1 /* 2 * INET An implementation of the TCP/IP protocol suite for the LINUX 3 * operating system. INET is implemented using the BSD Socket 4 * interface as the means of communication with the user level. 5 * 6 * Implementation of the Transmission Control Protocol(TCP). 7 * 8 * Authors: Ross Biro 9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 10 * Mark Evans, <evansmp@uhura.aston.ac.uk> 11 * Corey Minyard <wf-rch!minyard@relay.EU.net> 12 * Florian La Roche, <flla@stud.uni-sb.de> 13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 14 * Linus Torvalds, <torvalds@cs.helsinki.fi> 15 * Alan Cox, <gw4pts@gw4pts.ampr.org> 16 * Matthew Dillon, <dillon@apollo.west.oic.com> 17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 18 * Jorge Cwik, <jorge@laser.satlink.net> 19 * 20 * Fixes: 21 * Alan Cox : Numerous verify_area() calls 22 * Alan Cox : Set the ACK bit on a reset 23 * Alan Cox : Stopped it crashing if it closed while 24 * sk->inuse=1 and was trying to connect 25 * (tcp_err()). 26 * Alan Cox : All icmp error handling was broken 27 * pointers passed where wrong and the 28 * socket was looked up backwards. Nobody 29 * tested any icmp error code obviously. 30 * Alan Cox : tcp_err() now handled properly. It 31 * wakes people on errors. poll 32 * behaves and the icmp error race 33 * has gone by moving it into sock.c 34 * Alan Cox : tcp_send_reset() fixed to work for 35 * everything not just packets for 36 * unknown sockets. 37 * Alan Cox : tcp option processing. 38 * Alan Cox : Reset tweaked (still not 100%) [Had 39 * syn rule wrong] 40 * Herp Rosmanith : More reset fixes 41 * Alan Cox : No longer acks invalid rst frames. 42 * Acking any kind of RST is right out. 43 * Alan Cox : Sets an ignore me flag on an rst 44 * receive otherwise odd bits of prattle 45 * escape still 46 * Alan Cox : Fixed another acking RST frame bug. 47 * Should stop LAN workplace lockups. 48 * Alan Cox : Some tidyups using the new skb list 49 * facilities 50 * Alan Cox : sk->keepopen now seems to work 51 * Alan Cox : Pulls options out correctly on accepts 52 * Alan Cox : Fixed assorted sk->rqueue->next errors 53 * Alan Cox : PSH doesn't end a TCP read. Switched a 54 * bit to skb ops. 55 * Alan Cox : Tidied tcp_data to avoid a potential 56 * nasty. 57 * Alan Cox : Added some better commenting, as the 58 * tcp is hard to follow 59 * Alan Cox : Removed incorrect check for 20 * psh 60 * Michael O'Reilly : ack < copied bug fix. 61 * Johannes Stille : Misc tcp fixes (not all in yet). 62 * Alan Cox : FIN with no memory -> CRASH 63 * Alan Cox : Added socket option proto entries. 64 * Also added awareness of them to accept. 65 * Alan Cox : Added TCP options (SOL_TCP) 66 * Alan Cox : Switched wakeup calls to callbacks, 67 * so the kernel can layer network 68 * sockets. 69 * Alan Cox : Use ip_tos/ip_ttl settings. 70 * Alan Cox : Handle FIN (more) properly (we hope). 71 * Alan Cox : RST frames sent on unsynchronised 72 * state ack error. 73 * Alan Cox : Put in missing check for SYN bit. 74 * Alan Cox : Added tcp_select_window() aka NET2E 75 * window non shrink trick. 76 * Alan Cox : Added a couple of small NET2E timer 77 * fixes 78 * Charles Hedrick : TCP fixes 79 * Toomas Tamm : TCP window fixes 80 * Alan Cox : Small URG fix to rlogin ^C ack fight 81 * Charles Hedrick : Rewrote most of it to actually work 82 * Linus : Rewrote tcp_read() and URG handling 83 * completely 84 * Gerhard Koerting: Fixed some missing timer handling 85 * Matthew Dillon : Reworked TCP machine states as per RFC 86 * Gerhard Koerting: PC/TCP workarounds 87 * Adam Caldwell : Assorted timer/timing errors 88 * Matthew Dillon : Fixed another RST bug 89 * Alan Cox : Move to kernel side addressing changes. 90 * Alan Cox : Beginning work on TCP fastpathing 91 * (not yet usable) 92 * Arnt Gulbrandsen: Turbocharged tcp_check() routine. 93 * Alan Cox : TCP fast path debugging 94 * Alan Cox : Window clamping 95 * Michael Riepe : Bug in tcp_check() 96 * Matt Dillon : More TCP improvements and RST bug fixes 97 * Matt Dillon : Yet more small nasties remove from the 98 * TCP code (Be very nice to this man if 99 * tcp finally works 100%) 8) 100 * Alan Cox : BSD accept semantics. 101 * Alan Cox : Reset on closedown bug. 102 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto(). 103 * Michael Pall : Handle poll() after URG properly in 104 * all cases. 105 * Michael Pall : Undo the last fix in tcp_read_urg() 106 * (multi URG PUSH broke rlogin). 107 * Michael Pall : Fix the multi URG PUSH problem in 108 * tcp_readable(), poll() after URG 109 * works now. 110 * Michael Pall : recv(...,MSG_OOB) never blocks in the 111 * BSD api. 112 * Alan Cox : Changed the semantics of sk->socket to 113 * fix a race and a signal problem with 114 * accept() and async I/O. 115 * Alan Cox : Relaxed the rules on tcp_sendto(). 116 * Yury Shevchuk : Really fixed accept() blocking problem. 117 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for 118 * clients/servers which listen in on 119 * fixed ports. 120 * Alan Cox : Cleaned the above up and shrank it to 121 * a sensible code size. 122 * Alan Cox : Self connect lockup fix. 123 * Alan Cox : No connect to multicast. 124 * Ross Biro : Close unaccepted children on master 125 * socket close. 126 * Alan Cox : Reset tracing code. 127 * Alan Cox : Spurious resets on shutdown. 128 * Alan Cox : Giant 15 minute/60 second timer error 129 * Alan Cox : Small whoops in polling before an 130 * accept. 131 * Alan Cox : Kept the state trace facility since 132 * it's handy for debugging. 133 * Alan Cox : More reset handler fixes. 134 * Alan Cox : Started rewriting the code based on 135 * the RFC's for other useful protocol 136 * references see: Comer, KA9Q NOS, and 137 * for a reference on the difference 138 * between specifications and how BSD 139 * works see the 4.4lite source. 140 * A.N.Kuznetsov : Don't time wait on completion of tidy 141 * close. 142 * Linus Torvalds : Fin/Shutdown & copied_seq changes. 143 * Linus Torvalds : Fixed BSD port reuse to work first syn 144 * Alan Cox : Reimplemented timers as per the RFC 145 * and using multiple timers for sanity. 146 * Alan Cox : Small bug fixes, and a lot of new 147 * comments. 148 * Alan Cox : Fixed dual reader crash by locking 149 * the buffers (much like datagram.c) 150 * Alan Cox : Fixed stuck sockets in probe. A probe 151 * now gets fed up of retrying without 152 * (even a no space) answer. 153 * Alan Cox : Extracted closing code better 154 * Alan Cox : Fixed the closing state machine to 155 * resemble the RFC. 156 * Alan Cox : More 'per spec' fixes. 157 * Jorge Cwik : Even faster checksumming. 158 * Alan Cox : tcp_data() doesn't ack illegal PSH 159 * only frames. At least one pc tcp stack 160 * generates them. 161 * Alan Cox : Cache last socket. 162 * Alan Cox : Per route irtt. 163 * Matt Day : poll()->select() match BSD precisely on error 164 * Alan Cox : New buffers 165 * Marc Tamsky : Various sk->prot->retransmits and 166 * sk->retransmits misupdating fixed. 167 * Fixed tcp_write_timeout: stuck close, 168 * and TCP syn retries gets used now. 169 * Mark Yarvis : In tcp_read_wakeup(), don't send an 170 * ack if state is TCP_CLOSED. 171 * Alan Cox : Look up device on a retransmit - routes may 172 * change. Doesn't yet cope with MSS shrink right 173 * but it's a start! 174 * Marc Tamsky : Closing in closing fixes. 175 * Mike Shaver : RFC1122 verifications. 176 * Alan Cox : rcv_saddr errors. 177 * Alan Cox : Block double connect(). 178 * Alan Cox : Small hooks for enSKIP. 179 * Alexey Kuznetsov: Path MTU discovery. 180 * Alan Cox : Support soft errors. 181 * Alan Cox : Fix MTU discovery pathological case 182 * when the remote claims no mtu! 183 * Marc Tamsky : TCP_CLOSE fix. 184 * Colin (G3TNE) : Send a reset on syn ack replies in 185 * window but wrong (fixes NT lpd problems) 186 * Pedro Roque : Better TCP window handling, delayed ack. 187 * Joerg Reuter : No modification of locked buffers in 188 * tcp_do_retransmit() 189 * Eric Schenk : Changed receiver side silly window 190 * avoidance algorithm to BSD style 191 * algorithm. This doubles throughput 192 * against machines running Solaris, 193 * and seems to result in general 194 * improvement. 195 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD 196 * Willy Konynenberg : Transparent proxying support. 197 * Mike McLagan : Routing by source 198 * Keith Owens : Do proper merging with partial SKB's in 199 * tcp_do_sendmsg to avoid burstiness. 200 * Eric Schenk : Fix fast close down bug with 201 * shutdown() followed by close(). 202 * Andi Kleen : Make poll agree with SIGIO 203 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and 204 * lingertime == 0 (RFC 793 ABORT Call) 205 * Hirokazu Takahashi : Use copy_from_user() instead of 206 * csum_and_copy_from_user() if possible. 207 * 208 * This program is free software; you can redistribute it and/or 209 * modify it under the terms of the GNU General Public License 210 * as published by the Free Software Foundation; either version 211 * 2 of the License, or(at your option) any later version. 212 * 213 * Description of States: 214 * 215 * TCP_SYN_SENT sent a connection request, waiting for ack 216 * 217 * TCP_SYN_RECV received a connection request, sent ack, 218 * waiting for final ack in three-way handshake. 219 * 220 * TCP_ESTABLISHED connection established 221 * 222 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete 223 * transmission of remaining buffered data 224 * 225 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote 226 * to shutdown 227 * 228 * TCP_CLOSING both sides have shutdown but we still have 229 * data we have to finish sending 230 * 231 * TCP_TIME_WAIT timeout to catch resent junk before entering 232 * closed, can only be entered from FIN_WAIT2 233 * or CLOSING. Required because the other end 234 * may not have gotten our last ACK causing it 235 * to retransmit the data packet (which we ignore) 236 * 237 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for 238 * us to finish writing our data and to shutdown 239 * (we have to close() to move on to LAST_ACK) 240 * 241 * TCP_LAST_ACK out side has shutdown after remote has 242 * shutdown. There may still be data in our 243 * buffer that we have to finish sending 244 * 245 * TCP_CLOSE socket is finished 246 */ 247 248 #define pr_fmt(fmt) "TCP: " fmt 249 250 #include <linux/kernel.h> 251 #include <linux/module.h> 252 #include <linux/types.h> 253 #include <linux/fcntl.h> 254 #include <linux/poll.h> 255 #include <linux/init.h> 256 #include <linux/fs.h> 257 #include <linux/skbuff.h> 258 #include <linux/scatterlist.h> 259 #include <linux/splice.h> 260 #include <linux/net.h> 261 #include <linux/socket.h> 262 #include <linux/random.h> 263 #include <linux/bootmem.h> 264 #include <linux/highmem.h> 265 #include <linux/swap.h> 266 #include <linux/cache.h> 267 #include <linux/err.h> 268 #include <linux/crypto.h> 269 #include <linux/time.h> 270 #include <linux/slab.h> 271 272 #include <net/icmp.h> 273 #include <net/inet_common.h> 274 #include <net/tcp.h> 275 #include <net/xfrm.h> 276 #include <net/ip.h> 277 #include <net/sock.h> 278 279 #include <asm/uaccess.h> 280 #include <asm/ioctls.h> 281 #include <net/busy_poll.h> 282 283 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT; 284 285 int sysctl_tcp_min_tso_segs __read_mostly = 2; 286 287 int sysctl_tcp_autocorking __read_mostly = 1; 288 289 struct percpu_counter tcp_orphan_count; 290 EXPORT_SYMBOL_GPL(tcp_orphan_count); 291 292 long sysctl_tcp_mem[3] __read_mostly; 293 int sysctl_tcp_wmem[3] __read_mostly; 294 int sysctl_tcp_rmem[3] __read_mostly; 295 296 EXPORT_SYMBOL(sysctl_tcp_mem); 297 EXPORT_SYMBOL(sysctl_tcp_rmem); 298 EXPORT_SYMBOL(sysctl_tcp_wmem); 299 300 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */ 301 EXPORT_SYMBOL(tcp_memory_allocated); 302 303 /* 304 * Current number of TCP sockets. 