } 1782 EXPORT_SYMBOL(tcp_recvmsg); 1783 1784 void tcp_set_state(struct sock *sk, int state) 1785 { 1786 int oldstate = sk->sk_state; 1787 1788 switch (state) { 1789 case TCP_ESTABLISHED: 1790 if (oldstate != TCP_ESTABLISHED) 1791 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1792 break; 1793 1794 case TCP_CLOSE: 1795 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED) 1796 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS); 1797 1798 sk->sk_prot->unhash(sk); 1799 if (inet_csk(sk)->icsk_bind_hash && 1800 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK)) 1801 inet_put_port(sk); 1802 /* fall through */ 1803 default: 1804 if (oldstate == TCP_ESTABLISHED) 1805 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1806 } 1807 1808 /* Change state AFTER socket is unhashed to avoid closed 1809 * socket sitting in hash tables. 1810 */ 1811 sk->sk_state = state; 1812 1813 #ifdef STATE_TRACE 1814 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]); 1815 #endif 1816 } 1817 EXPORT_SYMBOL_GPL(tcp_set_state); 1818 1819 /* 1820 * State processing on a close. This implements the state shift for 1821 * sending our FIN frame. Note that we only send a FIN for some 1822 * states. A shutdown() may have already sent the FIN, or we may be 1823 * closed. 1824 */ 1825 1826 static const unsigned char new_state[16] = { 1827 /* current state: new state: action: */ 1828 /* (Invalid) */ TCP_CLOSE, 1829 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1830 /* TCP_SYN_SENT */ TCP_CLOSE, 1831 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1832 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1, 1833 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2, 1834 /* TCP_TIME_WAIT */ TCP_CLOSE, 1835 /* TCP_CLOSE */ TCP_CLOSE, 1836 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN, 1837 /* TCP_LAST_ACK */ TCP_LAST_ACK, 1838 /* TCP_LISTEN */ TCP_CLOSE, 1839 /* TCP_CLOSING */ TCP_CLOSING, 1840 }; 1841 1842 static int tcp_close_state(struct sock *sk) 1843 { 1844 int next = (int)new_state[sk->sk_state]; 1845 int ns = next & TCP_STATE_MASK; 1846 1847 tcp_set_state(sk, ns); 1848 1849 return next & TCP_ACTION_FIN; 1850 } 1851 1852 /* 1853 * Shutdown the sending side of a connection. Much like close except 1854 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD). 1855 */ 1856 1857 void tcp_shutdown(struct sock *sk, int how) 1858 { 1859 /* We need to grab some memory, and put together a FIN, 1860 * and then put it into the queue to be sent. 1861 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92. 1862 */ 1863 if (!(how & SEND_SHUTDOWN)) 1864 return; 1865 1866 /* If we've already sent a FIN, or it's a closed state, skip this. */ 1867 if ((1 << sk->sk_state) & 1868 (TCPF_ESTABLISHED | TCPF_SYN_SENT | 1869 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) { 1870 /* Clear out any half completed packets. FIN if needed. */ 1871 if (tcp_close_state(sk)) 1872 tcp_send_fin(sk); 1873 } 1874 } 1875 EXPORT_SYMBOL(tcp_shutdown); 1876 1877 void tcp_close(struct sock *sk, long timeout) 1878 { 1879 struct sk_buff *skb; 1880 int data_was_unread = 0; 1881 int state; 1882 1883 lock_sock(sk); 1884 sk->sk_shutdown = SHUTDOWN_MASK; 1885 1886 if (sk->sk_state == TCP_LISTEN) { 1887 tcp_set_state(sk, TCP_CLOSE); 1888 1889 /* Special case. */ 1890 inet_csk_listen_stop(sk); 1891 1892 goto adjudge_to_death; 1893 } 1894 1895 /* We need to flush the recv. buffs. We do this only on the 1896 * descriptor close, not protocol-sourced closes, because the 1897 * reader process may not have drained the data yet! 1898 */ 1899 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) { 1900 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq - 1901 tcp_hdr(skb)->fin; 1902 data_was_unread += len; 1903 __kfree_skb(skb); 1904 } 1905 1906 sk_mem_reclaim(sk); 1907 1908 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */ 1909 if (sk->sk_state == TCP_CLOSE) 1910 goto adjudge_to_death; 1911 1912 /* As outlined in RFC 2525, section 2.17, we send a RST here because 1913 * data was lost. To witness the awful effects of the old behavior of 1914 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk 1915 * GET in an FTP client, suspend the process, wait for the client to 1916 * advertise a zero window, then kill -9 the FTP client, wheee... 1917 * Note: timeout is always zero in such a case. 1918 */ 1919 if (data_was_unread) { 1920 /* Unread data was tossed, zap the connection. */ 1921 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE); 1922 tcp_set_state(sk, TCP_CLOSE); 1923 tcp_send_active_reset(sk, sk->sk_allocation); 1924 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) { 1925 /* Check zero linger _after_ checking for unread data. */ 1926 sk->sk_prot->disconnect(sk, 0); 1927 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 1928 } else if (tcp_close_state(sk)) { 1929 /* We FIN if the application ate all the data before 1930 * zapping the connection. 1931 */ 1932 1933 /* RED-PEN. Formally speaking, we have broken TCP state 1934 * machine. State transitions: 1935 * 1936 * TCP_ESTABLISHED -> TCP_FIN_WAIT1 1937 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible) 1938 * TCP_CLOSE_WAIT -> TCP_LAST_ACK 1939 * 1940 * are legal only when FIN has been sent (i.e. in window), 1941 * rather than queued out of window. Purists blame. 1942 * 1943 * F.e. "RFC state" is ESTABLISHED, 1944 * if Linux state is FIN-WAIT-1, but FIN is still not sent. 1945 * 1946 * The visible declinations are that sometimes 1947 * we enter time-wait state, when it is not required really 1948 * (harmless), do not send active resets, when they are 1949 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when 1950 * they look as CLOSING or LAST_ACK for Linux) 1951 * Probably, I missed some more holelets. 