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/inet_diag.h> 256 #include <linux/init.h> 257 #include <linux/fs.h> 258 #include <linux/skbuff.h> 259 #include <linux/scatterlist.h> 260 #include <linux/splice.h> 261 #include <linux/net.h> 262 #include <linux/socket.h> 263 #include <linux/random.h> 264 #include <linux/bootmem.h> 265 #include <linux/highmem.h> 266 #include <linux/swap.h> 267 #include <linux/cache.h> 268 #include <linux/err.h> 269 #include <linux/crypto.h> 270 #include <linux/time.h> 271 #include <linux/slab.h> 272 273 #include <net/icmp.h> 274 #include <net/inet_common.h> 275 #include <net/tcp.h> 276 #include <net/xfrm.h> 277 #include <net/ip.h> 278 #include <net/sock.h> 279 280 #include <asm/uaccess.h> 281 #include <asm/ioctls.h> 282 #include <net/busy_poll.h> 283 284 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT; 285 286 int sysctl_tcp_min_tso_segs __read_mostly = 2; 287 288 int sysctl_tcp_autocorking __read_mostly = 1; 289 290 struct percpu_counter tcp_orphan_count; 291 EXPORT_SYMBOL_GPL(tcp_orphan_count); 292 293 long sysctl_tcp_mem[3] __read_mostly; 294 int sysctl_tcp_wmem[3] __read_mostly; 295 int sysctl_tcp_rmem[3] __read_mostly; 296 297 EXPORT_SYMBOL(sysctl_tcp_mem); 298 EXPORT_SYMBOL(sysctl_tcp_rmem); 299 EXPORT_SYMBOL(sysctl_tcp_wmem); 300 301 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */ 302 EXPORT_SYMBOL(tcp_memory_allocated); 303 304 /* 305 * Current number of TCP sockets. 306 */ 307 struct percpu_counter tcp_sockets_allocated; 308 EXPORT_SYMBOL(tcp_sockets_allocated); 309 310 /* 311 * TCP splice context 312 */ 313 struct tcp_splice_state { 314 struct pipe_inode_info *pipe; 315 size_t len; 316 unsigned int flags; 317 }; 318 319 /* 320 * Pressure flag: try to collapse. 321 * Technical note: it is used by multiple contexts non atomically. 322 * All the __sk_mem_schedule() is of this nature: accounting 323 * is strict, actions are advisory and have some latency. 324 */ 325 int tcp_memory_pressure __read_mostly; 326 EXPORT_SYMBOL(tcp_memory_pressure); 327 328 void tcp_enter_memory_pressure(struct sock *sk) 329 { 330 if (!tcp_memory_pressure) { 331 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES); 332 tcp_memory_pressure = 1; 333 } 334 } 335 EXPORT_SYMBOL(tcp_enter_memory_pressure); 336 337 /* Convert seconds to retransmits based on initial and max timeout */ 338 static u8 secs_to_retrans(int seconds, int timeout, int rto_max) 339 { 340 u8 res = 0; 341 342 if (seconds > 0) { 343 int period = timeout; 344 345 res = 1; 346 while (seconds > period && res < 255) { 347 res++; 348 timeout <<= 1; 349 if (timeout > rto_max) 350 timeout = rto_max; 351 period += timeout; 352 } 353 } 354 return res; 355 } 356 357 /* Convert retransmits to seconds based on initial and max timeout */ 358 static int retrans_to_secs(u8 retrans, int timeout, int rto_max) 359 { 360 int period = 0; 361 362 if (retrans > 0) { 363 period = timeout; 364 while (--retrans) { 365 timeout <<= 1; 366 if (timeout > rto_max) 367 timeout = rto_max; 368 period += timeout; 369 } 370 } 371 return period; 372 } 373 374 /* Address-family independent initialization for a tcp_sock. 375 * 376 * NOTE: A lot of things set to zero explicitly by call to 377 * sk_alloc() so need not be done here. 378 */ 379 void tcp_init_sock(struct sock *sk) 380 { 381 struct inet_connection_sock *icsk = inet_csk(sk); 382 struct tcp_sock *tp = tcp_sk(sk); 383 384 __skb_queue_head_init(&tp->out_of_order_queue); 385 tcp_init_xmit_timers(sk); 386 tcp_prequeue_init(tp); 387 INIT_LIST_HEAD(&tp->tsq_node); 388 389 icsk->icsk_rto = TCP_TIMEOUT_INIT; 390 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT); 391 392 /* So many TCP implementations out there (incorrectly) count the 393 * initial SYN frame in their delayed-ACK and congestion control 394 * algorithms that we must have the following bandaid to talk 395 * efficiently to them. -DaveM 396 */ 397 tp->snd_cwnd = TCP_INIT_CWND; 398 399 /* See draft-stevens-tcpca-spec-01 for discussion of the 400 * initialization of these values. 401 */ 402 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 403 tp->snd_cwnd_clamp = ~0; 404 tp->mss_cache = TCP_MSS_DEFAULT; 405 u64_stats_init(&tp->syncp); 406 407 tp->reordering = sysctl_tcp_reordering; 408 tcp_enable_early_retrans(tp); 409 tcp_assign_congestion_control(sk); 410 411 tp->tsoffset = 0; 412 413 sk->sk_state = TCP_CLOSE; 414 415 sk->sk_write_space = sk_stream_write_space; 416 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE); 417 418 icsk->icsk_sync_mss = tcp_sync_mss; 419 420 sk->sk_sndbuf = sysctl_tcp_wmem[1]; 421 sk->sk_rcvbuf = sysctl_tcp_rmem[1]; 422 423 local_bh_disable(); 424 sock_update_memcg(sk); 425 sk_sockets_allocated_inc(sk); 426 local_bh_enable(); 427 } 428 EXPORT_SYMBOL(tcp_init_sock); 429 430 static void tcp_tx_timestamp(struct sock *sk, struct sk_buff *skb) 431 { 432 if (sk->sk_tsflags) { 433 struct skb_shared_info *shinfo = skb_shinfo(skb); 434 435 sock_tx_timestamp(sk, &shinfo->tx_flags); 436 if (shinfo->tx_flags & SKBTX_ANY_TSTAMP) 437 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1; 438 } 439 } 440 441 /* 442 * Wait for a TCP event. 443 * 444 * Note that we don't need to lock the socket, as the upper poll layers 445 * take care of normal races (between the test and the event) and we don't 446 * go look at any of the socket buffers directly. 447 */ 448 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait) 449 { 450 unsigned int mask; 451 struct sock *sk = sock->sk; 452 const struct tcp_sock *tp = tcp_sk(sk); 453 454 sock_rps_record_flow(sk); 455 456 sock_poll_wait(file, sk_sleep(sk), wait); 457 if (sk->sk_state == TCP_LISTEN) 458 return inet_csk_listen_poll(sk); 459 460 /* Socket is not locked. We are protected from async events 461 * by poll logic and correct handling of state changes 462 * made by other threads is impossible in any case. 463 */ 464 465 mask = 0; 466 467 /* 468 * POLLHUP is certainly not done right. But poll() doesn't 469 * have a notion of HUP in just one direction, and for a 470 * socket the read side is more interesting. 471 * 472 * Some poll() documentation says that POLLHUP is incompatible 473 * with the POLLOUT/POLLWR flags, so somebody should check this 474 * all. But careful, it tends to be safer to return too many 475 * bits than too few, and you can easily break real applications 476 * if you don't tell them that something has hung up! 477 * 478 * Check-me. 479 * 480 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and 481 * our fs/select.c). It means that after we received EOF, 482 * poll always returns immediately, making impossible poll() on write() 483 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP 484 * if and only if shutdown has been made in both directions. 485 * Actually, it is interesting to look how Solaris and DUX 486 * solve this dilemma. I would prefer, if POLLHUP were maskable, 487 * then we could set it on SND_SHUTDOWN. BTW examples given 488 * in Stevens' books assume exactly this behaviour, it explains 489 * why POLLHUP is incompatible with POLLOUT. --ANK 490 * 491 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent 492 * blocking on fresh not-connected or disconnected socket. --ANK 493 */ 494 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE) 495 mask |= POLLHUP; 496 if (sk->sk_shutdown & RCV_SHUTDOWN) 497 mask |= POLLIN | POLLRDNORM | POLLRDHUP; 498 499 /* Connected or passive Fast Open socket? */ 500 if (sk->sk_state != TCP_SYN_SENT && 501 (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk)) { 502 int target = sock_rcvlowat(sk, 0, INT_MAX); 503 504 if (tp->urg_seq == tp->copied_seq && 505 !sock_flag(sk, SOCK_URGINLINE) && 506 tp->urg_data) 507 target++; 508 509 /* Potential race condition. If read of tp below will 510 * escape above sk->sk_state, we can be illegally awaken 511 * in SYN_* states. */ 512 if (tp->rcv_nxt - tp->copied_seq >= target) 513 mask |= POLLIN | POLLRDNORM; 514 515 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) { 516 if (sk_stream_is_writeable(sk)) { 517 mask |= POLLOUT | POLLWRNORM; 518 } else { /* send SIGIO later */ 519 set_bit(SOCK_ASYNC_NOSPACE, 520 &sk->sk_socket->flags); 521 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 522 523 /* Race breaker. If space is freed after 524 * wspace test but before the flags are set, 525 * IO signal will be lost. Memory barrier 526 * pairs with the input side. 527 */ 528 smp_mb__after_atomic(); 529 if (sk_stream_is_writeable(sk)) 530 mask |= POLLOUT | POLLWRNORM; 531 } 532 } else 533 mask |= POLLOUT | POLLWRNORM; 534 535 if (tp->urg_data & TCP_URG_VALID) 536 mask |= POLLPRI; 537 } 538 /* This barrier is coupled with smp_wmb() in tcp_reset() */ 539 smp_rmb(); 540 if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue)) 541 mask |= POLLERR; 542 543 return mask; 544 } 545 EXPORT_SYMBOL(tcp_poll); 546 547 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg) 548 { 549 struct tcp_sock *tp = tcp_sk(sk); 550 int answ; 551 bool slow; 552 553 switch (cmd) { 554 case SIOCINQ: 555 if (sk->sk_state == TCP_LISTEN) 556 return -EINVAL; 557 558 slow = lock_sock_fast(sk); 559 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 560 answ = 0; 561 else if (sock_flag(sk, SOCK_URGINLINE) || 562 !tp->urg_data || 563 before(tp->urg_seq, tp->copied_seq) || 564 !before(tp->urg_seq, tp->rcv_nxt)) { 565 566 answ = tp->rcv_nxt - tp->copied_seq; 567 568 /* Subtract 1, if FIN was received */ 569 if (answ && sock_flag(sk, SOCK_DONE)) 570 answ--; 571 } else 572 answ = tp->urg_seq - tp->copied_seq; 573 unlock_sock_fast(sk, slow); 574 break; 575 case SIOCATMARK: 576 answ = tp->urg_data && tp->urg_seq == tp->copied_seq; 577 break; 578 case SIOCOUTQ: 579 if (sk->sk_state == TCP_LISTEN) 580 return -EINVAL; 581 582 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 583 answ = 0; 584 else 585 answ = tp->write_seq - tp->snd_una; 586 break; 587 case SIOCOUTQNSD: 588 if (sk->sk_state == TCP_LISTEN) 589 return -EINVAL; 590 591 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 592 answ = 0; 593 else 594 answ = tp->write_seq - tp->snd_nxt; 595 break; 596 default: 597 return -ENOIOCTLCMD; 598 } 599 600 return put_user(answ, (int __user *)arg); 601 } 602 EXPORT_SYMBOL(tcp_ioctl); 603 604 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb) 605 { 606 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 607 tp->pushed_seq = tp->write_seq; 608 } 609 610 static inline bool forced_push(const struct tcp_sock *tp) 611 { 612 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1)); 613 } 614 615 static void skb_entail(struct sock *sk, struct sk_buff *skb) 616 { 617 struct tcp_sock *tp = tcp_sk(sk); 618 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 619 620 skb->csum = 0; 621 tcb->seq = tcb->end_seq = tp->write_seq; 622 tcb->tcp_flags = TCPHDR_ACK; 623 tcb->sacked = 0; 624 __skb_header_release(skb); 625 tcp_add_write_queue_tail(sk, skb); 626 sk->sk_wmem_queued += skb->truesize; 627 sk_mem_charge(sk, skb->truesize); 628 if (tp->nonagle & TCP_NAGLE_PUSH) 629 tp->nonagle &= ~TCP_NAGLE_PUSH; 630 } 631 632 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags) 633 { 634 if (flags & MSG_OOB) 635 tp->snd_up = tp->write_seq; 636 } 637 638 /* If a not yet filled skb is pushed, do not send it if 639 * we have data packets in Qdisc or NIC queues : 640 * Because TX completion will happen shortly, it gives a chance 641 * to coalesce future sendmsg() payload into this skb, without 642 * need for a timer, and with no latency trade off. 643 * As packets containing data payload have a bigger truesize 644 * than pure acks (dataless) packets, the last checks prevent 645 * autocorking if we only have an ACK in Qdisc/NIC queues, 646 * or if TX completion was delayed after we processed ACK packet. 