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 } 455 456 if (tp->urg_data & TCP_URG_VALID) 457 mask |= POLLPRI; 458 } 459 return mask; 460 } 461 EXPORT_SYMBOL(tcp_poll); 462 463 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg) 464 { 465 struct tcp_sock *tp = tcp_sk(sk); 466 int answ; 467 468 switch (cmd) { 469 case SIOCINQ: 470 if (sk->sk_state == TCP_LISTEN) 471 return -EINVAL; 472 473 lock_sock(sk); 474 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 475 answ = 0; 476 else if (sock_flag(sk, SOCK_URGINLINE) || 477 !tp->urg_data || 478 before(tp->urg_seq, tp->copied_seq) || 479 !before(tp->urg_seq, tp->rcv_nxt)) { 480 struct sk_buff *skb; 481 482 answ = tp->rcv_nxt - tp->copied_seq; 483 484 /* Subtract 1, if FIN is in queue. */ 485 skb = skb_peek_tail(&sk->sk_receive_queue); 486 if (answ && skb) 487 answ -= tcp_hdr(skb)->fin; 488 } else 489 answ = tp->urg_seq - tp->copied_seq; 490 release_sock(sk); 491 break; 492 case SIOCATMARK: 493 answ = tp->urg_data && tp->urg_seq == tp->copied_seq; 494 break; 495 case SIOCOUTQ: 496 if (sk->sk_state == TCP_LISTEN) 497 return -EINVAL; 498 499 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 500 answ = 0; 501 else 502 answ = tp->write_seq - tp->snd_una; 503 break; 504 default: 505 return -ENOIOCTLCMD; 506 } 507 508 return put_user(answ, (int __user *)arg); 509 } 510 EXPORT_SYMBOL(tcp_ioctl); 511 512 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb) 513 { 514 TCP_SKB_CB(skb)->flags |= TCPHDR_PSH; 515 tp->pushed_seq = tp->write_seq; 516 } 517 518 static inline int forced_push(struct tcp_sock *tp) 519 { 520 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1)); 521 } 522 523 static inline void skb_entail(struct sock *sk, struct sk_buff *skb) 524 { 525 struct tcp_sock *tp = tcp_sk(sk); 526 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 527 528 skb->csum = 0; 529 tcb->seq = tcb->end_seq = tp->write_seq; 530 tcb->flags = TCPHDR_ACK; 531 tcb->sacked = 0; 532 skb_header_release(skb); 533 tcp_add_write_queue_tail(sk, skb); 534 sk->sk_wmem_queued += skb->truesize; 535 sk_mem_charge(sk, skb->truesize); 536 if (tp->nonagle & TCP_NAGLE_PUSH) 537 tp->nonagle &= ~TCP_NAGLE_PUSH; 538 } 539 540 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags) 541 { 542 if (flags & MSG_OOB) 543 tp->snd_up = tp->write_seq; 544 } 545 546 static inline void tcp_push(struct sock *sk, int flags, int mss_now, 547 int nonagle) 548 { 549 if (tcp_send_head(sk)) { 550 struct tcp_sock *tp = tcp_sk(sk); 551 552 if (!(flags & MSG_MORE) || forced_push(tp)) 553 tcp_mark_push(tp, tcp_write_queue_tail(sk)); 554 555 tcp_mark_urg(tp, flags); 556 __tcp_push_pending_frames(sk, mss_now, 557 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle); 558 } 559 } 560 561 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, 562 unsigned int offset, size_t len) 563 { 564 struct tcp_splice_state *tss = rd_desc->arg.data; 565 int ret; 566 567 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len), 568 tss->flags); 569 if (ret > 0) 570 rd_desc->count -= ret; 571 return ret; 572 } 573 574 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss) 575 { 576 /* Store TCP splice context information in read_descriptor_t. */ 577 read_descriptor_t rd_desc = { 578 .arg.data = tss, 579 .count = tss->len, 580 }; 581 582 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv); 583 } 584 585 /** 586 * tcp_splice_read - splice data from TCP socket to a pipe 587 * @sock: socket to splice from 588 * @ppos: position (not valid) 589 * @pipe: pipe to splice to 590 * @len: number of bytes to splice 591 * @flags: splice modifier flags 592 * 593 * Description: 594 * Will read pages from given socket and fill them into a pipe. 595 * 596 **/ 597 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos, 598 struct pipe_inode_info *pipe, size_t len, 599 unsigned int flags) 600 { 601 struct sock *sk = sock->sk; 602 struct tcp_splice_state tss = { 603 .pipe = pipe, 604 .len = len, 605 .flags = flags, 606 }; 607 long timeo; 608 ssize_t spliced; 609 int ret; 610 611 sock_rps_record_flow(sk); 612 /* 613 * We can't seek on a socket input 614 */ 615 if (unlikely(*ppos)) 616 return -ESPIPE; 617 618 ret = spliced = 0; 619 620 lock_sock(sk); 621 622 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK); 623 while (tss.len) { 624 ret = __tcp_splice_read(sk, &tss); 625 if (ret < 0) 626 break; 627 else if (!ret) { 628 if (spliced) 629 break; 630 if (sock_flag(sk, SOCK_DONE)) 631 break; 632 if (sk->sk_err) { 633 ret = sock_error(sk); 634 break; 635 } 636 if (sk->sk_shutdown & RCV_SHUTDOWN) 637 break; 638 if (sk->sk_state == TCP_CLOSE) { 639 /* 640 * This occurs when user tries to read 641 * from never connected socket. 642 */ 643 if (!sock_flag(sk, SOCK_DONE)) 644 ret = -ENOTCONN; 645 break; 646 } 647 if (!timeo) { 648 ret = -EAGAIN; 649 break; 650 } 651 sk_wait_data(sk, &timeo); 652 if (signal_pending(current)) { 653 ret = sock_intr_errno(timeo); 654 break; 655 } 656 continue; 657 } 658 tss.len -= ret; 659 spliced += ret; 660 661 if (!timeo) 662 break; 663 release_sock(sk); 664 lock_sock(sk); 665 666 if (sk->sk_err || sk->sk_state == TCP_CLOSE || 667 (sk->sk_shutdown & RCV_SHUTDOWN) || 668 signal_pending(current)) 669 break; 670 } 671 672 release_sock(sk); 673 674 if (spliced) 675 return spliced; 676 677 return ret; 678 } 679 EXPORT_SYMBOL(tcp_splice_read); 680 681 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp) 682 { 683 struct sk_buff *skb; 684 685 /* The TCP header must be at least 32-bit aligned. */ 686 size = ALIGN(size, 4); 687 688 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp); 689 if (skb) { 690 if (sk_wmem_schedule(sk, skb->truesize)) { 691 /* 692 * Make sure that we have exactly size bytes 693 * available to the caller, no more, no less. 694 */ 695 skb_reserve(skb, skb_tailroom(skb) - size); 696 return skb; 697 } 698 __kfree_skb(skb); 699 } else { 700 sk->sk_prot->enter_memory_pressure(sk); 701 sk_stream_moderate_sndbuf(sk); 702 } 703 return NULL; 704 } 705 706 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now, 707 int large_allowed) 708 { 709 struct tcp_sock *tp = tcp_sk(sk); 710 u32 xmit_size_goal, old_size_goal; 711 712 xmit_size_goal = mss_now; 713 714 if (large_allowed && sk_can_gso(sk)) { 715 xmit_size_goal = ((sk->sk_gso_max_size - 1) - 716 inet_csk(sk)->icsk_af_ops->net_header_len - 717 inet_csk(sk)->icsk_ext_hdr_len - 718 tp->tcp_header_len); 719 720 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal); 721 722 /* We try hard to avoid divides here */ 723 old_size_goal = tp->xmit_size_goal_segs * mss_now; 724 725 if (likely(old_size_goal <= xmit_size_goal && 726 old_size_goal + mss_now > xmit_size_goal)) { 727 xmit_size_goal = old_size_goal; 728 } else { 729 tp->xmit_size_goal_segs = xmit_size_goal / mss_now; 730 xmit_size_goal = tp->xmit_size_goal_segs * mss_now; 731 } 732 } 733 734 return max(xmit_size_goal, mss_now); 735 } 736 737 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags) 738 { 739 int mss_now; 740 741 mss_now = tcp_current_mss(sk); 742 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB)); 743 744 return mss_now; 745 } 746 747 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset, 748 size_t psize, int flags) 749 { 750 struct tcp_sock *tp = tcp_sk(sk); 751 int mss_now, size_goal; 752 int err; 753 ssize_t copied; 754 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 755 756 /* Wait for a connection to finish. */ 757 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 758 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0) 759 goto out_err; 760 761 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags); 762 763 mss_now = tcp_send_mss(sk, &size_goal, flags); 764 copied = 0; 765 766 err = -EPIPE; 767 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 768 goto out_err; 769 770 while (psize > 0) { 771 struct sk_buff *skb = tcp_write_queue_tail(sk); 772 struct page *page = pages[poffset / PAGE_SIZE]; 773 int copy, i, can_coalesce; 774 int offset = poffset % PAGE_SIZE; 775 int size = min_t(size_t, psize, PAGE_SIZE - offset); 776 777 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) { 778 new_segment: 779 if (!sk_stream_memory_free(sk)) 780 goto wait_for_sndbuf; 781 782 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation); 783 if (!skb) 784 goto wait_for_memory; 785 786 skb_entail(sk, skb); 787 copy = size_goal; 788 } 789 790 if (copy > size) 791 copy = size; 792 793 i = skb_shinfo(skb)->nr_frags; 794 can_coalesce = skb_can_coalesce(skb, i, page, offset); 795 if (!can_coalesce && i >= MAX_SKB_FRAGS) { 796 tcp_mark_push(tp, skb); 797 goto new_segment; 798 } 799 if (!sk_wmem_schedule(sk, copy)) 800 goto wait_for_memory; 801 802 if (can_coalesce) { 803 skb_shinfo(skb)->frags[i - 1].size += copy; 804 } else { 805 get_page(page); 806 skb_fill_page_desc(skb, i, page, offset, copy); 807 } 808 809 skb->len += copy; 810 skb->data_len += copy; 811 skb->truesize += copy; 812 sk->sk_wmem_queued += copy; 813 sk_mem_charge(sk, copy); 814 skb->ip_summed = CHECKSUM_PARTIAL; 815 tp->write_seq += copy; 816 TCP_SKB_CB(skb)->end_seq += copy; 817 skb_shinfo(skb)->gso_segs = 0; 818 819 if (!copied) 820 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH; 821 822 copied += copy; 823 poffset += copy; 824 if (!(psize -= copy)) 825 goto out; 826 827 if (skb->len < size_goal || (flags & MSG_OOB)) 828 continue; 829 830 if (forced_push(tp)) { 831 tcp_mark_push(tp, skb); 832 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 833 } else if (skb == tcp_send_head(sk)) 834 tcp_push_one(sk, mss_now); 835 continue; 836 837 wait_for_sndbuf: 838 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 839 wait_for_memory: 840 if (copied) 841 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH); 842 843 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0) 844 goto do_error; 845 846 mss_now = tcp_send_mss(sk, &size_goal, flags); 847 } 848 849 out: 850 if (copied) 851 tcp_push(sk, flags, mss_now, tp->nonagle); 852 return copied; 853 854 do_error: 855 if (copied) 856 goto out; 857 out_err: 858 return sk_stream_error(sk, flags, err); 859 } 860 861 int tcp_sendpage(struct sock *sk, struct page *page, int offset, 862 size_t size, int flags) 863 { 864 ssize_t res; 865 866 if (!