305 */ 306 struct percpu_counter tcp_sockets_allocated; 307 EXPORT_SYMBOL(tcp_sockets_allocated); 308 309 /* 310 * TCP splice context 311 */ 312 struct tcp_splice_state { 313 struct pipe_inode_info *pipe; 314 size_t len; 315 unsigned int flags; 316 }; 317 318 /* 319 * Pressure flag: try to collapse. 320 * Technical note: it is used by multiple contexts non atomically. 321 * All the __sk_mem_schedule() is of this nature: accounting 322 * is strict, actions are advisory and have some latency. 323 */ 324 int tcp_memory_pressure __read_mostly; 325 EXPORT_SYMBOL(tcp_memory_pressure); 326 327 void tcp_enter_memory_pressure(struct sock *sk) 328 { 329 if (!tcp_memory_pressure) { 330 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES); 331 tcp_memory_pressure = 1; 332 } 333 } 334 EXPORT_SYMBOL(tcp_enter_memory_pressure); 335 336 /* Convert seconds to retransmits based on initial and max timeout */ 337 static u8 secs_to_retrans(int seconds, int timeout, int rto_max) 338 { 339 u8 res = 0; 340 341 if (seconds > 0) { 342 int period = timeout; 343 344 res = 1; 345 while (seconds > period && res < 255) { 346 res++; 347 timeout <<= 1; 348 if (timeout > rto_max) 349 timeout = rto_max; 350 period += timeout; 351 } 352 } 353 return res; 354 } 355 356 /* Convert retransmits to seconds based on initial and max timeout */ 357 static int retrans_to_secs(u8 retrans, int timeout, int rto_max) 358 { 359 int period = 0; 360 361 if (retrans > 0) { 362 period = timeout; 363 while (--retrans) { 364 timeout <<= 1; 365 if (timeout > rto_max) 366 timeout = rto_max; 367 period += timeout; 368 } 369 } 370 return period; 371 } 372 373 /* Address-family independent initialization for a tcp_sock. 374 * 375 * NOTE: A lot of things set to zero explicitly by call to 376 * sk_alloc() so need not be done here. 377 */ 378 void tcp_init_sock(struct sock *sk) 379 { 380 struct inet_connection_sock *icsk = inet_csk(sk); 381 struct tcp_sock *tp = tcp_sk(sk); 382 383 __skb_queue_head_init(&tp->out_of_order_queue); 384 tcp_init_xmit_timers(sk); 385 tcp_prequeue_init(tp); 386 INIT_LIST_HEAD(&tp->tsq_node); 387 388 icsk->icsk_rto = TCP_TIMEOUT_INIT; 389 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT); 390 391 /* So many TCP implementations out there (incorrectly) count the 392 * initial SYN frame in their delayed-ACK and congestion control 393 * algorithms that we must have the following bandaid to talk 394 * efficiently to them. -DaveM 395 */ 396 tp->snd_cwnd = TCP_INIT_CWND; 397 398 /* See draft-stevens-tcpca-spec-01 for discussion of the 399 * initialization of these values. 400 */ 401 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 402 tp->snd_cwnd_clamp = ~0; 403 tp->mss_cache = TCP_MSS_DEFAULT; 404 405 tp->reordering = sysctl_tcp_reordering; 406 tcp_enable_early_retrans(tp); 407 tcp_assign_congestion_control(sk); 408 409 tp->tsoffset = 0; 410 411 sk->sk_state = TCP_CLOSE; 412 413 sk->sk_write_space = sk_stream_write_space; 414 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE); 415 416 icsk->icsk_sync_mss = tcp_sync_mss; 417 418 sk->sk_sndbuf = sysctl_tcp_wmem[1]; 419 sk->sk_rcvbuf = sysctl_tcp_rmem[1]; 420 421 local_bh_disable(); 422 sock_update_memcg(sk); 423 sk_sockets_allocated_inc(sk); 424 local_bh_enable(); 425 } 426 EXPORT_SYMBOL(tcp_init_sock); 427 428 static void tcp_tx_timestamp(struct sock *sk, struct sk_buff *skb) 429 { 430 if (sk->sk_tsflags) { 431 struct skb_shared_info *shinfo = skb_shinfo(skb); 432 433 sock_tx_timestamp(sk, &shinfo->tx_flags); 434 if (shinfo->tx_flags & SKBTX_ANY_TSTAMP) 435 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1; 436 } 437 } 438 439 /* 440 * Wait for a TCP event. 441 * 442 * Note that we don't need to lock the socket, as the upper poll layers 443 * take care of normal races (between the test and the event) and we don't 444 * go look at any of the socket buffers directly. 445 */ 446 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait) 447 { 448 unsigned int mask; 449 struct sock *sk = sock->sk; 450 const struct tcp_sock *tp = tcp_sk(sk); 451 452 sock_rps_record_flow(sk); 453 454 sock_poll_wait(file, sk_sleep(sk), wait); 455 if (sk->sk_state == TCP_LISTEN) 456 return inet_csk_listen_poll(sk); 457 458 /* Socket is not locked. We are protected from async events 459 * by poll logic and correct handling of state changes 460 * made by other threads is impossible in any case. 461 */ 462 463 mask = 0; 464 465 /* 466 * POLLHUP is certainly not done right. But poll() doesn't 467 * have a notion of HUP in just one direction, and for a 468 * socket the read side is more interesting. 469 * 470 * Some poll() documentation says that POLLHUP is incompatible 471 * with the POLLOUT/POLLWR flags, so somebody should check this 472 * all. But careful, it tends to be safer to return too many 473 * bits than too few, and you can easily break real applications 474 * if you don't tell them that something has hung up! 475 * 476 * Check-me. 477 * 478 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and 479 * our fs/select.c). It means that after we received EOF, 480 * poll always returns immediately, making impossible poll() on write() 481 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP 482 * if and only if shutdown has been made in both directions. 483 * Actually, it is interesting to look how Solaris and DUX 484 * solve this dilemma. I would prefer, if POLLHUP were maskable, 485 * then we could set it on SND_SHUTDOWN. BTW examples given 486 * in Stevens' books assume exactly this behaviour, it explains 487 * why POLLHUP is incompatible with POLLOUT. --ANK 488 * 489 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent 490 * blocking on fresh not-connected or disconnected socket. --ANK 491 */ 492 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE) 493 mask |= POLLHUP; 494 if (sk->sk_shutdown & RCV_SHUTDOWN) 495 mask |= POLLIN | POLLRDNORM | POLLRDHUP; 496 497 /* Connected or passive Fast Open socket? */ 498 if (sk->sk_state != TCP_SYN_SENT && 499 (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk != NULL)) { 500 int target = sock_rcvlowat(sk, 0, INT_MAX); 501 502 if (tp->urg_seq == tp->copied_seq && 503 !sock_flag(sk, SOCK_URGINLINE) && 504 tp->urg_data) 505 target++; 506 507 /* Potential race condition. If read of tp below will 508 * escape above sk->sk_state, we can be illegally awaken 509 * in SYN_* states. */ 510 if (tp->rcv_nxt - tp->copied_seq >= target) 511 mask |= POLLIN | POLLRDNORM; 512 513 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) { 514 if (sk_stream_is_writeable(sk)) { 515 mask |= POLLOUT | POLLWRNORM; 516 } else { /* send SIGIO later */ 517 set_bit(SOCK_ASYNC_NOSPACE, 518 &sk->sk_socket->flags); 519 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 520 521 /* Race breaker. If space is freed after 522 * wspace test but before the flags are set, 523 * IO signal will be lost. 524 */ 525 if (sk_stream_is_writeable(sk)) 526 mask |= POLLOUT | POLLWRNORM; 527 } 528 } else 529 mask |= POLLOUT | POLLWRNORM; 530 531 if (tp->urg_data & TCP_URG_VALID) 532 mask |= POLLPRI; 533 } 534 /* This barrier is coupled with smp_wmb() in tcp_reset() */ 535 smp_rmb(); 536 if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue)) 537 mask |= POLLERR; 538 539 return mask; 540 } 541 EXPORT_SYMBOL(tcp_poll); 542 543 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg) 544 { 545 struct tcp_sock *tp = tcp_sk(sk); 546 int answ; 547 bool slow; 548 549 switch (cmd) { 550 case SIOCINQ: 551 if (sk->sk_state == TCP_LISTEN) 552 return -EINVAL; 553 554 slow = lock_sock_fast(sk); 555 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 556 answ = 0; 557 else if (sock_flag(sk, SOCK_URGINLINE) || 558 !tp->urg_data || 559 before(tp->urg_seq, tp->copied_seq) || 560 !before(tp->urg_seq, tp->rcv_nxt)) { 561 562 answ = tp->rcv_nxt - tp->copied_seq; 563 564 /* Subtract 1, if FIN was received */ 565 if (answ && sock_flag(sk, SOCK_DONE)) 566 answ--; 567 } else 568 answ = tp->urg_seq - tp->copied_seq; 569 unlock_sock_fast(sk, slow); 570 break; 571 case SIOCATMARK: 572 answ = tp->urg_data && tp->urg_seq == tp->copied_seq; 573 break; 574 case SIOCOUTQ: 575 if (sk->sk_state == TCP_LISTEN) 576 return -EINVAL; 577 578 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 579 answ = 0; 580 else 581 answ = tp->write_seq - tp->snd_una; 582 break; 583 case SIOCOUTQNSD: 584 if (sk->sk_state == TCP_LISTEN) 585 return -EINVAL; 586 587 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 588 answ = 0; 589 else 590 answ = tp->write_seq - tp->snd_nxt; 591 break; 592 default: 593 return -ENOIOCTLCMD; 594 } 595 596 return put_user(answ, (int __user *)arg); 597 } 598 EXPORT_SYMBOL(tcp_ioctl); 599 600 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb) 601 { 602 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 603 tp->pushed_seq = tp->write_seq; 604 } 605 606 static inline bool forced_push(const struct tcp_sock *tp) 607 { 608 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1)); 609 } 610 611 static void skb_entail(struct sock *sk, struct sk_buff *skb) 612 { 613 struct tcp_sock *tp = tcp_sk(sk); 614 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 615 616 skb->csum = 0; 617 tcb->seq = tcb->end_seq = tp->write_seq; 618 tcb->tcp_flags = TCPHDR_ACK; 619 tcb->sacked = 0; 620 __skb_header_release(skb); 621 tcp_add_write_queue_tail(sk, skb); 622 sk->sk_wmem_queued += skb->truesize; 623 sk_mem_charge(sk, skb->truesize); 624 if (tp->nonagle & TCP_NAGLE_PUSH) 625 tp->nonagle &= ~TCP_NAGLE_PUSH; 626 } 627 628 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags) 629 { 630 if (flags & MSG_OOB) 631 tp->snd_up = tp->write_seq; 632 } 633 634 /* If a not yet filled skb is pushed, do not send it if 635 * we have data packets in Qdisc or NIC queues : 636 * Because TX completion will happen shortly, it gives a chance 637 * to coalesce future sendmsg() payload into this skb, without 638 * need for a timer, and with no latency trade off. 639 * As packets containing data payload have a bigger truesize 640 * than pure acks (dataless) packets, the last checks prevent 641 * autocorking if we only have an ACK in Qdisc/NIC queues, 642 * or if TX completion was delayed after we processed ACK packet. 643 */ 644 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb, 645 int size_goal) 646 { 647 return skb->len < size_goal && 648 sysctl_tcp_autocorking && 649 skb != tcp_write_queue_head(sk) && 650 atomic_read(&sk->sk_wmem_alloc) > skb->truesize; 651 } 652 653 static void tcp_push(struct sock *sk, int flags, int mss_now, 654 int nonagle, int size_goal) 655 { 656 struct tcp_sock *tp = tcp_sk(sk); 657 struct sk_buff *skb; 658 659 if (!tcp_send_head(sk)) 660 return; 661 662 skb = tcp_write_queue_tail(sk); 663 if (!