647 */ 648 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb, 649 int size_goal) 650 { 651 return skb->len < size_goal && 652 sysctl_tcp_autocorking && 653 skb != tcp_write_queue_head(sk) && 654 atomic_read(&sk->sk_wmem_alloc) > skb->truesize; 655 } 656 657 static void tcp_push(struct sock *sk, int flags, int mss_now, 658 int nonagle, int size_goal) 659 { 660 struct tcp_sock *tp = tcp_sk(sk); 661 struct sk_buff *skb; 662 663 if (!tcp_send_head(sk)) 664 return; 665 666 skb = tcp_write_queue_tail(sk); 667 if (!(flags & MSG_MORE) || forced_push(tp)) 668 tcp_mark_push(tp, skb); 669 670 tcp_mark_urg(tp, flags); 671 672 if (tcp_should_autocork(sk, skb, size_goal)) { 673 674 /* avoid atomic op if TSQ_THROTTLED bit is already set */ 675 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) { 676 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING); 677 set_bit(TSQ_THROTTLED, &tp->tsq_flags); 678 } 679 /* It is possible TX completion already happened 680 * before we set TSQ_THROTTLED. 681 */ 682 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize) 683 return; 684 } 685 686 if (flags & MSG_MORE) 687 nonagle = TCP_NAGLE_CORK; 688 689 __tcp_push_pending_frames(sk, mss_now, nonagle); 690 } 691 692 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, 693 unsigned int offset, size_t len) 694 { 695 struct tcp_splice_state *tss = rd_desc->arg.data; 696 int ret; 697 698 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe, 699 min(rd_desc->count, len), tss->flags, 700 skb_socket_splice); 701 if (ret > 0) 702 rd_desc->count -= ret; 703 return ret; 704 } 705 706 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss) 707 { 708 /* Store TCP splice context information in read_descriptor_t. */ 709 read_descriptor_t rd_desc = { 710 .arg.data = tss, 711 .count = tss->len, 712 }; 713 714 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv); 715 } 716 717 /** 718 * tcp_splice_read - splice data from TCP socket to a pipe 719 * @sock: socket to splice from 720 * @ppos: position (not valid) 721 * @pipe: pipe to splice to 722 * @len: number of bytes to splice 723 * @flags: splice modifier flags 724 * 725 * Description: 726 * Will read pages from given socket and fill them into a pipe. 727 * 728 **/ 729 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos, 730 struct pipe_inode_info *pipe, size_t len, 731 unsigned int flags) 732 { 733 struct sock *sk = sock->sk; 734 struct tcp_splice_state tss = { 735 .pipe = pipe, 736 .len = len, 737 .flags = flags, 738 }; 739 long timeo; 740 ssize_t spliced; 741 int ret; 742 743 sock_rps_record_flow(sk); 744 /* 745 * We can't seek on a socket input 746 */ 747 if (unlikely(*ppos)) 748 return -ESPIPE; 749 750 ret = spliced = 0; 751 752 lock_sock(sk); 753 754 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK); 755 while (tss.len) { 756 ret = __tcp_splice_read(sk, &tss); 757 if (ret < 0) 758 break; 759 else if (!ret) { 760 if (spliced) 761 break; 762 if (sock_flag(sk, SOCK_DONE)) 763 break; 764 if (sk->sk_err) { 765 ret = sock_error(sk); 766 break; 767 } 768 if (sk->sk_shutdown & RCV_SHUTDOWN) 769 break; 770 if (sk->sk_state == TCP_CLOSE) { 771 /* 772 * This occurs when user tries to read 773 * from never connected socket. 774 */ 775 if (!sock_flag(sk, SOCK_DONE)) 776 ret = -ENOTCONN; 777 break; 778 } 779 if (!timeo) { 780 ret = -EAGAIN; 781 break; 782 } 783 sk_wait_data(sk, &timeo); 784 if (signal_pending(current)) { 785 ret = sock_intr_errno(timeo); 786 break; 787 } 788 continue; 789 } 790 tss.len -= ret; 791 spliced += ret; 792 793 if (!timeo) 794 break; 795 release_sock(sk); 796 lock_sock(sk); 797 798 if (sk->sk_err || sk->sk_state == TCP_CLOSE || 799 (sk->sk_shutdown & RCV_SHUTDOWN) || 800 signal_pending(current)) 801 break; 802 } 803 804 release_sock(sk); 805 806 if (spliced) 807 return spliced; 808 809 return ret; 810 } 811 EXPORT_SYMBOL(tcp_splice_read); 812 813 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp, 814 bool force_schedule) 815 { 816 struct sk_buff *skb; 817 818 /* The TCP header must be at least 32-bit aligned. */ 819 size = ALIGN(size, 4); 820 821 if (unlikely(tcp_under_memory_pressure(sk))) 822 sk_mem_reclaim_partial(sk); 823 824 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp); 825 if (likely(skb)) { 826 bool mem_scheduled; 827 828 if (force_schedule) { 829 mem_scheduled = true; 830 sk_forced_mem_schedule(sk, skb->truesize); 831 } else { 832 mem_scheduled = sk_wmem_schedule(sk, skb->truesize); 833 } 834 if (likely(mem_scheduled)) { 835 skb_reserve(skb, sk->sk_prot->max_header); 836 /* 837 * Make sure that we have exactly size bytes 838 * available to the caller, no more, no less. 839 */ 840 skb->reserved_tailroom = skb->end - skb->tail - size; 841 return skb; 842 } 843 __kfree_skb(skb); 844 } else { 845 sk->sk_prot->enter_memory_pressure(sk); 846 sk_stream_moderate_sndbuf(sk); 847 } 848 return NULL; 849 } 850 851 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now, 852 int large_allowed) 853 { 854 struct tcp_sock *tp = tcp_sk(sk); 855 u32 new_size_goal, size_goal; 856 857 if (!large_allowed || !sk_can_gso(sk)) 858 return mss_now; 859 860 /* Note : tcp_tso_autosize() will eventually split this later */ 861 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER; 862 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal); 863 864 /* We try hard to avoid divides here */ 865 size_goal = tp->gso_segs * mss_now; 866 if (unlikely(new_size_goal < size_goal || 867 new_size_goal >= size_goal + mss_now)) { 868 tp->gso_segs = min_t(u16, new_size_goal / mss_now, 869 sk->sk_gso_max_segs); 870 size_goal = tp->gso_segs * mss_now; 871 } 872 873 return max(size_goal, mss_now); 874 } 875 876 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags) 877 { 878 int mss_now; 879 880 mss_now = tcp_current_mss(sk); 881 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB)); 882 883 return mss_now; 884 } 885 886 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset, 887 size_t size, int flags) 888 { 889 struct tcp_sock *tp = tcp_sk(sk); 890 int mss_now, size_goal; 891 int err; 892 ssize_t copied; 893 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 894 895 /* Wait for a connection to finish. One exception is TCP Fast Open 896 * (passive side) where data is allowed to be sent before a connection 897 * is fully established. 898 */ 899 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) && 900 !tcp_passive_fastopen(sk)) { 901 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0) 902 goto out_err; 903 } 904 905 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags); 906 907 mss_now = tcp_send_mss(sk, &size_goal, flags); 908 copied = 0; 909 910 err = -EPIPE; 911 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 912 goto out_err; 913 914 while (size > 0) { 915 struct sk_buff *skb = tcp_write_queue_tail(sk); 916 int copy, i; 917 bool can_coalesce; 918 919 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) { 920 new_segment: 921 if (!sk_stream_memory_free(sk)) 922 goto wait_for_sndbuf; 923 924 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, 925 skb_queue_empty(&sk->sk_write_queue)); 926 if (!skb) 927 goto wait_for_memory; 928 929 skb_entail(sk, skb); 930 copy = size_goal; 931 } 932 933 if (copy > size) 934 copy = size; 935 936 i = skb_shinfo(skb)->nr_frags; 937 can_coalesce = skb_can_coalesce(skb, i, page, offset); 938 if (!can_coalesce && i >= MAX_SKB_FRAGS) { 939 tcp_mark_push(tp, skb); 940 goto new_segment; 941 } 942 if (!sk_wmem_schedule(sk, copy)) 943 goto wait_for_memory; 944 945 if (can_coalesce) { 946 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 947 } else { 948 get_page(page); 949 skb_fill_page_desc(skb, i, page, offset, copy); 950 } 951 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG; 952 953 skb->len += copy; 954 skb->data_len += copy; 955 skb->truesize += copy; 956 sk->sk_wmem_queued += copy; 957 sk_mem_charge(sk, copy); 958 skb->ip_summed = CHECKSUM_PARTIAL; 959 tp->write_seq += copy; 960 TCP_SKB_CB(skb)->end_seq += copy; 961 tcp_skb_pcount_set(skb, 0); 962 963 if (!copied) 964 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 965 966 copied += copy; 967 offset += copy; 968 if (!(size -= copy)) { 969 tcp_tx_timestamp(sk, skb); 970 goto out; 971 } 972 973 if (skb->len < size_goal || (flags & MSG_OOB)) 974 continue; 975 976 if (forced_push(tp)) { 977 tcp_mark_push(tp, skb); 978 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 979 } else if (skb == tcp_send_head(sk)) 980 tcp_push_one(sk, mss_now); 981 continue; 982 983 wait_for_sndbuf: 984 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 985 wait_for_memory: 986 tcp_push(sk, flags & ~MSG_MORE, mss_now, 987 TCP_NAGLE_PUSH, size_goal); 988 989 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0) 990 goto do_error; 991 992 mss_now = tcp_send_mss(sk, &size_goal, flags); 993 } 994 995 out: 996 if (copied && !