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 <crypto/hash.h> 251 #include <linux/kernel.h> 252 #include <linux/module.h> 253 #include <linux/types.h> 254 #include <linux/fcntl.h> 255 #include <linux/poll.h> 256 #include <linux/inet_diag.h> 257 #include <linux/init.h> 258 #include <linux/fs.h> 259 #include <linux/skbuff.h> 260 #include <linux/scatterlist.h> 261 #include <linux/splice.h> 262 #include <linux/net.h> 263 #include <linux/socket.h> 264 #include <linux/random.h> 265 #include <linux/bootmem.h> 266 #include <linux/highmem.h> 267 #include <linux/swap.h> 268 #include <linux/cache.h> 269 #include <linux/err.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 <asm/unaligned.h> 283 #include <net/busy_poll.h> 284 285 int sysctl_tcp_min_tso_segs __read_mostly = 2; 286 287 int sysctl_tcp_autocorking __read_mostly = 1; 288 289 struct percpu_counter tcp_orphan_count; 290 EXPORT_SYMBOL_GPL(tcp_orphan_count); 291 292 long sysctl_tcp_mem[3] __read_mostly; 293 int sysctl_tcp_wmem[3] __read_mostly; 294 int sysctl_tcp_rmem[3] __read_mostly; 295 296 EXPORT_SYMBOL(sysctl_tcp_mem); 297 EXPORT_SYMBOL(sysctl_tcp_rmem); 298 EXPORT_SYMBOL(sysctl_tcp_wmem); 299 300 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */ 301 EXPORT_SYMBOL(tcp_memory_allocated); 302 303 /* 304 * Current number of TCP sockets. 305 */ 306 struct percpu_counter tcp_sockets_allocated; 307 EXPORT_SYMBOL(tcp_sockets_allocated); 308 309 /* 310 * TCP splice context 311 */ 312 struct tcp_splice_state { 313 struct pipe_inode_info *pipe; 314 size_t len; 315 unsigned int flags; 316 }; 317 318 /* 319 * Pressure flag: try to collapse. 320 * Technical note: it is used by multiple contexts non atomically. 321 * All the __sk_mem_schedule() is of this nature: accounting 322 * is strict, actions are advisory and have some latency. 323 */ 324 int tcp_memory_pressure __read_mostly; 325 EXPORT_SYMBOL(tcp_memory_pressure); 326 327 void tcp_enter_memory_pressure(struct sock *sk) 328 { 329 if (!tcp_memory_pressure) { 330 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES); 331 tcp_memory_pressure = 1; 332 } 333 } 334 EXPORT_SYMBOL(tcp_enter_memory_pressure); 335 336 /* Convert seconds to retransmits based on initial and max timeout */ 337 static u8 secs_to_retrans(int seconds, int timeout, int rto_max) 338 { 339 u8 res = 0; 340 341 if (seconds > 0) { 342 int period = timeout; 343 344 res = 1; 345 while (seconds > period && res < 255) { 346 res++; 347 timeout <<= 1; 348 if (timeout > rto_max) 349 timeout = rto_max; 350 period += timeout; 351 } 352 } 353 return res; 354 } 355 356 /* Convert retransmits to seconds based on initial and max timeout */ 357 static int retrans_to_secs(u8 retrans, int timeout, int rto_max) 358 { 359 int period = 0; 360 361 if (retrans > 0) { 362 period = timeout; 363 while (--retrans) { 364 timeout <<= 1; 365 if (timeout > rto_max) 366 timeout = rto_max; 367 period += timeout; 368 } 369 } 370 return period; 371 } 372 373 /* Address-family independent initialization for a tcp_sock. 374 * 375 * NOTE: A lot of things set to zero explicitly by call to 376 * sk_alloc() so need not be done here. 377 */ 378 void tcp_init_sock(struct sock *sk) 379 { 380 struct inet_connection_sock *icsk = inet_csk(sk); 381 struct tcp_sock *tp = tcp_sk(sk); 382 383 __skb_queue_head_init(&tp->out_of_order_queue); 384 tcp_init_xmit_timers(sk); 385 tcp_prequeue_init(tp); 386 INIT_LIST_HEAD(&tp->tsq_node); 387 388 icsk->icsk_rto = TCP_TIMEOUT_INIT; 389 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT); 390 tp->rtt_min[0].rtt = ~0U; 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 = sock_net(sk)->ipv4.