(flags & MSG_MORE) || forced_push(tp)) 664 tcp_mark_push(tp, skb); 665 666 tcp_mark_urg(tp, flags); 667 668 if (tcp_should_autocork(sk, skb, size_goal)) { 669 670 /* avoid atomic op if TSQ_THROTTLED bit is already set */ 671 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) { 672 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING); 673 set_bit(TSQ_THROTTLED, &tp->tsq_flags); 674 } 675 /* It is possible TX completion already happened 676 * before we set TSQ_THROTTLED. 677 */ 678 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize) 679 return; 680 } 681 682 if (flags & MSG_MORE) 683 nonagle = TCP_NAGLE_CORK; 684 685 __tcp_push_pending_frames(sk, mss_now, nonagle); 686 } 687 688 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, 689 unsigned int offset, size_t len) 690 { 691 struct tcp_splice_state *tss = rd_desc->arg.data; 692 int ret; 693 694 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len), 695 tss->flags); 696 if (ret > 0) 697 rd_desc->count -= ret; 698 return ret; 699 } 700 701 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss) 702 { 703 /* Store TCP splice context information in read_descriptor_t. */ 704 read_descriptor_t rd_desc = { 705 .arg.data = tss, 706 .count = tss->len, 707 }; 708 709 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv); 710 } 711 712 /** 713 * tcp_splice_read - splice data from TCP socket to a pipe 714 * @sock: socket to splice from 715 * @ppos: position (not valid) 716 * @pipe: pipe to splice to 717 * @len: number of bytes to splice 718 * @flags: splice modifier flags 719 * 720 * Description: 721 * Will read pages from given socket and fill them into a pipe. 722 * 723 **/ 724 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos, 725 struct pipe_inode_info *pipe, size_t len, 726 unsigned int flags) 727 { 728 struct sock *sk = sock->sk; 729 struct tcp_splice_state tss = { 730 .pipe = pipe, 731 .len = len, 732 .flags = flags, 733 }; 734 long timeo; 735 ssize_t spliced; 736 int ret; 737 738 sock_rps_record_flow(sk); 739 /* 740 * We can't seek on a socket input 741 */ 742 if (unlikely(*ppos)) 743 return -ESPIPE; 744 745 ret = spliced = 0; 746 747 lock_sock(sk); 748 749 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK); 750 while (tss.len) { 751 ret = __tcp_splice_read(sk, &tss); 752 if (ret < 0) 753 break; 754 else if (!ret) { 755 if (spliced) 756 break; 757 if (sock_flag(sk, SOCK_DONE)) 758 break; 759 if (sk->sk_err) { 760 ret = sock_error(sk); 761 break; 762 } 763 if (sk->sk_shutdown & RCV_SHUTDOWN) 764 break; 765 if (sk->sk_state == TCP_CLOSE) { 766 /* 767 * This occurs when user tries to read 768 * from never connected socket. 769 */ 770 if (!sock_flag(sk, SOCK_DONE)) 771 ret = -ENOTCONN; 772 break; 773 } 774 if (!timeo) { 775 ret = -EAGAIN; 776 break; 777 } 778 sk_wait_data(sk, &timeo); 779 if (signal_pending(current)) { 780 ret = sock_intr_errno(timeo); 781 break; 782 } 783 continue; 784 } 785 tss.len -= ret; 786 spliced += ret; 787 788 if (!timeo) 789 break; 790 release_sock(sk); 791 lock_sock(sk); 792 793 if (sk->sk_err || sk->sk_state == TCP_CLOSE || 794 (sk->sk_shutdown & RCV_SHUTDOWN) || 795 signal_pending(current)) 796 break; 797 } 798 799 release_sock(sk); 800 801 if (spliced) 802 return spliced; 803 804 return ret; 805 } 806 EXPORT_SYMBOL(tcp_splice_read); 807 808 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp) 809 { 810 struct sk_buff *skb; 811 812 /* The TCP header must be at least 32-bit aligned. */ 813 size = ALIGN(size, 4); 814 815 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp); 816 if (skb) { 817 if (sk_wmem_schedule(sk, skb->truesize)) { 818 skb_reserve(skb, sk->sk_prot->max_header); 819 /* 820 * Make sure that we have exactly size bytes 821 * available to the caller, no more, no less. 822 */ 823 skb->reserved_tailroom = skb->end - skb->tail - size; 824 return skb; 825 } 826 __kfree_skb(skb); 827 } else { 828 sk->sk_prot->enter_memory_pressure(sk); 829 sk_stream_moderate_sndbuf(sk); 830 } 831 return NULL; 832 } 833 834 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now, 835 int large_allowed) 836 { 837 struct tcp_sock *tp = tcp_sk(sk); 838 u32 new_size_goal, size_goal, hlen; 839 840 if (!large_allowed || !sk_can_gso(sk)) 841 return mss_now; 842 843 /* Maybe we should/could use sk->sk_prot->max_header here ? */ 844 hlen = inet_csk(sk)->icsk_af_ops->net_header_len + 845 inet_csk(sk)->icsk_ext_hdr_len + 846 tp->tcp_header_len; 847 848 new_size_goal = sk->sk_gso_max_size - 1 - hlen; 849 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal); 850 851 /* We try hard to avoid divides here */ 852 size_goal = tp->gso_segs * mss_now; 853 if (unlikely(new_size_goal < size_goal || 854 new_size_goal >= size_goal + mss_now)) { 855 tp->gso_segs = min_t(u16, new_size_goal / mss_now, 856 sk->sk_gso_max_segs); 857 size_goal = tp->gso_segs * mss_now; 858 } 859 860 return max(size_goal, mss_now); 861 } 862 863 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags) 864 { 865 int mss_now; 866 867 mss_now = tcp_current_mss(sk); 868 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB)); 869 870 return mss_now; 871 } 872 873 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset, 874 size_t size, int flags) 875 { 876 struct tcp_sock *tp = tcp_sk(sk); 877 int mss_now, size_goal; 878 int err; 879 ssize_t copied; 880 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 881 882 /* Wait for a connection to finish. One exception is TCP Fast Open 883 * (passive side) where data is allowed to be sent before a connection 884 * is fully established. 885 */ 886 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) && 887 !tcp_passive_fastopen(sk)) { 888 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0) 889 goto out_err; 890 } 891 892 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags); 893 894 mss_now = tcp_send_mss(sk, &size_goal, flags); 895 copied = 0; 896 897 err = -EPIPE; 898 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 899 goto out_err; 900 901 while (size > 0) { 902 struct sk_buff *skb = tcp_write_queue_tail(sk); 903 int copy, i; 904 bool can_coalesce; 905 906 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) { 907 new_segment: 908 if (!sk_stream_memory_free(sk)) 909 goto wait_for_sndbuf; 910 911 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation); 912 if (!skb) 913 goto wait_for_memory; 914 915 skb_entail(sk, skb); 916 copy = size_goal; 917 } 918 919 if (copy > size) 920 copy = size; 921 922 i = skb_shinfo(skb)->nr_frags; 923 can_coalesce = skb_can_coalesce(skb, i, page, offset); 924 if (!can_coalesce && i >= MAX_SKB_FRAGS) { 925 tcp_mark_push(tp, skb); 926 goto new_segment; 927 } 928 if (!sk_wmem_schedule(sk, copy)) 929 goto wait_for_memory; 930 931 if (can_coalesce) { 932 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 933 } else { 934 get_page(page); 935 skb_fill_page_desc(skb, i, page, offset, copy); 936 } 937 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG; 938 939 skb->len += copy; 940 skb->data_len += copy; 941 skb->truesize += copy; 942 sk->sk_wmem_queued += copy; 943 sk_mem_charge(sk, copy); 944 skb->ip_summed = CHECKSUM_PARTIAL; 945 tp->write_seq += copy; 946 TCP_SKB_CB(skb)->end_seq += copy; 947 tcp_skb_pcount_set(skb, 0); 948 949 if (!copied) 950 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 951 952 copied += copy; 953 offset += copy; 954 if (!(size -= copy)) { 955 tcp_tx_timestamp(sk, skb); 956 goto out; 957 } 958 959 if (skb->len < size_goal || (flags & MSG_OOB)) 960 continue; 961 962 if (forced_push(tp)) { 963 tcp_mark_push(tp, skb); 964 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 965 } else if (skb == tcp_send_head(sk)) 966 tcp_push_one(sk, mss_now); 967 continue; 968 969 wait_for_sndbuf: 970 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 971 wait_for_memory: 972 tcp_push(sk, flags & ~MSG_MORE, mss_now, 973 TCP_NAGLE_PUSH, size_goal); 974 975 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0) 976 goto do_error; 977 978 mss_now = tcp_send_mss(sk, &size_goal, flags); 979 } 980 981 out: 982 if (copied && !(flags & MSG_SENDPAGE_NOTLAST)) 983 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal); 984 return copied; 985 986 do_error: 987 if (copied) 988 goto out; 989 out_err: 990 return sk_stream_error(sk, flags, err); 991 } 992 993 int tcp_sendpage(struct sock *sk, struct page *page, int offset, 994 size_t size, int flags) 995 { 996 ssize_t res; 997 998 if (!(sk->sk_route_caps & NETIF_F_SG) || 999 !(sk->sk_route_caps & NETIF_F_ALL_CSUM)) 1000 return sock_no_sendpage(sk->sk_socket, page, offset, size, 1001 flags); 1002 1003 lock_sock(sk); 1004 res = do_tcp_sendpages(sk, page, offset, size, flags); 1005 release_sock(sk); 1006 return res; 1007 } 1008 EXPORT_SYMBOL(tcp_sendpage); 1009 1010 static inline int select_size(const struct sock *sk, bool sg) 1011 { 1012 const struct tcp_sock *tp = tcp_sk(sk); 1013 int tmp = tp->mss_cache; 1014 1015 if (sg) { 1016 if (sk_can_gso(sk)) { 1017 /* Small frames wont use a full page: 1018 * Payload will immediately follow tcp header. 1019 */ 1020 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER); 1021 } else { 1022 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER); 1023 1024 if (tmp >= pgbreak && 1025 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE) 1026 tmp = pgbreak; 1027 } 1028 } 1029 1030 return tmp; 1031 } 1032 1033 void tcp_free_fastopen_req(struct tcp_sock *tp) 1034 { 1035 if (tp->fastopen_req != NULL) { 1036 kfree(tp->fastopen_req); 1037 tp->fastopen_req = NULL; 1038 } 1039 } 1040 1041 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, 1042 int *copied, size_t size) 1043 { 1044 struct tcp_sock *tp = tcp_sk(sk); 1045 int err, flags; 1046 1047 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE)) 1048 return -EOPNOTSUPP; 1049 if (tp->fastopen_req != NULL) 1050 return -EALREADY; /* Another Fast Open is in progress */ 1051 1052 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request), 1053 sk->sk_allocation); 1054 if (unlikely(tp->fastopen_req == NULL)) 1055 return -ENOBUFS; 1056 tp->fastopen_req->data = msg; 1057 tp->fastopen_req->size = size; 1058 1059 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0; 1060 err = __inet_stream_connect(sk->sk_socket, msg->msg_name, 1061 msg->msg_namelen, flags); 1062 *copied = tp->fastopen_req->copied; 1063 tcp_free_fastopen_req(tp); 1064 return err; 1065 } 1066 1067 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg, 1068 size_t size) 1069 { 1070 struct tcp_sock *tp = tcp_sk(sk); 1071 struct sk_buff *skb; 1072 int flags, err, copied = 0; 1073 int mss_now = 0, size_goal, copied_syn = 0; 1074 bool sg; 1075 long timeo; 1076 1077 lock_sock(sk); 1078 1079 flags = msg->msg_flags; 1080 if (flags & MSG_FASTOPEN) { 1081 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size); 1082 if (err == -EINPROGRESS && copied_syn > 0) 1083 goto out; 1084 else if (err) 1085 goto out_err; 1086 } 1087 1088 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 1089 1090 /* Wait for a connection to finish. One exception is TCP Fast Open 1091 * (passive side) where data is allowed to be sent before a connection 1092 * is fully established. 