(flags & MSG_SENDPAGE_NOTLAST)) 997 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal); 998 return copied; 999 1000 do_error: 1001 if (copied) 1002 goto out; 1003 out_err: 1004 /* make sure we wake any epoll edge trigger waiter */ 1005 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN)) 1006 sk->sk_write_space(sk); 1007 return sk_stream_error(sk, flags, err); 1008 } 1009 1010 int tcp_sendpage(struct sock *sk, struct page *page, int offset, 1011 size_t size, int flags) 1012 { 1013 ssize_t res; 1014 1015 if (!(sk->sk_route_caps & NETIF_F_SG) || 1016 !(sk->sk_route_caps & NETIF_F_ALL_CSUM)) 1017 return sock_no_sendpage(sk->sk_socket, page, offset, size, 1018 flags); 1019 1020 lock_sock(sk); 1021 res = do_tcp_sendpages(sk, page, offset, size, flags); 1022 release_sock(sk); 1023 return res; 1024 } 1025 EXPORT_SYMBOL(tcp_sendpage); 1026 1027 static inline int select_size(const struct sock *sk, bool sg) 1028 { 1029 const struct tcp_sock *tp = tcp_sk(sk); 1030 int tmp = tp->mss_cache; 1031 1032 if (sg) { 1033 if (sk_can_gso(sk)) { 1034 /* Small frames wont use a full page: 1035 * Payload will immediately follow tcp header. 1036 */ 1037 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER); 1038 } else { 1039 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER); 1040 1041 if (tmp >= pgbreak && 1042 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE) 1043 tmp = pgbreak; 1044 } 1045 } 1046 1047 return tmp; 1048 } 1049 1050 void tcp_free_fastopen_req(struct tcp_sock *tp) 1051 { 1052 if (tp->fastopen_req) { 1053 kfree(tp->fastopen_req); 1054 tp->fastopen_req = NULL; 1055 } 1056 } 1057 1058 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, 1059 int *copied, size_t size) 1060 { 1061 struct tcp_sock *tp = tcp_sk(sk); 1062 int err, flags; 1063 1064 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE)) 1065 return -EOPNOTSUPP; 1066 if (tp->fastopen_req) 1067 return -EALREADY; /* Another Fast Open is in progress */ 1068 1069 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request), 1070 sk->sk_allocation); 1071 if (unlikely(!tp->fastopen_req)) 1072 return -ENOBUFS; 1073 tp->fastopen_req->data = msg; 1074 tp->fastopen_req->size = size; 1075 1076 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0; 1077 err = __inet_stream_connect(sk->sk_socket, msg->msg_name, 1078 msg->msg_namelen, flags); 1079 *copied = tp->fastopen_req->copied; 1080 tcp_free_fastopen_req(tp); 1081 return err; 1082 } 1083 1084 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size) 1085 { 1086 struct tcp_sock *tp = tcp_sk(sk); 1087 struct sk_buff *skb; 1088 int flags, err, copied = 0; 1089 int mss_now = 0, size_goal, copied_syn = 0; 1090 bool sg; 1091 long timeo; 1092 1093 lock_sock(sk); 1094 1095 flags = msg->msg_flags; 1096 if (flags & MSG_FASTOPEN) { 1097 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size); 1098 if (err == -EINPROGRESS && copied_syn > 0) 1099 goto out; 1100 else if (err) 1101 goto out_err; 1102 } 1103 1104 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 1105 1106 /* Wait for a connection to finish. One exception is TCP Fast Open 1107 * (passive side) where data is allowed to be sent before a connection 1108 * is fully established. 1109 */ 1110 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) && 1111 !tcp_passive_fastopen(sk)) { 1112 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0) 1113 goto do_error; 1114 } 1115 1116 if (unlikely(tp->repair)) { 1117 if (tp->repair_queue == TCP_RECV_QUEUE) { 1118 copied = tcp_send_rcvq(sk, msg, size); 1119 goto out_nopush; 1120 } 1121 1122 err = -EINVAL; 1123 if (tp->repair_queue == TCP_NO_QUEUE) 1124 goto out_err; 1125 1126 /* 'common' sending to sendq */ 1127 } 1128 1129 /* This should be in poll */ 1130 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags); 1131 1132 mss_now = tcp_send_mss(sk, &size_goal, flags); 1133 1134 /* Ok commence sending. */ 1135 copied = 0; 1136 1137 err = -EPIPE; 1138 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 1139 goto out_err; 1140 1141 sg = !!(sk->sk_route_caps & NETIF_F_SG); 1142 1143 while (msg_data_left(msg)) { 1144 int copy = 0; 1145 int max = size_goal; 1146 1147 skb = tcp_write_queue_tail(sk); 1148 if (tcp_send_head(sk)) { 1149 if (skb->ip_summed == CHECKSUM_NONE) 1150 max = mss_now; 1151 copy = max - skb->len; 1152 } 1153 1154 if (copy <= 0) { 1155 new_segment: 1156 /* Allocate new segment. If the interface is SG, 1157 * allocate skb fitting to single page. 1158 */ 1159 if (!sk_stream_memory_free(sk)) 1160 goto wait_for_sndbuf; 1161 1162 skb = sk_stream_alloc_skb(sk, 1163 select_size(sk, sg), 1164 sk->sk_allocation, 1165 skb_queue_empty(&sk->sk_write_queue)); 1166 if (!skb) 1167 goto wait_for_memory; 1168 1169 /* 1170 * Check whether we can use HW checksum. 1171 */ 1172 if (sk->sk_route_caps & NETIF_F_ALL_CSUM) 1173 skb->ip_summed = CHECKSUM_PARTIAL; 1174 1175 skb_entail(sk, skb); 1176 copy = size_goal; 1177 max = size_goal; 1178 1179 /* All packets are restored as if they have 1180 * already been sent. skb_mstamp isn't set to 1181 * avoid wrong rtt estimation. 1182 */ 1183 if (tp->repair) 1184 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED; 1185 } 1186 1187 /* Try to append data to the end of skb. */ 1188 if (copy > msg_data_left(msg)) 1189 copy = msg_data_left(msg); 1190 1191 /* Where to copy to? */ 1192 if (skb_availroom(skb) > 0) { 1193 /* We have some space in skb head. Superb! */ 1194 copy = min_t(int, copy, skb_availroom(skb)); 1195 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy); 1196 if (err) 1197 goto do_fault; 1198 } else { 1199 bool merge = true; 1200 int i = skb_shinfo(skb)->nr_frags; 1201 struct page_frag *pfrag = sk_page_frag(sk); 1202 1203 if (!sk_page_frag_refill(sk, pfrag)) 1204 goto wait_for_memory; 1205 1206 if (!skb_can_coalesce(skb, i, pfrag->page, 1207 pfrag->offset)) { 1208 if (i == MAX_SKB_FRAGS || !sg) { 1209 tcp_mark_push(tp, skb); 1210 goto new_segment; 1211 } 1212 merge = false; 1213 } 1214 1215 copy = min_t(int, copy, pfrag->size - pfrag->offset); 1216 1217 if (!sk_wmem_schedule(sk, copy)) 1218 goto wait_for_memory; 1219 1220 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb, 1221 pfrag->page, 1222 pfrag->offset, 1223 copy); 1224 if (err) 1225 goto do_error; 1226 1227 /* Update the skb. */ 1228 if (merge) { 1229 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 1230 } else { 1231 skb_fill_page_desc(skb, i, pfrag->page, 1232 pfrag->offset, copy); 1233 get_page(pfrag->page); 1234 } 1235 pfrag->offset += copy; 1236 } 1237 1238 if (!copied) 1239 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 1240 1241 tp->write_seq += copy; 1242 TCP_SKB_CB(skb)->end_seq += copy; 1243 tcp_skb_pcount_set(skb, 0); 1244 1245 copied += copy; 1246 if (!msg_data_left(msg)) { 1247 tcp_tx_timestamp(sk, skb); 1248 goto out; 1249 } 1250 1251 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair)) 1252 continue; 1253 1254 if (forced_push(tp)) { 1255 tcp_mark_push(tp, skb); 1256 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 1257 } else if (skb == tcp_send_head(sk)) 1258 tcp_push_one(sk, mss_now); 1259 continue; 1260 1261 wait_for_sndbuf: 1262 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1263 wait_for_memory: 1264 if (copied) 1265 tcp_push(sk, flags & ~MSG_MORE, mss_now, 1266 TCP_NAGLE_PUSH, size_goal); 1267 1268 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0) 1269 goto do_error; 1270 1271 mss_now = tcp_send_mss(sk, &size_goal, flags); 1272 } 1273 1274 out: 1275 if (copied) 1276 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal); 1277 out_nopush: 1278 release_sock(sk); 1279 return copied + copied_syn; 1280 1281 do_fault: 1282 if (!skb->len) { 1283 tcp_unlink_write_queue(skb, sk); 1284 /* It is the one place in all of TCP, except connection 1285 * reset, where we can be unlinking the send_head. 1286 */ 1287 tcp_check_send_head(sk, skb); 1288 sk_wmem_free_skb(sk, skb); 1289 } 1290 1291 do_error: 1292 if (copied + copied_syn) 1293 goto out; 1294 out_err: 1295 err = sk_stream_error(sk, flags, err); 1296 /* make sure we wake any epoll edge trigger waiter */ 1297 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN)) 1298 sk->sk_write_space(sk); 1299 release_sock(sk); 1300 return err; 1301 } 1302 EXPORT_SYMBOL(tcp_sendmsg); 1303 1304 /* 1305 * Handle reading urgent data. BSD has very simple semantics for 1306 * this, no blocking and very strange errors 8) 1307 */ 1308 1309 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags) 1310 { 1311 struct tcp_sock *tp = tcp_sk(sk); 1312 1313 /* No URG data to read. */ 1314 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data || 1315 tp->urg_data == TCP_URG_READ) 1316 return -EINVAL; /* Yes this is right ! */ 1317 1318 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE)) 1319 return -ENOTCONN; 1320 1321 if (tp->urg_data & TCP_URG_VALID) { 1322 int err = 0; 1323 char c = tp->urg_data; 1324 1325 if (!(flags & MSG_PEEK)) 1326 tp->urg_data = TCP_URG_READ; 1327 1328 /* Read urgent data. */ 1329 msg->msg_flags |= MSG_OOB; 1330 1331 if (len > 0) { 1332 if (!(flags & MSG_TRUNC)) 1333 err = memcpy_to_msg(msg, &c, 1); 1334 len = 1; 1335 } else 1336 msg->msg_flags |= MSG_TRUNC; 1337 1338 return err ? -EFAULT : len; 1339 } 1340 1341 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 1342 return 0; 1343 1344 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and 1345 * the available implementations agree in this case: 1346 * this call should never block, independent of the 1347 * blocking state of the socket. 1348 * Mike <pall@rz.uni-karlsruhe.de> 1349 */ 1350 return -EAGAIN; 1351 } 1352 1353 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len) 1354 { 1355 struct sk_buff *skb; 1356 int copied = 0, err = 0; 1357 1358 /* XXX -- need to support SO_PEEK_OFF */ 1359 1360 skb_queue_walk(&sk->sk_write_queue, skb) { 1361 err = skb_copy_datagram_msg(skb, 0, msg, skb->len); 1362 if (err) 1363 break; 1364 1365 copied += skb->len; 1366 } 1367 1368 return err ?: copied; 1369 } 1370 1371 /* Clean up the receive buffer for full frames taken by the user, 1372 * then send an ACK if necessary. COPIED is the number of bytes 1373 * tcp_recvmsg has given to the user so far, it speeds up the 1374 * calculation of whether or not we must ACK for the sake of 1375 * a window update. 