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 if (mem_cgroup_sockets_enabled) 425 sock_update_memcg(sk); 426 sk_sockets_allocated_inc(sk); 427 local_bh_enable(); 428 } 429 EXPORT_SYMBOL(tcp_init_sock); 430 431 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags, struct sk_buff *skb) 432 { 433 if (tsflags) { 434 struct skb_shared_info *shinfo = skb_shinfo(skb); 435 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 436 437 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags); 438 if (tsflags & SOF_TIMESTAMPING_TX_ACK) 439 tcb->txstamp_ack = 1; 440 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK) 441 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1; 442 } 443 } 444 445 /* 446 * Wait for a TCP event. 447 * 448 * Note that we don't need to lock the socket, as the upper poll layers 449 * take care of normal races (between the test and the event) and we don't 450 * go look at any of the socket buffers directly. 451 */ 452 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait) 453 { 454 unsigned int mask; 455 struct sock *sk = sock->sk; 456 const struct tcp_sock *tp = tcp_sk(sk); 457 int state; 458 459 sock_rps_record_flow(sk); 460 461 sock_poll_wait(file, sk_sleep(sk), wait); 462 463 state = sk_state_load(sk); 464 if (state == TCP_LISTEN) 465 return inet_csk_listen_poll(sk); 466 467 /* Socket is not locked. We are protected from async events 468 * by poll logic and correct handling of state changes 469 * made by other threads is impossible in any case. 470 */ 471 472 mask = 0; 473 474 /* 475 * POLLHUP is certainly not done right. But poll() doesn't 476 * have a notion of HUP in just one direction, and for a 477 * socket the read side is more interesting. 478 * 479 * Some poll() documentation says that POLLHUP is incompatible 480 * with the POLLOUT/POLLWR flags, so somebody should check this 481 * all. But careful, it tends to be safer to return too many 482 * bits than too few, and you can easily break real applications 483 * if you don't tell them that something has hung up! 484 * 485 * Check-me. 486 * 487 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and 488 * our fs/select.c). It means that after we received EOF, 489 * poll always returns immediately, making impossible poll() on write() 490 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP 491 * if and only if shutdown has been made in both directions. 492 * Actually, it is interesting to look how Solaris and DUX 493 * solve this dilemma. I would prefer, if POLLHUP were maskable, 494 * then we could set it on SND_SHUTDOWN. BTW examples given 495 * in Stevens' books assume exactly this behaviour, it explains 496 * why POLLHUP is incompatible with POLLOUT. --ANK 497 * 498 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent 499 * blocking on fresh not-connected or disconnected socket. --ANK 500 */ 501 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE) 502 mask |= POLLHUP; 503 if (sk->sk_shutdown & RCV_SHUTDOWN) 504 mask |= POLLIN | POLLRDNORM | POLLRDHUP; 505 506 /* Connected or passive Fast Open socket? */ 507 if (state != TCP_SYN_SENT && 508 (state != TCP_SYN_RECV || tp->fastopen_rsk)) { 509 int target = sock_rcvlowat(sk, 0, INT_MAX); 510 511 if (tp->urg_seq == tp->copied_seq && 512 !sock_flag(sk, SOCK_URGINLINE) && 513 tp->urg_data) 514 target++; 515 516 if (tp->rcv_nxt - tp->copied_seq >= target) 517 mask |= POLLIN | POLLRDNORM; 518 519 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) { 520 if (sk_stream_is_writeable(sk)) { 521 mask |= POLLOUT | POLLWRNORM; 522 } else { /* send SIGIO later */ 523 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk); 524 