1952 * --ANK 1953 */ 1954 tcp_send_fin(sk); 1955 } 1956 1957 sk_stream_wait_close(sk, timeout); 1958 1959 adjudge_to_death: 1960 state = sk->sk_state; 1961 sock_hold(sk); 1962 sock_orphan(sk); 1963 1964 /* It is the last release_sock in its life. It will remove backlog. */ 1965 release_sock(sk); 1966 1967 1968 /* Now socket is owned by kernel and we acquire BH lock 1969 to finish close. No need to check for user refs. 1970 */ 1971 local_bh_disable(); 1972 bh_lock_sock(sk); 1973 WARN_ON(sock_owned_by_user(sk)); 1974 1975 percpu_counter_inc(sk->sk_prot->orphan_count); 1976 1977 /* Have we already been destroyed by a softirq or backlog? */ 1978 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE) 1979 goto out; 1980 1981 /* This is a (useful) BSD violating of the RFC. There is a 1982 * problem with TCP as specified in that the other end could 1983 * keep a socket open forever with no application left this end. 1984 * We use a 3 minute timeout (about the same as BSD) then kill 1985 * our end. If they send after that then tough - BUT: long enough 1986 * that we won't make the old 4*rto = almost no time - whoops 1987 * reset mistake. 1988 * 1989 * Nope, it was not mistake. It is really desired behaviour 1990 * f.e. on http servers, when such sockets are useless, but 1991 * consume significant resources. Let's do it with special 1992 * linger2 option. --ANK 1993 */ 1994 1995 if (sk->sk_state == TCP_FIN_WAIT2) { 1996 struct tcp_sock *tp = tcp_sk(sk); 1997 if (tp->linger2 < 0) { 1998 tcp_set_state(sk, TCP_CLOSE); 1999 tcp_send_active_reset(sk, GFP_ATOMIC); 2000 NET_INC_STATS_BH(sock_net(sk), 2001 LINUX_MIB_TCPABORTONLINGER); 2002 } else { 2003 const int tmo = tcp_fin_time(sk); 2004 2005 if (tmo > TCP_TIMEWAIT_LEN) { 2006 inet_csk_reset_keepalive_timer(sk, 2007 tmo - TCP_TIMEWAIT_LEN); 2008 } else { 2009 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 2010 goto out; 2011 } 2012 } 2013 } 2014 if (sk->sk_state != TCP_CLOSE) { 2015 sk_mem_reclaim(sk); 2016 if (tcp_too_many_orphans(sk, 0)) { 2017 if (net_ratelimit()) 2018 printk(KERN_INFO "TCP: too many of orphaned " 2019 "sockets\n"); 2020 tcp_set_state(sk, TCP_CLOSE); 2021 tcp_send_active_reset(sk, GFP_ATOMIC); 2022 NET_INC_STATS_BH(sock_net(sk), 2023 LINUX_MIB_TCPABORTONMEMORY); 2024 } 2025 } 2026 2027 if (sk->sk_state == TCP_CLOSE) 2028 inet_csk_destroy_sock(sk); 2029 /* Otherwise, socket is reprieved until protocol close. */ 2030 2031 out: 2032 bh_unlock_sock(sk); 2033 local_bh_enable(); 2034 sock_put(sk); 2035 } 2036 EXPORT_SYMBOL(tcp_close); 2037 2038 /* These states need RST on ABORT according to RFC793 */ 2039 2040 static inline int tcp_need_reset(int state) 2041 { 2042 return (1 << state) & 2043 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 | 2044 TCPF_FIN_WAIT2 | TCPF_SYN_RECV); 2045 } 2046 2047 int tcp_disconnect(struct sock *sk, int flags) 2048 { 2049 struct inet_sock *inet = inet_sk(sk); 2050 struct inet_connection_sock *icsk = inet_csk(sk); 2051 struct tcp_sock *tp = tcp_sk(sk); 2052 int err = 0; 2053 int old_state = sk->sk_state; 2054 2055 if (old_state != TCP_CLOSE) 2056 tcp_set_state(sk, TCP_CLOSE); 2057 2058 /* ABORT function of RFC793 */ 2059 if (old_state == TCP_LISTEN) { 2060 inet_csk_listen_stop(sk); 2061 } else if (tcp_need_reset(old_state) || 2062 (tp->snd_nxt != tp->write_seq && 2063 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) { 2064 /* The last check adjusts for discrepancy of Linux wrt. RFC 2065 * states 2066 */ 2067 tcp_send_active_reset(sk, gfp_any()); 2068 sk->sk_err = ECONNRESET; 2069 } else if (old_state == TCP_SYN_SENT) 2070 sk->sk_err = ECONNRESET; 2071 2072 tcp_clear_xmit_timers(sk); 2073 __skb_queue_purge(&sk->sk_receive_queue); 2074 tcp_write_queue_purge(sk); 2075 __skb_queue_purge(&tp->out_of_order_queue); 2076 #ifdef CONFIG_NET_DMA 2077 __skb_queue_purge(&sk->sk_async_wait_queue); 2078 #endif 2079 2080 inet->inet_dport = 0; 2081 2082 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK)) 2083 inet_reset_saddr(sk); 2084 2085 sk->sk_shutdown = 0; 2086 sock_reset_flag(sk, SOCK_DONE); 2087 tp->srtt = 0; 2088 if ((tp->write_seq += tp->max_window + 2) == 0) 2089 tp->write_seq = 1; 2090 icsk->icsk_backoff = 0; 2091 tp->snd_cwnd = 2; 2092 icsk->icsk_probes_out = 0; 2093 tp->packets_out = 0; 2094 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 2095 tp->snd_cwnd_cnt = 0; 2096 tp->bytes_acked = 0; 2097 tp->window_clamp = 0; 2098 tcp_set_ca_state(sk, TCP_CA_Open); 2099 tcp_clear_retrans(tp); 2100 inet_csk_delack_init(sk); 2101 tcp_init_send_head(sk); 2102 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt)); 2103 __sk_dst_reset(sk); 2104 2105 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash); 2106 2107 sk->sk_error_report(sk); 2108 return err; 2109 } 2110 EXPORT_SYMBOL(tcp_disconnect); 2111 2112 /* 2113 * Socket option code for TCP. 2114 */ 2115 static int do_tcp_setsockopt(struct sock *sk, int level, 2116 int optname, char __user *optval, unsigned int optlen) 2117 { 2118 struct tcp_sock *tp = tcp_sk(sk); 2119 struct inet_connection_sock *icsk = inet_csk(sk); 2120 int val; 2121 int err = 0; 2122 2123 /* These are data/string values, all the others are ints */ 2124 switch (optname) { 2125 case TCP_CONGESTION: { 2126 char name[TCP_CA_NAME_MAX]; 2127 2128 if (optlen < 1) 2129 return -EINVAL; 2130 2131 val = strncpy_from_user(name, optval, 2132 min_t(long, TCP_CA_NAME_MAX-1, optlen)); 2133 if (val < 0) 2134 return -EFAULT; 2135 name[val] = 0; 2136 2137 lock_sock(sk); 2138 err = tcp_set_congestion_control(sk, name); 2139 release_sock(sk); 2140 return err; 2141 } 2142 case TCP_COOKIE_TRANSACTIONS: { 2143 struct tcp_cookie_transactions ctd; 2144 struct tcp_cookie_values *cvp = NULL; 2145 2146 if (sizeof(ctd) > optlen) 2147 return -EINVAL; 2148 if (copy_from_user(&ctd, optval, sizeof(ctd))) 2149 return -EFAULT; 2150 2151 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) || 2152 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED) 2153 return -EINVAL; 2154 2155 if (ctd.tcpct_cookie_desired == 0) { 2156 /* default to global value */ 2157 } else if ((0x1 & ctd.tcpct_cookie_desired) || 2158 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX || 2159 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) { 2160 return -EINVAL; 2161 } 2162 2163 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) { 2164 /* Supercedes all other values */ 2165 lock_sock(sk); 2166 if (tp->cookie_values != NULL) { 2167 kref_put(&tp->cookie_values->kref, 2168 tcp_cookie_values_release); 2169 tp->cookie_values = NULL; 2170 } 2171 tp->rx_opt.cookie_in_always = 0; /* false */ 2172 tp->rx_opt.cookie_out_never = 1; /* true */ 2173 release_sock(sk); 2174 return err; 2175 } 2176 2177 /* Allocate ancillary memory before locking. 2178 */ 2179 if (ctd.tcpct_used > 0 || 2180 (tp->cookie_values == NULL && 2181 (sysctl_tcp_cookie_size > 0 || 2182 ctd.tcpct_cookie_desired > 0 || 2183 ctd.tcpct_s_data_desired > 0))) { 2184 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used, 2185 GFP_KERNEL); 2186 if (cvp == NULL) 2187 return -ENOMEM; 2188 2189 kref_init(&cvp->kref); 2190 } 2191 lock_sock(sk); 2192 tp->rx_opt.cookie_in_always = 2193 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags); 2194 tp->rx_opt.cookie_out_never = 0; /* false */ 2195 2196 if (tp->cookie_values != NULL) { 2197 if (cvp != NULL) { 2198 /* Changed values are recorded by a changed 2199 * pointer, ensuring the cookie will differ, 2200 * without separately hashing each value later. 2201 */ 2202 kref_put(&tp->cookie_values->kref, 2203 tcp_cookie_values_release); 2204 } else { 2205 cvp = tp->cookie_values; 2206 } 2207 } 2208 2209 if (cvp != NULL) { 2210 cvp->cookie_desired = ctd.tcpct_cookie_desired; 2211 2212 if (ctd.tcpct_used > 0) { 2213 memcpy(cvp->s_data_payload, ctd.tcpct_value, 2214 ctd.tcpct_used); 2215 cvp->s_data_desired = ctd.tcpct_used; 2216 cvp->s_data_constant = 1; /* true */ 2217 } else { 2218 /* No constant payload data. */ 2219 cvp->s_data_desired = ctd.tcpct_s_data_desired; 2220 cvp->s_data_constant = 0; /* false */ 2221 } 2222 2223 tp->cookie_values = cvp; 2224 } 2225 release_sock(sk); 2226 return err; 2227 } 2228 default: 2229 /* fallthru */ 2230 break; 2231 } 2232 2233 if (optlen < sizeof(int)) 2234 return -EINVAL; 2235 2236 if (get_user(val, (int __user *)optval)) 2237 return -EFAULT; 2238 2239 lock_sock(sk); 2240 2241 switch (optname) { 2242 case TCP_MAXSEG: 2243 /* Values greater than interface MTU won't take effect. However 2244 * at the point when this call is done we typically don't yet 2245 * know which interface is going to be used */ 2246 if (val < 8 || val > MAX_TCP_WINDOW) { 2247 err = -EINVAL; 2248 break; 2249 } 2250 tp->rx_opt.user_mss = val; 2251 break; 2252 2253 case TCP_NODELAY: 2254 if (val) { 2255 /* TCP_NODELAY is weaker than TCP_CORK, so that 2256 * this option on corked socket is remembered, but 2257 * it is not activated until cork is cleared. 2258 * 2259 * However, when TCP_NODELAY is set we make 2260 * an explicit push, which overrides even TCP_CORK 2261 * for currently queued segments. 2262 */ 2263 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH; 2264 tcp_push_pending_frames(sk); 2265 } else { 2266 tp->nonagle &= ~TCP_NAGLE_OFF; 2267 } 2268 break; 2269 2270 case TCP_THIN_LINEAR_TIMEOUTS: 2271 if (val < 0 || val > 1) 2272 err = -EINVAL; 2273 else 2274 tp->thin_lto = val; 2275 break; 2276 2277 case TCP_THIN_DUPACK: 2278 if (val < 0 || val > 1) 2279 err = -EINVAL; 2280 else 2281 tp->thin_dupack = val; 2282 break; 2283 2284 case TCP_CORK: 2285 /* When set indicates to always queue non-full frames. 2286 * Later the user clears this option and we transmit 2287 * any pending partial frames in the queue. This is 2288 * meant to be used alongside sendfile() to get properly 2289 * filled frames when the user (for example) must write 2290 * out headers with a write() call first and then use 2291 * sendfile to send out the data parts. 2292 * 2293 * TCP_CORK can be set together with TCP_NODELAY and it is 2294 * stronger than TCP_NODELAY. 2295 */ 2296 if (val) { 2297 tp->nonagle |= TCP_NAGLE_CORK; 2298 } else { 2299 tp->nonagle &= ~TCP_NAGLE_CORK; 2300 if (tp->nonagle&TCP_NAGLE_OFF) 2301 tp->nonagle |= TCP_NAGLE_PUSH; 2302 tcp_push_pending_frames(sk); 2303 } 2304 break; 2305 2306 case TCP_KEEPIDLE: 2307 if (val < 1 || val > MAX_TCP_KEEPIDLE) 2308 err = -EINVAL; 2309 else { 2310 tp->keepalive_time = val * HZ; 2311 if (sock_flag(sk, SOCK_KEEPOPEN) && 2312 !((1 << sk->sk_state) & 2313 (TCPF_CLOSE | TCPF_LISTEN))) { 2314 u32 elapsed = keepalive_time_elapsed(tp); 2315 if (tp->keepalive_time > elapsed) 2316 elapsed = tp->keepalive_time - elapsed; 2317 else 2318 elapsed = 0; 2319 inet_csk_reset_keepalive_timer(sk, elapsed); 2320 } 2321 } 2322 break; 2323 case TCP_KEEPINTVL: 2324 if (val < 1 || val > MAX_TCP_KEEPINTVL) 2325 err = -EINVAL; 2326 else 2327 tp->keepalive_intvl = val * HZ; 2328 break; 2329 case TCP_KEEPCNT: 2330 if (val < 1 || val > MAX_TCP_KEEPCNT) 2331 err = -EINVAL; 2332 else 2333 tp->keepalive_probes = val; 2334 break; 2335 case TCP_SYNCNT: 2336 if (val < 1 || val > MAX_TCP_SYNCNT) 2337 err = -EINVAL; 2338 else 2339 icsk->icsk_syn_retries = val; 2340 break; 2341 2342 case TCP_LINGER2: 2343 if (val < 0) 2344 tp->linger2 = -1; 2345 else if (val > sysctl_tcp_fin_timeout / HZ) 2346 tp->linger2 = 0; 2347 else 2348 tp->linger2 = val * HZ; 2349 break; 2350 2351 case TCP_DEFER_ACCEPT: 2352 /* Translate value in seconds to number of retransmits */ 2353 icsk->icsk_accept_queue.rskq_defer_accept = 2354 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ, 2355 TCP_RTO_MAX / HZ); 2356 break; 2357 2358 case TCP_WINDOW_CLAMP: 2359 if (!val) { 2360 if (sk->sk_state != TCP_CLOSE) { 2361 err = -EINVAL; 2362 break; 2363 } 2364 tp->window_clamp = 0; 2365 } else 2366 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ? 