1376 */ 1377 static void tcp_cleanup_rbuf(struct sock *sk, int copied) 1378 { 1379 struct tcp_sock *tp = tcp_sk(sk); 1380 bool time_to_ack = false; 1381 1382 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 1383 1384 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq), 1385 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n", 1386 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt); 1387 1388 if (inet_csk_ack_scheduled(sk)) { 1389 const struct inet_connection_sock *icsk = inet_csk(sk); 1390 /* Delayed ACKs frequently hit locked sockets during bulk 1391 * receive. */ 1392 if (icsk->icsk_ack.blocked || 1393 /* Once-per-two-segments ACK was not sent by tcp_input.c */ 1394 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss || 1395 /* 1396 * If this read emptied read buffer, we send ACK, if 1397 * connection is not bidirectional, user drained 1398 * receive buffer and there was a small segment 1399 * in queue. 1400 */ 1401 (copied > 0 && 1402 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) || 1403 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) && 1404 !icsk->icsk_ack.pingpong)) && 1405 !atomic_read(&sk->sk_rmem_alloc))) 1406 time_to_ack = true; 1407 } 1408 1409 /* We send an ACK if we can now advertise a non-zero window 1410 * which has been raised "significantly". 1411 * 1412 * Even if window raised up to infinity, do not send window open ACK 1413 * in states, where we will not receive more. It is useless. 1414 */ 1415 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) { 1416 __u32 rcv_window_now = tcp_receive_window(tp); 1417 1418 /* Optimize, __tcp_select_window() is not cheap. */ 1419 if (2*rcv_window_now <= tp->window_clamp) { 1420 __u32 new_window = __tcp_select_window(sk); 1421 1422 /* Send ACK now, if this read freed lots of space 1423 * in our buffer. Certainly, new_window is new window. 1424 * We can advertise it now, if it is not less than current one. 1425 * "Lots" means "at least twice" here. 1426 */ 1427 if (new_window && new_window >= 2 * rcv_window_now) 1428 time_to_ack = true; 1429 } 1430 } 1431 if (time_to_ack) 1432 tcp_send_ack(sk); 1433 } 1434 1435 static void tcp_prequeue_process(struct sock *sk) 1436 { 1437 struct sk_buff *skb; 1438 struct tcp_sock *tp = tcp_sk(sk); 1439 1440 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED); 1441 1442 /* RX process wants to run with disabled BHs, though it is not 1443 * necessary */ 1444 local_bh_disable(); 1445 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL) 1446 sk_backlog_rcv(sk, skb); 1447 local_bh_enable(); 1448 1449 /* Clear memory counter. */ 1450 tp->ucopy.memory = 0; 1451 } 1452 1453 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off) 1454 { 1455 struct sk_buff *skb; 1456 u32 offset; 1457 1458 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) { 1459 offset = seq - TCP_SKB_CB(skb)->seq; 1460 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 1461 offset--; 1462 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) { 1463 *off = offset; 1464 return skb; 1465 } 1466 /* This looks weird, but this can happen if TCP collapsing 1467 * splitted a fat GRO packet, while we released socket lock 1468 * in skb_splice_bits() 1469 */ 1470 sk_eat_skb(sk, skb); 1471 } 1472 return NULL; 1473 } 1474 1475 /* 1476 * This routine provides an alternative to tcp_recvmsg() for routines 1477 * that would like to handle copying from skbuffs directly in 'sendfile' 1478 * fashion. 1479 * Note: 1480 * - It is assumed that the socket was locked by the caller. 1481 * - The routine does not block. 1482 * - At present, there is no support for reading OOB data 1483 * or for 'peeking' the socket using this routine 1484 * (although both would be easy to implement). 1485 */ 1486 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, 1487 sk_read_actor_t recv_actor) 1488 { 1489 struct sk_buff *skb; 1490 struct tcp_sock *tp = tcp_sk(sk); 1491 u32 seq = tp->copied_seq; 1492 u32 offset; 1493 int copied = 0; 1494 1495 if (sk->sk_state == TCP_LISTEN) 1496 return -ENOTCONN; 1497 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) { 1498 if (offset < skb->len) { 1499 int used; 1500 size_t len; 1501 1502 len = skb->len - offset; 1503 /* Stop reading if we hit a patch of urgent data */ 1504 if (tp->urg_data) { 1505 u32 urg_offset = tp->urg_seq - seq; 1506 if (urg_offset < len) 1507 len = urg_offset; 1508 if (!len) 1509 break; 1510 } 1511 used = recv_actor(desc, skb, offset, len); 1512 if (used <= 0) { 1513 if (!copied) 1514 copied = used; 1515 break; 1516 } else if (used <= len) { 1517 seq += used; 1518 copied += used; 1519 offset += used; 1520 } 1521 /* If recv_actor drops the lock (e.g. TCP splice 1522 * receive) the skb pointer might be invalid when 1523 * getting here: tcp_collapse might have deleted it 1524 * while aggregating skbs from the socket queue. 1525 */ 1526 skb = tcp_recv_skb(sk, seq - 1, &offset); 1527 if (!skb) 1528 break; 1529 /* TCP coalescing might have appended data to the skb. 1530 * Try to splice more frags 1531 */ 1532 if (offset + 1 != skb->len) 1533 continue; 1534 } 1535 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) { 1536 sk_eat_skb(sk, skb); 1537 ++seq; 1538 break; 1539 } 1540 sk_eat_skb(sk, skb); 1541 if (!desc->count) 1542 break; 1543 tp->copied_seq = seq; 1544 } 1545 tp->copied_seq = seq; 1546 1547 tcp_rcv_space_adjust(sk); 1548 1549 /* Clean up data we have read: This will do ACK frames. */ 1550 if (copied > 0) { 1551 tcp_recv_skb(sk, seq, &offset); 1552 tcp_cleanup_rbuf(sk, copied); 1553 } 1554 return copied; 1555 } 1556 EXPORT_SYMBOL(tcp_read_sock); 1557 1558 /* 1559 * This routine copies from a sock struct into the user buffer. 1560 * 1561 * Technical note: in 2.3 we work on _locked_ socket, so that 1562 * tricks with *seq access order and skb->users are not required. 1563 * Probably, code can be easily improved even more. 1564 */ 1565 1566 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock, 1567 int flags, int *addr_len) 1568 { 1569 struct tcp_sock *tp = tcp_sk(sk); 1570 int copied = 0; 1571 u32 peek_seq; 1572 u32 *seq; 1573 unsigned long used; 1574 int err; 1575 int target; /* Read at least this many bytes */ 1576 long timeo; 1577 struct task_struct *user_recv = NULL; 1578 struct sk_buff *skb; 1579 u32 urg_hole = 0; 1580 1581 if (unlikely(flags & MSG_ERRQUEUE)) 1582 return inet_recv_error(sk, msg, len, addr_len); 1583 1584 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) && 1585 (sk->sk_state == TCP_ESTABLISHED)) 1586 sk_busy_loop(sk, nonblock); 1587 1588 lock_sock(sk); 1589 1590 err = -ENOTCONN; 1591 if (sk->sk_state == TCP_LISTEN) 1592 goto out; 1593 1594 timeo = sock_rcvtimeo(sk, nonblock); 1595 1596 /* Urgent data needs to be handled specially. */ 1597 if (flags & MSG_OOB) 1598 goto recv_urg; 1599 1600 if (unlikely(tp->repair)) { 1601 err = -EPERM; 1602 if (!(flags & MSG_PEEK)) 1603 goto out; 1604 1605 if (tp->repair_queue == TCP_SEND_QUEUE) 1606 goto recv_sndq; 1607 1608 err = -EINVAL; 1609 if (tp->repair_queue == TCP_NO_QUEUE) 1610 goto out; 1611 1612 /* 'common' recv queue MSG_PEEK-ing */ 1613 } 1614 1615 seq = &tp->copied_seq; 1616 if (flags & MSG_PEEK) { 1617 peek_seq = tp->copied_seq; 1618 seq = &peek_seq; 1619 } 1620 1621 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 1622 1623 do { 1624 u32 offset; 1625 1626 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */ 1627 if (tp->urg_data && tp->urg_seq == *seq) { 1628 if (copied) 1629 break; 1630 if (signal_pending(current)) { 1631 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN; 1632 break; 1633 } 1634 } 1635 1636 /* Next get a buffer. */ 1637 1638 skb_queue_walk(&sk->sk_receive_queue, skb) { 1639 /* Now that we have two receive queues this 1640 * shouldn't happen. 1641 */ 1642 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq), 1643 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n", 1644 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, 1645 flags)) 1646 break; 1647 1648 offset = *seq - TCP_SKB_CB(skb)->seq; 1649 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 1650 offset--; 1651 if (offset < skb->len) 1652 goto found_ok_skb; 1653 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1654 goto found_fin_ok; 1655 WARN(!(flags & MSG_PEEK), 1656 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n", 1657 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags); 1658 } 1659 1660 /* Well, if we have backlog, try to process it now yet. */ 1661 1662 if (copied >= target && !sk->sk_backlog.tail) 1663 break; 1664 1665 if (copied) { 1666 if (sk->sk_err || 1667 sk->sk_state == TCP_CLOSE || 1668 (sk->sk_shutdown & RCV_SHUTDOWN) || 1669 !timeo || 1670 signal_pending(current)) 1671 break; 1672 } else { 1673 if (sock_flag(sk, SOCK_DONE)) 1674 break; 1675 1676 if (sk->sk_err) { 1677 copied = sock_error(sk); 1678 break; 1679 } 1680 1681 if (sk->sk_shutdown & RCV_SHUTDOWN) 1682 break; 1683 1684 if (sk->sk_state == TCP_CLOSE) { 1685 if (!sock_flag(sk, SOCK_DONE)) { 1686 /* This occurs when user tries to read 1687 * from never connected socket. 1688 */ 1689 copied = -ENOTCONN; 1690 break; 1691 } 1692 break; 1693 } 1694 1695 if (!timeo) { 1696 copied = -EAGAIN; 1697 break; 1698 } 1699 1700 if (signal_pending(current)) { 1701 copied = sock_intr_errno(timeo); 1702 break; 1703 } 1704 } 1705 1706 tcp_cleanup_rbuf(sk, copied); 1707 1708 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) { 1709 /* Install new reader */ 1710 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) { 1711 user_recv = current; 1712 tp->ucopy.task = user_recv; 1713 tp->ucopy.msg = msg; 1714 } 1715 1716 tp->ucopy.len = len; 1717 1718 WARN_ON(tp->copied_seq != tp->rcv_nxt && 1719 !(flags & (MSG_PEEK | MSG_TRUNC))); 1720 1721 /* Ugly... If prequeue is not empty, we have to 1722 * process it before releasing socket, otherwise 1723 * order will be broken at second iteration. 1724 * More elegant solution is required!!! 1725 * 1726 * Look: we have the following (pseudo)queues: 1727 * 1728 * 1. packets in flight 1729 * 2. backlog 1730 * 3. prequeue 1731 * 4. receive_queue 1732 * 1733 * Each queue can be processed only if the next ones 1734 * are empty. At this point we have empty receive_queue. 1735 * But prequeue _can_ be not empty after 2nd iteration, 1736 * when we jumped to start of loop because backlog 1737 * processing added something to receive_queue. 