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 525 526 /* Race breaker. If space is freed after 527 * wspace test but before the flags are set, 528 * IO signal will be lost. Memory barrier 529 * pairs with the input side. 530 */ 531 smp_mb__after_atomic(); 532 if (sk_stream_is_writeable(sk)) 533 mask |= POLLOUT | POLLWRNORM; 534 } 535 } else 536 mask |= POLLOUT | POLLWRNORM; 537 538 if (tp->urg_data & TCP_URG_VALID) 539 mask |= POLLPRI; 540 } 541 /* This barrier is coupled with smp_wmb() in tcp_reset() */ 542 smp_rmb(); 543 if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue)) 544 mask |= POLLERR; 545 546 return mask; 547 } 548 EXPORT_SYMBOL(tcp_poll); 549 550 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg) 551 { 552 struct tcp_sock *tp = tcp_sk(sk); 553 int answ; 554 bool slow; 555 556 switch (cmd) { 557 case SIOCINQ: 558 if (sk->sk_state == TCP_LISTEN) 559 return -EINVAL; 560 561 slow = lock_sock_fast(sk); 562 answ = tcp_inq(sk); 563 unlock_sock_fast(sk, slow); 564 break; 565 case SIOCATMARK: 566 answ = tp->urg_data && tp->urg_seq == tp->copied_seq; 567 break; 568 case SIOCOUTQ: 569 if (sk->sk_state == TCP_LISTEN) 570 return -EINVAL; 571 572 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 573 answ = 0; 574 else 575 answ = tp->write_seq - tp->snd_una; 576 break; 577 case SIOCOUTQNSD: 578 if (sk->sk_state == TCP_LISTEN) 579 return -EINVAL; 580 581 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 582 answ = 0; 583 else 584 answ = tp->write_seq - tp->snd_nxt; 585 break; 586 default: 587 return -ENOIOCTLCMD; 588 } 589 590 return put_user(answ, (int __user *)arg); 591 } 592 EXPORT_SYMBOL(tcp_ioctl); 593 594 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb) 595 { 596 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 597 tp->pushed_seq = tp->write_seq; 598 } 599 600 static inline bool forced_push(const struct tcp_sock *tp) 601 { 602 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1)); 603 } 604 605 static void skb_entail(struct sock *sk, struct sk_buff *skb) 606 { 607 struct tcp_sock *tp = tcp_sk(sk); 608 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 609 610 skb->csum = 0; 611 tcb->seq = tcb->end_seq = tp->write_seq; 612 tcb->tcp_flags = TCPHDR_ACK; 613 tcb->sacked = 0; 614 __skb_header_release(skb); 615 tcp_add_write_queue_tail(sk, skb); 616 sk->sk_wmem_queued += skb->truesize; 617 sk_mem_charge(sk, skb->truesize); 618 if (tp->nonagle & TCP_NAGLE_PUSH) 619 tp->nonagle &= ~TCP_NAGLE_PUSH; 620 621 tcp_slow_start_after_idle_check(sk); 622 } 623 624 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags) 625 { 626 if (flags & MSG_OOB) 627 tp->snd_up = tp->write_seq; 628 } 629 630 /* If a not yet filled skb is pushed, do not send it if 631 * we have data packets in Qdisc or NIC queues : 632 * Because TX completion will happen shortly, it gives a chance 633 * to coalesce future sendmsg() payload into this skb, without 634 * need for a timer, and with no latency trade off. 635 * As packets containing data payload have a bigger truesize 636 * than pure acks (dataless) packets, the last checks prevent 637 * autocorking if we only have an ACK in Qdisc/NIC queues, 638 * or if TX completion was delayed after we processed ACK packet. 639 */ 640 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb, 641 int size_goal) 642 { 643 return skb->len < size_goal && 644 sysctl_tcp_autocorking && 645 skb != tcp_write_queue_head(sk) && 646 atomic_read(&sk->sk_wmem_alloc) > skb->truesize; 647 } 648 649 static void tcp_push(struct sock *sk, int flags, int mss_now, 650 int nonagle, int size_goal) 651 { 652 struct tcp_sock *tp = tcp_sk(sk); 653 struct sk_buff *skb; 654 655 if (!tcp_send_head(sk)) 656 return; 657 658 skb = tcp_write_queue_tail(sk); 659 if (!