2367 SOCK_MIN_RCVBUF / 2 : val; 2368 break; 2369 2370 case TCP_QUICKACK: 2371 if (!val) { 2372 icsk->icsk_ack.pingpong = 1; 2373 } else { 2374 icsk->icsk_ack.pingpong = 0; 2375 if ((1 << sk->sk_state) & 2376 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) && 2377 inet_csk_ack_scheduled(sk)) { 2378 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 2379 tcp_cleanup_rbuf(sk, 1); 2380 if (!(val & 1)) 2381 icsk->icsk_ack.pingpong = 1; 2382 } 2383 } 2384 break; 2385 2386 #ifdef CONFIG_TCP_MD5SIG 2387 case TCP_MD5SIG: 2388 /* Read the IP->Key mappings from userspace */ 2389 err = tp->af_specific->md5_parse(sk, optval, optlen); 2390 break; 2391 #endif 2392 2393 default: 2394 err = -ENOPROTOOPT; 2395 break; 2396 } 2397 2398 release_sock(sk); 2399 return err; 2400 } 2401 2402 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval, 2403 unsigned int optlen) 2404 { 2405 struct inet_connection_sock *icsk = inet_csk(sk); 2406 2407 if (level != SOL_TCP) 2408 return icsk->icsk_af_ops->setsockopt(sk, level, optname, 2409 optval, optlen); 2410 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2411 } 2412 EXPORT_SYMBOL(tcp_setsockopt); 2413 2414 #ifdef CONFIG_COMPAT 2415 int compat_tcp_setsockopt(struct sock *sk, int level, int optname, 2416 char __user *optval, unsigned int optlen) 2417 { 2418 if (level != SOL_TCP) 2419 return inet_csk_compat_setsockopt(sk, level, optname, 2420 optval, optlen); 2421 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2422 } 2423 EXPORT_SYMBOL(compat_tcp_setsockopt); 2424 #endif 2425 2426 /* Return information about state of tcp endpoint in API format. */ 2427 void tcp_get_info(struct sock *sk, struct tcp_info *info) 2428 { 2429 struct tcp_sock *tp = tcp_sk(sk); 2430 const struct inet_connection_sock *icsk = inet_csk(sk); 2431 u32 now = tcp_time_stamp; 2432 2433 memset(info, 0, sizeof(*info)); 2434 2435 info->tcpi_state = sk->sk_state; 2436 info->tcpi_ca_state = icsk->icsk_ca_state; 2437 info->tcpi_retransmits = icsk->icsk_retransmits; 2438 info->tcpi_probes = icsk->icsk_probes_out; 2439 info->tcpi_backoff = icsk->icsk_backoff; 2440 2441 if (tp->rx_opt.tstamp_ok) 2442 info->tcpi_options |= TCPI_OPT_TIMESTAMPS; 2443 if (tcp_is_sack(tp)) 2444 info->tcpi_options |= TCPI_OPT_SACK; 2445 if (tp->rx_opt.wscale_ok) { 2446 info->tcpi_options |= TCPI_OPT_WSCALE; 2447 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale; 2448 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale; 2449 } 2450 2451 if (tp->ecn_flags&TCP_ECN_OK) 2452 info->tcpi_options |= TCPI_OPT_ECN; 2453 2454 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto); 2455 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato); 2456 info->tcpi_snd_mss = tp->mss_cache; 2457 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss; 2458 2459 if (sk->sk_state == TCP_LISTEN) { 2460 info->tcpi_unacked = sk->sk_ack_backlog; 2461 info->tcpi_sacked = sk->sk_max_ack_backlog; 2462 } else { 2463 info->tcpi_unacked = tp->packets_out; 2464 info->tcpi_sacked = tp->sacked_out; 2465 } 2466 info->tcpi_lost = tp->lost_out; 2467 info->tcpi_retrans = tp->retrans_out; 2468 info->tcpi_fackets = tp->fackets_out; 2469 2470 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime); 2471 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime); 2472 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp); 2473 2474 info->tcpi_pmtu = icsk->icsk_pmtu_cookie; 2475 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh; 2476 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3; 2477 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2; 2478 info->tcpi_snd_ssthresh = tp->snd_ssthresh; 2479 info->tcpi_snd_cwnd = tp->snd_cwnd; 2480 info->tcpi_advmss = tp->advmss; 2481 info->tcpi_reordering = tp->reordering; 2482 2483 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3; 2484 info->tcpi_rcv_space = tp->rcvq_space.space; 2485 2486 info->tcpi_total_retrans = tp->total_retrans; 2487 } 2488 EXPORT_SYMBOL_GPL(tcp_get_info); 2489 2490 static int do_tcp_getsockopt(struct sock *sk, int level, 2491 int optname, char __user *optval, int __user *optlen) 2492 { 2493 struct inet_connection_sock *icsk = inet_csk(sk); 2494 struct tcp_sock *tp = tcp_sk(sk); 2495 int val, len; 2496 2497 if (get_user(len, optlen)) 2498 return -EFAULT; 2499 2500 len = min_t(unsigned int, len, sizeof(int)); 2501 2502 if (len < 0) 2503 return -EINVAL; 2504 2505 switch (optname) { 2506 case TCP_MAXSEG: 2507 val = tp->mss_cache; 2508 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 2509 val = tp->rx_opt.user_mss; 2510 break; 2511 case TCP_NODELAY: 2512 val = !!(tp->nonagle&TCP_NAGLE_OFF); 2513 break; 2514 case TCP_CORK: 2515 val = !!