1738 * We cannot release_sock(), because backlog contains 1739 * packets arrived _after_ prequeued ones. 1740 * 1741 * Shortly, algorithm is clear --- to process all 1742 * the queues in order. We could make it more directly, 1743 * requeueing packets from backlog to prequeue, if 1744 * is not empty. It is more elegant, but eats cycles, 1745 * unfortunately. 1746 */ 1747 if (!skb_queue_empty(&tp->ucopy.prequeue)) 1748 goto do_prequeue; 1749 1750 /* __ Set realtime policy in scheduler __ */ 1751 } 1752 1753 if (copied >= target) { 1754 /* Do not sleep, just process backlog. */ 1755 release_sock(sk); 1756 lock_sock(sk); 1757 } else 1758 sk_wait_data(sk, &timeo); 1759 1760 if (user_recv) { 1761 int chunk; 1762 1763 /* __ Restore normal policy in scheduler __ */ 1764 1765 if ((chunk = len - tp->ucopy.len) != 0) { 1766 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk); 1767 len -= chunk; 1768 copied += chunk; 1769 } 1770 1771 if (tp->rcv_nxt == tp->copied_seq && 1772 !skb_queue_empty(&tp->ucopy.prequeue)) { 1773 do_prequeue: 1774 tcp_prequeue_process(sk); 1775 1776 if ((chunk = len - tp->ucopy.len) != 0) { 1777 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1778 len -= chunk; 1779 copied += chunk; 1780 } 1781 } 1782 } 1783 if ((flags & MSG_PEEK) && 1784 (peek_seq - copied - urg_hole != tp->copied_seq)) { 1785 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n", 1786 current->comm, 1787 task_pid_nr(current)); 1788 peek_seq = tp->copied_seq; 1789 } 1790 continue; 1791 1792 found_ok_skb: 1793 /* Ok so how much can we use? */ 1794 used = skb->len - offset; 1795 if (len < used) 1796 used = len; 1797 1798 /* Do we have urgent data here? */ 1799 if (tp->urg_data) { 1800 u32 urg_offset = tp->urg_seq - *seq; 1801 if (urg_offset < used) { 1802 if (!urg_offset) { 1803 if (!sock_flag(sk, SOCK_URGINLINE)) { 1804 ++*seq; 1805 urg_hole++; 1806 offset++; 1807 used--; 1808 if (!used) 1809 goto skip_copy; 1810 } 1811 } else 1812 used = urg_offset; 1813 } 1814 } 1815 1816 if (!(flags & MSG_TRUNC)) { 1817 err = skb_copy_datagram_msg(skb, offset, msg, used); 1818 if (err) { 1819 /* Exception. Bailout! */ 1820 if (!copied) 1821 copied = -EFAULT; 1822 break; 1823 } 1824 } 1825 1826 *seq += used; 1827 copied += used; 1828 len -= used; 1829 1830 tcp_rcv_space_adjust(sk); 1831 1832 skip_copy: 1833 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) { 1834 tp->urg_data = 0; 1835 tcp_fast_path_check(sk); 1836 } 1837 if (used + offset < skb->len) 1838 continue; 1839 1840 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1841 goto found_fin_ok; 1842 if (!(flags & MSG_PEEK)) 1843 sk_eat_skb(sk, skb); 1844 continue; 1845 1846 found_fin_ok: 1847 /* Process the FIN. */ 1848 ++*seq; 1849 if (!(flags & MSG_PEEK)) 1850 sk_eat_skb(sk, skb); 1851 break; 1852 } while (len > 0); 1853 1854 if (user_recv) { 1855 if (!skb_queue_empty(&tp->ucopy.prequeue)) { 1856 int chunk; 1857 1858 tp->ucopy.len = copied > 0 ? len : 0; 1859 1860 tcp_prequeue_process(sk); 1861 1862 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) { 1863 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1864 len -= chunk; 1865 copied += chunk; 1866 } 1867 } 1868 1869 tp->ucopy.task = NULL; 1870 tp->ucopy.len = 0; 1871 } 1872 1873 /* According to UNIX98, msg_name/msg_namelen are ignored 1874 * on connected socket. I was just happy when found this 8) --ANK 1875 */ 1876 1877 /* Clean up data we have read: This will do ACK frames. */ 1878 tcp_cleanup_rbuf(sk, copied); 1879 1880 release_sock(sk); 1881 return copied; 1882 1883 out: 1884 release_sock(sk); 1885 return err; 1886 1887 recv_urg: 1888 err = tcp_recv_urg(sk, msg, len, flags); 1889 goto out; 1890 1891 recv_sndq: 1892 err = tcp_peek_sndq(sk, msg, len); 1893 goto out; 1894 } 1895 EXPORT_SYMBOL(tcp_recvmsg); 1896 1897 void tcp_set_state(struct sock *sk, int state) 1898 { 1899 int oldstate = sk->sk_state; 1900 1901 switch (state) { 1902 case TCP_ESTABLISHED: 1903 if (oldstate != TCP_ESTABLISHED) 1904 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1905 break; 1906 1907 case TCP_CLOSE: 1908 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED) 1909 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS); 1910 1911 sk->sk_prot->unhash(sk); 1912 if (inet_csk(sk)->icsk_bind_hash && 1913 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK)) 1914 inet_put_port(sk); 1915 /* fall through */ 1916 default: 1917 if (oldstate == TCP_ESTABLISHED) 1918 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1919 } 1920 1921 /* Change state AFTER socket is unhashed to avoid closed 1922 * socket sitting in hash tables. 1923 */ 1924 sk->sk_state = state; 1925 1926 #ifdef STATE_TRACE 1927 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]); 1928 #endif 1929 } 1930 EXPORT_SYMBOL_GPL(tcp_set_state); 1931 1932 /* 1933 * State processing on a close. This implements the state shift for 1934 * sending our FIN frame. Note that we only send a FIN for some 1935 * states. A shutdown() may have already sent the FIN, or we may be 1936 * closed. 1937 */ 1938 1939 static const unsigned char new_state[16] = { 1940 /* current state: new state: action: */ 1941 [0 /* (Invalid) */] = TCP_CLOSE, 1942 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1943 [TCP_SYN_SENT] = TCP_CLOSE, 1944 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1945 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1, 1946 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2, 1947 [TCP_TIME_WAIT] = TCP_CLOSE, 1948 [TCP_CLOSE] = TCP_CLOSE, 1949 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN, 1950 [TCP_LAST_ACK] = TCP_LAST_ACK, 1951 [TCP_LISTEN] = TCP_CLOSE, 1952 [TCP_CLOSING] = TCP_CLOSING, 1953 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */ 1954 }; 1955 1956 static int tcp_close_state(struct sock *sk) 1957 { 1958 int next = (int)new_state[sk->sk_state]; 1959 int ns = next & TCP_STATE_MASK; 1960 1961 tcp_set_state(sk, ns); 1962 1963 return next & TCP_ACTION_FIN; 1964 } 1965 1966 /* 1967 * Shutdown the sending side of a connection. Much like close except 1968 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD). 1969 */ 1970 1971 void tcp_shutdown(struct sock *sk, int how) 1972 { 1973 /* We need to grab some memory, and put together a FIN, 1974 * and then put it into the queue to be sent. 1975 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92. 1976 */ 1977 if (!(how & SEND_SHUTDOWN)) 1978 return; 1979 1980 /* If we've already sent a FIN, or it's a closed state, skip this. */ 1981 if ((1 << sk->sk_state) & 1982 (TCPF_ESTABLISHED | TCPF_SYN_SENT | 1983 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) { 1984 /* Clear out any half completed packets. FIN if needed. */ 1985 if (tcp_close_state(sk)) 1986 tcp_send_fin(sk); 1987 } 1988 } 1989 EXPORT_SYMBOL(tcp_shutdown); 1990 1991 bool tcp_check_oom(struct sock *sk, int shift) 1992 { 1993 bool too_many_orphans, out_of_socket_memory; 1994 1995 too_many_orphans = tcp_too_many_orphans(sk, shift); 1996 out_of_socket_memory = tcp_out_of_memory(sk); 1997 1998 if (too_many_orphans) 1999 net_info_ratelimited("too many orphaned sockets\n"); 2000 if (out_of_socket_memory) 2001 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n"); 2002 return too_many_orphans || out_of_socket_memory; 2003 } 2004 2005 void tcp_close(struct sock *sk, long timeout) 2006 { 2007 struct sk_buff *skb; 2008 int data_was_unread = 0; 2009 int state; 2010 2011 lock_sock(sk); 2012 sk->sk_shutdown = SHUTDOWN_MASK; 2013 2014 if (sk->sk_state == TCP_LISTEN) { 2015 tcp_set_state(sk, TCP_CLOSE); 2016 2017 /* Special case. */ 2018 inet_csk_listen_stop(sk); 2019 2020 goto adjudge_to_death; 2021 } 2022 2023 /* We need to flush the recv. buffs. We do this only on the 2024 * descriptor close, not protocol-sourced closes, because the 2025 * reader process may not have drained the data yet! 2026 */ 2027 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) { 2028 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq; 2029 2030 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 2031 len--; 2032 data_was_unread += len; 2033 __kfree_skb(skb); 2034 } 2035 2036 sk_mem_reclaim(sk); 2037 2038 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */ 2039 if (sk->sk_state == TCP_CLOSE) 2040 goto adjudge_to_death; 2041 2042 /* As outlined in RFC 2525, section 2.17, we send a RST here because 2043 * data was lost. To witness the awful effects of the old behavior of 2044 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk 2045 * GET in an FTP client, suspend the process, wait for the client to 2046 * advertise a zero window, then kill -9 the FTP client, wheee... 2047 * Note: timeout is always zero in such a case. 2048 */ 2049 if (unlikely(tcp_sk(sk)->repair)) { 2050 sk->sk_prot->disconnect(sk, 0); 2051 } else if (data_was_unread) { 2052 /* Unread data was tossed, zap the connection. */ 2053 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE); 2054 tcp_set_state(sk, TCP_CLOSE); 2055 tcp_send_active_reset(sk, sk->sk_allocation); 2056 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) { 2057 /* Check zero linger _after_ checking for unread data. */ 2058 sk->sk_prot->disconnect(sk, 0); 2059 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 2060 } else if (tcp_close_state(sk)) { 2061 /* We FIN if the application ate all the data before 2062 * zapping the connection. 2063 */ 2064 2065 /* RED-PEN. Formally speaking, we have broken TCP state 2066 * machine. State transitions: 2067 * 2068 * TCP_ESTABLISHED -> TCP_FIN_WAIT1 2069 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible) 2070 * TCP_CLOSE_WAIT -> TCP_LAST_ACK 2071 * 2072 * are legal only when FIN has been sent (i.e. in window), 2073 * rather than queued out of window. Purists blame. 2074 * 2075 * F.e. "RFC state" is ESTABLISHED, 2076 * if Linux state is FIN-WAIT-1, but FIN is still not sent. 2077 * 2078 * The visible declinations are that sometimes 2079 * we enter time-wait state, when it is not required really 2080 * (harmless), do not send active resets, when they are 2081 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when 2082 * they look as CLOSING or LAST_ACK for Linux) 2083 * Probably, I missed some more holelets. 