(flags & MSG_MORE) || forced_push(tp)) 660 tcp_mark_push(tp, skb); 661 662 tcp_mark_urg(tp, flags); 663 664 if (tcp_should_autocork(sk, skb, size_goal)) { 665 666 /* avoid atomic op if TSQ_THROTTLED bit is already set */ 667 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) { 668 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING); 669 set_bit(TSQ_THROTTLED, &tp->tsq_flags); 670 } 671 /* It is possible TX completion already happened 672 * before we set TSQ_THROTTLED. 673 */ 674 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize) 675 return; 676 } 677 678 if (flags & MSG_MORE) 679 nonagle = TCP_NAGLE_CORK; 680 681 __tcp_push_pending_frames(sk, mss_now, nonagle); 682 } 683 684 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, 685 unsigned int offset, size_t len) 686 { 687 struct tcp_splice_state *tss = rd_desc->arg.data; 688 int ret; 689 690 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe, 691 min(rd_desc->count, len), tss->flags, 692 skb_socket_splice); 693 if (ret > 0) 694 rd_desc->count -= ret; 695 return ret; 696 } 697 698 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss) 699 { 700 /* Store TCP splice context information in read_descriptor_t. */ 701 read_descriptor_t rd_desc = { 702 .arg.data = tss, 703 .count = tss->len, 704 }; 705 706 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv); 707 } 708 709 /** 710 * tcp_splice_read - splice data from TCP socket to a pipe 711 * @sock: socket to splice from 712 * @ppos: position (not valid) 713 * @pipe: pipe to splice to 714 * @len: number of bytes to splice 715 * @flags: splice modifier flags 716 * 717 * Description: 718 * Will read pages from given socket and fill them into a pipe. 719 * 720 **/ 721 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos, 722 struct pipe_inode_info *pipe, size_t len, 723 unsigned int flags) 724 { 725 struct sock *sk = sock->sk; 726 struct tcp_splice_state tss = { 727 .pipe = pipe, 728 .len = len, 729 .flags = flags, 730 }; 731 long timeo; 732 ssize_t spliced; 733 int ret; 734 735 sock_rps_record_flow(sk); 736 /* 737 * We can't seek on a socket input 738 */ 739 if (unlikely(*ppos)) 740 return -ESPIPE; 741 742 ret = spliced = 0; 743 744 lock_sock(sk); 745 746 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK); 747 while (tss.len) { 748 ret = __tcp_splice_read(sk, &tss); 749 if (ret < 0) 750 break; 751 else if (!ret) { 752 if (spliced) 753 break; 754 if (sock_flag(sk, SOCK_DONE)) 755 break; 756 if (sk->sk_err) { 757 ret = sock_error(sk); 758 break; 759 } 760 if (sk->sk_shutdown & RCV_SHUTDOWN) 761 break; 762 if (sk->sk_state == TCP_CLOSE) { 763 /* 764 * This occurs when user tries to read 765 * from never connected socket. 766 */ 767 if (!sock_flag(sk, SOCK_DONE)) 768 ret = -ENOTCONN; 769 break; 770 } 771 if (!timeo) { 772 ret = -EAGAIN; 773 break; 774 } 775 sk_wait_data(sk, &timeo, NULL); 776 if (signal_pending(current)) { 777 ret = sock_intr_errno(timeo); 778 break; 779 } 780 continue; 781 } 782 tss.len -= ret; 783 spliced += ret; 784 785 if (!timeo) 786 break; 787 release_sock(sk); 788 lock_sock(sk); 789 790 if (sk->sk_err || sk->sk_state == TCP_CLOSE || 791 (sk->sk_shutdown & RCV_SHUTDOWN) || 792 signal_pending(current)) 793 break; 794 } 795 796 release_sock(sk); 797 798 if (spliced) 799 return spliced; 800 801 return ret; 802 } 803 EXPORT_SYMBOL(tcp_splice_read); 804 805 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp, 806 bool force_schedule) 807 { 808 struct sk_buff *skb; 809 810 /* The TCP header must be at least 