(sk->sk_route_caps & NETIF_F_SG) || 867 !(sk->sk_route_caps & NETIF_F_ALL_CSUM)) 868 return sock_no_sendpage(sk->sk_socket, page, offset, size, 869 flags); 870 871 lock_sock(sk); 872 TCP_CHECK_TIMER(sk); 873 res = do_tcp_sendpages(sk, &page, offset, size, flags); 874 TCP_CHECK_TIMER(sk); 875 release_sock(sk); 876 return res; 877 } 878 EXPORT_SYMBOL(tcp_sendpage); 879 880 #define TCP_PAGE(sk) (sk->sk_sndmsg_page) 881 #define TCP_OFF(sk) (sk->sk_sndmsg_off) 882 883 static inline int select_size(struct sock *sk, int sg) 884 { 885 struct tcp_sock *tp = tcp_sk(sk); 886 int tmp = tp->mss_cache; 887 888 if (sg) { 889 if (sk_can_gso(sk)) 890 tmp = 0; 891 else { 892 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER); 893 894 if (tmp >= pgbreak && 895 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE) 896 tmp = pgbreak; 897 } 898 } 899 900 return tmp; 901 } 902 903 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg, 904 size_t size) 905 { 906 struct iovec *iov; 907 struct tcp_sock *tp = tcp_sk(sk); 908 struct sk_buff *skb; 909 int iovlen, flags; 910 int mss_now, size_goal; 911 int sg, err, copied; 912 long timeo; 913 914 lock_sock(sk); 915 TCP_CHECK_TIMER(sk); 916 917 flags = msg->msg_flags; 918 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 919 920 /* Wait for a connection to finish. */ 921 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 922 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0) 923 goto out_err; 924 925 /* This should be in poll */ 926 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags); 927 928 mss_now = tcp_send_mss(sk, &size_goal, flags); 929 930 /* Ok commence sending. */ 931 iovlen = msg->msg_iovlen; 932 iov = msg->msg_iov; 933 copied = 0; 934 935 err = -EPIPE; 936 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 937 goto out_err; 938 939 sg = sk->sk_route_caps & NETIF_F_SG; 940 941 while (--iovlen >= 0) { 942 int seglen = iov->iov_len; 943 unsigned char __user *from = iov->iov_base; 944 945 iov++; 946 947 while (seglen > 0) { 948 int copy = 0; 949 int max = size_goal; 950 951 skb = tcp_write_queue_tail(sk); 952 if (tcp_send_head(sk)) { 953 if (skb->ip_summed == CHECKSUM_NONE) 954 max = mss_now; 955 copy = max - skb->len; 956 } 957 958 if (copy <= 0) { 959 new_segment: 960 /* Allocate new segment. If the interface is SG, 961 * allocate skb fitting to single page. 962 */ 963 if (!sk_stream_memory_free(sk)) 964 goto wait_for_sndbuf; 965 966 skb = sk_stream_alloc_skb(sk, 967 select_size(sk, sg), 968 sk->sk_allocation); 969 if (!skb) 970 goto wait_for_memory; 971 972 /* 973 * Check whether we can use HW checksum. 974 */ 975 if (sk->sk_route_caps & NETIF_F_ALL_CSUM) 976 skb->ip_summed = CHECKSUM_PARTIAL; 977 978 skb_entail(sk, skb); 979 copy = size_goal; 980 max = size_goal; 981 } 982 983 /* Try to append data to the end of skb. */ 984 if (copy > seglen) 985 copy = seglen; 986 987 /* Where to copy to? */ 988 if (skb_tailroom(skb) > 0) { 989 /* We have some space in skb head. Superb! */ 990 if (copy > skb_tailroom(skb)) 991 copy = skb_tailroom(skb); 992 if ((err = skb_add_data(skb, from, copy)) != 0) 993 goto do_fault; 994 } else { 995 int merge = 0; 996 int i = skb_shinfo(skb)->nr_frags; 997 struct page *page = TCP_PAGE(sk); 998 int off = TCP_OFF(sk); 999 1000 if (skb_can_coalesce(skb, i, page, off) && 1001 off != PAGE_SIZE) { 1002 /* We can extend the last page 1003 * fragment. */ 1004 merge = 1; 1005 } else if (i == MAX_SKB_FRAGS || !sg) { 1006 /* Need to add new fragment and cannot 1007 * do this because interface is non-SG, 1008 * or because all the page slots are 1009 * busy. */ 1010 tcp_mark_push(tp, skb); 1011 goto new_segment; 1012 } else if (page) { 1013 if (off == PAGE_SIZE) { 1014 put_page(page); 1015 TCP_PAGE(sk) = page = NULL; 1016 off = 0; 1017 } 1018 } else 1019 off = 0; 1020 1021 if (copy > PAGE_SIZE - off) 1022 copy = PAGE_SIZE - off; 1023 1024 if (!sk_wmem_schedule(sk, copy)) 1025 goto wait_for_memory; 1026 1027 if (!page) { 1028 /* Allocate new cache page. */ 1029 if (!(page = sk_stream_alloc_page(sk))) 1030 goto wait_for_memory; 1031 } 1032 1033 /* Time to copy data. We are close to 1034 * the end! */ 1035 err = skb_copy_to_page(sk, from, skb, page, 1036 off, copy); 1037 if (err) { 1038 /* If this page was new, give it to the 1039 * socket so it does not get leaked. 1040 */ 1041 if (!TCP_PAGE(sk)) { 1042 TCP_PAGE(sk) = page; 1043 TCP_OFF(sk) = 0; 1044 } 1045 goto do_error; 1046 } 1047 1048 /* Update the skb. */ 1049 if (merge) { 1050 skb_shinfo(skb)->frags[i - 1].size += 1051 copy; 1052 } else { 1053 skb_fill_page_desc(skb, i, page, off, copy); 1054 if (TCP_PAGE(sk)) { 1055 get_page(page); 1056 } else if (off + copy < PAGE_SIZE) { 1057 get_page(page); 1058 TCP_PAGE(sk) = page; 1059 } 1060 } 1061 1062 TCP_OFF(sk) = off + copy; 1063 } 1064 1065 if (!copied) 1066 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH; 1067 1068 tp->write_seq += copy; 1069 TCP_SKB_CB(skb)->end_seq += copy; 1070 skb_shinfo(skb)->gso_segs = 0; 1071 1072 from += copy; 1073 copied += copy; 1074 if ((seglen -= copy) == 0 && iovlen == 0) 1075 goto out; 1076 1077 if (skb->len < max || (flags & MSG_OOB)) 1078 continue; 1079 1080 if (forced_push(tp)) { 1081 tcp_mark_push(tp, skb); 1082 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 1083 } else if (skb == tcp_send_head(sk)) 1084 tcp_push_one(sk, mss_now); 1085 continue; 1086 1087 wait_for_sndbuf: 1088 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1089 wait_for_memory: 1090 if (copied) 1091 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH); 1092 1093 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0) 1094 goto do_error; 1095 1096 mss_now = tcp_send_mss(sk, &size_goal, flags); 1097 } 1098 } 1099 1100 out: 1101 if (copied) 1102 tcp_push(sk, flags, mss_now, tp->nonagle); 1103 TCP_CHECK_TIMER(sk); 1104 release_sock(sk); 1105 return copied; 1106 1107 do_fault: 1108 if (!skb->len) { 1109 tcp_unlink_write_queue(skb, sk); 1110 /* It is the one place in all of TCP, except connection 1111 * reset, where we can be unlinking the send_head. 1112 */ 1113 tcp_check_send_head(sk, skb); 1114 sk_wmem_free_skb(sk, skb); 1115 } 1116 1117 do_error: 1118 if (copied) 1119 goto out; 1120 out_err: 1121 err = sk_stream_error(sk, flags, err); 1122 TCP_CHECK_TIMER(sk); 1123 release_sock(sk); 1124 return err; 1125 } 1126 EXPORT_SYMBOL(tcp_sendmsg); 1127 1128 /* 1129 * Handle reading urgent data. BSD has very simple semantics for 1130 * this, no blocking and very strange errors 8) 1131 */ 1132 1133 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags) 1134 { 1135 struct tcp_sock *tp = tcp_sk(sk); 1136 1137 /* No URG data to read. */ 1138 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data || 1139 tp->urg_data == TCP_URG_READ) 1140 return -EINVAL; /* Yes this is right ! */ 1141 1142 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE)) 1143 return -ENOTCONN; 1144 1145 if (tp->urg_data & TCP_URG_VALID) { 1146 int err = 0; 1147 char c = tp->urg_data; 1148 1149 if (!(flags & MSG_PEEK)) 1150 tp->urg_data = TCP_URG_READ; 1151 1152 /* Read urgent data. */ 1153 msg->msg_flags |= MSG_OOB; 1154 1155 if (len > 0) { 1156 if (!(flags & MSG_TRUNC)) 1157 err = memcpy_toiovec(msg->msg_iov, &c, 1); 1158 len = 1; 1159 } else 1160 msg->msg_flags |= MSG_TRUNC; 1161 1162 return err ? -EFAULT : len; 1163 } 1164 1165 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 1166 return 0; 1167 1168 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and 1169 * the available implementations agree in this case: 1170 * this call should never block, independent of the 1171 * blocking state of the socket. 1172 * Mike <pall@rz.uni-karlsruhe.de> 1173 */ 1174 return -EAGAIN; 1175 } 1176 1177 /* Clean up the receive buffer for full frames taken by the user, 1178 * then send an ACK if necessary. COPIED is the number of bytes 1179 * tcp_recvmsg has given to the user so far, it speeds up the 1180 * calculation of whether or not we must ACK for the sake of 1181 * a window update. 1182 */ 1183 void tcp_cleanup_rbuf(struct sock *sk, int copied) 1184 { 1185 struct tcp_sock *tp = tcp_sk(sk); 1186 int time_to_ack = 0; 1187 1188 #if TCP_DEBUG 1189 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 1190 1191 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq), 1192 KERN_INFO "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n", 1193 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt); 1194 #endif 1195 1196 if (inet_csk_ack_scheduled(sk)) { 1197 const struct inet_connection_sock *icsk = inet_csk(sk); 1198 /* Delayed ACKs frequently hit locked sockets during bulk 1199 * receive. */ 1200 if (icsk->icsk_ack.blocked || 1201 /* Once-per-two-segments ACK was not sent by tcp_input.c */ 1202 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss || 1203 /* 1204 * If this read emptied read buffer, we send ACK, if 1205 * connection is not bidirectional, user drained 1206 * receive buffer and there was a small segment 1207 * in queue. 1208 */ 1209 (copied > 0 && 1210 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) || 1211 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) && 1212 !icsk->icsk_ack.pingpong)) && 1213 !atomic_read(&sk->sk_rmem_alloc))) 1214 time_to_ack = 1; 1215 } 1216 1217 /* We send an ACK if we can now advertise a non-zero window 1218 * which has been raised "significantly". 1219 * 1220 * Even if window raised up to infinity, do not send window open ACK 1221 * in states, where we will not receive more. It is useless. 1222 */ 1223 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) { 1224 __u32 rcv_window_now = tcp_receive_window(tp); 1225 1226 /* Optimize, __tcp_select_window() is not cheap. */ 1227 if (2*rcv_window_now <= tp->window_clamp) { 1228 __u32 new_window = __tcp_select_window(sk); 1229 1230 /* Send ACK now, if this read freed lots of space 1231 * in our buffer. Certainly, new_window is new window. 