1093 */ 1094 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) && 1095 !tcp_passive_fastopen(sk)) { 1096 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0) 1097 goto do_error; 1098 } 1099 1100 if (unlikely(tp->repair)) { 1101 if (tp->repair_queue == TCP_RECV_QUEUE) { 1102 copied = tcp_send_rcvq(sk, msg, size); 1103 goto out_nopush; 1104 } 1105 1106 err = -EINVAL; 1107 if (tp->repair_queue == TCP_NO_QUEUE) 1108 goto out_err; 1109 1110 /* 'common' sending to sendq */ 1111 } 1112 1113 /* This should be in poll */ 1114 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags); 1115 1116 mss_now = tcp_send_mss(sk, &size_goal, flags); 1117 1118 /* Ok commence sending. */ 1119 copied = 0; 1120 1121 err = -EPIPE; 1122 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 1123 goto out_err; 1124 1125 sg = !!(sk->sk_route_caps & NETIF_F_SG); 1126 1127 while (iov_iter_count(&msg->msg_iter)) { 1128 int copy = 0; 1129 int max = size_goal; 1130 1131 skb = tcp_write_queue_tail(sk); 1132 if (tcp_send_head(sk)) { 1133 if (skb->ip_summed == CHECKSUM_NONE) 1134 max = mss_now; 1135 copy = max - skb->len; 1136 } 1137 1138 if (copy <= 0) { 1139 new_segment: 1140 /* Allocate new segment. If the interface is SG, 1141 * allocate skb fitting to single page. 1142 */ 1143 if (!sk_stream_memory_free(sk)) 1144 goto wait_for_sndbuf; 1145 1146 skb = sk_stream_alloc_skb(sk, 1147 select_size(sk, sg), 1148 sk->sk_allocation); 1149 if (!skb) 1150 goto wait_for_memory; 1151 1152 /* 1153 * Check whether we can use HW checksum. 1154 */ 1155 if (sk->sk_route_caps & NETIF_F_ALL_CSUM) 1156 skb->ip_summed = CHECKSUM_PARTIAL; 1157 1158 skb_entail(sk, skb); 1159 copy = size_goal; 1160 max = size_goal; 1161 1162 /* All packets are restored as if they have 1163 * already been sent. skb_mstamp isn't set to 1164 * avoid wrong rtt estimation. 1165 */ 1166 if (tp->repair) 1167 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED; 1168 } 1169 1170 /* Try to append data to the end of skb. */ 1171 if (copy > iov_iter_count(&msg->msg_iter)) 1172 copy = iov_iter_count(&msg->msg_iter); 1173 1174 /* Where to copy to? */ 1175 if (skb_availroom(skb) > 0) { 1176 /* We have some space in skb head. Superb! */ 1177 copy = min_t(int, copy, skb_availroom(skb)); 1178 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy); 1179 if (err) 1180 goto do_fault; 1181 } else { 1182 bool merge = true; 1183 int i = skb_shinfo(skb)->nr_frags; 1184 struct page_frag *pfrag = sk_page_frag(sk); 1185 1186 if (!sk_page_frag_refill(sk, pfrag)) 1187 goto wait_for_memory; 1188 1189 if (!skb_can_coalesce(skb, i, pfrag->page, 1190 pfrag->offset)) { 1191 if (i == MAX_SKB_FRAGS || !sg) { 1192 tcp_mark_push(tp, skb); 1193 goto new_segment; 1194 } 1195 merge = false; 1196 } 1197 1198 copy = min_t(int, copy, pfrag->size - pfrag->offset); 1199 1200 if (!sk_wmem_schedule(sk, copy)) 1201 goto wait_for_memory; 1202 1203 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb, 1204 pfrag->page, 1205 pfrag->offset, 1206 copy); 1207 if (err) 1208 goto do_error; 1209 1210 /* Update the skb. */ 1211 if (merge) { 1212 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 1213 } else { 1214 skb_fill_page_desc(skb, i, pfrag->page, 1215 pfrag->offset, copy); 1216 get_page(pfrag->page); 1217 } 1218 pfrag->offset += copy; 1219 } 1220 1221 if (!copied) 1222 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 1223 1224 tp->write_seq += copy; 1225 TCP_SKB_CB(skb)->end_seq += copy; 1226 tcp_skb_pcount_set(skb, 0); 1227 1228 copied += copy; 1229 if (!iov_iter_count(&msg->msg_iter)) { 1230 tcp_tx_timestamp(sk, skb); 1231 goto out; 1232 } 1233 1234 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair)) 1235 continue; 1236 1237 if (forced_push(tp)) { 1238 tcp_mark_push(tp, skb); 1239 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 1240 } else if (skb == tcp_send_head(sk)) 1241 tcp_push_one(sk, mss_now); 1242 continue; 1243 1244 wait_for_sndbuf: 1245 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1246 wait_for_memory: 1247 if (copied) 1248 tcp_push(sk, flags & ~MSG_MORE, mss_now, 1249 TCP_NAGLE_PUSH, size_goal); 1250 1251 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0) 1252 goto do_error; 1253 1254 mss_now = tcp_send_mss(sk, &size_goal, flags); 1255 } 1256 1257 out: 1258 if (copied) 1259 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal); 1260 out_nopush: 1261 release_sock(sk); 1262 return copied + copied_syn; 1263 1264 do_fault: 1265 if (!skb->len) { 1266 tcp_unlink_write_queue(skb, sk); 1267 /* It is the one place in all of TCP, except connection 1268 * reset, where we can be unlinking the send_head. 1269 */ 1270 tcp_check_send_head(sk, skb); 1271 sk_wmem_free_skb(sk, skb); 1272 } 1273 1274 do_error: 1275 if (copied + copied_syn) 1276 goto out; 1277 out_err: 1278 err = sk_stream_error(sk, flags, err); 1279 release_sock(sk); 1280 return err; 1281 } 1282 EXPORT_SYMBOL(tcp_sendmsg); 1283 1284 /* 1285 * Handle reading urgent data. BSD has very simple semantics for 1286 * this, no blocking and very strange errors 8) 1287 */ 1288 1289 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags) 1290 { 1291 struct tcp_sock *tp = tcp_sk(sk); 1292 1293 /* No URG data to read. */ 1294 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data || 1295 tp->urg_data == TCP_URG_READ) 1296 return -EINVAL; /* Yes this is right ! */ 1297 1298 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE)) 1299 return -ENOTCONN; 1300 1301 if (tp->urg_data & TCP_URG_VALID) { 1302 int err = 0; 1303 char c = tp->urg_data; 1304 1305 if (!(flags & MSG_PEEK)) 1306 tp->urg_data = TCP_URG_READ; 1307 1308 /* Read urgent data. */ 1309 msg->msg_flags |= MSG_OOB; 1310 1311 if (len > 0) { 1312 if (!(flags & MSG_TRUNC)) 1313 err = memcpy_to_msg(msg, &c, 1); 1314 len = 1; 1315 } else 1316 msg->msg_flags |= MSG_TRUNC; 1317 1318 return err ? -EFAULT : len; 1319 } 1320 1321 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 1322 return 0; 1323 1324 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and 1325 * the available implementations agree in this case: 1326 * this call should never block, independent of the 1327 * blocking state of the socket. 1328 * Mike <pall@rz.uni-karlsruhe.de> 1329 */ 1330 return -EAGAIN; 1331 } 1332 1333 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len) 1334 { 1335 struct sk_buff *skb; 1336 int copied = 0, err = 0; 1337 1338 /* XXX -- need to support SO_PEEK_OFF */ 1339 1340 skb_queue_walk(&sk->sk_write_queue, skb) { 1341 err = skb_copy_datagram_msg(skb, 0, msg, skb->len); 1342 if (err) 1343 break; 1344 1345 copied += skb->len; 1346 } 1347 1348 return err ?: copied; 1349 } 1350 1351 /* Clean up the receive buffer for full frames taken by the user, 1352 * then send an ACK if necessary. COPIED is the number of bytes 1353 * tcp_recvmsg has given to the user so far, it speeds up the 1354 * calculation of whether or not we must ACK for the sake of 1355 * a window update. 1356 */ 1357 static void tcp_cleanup_rbuf(struct sock *sk, int copied) 1358 { 1359 struct tcp_sock *tp = tcp_sk(sk); 1360 bool time_to_ack = false; 1361 1362 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 1363 1364 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq), 1365 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n", 1366 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt); 1367 1368 if (inet_csk_ack_scheduled(sk)) { 1369 const struct inet_connection_sock *icsk = inet_csk(sk); 1370 /* Delayed ACKs frequently hit locked sockets during bulk 1371 * receive. */ 1372 if (icsk->icsk_ack.blocked || 1373 /* Once-per-two-segments ACK was not sent by tcp_input.c */ 1374 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss || 1375 /* 1376 * If this read emptied read buffer, we send ACK, if 1377 * connection is not bidirectional, user drained 1378 * receive buffer and there was a small segment 1379 * in queue. 1380 */ 1381 (copied > 0 && 1382 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) || 1383 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) && 1384 !icsk->icsk_ack.pingpong)) && 1385 !atomic_read(&sk->sk_rmem_alloc))) 1386 time_to_ack = true; 1387 } 1388 1389 /* We send an ACK if we can now advertise a non-zero window 1390 * which has been raised "significantly". 1391 * 1392 * Even if window raised up to infinity, do not send window open ACK 1393 * in states, where we will not receive more. It is useless. 1394 */ 1395 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) { 1396 __u32 rcv_window_now = tcp_receive_window(tp); 1397 1398 /* Optimize, __tcp_select_window() is not cheap. */ 1399 if (2*rcv_window_now <= tp->window_clamp) { 1400 __u32 new_window = __tcp_select_window(sk); 1401 1402 /* Send ACK now, if this read freed lots of space 1403 * in our buffer. Certainly, new_window is new window. 1404 * We can advertise it now, if it is not less than current one. 1405 * "Lots" means "at least twice" here. 1406 */ 1407 if (new_window && new_window >= 2 * rcv_window_now) 1408 time_to_ack = true; 1409 } 1410 } 1411 if (time_to_ack) 1412 tcp_send_ack(sk); 1413 } 1414 1415 static void tcp_prequeue_process(struct sock *sk) 1416 { 1417 struct sk_buff *skb; 1418 struct tcp_sock *tp = tcp_sk(sk); 1419 1420 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED); 1421 1422 /* RX process wants to run with disabled BHs, though it is not 1423 * necessary */ 1424 local_bh_disable(); 1425 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL) 1426 sk_backlog_rcv(sk, skb); 1427 local_bh_enable(); 1428 1429 /* Clear memory counter. */ 1430 tp->ucopy.memory = 0; 1431 } 1432 1433 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off) 1434 { 1435 struct sk_buff *skb; 1436 u32 offset; 1437 1438 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) { 1439 offset = seq - TCP_SKB_CB(skb)->seq; 1440 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 1441 offset--; 1442 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) { 1443 *off = offset; 1444 return skb; 1445 } 1446 /* This looks weird, but this can happen if TCP collapsing 1447 * splitted a fat GRO packet, while we released socket lock 1448 * in skb_splice_bits() 1449 */ 1450 sk_eat_skb(sk, skb); 1451 } 1452 return NULL; 1453 } 1454 1455 /* 1456 * This routine provides an alternative to tcp_recvmsg() for routines 1457 * that would like to handle copying from skbuffs directly in 'sendfile' 1458 * fashion. 1459 * Note: 1460 * - It is assumed that the socket was locked by the caller. 1461 * - The routine does not block. 1462 * - At present, there is no support for reading OOB data 1463 * or for 'peeking' the socket using this routine 1464 * (although both would be easy to implement). 1465 */ 1466 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, 1467 sk_read_actor_t recv_actor) 1468 { 1469 struct sk_buff *skb; 1470 struct tcp_sock *tp = tcp_sk(sk); 1471 u32 seq = tp->copied_seq; 1472 u32 offset; 1473 int copied = 0; 1474 1475 if (sk->sk_state == TCP_LISTEN) 1476 return -ENOTCONN; 1477 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) { 1478 if (offset < skb->len) { 1479 int used; 1480 size_t len; 1481 1482 len = skb->len - offset; 1483 /* Stop reading if we hit a patch of urgent data */ 1484 if (tp->urg_data) { 1485 u32 urg_offset = tp->urg_seq - seq; 1486 if (urg_offset < len) 1487 len = urg_offset; 1488 if (!len) 1489 break; 1490 } 1491 used = recv_actor(desc, skb, offset, len); 1492 if (used <= 0) { 1493 if (!copied) 1494 copied = used; 1495 break; 1496 } else if (used <= len) { 1497 seq += used; 1498 copied += used; 1499 offset += used; 1500 } 1501 /* If recv_actor drops the lock (e.g. TCP splice 1502 * receive) the skb pointer might be invalid when 1503 * getting here: tcp_collapse might have deleted it 1504 * while aggregating skbs from the socket queue. 