2084 * --ANK 2085 * XXX (TFO) - To start off we don't support SYN+ACK+FIN 2086 * in a single packet! (May consider it later but will 2087 * probably need API support or TCP_CORK SYN-ACK until 2088 * data is written and socket is closed.) 2089 */ 2090 tcp_send_fin(sk); 2091 } 2092 2093 sk_stream_wait_close(sk, timeout); 2094 2095 adjudge_to_death: 2096 state = sk->sk_state; 2097 sock_hold(sk); 2098 sock_orphan(sk); 2099 2100 /* It is the last release_sock in its life. It will remove backlog. */ 2101 release_sock(sk); 2102 2103 2104 /* Now socket is owned by kernel and we acquire BH lock 2105 to finish close. No need to check for user refs. 2106 */ 2107 local_bh_disable(); 2108 bh_lock_sock(sk); 2109 WARN_ON(sock_owned_by_user(sk)); 2110 2111 percpu_counter_inc(sk->sk_prot->orphan_count); 2112 2113 /* Have we already been destroyed by a softirq or backlog? */ 2114 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE) 2115 goto out; 2116 2117 /* This is a (useful) BSD violating of the RFC. There is a 2118 * problem with TCP as specified in that the other end could 2119 * keep a socket open forever with no application left this end. 2120 * We use a 1 minute timeout (about the same as BSD) then kill 2121 * our end. If they send after that then tough - BUT: long enough 2122 * that we won't make the old 4*rto = almost no time - whoops 2123 * reset mistake. 2124 * 2125 * Nope, it was not mistake. It is really desired behaviour 2126 * f.e. on http servers, when such sockets are useless, but 2127 * consume significant resources. Let's do it with special 2128 * linger2 option. --ANK 2129 */ 2130 2131 if (sk->sk_state == TCP_FIN_WAIT2) { 2132 struct tcp_sock *tp = tcp_sk(sk); 2133 if (tp->linger2 < 0) { 2134 tcp_set_state(sk, TCP_CLOSE); 2135 tcp_send_active_reset(sk, GFP_ATOMIC); 2136 NET_INC_STATS_BH(sock_net(sk), 2137 LINUX_MIB_TCPABORTONLINGER); 2138 } else { 2139 const int tmo = tcp_fin_time(sk); 2140 2141 if (tmo > TCP_TIMEWAIT_LEN) { 2142 inet_csk_reset_keepalive_timer(sk, 2143 tmo - TCP_TIMEWAIT_LEN); 2144 } else { 2145 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 2146 goto out; 2147 } 2148 } 2149 } 2150 if (sk->sk_state != TCP_CLOSE) { 2151 sk_mem_reclaim(sk); 2152 if (tcp_check_oom(sk, 0)) { 2153 tcp_set_state(sk, TCP_CLOSE); 2154 tcp_send_active_reset(sk, GFP_ATOMIC); 2155 NET_INC_STATS_BH(sock_net(sk), 2156 LINUX_MIB_TCPABORTONMEMORY); 2157 } 2158 } 2159 2160 if (sk->sk_state == TCP_CLOSE) { 2161 struct request_sock *req = tcp_sk(sk)->fastopen_rsk; 2162 /* We could get here with a non-NULL req if the socket is 2163 * aborted (e.g., closed with unread data) before 3WHS 2164 * finishes. 2165 */ 2166 if (req) 2167 reqsk_fastopen_remove(sk, req, false); 2168 inet_csk_destroy_sock(sk); 2169 } 2170 /* Otherwise, socket is reprieved until protocol close. */ 2171 2172 out: 2173 bh_unlock_sock(sk); 2174 local_bh_enable(); 2175 sock_put(sk); 2176 } 2177 EXPORT_SYMBOL(tcp_close); 2178 2179 /* These states need RST on ABORT according to RFC793 */ 2180 2181 static inline bool tcp_need_reset(int state) 2182 { 2183 return (1 << state) & 2184 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 | 2185 TCPF_FIN_WAIT2 | TCPF_SYN_RECV); 2186 } 2187 2188 int tcp_disconnect(struct sock *sk, int flags) 2189 { 2190 struct inet_sock *inet = inet_sk(sk); 2191 struct inet_connection_sock *icsk = inet_csk(sk); 2192 struct tcp_sock *tp = tcp_sk(sk); 2193 int err = 0; 2194 int old_state = sk->sk_state; 2195 2196 if (old_state != TCP_CLOSE) 2197 tcp_set_state(sk, TCP_CLOSE); 2198 2199 /* ABORT function of RFC793 */ 2200 if (old_state == TCP_LISTEN) { 2201 inet_csk_listen_stop(sk); 2202 } else if (unlikely(tp->repair)) { 2203 sk->sk_err = ECONNABORTED; 2204 } else if (tcp_need_reset(old_state) || 2205 (tp->snd_nxt != tp->write_seq && 2206 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) { 2207 /* The last check adjusts for discrepancy of Linux wrt. RFC 2208 * states 2209 */ 2210 tcp_send_active_reset(sk, gfp_any()); 2211 sk->sk_err = ECONNRESET; 2212 } else if (old_state == TCP_SYN_SENT) 2213 sk->sk_err = ECONNRESET; 2214 2215 tcp_clear_xmit_timers(sk); 2216 __skb_queue_purge(&sk->sk_receive_queue); 2217 tcp_write_queue_purge(sk); 2218 __skb_queue_purge(&tp->out_of_order_queue); 2219 2220 inet->inet_dport = 0; 2221 2222 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK)) 2223 inet_reset_saddr(sk); 2224 2225 sk->sk_shutdown = 0; 2226 sock_reset_flag(sk, SOCK_DONE); 2227 tp->srtt_us = 0; 2228 if ((tp->write_seq += tp->max_window + 2) == 0) 2229 tp->write_seq = 1; 2230 icsk->icsk_backoff = 0; 2231 tp->snd_cwnd = 2; 2232 icsk->icsk_probes_out = 0; 2233 tp->packets_out = 0; 2234 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 2235 tp->snd_cwnd_cnt = 0; 2236 tp->window_clamp = 0; 2237 tcp_set_ca_state(sk, TCP_CA_Open); 2238 tcp_clear_retrans(tp); 2239 inet_csk_delack_init(sk); 2240 tcp_init_send_head(sk); 2241 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt)); 2242 __sk_dst_reset(sk); 2243 2244 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash); 2245 2246 sk->sk_error_report(sk); 2247 return err; 2248 } 2249 EXPORT_SYMBOL(tcp_disconnect); 2250 2251 void tcp_sock_destruct(struct sock *sk) 2252 { 2253 inet_sock_destruct(sk); 2254 2255 kfree(inet_csk(sk)->icsk_accept_queue.fastopenq); 2256 } 2257 2258 static inline bool tcp_can_repair_sock(const struct sock *sk) 2259 { 2260 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) && 2261 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED)); 2262 } 2263 2264 static int tcp_repair_options_est(struct tcp_sock *tp, 2265 struct tcp_repair_opt __user *optbuf, unsigned int len) 2266 { 2267 struct tcp_repair_opt opt; 2268 2269 while (len >= sizeof(opt)) { 2270 if (copy_from_user(&opt, optbuf, sizeof(opt))) 2271 return -EFAULT; 2272 2273 optbuf++; 2274 len -= sizeof(opt); 2275 2276 switch (opt.opt_code) { 2277 case TCPOPT_MSS: 2278 tp->rx_opt.mss_clamp = opt.opt_val; 2279 break; 2280 case TCPOPT_WINDOW: 2281 { 2282 u16 snd_wscale = opt.opt_val & 0xFFFF; 2283 u16 rcv_wscale = opt.opt_val >> 16; 2284 2285 if (snd_wscale > 14 || rcv_wscale > 14) 2286 return -EFBIG; 2287 2288 tp->rx_opt.snd_wscale = snd_wscale; 2289 tp->rx_opt.rcv_wscale = rcv_wscale; 2290 tp->rx_opt.wscale_ok = 1; 2291 } 2292 break; 2293 case TCPOPT_SACK_PERM: 2294 if (opt.opt_val != 0) 2295 return -EINVAL; 2296 2297 tp->rx_opt.sack_ok |= TCP_SACK_SEEN; 2298 if (sysctl_tcp_fack) 2299 tcp_enable_fack(tp); 2300 break; 2301 case TCPOPT_TIMESTAMP: 2302 if (opt.opt_val != 0) 2303 return -EINVAL; 2304 2305 tp->rx_opt.tstamp_ok = 1; 2306 break; 2307 } 2308 } 2309 2310 return 0; 2311 } 2312 2313 /* 2314 * Socket option code for TCP. 2315 */ 2316 static int do_tcp_setsockopt(struct sock *sk, int level, 2317 int optname, char __user *optval, unsigned int optlen) 2318 { 2319 struct tcp_sock *tp = tcp_sk(sk); 2320 struct inet_connection_sock *icsk = inet_csk(sk); 2321 int val; 2322 int err = 0; 2323 2324 /* These are data/string values, all the others are ints */ 2325 switch (optname) { 2326 case TCP_CONGESTION: { 2327 char name[TCP_CA_NAME_MAX]; 2328 2329 if (optlen < 1) 2330 return -EINVAL; 2331 2332 val = strncpy_from_user(name, optval, 2333 min_t(long, TCP_CA_NAME_MAX-1, optlen)); 2334 if (val < 0) 2335 return -EFAULT; 2336 name[val] = 0; 2337 2338 lock_sock(sk); 2339 err = tcp_set_congestion_control(sk, name); 2340 release_sock(sk); 2341 return err; 2342 } 2343 default: 2344 /* fallthru */ 2345 break; 2346 } 2347 2348 if (optlen < sizeof(int)) 2349 return -EINVAL; 2350 2351 if (get_user(val, (int __user *)optval)) 2352 return -EFAULT; 2353 2354 lock_sock(sk); 2355 2356 switch (optname) { 2357 case TCP_MAXSEG: 2358 /* Values greater than interface MTU won't take effect. However 2359 * at the point when this call is done we typically don't yet 2360 * know which interface is going to be used */ 2361 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) { 2362 err = -EINVAL; 2363 break; 2364 } 2365 tp->rx_opt.user_mss = val; 2366 break; 2367 2368 case TCP_NODELAY: 2369 if (val) { 2370 /* TCP_NODELAY is weaker than TCP_CORK, so that 2371 * this option on corked socket is remembered, but 2372 * it is not activated until cork is cleared. 2373 * 2374 * However, when TCP_NODELAY is set we make 2375 * an explicit push, which overrides even TCP_CORK 2376 * for currently queued segments. 2377 */ 2378 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH; 2379 tcp_push_pending_frames(sk); 2380 } else { 2381 tp->nonagle &= ~TCP_NAGLE_OFF; 2382 } 2383 break; 2384 2385 case TCP_THIN_LINEAR_TIMEOUTS: 2386 if (val < 0 || val > 1) 2387 err = -EINVAL; 2388 else 2389 tp->thin_lto = val; 2390 break; 2391 2392 case TCP_THIN_DUPACK: 2393 if (val < 0 || val > 1) 2394 err = -EINVAL; 2395 else { 2396 tp->thin_dupack = val; 2397 if (tp->thin_dupack) 2398 tcp_disable_early_retrans(tp); 2399 } 2400 break; 2401 2402 case TCP_REPAIR: 2403 if (!tcp_can_repair_sock(sk)) 2404 err = -EPERM; 2405 else if (val == 1) { 2406 tp->repair = 1; 2407 sk->sk_reuse = SK_FORCE_REUSE; 2408 tp->repair_queue = TCP_NO_QUEUE; 2409 } else if (val == 0) { 2410 tp->repair = 0; 2411 sk->sk_reuse = SK_NO_REUSE; 2412 tcp_send_window_probe(sk); 2413 } else 2414 err = -EINVAL; 2415 2416 break; 2417 2418 case TCP_REPAIR_QUEUE: 2419 if (!tp->repair) 2420 err = -EPERM; 2421 else if (val < TCP_QUEUES_NR) 2422 tp->repair_queue = val; 2423 else 2424 err = -EINVAL; 2425 break; 2426 2427 case TCP_QUEUE_SEQ: 2428 if (sk->sk_state != TCP_CLOSE) 2429 err = -EPERM; 2430 else if (tp->repair_queue == TCP_SEND_QUEUE) 2431 tp->write_seq = val; 2432 else if (tp->repair_queue == TCP_RECV_QUEUE) 2433 tp->rcv_nxt = val; 2434 else 2435 err = -EINVAL; 2436 break; 2437 2438 case TCP_REPAIR_OPTIONS: 2439 if (!tp->repair) 2440 err = -EINVAL; 2441 else if (sk->sk_state == TCP_ESTABLISHED) 2442 err = tcp_repair_options_est(tp, 2443 (struct tcp_repair_opt __user *)optval, 2444 optlen); 2445 else 2446 err = -EPERM; 2447 break; 2448 2449 case TCP_CORK: 2450 /* When set indicates to always queue non-full frames. 2451 * Later the user clears this option and we transmit 2452 * any pending partial frames in the queue. This is 2453 * meant to be used alongside sendfile() to get properly 2454 * filled frames when the user (for example) must write 2455 * out headers with a write() call first and then use 2456 * sendfile to send out the data parts. 