32-bit aligned. */ 811 size = ALIGN(size, 4); 812 813 if (unlikely(tcp_under_memory_pressure(sk))) 814 sk_mem_reclaim_partial(sk); 815 816 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp); 817 if (likely(skb)) { 818 bool mem_scheduled; 819 820 if (force_schedule) { 821 mem_scheduled = true; 822 sk_forced_mem_schedule(sk, skb->truesize); 823 } else { 824 mem_scheduled = sk_wmem_schedule(sk, skb->truesize); 825 } 826 if (likely(mem_scheduled)) { 827 skb_reserve(skb, sk->sk_prot->max_header); 828 /* 829 * Make sure that we have exactly size bytes 830 * available to the caller, no more, no less. 831 */ 832 skb->reserved_tailroom = skb->end - skb->tail - size; 833 return skb; 834 } 835 __kfree_skb(skb); 836 } else { 837 sk->sk_prot->enter_memory_pressure(sk); 838 sk_stream_moderate_sndbuf(sk); 839 } 840 return NULL; 841 } 842 843 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now, 844 int large_allowed) 845 { 846 struct tcp_sock *tp = tcp_sk(sk); 847 u32 new_size_goal, size_goal; 848 849 if (!large_allowed || !sk_can_gso(sk)) 850 return mss_now; 851 852 /* Note : tcp_tso_autosize() will eventually split this later */ 853 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER; 854 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal); 855 856 /* We try hard to avoid divides here */ 857 size_goal = tp->gso_segs * mss_now; 858 if (unlikely(new_size_goal < size_goal || 859 new_size_goal >= size_goal + mss_now)) { 860 tp->gso_segs = min_t(u16, new_size_goal / mss_now, 861 sk->sk_gso_max_segs); 862 size_goal = tp->gso_segs * mss_now; 863 } 864 865 return max(size_goal, mss_now); 866 } 867 868 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags) 869 { 870 int mss_now; 871 872 mss_now = tcp_current_mss(sk); 873 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB)); 874 875 return mss_now; 876 } 877 878 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset, 879 size_t size, int flags) 880 { 881 struct tcp_sock *tp = tcp_sk(sk); 882 int mss_now, size_goal; 883 int err; 884 ssize_t copied; 885 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 886 887 /* Wait for a connection to finish. One exception is TCP Fast Open 888 * (passive side) where data is allowed to be sent before a connection 889 * is fully established. 890 */ 891 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) && 892 !tcp_passive_fastopen(sk)) { 893 err = sk_stream_wait_connect(sk, &timeo); 894 if (err != 0) 895 goto out_err; 896 } 897 898 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk); 899 900 mss_now = tcp_send_mss(sk, &size_goal, flags); 901 copied = 0; 902 903 err = -EPIPE; 904 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 905 goto out_err; 906 907 while (size > 0) { 908 struct sk_buff *skb = tcp_write_queue_tail(sk); 909 int copy, i; 910 bool can_coalesce; 911 912 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0 || 913 !tcp_skb_can_collapse_to(skb)) { 914 new_segment: 915 if (!sk_stream_memory_free(sk)) 916 goto wait_for_sndbuf; 917 918 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, 919 skb_queue_empty(&sk->sk_write_queue)); 920 if (!skb) 921 goto wait_for_memory; 922 923 skb_entail(sk, skb); 924 copy = size_goal; 925 } 926 927 if (copy > size) 928 copy = size; 929 930 i = skb_shinfo(skb)->nr_frags; 931 can_coalesce = skb_can_coalesce(skb, i, page, offset); 932 if (!can_coalesce && i >= sysctl_max_skb_frags) { 933 tcp_mark_push(tp, skb); 934 goto new_segment; 935 } 936 if (!sk_wmem_schedule(sk, copy)) 937 goto wait_for_memory; 938 939 if (can_coalesce) { 940 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 941 } else { 942 get_page(page); 