1232 * We can advertise it now, if it is not less than current one. 1233 * "Lots" means "at least twice" here. 1234 */ 1235 if (new_window && new_window >= 2 * rcv_window_now) 1236 time_to_ack = 1; 1237 } 1238 } 1239 if (time_to_ack) 1240 tcp_send_ack(sk); 1241 } 1242 1243 static void tcp_prequeue_process(struct sock *sk) 1244 { 1245 struct sk_buff *skb; 1246 struct tcp_sock *tp = tcp_sk(sk); 1247 1248 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED); 1249 1250 /* RX process wants to run with disabled BHs, though it is not 1251 * necessary */ 1252 local_bh_disable(); 1253 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL) 1254 sk_backlog_rcv(sk, skb); 1255 local_bh_enable(); 1256 1257 /* Clear memory counter. */ 1258 tp->ucopy.memory = 0; 1259 } 1260 1261 #ifdef CONFIG_NET_DMA 1262 static void tcp_service_net_dma(struct sock *sk, bool wait) 1263 { 1264 dma_cookie_t done, used; 1265 dma_cookie_t last_issued; 1266 struct tcp_sock *tp = tcp_sk(sk); 1267 1268 if (!tp->ucopy.dma_chan) 1269 return; 1270 1271 last_issued = tp->ucopy.dma_cookie; 1272 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1273 1274 do { 1275 if (dma_async_memcpy_complete(tp->ucopy.dma_chan, 1276 last_issued, &done, 1277 &used) == DMA_SUCCESS) { 1278 /* Safe to free early-copied skbs now */ 1279 __skb_queue_purge(&sk->sk_async_wait_queue); 1280 break; 1281 } else { 1282 struct sk_buff *skb; 1283 while ((skb = skb_peek(&sk->sk_async_wait_queue)) && 1284 (dma_async_is_complete(skb->dma_cookie, done, 1285 used) == DMA_SUCCESS)) { 1286 __skb_dequeue(&sk->sk_async_wait_queue); 1287 kfree_skb(skb); 1288 } 1289 } 1290 } while (wait); 1291 } 1292 #endif 1293 1294 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off) 1295 { 1296 struct sk_buff *skb; 1297 u32 offset; 1298 1299 skb_queue_walk(&sk->sk_receive_queue, skb) { 1300 offset = seq - TCP_SKB_CB(skb)->seq; 1301 if (tcp_hdr(skb)->syn) 1302 offset--; 1303 if (offset < skb->len || tcp_hdr(skb)->fin) { 1304 *off = offset; 1305 return skb; 1306 } 1307 } 1308 return NULL; 1309 } 1310 1311 /* 1312 * This routine provides an alternative to tcp_recvmsg() for routines 1313 * that would like to handle copying from skbuffs directly in 'sendfile' 1314 * fashion. 1315 * Note: 1316 * - It is assumed that the socket was locked by the caller. 1317 * - The routine does not block. 1318 * - At present, there is no support for reading OOB data 1319 * or for 'peeking' the socket using this routine 1320 * (although both would be easy to implement). 1321 */ 1322 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, 1323 sk_read_actor_t recv_actor) 1324 { 1325 struct sk_buff *skb; 1326 struct tcp_sock *tp = tcp_sk(sk); 1327 u32 seq = tp->copied_seq; 1328 u32 offset; 1329 int copied = 0; 1330 1331 if (sk->sk_state == TCP_LISTEN) 1332 return -ENOTCONN; 1333 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) { 1334 if (offset < skb->len) { 1335 int used; 1336 size_t len; 1337 1338 len = skb->len - offset; 1339 /* Stop reading if we hit a patch of urgent data */ 1340 if (tp->urg_data) { 1341 u32 urg_offset = tp->urg_seq - seq; 1342 if (urg_offset < len) 1343 len = urg_offset; 1344 if (!len) 1345 break; 1346 } 1347 used = recv_actor(desc, skb, offset, len); 1348 if (used < 0) { 1349 if (!copied) 1350 copied = used; 1351 break; 1352 } else if (used <= len) { 1353 seq += used; 1354 copied += used; 1355 offset += used; 1356 } 1357 /* 1358 * If recv_actor drops the lock (e.g. TCP splice 1359 * receive) the skb pointer might be invalid when 1360 * getting here: tcp_collapse might have deleted it 1361 * while aggregating skbs from the socket queue. 1362 */ 1363 skb = tcp_recv_skb(sk, seq-1, &offset); 1364 if (!skb || (offset+1 != skb->len)) 1365 break; 1366 } 1367 if (tcp_hdr(skb)->fin) { 1368 sk_eat_skb(sk, skb, 0); 1369 ++seq; 1370 break; 1371 } 1372 sk_eat_skb(sk, skb, 0); 1373 if (!desc->count) 1374 break; 1375 tp->copied_seq = seq; 1376 } 1377 tp->copied_seq = seq; 1378 1379 tcp_rcv_space_adjust(sk); 1380 1381 /* Clean up data we have read: This will do ACK frames. */ 1382 if (copied > 0) 1383 tcp_cleanup_rbuf(sk, copied); 1384 return copied; 1385 } 1386 EXPORT_SYMBOL(tcp_read_sock); 1387 1388 /* 1389 * This routine copies from a sock struct into the user buffer. 1390 * 1391 * Technical note: in 2.3 we work on _locked_ socket, so that 1392 * tricks with *seq access order and skb->users are not required. 1393 * Probably, code can be easily improved even more. 1394 */ 1395 1396 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg, 1397 size_t len, int nonblock, int flags, int *addr_len) 1398 { 1399 struct tcp_sock *tp = tcp_sk(sk); 1400 int copied = 0; 1401 u32 peek_seq; 1402 u32 *seq; 1403 unsigned long used; 1404 int err; 1405 int target; /* Read at least this many bytes */ 1406 long timeo; 1407 struct task_struct *user_recv = NULL; 1408 int copied_early = 0; 1409 struct sk_buff *skb; 1410 u32 urg_hole = 0; 1411 1412 lock_sock(sk); 1413 1414 TCP_CHECK_TIMER(sk); 1415 1416 err = -ENOTCONN; 1417 if (sk->sk_state == TCP_LISTEN) 1418 goto out; 1419 1420 timeo = sock_rcvtimeo(sk, nonblock); 1421 1422 /* Urgent data needs to be handled specially. */ 1423 if (flags & MSG_OOB) 1424 goto recv_urg; 1425 1426 seq = &tp->copied_seq; 1427 if (flags & MSG_PEEK) { 1428 peek_seq = tp->copied_seq; 1429 seq = &peek_seq; 1430 } 1431 1432 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 1433 1434 #ifdef CONFIG_NET_DMA 1435 tp->ucopy.dma_chan = NULL; 1436 preempt_disable(); 1437 skb = skb_peek_tail(&sk->sk_receive_queue); 1438 { 1439 int available = 0; 1440 1441 if (skb) 1442 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq); 1443 if ((available < target) && 1444 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) && 1445 !sysctl_tcp_low_latency && 1446 dma_find_channel(DMA_MEMCPY)) { 1447 preempt_enable_no_resched(); 1448 tp->ucopy.pinned_list = 1449 dma_pin_iovec_pages(msg->msg_iov, len); 1450 } else { 1451 preempt_enable_no_resched(); 1452 } 1453 } 1454 #endif 1455 1456 do { 1457 u32 offset; 1458 1459 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */ 1460 if (tp->urg_data && tp->urg_seq == *seq) { 1461 if (copied) 1462 break; 1463 if (signal_pending(current)) { 1464 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN; 1465 break; 1466 } 1467 } 1468 1469 /* Next get a buffer. */ 1470 1471 skb_queue_walk(&sk->sk_receive_queue, skb) { 1472 /* Now that we have two receive queues this 1473 * shouldn't happen. 1474 */ 1475 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq), 1476 KERN_INFO "recvmsg bug: copied %X " 1477 "seq %X rcvnxt %X fl %X\n", *seq, 1478 TCP_SKB_CB(skb)->seq, tp->rcv_nxt, 1479 flags)) 1480 break; 1481 1482 offset = *seq - TCP_SKB_CB(skb)->seq; 1483 if (tcp_hdr(skb)->syn) 1484 offset--; 1485 if (offset < skb->len) 1486 goto found_ok_skb; 1487 if (tcp_hdr(skb)->fin) 1488 goto found_fin_ok; 1489 WARN(!(flags & MSG_PEEK), KERN_INFO "recvmsg bug 2: " 1490 "copied %X seq %X rcvnxt %X fl %X\n", 1491 *seq, TCP_SKB_CB(skb)->seq, 1492 tp->rcv_nxt, flags); 1493 } 1494 1495 /* Well, if we have backlog, try to process it now yet. */ 1496 1497 if (copied >= target && !sk->sk_backlog.tail) 1498 break; 1499 1500 if (copied) { 1501 if (sk->sk_err || 1502 sk->sk_state == TCP_CLOSE || 1503 (sk->sk_shutdown & RCV_SHUTDOWN) || 1504 !timeo || 1505 signal_pending(current)) 1506 break; 1507 } else { 1508 if (sock_flag(sk, SOCK_DONE)) 1509 break; 1510 1511 if (sk->sk_err) { 1512 copied = sock_error(sk); 1513 break; 1514 } 1515 1516 if (sk->sk_shutdown & RCV_SHUTDOWN) 1517 break; 1518 1519 if (sk->sk_state == TCP_CLOSE) { 1520 if (!sock_flag(sk, SOCK_DONE)) { 1521 /* This occurs when user tries to read 1522 * from never connected socket. 1523 */ 1524 copied = -ENOTCONN; 1525 break; 1526 } 1527 break; 1528 } 1529 1530 if (!timeo) { 1531 copied = -EAGAIN; 1532 break; 1533 } 1534 1535 if (signal_pending(current)) { 1536 copied = sock_intr_errno(timeo); 1537 break; 1538 } 1539 } 1540 1541 tcp_cleanup_rbuf(sk, copied); 1542 1543 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) { 1544 /* Install new reader */ 1545 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) { 1546 user_recv = current; 1547 tp->ucopy.task = user_recv; 1548 tp->ucopy.iov = msg->msg_iov; 1549 } 1550 1551 tp->ucopy.len = len; 1552 1553 WARN_ON(tp->copied_seq != tp->rcv_nxt && 1554 !(flags & (MSG_PEEK | MSG_TRUNC))); 1555 1556 /* Ugly... If prequeue is not empty, we have to 1557 * process it before releasing socket, otherwise 1558 * order will be broken at second iteration. 1559 * More elegant solution is required!!! 1560 * 1561 * Look: we have the following (pseudo)queues: 1562 * 1563 * 1. packets in flight 1564 * 2. backlog 1565 * 3. prequeue 1566 * 4. receive_queue 1567 * 1568 * Each queue can be processed only if the next ones 1569 * are empty. At this point we have empty receive_queue. 1570 * But prequeue _can_ be not empty after 2nd iteration, 1571 * when we jumped to start of loop because backlog 1572 * processing added something to receive_queue. 1573 * We cannot release_sock(), because backlog contains 1574 * packets arrived _after_ prequeued ones. 1575 * 1576 * Shortly, algorithm is clear --- to process all 1577 * the queues in order. We could make it more directly, 1578 * requeueing packets from backlog to prequeue, if 1579 * is not empty. It is more elegant, but eats cycles, 1580 * unfortunately. 