1505 */ 1506 skb = tcp_recv_skb(sk, seq - 1, &offset); 1507 if (!skb) 1508 break; 1509 /* TCP coalescing might have appended data to the skb. 1510 * Try to splice more frags 1511 */ 1512 if (offset + 1 != skb->len) 1513 continue; 1514 } 1515 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) { 1516 sk_eat_skb(sk, skb); 1517 ++seq; 1518 break; 1519 } 1520 sk_eat_skb(sk, skb); 1521 if (!desc->count) 1522 break; 1523 tp->copied_seq = seq; 1524 } 1525 tp->copied_seq = seq; 1526 1527 tcp_rcv_space_adjust(sk); 1528 1529 /* Clean up data we have read: This will do ACK frames. */ 1530 if (copied > 0) { 1531 tcp_recv_skb(sk, seq, &offset); 1532 tcp_cleanup_rbuf(sk, copied); 1533 } 1534 return copied; 1535 } 1536 EXPORT_SYMBOL(tcp_read_sock); 1537 1538 /* 1539 * This routine copies from a sock struct into the user buffer. 1540 * 1541 * Technical note: in 2.3 we work on _locked_ socket, so that 1542 * tricks with *seq access order and skb->users are not required. 1543 * Probably, code can be easily improved even more. 1544 */ 1545 1546 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg, 1547 size_t len, int nonblock, int flags, int *addr_len) 1548 { 1549 struct tcp_sock *tp = tcp_sk(sk); 1550 int copied = 0; 1551 u32 peek_seq; 1552 u32 *seq; 1553 unsigned long used; 1554 int err; 1555 int target; /* Read at least this many bytes */ 1556 long timeo; 1557 struct task_struct *user_recv = NULL; 1558 struct sk_buff *skb; 1559 u32 urg_hole = 0; 1560 1561 if (unlikely(flags & MSG_ERRQUEUE)) 1562 return inet_recv_error(sk, msg, len, addr_len); 1563 1564 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) && 1565 (sk->sk_state == TCP_ESTABLISHED)) 1566 sk_busy_loop(sk, nonblock); 1567 1568 lock_sock(sk); 1569 1570 err = -ENOTCONN; 1571 if (sk->sk_state == TCP_LISTEN) 1572 goto out; 1573 1574 timeo = sock_rcvtimeo(sk, nonblock); 1575 1576 /* Urgent data needs to be handled specially. */ 1577 if (flags & MSG_OOB) 1578 goto recv_urg; 1579 1580 if (unlikely(tp->repair)) { 1581 err = -EPERM; 1582 if (!(flags & MSG_PEEK)) 1583 goto out; 1584 1585 if (tp->repair_queue == TCP_SEND_QUEUE) 1586 goto recv_sndq; 1587 1588 err = -EINVAL; 1589 if (tp->repair_queue == TCP_NO_QUEUE) 1590 goto out; 1591 1592 /* 'common' recv queue MSG_PEEK-ing */ 1593 } 1594 1595 seq = &tp->copied_seq; 1596 if (flags & MSG_PEEK) { 1597 peek_seq = tp->copied_seq; 1598 seq = &peek_seq; 1599 } 1600 1601 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 1602 1603 do { 1604 u32 offset; 1605 1606 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */ 1607 if (tp->urg_data && tp->urg_seq == *seq) { 1608 if (copied) 1609 break; 1610 if (signal_pending(current)) { 1611 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN; 1612 break; 1613 } 1614 } 1615 1616 /* Next get a buffer. */ 1617 1618 skb_queue_walk(&sk->sk_receive_queue, skb) { 1619 /* Now that we have two receive queues this 1620 * shouldn't happen. 1621 */ 1622 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq), 1623 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n", 1624 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, 1625 flags)) 1626 break; 1627 1628 offset = *seq - TCP_SKB_CB(skb)->seq; 1629 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 1630 offset--; 1631 if (offset < skb->len) 1632 goto found_ok_skb; 1633 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1634 goto found_fin_ok; 1635 WARN(!(flags & MSG_PEEK), 1636 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n", 1637 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags); 1638 } 1639 1640 /* Well, if we have backlog, try to process it now yet. */ 1641 1642 if (copied >= target && !sk->sk_backlog.tail) 1643 break; 1644 1645 if (copied) { 1646 if (sk->sk_err || 1647 sk->sk_state == TCP_CLOSE || 1648 (sk->sk_shutdown & RCV_SHUTDOWN) || 1649 !timeo || 1650 signal_pending(current)) 1651 break; 1652 } else { 1653 if (sock_flag(sk, SOCK_DONE)) 1654 break; 1655 1656 if (sk->sk_err) { 1657 copied = sock_error(sk); 1658 break; 1659 } 1660 1661 if (sk->sk_shutdown & RCV_SHUTDOWN) 1662 break; 1663 1664 if (sk->sk_state == TCP_CLOSE) { 1665 if (!sock_flag(sk, SOCK_DONE)) { 1666 /* This occurs when user tries to read 1667 * from never connected socket. 1668 */ 1669 copied = -ENOTCONN; 1670 break; 1671 } 1672 break; 1673 } 1674 1675 if (!timeo) { 1676 copied = -EAGAIN; 1677 break; 1678 } 1679 1680 if (signal_pending(current)) { 1681 copied = sock_intr_errno(timeo); 1682 break; 1683 } 1684 } 1685 1686 tcp_cleanup_rbuf(sk, copied); 1687 1688 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) { 1689 /* Install new reader */ 1690 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) { 1691 user_recv = current; 1692 tp->ucopy.task = user_recv; 1693 tp->ucopy.msg = msg; 1694 } 1695 1696 tp->ucopy.len = len; 1697 1698 WARN_ON(tp->copied_seq != tp->rcv_nxt && 1699 !(flags & (MSG_PEEK | MSG_TRUNC))); 1700 1701 /* Ugly... If prequeue is not empty, we have to 1702 * process it before releasing socket, otherwise 1703 * order will be broken at second iteration. 1704 * More elegant solution is required!!! 1705 * 1706 * Look: we have the following (pseudo)queues: 1707 * 1708 * 1. packets in flight 1709 * 2. backlog 1710 * 3. prequeue 1711 * 4. receive_queue 1712 * 1713 * Each queue can be processed only if the next ones 1714 * are empty. At this point we have empty receive_queue. 1715 * But prequeue _can_ be not empty after 2nd iteration, 1716 * when we jumped to start of loop because backlog 1717 * processing added something to receive_queue. 1718 * We cannot release_sock(), because backlog contains 1719 * packets arrived _after_ prequeued ones. 1720 * 1721 * Shortly, algorithm is clear --- to process all 1722 * the queues in order. We could make it more directly, 1723 * requeueing packets from backlog to prequeue, if 1724 * is not empty. It is more elegant, but eats cycles, 1725 * unfortunately. 1726 */ 1727 if (!skb_queue_empty(&tp->ucopy.prequeue)) 1728 goto do_prequeue; 1729 1730 /* __ Set realtime policy in scheduler __ */ 1731 } 1732 1733 if (copied >= target) { 1734 /* Do not sleep, just process backlog. */ 1735 release_sock(sk); 1736 lock_sock(sk); 1737 } else 1738 sk_wait_data(sk, &timeo); 1739 1740 if (user_recv) { 1741 int chunk; 1742 1743 /* __ Restore normal policy in scheduler __ */ 1744 1745 if ((chunk = len - tp->ucopy.len) != 0) { 1746 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk); 1747 len -= chunk; 1748 copied += chunk; 1749 } 1750 1751 if (tp->rcv_nxt == tp->copied_seq && 1752 !skb_queue_empty(&tp->ucopy.prequeue)) { 1753 do_prequeue: 1754 tcp_prequeue_process(sk); 1755 1756 if ((chunk = len - tp->ucopy.len) != 0) { 1757 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1758 len -= chunk; 1759 copied += chunk; 1760 } 1761 } 1762 } 1763 if ((flags & MSG_PEEK) && 1764 (peek_seq - copied - urg_hole != tp->copied_seq)) { 1765 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n", 1766 current->comm, 1767 task_pid_nr(current)); 1768 peek_seq = tp->copied_seq; 1769 } 1770 continue; 1771 1772 found_ok_skb: 1773 /* Ok so how much can we use? */ 1774 used = skb->len - offset; 1775 if (len < used) 1776 used = len; 1777 1778 /* Do we have urgent data here? */ 1779 if (tp->urg_data) { 1780 u32 urg_offset = tp->urg_seq - *seq; 1781 if (urg_offset < used) { 1782 if (!urg_offset) { 1783 if (!sock_flag(sk, SOCK_URGINLINE)) { 1784 ++*seq; 1785 urg_hole++; 1786 offset++; 1787 used--; 1788 if (!used) 1789 goto skip_copy; 1790 } 1791 } else 1792 used = urg_offset; 1793 } 1794 } 1795 1796 if (!(flags & MSG_TRUNC)) { 1797 err = skb_copy_datagram_msg(skb, offset, msg, used); 1798 if (err) { 1799 /* Exception. Bailout! */ 1800 if (!copied) 1801 copied = -EFAULT; 1802 break; 1803 } 1804 } 1805 1806 *seq += used; 1807 copied += used; 1808 len -= used; 1809 1810 tcp_rcv_space_adjust(sk); 1811 1812 skip_copy: 1813 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) { 1814 tp->urg_data = 0; 1815 tcp_fast_path_check(sk); 1816 } 1817 if (used + offset < skb->len) 1818 continue; 1819 1820 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1821 goto found_fin_ok; 1822 if (!(flags & MSG_PEEK)) 1823 sk_eat_skb(sk, skb); 1824 continue; 1825 1826 found_fin_ok: 1827 /* Process the FIN. */ 1828 ++*seq; 1829 if (!(flags & MSG_PEEK)) 1830 sk_eat_skb(sk, skb); 1831 break; 1832 } while (len > 0); 1833 1834 if (user_recv) { 1835 if (!skb_queue_empty(&tp->ucopy.prequeue)) { 1836 int chunk; 1837 1838 tp->ucopy.len = copied > 0 ? len : 0; 1839 1840 tcp_prequeue_process(sk); 1841 1842 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) { 1843 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1844 len -= chunk; 1845 copied += chunk; 1846 } 1847 } 1848 1849 tp->ucopy.task = NULL; 1850 tp->ucopy.len = 0; 1851 } 1852 1853 /* According to UNIX98, msg_name/msg_namelen are ignored 1854 * on connected socket. I was just happy when found this 8) --ANK 1855 */ 1856 1857 /* Clean up data we have read: This will do ACK frames. */ 1858 tcp_cleanup_rbuf(sk, copied); 1859 1860 release_sock(sk); 1861 return copied; 1862 1863 out: 1864 release_sock(sk); 1865 return err; 1866 1867 recv_urg: 1868 err = tcp_recv_urg(sk, msg, len, flags); 1869 goto out; 1870 1871 recv_sndq: 1872 err = tcp_peek_sndq(sk, msg, len); 1873 goto out; 1874 } 1875 EXPORT_SYMBOL(tcp_recvmsg); 1876 1877 void tcp_set_state(struct sock *sk, int state) 1878 { 1879 int oldstate = sk->sk_state; 1880 1881 switch (state) { 1882 case TCP_ESTABLISHED: 1883 if (oldstate != TCP_ESTABLISHED) 1884 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1885 break; 1886 1887 case TCP_CLOSE: 1888 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED) 1889 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS); 1890 1891 sk->sk_prot->unhash(sk); 1892 if (inet_csk(sk)->icsk_bind_hash && 1893 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK)) 1894 inet_put_port(sk); 1895 /* fall through */ 1896 default: 1897 if (oldstate == TCP_ESTABLISHED) 1898 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1899 } 1900 1901 /* Change state AFTER socket is unhashed to avoid closed 1902 * socket sitting in hash tables. 1903 */ 1904 sk->sk_state = state; 1905 1906 #ifdef STATE_TRACE 1907 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]); 1908 #endif 1909 } 1910 EXPORT_SYMBOL_GPL(tcp_set_state); 1911 1912 /* 1913 * State processing on a close. This implements the state shift for 1914 * sending our FIN frame. Note that we only send a FIN for some 1915 * states. A shutdown() may have already sent the FIN, or we may be 1916 * closed. 