943 skb_fill_page_desc(skb, i, page, offset, copy); 944 } 945 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG; 946 947 skb->len += copy; 948 skb->data_len += copy; 949 skb->truesize += copy; 950 sk->sk_wmem_queued += copy; 951 sk_mem_charge(sk, copy); 952 skb->ip_summed = CHECKSUM_PARTIAL; 953 tp->write_seq += copy; 954 TCP_SKB_CB(skb)->end_seq += copy; 955 tcp_skb_pcount_set(skb, 0); 956 957 if (!copied) 958 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 959 960 copied += copy; 961 offset += copy; 962 size -= copy; 963 if (!size) { 964 tcp_tx_timestamp(sk, sk->sk_tsflags, skb); 965 goto out; 966 } 967 968 if (skb->len < size_goal || (flags & MSG_OOB)) 969 continue; 970 971 if (forced_push(tp)) { 972 tcp_mark_push(tp, skb); 973 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 974 } else if (skb == tcp_send_head(sk)) 975 tcp_push_one(sk, mss_now); 976 continue; 977 978 wait_for_sndbuf: 979 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 980 wait_for_memory: 981 tcp_push(sk, flags & ~MSG_MORE, mss_now, 982 TCP_NAGLE_PUSH, size_goal); 983 984 err = sk_stream_wait_memory(sk, &timeo); 985 if (err != 0) 986 goto do_error; 987 988 mss_now = tcp_send_mss(sk, &size_goal, flags); 989 } 990 991 out: 992 if (copied && !(flags & MSG_SENDPAGE_NOTLAST)) 993 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal); 994 return copied; 995 996 do_error: 997 if (copied) 998 goto out; 999 out_err: 1000 /* make sure we wake any epoll edge trigger waiter */ 1001 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN)) 1002 sk->sk_write_space(sk); 1003 return sk_stream_error(sk, flags, err); 1004 } 1005 1006 int tcp_sendpage(struct sock *sk, struct page *page, int offset, 1007 size_t size, int flags) 1008 { 1009 ssize_t res; 1010 1011 if (!(sk->sk_route_caps & NETIF_F_SG) || 1012 !sk_check_csum_caps(sk)) 1013 return sock_no_sendpage(sk->sk_socket, page, offset, size, 1014 flags); 1015 1016 lock_sock(sk); 1017 res = do_tcp_sendpages(sk, page, offset, size, flags); 1018 release_sock(sk); 1019 return res; 1020 } 1021 EXPORT_SYMBOL(tcp_sendpage); 1022 1023 static inline int select_size(const struct sock *sk, bool sg) 1024 { 1025 const struct tcp_sock *tp = tcp_sk(sk); 1026 int tmp = tp->mss_cache; 1027 1028 if (sg) { 1029 if (sk_can_gso(sk)) { 1030 /* Small frames wont use a full page: 1031 * Payload will immediately follow tcp header. 1032 */ 1033 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER); 1034 } else { 1035 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER); 1036 1037 if (tmp >= pgbreak && 1038 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE) 1039 tmp = pgbreak; 1040 } 1041 } 1042 1043 return tmp; 1044 } 1045 1046 void tcp_free_fastopen_req(struct tcp_sock *tp) 1047 { 1048 if (tp->fastopen_req) { 1049 kfree(tp->fastopen_req); 1050 tp->fastopen_req = NULL; 1051 } 1052 } 1053 1054 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, 1055 int *copied, size_t size) 1056 { 1057 struct tcp_sock *tp = tcp_sk(sk); 1058 int err, flags; 1059 1060 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE)) 1061 return -EOPNOTSUPP; 1062 if (tp->fastopen_req) 1063 return -EALREADY; /* Another Fast Open is in progress */ 1064 1065 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request), 1066 sk->sk_allocation); 1067 if (unlikely(!tp->fastopen_req)) 1068 return -ENOBUFS; 1069 tp->fastopen_req->data = msg; 1070 tp->fastopen_req->size = size; 1071 1072 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0; 1073 err = __inet_stream_connect(sk->sk_socket, msg->msg_name, 1074 msg->msg_namelen, flags); 1075 *copied = tp->fastopen_req->copied; 1076 