1581 */ 1582 if (!skb_queue_empty(&tp->ucopy.prequeue)) 1583 goto do_prequeue; 1584 1585 /* __ Set realtime policy in scheduler __ */ 1586 } 1587 1588 #ifdef CONFIG_NET_DMA 1589 if (tp->ucopy.dma_chan) 1590 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1591 #endif 1592 if (copied >= target) { 1593 /* Do not sleep, just process backlog. */ 1594 release_sock(sk); 1595 lock_sock(sk); 1596 } else 1597 sk_wait_data(sk, &timeo); 1598 1599 #ifdef CONFIG_NET_DMA 1600 tcp_service_net_dma(sk, false); /* Don't block */ 1601 tp->ucopy.wakeup = 0; 1602 #endif 1603 1604 if (user_recv) { 1605 int chunk; 1606 1607 /* __ Restore normal policy in scheduler __ */ 1608 1609 if ((chunk = len - tp->ucopy.len) != 0) { 1610 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk); 1611 len -= chunk; 1612 copied += chunk; 1613 } 1614 1615 if (tp->rcv_nxt == tp->copied_seq && 1616 !skb_queue_empty(&tp->ucopy.prequeue)) { 1617 do_prequeue: 1618 tcp_prequeue_process(sk); 1619 1620 if ((chunk = len - tp->ucopy.len) != 0) { 1621 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1622 len -= chunk; 1623 copied += chunk; 1624 } 1625 } 1626 } 1627 if ((flags & MSG_PEEK) && 1628 (peek_seq - copied - urg_hole != tp->copied_seq)) { 1629 if (net_ratelimit()) 1630 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n", 1631 current->comm, task_pid_nr(current)); 1632 peek_seq = tp->copied_seq; 1633 } 1634 continue; 1635 1636 found_ok_skb: 1637 /* Ok so how much can we use? */ 1638 used = skb->len - offset; 1639 if (len < used) 1640 used = len; 1641 1642 /* Do we have urgent data here? */ 1643 if (tp->urg_data) { 1644 u32 urg_offset = tp->urg_seq - *seq; 1645 if (urg_offset < used) { 1646 if (!urg_offset) { 1647 if (!sock_flag(sk, SOCK_URGINLINE)) { 1648 ++*seq; 1649 urg_hole++; 1650 offset++; 1651 used--; 1652 if (!used) 1653 goto skip_copy; 1654 } 1655 } else 1656 used = urg_offset; 1657 } 1658 } 1659 1660 if (!(flags & MSG_TRUNC)) { 1661 #ifdef CONFIG_NET_DMA 1662 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 1663 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY); 1664 1665 if (tp->ucopy.dma_chan) { 1666 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec( 1667 tp->ucopy.dma_chan, skb, offset, 1668 msg->msg_iov, used, 1669 tp->ucopy.pinned_list); 1670 1671 if (tp->ucopy.dma_cookie < 0) { 1672 1673 printk(KERN_ALERT "dma_cookie < 0\n"); 1674 1675 /* Exception. Bailout! */ 1676 if (!copied) 1677 copied = -EFAULT; 1678 break; 1679 } 1680 1681 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1682 1683 if ((offset + used) == skb->len) 1684 copied_early = 1; 1685 1686 } else 1687 #endif 1688 { 1689 err = skb_copy_datagram_iovec(skb, offset, 1690 msg->msg_iov, used); 1691 if (err) { 1692 /* Exception. Bailout! */ 1693 if (!copied) 1694 copied = -EFAULT; 1695 break; 1696 } 1697 } 1698 } 1699 1700 *seq += used; 1701 copied += used; 1702 len -= used; 1703 1704 tcp_rcv_space_adjust(sk); 1705 1706 skip_copy: 1707 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) { 1708 tp->urg_data = 0; 1709 tcp_fast_path_check(sk); 1710 } 1711 if (used + offset < skb->len) 1712 continue; 1713 1714 if (tcp_hdr(skb)->fin) 1715 goto found_fin_ok; 1716 if (!(flags & MSG_PEEK)) { 1717 sk_eat_skb(sk, skb, copied_early); 1718 copied_early = 0; 1719 } 1720 continue; 1721 1722 found_fin_ok: 1723 /* Process the FIN. */ 1724 ++*seq; 1725 if (!(flags & MSG_PEEK)) { 1726 sk_eat_skb(sk, skb, copied_early); 1727 copied_early = 0; 1728 } 1729 break; 1730 } while (len > 0); 1731 1732 if (user_recv) { 1733 if (!skb_queue_empty(&tp->ucopy.prequeue)) { 1734 int chunk; 1735 1736 tp->ucopy.len = copied > 0 ? len : 0; 1737 1738 tcp_prequeue_process(sk); 1739 1740 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) { 1741 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1742 len -= chunk; 1743 copied += chunk; 1744 } 1745 } 1746 1747 tp->ucopy.task = NULL; 1748 tp->ucopy.len = 0; 1749 } 1750 1751 #ifdef CONFIG_NET_DMA 1752 tcp_service_net_dma(sk, true); /* Wait for queue to drain */ 1753 tp->ucopy.dma_chan = NULL; 1754 1755 if (tp->ucopy.pinned_list) { 1756 dma_unpin_iovec_pages(tp->ucopy.pinned_list); 1757 tp->ucopy.pinned_list = NULL; 1758 } 1759 #endif 1760 1761 /* According to UNIX98, msg_name/msg_namelen are ignored 1762 * on connected socket. I was just happy when found this 8) --ANK 1763 */ 1764 1765 /* Clean up data we have read: This will do ACK frames. */ 1766 tcp_cleanup_rbuf(sk, copied); 1767 1768 TCP_CHECK_TIMER(sk); 1769 release_sock(sk); 1770 return copied; 1771 1772 out: 1773 TCP_CHECK_TIMER(sk); 1774 release_sock(sk); 1775 return err; 1776 1777 recv_urg: 1778 err = tcp_recv_urg(sk, msg, len, flags); 1779 goto out; 1780 } 1781 EXPORT_SYMBOL(tcp_recvmsg); 1782 1783 void tcp_set_state(struct sock *sk, int state) 1784 { 1785 int oldstate = sk->sk_state; 1786 1787 switch (state) { 1788 case TCP_ESTABLISHED: 1789 if (oldstate != TCP_ESTABLISHED) 1790 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1791 break; 1792 1793 case TCP_CLOSE: 1794 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED) 1795 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS); 1796 1797 sk->sk_prot->unhash(sk); 1798 if (inet_csk(sk)->icsk_bind_hash && 1799 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK)) 1800 inet_put_port(sk); 1801 /* fall through */ 1802 default: 1803 if (oldstate == TCP_ESTABLISHED) 1804 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1805 } 1806 1807 /* Change state AFTER socket is unhashed to avoid closed 1808 * socket sitting in hash tables. 1809 */ 1810 sk->sk_state = state; 1811 1812 #ifdef STATE_TRACE 1813 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]); 1814 #endif 1815 } 1816 EXPORT_SYMBOL_GPL(tcp_set_state); 1817 1818 /* 1819 * State processing on a close. This implements the state shift for 1820 * sending our FIN frame. Note that we only send a FIN for some 1821 * states. A shutdown() may have already sent the FIN, or we may be 1822 * closed. 1823 */ 1824 1825 static const unsigned char new_state[16] = { 1826 /* current state: new state: action: */ 1827 /* (Invalid) */ TCP_CLOSE, 1828 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1829 /* TCP_SYN_SENT */ TCP_CLOSE, 1830 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1831 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1, 1832 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2, 1833 /* TCP_TIME_WAIT */ TCP_CLOSE, 1834 /* TCP_CLOSE */ TCP_CLOSE, 1835 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN, 1836 /* TCP_LAST_ACK */ TCP_LAST_ACK, 1837 /* TCP_LISTEN */ TCP_CLOSE, 1838 /* TCP_CLOSING */ TCP_CLOSING, 1839 }; 1840 1841 static int tcp_close_state(struct sock *sk) 1842 { 1843 int next = (int)new_state[sk->sk_state]; 1844 int ns = next & TCP_STATE_MASK; 1845 1846 tcp_set_state(sk, ns); 1847 1848 return next & TCP_ACTION_FIN; 1849 } 1850 1851 /* 1852 * Shutdown the sending side of a connection. Much like close except 1853 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD). 1854 */ 1855 1856 void tcp_shutdown(struct sock *sk, int how) 1857 { 1858 /* We need to grab some memory, and put together a FIN, 1859 * and then put it into the queue to be sent. 1860 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92. 1861 */ 1862 if (!(how & SEND_SHUTDOWN)) 1863 return; 1864 1865 /* If we've already sent a FIN, or it's a closed state, skip this. */ 1866 if ((1 << sk->sk_state) & 1867 (TCPF_ESTABLISHED | TCPF_SYN_SENT | 1868 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) { 1869 /* Clear out any half completed packets. FIN if needed. */ 1870 if (tcp_close_state(sk)) 1871 tcp_send_fin(sk); 1872 } 1873 } 1874 EXPORT_SYMBOL(tcp_shutdown); 1875 1876 void tcp_close(struct sock *sk, long timeout) 1877 { 1878 struct sk_buff *skb; 1879 int data_was_unread = 0; 1880 int state; 1881 1882 lock_sock(sk); 1883 sk->sk_shutdown = SHUTDOWN_MASK; 1884 1885 if (sk->sk_state == TCP_LISTEN) { 1886 tcp_set_state(sk, TCP_CLOSE); 1887 1888 /* Special case. */ 1889 inet_csk_listen_stop(sk); 1890 1891 goto adjudge_to_death; 1892 } 1893 1894 /* We need to flush the recv. buffs. We do this only on the 1895 * descriptor close, not protocol-sourced closes, because the 1896 * reader process may not have drained the data yet! 1897 */ 1898 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) { 1899 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq - 1900 tcp_hdr(skb)->fin; 1901 data_was_unread += len; 1902 __kfree_skb(skb); 1903 } 1904 1905 sk_mem_reclaim(sk); 1906 1907 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */ 1908 if (sk->sk_state == TCP_CLOSE) 1909 goto adjudge_to_death; 1910 1911 /* As outlined in RFC 2525, section 2.17, we send a RST here because 1912 * data was lost. To witness the awful effects of the old behavior of 1913 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk 1914 * GET in an FTP client, suspend the process, wait for the client to 1915 * advertise a zero window, then kill -9 the FTP client, wheee... 1916 * Note: timeout is always zero in such a case. 1917 */ 1918 if (data_was_unread) { 1919 /* Unread data was tossed, zap the connection. */ 1920 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE); 1921 tcp_set_state(sk, TCP_CLOSE); 1922 tcp_send_active_reset(sk, sk->sk_allocation); 1923 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) { 1924 /* Check zero linger _after_ checking for unread data. */ 1925 sk->sk_prot->disconnect(sk, 0); 1926 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 1927 } else if (tcp_close_state(sk)) { 1928 /* We FIN if the application ate all the data before 1929 * zapping the connection. 1930 */ 1931 1932 /* RED-PEN. Formally speaking, we have broken TCP state 1933 * machine. State transitions: 1934 * 1935 * TCP_ESTABLISHED -> TCP_FIN_WAIT1 1936 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible) 1937 * TCP_CLOSE_WAIT -> TCP_LAST_ACK 1938 * 1939 * are legal only when FIN has been sent (i.e. in window), 1940 * rather than queued out of window. Purists blame. 1941 * 1942 * F.e. "RFC state" is ESTABLISHED, 1943 * if Linux state is FIN-WAIT-1, but FIN is still not sent. 1944 * 1945 * The visible declinations are that sometimes 1946 * we enter time-wait state, when it is not required really 1947 * (harmless), do not send active resets, when they are 1948 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when 1949 * they look as CLOSING or LAST_ACK for Linux) 1950 * Probably, I missed some more holelets. 