1917 */ 1918 1919 static const unsigned char new_state[16] = { 1920 /* current state: new state: action: */ 1921 /* (Invalid) */ TCP_CLOSE, 1922 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1923 /* TCP_SYN_SENT */ TCP_CLOSE, 1924 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1925 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1, 1926 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2, 1927 /* TCP_TIME_WAIT */ TCP_CLOSE, 1928 /* TCP_CLOSE */ TCP_CLOSE, 1929 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN, 1930 /* TCP_LAST_ACK */ TCP_LAST_ACK, 1931 /* TCP_LISTEN */ TCP_CLOSE, 1932 /* TCP_CLOSING */ TCP_CLOSING, 1933 }; 1934 1935 static int tcp_close_state(struct sock *sk) 1936 { 1937 int next = (int)new_state[sk->sk_state]; 1938 int ns = next & TCP_STATE_MASK; 1939 1940 tcp_set_state(sk, ns); 1941 1942 return next & TCP_ACTION_FIN; 1943 } 1944 1945 /* 1946 * Shutdown the sending side of a connection. Much like close except 1947 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD). 1948 */ 1949 1950 void tcp_shutdown(struct sock *sk, int how) 1951 { 1952 /* We need to grab some memory, and put together a FIN, 1953 * and then put it into the queue to be sent. 1954 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92. 1955 */ 1956 if (!(how & SEND_SHUTDOWN)) 1957 return; 1958 1959 /* If we've already sent a FIN, or it's a closed state, skip this. */ 1960 if ((1 << sk->sk_state) & 1961 (TCPF_ESTABLISHED | TCPF_SYN_SENT | 1962 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) { 1963 /* Clear out any half completed packets. FIN if needed. */ 1964 if (tcp_close_state(sk)) 1965 tcp_send_fin(sk); 1966 } 1967 } 1968 EXPORT_SYMBOL(tcp_shutdown); 1969 1970 bool tcp_check_oom(struct sock *sk, int shift) 1971 { 1972 bool too_many_orphans, out_of_socket_memory; 1973 1974 too_many_orphans = tcp_too_many_orphans(sk, shift); 1975 out_of_socket_memory = tcp_out_of_memory(sk); 1976 1977 if (too_many_orphans) 1978 net_info_ratelimited("too many orphaned sockets\n"); 1979 if (out_of_socket_memory) 1980 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n"); 1981 return too_many_orphans || out_of_socket_memory; 1982 } 1983 1984 void tcp_close(struct sock *sk, long timeout) 1985 { 1986 struct sk_buff *skb; 1987 int data_was_unread = 0; 1988 int state; 1989 1990 lock_sock(sk); 1991 sk->sk_shutdown = SHUTDOWN_MASK; 1992 1993 if (sk->sk_state == TCP_LISTEN) { 1994 tcp_set_state(sk, TCP_CLOSE); 1995 1996 /* Special case. */ 1997 inet_csk_listen_stop(sk); 1998 1999 goto adjudge_to_death; 2000 } 2001 2002 /* We need to flush the recv. buffs. We do this only on the 2003 * descriptor close, not protocol-sourced closes, because the 2004 * reader process may not have drained the data yet! 2005 */ 2006 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) { 2007 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq; 2008 2009 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 2010 len--; 2011 data_was_unread += len; 2012 __kfree_skb(skb); 2013 } 2014 2015 sk_mem_reclaim(sk); 2016 2017 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */ 2018 if (sk->sk_state == TCP_CLOSE) 2019 goto adjudge_to_death; 2020 2021 /* As outlined in RFC 2525, section 2.17, we send a RST here because 2022 * data was lost. To witness the awful effects of the old behavior of 2023 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk 2024 * GET in an FTP client, suspend the process, wait for the client to 2025 * advertise a zero window, then kill -9 the FTP client, wheee... 2026 * Note: timeout is always zero in such a case. 2027 */ 2028 if (unlikely(tcp_sk(sk)->repair)) { 2029 sk->sk_prot->disconnect(sk, 0); 2030 } else if (data_was_unread) { 2031 /* Unread data was tossed, zap the connection. */ 2032 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE); 2033 tcp_set_state(sk, TCP_CLOSE); 2034 tcp_send_active_reset(sk, sk->sk_allocation); 2035 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) { 2036 /* Check zero linger _after_ checking for unread data. */ 2037 sk->sk_prot->disconnect(sk, 0); 2038 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 2039 } else if (tcp_close_state(sk)) { 2040 /* We FIN if the application ate all the data before 2041 * zapping the connection. 2042 */ 2043 2044 /* RED-PEN. Formally speaking, we have broken TCP state 2045 * machine. State transitions: 2046 * 2047 * TCP_ESTABLISHED -> TCP_FIN_WAIT1 2048 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible) 2049 * TCP_CLOSE_WAIT -> TCP_LAST_ACK 2050 * 2051 * are legal only when FIN has been sent (i.e. in window), 2052 * rather than queued out of window. Purists blame. 2053 * 2054 * F.e. "RFC state" is ESTABLISHED, 2055 * if Linux state is FIN-WAIT-1, but FIN is still not sent. 2056 * 2057 * The visible declinations are that sometimes 2058 * we enter time-wait state, when it is not required really 2059 * (harmless), do not send active resets, when they are 2060 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when 2061 * they look as CLOSING or LAST_ACK for Linux) 2062 * Probably, I missed some more holelets. 2063 * --ANK 2064 * XXX (TFO) - To start off we don't support SYN+ACK+FIN 2065 * in a single packet! (May consider it later but will 2066 * probably need API support or TCP_CORK SYN-ACK until 2067 * data is written and socket is closed.) 2068 */ 2069 tcp_send_fin(sk); 2070 } 2071 2072 sk_stream_wait_close(sk, timeout); 2073 2074 adjudge_to_death: 2075 state = sk->sk_state; 2076 sock_hold(sk); 2077 sock_orphan(sk); 2078 2079 /* It is the last release_sock in its life. It will remove backlog. */ 2080 release_sock(sk); 2081 2082 2083 /* Now socket is owned by kernel and we acquire BH lock 2084 to finish close. No need to check for user refs. 2085 */ 2086 local_bh_disable(); 2087 bh_lock_sock(sk); 2088 WARN_ON(sock_owned_by_user(sk)); 2089 2090 percpu_counter_inc(sk->sk_prot->orphan_count); 2091 2092 /* Have we already been destroyed by a softirq or backlog? */ 2093 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE) 2094 goto out; 2095 2096 /* This is a (useful) BSD violating of the RFC. There is a 2097 * problem with TCP as specified in that the other end could 2098 * keep a socket open forever with no application left this end. 2099 * We use a 1 minute timeout (about the same as BSD) then kill 2100 * our end. If they send after that then tough - BUT: long enough 2101 * that we won't make the old 4*rto = almost no time - whoops 2102 * reset mistake. 2103 * 2104 * Nope, it was not mistake. It is really desired behaviour 2105 * f.e. on http servers, when such sockets are useless, but 2106 * consume significant resources. Let's do it with special 2107 * linger2 option. --ANK 2108 */ 2109 2110 if (sk->sk_state == TCP_FIN_WAIT2) { 2111 struct tcp_sock *tp = tcp_sk(sk); 2112 if (tp->linger2 < 0) { 2113 tcp_set_state(sk, TCP_CLOSE); 2114 tcp_send_active_reset(sk, GFP_ATOMIC); 2115 NET_INC_STATS_BH(sock_net(sk), 2116 LINUX_MIB_TCPABORTONLINGER); 2117 } else { 2118 const int tmo = tcp_fin_time(sk); 2119 2120 if (tmo > TCP_TIMEWAIT_LEN) { 2121 inet_csk_reset_keepalive_timer(sk, 2122 tmo - TCP_TIMEWAIT_LEN); 2123 } else { 2124 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 2125 goto out; 2126 } 2127 } 2128 } 2129 if (sk->sk_state != TCP_CLOSE) { 2130 sk_mem_reclaim(sk); 2131 if (tcp_check_oom(sk, 0)) { 2132 tcp_set_state(sk, TCP_CLOSE); 2133 tcp_send_active_reset(sk, GFP_ATOMIC); 2134 NET_INC_STATS_BH(sock_net(sk), 2135 LINUX_MIB_TCPABORTONMEMORY); 2136 } 2137 } 2138 2139 if (sk->sk_state == TCP_CLOSE) { 2140 struct request_sock *req = tcp_sk(sk)->fastopen_rsk; 2141 /* We could get here with a non-NULL req if the socket is 2142 * aborted (e.g., closed with unread data) before 3WHS 2143 * finishes. 2144 */ 2145 if (req != NULL) 2146 reqsk_fastopen_remove(sk, req, false); 2147 inet_csk_destroy_sock(sk); 2148 } 2149 /* Otherwise, socket is reprieved until protocol close. */ 2150 2151 out: 2152 bh_unlock_sock(sk); 2153 local_bh_enable(); 2154 sock_put(sk); 2155 } 2156 EXPORT_SYMBOL(tcp_close); 2157 2158 /* These states need RST on ABORT according to RFC793 */ 2159 2160 static inline bool tcp_need_reset(int state) 2161 { 2162 return (1 << state) & 2163 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 | 2164 TCPF_FIN_WAIT2 | TCPF_SYN_RECV); 2165 } 2166 2167 int tcp_disconnect(struct sock *sk, int flags) 2168 { 2169 struct inet_sock *inet = inet_sk(sk); 2170 struct inet_connection_sock *icsk = inet_csk(sk); 2171 struct tcp_sock *tp = tcp_sk(sk); 2172 int err = 0; 2173 int old_state = sk->sk_state; 2174 2175 if (old_state != TCP_CLOSE) 2176 tcp_set_state(sk, TCP_CLOSE); 2177 2178 /* ABORT function of RFC793 */ 2179 if (old_state == TCP_LISTEN) { 2180 inet_csk_listen_stop(sk); 2181 } else if (unlikely(tp->repair)) { 2182 sk->sk_err = ECONNABORTED; 2183 } else if (tcp_need_reset(old_state) || 2184 (tp->snd_nxt != tp->write_seq && 2185 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) { 2186 /* The last check adjusts for discrepancy of Linux wrt. RFC 2187 * states 2188 */ 2189 tcp_send_active_reset(sk, gfp_any()); 2190 sk->sk_err = ECONNRESET; 2191 } else if (old_state == TCP_SYN_SENT) 2192 sk->sk_err = ECONNRESET; 2193 2194 tcp_clear_xmit_timers(sk); 2195 __skb_queue_purge(&sk->sk_receive_queue); 2196 tcp_write_queue_purge(sk); 2197 __skb_queue_purge(&tp->out_of_order_queue); 2198 2199 inet->inet_dport = 0; 2200 2201 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK)) 2202 inet_reset_saddr(sk); 2203 2204 sk->sk_shutdown = 0; 2205 sock_reset_flag(sk, SOCK_DONE); 2206 tp->srtt_us = 0; 2207 if ((tp->write_seq += tp->max_window + 2) == 0) 2208 tp->write_seq = 1; 2209 icsk->icsk_backoff = 0; 2210 tp->snd_cwnd = 2; 2211 icsk->icsk_probes_out = 0; 2212 tp->packets_out = 0; 2213 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 2214 tp->snd_cwnd_cnt = 0; 2215 tp->window_clamp = 0; 2216 tcp_set_ca_state(sk, TCP_CA_Open); 2217 tcp_clear_retrans(tp); 2218 inet_csk_delack_init(sk); 2219 tcp_init_send_head(sk); 2220 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt)); 2221 __sk_dst_reset(sk); 2222 2223 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash); 2224 2225 sk->sk_error_report(sk); 2226 return err; 2227 } 2228 EXPORT_SYMBOL(tcp_disconnect); 2229 2230 void tcp_sock_destruct(struct sock *sk) 2231 { 2232 inet_sock_destruct(sk); 2233 2234 kfree(inet_csk(sk)->icsk_accept_queue.fastopenq); 2235 } 2236 2237 static inline bool tcp_can_repair_sock(const struct sock *sk) 2238 { 2239 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) && 2240 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED)); 2241 } 2242 2243 static int tcp_repair_options_est(struct tcp_sock *tp, 2244 struct tcp_repair_opt __user *optbuf, unsigned int len) 2245 { 2246 struct tcp_repair_opt opt; 2247 2248 while (len >= sizeof(opt)) { 2249 if (copy_from_user(&opt, optbuf, sizeof(opt))) 2250 return -EFAULT; 2251 2252 optbuf++; 2253 len -= sizeof(opt); 2254 2255 switch (opt.opt_code) { 2256 case TCPOPT_MSS: 2257 tp->rx_opt.mss_clamp = opt.opt_val; 2258 break; 2259 case TCPOPT_WINDOW: 2260 { 2261 u16 snd_wscale = opt.opt_val & 0xFFFF; 2262 u16 rcv_wscale = opt.opt_val >> 16; 2263 2264 if (snd_wscale > 14 || rcv_wscale > 14) 2265 return -EFBIG; 2266 2267 tp->rx_opt.snd_wscale = snd_wscale; 2268 tp->rx_opt.rcv_wscale = rcv_wscale; 2269 tp->rx_opt.wscale_ok = 1; 2270 } 2271 break; 2272 case TCPOPT_SACK_PERM: 2273 if (opt.opt_val != 0) 2274 return -EINVAL; 2275 2276 tp->rx_opt.sack_ok |= TCP_SACK_SEEN; 2277 if (sysctl_tcp_fack) 2278 tcp_enable_fack(tp); 2279 break; 2280 case TCPOPT_TIMESTAMP: 2281 if (opt.opt_val != 0) 2282 return -EINVAL; 2283 2284 tp->rx_opt.tstamp_ok = 1; 2285 break; 2286 } 2287 } 2288 2289 return 0; 2290 } 2291 2292 /* 2293 * Socket option code for TCP. 2294 */ 2295 static int do_tcp_setsockopt(struct sock *sk, int level, 2296 int optname, char __user *optval, unsigned int optlen) 2297 { 2298 struct tcp_sock *tp = tcp_sk(sk); 2299 struct inet_connection_sock *icsk = inet_csk(sk); 2300 int val; 2301 int err = 0; 2302 2303 /* These are data/string values, all the others are ints */ 2304 switch (optname) { 2305 case TCP_CONGESTION: { 2306 char name[TCP_CA_NAME_MAX]; 2307 2308 if (optlen < 1) 2309 return -EINVAL; 2310 2311 val = strncpy_from_user(name, optval, 2312 min_t(long, TCP_CA_NAME_MAX-1, optlen)); 2313 if (val < 0) 2314 return -EFAULT; 2315 name[val] = 0; 2316 2317 lock_sock(sk); 2318 err = tcp_set_congestion_control(sk, name); 2319 release_sock(sk); 2320 return err; 2321 } 2322 default: 2323 /* fallthru */ 2324 break; 2325 } 2326 2327 if (optlen < sizeof(int)) 2328 return -EINVAL; 2329 2330 if (get_user(val, (int __user *)optval)) 2331 return -EFAULT; 2332 2333 lock_sock(sk); 2334 2335 switch (optname) { 2336 case TCP_MAXSEG: 2337 /* Values greater than interface MTU won't take effect. However 2338 * at the point when this call is done we typically don't yet 2339 * know which interface is going to be used */ 2340 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) { 2341 err = -EINVAL; 2342 break; 2343 } 2344 tp->rx_opt.user_mss = val; 2345 break; 2346 2347 case TCP_NODELAY: 2348 if (val) { 2349 /* TCP_NODELAY is weaker than TCP_CORK, so that 2350 * this option on corked socket is remembered, but 2351 * it is not activated until cork is cleared. 