(tp->nonagle&TCP_NAGLE_CORK); 2516 break; 2517 case TCP_KEEPIDLE: 2518 val = keepalive_time_when(tp) / HZ; 2519 break; 2520 case TCP_KEEPINTVL: 2521 val = keepalive_intvl_when(tp) / HZ; 2522 break; 2523 case TCP_KEEPCNT: 2524 val = keepalive_probes(tp); 2525 break; 2526 case TCP_SYNCNT: 2527 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries; 2528 break; 2529 case TCP_LINGER2: 2530 val = tp->linger2; 2531 if (val >= 0) 2532 val = (val ? : sysctl_tcp_fin_timeout) / HZ; 2533 break; 2534 case TCP_DEFER_ACCEPT: 2535 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept, 2536 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ); 2537 break; 2538 case TCP_WINDOW_CLAMP: 2539 val = tp->window_clamp; 2540 break; 2541 case TCP_INFO: { 2542 struct tcp_info info; 2543 2544 if (get_user(len, optlen)) 2545 return -EFAULT; 2546 2547 tcp_get_info(sk, &info); 2548 2549 len = min_t(unsigned int, len, sizeof(info)); 2550 if (put_user(len, optlen)) 2551 return -EFAULT; 2552 if (copy_to_user(optval, &info, len)) 2553 return -EFAULT; 2554 return 0; 2555 } 2556 case TCP_QUICKACK: 2557 val = !icsk->icsk_ack.pingpong; 2558 break; 2559 2560 case TCP_CONGESTION: 2561 if (get_user(len, optlen)) 2562 return -EFAULT; 2563 len = min_t(unsigned int, len, TCP_CA_NAME_MAX); 2564 if (put_user(len, optlen)) 2565 return -EFAULT; 2566 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len)) 2567 return -EFAULT; 2568 return 0; 2569 2570 case TCP_COOKIE_TRANSACTIONS: { 2571 struct tcp_cookie_transactions ctd; 2572 struct tcp_cookie_values *cvp = tp->cookie_values; 2573 2574 if (get_user(len, optlen)) 2575 return -EFAULT; 2576 if (len < sizeof(ctd)) 2577 return -EINVAL; 2578 2579 memset(&ctd, 0, sizeof(ctd)); 2580 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ? 2581 TCP_COOKIE_IN_ALWAYS : 0) 2582 | (tp->rx_opt.cookie_out_never ? 2583 TCP_COOKIE_OUT_NEVER : 0); 2584 2585 if (cvp != NULL) { 2586 ctd.tcpct_flags |= (cvp->s_data_in ? 2587 TCP_S_DATA_IN : 0) 2588 | (cvp->s_data_out ? 2589 TCP_S_DATA_OUT : 0); 2590 2591 ctd.tcpct_cookie_desired = cvp->cookie_desired; 2592 ctd.tcpct_s_data_desired = cvp->s_data_desired; 2593 2594 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0], 2595 cvp->cookie_pair_size); 2596 ctd.tcpct_used = cvp->cookie_pair_size; 2597 } 2598 2599 if (put_user(sizeof(ctd), optlen)) 2600 return -EFAULT; 2601 if (copy_to_user(optval, &ctd, sizeof(ctd))) 2602 return -EFAULT; 2603 return 0; 2604 } 2605 case TCP_THIN_LINEAR_TIMEOUTS: 2606 val = tp->thin_lto; 2607 break; 2608 case TCP_THIN_DUPACK: 2609 val = tp->thin_dupack; 2610 break; 2611 default: 2612 return -ENOPROTOOPT; 2613 } 2614 2615 if (put_user(len, optlen)) 2616 return -EFAULT; 2617 if (copy_to_user(optval, &val, len)) 2618 return -EFAULT; 2619 return 0; 2620 } 2621 2622 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval, 2623 int __user *optlen) 2624 { 2625 struct inet_connection_sock *icsk = inet_csk(sk); 2626 2627 if (level != SOL_TCP) 2628 return icsk->icsk_af_ops->getsockopt(sk, level, optname, 2629 optval, optlen); 2630 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2631 } 2632 EXPORT_SYMBOL(tcp_getsockopt); 2633 2634 #ifdef CONFIG_COMPAT 2635 int compat_tcp_getsockopt(struct sock *sk, int level, int optname, 2636 char __user *optval, int __user *optlen) 2637 { 2638 if (level != SOL_TCP) 2639 return inet_csk_compat_getsockopt(sk, level, optname, 2640 optval, optlen); 2641 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2642 } 2643 EXPORT_SYMBOL(compat_tcp_getsockopt); 2644 #endif 2645 2646 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features) 2647 { 2648 struct sk_buff *segs = ERR_PTR(-EINVAL); 2649 struct tcphdr *th; 2650 unsigned thlen; 2651 unsigned int seq; 2652 __be32 delta; 2653 unsigned int oldlen; 2654 unsigned int mss; 2655 2656 if (!pskb_may_pull(skb, sizeof(*th))) 2657 goto out; 2658 2659 th = tcp_hdr(skb); 2660 thlen = th->doff * 4; 2661 if (thlen < sizeof(*th)) 2662 goto out; 2663 2664 if (!pskb_may_pull(skb, thlen)) 2665 goto out; 2666 2667 oldlen = (u16)~skb->len; 2668 __skb_pull(skb, thlen); 2669 2670 mss = skb_shinfo(skb)->gso_size; 2671 if (unlikely(skb->len <= mss)) 2672 goto out; 2673 2674 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) { 2675 /* Packet is from an untrusted source, reset gso_segs. */ 2676 int type = skb_shinfo(skb)->gso_type; 2677 2678 if (unlikely(type & 2679 ~(SKB_GSO_TCPV4 | 2680 SKB_GSO_DODGY | 2681 SKB_GSO_TCP_ECN | 2682 SKB_GSO_TCPV6 | 2683 0) || 2684 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))) 2685 goto out; 2686 2687 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss); 2688 2689 segs = NULL; 2690 goto out; 2691 } 2692 2693 segs = skb_segment(skb, features); 2694 if (IS_ERR(segs)) 2695 goto out; 2696 2697 delta = htonl(oldlen + (thlen + mss)); 2698 2699 skb = segs; 2700 th = tcp_hdr(skb); 2701 seq = ntohl(th->seq); 2702 2703 do { 2704 th->fin = th->psh = 0; 2705 2706 th->check = ~csum_fold((__force __wsum)((__force u32)th->check + 2707 (__force u32)delta)); 2708 if (skb->ip_summed != CHECKSUM_PARTIAL) 2709 th->check = 2710 csum_fold(csum_partial(skb_transport_header(skb), 2711 thlen, skb->csum)); 2712 2713 seq += mss; 2714 skb = skb->next; 2715 th = tcp_hdr(skb); 2716 2717 th->seq = htonl(seq); 2718 th->cwr = 0; 2719 } while (skb->next); 2720 2721 delta = htonl(oldlen + (skb->tail - skb->transport_header) + 2722 skb->data_len); 2723 th->check = ~csum_fold((__force __wsum)((__force u32)th->check + 2724 (__force u32)delta)); 2725 if (skb->ip_summed != CHECKSUM_PARTIAL) 2726 th->check = csum_fold(csum_partial(skb_transport_header(skb), 2727 thlen, skb->csum)); 2728 2729 out: 2730 return segs; 2731 } 2732 EXPORT_SYMBOL(tcp_tso_segment); 2733 2734 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb) 2735 { 2736 struct sk_buff **pp = NULL; 2737 struct sk_buff *p; 2738 struct tcphdr *th; 2739 struct tcphdr *th2; 2740 unsigned int len; 2741 unsigned int thlen; 2742 __be32 flags; 2743 unsigned int mss = 1; 2744 unsigned int hlen; 2745 unsigned int off; 2746 int flush = 1; 2747 int i; 2748 2749 off = skb_gro_offset(skb); 2750 hlen = off + sizeof(*th); 2751 th = skb_gro_header_fast(skb, off); 2752 if (skb_gro_header_hard(skb, hlen)) { 2753 th = skb_gro_header_slow(skb, hlen, off); 2754 if (unlikely(!th)) 2755 goto out; 2756 } 2757 2758 thlen = th->doff * 4; 2759 if (thlen < sizeof(*th)) 2760 goto out; 2761 2762 hlen = off + thlen; 2763 if (skb_gro_header_hard(skb, hlen)) { 2764 th = skb_gro_header_slow(skb, hlen, off); 2765 if (unlikely(!th)) 2766 goto out; 2767 } 2768 2769 skb_gro_pull(skb, thlen); 2770 2771 len = skb_gro_len(skb); 2772 flags = tcp_flag_word(th); 2773 2774 for (; (p = *head); head = &p->next) { 2775 if (!NAPI_GRO_CB(p)->same_flow) 2776 continue; 2777 2778 th2 = tcp_hdr(p); 2779 2780 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) { 2781 NAPI_GRO_CB(p)->same_flow = 0; 2782 continue; 2783 } 2784 2785 goto found; 2786 } 2787 2788 goto out_check_final; 2789 2790 found: 2791 flush = NAPI_GRO_CB(p)->flush; 2792 flush |= (__force int)(flags & TCP_FLAG_CWR); 2793 flush |= (__force int)((flags ^ tcp_flag_word(th2)) & 2794 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH)); 2795 flush |= (__force int)(th->ack_seq ^ th2->ack_seq); 2796 for (i = sizeof(*th); i < thlen; i += 4) 2797 flush |= *(u32 *)((u8 *)th + i) ^ 2798 *(u32 *)((u8 *)th2 + i); 2799 2800 mss = skb_shinfo(p)->gso_size; 2801 2802 flush |= (len - 1) >= mss; 2803 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq); 2804 2805 if (flush || skb_gro_receive(head, skb)) { 2806 mss = 1; 2807 goto out_check_final; 2808 } 2809 2810 p = *head; 2811 th2 = tcp_hdr(p); 2812 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH); 2813 2814 out_check_final: 2815 flush = len < mss; 2816 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH | 2817 TCP_FLAG_RST | TCP_FLAG_SYN | 2818 TCP_FLAG_FIN)); 2819 2820 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush)) 2821 pp = head; 2822 2823 out: 2824 NAPI_GRO_CB(skb)->flush |= flush; 2825 2826 return pp; 2827 } 2828 EXPORT_SYMBOL(tcp_gro_receive); 2829 2830 int tcp_gro_complete(struct sk_buff *skb) 2831 { 2832 struct tcphdr *th = tcp_hdr(skb); 2833 2834 skb->csum_start = skb_transport_header(skb) - skb->head; 2835 skb->csum_offset = offsetof(struct tcphdr, check); 2836 skb->ip_summed = CHECKSUM_PARTIAL; 2837 2838 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count; 2839 2840 if (th->cwr) 2841 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 2842 2843 return 0; 2844 } 2845 EXPORT_SYMBOL(tcp_gro_complete); 2846 2847 #ifdef CONFIG_TCP_MD5SIG 2848 static unsigned long tcp_md5sig_users; 2849 static struct tcp_md5sig_pool * __percpu *tcp_md5sig_pool; 2850 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock); 2851 2852 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool * __percpu *pool) 2853 { 2854 int cpu; 2855 for_each_possible_cpu(cpu) { 2856 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu); 2857 if (p) { 2858 if (p->md5_desc.tfm) 2859 crypto_free_hash(p->md5_desc.tfm); 2860 kfree(p); 2861 } 2862 } 2863 free_percpu(pool); 2864 } 2865 2866 void tcp_free_md5sig_pool(void) 2867 { 2868 struct tcp_md5sig_pool * __percpu *pool = NULL; 2869 2870 spin_lock_bh(&tcp_md5sig_pool_lock); 2871 if (--tcp_md5sig_users == 0) { 2872 pool = tcp_md5sig_pool; 2873 tcp_md5sig_pool = NULL; 2874 } 2875 spin_unlock_bh(&tcp_md5sig_pool_lock); 2876 if (pool) 2877 __tcp_free_md5sig_pool(pool); 2878 } 2879 EXPORT_SYMBOL(tcp_free_md5sig_pool); 2880 2881 static struct tcp_md5sig_pool * __percpu * 2882 __tcp_alloc_md5sig_pool(struct sock *sk) 2883 { 2884 int cpu; 2885 struct tcp_md5sig_pool * __percpu *pool; 2886 2887 pool = alloc_percpu(struct tcp_md5sig_pool *); 2888 if (!pool) 2889 return NULL; 2890 2891 for_each_possible_cpu(cpu) { 2892 struct tcp_md5sig_pool *p; 2893 struct crypto_hash *hash; 2894 2895 p = kzalloc(sizeof(*p), sk->sk_allocation); 2896 if (!p) 2897 goto out_free; 2898 *per_cpu_ptr(pool, cpu) = p; 2899 2900 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC); 2901 if (!hash || IS_ERR(hash)) 2902 goto out_free; 2903 2904 p->md5_desc.tfm = hash; 2905 } 2906 return pool; 2907 out_free: 2908 __tcp_free_md5sig_pool(pool); 2909 return NULL; 2910 } 2911 2912 struct tcp_md5sig_pool * __percpu *tcp_alloc_md5sig_pool(struct sock *sk) 2913 { 2914 struct tcp_md5sig_pool * __percpu *pool; 2915 int alloc = 0; 2916 2917 retry: 2918 spin_lock_bh(&tcp_md5sig_pool_lock); 2919 pool = tcp_md5sig_pool; 2920 if (tcp_md5sig_users++ == 0) { 2921 alloc = 1; 2922 spin_unlock_bh(&tcp_md5sig_pool_lock); 2923 } else if (!pool) { 2924 tcp_md5sig_users--; 2925 spin_unlock_bh(&tcp_md5sig_pool_lock); 2926 cpu_relax(); 2927 goto retry; 2928 } else 2929 spin_unlock_bh(&tcp_md5sig_pool_lock); 2930 2931 if (alloc) { 2932 /* we cannot hold spinlock here because this may sleep. */ 2933 struct tcp_md5sig_pool * __percpu *p; 2934 2935 p = __tcp_alloc_md5sig_pool(sk); 2936 spin_lock_bh(&tcp_md5sig_pool_lock); 2937 if (!p) { 2938 tcp_md5sig_users--; 2939 spin_unlock_bh(&tcp_md5sig_pool_lock); 2940 return NULL; 2941 } 2942 pool = tcp_md5sig_pool; 2943 if (pool) { 2944 /* oops, it has already been assigned. */ 2945 spin_unlock_bh(&tcp_md5sig_pool_lock); 2946 __tcp_free_md5sig_pool(p); 2947 } else { 2948 tcp_md5sig_pool = pool = p; 2949 spin_unlock_bh(&tcp_md5sig_pool_lock); 2950 } 2951 } 2952 return pool; 2953 } 2954 EXPORT_SYMBOL(tcp_alloc_md5sig_pool); 2955 2956 2957 /** 2958 * tcp_get_md5sig_pool - get md5sig_pool for this user 2959 * 2960 * We use percpu structure, so if we succeed, we exit with preemption 2961 * and BH disabled, to make sure another thread or softirq handling 2962 * wont try to get same context. 