1951 * --ANK 1952 */ 1953 tcp_send_fin(sk); 1954 } 1955 1956 sk_stream_wait_close(sk, timeout); 1957 1958 adjudge_to_death: 1959 state = sk->sk_state; 1960 sock_hold(sk); 1961 sock_orphan(sk); 1962 1963 /* It is the last release_sock in its life. It will remove backlog. */ 1964 release_sock(sk); 1965 1966 1967 /* Now socket is owned by kernel and we acquire BH lock 1968 to finish close. No need to check for user refs. 1969 */ 1970 local_bh_disable(); 1971 bh_lock_sock(sk); 1972 WARN_ON(sock_owned_by_user(sk)); 1973 1974 percpu_counter_inc(sk->sk_prot->orphan_count); 1975 1976 /* Have we already been destroyed by a softirq or backlog? */ 1977 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE) 1978 goto out; 1979 1980 /* This is a (useful) BSD violating of the RFC. There is a 1981 * problem with TCP as specified in that the other end could 1982 * keep a socket open forever with no application left this end. 1983 * We use a 3 minute timeout (about the same as BSD) then kill 1984 * our end. If they send after that then tough - BUT: long enough 1985 * that we won't make the old 4*rto = almost no time - whoops 1986 * reset mistake. 1987 * 1988 * Nope, it was not mistake. It is really desired behaviour 1989 * f.e. on http servers, when such sockets are useless, but 1990 * consume significant resources. Let's do it with special 1991 * linger2 option. --ANK 1992 */ 1993 1994 if (sk->sk_state == TCP_FIN_WAIT2) { 1995 struct tcp_sock *tp = tcp_sk(sk); 1996 if (tp->linger2 < 0) { 1997 tcp_set_state(sk, TCP_CLOSE); 1998 tcp_send_active_reset(sk, GFP_ATOMIC); 1999 NET_INC_STATS_BH(sock_net(sk), 2000 LINUX_MIB_TCPABORTONLINGER); 2001 } else { 2002 const int tmo = tcp_fin_time(sk); 2003 2004 if (tmo > TCP_TIMEWAIT_LEN) { 2005 inet_csk_reset_keepalive_timer(sk, 2006 tmo - TCP_TIMEWAIT_LEN); 2007 } else { 2008 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 2009 goto out; 2010 } 2011 } 2012 } 2013 if (sk->sk_state != TCP_CLOSE) { 2014 int orphan_count = percpu_counter_read_positive( 2015 sk->sk_prot->orphan_count); 2016 2017 sk_mem_reclaim(sk); 2018 if (tcp_too_many_orphans(sk, orphan_count)) { 2019 if (net_ratelimit()) 2020 printk(KERN_INFO "TCP: too many of orphaned " 2021 "sockets\n"); 2022 tcp_set_state(sk, TCP_CLOSE); 2023 tcp_send_active_reset(sk, GFP_ATOMIC); 2024 NET_INC_STATS_BH(sock_net(sk), 2025 LINUX_MIB_TCPABORTONMEMORY); 2026 } 2027 } 2028 2029 if (sk->sk_state == TCP_CLOSE) 2030 inet_csk_destroy_sock(sk); 2031 /* Otherwise, socket is reprieved until protocol close. */ 2032 2033 out: 2034 bh_unlock_sock(sk); 2035 local_bh_enable(); 2036 sock_put(sk); 2037 } 2038 EXPORT_SYMBOL(tcp_close); 2039 2040 /* These states need RST on ABORT according to RFC793 */ 2041 2042 static inline int tcp_need_reset(int state) 2043 { 2044 return (1 << state) & 2045 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 | 2046 TCPF_FIN_WAIT2 | TCPF_SYN_RECV); 2047 } 2048 2049 int tcp_disconnect(struct sock *sk, int flags) 2050 { 2051 struct inet_sock *inet = inet_sk(sk); 2052 struct inet_connection_sock *icsk = inet_csk(sk); 2053 struct tcp_sock *tp = tcp_sk(sk); 2054 int err = 0; 2055 int old_state = sk->sk_state; 2056 2057 if (old_state != TCP_CLOSE) 2058 tcp_set_state(sk, TCP_CLOSE); 2059 2060 /* ABORT function of RFC793 */ 2061 if (old_state == TCP_LISTEN) { 2062 inet_csk_listen_stop(sk); 2063 } else if (tcp_need_reset(old_state) || 2064 (tp->snd_nxt != tp->write_seq && 2065 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) { 2066 /* The last check adjusts for discrepancy of Linux wrt. RFC 2067 * states 2068 */ 2069 tcp_send_active_reset(sk, gfp_any()); 2070 sk->sk_err = ECONNRESET; 2071 } else if (old_state == TCP_SYN_SENT) 2072 sk->sk_err = ECONNRESET; 2073 2074 tcp_clear_xmit_timers(sk); 2075 __skb_queue_purge(&sk->sk_receive_queue); 2076 tcp_write_queue_purge(sk); 2077 __skb_queue_purge(&tp->out_of_order_queue); 2078 #ifdef CONFIG_NET_DMA 2079 __skb_queue_purge(&sk->sk_async_wait_queue); 2080 #endif 2081 2082 inet->inet_dport = 0; 2083 2084 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK)) 2085 inet_reset_saddr(sk); 2086 2087 sk->sk_shutdown = 0; 2088 sock_reset_flag(sk, SOCK_DONE); 2089 tp->srtt = 0; 2090 if ((tp->write_seq += tp->max_window + 2) == 0) 2091 tp->write_seq = 1; 2092 icsk->icsk_backoff = 0; 2093 tp->snd_cwnd = 2; 2094 icsk->icsk_probes_out = 0; 2095 tp->packets_out = 0; 2096 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 2097 tp->snd_cwnd_cnt = 0; 2098 tp->bytes_acked = 0; 2099 tp->window_clamp = 0; 2100 tcp_set_ca_state(sk, TCP_CA_Open); 2101 tcp_clear_retrans(tp); 2102 inet_csk_delack_init(sk); 2103 tcp_init_send_head(sk); 2104 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt)); 2105 __sk_dst_reset(sk); 2106 2107 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash); 2108 2109 sk->sk_error_report(sk); 2110 return err; 2111 } 2112 EXPORT_SYMBOL(tcp_disconnect); 2113 2114 /* 2115 * Socket option code for TCP. 2116 */ 2117 static int do_tcp_setsockopt(struct sock *sk, int level, 2118 int optname, char __user *optval, unsigned int optlen) 2119 { 2120 struct tcp_sock *tp = tcp_sk(sk); 2121 struct inet_connection_sock *icsk = inet_csk(sk); 2122 int val; 2123 int err = 0; 2124 2125 /* These are data/string values, all the others are ints */ 2126 switch (optname) { 2127 case TCP_CONGESTION: { 2128 char name[TCP_CA_NAME_MAX]; 2129 2130 if (optlen < 1) 2131 return -EINVAL; 2132 2133 val = strncpy_from_user(name, optval, 2134 min_t(long, TCP_CA_NAME_MAX-1, optlen)); 2135 if (val < 0) 2136 return -EFAULT; 2137 name[val] = 0; 2138 2139 lock_sock(sk); 2140 err = tcp_set_congestion_control(sk, name); 2141 release_sock(sk); 2142 return err; 2143 } 2144 case TCP_COOKIE_TRANSACTIONS: { 2145 struct tcp_cookie_transactions ctd; 2146 struct tcp_cookie_values *cvp = NULL; 2147 2148 if (sizeof(ctd) > optlen) 2149 return -EINVAL; 2150 if (copy_from_user(&ctd, optval, sizeof(ctd))) 2151 return -EFAULT; 2152 2153 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) || 2154 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED) 2155 return -EINVAL; 2156 2157 if (ctd.tcpct_cookie_desired == 0) { 2158 /* default to global value */ 2159 } else if ((0x1 & ctd.tcpct_cookie_desired) || 2160 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX || 2161 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) { 2162 return -EINVAL; 2163 } 2164 2165 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) { 2166 /* Supercedes all other values */ 2167 lock_sock(sk); 2168 if (tp->cookie_values != NULL) { 2169 kref_put(&tp->cookie_values->kref, 2170 tcp_cookie_values_release); 2171 tp->cookie_values = NULL; 2172 } 2173 tp->rx_opt.cookie_in_always = 0; /* false */ 2174 tp->rx_opt.cookie_out_never = 1; /* true */ 2175 release_sock(sk); 2176 return err; 2177 } 2178 2179 /* Allocate ancillary memory before locking. 2180 */ 2181 if (ctd.tcpct_used > 0 || 2182 (tp->cookie_values == NULL && 2183 (sysctl_tcp_cookie_size > 0 || 2184 ctd.tcpct_cookie_desired > 0 || 2185 ctd.tcpct_s_data_desired > 0))) { 2186 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used, 2187 GFP_KERNEL); 2188 if (cvp == NULL) 2189 return -ENOMEM; 2190 2191 kref_init(&cvp->kref); 2192 } 2193 lock_sock(sk); 2194 tp->rx_opt.cookie_in_always = 2195 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags); 2196 tp->rx_opt.cookie_out_never = 0; /* false */ 2197 2198 if (tp->cookie_values != NULL) { 2199 if (cvp != NULL) { 2200 /* Changed values are recorded by a changed 2201 * pointer, ensuring the cookie will differ, 2202 * without separately hashing each value later. 2203 */ 2204 kref_put(&tp->cookie_values->kref, 2205 tcp_cookie_values_release); 2206 } else { 2207 cvp = tp->cookie_values; 2208 } 2209 } 2210 2211 if (cvp != NULL) { 2212 cvp->cookie_desired = ctd.tcpct_cookie_desired; 2213 2214 if (ctd.tcpct_used > 0) { 2215 memcpy(cvp->s_data_payload, ctd.tcpct_value, 2216 ctd.tcpct_used); 2217 cvp->s_data_desired = ctd.tcpct_used; 2218 cvp->s_data_constant = 1; /* true */ 2219 } else { 2220 /* No constant payload data. */ 2221 cvp->s_data_desired = ctd.tcpct_s_data_desired; 2222 cvp->s_data_constant = 0; /* false */ 2223 } 2224 2225 tp->cookie_values = cvp; 2226 } 2227 release_sock(sk); 2228 return err; 2229 } 2230 default: 2231 /* fallthru */ 2232 break; 2233 } 2234 2235 if (optlen < sizeof(int)) 2236 return -EINVAL; 2237 2238 if (get_user(val, (int __user *)optval)) 2239 return -EFAULT; 2240 2241 lock_sock(sk); 2242 2243 switch (optname) { 2244 case TCP_MAXSEG: 2245 /* Values greater than interface MTU won't take effect. However 2246 * at the point when this call is done we typically don't yet 2247 * know which interface is going to be used */ 2248 if (val < 8 || val > MAX_TCP_WINDOW) { 2249 err = -EINVAL; 2250 break; 2251 } 2252 tp->rx_opt.user_mss = val; 2253 break; 2254 2255 case TCP_NODELAY: 2256 if (val) { 2257 /* TCP_NODELAY is weaker than TCP_CORK, so that 2258 * this option on corked socket is remembered, but 2259 * it is not activated until cork is cleared. 2260 * 2261 * However, when TCP_NODELAY is set we make 2262 * an explicit push, which overrides even TCP_CORK 2263 * for currently queued segments. 2264 */ 2265 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH; 2266 tcp_push_pending_frames(sk); 2267 } else { 2268 tp->nonagle &= ~TCP_NAGLE_OFF; 2269 } 2270 break; 2271 2272 case TCP_THIN_LINEAR_TIMEOUTS: 2273 if (val < 0 || val > 1) 2274 err = -EINVAL; 2275 else 2276 tp->thin_lto = val; 2277 break; 2278 2279 case TCP_THIN_DUPACK: 2280 if (val < 0 || val > 1) 2281 err = -EINVAL; 2282 else 2283 tp->thin_dupack = val; 2284 break; 2285 2286 case TCP_CORK: 2287 /* When set indicates to always queue non-full frames. 2288 * Later the user clears this option and we transmit 2289 * any pending partial frames in the queue. This is 2290 * meant to be used alongside sendfile() to get properly 2291 * filled frames when the user (for example) must write 2292 * out headers with a write() call first and then use 2293 * sendfile to send out the data parts. 2294 * 2295 * TCP_CORK can be set together with TCP_NODELAY and it is 2296 * stronger than TCP_NODELAY. 