2352 * 2353 * However, when TCP_NODELAY is set we make 2354 * an explicit push, which overrides even TCP_CORK 2355 * for currently queued segments. 2356 */ 2357 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH; 2358 tcp_push_pending_frames(sk); 2359 } else { 2360 tp->nonagle &= ~TCP_NAGLE_OFF; 2361 } 2362 break; 2363 2364 case TCP_THIN_LINEAR_TIMEOUTS: 2365 if (val < 0 || val > 1) 2366 err = -EINVAL; 2367 else 2368 tp->thin_lto = val; 2369 break; 2370 2371 case TCP_THIN_DUPACK: 2372 if (val < 0 || val > 1) 2373 err = -EINVAL; 2374 else { 2375 tp->thin_dupack = val; 2376 if (tp->thin_dupack) 2377 tcp_disable_early_retrans(tp); 2378 } 2379 break; 2380 2381 case TCP_REPAIR: 2382 if (!tcp_can_repair_sock(sk)) 2383 err = -EPERM; 2384 else if (val == 1) { 2385 tp->repair = 1; 2386 sk->sk_reuse = SK_FORCE_REUSE; 2387 tp->repair_queue = TCP_NO_QUEUE; 2388 } else if (val == 0) { 2389 tp->repair = 0; 2390 sk->sk_reuse = SK_NO_REUSE; 2391 tcp_send_window_probe(sk); 2392 } else 2393 err = -EINVAL; 2394 2395 break; 2396 2397 case TCP_REPAIR_QUEUE: 2398 if (!tp->repair) 2399 err = -EPERM; 2400 else if (val < TCP_QUEUES_NR) 2401 tp->repair_queue = val; 2402 else 2403 err = -EINVAL; 2404 break; 2405 2406 case TCP_QUEUE_SEQ: 2407 if (sk->sk_state != TCP_CLOSE) 2408 err = -EPERM; 2409 else if (tp->repair_queue == TCP_SEND_QUEUE) 2410 tp->write_seq = val; 2411 else if (tp->repair_queue == TCP_RECV_QUEUE) 2412 tp->rcv_nxt = val; 2413 else 2414 err = -EINVAL; 2415 break; 2416 2417 case TCP_REPAIR_OPTIONS: 2418 if (!tp->repair) 2419 err = -EINVAL; 2420 else if (sk->sk_state == TCP_ESTABLISHED) 2421 err = tcp_repair_options_est(tp, 2422 (struct tcp_repair_opt __user *)optval, 2423 optlen); 2424 else 2425 err = -EPERM; 2426 break; 2427 2428 case TCP_CORK: 2429 /* When set indicates to always queue non-full frames. 2430 * Later the user clears this option and we transmit 2431 * any pending partial frames in the queue. This is 2432 * meant to be used alongside sendfile() to get properly 2433 * filled frames when the user (for example) must write 2434 * out headers with a write() call first and then use 2435 * sendfile to send out the data parts. 2436 * 2437 * TCP_CORK can be set together with TCP_NODELAY and it is 2438 * stronger than TCP_NODELAY. 2439 */ 2440 if (val) { 2441 tp->nonagle |= TCP_NAGLE_CORK; 2442 } else { 2443 tp->nonagle &= ~TCP_NAGLE_CORK; 2444 if (tp->nonagle&TCP_NAGLE_OFF) 2445 tp->nonagle |= TCP_NAGLE_PUSH; 2446 tcp_push_pending_frames(sk); 2447 } 2448 break; 2449 2450 case TCP_KEEPIDLE: 2451 if (val < 1 || val > MAX_TCP_KEEPIDLE) 2452 err = -EINVAL; 2453 else { 2454 tp->keepalive_time = val * HZ; 2455 if (sock_flag(sk, SOCK_KEEPOPEN) && 2456 !((1 << sk->sk_state) & 2457 (TCPF_CLOSE | TCPF_LISTEN))) { 2458 u32 elapsed = keepalive_time_elapsed(tp); 2459 if (tp->keepalive_time > elapsed) 2460 elapsed = tp->keepalive_time - elapsed; 2461 else 2462 elapsed = 0; 2463 inet_csk_reset_keepalive_timer(sk, elapsed); 2464 } 2465 } 2466 break; 2467 case TCP_KEEPINTVL: 2468 if (val < 1 || val > MAX_TCP_KEEPINTVL) 2469 err = -EINVAL; 2470 else 2471 tp->keepalive_intvl = val * HZ; 2472 break; 2473 case TCP_KEEPCNT: 2474 if (val < 1 || val > MAX_TCP_KEEPCNT) 2475 err = -EINVAL; 2476 else 2477 tp->keepalive_probes = val; 2478 break; 2479 case TCP_SYNCNT: 2480 if (val < 1 || val > MAX_TCP_SYNCNT) 2481 err = -EINVAL; 2482 else 2483 icsk->icsk_syn_retries = val; 2484 break; 2485 2486 case TCP_LINGER2: 2487 if (val < 0) 2488 tp->linger2 = -1; 2489 else if (val > sysctl_tcp_fin_timeout / HZ) 2490 tp->linger2 = 0; 2491 else 2492 tp->linger2 = val * HZ; 2493 break; 2494 2495 case TCP_DEFER_ACCEPT: 2496 /* Translate value in seconds to number of retransmits */ 2497 icsk->icsk_accept_queue.rskq_defer_accept = 2498 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ, 2499 TCP_RTO_MAX / HZ); 2500 break; 2501 2502 case TCP_WINDOW_CLAMP: 2503 if (!val) { 2504 if (sk->sk_state != TCP_CLOSE) { 2505 err = -EINVAL; 2506 break; 2507 } 2508 tp->window_clamp = 0; 2509 } else 2510 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ? 2511 SOCK_MIN_RCVBUF / 2 : val; 2512 break; 2513 2514 case TCP_QUICKACK: 2515 if (!val) { 2516 icsk->icsk_ack.pingpong = 1; 2517 } else { 2518 icsk->icsk_ack.pingpong = 0; 2519 if ((1 << sk->sk_state) & 2520 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) && 2521 inet_csk_ack_scheduled(sk)) { 2522 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 2523 tcp_cleanup_rbuf(sk, 1); 2524 if (!(val & 1)) 2525 icsk->icsk_ack.pingpong = 1; 2526 } 2527 } 2528 break; 2529 2530 #ifdef CONFIG_TCP_MD5SIG 2531 case TCP_MD5SIG: 2532 /* Read the IP->Key mappings from userspace */ 2533 err = tp->af_specific->md5_parse(sk, optval, optlen); 2534 break; 2535 #endif 2536 case TCP_USER_TIMEOUT: 2537 /* Cap the max time in ms TCP will retry or probe the window 2538 * before giving up and aborting (ETIMEDOUT) a connection. 2539 */ 2540 if (val < 0) 2541 err = -EINVAL; 2542 else 2543 icsk->icsk_user_timeout = msecs_to_jiffies(val); 2544 break; 2545 2546 case TCP_FASTOPEN: 2547 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE | 2548 TCPF_LISTEN))) 2549 err = fastopen_init_queue(sk, val); 2550 else 2551 err = -EINVAL; 2552 break; 2553 case TCP_TIMESTAMP: 2554 if (!tp->repair) 2555 err = -EPERM; 2556 else 2557 tp->tsoffset = val - tcp_time_stamp; 2558 break; 2559 case TCP_NOTSENT_LOWAT: 2560 tp->notsent_lowat = val; 2561 sk->sk_write_space(sk); 2562 break; 2563 default: 2564 err = -ENOPROTOOPT; 2565 break; 2566 } 2567 2568 release_sock(sk); 2569 return err; 2570 } 2571 2572 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval, 2573 unsigned int optlen) 2574 { 2575 const struct inet_connection_sock *icsk = inet_csk(sk); 2576 2577 if (level != SOL_TCP) 2578 return icsk->icsk_af_ops->setsockopt(sk, level, optname, 2579 optval, optlen); 2580 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2581 } 2582 EXPORT_SYMBOL(tcp_setsockopt); 2583 2584 #ifdef CONFIG_COMPAT 2585 int compat_tcp_setsockopt(struct sock *sk, int level, int optname, 2586 char __user *optval, unsigned int optlen) 2587 { 2588 if (level != SOL_TCP) 2589 return inet_csk_compat_setsockopt(sk, level, optname, 2590 optval, optlen); 2591 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2592 } 2593 EXPORT_SYMBOL(compat_tcp_setsockopt); 2594 #endif 2595 2596 /* Return information about state of tcp endpoint in API format. */ 2597 void tcp_get_info(const struct sock *sk, struct tcp_info *info) 2598 { 2599 const struct tcp_sock *tp = tcp_sk(sk); 2600 const struct inet_connection_sock *icsk = inet_csk(sk); 2601 u32 now = tcp_time_stamp; 2602 2603 memset(info, 0, sizeof(*info)); 2604 2605 info->tcpi_state = sk->sk_state; 2606 info->tcpi_ca_state = icsk->icsk_ca_state; 2607 info->tcpi_retransmits = icsk->icsk_retransmits; 2608 info->tcpi_probes = icsk->icsk_probes_out; 2609 info->tcpi_backoff = icsk->icsk_backoff; 2610 2611 if (tp->rx_opt.tstamp_ok) 2612 info->tcpi_options |= TCPI_OPT_TIMESTAMPS; 2613 if (tcp_is_sack(tp)) 2614 info->tcpi_options |= TCPI_OPT_SACK; 2615 if (tp->rx_opt.wscale_ok) { 2616 info->tcpi_options |= TCPI_OPT_WSCALE; 2617 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale; 2618 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale; 2619 } 2620 2621 if (tp->ecn_flags & TCP_ECN_OK) 2622 info->tcpi_options |= TCPI_OPT_ECN; 2623 if (tp->ecn_flags & TCP_ECN_SEEN) 2624 info->tcpi_options |= TCPI_OPT_ECN_SEEN; 2625 if (tp->syn_data_acked) 2626 info->tcpi_options |= TCPI_OPT_SYN_DATA; 2627 2628 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto); 2629 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato); 2630 info->tcpi_snd_mss = tp->mss_cache; 2631 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss; 2632 2633 if (sk->sk_state == TCP_LISTEN) { 2634 info->tcpi_unacked = sk->sk_ack_backlog; 2635 info->tcpi_sacked = sk->sk_max_ack_backlog; 2636 } else { 2637 info->tcpi_unacked = tp->packets_out; 2638 info->tcpi_sacked = tp->sacked_out; 2639 } 2640 info->tcpi_lost = tp->lost_out; 2641 info->tcpi_retrans = tp->retrans_out; 2642 info->tcpi_fackets = tp->fackets_out; 2643 2644 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime); 2645 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime); 2646 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp); 2647 2648 info->tcpi_pmtu = icsk->icsk_pmtu_cookie; 2649 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh; 2650 info->tcpi_rtt = tp->srtt_us >> 3; 2651 info->tcpi_rttvar = tp->mdev_us >> 2; 2652 info->tcpi_snd_ssthresh = tp->snd_ssthresh; 2653 info->tcpi_snd_cwnd = tp->snd_cwnd; 2654 info->tcpi_advmss = tp->advmss; 2655 info->tcpi_reordering = tp->reordering; 2656 2657 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3; 2658 info->tcpi_rcv_space = tp->rcvq_space.space; 2659 2660 info->tcpi_total_retrans = tp->total_retrans; 2661 2662 info->tcpi_pacing_rate = sk->sk_pacing_rate != ~0U ? 2663 sk->sk_pacing_rate : ~0ULL; 2664 info->tcpi_max_pacing_rate = sk->sk_max_pacing_rate != ~0U ? 2665 sk->sk_max_pacing_rate : ~0ULL; 2666 } 2667 EXPORT_SYMBOL_GPL(tcp_get_info); 2668 2669 static int do_tcp_getsockopt(struct sock *sk, int level, 2670 int optname, char __user *optval, int __user *optlen) 2671 { 2672 struct inet_connection_sock *icsk = inet_csk(sk); 2673 struct tcp_sock *tp = tcp_sk(sk); 2674 int val, len; 2675 2676 if (get_user(len, optlen)) 2677 return -EFAULT; 2678 2679 len = min_t(unsigned int, len, sizeof(int)); 2680 2681 if (len < 0) 2682 return -EINVAL; 2683 2684 switch (optname) { 2685 case TCP_MAXSEG: 2686 val = tp->mss_cache; 2687 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 2688 val = tp->rx_opt.user_mss; 2689 if (tp->repair) 2690 val = tp->rx_opt.mss_clamp; 2691 break; 2692 case TCP_NODELAY: 2693 val = !!