2963 */ 2964 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void) 2965 { 2966 struct tcp_md5sig_pool * __percpu *p; 2967 2968 local_bh_disable(); 2969 2970 spin_lock(&tcp_md5sig_pool_lock); 2971 p = tcp_md5sig_pool; 2972 if (p) 2973 tcp_md5sig_users++; 2974 spin_unlock(&tcp_md5sig_pool_lock); 2975 2976 if (p) 2977 return *this_cpu_ptr(p); 2978 2979 local_bh_enable(); 2980 return NULL; 2981 } 2982 EXPORT_SYMBOL(tcp_get_md5sig_pool); 2983 2984 void tcp_put_md5sig_pool(void) 2985 { 2986 local_bh_enable(); 2987 tcp_free_md5sig_pool(); 2988 } 2989 EXPORT_SYMBOL(tcp_put_md5sig_pool); 2990 2991 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp, 2992 struct tcphdr *th) 2993 { 2994 struct scatterlist sg; 2995 int err; 2996 2997 __sum16 old_checksum = th->check; 2998 th->check = 0; 2999 /* options aren't included in the hash */ 3000 sg_init_one(&sg, th, sizeof(struct tcphdr)); 3001 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr)); 3002 th->check = old_checksum; 3003 return err; 3004 } 3005 EXPORT_SYMBOL(tcp_md5_hash_header); 3006 3007 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp, 3008 struct sk_buff *skb, unsigned header_len) 3009 { 3010 struct scatterlist sg; 3011 const struct tcphdr *tp = tcp_hdr(skb); 3012 struct hash_desc *desc = &hp->md5_desc; 3013 unsigned i; 3014 const unsigned head_data_len = skb_headlen(skb) > header_len ? 3015 skb_headlen(skb) - header_len : 0; 3016 const struct skb_shared_info *shi = skb_shinfo(skb); 3017 struct sk_buff *frag_iter; 3018 3019 sg_init_table(&sg, 1); 3020 3021 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len); 3022 if (crypto_hash_update(desc, &sg, head_data_len)) 3023 return 1; 3024 3025 for (i = 0; i < shi->nr_frags; ++i) { 3026 const struct skb_frag_struct *f = &shi->frags[i]; 3027 sg_set_page(&sg, f->page, f->size, f->page_offset); 3028 if (crypto_hash_update(desc, &sg, f->size)) 3029 return 1; 3030 } 3031 3032 skb_walk_frags(skb, frag_iter) 3033 if (tcp_md5_hash_skb_data(hp, frag_iter, 0)) 3034 return 1; 3035 3036 return 0; 3037 } 3038 EXPORT_SYMBOL(tcp_md5_hash_skb_data); 3039 3040 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key) 3041 { 3042 struct scatterlist sg; 3043 3044 sg_init_one(&sg, key->key, key->keylen); 3045 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen); 3046 } 3047 EXPORT_SYMBOL(tcp_md5_hash_key); 3048 3049 #endif 3050 3051 /** 3052 * Each Responder maintains up to two secret values concurrently for 3053 * efficient secret rollover. Each secret value has 4 states: 3054 * 3055 * Generating. (tcp_secret_generating != tcp_secret_primary) 3056 * Generates new Responder-Cookies, but not yet used for primary 3057 * verification. This is a short-term state, typically lasting only 3058 * one round trip time (RTT). 3059 * 3060 * Primary. (tcp_secret_generating == tcp_secret_primary) 3061 * Used both for generation and primary verification. 3062 * 3063 * Retiring. (tcp_secret_retiring != tcp_secret_secondary) 3064 * Used for verification, until the first failure that can be 3065 * verified by the newer Generating secret. At that time, this 3066 * cookie's state is changed to Secondary, and the Generating 3067 * cookie's state is changed to Primary. This is a short-term state, 3068 * typically lasting only one round trip time (RTT). 3069 * 3070 * Secondary. (tcp_secret_retiring == tcp_secret_secondary) 3071 * Used for secondary verification, after primary verification 3072 * failures. This state lasts no more than twice the Maximum Segment 3073 * Lifetime (2MSL). Then, the secret is discarded. 3074 */ 3075 struct tcp_cookie_secret { 3076 /* The secret is divided into two parts. The digest part is the 3077 * equivalent of previously hashing a secret and saving the state, 3078 * and serves as an initialization vector (IV). The message part 3079 * serves as the trailing secret. 3080 */ 3081 u32 secrets[COOKIE_WORKSPACE_WORDS]; 3082 unsigned long expires; 3083 }; 3084 3085 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL) 3086 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2) 3087 #define TCP_SECRET_LIFE (HZ * 600) 3088 3089 static struct tcp_cookie_secret tcp_secret_one; 3090 static struct tcp_cookie_secret tcp_secret_two; 3091 3092 /* Essentially a circular list, without dynamic allocation. */ 3093 static struct tcp_cookie_secret *tcp_secret_generating; 3094 static struct tcp_cookie_secret *tcp_secret_primary; 3095 static struct tcp_cookie_secret *tcp_secret_retiring; 3096 static struct tcp_cookie_secret *tcp_secret_secondary; 3097 3098 static DEFINE_SPINLOCK(tcp_secret_locker); 3099 3100 /* Select a pseudo-random word in the cookie workspace. 3101 */ 3102 static inline u32 tcp_cookie_work(const u32 *ws, const int n) 3103 { 3104 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])]; 3105 } 3106 3107 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed. 3108 * Called in softirq context. 3109 * Returns: 0 for success. 3110 */ 3111 int tcp_cookie_generator(u32 *bakery) 3112 { 3113 unsigned long jiffy = jiffies; 3114 3115 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) { 3116 spin_lock_bh(&tcp_secret_locker); 3117 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) { 3118 /* refreshed by another */ 3119 memcpy(bakery, 3120 &tcp_secret_generating->secrets[0], 3121 COOKIE_WORKSPACE_WORDS); 3122 } else { 3123 /* still needs refreshing */ 3124 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS); 3125 3126 /* The first time, paranoia assumes that the 3127 * randomization function isn't as strong. But, 3128 * this secret initialization is delayed until 3129 * the last possible moment (packet arrival). 3130 * Although that time is observable, it is 3131 * unpredictably variable. Mash in the most 3132 * volatile clock bits available, and expire the 3133 * secret extra quickly. 