2297 */ 2298 if (val) { 2299 tp->nonagle |= TCP_NAGLE_CORK; 2300 } else { 2301 tp->nonagle &= ~TCP_NAGLE_CORK; 2302 if (tp->nonagle&TCP_NAGLE_OFF) 2303 tp->nonagle |= TCP_NAGLE_PUSH; 2304 tcp_push_pending_frames(sk); 2305 } 2306 break; 2307 2308 case TCP_KEEPIDLE: 2309 if (val < 1 || val > MAX_TCP_KEEPIDLE) 2310 err = -EINVAL; 2311 else { 2312 tp->keepalive_time = val * HZ; 2313 if (sock_flag(sk, SOCK_KEEPOPEN) && 2314 !((1 << sk->sk_state) & 2315 (TCPF_CLOSE | TCPF_LISTEN))) { 2316 u32 elapsed = keepalive_time_elapsed(tp); 2317 if (tp->keepalive_time > elapsed) 2318 elapsed = tp->keepalive_time - elapsed; 2319 else 2320 elapsed = 0; 2321 inet_csk_reset_keepalive_timer(sk, elapsed); 2322 } 2323 } 2324 break; 2325 case TCP_KEEPINTVL: 2326 if (val < 1 || val > MAX_TCP_KEEPINTVL) 2327 err = -EINVAL; 2328 else 2329 tp->keepalive_intvl = val * HZ; 2330 break; 2331 case TCP_KEEPCNT: 2332 if (val < 1 || val > MAX_TCP_KEEPCNT) 2333 err = -EINVAL; 2334 else 2335 tp->keepalive_probes = val; 2336 break; 2337 case TCP_SYNCNT: 2338 if (val < 1 || val > MAX_TCP_SYNCNT) 2339 err = -EINVAL; 2340 else 2341 icsk->icsk_syn_retries = val; 2342 break; 2343 2344 case TCP_LINGER2: 2345 if (val < 0) 2346 tp->linger2 = -1; 2347 else if (val > sysctl_tcp_fin_timeout / HZ) 2348 tp->linger2 = 0; 2349 else 2350 tp->linger2 = val * HZ; 2351 break; 2352 2353 case TCP_DEFER_ACCEPT: 2354 /* Translate value in seconds to number of retransmits */ 2355 icsk->icsk_accept_queue.rskq_defer_accept = 2356 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ, 2357 TCP_RTO_MAX / HZ); 2358 break; 2359 2360 case TCP_WINDOW_CLAMP: 2361 if (!val) { 2362 if (sk->sk_state != TCP_CLOSE) { 2363 err = -EINVAL; 2364 break; 2365 } 2366 tp->window_clamp = 0; 2367 } else 2368 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ? 2369 SOCK_MIN_RCVBUF / 2 : val; 2370 break; 2371 2372 case TCP_QUICKACK: 2373 if (!val) { 2374 icsk->icsk_ack.pingpong = 1; 2375 } else { 2376 icsk->icsk_ack.pingpong = 0; 2377 if ((1 << sk->sk_state) & 2378 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) && 2379 inet_csk_ack_scheduled(sk)) { 2380 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 2381 tcp_cleanup_rbuf(sk, 1); 2382 if (!(val & 1)) 2383 icsk->icsk_ack.pingpong = 1; 2384 } 2385 } 2386 break; 2387 2388 #ifdef CONFIG_TCP_MD5SIG 2389 case TCP_MD5SIG: 2390 /* Read the IP->Key mappings from userspace */ 2391 err = tp->af_specific->md5_parse(sk, optval, optlen); 2392 break; 2393 #endif 2394 2395 default: 2396 err = -ENOPROTOOPT; 2397 break; 2398 } 2399 2400 release_sock(sk); 2401 return err; 2402 } 2403 2404 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval, 2405 unsigned int optlen) 2406 { 2407 struct inet_connection_sock *icsk = inet_csk(sk); 2408 2409 if (level != SOL_TCP) 2410 return icsk->icsk_af_ops->setsockopt(sk, level, optname, 2411 optval, optlen); 2412 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2413 } 2414 EXPORT_SYMBOL(tcp_setsockopt); 2415 2416 #ifdef CONFIG_COMPAT 2417 int compat_tcp_setsockopt(struct sock *sk, int level, int optname, 2418 char __user *optval, unsigned int optlen) 2419 { 2420 if (level != SOL_TCP) 2421 return inet_csk_compat_setsockopt(sk, level, optname, 2422 optval, optlen); 2423 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2424 } 2425 EXPORT_SYMBOL(compat_tcp_setsockopt); 2426 #endif 2427 2428 /* Return information about state of tcp endpoint in API format. */ 2429 void tcp_get_info(struct sock *sk, struct tcp_info *info) 2430 { 2431 struct tcp_sock *tp = tcp_sk(sk); 2432 const struct inet_connection_sock *icsk = inet_csk(sk); 2433 u32 now = tcp_time_stamp; 2434 2435 memset(info, 0, sizeof(*info)); 2436 2437 info->tcpi_state = sk->sk_state; 2438 info->tcpi_ca_state = icsk->icsk_ca_state; 2439 info->tcpi_retransmits = icsk->icsk_retransmits; 2440 info->tcpi_probes = icsk->icsk_probes_out; 2441 info->tcpi_backoff = icsk->icsk_backoff; 2442 2443 if (tp->rx_opt.tstamp_ok) 2444 info->tcpi_options |= TCPI_OPT_TIMESTAMPS; 2445 if (tcp_is_sack(tp)) 2446 info->tcpi_options |= TCPI_OPT_SACK; 2447 if (tp->rx_opt.wscale_ok) { 2448 info->tcpi_options |= TCPI_OPT_WSCALE; 2449 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale; 2450 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale; 2451 } 2452 2453 if (tp->ecn_flags&TCP_ECN_OK) 2454 info->tcpi_options |= TCPI_OPT_ECN; 2455 2456 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto); 2457 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato); 2458 info->tcpi_snd_mss = tp->mss_cache; 2459 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss; 2460 2461 if (sk->sk_state == TCP_LISTEN) { 2462 info->tcpi_unacked = sk->sk_ack_backlog; 2463 info->tcpi_sacked = sk->sk_max_ack_backlog; 2464 } else { 2465 info->tcpi_unacked = tp->packets_out; 2466 info->tcpi_sacked = tp->sacked_out; 2467 } 2468 info->tcpi_lost = tp->lost_out; 2469 info->tcpi_retrans = tp->retrans_out; 2470 info->tcpi_fackets = tp->fackets_out; 2471 2472 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime); 2473 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime); 2474 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp); 2475 2476 info->tcpi_pmtu = icsk->icsk_pmtu_cookie; 2477 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh; 2478 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3; 2479 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2; 2480 info->tcpi_snd_ssthresh = tp->snd_ssthresh; 2481 info->tcpi_snd_cwnd = tp->snd_cwnd; 2482 info->tcpi_advmss = tp->advmss; 2483 info->tcpi_reordering = tp->reordering; 2484 2485 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3; 2486 info->tcpi_rcv_space = tp->rcvq_space.space; 2487 2488 info->tcpi_total_retrans = tp->total_retrans; 2489 } 2490 EXPORT_SYMBOL_GPL(tcp_get_info); 2491 2492 static int do_tcp_getsockopt(struct sock *sk, int level, 2493 int optname, char __user *optval, int __user *optlen) 2494 { 2495 struct inet_connection_sock *icsk = inet_csk(sk); 2496 struct tcp_sock *tp = tcp_sk(sk); 2497 int val, len; 2498 2499 if (get_user(len, optlen)) 2500 return -EFAULT; 2501 2502 len = min_t(unsigned int, len, sizeof(int)); 2503 2504 if (len < 0) 2505 return -EINVAL; 2506 2507 switch (optname) { 2508 case TCP_MAXSEG: 2509 val = tp->mss_cache; 2510 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 2511 val = tp->rx_opt.user_mss; 2512 break; 2513 case TCP_NODELAY: 2514 val = !!(tp->nonagle&TCP_NAGLE_OFF); 2515 break; 2516 case TCP_CORK: 2517 val = !!(tp->nonagle&TCP_NAGLE_CORK); 2518 break; 2519 case TCP_KEEPIDLE: 2520 val = keepalive_time_when(tp) / HZ; 2521 break; 2522 case TCP_KEEPINTVL: 2523 val = keepalive_intvl_when(tp) / HZ; 2524 break; 2525 case TCP_KEEPCNT: 2526 val = keepalive_probes(tp); 2527 break; 2528 case TCP_SYNCNT: 2529 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries; 2530 break; 2531 case TCP_LINGER2: 2532 val = tp->linger2; 2533 if (val >= 0) 2534 val = (val ? : sysctl_tcp_fin_timeout) / HZ; 2535 break; 2536 case TCP_DEFER_ACCEPT: 2537 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept, 2538 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ); 2539 break; 2540 case TCP_WINDOW_CLAMP: 2541 val = tp->window_clamp; 2542 break; 2543 case TCP_INFO: { 2544 struct tcp_info info; 2545 2546 if (get_user(len, optlen)) 2547 return -EFAULT; 2548 2549 tcp_get_info(sk, &info); 2550 2551 len = min_t(unsigned int, len, sizeof(info)); 2552 if (put_user(len, optlen)) 2553 return -EFAULT; 2554 if (copy_to_user(optval, &info, len)) 2555 return -EFAULT; 2556 return 0; 2557 } 2558 case TCP_QUICKACK: 2559 val = !icsk->icsk_ack.pingpong; 2560 break; 2561 2562 case TCP_CONGESTION: 2563 if (get_user(len, optlen)) 2564 return -EFAULT; 2565 len = min_t(unsigned int, len, TCP_CA_NAME_MAX); 2566 if (put_user(len, optlen)) 2567 return -EFAULT; 2568 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len)) 2569 return -EFAULT; 2570 return 0; 2571 2572 case TCP_COOKIE_TRANSACTIONS: { 2573 struct tcp_cookie_transactions ctd; 2574 struct tcp_cookie_values *cvp = tp->cookie_values; 2575 2576 if (get_user(len, optlen)) 2577 return -EFAULT; 2578 if (len < sizeof(ctd)) 2579 return -EINVAL; 2580 2581 memset(&ctd, 0, sizeof(ctd)); 2582 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ? 2583 TCP_COOKIE_IN_ALWAYS : 0) 2584 | (tp->rx_opt.cookie_out_never ? 2585 TCP_COOKIE_OUT_NEVER : 0); 2586 2587 if (cvp != NULL) { 2588 ctd.tcpct_flags |= (cvp->s_data_in ? 2589 TCP_S_DATA_IN : 0) 2590 | (cvp->s_data_out ? 2591 TCP_S_DATA_OUT : 0); 2592 2593 ctd.tcpct_cookie_desired = cvp->cookie_desired; 2594 ctd.tcpct_s_data_desired = cvp->s_data_desired; 2595 2596 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0], 2597 cvp->cookie_pair_size); 2598 ctd.tcpct_used = cvp->cookie_pair_size; 2599 } 2600 2601 if (put_user(sizeof(ctd), optlen)) 2602 return -EFAULT; 2603 if (copy_to_user(optval, &ctd, sizeof(ctd))) 2604 return -EFAULT; 2605 return 0; 2606 } 2607 case TCP_THIN_LINEAR_TIMEOUTS: 2608 val = tp->thin_lto; 2609 break; 2610 case TCP_THIN_DUPACK: 2611 val = tp->thin_dupack; 2612 break; 2613 default: 2614 return -ENOPROTOOPT; 2615 } 2616 2617 if (put_user(len, optlen)) 2618 return -EFAULT; 2619 if (copy_to_user(optval, &val, len)) 2620 return -EFAULT; 2621 return 0; 2622 } 2623 2624 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval, 2625 int __user *optlen) 2626 { 2627 struct inet_connection_sock *icsk = inet_csk(sk); 2628 2629 if (level != SOL_TCP) 2630 return icsk->icsk_af_ops->getsockopt(sk, level, optname, 2631 optval, optlen); 2632 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2633 } 2634 EXPORT_SYMBOL(tcp_getsockopt); 2635 2636 #ifdef CONFIG_COMPAT 2637 int compat_tcp_getsockopt(struct sock *sk, int level, int optname, 2638 char __user *optval, int __user *optlen) 2639 { 2640 if (level != SOL_TCP) 2641 return inet_csk_compat_getsockopt(sk, level, optname, 2642 optval, optlen); 2643 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2644 } 2645 EXPORT_SYMBOL(compat_tcp_getsockopt); 2646 #endif 2647 2648 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features) 