(tp->nonagle&TCP_NAGLE_OFF); 2694 break; 2695 case TCP_CORK: 2696 val = !!(tp->nonagle&TCP_NAGLE_CORK); 2697 break; 2698 case TCP_KEEPIDLE: 2699 val = keepalive_time_when(tp) / HZ; 2700 break; 2701 case TCP_KEEPINTVL: 2702 val = keepalive_intvl_when(tp) / HZ; 2703 break; 2704 case TCP_KEEPCNT: 2705 val = keepalive_probes(tp); 2706 break; 2707 case TCP_SYNCNT: 2708 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries; 2709 break; 2710 case TCP_LINGER2: 2711 val = tp->linger2; 2712 if (val >= 0) 2713 val = (val ? : sysctl_tcp_fin_timeout) / HZ; 2714 break; 2715 case TCP_DEFER_ACCEPT: 2716 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept, 2717 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ); 2718 break; 2719 case TCP_WINDOW_CLAMP: 2720 val = tp->window_clamp; 2721 break; 2722 case TCP_INFO: { 2723 struct tcp_info info; 2724 2725 if (get_user(len, optlen)) 2726 return -EFAULT; 2727 2728 tcp_get_info(sk, &info); 2729 2730 len = min_t(unsigned int, len, sizeof(info)); 2731 if (put_user(len, optlen)) 2732 return -EFAULT; 2733 if (copy_to_user(optval, &info, len)) 2734 return -EFAULT; 2735 return 0; 2736 } 2737 case TCP_QUICKACK: 2738 val = !icsk->icsk_ack.pingpong; 2739 break; 2740 2741 case TCP_CONGESTION: 2742 if (get_user(len, optlen)) 2743 return -EFAULT; 2744 len = min_t(unsigned int, len, TCP_CA_NAME_MAX); 2745 if (put_user(len, optlen)) 2746 return -EFAULT; 2747 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len)) 2748 return -EFAULT; 2749 return 0; 2750 2751 case TCP_THIN_LINEAR_TIMEOUTS: 2752 val = tp->thin_lto; 2753 break; 2754 case TCP_THIN_DUPACK: 2755 val = tp->thin_dupack; 2756 break; 2757 2758 case TCP_REPAIR: 2759 val = tp->repair; 2760 break; 2761 2762 case TCP_REPAIR_QUEUE: 2763 if (tp->repair) 2764 val = tp->repair_queue; 2765 else 2766 return -EINVAL; 2767 break; 2768 2769 case TCP_QUEUE_SEQ: 2770 if (tp->repair_queue == TCP_SEND_QUEUE) 2771 val = tp->write_seq; 2772 else if (tp->repair_queue == TCP_RECV_QUEUE) 2773 val = tp->rcv_nxt; 2774 else 2775 return -EINVAL; 2776 break; 2777 2778 case TCP_USER_TIMEOUT: 2779 val = jiffies_to_msecs(icsk->icsk_user_timeout); 2780 break; 2781 2782 case TCP_FASTOPEN: 2783 if (icsk->icsk_accept_queue.fastopenq != NULL) 2784 val = icsk->icsk_accept_queue.fastopenq->max_qlen; 2785 else 2786 val = 0; 2787 break; 2788 2789 case TCP_TIMESTAMP: 2790 val = tcp_time_stamp + tp->tsoffset; 2791 break; 2792 case TCP_NOTSENT_LOWAT: 2793 val = tp->notsent_lowat; 2794 break; 2795 default: 2796 return -ENOPROTOOPT; 2797 } 2798 2799 if (put_user(len, optlen)) 2800 return -EFAULT; 2801 if (copy_to_user(optval, &val, len)) 2802 return -EFAULT; 2803 return 0; 2804 } 2805 2806 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval, 2807 int __user *optlen) 2808 { 2809 struct inet_connection_sock *icsk = inet_csk(sk); 2810 2811 if (level != SOL_TCP) 2812 return icsk->icsk_af_ops->getsockopt(sk, level, optname, 2813 optval, optlen); 2814 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2815 } 2816 EXPORT_SYMBOL(tcp_getsockopt); 2817 2818 #ifdef CONFIG_COMPAT 2819 int compat_tcp_getsockopt(struct sock *sk, int level, int optname, 2820 char __user *optval, int __user *optlen) 2821 { 2822 if (level != SOL_TCP) 2823 return inet_csk_compat_getsockopt(sk, level, optname, 2824 optval, optlen); 2825 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2826 } 2827 EXPORT_SYMBOL(compat_tcp_getsockopt); 2828 #endif 2829 2830 #ifdef CONFIG_TCP_MD5SIG 2831 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool); 2832 static DEFINE_MUTEX(tcp_md5sig_mutex); 2833 static bool tcp_md5sig_pool_populated = false; 2834 2835 static void __tcp_alloc_md5sig_pool(void) 2836 { 2837 int cpu; 2838 2839 for_each_possible_cpu(cpu) { 2840 if (!per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm) { 2841 struct crypto_hash *hash; 2842 2843 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC); 2844 if (IS_ERR_OR_NULL(hash)) 2845 return; 2846 per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm = hash; 2847 } 2848 } 2849 /* before setting tcp_md5sig_pool_populated, we must commit all writes 2850 * to memory. See smp_rmb() in tcp_get_md5sig_pool() 2851 */ 2852 smp_wmb(); 2853 tcp_md5sig_pool_populated = true; 2854 } 2855 2856 bool tcp_alloc_md5sig_pool(void) 2857 { 2858 if (unlikely(!tcp_md5sig_pool_populated)) { 2859 mutex_lock(&tcp_md5sig_mutex); 2860 2861 if (!tcp_md5sig_pool_populated) 2862 __tcp_alloc_md5sig_pool(); 2863 2864 mutex_unlock(&tcp_md5sig_mutex); 2865 } 2866 return tcp_md5sig_pool_populated; 2867 } 2868 EXPORT_SYMBOL(tcp_alloc_md5sig_pool); 2869 2870 2871 /** 2872 * tcp_get_md5sig_pool - get md5sig_pool for this user 2873 * 2874 * We use percpu structure, so if we succeed, we exit with preemption 2875 * and BH disabled, to make sure another thread or softirq handling 2876 * wont try to get same context. 2877 */ 2878 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void) 2879 { 2880 local_bh_disable(); 2881 2882 if (tcp_md5sig_pool_populated) { 2883 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */ 2884 smp_rmb(); 2885 return this_cpu_ptr(&tcp_md5sig_pool); 2886 } 2887 local_bh_enable(); 2888 return NULL; 2889 } 2890 EXPORT_SYMBOL(tcp_get_md5sig_pool); 2891 2892 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp, 2893 const struct tcphdr *th) 2894 { 2895 struct scatterlist sg; 2896 struct tcphdr hdr; 2897 int err; 2898 2899 /* We are not allowed to change tcphdr, make a local copy */ 2900 memcpy(&hdr, th, sizeof(hdr)); 2901 hdr.check = 0; 2902 2903 /* options aren't included in the hash */ 2904 sg_init_one(&sg, &hdr, sizeof(hdr)); 2905 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr)); 2906 return err; 2907 } 2908 EXPORT_SYMBOL(tcp_md5_hash_header); 2909 2910 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp, 2911 const struct sk_buff *skb, unsigned int header_len) 2912 { 2913 struct scatterlist sg; 2914 const struct tcphdr *tp = tcp_hdr(skb); 2915 struct hash_desc *desc = &hp->md5_desc; 2916 unsigned int i; 2917 const unsigned int head_data_len = skb_headlen(skb) > header_len ? 2918 skb_headlen(skb) - header_len : 0; 2919 const struct skb_shared_info *shi = skb_shinfo(skb); 2920 struct sk_buff *frag_iter; 2921 2922 sg_init_table(&sg, 1); 2923 2924 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len); 2925 if (crypto_hash_update(desc, &sg, head_data_len)) 2926 return 1; 2927 2928 for (i = 0; i < shi->nr_frags; ++i) { 2929 const struct skb_frag_struct *f = &shi->frags[i]; 2930 unsigned int offset = f->page_offset; 2931 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT); 2932 2933 sg_set_page(&sg, page, skb_frag_size(f), 2934 offset_in_page(offset)); 2935 if (crypto_hash_update(desc, &sg, skb_frag_size(f))) 2936 return 1; 2937 } 2938 2939 skb_walk_frags(skb, frag_iter) 2940 if (tcp_md5_hash_skb_data(hp, frag_iter, 0)) 2941 return 1; 2942 2943 return 0; 2944 } 2945 EXPORT_SYMBOL(tcp_md5_hash_skb_data); 2946 2947 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key) 2948 { 2949 struct scatterlist sg; 2950 2951 sg_init_one(&sg, key->key, key->keylen); 2952 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen); 2953 } 2954 EXPORT_SYMBOL(tcp_md5_hash_key); 2955 2956 #endif 2957 2958 void tcp_done(struct sock *sk) 2959 { 2960 struct request_sock *req = tcp_sk(sk)->fastopen_rsk; 2961 2962 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV) 2963 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 2964 2965 tcp_set_state(sk, TCP_CLOSE); 2966 tcp_clear_xmit_timers(sk); 2967 if (req != NULL) 2968 reqsk_fastopen_remove(sk, req, false); 2969 2970 sk->sk_shutdown = SHUTDOWN_MASK; 2971 2972 if (!sock_flag(sk, SOCK_DEAD)) 2973 sk->sk_state_change(sk); 2974 else 2975 inet_csk_destroy_sock(sk); 2976 } 2977 EXPORT_SYMBOL_GPL(tcp_done); 2978 2979 extern struct tcp_congestion_ops tcp_reno; 2980 2981 static __initdata unsigned long thash_entries; 2982 static int __init set_thash_entries(char *str) 2983 { 2984 ssize_t ret; 2985 2986 if (!str) 2987 return 0; 2988 2989 ret = kstrtoul(str, 0, &thash_entries); 2990 if (ret) 2991 return 0; 2992 2993 return 1; 2994 } 2995 __setup("thash_entries=", set_thash_entries); 2996 2997 static void __init tcp_init_mem(void) 2998 { 2999 unsigned long limit = nr_free_buffer_pages() / 8; 3000 limit = max(limit, 128UL); 3001 sysctl_tcp_mem[0] = limit / 4 * 3; 3002 sysctl_tcp_mem[1] = limit; 3003 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; 3004 } 3005 3006 void __init tcp_init(void) 3007 { 3008 struct sk_buff *skb = NULL; 3009 unsigned long limit; 3010 int max_rshare, max_wshare, cnt; 3011 unsigned int i; 3012 3013 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb)); 3014 3015 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL); 3016 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL); 3017 tcp_hashinfo.bind_bucket_cachep = 3018 kmem_cache_create("tcp_bind_bucket", 3019 sizeof(struct inet_bind_bucket), 0, 3020 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL); 3021 3022 /* Size and allocate the main established and bind bucket 3023 * hash tables. 3024 * 3025 * The methodology is similar to that of the buffer cache. 3026 */ 3027 tcp_hashinfo.ehash = 3028 alloc_large_system_hash("TCP established", 3029 sizeof(struct inet_ehash_bucket), 3030 thash_entries, 3031 17, /* one slot per 128 KB of memory */ 3032 0, 3033 NULL, 3034 &tcp_hashinfo.ehash_mask, 3035 0, 3036 thash_entries ? 0 : 512 * 1024); 3037 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) 3038 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i); 3039 3040 if (inet_ehash_locks_alloc(&tcp_hashinfo)) 3041 panic("TCP: failed to alloc ehash_locks"); 3042 tcp_hashinfo.bhash = 3043 alloc_large_system_hash("TCP bind", 3044 sizeof(struct inet_bind_hashbucket), 3045 tcp_hashinfo.ehash_mask + 1, 3046 17, /* one slot per 128 KB of memory */ 3047 0, 3048 &tcp_hashinfo.bhash_size, 3049 NULL, 3050 0, 3051 64 * 1024); 3052 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size; 3053 for (i = 0; i < tcp_hashinfo.bhash_size; i++) { 3054 spin_lock_init(&tcp_hashinfo.bhash[i].lock); 3055 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain); 3056 } 3057 3058 3059 cnt = tcp_hashinfo.ehash_mask + 1; 3060 3061 tcp_death_row.sysctl_max_tw_buckets = cnt / 2; 3062 sysctl_tcp_max_orphans = cnt / 2; 3063 sysctl_max_syn_backlog = max(128, cnt / 256); 3064 3065 tcp_init_mem(); 3066 /* Set per-socket limits to no more than 1/128 the pressure threshold */ 3067 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7); 3068 max_wshare = min(4UL*1024*1024, limit); 3069 max_rshare = min(6UL*1024*1024, limit); 3070 3071 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM; 3072 sysctl_tcp_wmem[1] = 16*1024; 3073 sysctl_tcp_wmem[2] = max(64*1024, max_wshare); 3074 3075 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM; 3076 sysctl_tcp_rmem[1] = 87380; 3077 sysctl_tcp_rmem[2] = max(87380, max_rshare); 3078 3079 pr_info("Hash tables configured (established %u bind %u)\n", 3080 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size); 3081 3082 tcp_metrics_init(); 3083 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0); 3084 tcp_tasklet_init(); 3085 } 3086