3134 */ 3135 if (unlikely(tcp_secret_primary->expires == 3136 tcp_secret_secondary->expires)) { 3137 struct timespec tv; 3138 3139 getnstimeofday(&tv); 3140 bakery[COOKIE_DIGEST_WORDS+0] ^= 3141 (u32)tv.tv_nsec; 3142 3143 tcp_secret_secondary->expires = jiffy 3144 + TCP_SECRET_1MSL 3145 + (0x0f & tcp_cookie_work(bakery, 0)); 3146 } else { 3147 tcp_secret_secondary->expires = jiffy 3148 + TCP_SECRET_LIFE 3149 + (0xff & tcp_cookie_work(bakery, 1)); 3150 tcp_secret_primary->expires = jiffy 3151 + TCP_SECRET_2MSL 3152 + (0x1f & tcp_cookie_work(bakery, 2)); 3153 } 3154 memcpy(&tcp_secret_secondary->secrets[0], 3155 bakery, COOKIE_WORKSPACE_WORDS); 3156 3157 rcu_assign_pointer(tcp_secret_generating, 3158 tcp_secret_secondary); 3159 rcu_assign_pointer(tcp_secret_retiring, 3160 tcp_secret_primary); 3161 /* 3162 * Neither call_rcu() nor synchronize_rcu() needed. 3163 * Retiring data is not freed. It is replaced after 3164 * further (locked) pointer updates, and a quiet time 3165 * (minimum 1MSL, maximum LIFE - 2MSL). 3166 */ 3167 } 3168 spin_unlock_bh(&tcp_secret_locker); 3169 } else { 3170 rcu_read_lock_bh(); 3171 memcpy(bakery, 3172 &rcu_dereference(tcp_secret_generating)->secrets[0], 3173 COOKIE_WORKSPACE_WORDS); 3174 rcu_read_unlock_bh(); 3175 } 3176 return 0; 3177 } 3178 EXPORT_SYMBOL(tcp_cookie_generator); 3179 3180 void tcp_done(struct sock *sk) 3181 { 3182 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV) 3183 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 3184 3185 tcp_set_state(sk, TCP_CLOSE); 3186 tcp_clear_xmit_timers(sk); 3187 3188 sk->sk_shutdown = SHUTDOWN_MASK; 3189 3190 if (!sock_flag(sk, SOCK_DEAD)) 3191 sk->sk_state_change(sk); 3192 else 3193 inet_csk_destroy_sock(sk); 3194 } 3195 EXPORT_SYMBOL_GPL(tcp_done); 3196 3197 extern struct tcp_congestion_ops tcp_reno; 3198 3199 static __initdata unsigned long thash_entries; 3200 static int __init set_thash_entries(char *str) 3201 { 3202 if (!str) 3203 return 0; 3204 thash_entries = simple_strtoul(str, &str, 0); 3205 return 1; 3206 } 3207 __setup("thash_entries=", set_thash_entries); 3208 3209 void __init tcp_init(void) 3210 { 3211 struct sk_buff *skb = NULL; 3212 unsigned long nr_pages, limit; 3213 int i, max_share, cnt; 3214 unsigned long jiffy = jiffies; 3215 3216 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb)); 3217 3218 percpu_counter_init(&tcp_sockets_allocated, 0); 3219 percpu_counter_init(&tcp_orphan_count, 0); 3220 tcp_hashinfo.bind_bucket_cachep = 3221 kmem_cache_create("tcp_bind_bucket", 3222 sizeof(struct inet_bind_bucket), 0, 3223 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL); 3224 3225 /* Size and allocate the main established and bind bucket 3226 * hash tables. 3227 * 3228 * The methodology is similar to that of the buffer cache. 3229 */ 3230 tcp_hashinfo.ehash = 3231 alloc_large_system_hash("TCP established", 3232 sizeof(struct inet_ehash_bucket), 3233 thash_entries, 3234 (totalram_pages >= 128 * 1024) ? 3235 13 : 15, 3236 0, 3237 NULL, 3238 &tcp_hashinfo.ehash_mask, 3239 thash_entries ? 0 : 512 * 1024); 3240 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) { 3241 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i); 3242 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i); 3243 } 3244 if (inet_ehash_locks_alloc(&tcp_hashinfo)) 3245 panic("TCP: failed to alloc ehash_locks"); 3246 tcp_hashinfo.bhash = 3247 alloc_large_system_hash("TCP bind", 3248 sizeof(struct inet_bind_hashbucket), 3249 tcp_hashinfo.ehash_mask + 1, 3250 (totalram_pages >= 128 * 1024) ? 3251 13 : 15, 3252 0, 3253 &tcp_hashinfo.bhash_size, 3254 NULL, 3255 64 * 1024); 3256 tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size; 3257 for (i = 0; i < tcp_hashinfo.bhash_size; i++) { 3258 spin_lock_init(&tcp_hashinfo.bhash[i].lock); 3259 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain); 3260 } 3261 3262 3263 cnt = tcp_hashinfo.ehash_mask + 1; 3264 3265 tcp_death_row.sysctl_max_tw_buckets = cnt / 2; 3266 sysctl_tcp_max_orphans = cnt / 2; 3267 sysctl_max_syn_backlog = max(128, cnt / 256); 3268 3269 /* Set the pressure threshold to be a fraction of global memory that 3270 * is up to 1/2 at 256 MB, decreasing toward zero with the amount of 3271 * memory, with a floor of 128 pages. 3272 */ 3273 nr_pages = totalram_pages - totalhigh_pages; 3274 limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT); 3275 limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11); 3276 limit = max(limit, 128UL); 3277 sysctl_tcp_mem[0] = limit / 4 * 3; 3278 sysctl_tcp_mem[1] = limit; 3279 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; 3280 3281 /* Set per-socket limits to no more than 1/128 the pressure threshold */ 3282 limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7); 3283 max_share = min(4UL*1024*1024, limit); 3284 3285 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM; 3286 sysctl_tcp_wmem[1] = 16*1024; 3287 sysctl_tcp_wmem[2] = max(64*1024, max_share); 3288 3289 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM; 3290 sysctl_tcp_rmem[1] = 87380; 3291 sysctl_tcp_rmem[2] = max(87380, max_share); 3292 3293 printk(KERN_INFO "TCP: Hash tables configured " 3294 "(established %u bind %u)\n", 3295 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size); 3296 3297 tcp_register_congestion_control(&tcp_reno); 3298 3299 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets)); 3300 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets)); 3301 tcp_secret_one.expires = jiffy; /* past due */ 3302 tcp_secret_two.expires = jiffy; /* past due */ 3303 tcp_secret_generating = &tcp_secret_one; 3304 tcp_secret_primary = &tcp_secret_one; 3305 tcp_secret_retiring = &tcp_secret_two; 3306 tcp_secret_secondary = &tcp_secret_two; 3307 } 3308