2457 * 2458 * TCP_CORK can be set together with TCP_NODELAY and it is 2459 * stronger than TCP_NODELAY. 2460 */ 2461 if (val) { 2462 tp->nonagle |= TCP_NAGLE_CORK; 2463 } else { 2464 tp->nonagle &= ~TCP_NAGLE_CORK; 2465 if (tp->nonagle&TCP_NAGLE_OFF) 2466 tp->nonagle |= TCP_NAGLE_PUSH; 2467 tcp_push_pending_frames(sk); 2468 } 2469 break; 2470 2471 case TCP_KEEPIDLE: 2472 if (val < 1 || val > MAX_TCP_KEEPIDLE) 2473 err = -EINVAL; 2474 else { 2475 tp->keepalive_time = val * HZ; 2476 if (sock_flag(sk, SOCK_KEEPOPEN) && 2477 !((1 << sk->sk_state) & 2478 (TCPF_CLOSE | TCPF_LISTEN))) { 2479 u32 elapsed = keepalive_time_elapsed(tp); 2480 if (tp->keepalive_time > elapsed) 2481 elapsed = tp->keepalive_time - elapsed; 2482 else 2483 elapsed = 0; 2484 inet_csk_reset_keepalive_timer(sk, elapsed); 2485 } 2486 } 2487 break; 2488 case TCP_KEEPINTVL: 2489 if (val < 1 || val > MAX_TCP_KEEPINTVL) 2490 err = -EINVAL; 2491 else 2492 tp->keepalive_intvl = val * HZ; 2493 break; 2494 case TCP_KEEPCNT: 2495 if (val < 1 || val > MAX_TCP_KEEPCNT) 2496 err = -EINVAL; 2497 else 2498 tp->keepalive_probes = val; 2499 break; 2500 case TCP_SYNCNT: 2501 if (val < 1 || val > MAX_TCP_SYNCNT) 2502 err = -EINVAL; 2503 else 2504 icsk->icsk_syn_retries = val; 2505 break; 2506 2507 case TCP_SAVE_SYN: 2508 if (val < 0 || val > 1) 2509 err = -EINVAL; 2510 else 2511 tp->save_syn = val; 2512 break; 2513 2514 case TCP_LINGER2: 2515 if (val < 0) 2516 tp->linger2 = -1; 2517 else if (val > sysctl_tcp_fin_timeout / HZ) 2518 tp->linger2 = 0; 2519 else 2520 tp->linger2 = val * HZ; 2521 break; 2522 2523 case TCP_DEFER_ACCEPT: 2524 /* Translate value in seconds to number of retransmits */ 2525 icsk->icsk_accept_queue.rskq_defer_accept = 2526 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ, 2527 TCP_RTO_MAX / HZ); 2528 break; 2529 2530 case TCP_WINDOW_CLAMP: 2531 if (!val) { 2532 if (sk->sk_state != TCP_CLOSE) { 2533 err = -EINVAL; 2534 break; 2535 } 2536 tp->window_clamp = 0; 2537 } else 2538 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ? 2539 SOCK_MIN_RCVBUF / 2 : val; 2540 break; 2541 2542 case TCP_QUICKACK: 2543 if (!val) { 2544 icsk->icsk_ack.pingpong = 1; 2545 } else { 2546 icsk->icsk_ack.pingpong = 0; 2547 if ((1 << sk->sk_state) & 2548 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) && 2549 inet_csk_ack_scheduled(sk)) { 2550 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 2551 tcp_cleanup_rbuf(sk, 1); 2552 if (!(val & 1)) 2553 icsk->icsk_ack.pingpong = 1; 2554 } 2555 } 2556 break; 2557 2558 #ifdef CONFIG_TCP_MD5SIG 2559 case TCP_MD5SIG: 2560 /* Read the IP->Key mappings from userspace */ 2561 err = tp->af_specific->md5_parse(sk, optval, optlen); 2562 break; 2563 #endif 2564 case TCP_USER_TIMEOUT: 2565 /* Cap the max time in ms TCP will retry or probe the window 2566 * before giving up and aborting (ETIMEDOUT) a connection. 2567 */ 2568 if (val < 0) 2569 err = -EINVAL; 2570 else 2571 icsk->icsk_user_timeout = msecs_to_jiffies(val); 2572 break; 2573 2574 case TCP_FASTOPEN: 2575 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE | 2576 TCPF_LISTEN))) { 2577 tcp_fastopen_init_key_once(true); 2578 2579 err = fastopen_init_queue(sk, val); 2580 } else { 2581 err = -EINVAL; 2582 } 2583 break; 2584 case TCP_TIMESTAMP: 2585 if (!tp->repair) 2586 err = -EPERM; 2587 else 2588 tp->tsoffset = val - tcp_time_stamp; 2589 break; 2590 case TCP_NOTSENT_LOWAT: 2591 tp->notsent_lowat = val; 2592 sk->sk_write_space(sk); 2593 break; 2594 default: 2595 err = -ENOPROTOOPT; 2596 break; 2597 } 2598 2599 release_sock(sk); 2600 return err; 2601 } 2602 2603 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval, 2604 unsigned int optlen) 2605 { 2606 const struct inet_connection_sock *icsk = inet_csk(sk); 2607 2608 if (level != SOL_TCP) 2609 return icsk->icsk_af_ops->setsockopt(sk, level, optname, 2610 optval, optlen); 2611 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2612 } 2613 EXPORT_SYMBOL(tcp_setsockopt); 2614 2615 #ifdef CONFIG_COMPAT 2616 int compat_tcp_setsockopt(struct sock *sk, int level, int optname, 2617 char __user *optval, unsigned int optlen) 2618 { 2619 if (level != SOL_TCP) 2620 return inet_csk_compat_setsockopt(sk, level, optname, 2621 optval, optlen); 2622 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2623 } 2624 EXPORT_SYMBOL(compat_tcp_setsockopt); 2625 #endif 2626 2627 /* Return information about state of tcp endpoint in API format. */ 2628 void tcp_get_info(struct sock *sk, struct tcp_info *info) 2629 { 2630 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */ 2631 const struct inet_connection_sock *icsk = inet_csk(sk); 2632 u32 now = tcp_time_stamp; 2633 unsigned int start; 2634 u32 rate; 2635 2636 memset(info, 0, sizeof(*info)); 2637 if (sk->sk_type != SOCK_STREAM) 2638 return; 2639 2640 info->tcpi_state = sk->sk_state; 2641 info->tcpi_ca_state = icsk->icsk_ca_state; 2642 info->tcpi_retransmits = icsk->icsk_retransmits; 2643 info->tcpi_probes = icsk->icsk_probes_out; 2644 info->tcpi_backoff = icsk->icsk_backoff; 2645 2646 if (tp->rx_opt.tstamp_ok) 2647 info->tcpi_options |= TCPI_OPT_TIMESTAMPS; 2648 if (tcp_is_sack(tp)) 2649 info->tcpi_options |= TCPI_OPT_SACK; 2650 if (tp->rx_opt.wscale_ok) { 2651 info->tcpi_options |= TCPI_OPT_WSCALE; 2652 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale; 2653 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale; 2654 } 2655 2656 if (tp->ecn_flags & TCP_ECN_OK) 2657 info->tcpi_options |= TCPI_OPT_ECN; 2658 if (tp->ecn_flags & TCP_ECN_SEEN) 2659 info->tcpi_options |= TCPI_OPT_ECN_SEEN; 2660 if (tp->syn_data_acked) 2661 info->tcpi_options |= TCPI_OPT_SYN_DATA; 2662 2663 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto); 2664 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato); 2665 info->tcpi_snd_mss = tp->mss_cache; 2666 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss; 2667 2668 if (sk->sk_state == TCP_LISTEN) { 2669 info->tcpi_unacked = sk->sk_ack_backlog; 2670 info->tcpi_sacked = sk->sk_max_ack_backlog; 2671 } else { 2672 info->tcpi_unacked = tp->packets_out; 2673 info->tcpi_sacked = tp->sacked_out; 2674 } 2675 info->tcpi_lost = tp->lost_out; 2676 info->tcpi_retrans = tp->retrans_out; 2677 info->tcpi_fackets = tp->fackets_out; 2678 2679 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime); 2680 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime); 2681 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp); 2682 2683 info->tcpi_pmtu = icsk->icsk_pmtu_cookie; 2684 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh; 2685 info->tcpi_rtt = tp->srtt_us >> 3; 2686 info->tcpi_rttvar = tp->mdev_us >> 2; 2687 info->tcpi_snd_ssthresh = tp->snd_ssthresh; 2688 info->tcpi_snd_cwnd = tp->snd_cwnd; 2689 info->tcpi_advmss = tp->advmss; 2690 info->tcpi_reordering = tp->reordering; 2691 2692 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3; 2693 info->tcpi_rcv_space = tp->rcvq_space.space; 2694 2695 info->tcpi_total_retrans = tp->total_retrans; 2696 2697 rate = READ_ONCE(sk->sk_pacing_rate); 2698 info->tcpi_pacing_rate = rate != ~0U ? rate : ~0ULL; 2699 2700 rate = READ_ONCE(sk->sk_max_pacing_rate); 2701 info->tcpi_max_pacing_rate = rate != ~0U ? rate : ~0ULL; 2702 2703 do { 2704 start = u64_stats_fetch_begin_irq(&tp->syncp); 2705 info->tcpi_bytes_acked = tp->bytes_acked; 2706 info->tcpi_bytes_received = tp->bytes_received; 2707 } while (u64_stats_fetch_retry_irq(&tp->syncp, start)); 2708 info->tcpi_segs_out = tp->segs_out; 2709 info->tcpi_segs_in = tp->segs_in; 2710 } 2711 EXPORT_SYMBOL_GPL(tcp_get_info); 2712 2713 static int do_tcp_getsockopt(struct sock *sk, int level, 2714 int optname, char __user *optval, int __user *optlen) 2715 { 2716 struct inet_connection_sock *icsk = inet_csk(sk); 2717 struct tcp_sock *tp = tcp_sk(sk); 2718 int val, len; 2719 2720 if (get_user(len, optlen)) 2721 return -EFAULT; 2722 2723 len = min_t(unsigned int, len, sizeof(int)); 2724 2725 if (len < 0) 2726 return -EINVAL; 2727 2728 switch (optname) { 2729 case TCP_MAXSEG: 2730 val = tp->mss_cache; 2731 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 2732 val = tp->rx_opt.user_mss; 2733 if (tp->repair) 2734 val = tp->rx_opt.mss_clamp; 2735 break; 2736 case TCP_NODELAY: 2737 val = !!(tp->nonagle&TCP_NAGLE_OFF); 2738 break; 2739 case TCP_CORK: 2740 val = !!(tp->nonagle&TCP_NAGLE_CORK); 2741 break; 2742 case TCP_KEEPIDLE: 2743 val = keepalive_time_when(tp) / HZ; 2744 break; 2745 case TCP_KEEPINTVL: 2746 val = keepalive_intvl_when(tp) / HZ; 2747 break; 2748 case TCP_KEEPCNT: 2749 val = keepalive_probes(tp); 2750 break; 2751 case TCP_SYNCNT: 2752 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries; 2753 break; 2754 case TCP_LINGER2: 2755 val = tp->linger2; 2756 if (val >= 0) 2757 val = (val ? : sysctl_tcp_fin_timeout) / HZ; 2758 break; 2759 case TCP_DEFER_ACCEPT: 2760 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept, 2761 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ); 2762 break; 2763 case TCP_WINDOW_CLAMP: 2764 val = tp->window_clamp; 2765 break; 2766 case TCP_INFO: { 2767 struct tcp_info info; 2768 2769 if (get_user(len, optlen)) 2770 return -EFAULT; 2771 2772 tcp_get_info(sk, &info); 2773 2774 len = min_t(unsigned int, len, sizeof(info)); 2775 if (put_user(len, optlen)) 2776 return -EFAULT; 2777 if (copy_to_user(optval, &info, len)) 2778 return -EFAULT; 2779 return 0; 2780 } 2781 case TCP_CC_INFO: { 2782 const struct tcp_congestion_ops *ca_ops; 2783 union tcp_cc_info info; 2784 size_t sz = 0; 2785 int attr; 2786 2787 if (get_user(len, optlen)) 2788 return -EFAULT; 2789 2790 ca_ops = icsk->icsk_ca_ops; 2791 if (ca_ops && ca_ops->get_info) 2792 sz = ca_ops->get_info(sk, ~0U, &attr, &info); 2793 2794 len = min_t(unsigned int, len, sz); 2795 if (put_user(len, optlen)) 2796 return -EFAULT; 2797 if (copy_to_user(optval, &info, len)) 2798 return -EFAULT; 2799 return 0; 2800 } 2801 case TCP_QUICKACK: 2802 val = !icsk->icsk_ack.pingpong; 2803 break; 2804 2805 case TCP_CONGESTION: 2806 if (get_user(len, optlen)) 2807 return -EFAULT; 2808 len = min_t(unsigned int, len, TCP_CA_NAME_MAX); 2809 if (put_user(len, optlen)) 2810 return -EFAULT; 2811 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len)) 2812 return -EFAULT; 2813 return 0; 2814 2815 case TCP_THIN_LINEAR_TIMEOUTS: 2816 val = tp->thin_lto; 2817 break; 2818 case TCP_THIN_DUPACK: 2819 val = tp->thin_dupack; 