2649 { 2650 struct sk_buff *segs = ERR_PTR(-EINVAL); 2651 struct tcphdr *th; 2652 unsigned thlen; 2653 unsigned int seq; 2654 __be32 delta; 2655 unsigned int oldlen; 2656 unsigned int mss; 2657 2658 if (!pskb_may_pull(skb, sizeof(*th))) 2659 goto out; 2660 2661 th = tcp_hdr(skb); 2662 thlen = th->doff * 4; 2663 if (thlen < sizeof(*th)) 2664 goto out; 2665 2666 if (!pskb_may_pull(skb, thlen)) 2667 goto out; 2668 2669 oldlen = (u16)~skb->len; 2670 __skb_pull(skb, thlen); 2671 2672 mss = skb_shinfo(skb)->gso_size; 2673 if (unlikely(skb->len <= mss)) 2674 goto out; 2675 2676 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) { 2677 /* Packet is from an untrusted source, reset gso_segs. */ 2678 int type = skb_shinfo(skb)->gso_type; 2679 2680 if (unlikely(type & 2681 ~(SKB_GSO_TCPV4 | 2682 SKB_GSO_DODGY | 2683 SKB_GSO_TCP_ECN | 2684 SKB_GSO_TCPV6 | 2685 0) || 2686 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))) 2687 goto out; 2688 2689 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss); 2690 2691 segs = NULL; 2692 goto out; 2693 } 2694 2695 segs = skb_segment(skb, features); 2696 if (IS_ERR(segs)) 2697 goto out; 2698 2699 delta = htonl(oldlen + (thlen + mss)); 2700 2701 skb = segs; 2702 th = tcp_hdr(skb); 2703 seq = ntohl(th->seq); 2704 2705 do { 2706 th->fin = th->psh = 0; 2707 2708 th->check = ~csum_fold((__force __wsum)((__force u32)th->check + 2709 (__force u32)delta)); 2710 if (skb->ip_summed != CHECKSUM_PARTIAL) 2711 th->check = 2712 csum_fold(csum_partial(skb_transport_header(skb), 2713 thlen, skb->csum)); 2714 2715 seq += mss; 2716 skb = skb->next; 2717 th = tcp_hdr(skb); 2718 2719 th->seq = htonl(seq); 2720 th->cwr = 0; 2721 } while (skb->next); 2722 2723 delta = htonl(oldlen + (skb->tail - skb->transport_header) + 2724 skb->data_len); 2725 th->check = ~csum_fold((__force __wsum)((__force u32)th->check + 2726 (__force u32)delta)); 2727 if (skb->ip_summed != CHECKSUM_PARTIAL) 2728 th->check = csum_fold(csum_partial(skb_transport_header(skb), 2729 thlen, skb->csum)); 2730 2731 out: 2732 return segs; 2733 } 2734 EXPORT_SYMBOL(tcp_tso_segment); 2735 2736 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb) 2737 { 2738 struct sk_buff **pp = NULL; 2739 struct sk_buff *p; 2740 struct tcphdr *th; 2741 struct tcphdr *th2; 2742 unsigned int len; 2743 unsigned int thlen; 2744 __be32 flags; 2745 unsigned int mss = 1; 2746 unsigned int hlen; 2747 unsigned int off; 2748 int flush = 1; 2749 int i; 2750 2751 off = skb_gro_offset(skb); 2752 hlen = off + sizeof(*th); 2753 th = skb_gro_header_fast(skb, off); 2754 if (skb_gro_header_hard(skb, hlen)) { 2755 th = skb_gro_header_slow(skb, hlen, off); 2756 if (unlikely(!th)) 2757 goto out; 2758 } 2759 2760 thlen = th->doff * 4; 2761 if (thlen < sizeof(*th)) 2762 goto out; 2763 2764 hlen = off + thlen; 2765 if (skb_gro_header_hard(skb, hlen)) { 2766 th = skb_gro_header_slow(skb, hlen, off); 2767 if (unlikely(!th)) 2768 goto out; 2769 } 2770 2771 skb_gro_pull(skb, thlen); 2772 2773 len = skb_gro_len(skb); 2774 flags = tcp_flag_word(th); 2775 2776 for (; (p = *head); head = &p->next) { 2777 if (!NAPI_GRO_CB(p)->same_flow) 2778 continue; 2779 2780 th2 = tcp_hdr(p); 2781 2782 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) { 2783 NAPI_GRO_CB(p)->same_flow = 0; 2784 continue; 2785 } 2786 2787 goto found; 2788 } 2789 2790 goto out_check_final; 2791 2792 found: 2793 flush = NAPI_GRO_CB(p)->flush; 2794 flush |= (__force int)(flags & TCP_FLAG_CWR); 2795 flush |= (__force int)((flags ^ tcp_flag_word(th2)) & 2796 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH)); 2797 flush |= (__force int)(th->ack_seq ^ th2->ack_seq); 2798 for (i = sizeof(*th); i < thlen; i += 4) 2799 flush |= *(u32 *)((u8 *)th + i) ^ 2800 *(u32 *)((u8 *)th2 + i); 2801 2802 mss = skb_shinfo(p)->gso_size; 2803 2804 flush |= (len - 1) >= mss; 2805 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq); 2806 2807 if (flush || skb_gro_receive(head, skb)) { 2808 mss = 1; 2809 goto out_check_final; 2810 } 2811 2812 p = *head; 2813 th2 = tcp_hdr(p); 2814 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH); 2815 2816 out_check_final: 2817 flush = len < mss; 2818 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH | 2819 TCP_FLAG_RST | TCP_FLAG_SYN | 2820 TCP_FLAG_FIN)); 2821 2822 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush)) 2823 pp = head; 2824 2825 out: 2826 NAPI_GRO_CB(skb)->flush |= flush; 2827 2828 return pp; 2829 } 2830 EXPORT_SYMBOL(tcp_gro_receive); 2831 2832 int tcp_gro_complete(struct sk_buff *skb) 2833 { 2834 struct tcphdr *th = tcp_hdr(skb); 2835 2836 skb->csum_start = skb_transport_header(skb) - skb->head; 2837 skb->csum_offset = offsetof(struct tcphdr, check); 2838 skb->ip_summed = CHECKSUM_PARTIAL; 2839 2840 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count; 2841 2842 if (th->cwr) 2843 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 2844 2845 return 0; 2846 } 2847 EXPORT_SYMBOL(tcp_gro_complete); 2848 2849 #ifdef CONFIG_TCP_MD5SIG 2850 static unsigned long tcp_md5sig_users; 2851 static struct tcp_md5sig_pool * __percpu *tcp_md5sig_pool; 2852 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock); 2853 2854 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool * __percpu *pool) 2855 { 2856 int cpu; 2857 for_each_possible_cpu(cpu) { 2858 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu); 2859 if (p) { 2860 if (p->md5_desc.tfm) 2861 crypto_free_hash(p->md5_desc.tfm); 2862 kfree(p); 2863 } 2864 } 2865 free_percpu(pool); 2866 } 2867 2868 void tcp_free_md5sig_pool(void) 2869 { 2870 struct tcp_md5sig_pool * __percpu *pool = NULL; 2871 2872 spin_lock_bh(&tcp_md5sig_pool_lock); 2873 if (--tcp_md5sig_users == 0) { 2874 pool = tcp_md5sig_pool; 2875 tcp_md5sig_pool = NULL; 2876 } 2877 spin_unlock_bh(&tcp_md5sig_pool_lock); 2878 if (pool) 2879 __tcp_free_md5sig_pool(pool); 2880 } 2881 EXPORT_SYMBOL(tcp_free_md5sig_pool); 2882 2883 static struct tcp_md5sig_pool * __percpu * 2884 __tcp_alloc_md5sig_pool(struct sock *sk) 2885 { 2886 int cpu; 2887 struct tcp_md5sig_pool * __percpu *pool; 2888 2889 pool = alloc_percpu(struct tcp_md5sig_pool *); 2890 if (!pool) 2891 return NULL; 2892 2893 for_each_possible_cpu(cpu) { 2894 struct tcp_md5sig_pool *p; 2895 struct crypto_hash *hash; 2896 2897 p = kzalloc(sizeof(*p), sk->sk_allocation); 2898 if (!p) 2899 goto out_free; 2900 *per_cpu_ptr(pool, cpu) = p; 2901 2902 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC); 2903 if (!hash || IS_ERR(hash)) 2904 goto out_free; 2905 2906 p->md5_desc.tfm = hash; 2907 } 2908 return pool; 2909 out_free: 2910 __tcp_free_md5sig_pool(pool); 2911 return NULL; 2912 } 2913 2914 struct tcp_md5sig_pool * __percpu *tcp_alloc_md5sig_pool(struct sock *sk) 2915 { 2916 struct tcp_md5sig_pool * __percpu *pool; 2917 int alloc = 0; 2918 2919 retry: 2920 spin_lock_bh(&tcp_md5sig_pool_lock); 2921 pool = tcp_md5sig_pool; 2922 if (tcp_md5sig_users++ == 0) { 2923 alloc = 1; 2924 spin_unlock_bh(&tcp_md5sig_pool_lock); 2925 } else if (!pool) { 2926 tcp_md5sig_users--; 2927 spin_unlock_bh(&tcp_md5sig_pool_lock); 2928 cpu_relax(); 2929 goto retry; 2930 } else 2931 spin_unlock_bh(&tcp_md5sig_pool_lock); 2932 2933 if (alloc) { 2934 /* we cannot hold spinlock here because this may sleep. */ 2935 struct tcp_md5sig_pool * __percpu *p; 2936 2937 p = __tcp_alloc_md5sig_pool(sk); 2938 spin_lock_bh(&tcp_md5sig_pool_lock); 2939 if (!p) { 2940 tcp_md5sig_users--; 2941 spin_unlock_bh(&tcp_md5sig_pool_lock); 2942 return NULL; 2943 } 2944 pool = tcp_md5sig_pool; 2945 if (pool) { 2946 /* oops, it has already been assigned. */ 2947 spin_unlock_bh(&tcp_md5sig_pool_lock); 2948 __tcp_free_md5sig_pool(p); 2949 } else { 2950 tcp_md5sig_pool = pool = p; 2951 spin_unlock_bh(&tcp_md5sig_pool_lock); 2952 } 2953 } 2954 return pool; 2955 } 2956 EXPORT_SYMBOL(tcp_alloc_md5sig_pool); 2957 2958 2959 /** 2960 * tcp_get_md5sig_pool - get md5sig_pool for this user 2961 * 2962 * We use percpu structure, so if we succeed, we exit with preemption 2963 * and BH disabled, to make sure another thread or softirq handling 2964 * wont try to get same context. 2965 */ 2966 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void) 2967 { 2968 struct tcp_md5sig_pool * __percpu *p; 2969 2970 local_bh_disable(); 2971 2972 spin_lock(&tcp_md5sig_pool_lock); 2973 p = tcp_md5sig_pool; 2974 if (p) 2975 tcp_md5sig_users++; 2976 spin_unlock(&tcp_md5sig_pool_lock); 2977 2978 if (p) 2979 return *this_cpu_ptr(p); 2980 2981 local_bh_enable(); 2982 return NULL; 2983 } 2984 EXPORT_SYMBOL(tcp_get_md5sig_pool); 2985 2986 void tcp_put_md5sig_pool(void) 2987 { 2988 local_bh_enable(); 2989 tcp_free_md5sig_pool(); 2990 } 2991 EXPORT_SYMBOL(tcp_put_md5sig_pool); 2992 2993 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp, 2994 struct tcphdr *th) 2995 { 2996 struct scatterlist sg; 2997 int err; 2998 2999 __sum16 old_checksum = th->check; 3000 th->check = 0; 3001 /* options aren't included in the hash */ 3002 sg_init_one(&sg, th, sizeof(struct tcphdr)); 3003 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr)); 3004 th->check = old_checksum; 3005 return err; 3006 } 3007 EXPORT_SYMBOL(tcp_md5_hash_header); 3008 3009 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp, 3010 struct sk_buff *skb, unsigned header_len) 3011 { 3012 struct scatterlist sg; 3013 const struct tcphdr *tp = tcp_hdr(skb); 3014 struct hash_desc *desc = &hp->md5_desc; 3015 unsigned i; 3016 const unsigned head_data_len = skb_headlen(skb) > header_len ? 