2820 break; 2821 2822 case TCP_REPAIR: 2823 val = tp->repair; 2824 break; 2825 2826 case TCP_REPAIR_QUEUE: 2827 if (tp->repair) 2828 val = tp->repair_queue; 2829 else 2830 return -EINVAL; 2831 break; 2832 2833 case TCP_QUEUE_SEQ: 2834 if (tp->repair_queue == TCP_SEND_QUEUE) 2835 val = tp->write_seq; 2836 else if (tp->repair_queue == TCP_RECV_QUEUE) 2837 val = tp->rcv_nxt; 2838 else 2839 return -EINVAL; 2840 break; 2841 2842 case TCP_USER_TIMEOUT: 2843 val = jiffies_to_msecs(icsk->icsk_user_timeout); 2844 break; 2845 2846 case TCP_FASTOPEN: 2847 if (icsk->icsk_accept_queue.fastopenq) 2848 val = icsk->icsk_accept_queue.fastopenq->max_qlen; 2849 else 2850 val = 0; 2851 break; 2852 2853 case TCP_TIMESTAMP: 2854 val = tcp_time_stamp + tp->tsoffset; 2855 break; 2856 case TCP_NOTSENT_LOWAT: 2857 val = tp->notsent_lowat; 2858 break; 2859 case TCP_SAVE_SYN: 2860 val = tp->save_syn; 2861 break; 2862 case TCP_SAVED_SYN: { 2863 if (get_user(len, optlen)) 2864 return -EFAULT; 2865 2866 lock_sock(sk); 2867 if (tp->saved_syn) { 2868 if (len < tp->saved_syn[0]) { 2869 if (put_user(tp->saved_syn[0], optlen)) { 2870 release_sock(sk); 2871 return -EFAULT; 2872 } 2873 release_sock(sk); 2874 return -EINVAL; 2875 } 2876 len = tp->saved_syn[0]; 2877 if (put_user(len, optlen)) { 2878 release_sock(sk); 2879 return -EFAULT; 2880 } 2881 if (copy_to_user(optval, tp->saved_syn + 1, len)) { 2882 release_sock(sk); 2883 return -EFAULT; 2884 } 2885 tcp_saved_syn_free(tp); 2886 release_sock(sk); 2887 } else { 2888 release_sock(sk); 2889 len = 0; 2890 if (put_user(len, optlen)) 2891 return -EFAULT; 2892 } 2893 return 0; 2894 } 2895 default: 2896 return -ENOPROTOOPT; 2897 } 2898 2899 if (put_user(len, optlen)) 2900 return -EFAULT; 2901 if (copy_to_user(optval, &val, len)) 2902 return -EFAULT; 2903 return 0; 2904 } 2905 2906 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval, 2907 int __user *optlen) 2908 { 2909 struct inet_connection_sock *icsk = inet_csk(sk); 2910 2911 if (level != SOL_TCP) 2912 return icsk->icsk_af_ops->getsockopt(sk, level, optname, 2913 optval, optlen); 2914 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2915 } 2916 EXPORT_SYMBOL(tcp_getsockopt); 2917 2918 #ifdef CONFIG_COMPAT 2919 int compat_tcp_getsockopt(struct sock *sk, int level, int optname, 2920 char __user *optval, int __user *optlen) 2921 { 2922 if (level != SOL_TCP) 2923 return inet_csk_compat_getsockopt(sk, level, optname, 2924 optval, optlen); 2925 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2926 } 2927 EXPORT_SYMBOL(compat_tcp_getsockopt); 2928 #endif 2929 2930 #ifdef CONFIG_TCP_MD5SIG 2931 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool); 2932 static DEFINE_MUTEX(tcp_md5sig_mutex); 2933 static bool tcp_md5sig_pool_populated = false; 2934 2935 static void __tcp_alloc_md5sig_pool(void) 2936 { 2937 int cpu; 2938 2939 for_each_possible_cpu(cpu) { 2940 if (!per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm) { 2941 struct crypto_hash *hash; 2942 2943 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC); 2944 if (IS_ERR_OR_NULL(hash)) 2945 return; 2946 per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm = hash; 2947 } 2948 } 2949 /* before setting tcp_md5sig_pool_populated, we must commit all writes 2950 * to memory. See smp_rmb() in tcp_get_md5sig_pool() 2951 */ 2952 smp_wmb(); 2953 tcp_md5sig_pool_populated = true; 2954 } 2955 2956 bool tcp_alloc_md5sig_pool(void) 2957 { 2958 if (unlikely(!tcp_md5sig_pool_populated)) { 2959 mutex_lock(&tcp_md5sig_mutex); 2960 2961 if (!tcp_md5sig_pool_populated) 2962 __tcp_alloc_md5sig_pool(); 2963 2964 mutex_unlock(&tcp_md5sig_mutex); 2965 } 2966 return tcp_md5sig_pool_populated; 2967 } 2968 EXPORT_SYMBOL(tcp_alloc_md5sig_pool); 2969 2970 2971 /** 2972 * tcp_get_md5sig_pool - get md5sig_pool for this user 2973 * 2974 * We use percpu structure, so if we succeed, we exit with preemption 2975 * and BH disabled, to make sure another thread or softirq handling 2976 * wont try to get same context. 2977 */ 2978 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void) 2979 { 2980 local_bh_disable(); 2981 2982 if (tcp_md5sig_pool_populated) { 2983 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */ 2984 smp_rmb(); 2985 return this_cpu_ptr(&tcp_md5sig_pool); 2986 } 2987 local_bh_enable(); 2988 return NULL; 2989 } 2990 EXPORT_SYMBOL(tcp_get_md5sig_pool); 2991 2992 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp, 2993 const struct tcphdr *th) 2994 { 2995 struct scatterlist sg; 2996 struct tcphdr hdr; 2997 int err; 2998 2999 /* We are not allowed to change tcphdr, make a local copy */ 3000 memcpy(&hdr, th, sizeof(hdr)); 3001 hdr.check = 0; 3002 3003 /* options aren't included in the hash */ 3004 sg_init_one(&sg, &hdr, sizeof(hdr)); 3005 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr)); 3006 return err; 3007 } 3008 EXPORT_SYMBOL(tcp_md5_hash_header); 3009 3010 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp, 3011 const struct sk_buff *skb, unsigned int header_len) 3012 { 3013 struct scatterlist sg; 3014 const struct tcphdr *tp = tcp_hdr(skb); 3015 struct hash_desc *desc = &hp->md5_desc; 3016 unsigned int i; 3017 const unsigned int head_data_len = skb_headlen(skb) > header_len ? 3018 skb_headlen(skb) - header_len : 0; 3019 const struct skb_shared_info *shi = skb_shinfo(skb); 3020 struct sk_buff *frag_iter; 3021 3022 sg_init_table(&sg, 1); 3023 3024 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len); 3025 if (crypto_hash_update(desc, &sg, head_data_len)) 3026 return 1; 3027 3028 for (i = 0; i < shi->nr_frags; ++i) { 3029 const struct skb_frag_struct *f = &shi->frags[i]; 3030 unsigned int offset = f->page_offset; 3031 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT); 3032 3033 sg_set_page(&sg, page, skb_frag_size(f), 3034 offset_in_page(offset)); 3035 if (crypto_hash_update(desc, &sg, skb_frag_size(f))) 3036 return 1; 3037 } 3038 3039 skb_walk_frags(skb, frag_iter) 3040 if (tcp_md5_hash_skb_data(hp, frag_iter, 0)) 3041 return 1; 3042 3043 return 0; 3044 } 3045 EXPORT_SYMBOL(tcp_md5_hash_skb_data); 3046 3047 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key) 3048 { 3049 struct scatterlist sg; 3050 3051 sg_init_one(&sg, key->key, key->keylen); 3052 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen); 3053 } 3054 EXPORT_SYMBOL(tcp_md5_hash_key); 3055 3056 #endif 3057 3058 void tcp_done(struct sock *sk) 3059 { 3060 struct request_sock *req = tcp_sk(sk)->fastopen_rsk; 3061 3062 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV) 3063 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 3064 3065 tcp_set_state(sk, TCP_CLOSE); 3066 tcp_clear_xmit_timers(sk); 3067 if (req) 3068 reqsk_fastopen_remove(sk, req, false); 3069 3070 sk->sk_shutdown = SHUTDOWN_MASK; 3071 3072 if (!sock_flag(sk, SOCK_DEAD)) 3073 sk->sk_state_change(sk); 3074 else 3075 inet_csk_destroy_sock(sk); 3076 } 3077 EXPORT_SYMBOL_GPL(tcp_done); 3078 3079 extern struct tcp_congestion_ops tcp_reno; 3080 3081 static __initdata unsigned long thash_entries; 3082 static int __init set_thash_entries(char *str) 3083 { 3084 ssize_t ret; 3085 3086 if (!str) 3087 return 0; 3088 3089 ret = kstrtoul(str, 0, &thash_entries); 3090 if (ret) 3091 return 0; 3092 3093 return 1; 3094 } 3095 __setup("thash_entries=", set_thash_entries); 3096 3097 static void __init tcp_init_mem(void) 3098 { 3099 unsigned long limit = nr_free_buffer_pages() / 16; 3100 3101 limit = max(limit, 128UL); 3102 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */ 3103 sysctl_tcp_mem[1] = limit; /* 6.25 % */ 3104 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */ 3105 } 3106 3107 void __init tcp_init(void) 3108 { 3109 unsigned long limit; 3110 int max_rshare, max_wshare, cnt; 3111 unsigned int i; 3112 3113 sock_skb_cb_check_size(sizeof(struct tcp_skb_cb)); 3114 3115 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL); 3116 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL); 3117 tcp_hashinfo.bind_bucket_cachep = 3118 kmem_cache_create("tcp_bind_bucket", 3119 sizeof(struct inet_bind_bucket), 0, 3120 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL); 3121 3122 /* Size and allocate the main established and bind bucket 3123 * hash tables. 3124 * 3125 * The methodology is similar to that of the buffer cache. 3126 */ 3127 tcp_hashinfo.ehash = 3128 alloc_large_system_hash("TCP established", 3129 sizeof(struct inet_ehash_bucket), 3130 thash_entries, 3131 17, /* one slot per 128 KB of memory */ 3132 0, 3133 NULL, 3134 &tcp_hashinfo.ehash_mask, 3135 0, 3136 thash_entries ? 0 : 512 * 1024); 3137 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) 3138 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i); 3139 3140 if (inet_ehash_locks_alloc(&tcp_hashinfo)) 3141 panic("TCP: failed to alloc ehash_locks"); 3142 tcp_hashinfo.bhash = 3143 alloc_large_system_hash("TCP bind", 3144 sizeof(struct inet_bind_hashbucket), 3145 tcp_hashinfo.ehash_mask + 1, 3146 17, /* one slot per 128 KB of memory */ 3147 0, 3148 &tcp_hashinfo.bhash_size, 3149 NULL, 3150 0, 3151 64 * 1024); 3152 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size; 3153 for (i = 0; i < tcp_hashinfo.bhash_size; i++) { 3154 spin_lock_init(&tcp_hashinfo.bhash[i].lock); 3155 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain); 3156 } 3157 3158 3159 cnt = tcp_hashinfo.ehash_mask + 1; 3160 3161 tcp_death_row.sysctl_max_tw_buckets = cnt / 2; 3162 sysctl_tcp_max_orphans = cnt / 2; 3163 sysctl_max_syn_backlog = max(128, cnt / 256); 3164 3165 tcp_init_mem(); 3166 /* Set per-socket limits to no more than 1/128 the pressure threshold */ 3167 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7); 3168 max_wshare = min(4UL*1024*1024, limit); 3169 max_rshare = min(6UL*1024*1024, limit); 3170 3171 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM; 3172 sysctl_tcp_wmem[1] = 16*1024; 3173 sysctl_tcp_wmem[2] = max(64*1024, max_wshare); 3174 3175 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM; 3176 sysctl_tcp_rmem[1] = 87380; 3177 sysctl_tcp_rmem[2] = max(87380, max_rshare); 3178 3179 pr_info("Hash tables configured (established %u bind %u)\n", 3180 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size); 3181 3182 tcp_metrics_init(); 3183 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0); 3184 tcp_tasklet_init(); 3185 } 3186