3017 skb_headlen(skb) - header_len : 0; 3018 const struct skb_shared_info *shi = skb_shinfo(skb); 3019 struct sk_buff *frag_iter; 3020 3021 sg_init_table(&sg, 1); 3022 3023 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len); 3024 if (crypto_hash_update(desc, &sg, head_data_len)) 3025 return 1; 3026 3027 for (i = 0; i < shi->nr_frags; ++i) { 3028 const struct skb_frag_struct *f = &shi->frags[i]; 3029 sg_set_page(&sg, f->page, f->size, f->page_offset); 3030 if (crypto_hash_update(desc, &sg, f->size)) 3031 return 1; 3032 } 3033 3034 skb_walk_frags(skb, frag_iter) 3035 if (tcp_md5_hash_skb_data(hp, frag_iter, 0)) 3036 return 1; 3037 3038 return 0; 3039 } 3040 EXPORT_SYMBOL(tcp_md5_hash_skb_data); 3041 3042 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key) 3043 { 3044 struct scatterlist sg; 3045 3046 sg_init_one(&sg, key->key, key->keylen); 3047 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen); 3048 } 3049 EXPORT_SYMBOL(tcp_md5_hash_key); 3050 3051 #endif 3052 3053 /** 3054 * Each Responder maintains up to two secret values concurrently for 3055 * efficient secret rollover. Each secret value has 4 states: 3056 * 3057 * Generating. (tcp_secret_generating != tcp_secret_primary) 3058 * Generates new Responder-Cookies, but not yet used for primary 3059 * verification. This is a short-term state, typically lasting only 3060 * one round trip time (RTT). 3061 * 3062 * Primary. (tcp_secret_generating == tcp_secret_primary) 3063 * Used both for generation and primary verification. 3064 * 3065 * Retiring. (tcp_secret_retiring != tcp_secret_secondary) 3066 * Used for verification, until the first failure that can be 3067 * verified by the newer Generating secret. At that time, this 3068 * cookie's state is changed to Secondary, and the Generating 3069 * cookie's state is changed to Primary. This is a short-term state, 3070 * typically lasting only one round trip time (RTT). 3071 * 3072 * Secondary. (tcp_secret_retiring == tcp_secret_secondary) 3073 * Used for secondary verification, after primary verification 3074 * failures. This state lasts no more than twice the Maximum Segment 3075 * Lifetime (2MSL). Then, the secret is discarded. 3076 */ 3077 struct tcp_cookie_secret { 3078 /* The secret is divided into two parts. The digest part is the 3079 * equivalent of previously hashing a secret and saving the state, 3080 * and serves as an initialization vector (IV). The message part 3081 * serves as the trailing secret. 3082 */ 3083 u32 secrets[COOKIE_WORKSPACE_WORDS]; 3084 unsigned long expires; 3085 }; 3086 3087 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL) 3088 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2) 3089 #define TCP_SECRET_LIFE (HZ * 600) 3090 3091 static struct tcp_cookie_secret tcp_secret_one; 3092 static struct tcp_cookie_secret tcp_secret_two; 3093 3094 /* Essentially a circular list, without dynamic allocation. */ 3095 static struct tcp_cookie_secret *tcp_secret_generating; 3096 static struct tcp_cookie_secret *tcp_secret_primary; 3097 static struct tcp_cookie_secret *tcp_secret_retiring; 3098 static struct tcp_cookie_secret *tcp_secret_secondary; 3099 3100 static DEFINE_SPINLOCK(tcp_secret_locker); 3101 3102 /* Select a pseudo-random word in the cookie workspace. 3103 */ 3104 static inline u32 tcp_cookie_work(const u32 *ws, const int n) 3105 { 3106 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])]; 3107 } 3108 3109 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed. 3110 * Called in softirq context. 3111 * Returns: 0 for success. 3112 */ 3113 int tcp_cookie_generator(u32 *bakery) 3114 { 3115 unsigned long jiffy = jiffies; 3116 3117 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) { 3118 spin_lock_bh(&tcp_secret_locker); 3119 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) { 3120 /* refreshed by another */ 3121 memcpy(bakery, 3122 &tcp_secret_generating->secrets[0], 3123 COOKIE_WORKSPACE_WORDS); 3124 } else { 3125 /* still needs refreshing */ 3126 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS); 3127 3128 /* The first time, paranoia assumes that the 3129 * randomization function isn't as strong. But, 3130 * this secret initialization is delayed until 3131 * the last possible moment (packet arrival). 3132 * Although that time is observable, it is 3133 * unpredictably variable. Mash in the most 3134 * volatile clock bits available, and expire the 3135 * secret extra quickly. 3136 */ 3137 if (unlikely(tcp_secret_primary->expires == 3138 tcp_secret_secondary->expires)) { 3139 struct timespec tv; 3140 3141 getnstimeofday(&tv); 3142 bakery[COOKIE_DIGEST_WORDS+0] ^= 3143 (u32)tv.tv_nsec; 3144 3145 tcp_secret_secondary->expires = jiffy 3146 + TCP_SECRET_1MSL 3147 + (0x0f & tcp_cookie_work(bakery, 0)); 3148 } else { 3149 tcp_secret_secondary->expires = jiffy 3150 + TCP_SECRET_LIFE 3151 + (0xff & tcp_cookie_work(bakery, 1)); 3152 tcp_secret_primary->expires = jiffy 3153 + TCP_SECRET_2MSL 3154 + (0x1f & tcp_cookie_work(bakery, 2)); 3155 } 3156 memcpy(&tcp_secret_secondary->secrets[0], 3157 bakery, COOKIE_WORKSPACE_WORDS); 3158 3159 rcu_assign_pointer(tcp_secret_generating, 3160 tcp_secret_secondary); 3161 rcu_assign_pointer(tcp_secret_retiring, 3162 tcp_secret_primary); 3163 /* 3164 * Neither call_rcu() nor synchronize_rcu() needed. 3165 * Retiring data is not freed. It is replaced after 3166 * further (locked) pointer updates, and a quiet time 3167 * (minimum 1MSL, maximum LIFE - 2MSL). 3168 */ 3169 } 3170 spin_unlock_bh(&tcp_secret_locker); 3171 } else { 3172 rcu_read_lock_bh(); 3173 memcpy(bakery, 3174 &rcu_dereference(tcp_secret_generating)->secrets[0], 3175 COOKIE_WORKSPACE_WORDS); 3176 rcu_read_unlock_bh(); 3177 } 3178 return 0; 3179 } 3180 EXPORT_SYMBOL(tcp_cookie_generator); 3181 3182 void tcp_done(struct sock *sk) 3183 { 3184 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV) 3185 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 3186 3187 tcp_set_state(sk, TCP_CLOSE); 3188 tcp_clear_xmit_timers(sk); 3189 3190 sk->sk_shutdown = SHUTDOWN_MASK; 3191 3192 if (!sock_flag(sk, SOCK_DEAD)) 3193 sk->sk_state_change(sk); 3194 else 3195 inet_csk_destroy_sock(sk); 3196 } 3197 EXPORT_SYMBOL_GPL(tcp_done); 3198 3199 extern struct tcp_congestion_ops tcp_reno; 3200 3201 static __initdata unsigned long thash_entries; 3202 static int __init set_thash_entries(char *str) 3203 { 3204 if (!str) 3205 return 0; 3206 thash_entries = simple_strtoul(str, &str, 0); 3207 return 1; 3208 } 3209 __setup("thash_entries=", set_thash_entries); 3210 3211 void __init tcp_init(void) 3212 { 3213 struct sk_buff *skb = NULL; 3214 unsigned long nr_pages, limit; 3215 int order, i, max_share; 3216 unsigned long jiffy = jiffies; 3217 3218 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb)); 3219 3220 percpu_counter_init(&tcp_sockets_allocated, 0); 3221 percpu_counter_init(&tcp_orphan_count, 0); 3222 tcp_hashinfo.bind_bucket_cachep = 3223 kmem_cache_create("tcp_bind_bucket", 3224 sizeof(struct inet_bind_bucket), 0, 3225 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL); 3226 3227 /* Size and allocate the main established and bind bucket 3228 * hash tables. 3229 * 3230 * The methodology is similar to that of the buffer cache. 3231 */ 3232 tcp_hashinfo.ehash = 3233 alloc_large_system_hash("TCP established", 3234 sizeof(struct inet_ehash_bucket), 3235 thash_entries, 3236 (totalram_pages >= 128 * 1024) ? 3237 13 : 15, 3238 0, 3239 NULL, 3240 &tcp_hashinfo.ehash_mask, 3241 thash_entries ? 0 : 512 * 1024); 3242 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) { 3243 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i); 3244 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i); 3245 } 3246 if (inet_ehash_locks_alloc(&tcp_hashinfo)) 3247 panic("TCP: failed to alloc ehash_locks"); 3248 tcp_hashinfo.bhash = 3249 alloc_large_system_hash("TCP bind", 3250 sizeof(struct inet_bind_hashbucket), 3251 tcp_hashinfo.ehash_mask + 1, 3252 (totalram_pages >= 128 * 1024) ? 3253 13 : 15, 3254 0, 3255 &tcp_hashinfo.bhash_size, 3256 NULL, 3257 64 * 1024); 3258 tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size; 3259 for (i = 0; i < tcp_hashinfo.bhash_size; i++) { 3260 spin_lock_init(&tcp_hashinfo.bhash[i].lock); 3261 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain); 3262 } 3263 3264 /* Try to be a bit smarter and adjust defaults depending 3265 * on available memory. 3266 */ 3267 for (order = 0; ((1 << order) << PAGE_SHIFT) < 3268 (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket)); 3269 order++) 3270 ; 3271 if (order >= 4) { 3272 tcp_death_row.sysctl_max_tw_buckets = 180000; 3273 sysctl_tcp_max_orphans = 4096 << (order - 4); 3274 sysctl_max_syn_backlog = 1024; 3275 } else if (order < 3) { 3276 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order); 3277 sysctl_tcp_max_orphans >>= (3 - order); 3278 sysctl_max_syn_backlog = 128; 3279 } 3280 3281 /* Set the pressure threshold to be a fraction of global memory that 3282 * is up to 1/2 at 256 MB, decreasing toward zero with the amount of 3283 * memory, with a floor of 128 pages. 3284 */ 3285 nr_pages = totalram_pages - totalhigh_pages; 3286 limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT); 3287 limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11); 3288 limit = max(limit, 128UL); 3289 sysctl_tcp_mem[0] = limit / 4 * 3; 3290 sysctl_tcp_mem[1] = limit; 3291 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; 3292 3293 /* Set per-socket limits to no more than 1/128 the pressure threshold */ 3294 limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7); 3295 max_share = min(4UL*1024*1024, limit); 3296 3297 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM; 3298 sysctl_tcp_wmem[1] = 16*1024; 3299 sysctl_tcp_wmem[2] = max(64*1024, max_share); 3300 3301 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM; 3302 sysctl_tcp_rmem[1] = 87380; 3303 sysctl_tcp_rmem[2] = max(87380, max_share); 3304 3305 printk(KERN_INFO "TCP: Hash tables configured " 3306 "(established %u bind %u)\n", 3307 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size); 3308 3309 tcp_register_congestion_control(&tcp_reno); 3310 3311 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets)); 3312 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets)); 3313 tcp_secret_one.expires = jiffy; /* past due */ 3314 tcp_secret_two.expires = jiffy; /* past due */ 3315 tcp_secret_generating = &tcp_secret_one; 3316 tcp_secret_primary = &tcp_secret_one; 3317 tcp_secret_retiring = &tcp_secret_two; 3318 tcp_secret_secondary = &tcp_secret_two; 3319 } 3320