tcp_free_fastopen_req(tp); 1077 return err; 1078 } 1079 1080 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size) 1081 { 1082 struct tcp_sock *tp = tcp_sk(sk); 1083 struct sk_buff *skb; 1084 struct sockcm_cookie sockc; 1085 int flags, err, copied = 0; 1086 int mss_now = 0, size_goal, copied_syn = 0; 1087 bool process_backlog = false; 1088 bool sg; 1089 long timeo; 1090 1091 lock_sock(sk); 1092 1093 flags = msg->msg_flags; 1094 if (flags & MSG_FASTOPEN) { 1095 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size); 1096 if (err == -EINPROGRESS && copied_syn > 0) 1097 goto out; 1098 else if (err) 1099 goto out_err; 1100 } 1101 1102 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 1103 1104 /* Wait for a connection to finish. One exception is TCP Fast Open 1105 * (passive side) where data is allowed to be sent before a connection 1106 * is fully established. 1107 */ 1108 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) && 1109 !tcp_passive_fastopen(sk)) { 1110 err = sk_stream_wait_connect(sk, &timeo); 1111 if (err != 0) 1112 goto do_error; 1113 } 1114 1115 if (unlikely(tp->repair)) { 1116 if (tp->repair_queue == TCP_RECV_QUEUE) { 1117 copied = tcp_send_rcvq(sk, msg, size); 1118 goto out_nopush; 1119 } 1120 1121 err = -EINVAL; 1122 if (tp->repair_queue == TCP_NO_QUEUE) 1123 goto out_err; 1124 1125 /* 'common' sending to sendq */ 1126 } 1127 1128 sockc.tsflags = sk->sk_tsflags; 1129 if (msg->msg_controllen) { 1130 err = sock_cmsg_send(sk, msg, &sockc); 1131 if (unlikely(err)) { 1132 err = -EINVAL; 1133 goto out_err; 1134 } 1135 } 1136 1137 /* This should be in poll */ 1138 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk); 1139 1140 /* Ok commence sending. */ 1141 copied = 0; 1142 1143 restart: 1144 mss_now = tcp_send_mss(sk, &size_goal, flags); 1145 1146 err = -EPIPE; 1147 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 1148 goto out_err; 1149 1150 sg = !!(sk->sk_route_caps & NETIF_F_SG); 1151 1152 while (msg_data_left(msg)) { 1153 int copy = 0; 1154 int max = size_goal; 1155 1156 skb = tcp_write_queue_tail(sk); 1157 if (tcp_send_head(sk)) { 1158 if (skb->ip_summed == CHECKSUM_NONE) 1159 max = mss_now; 1160 copy = max - skb->len; 1161 } 1162 1163 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) { 1164 new_segment: 1165 /* Allocate new segment. If the interface is SG, 1166 * allocate skb fitting to single page. 1167 */ 1168 if (!sk_stream_memory_free(sk)) 1169 goto wait_for_sndbuf; 1170 1171 if (process_backlog && sk_flush_backlog(sk)) { 1172 process_backlog = false; 1173 goto restart; 1174 } 1175 skb = sk_stream_alloc_skb(sk, 1176 select_size(sk, sg), 1177 sk->sk_allocation, 1178 skb_queue_empty(&sk->sk_write_queue)); 1179 if (!skb) 1180 goto wait_for_memory; 1181 1182 process_backlog = true; 1183 /* 1184 * Check whether we can use HW checksum. 1185 */ 1186 if (sk_check_csum_caps(sk)) 1187 skb->ip_summed = CHECKSUM_PARTIAL; 1188 1189 skb_entail(sk, skb); 1190 copy = size_goal; 1191 max = size_goal; 1192 1193 /* All packets are restored as if they have 1194 * already been sent. skb_mstamp isn't set to 1195 * avoid wrong rtt estimation. 1196 */ 1197 if (tp->repair) 1198 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED; 1199 } 1200 1201 /* Try to append data to the end of skb. */ 1202 if (copy > msg_data_left(msg)) 1203 copy = msg_data_left(msg); 1204 1205 /* Where to copy to? */ 1206 if (skb_availroom(skb) > 0) { 1207 /* We have some space in skb head. Superb! */ 1208 copy = min_t(int, copy, skb_availroom(skb)); 1209 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy); 1210 if (err) 1211 goto do_fault; 1212 } else { 1213 bool merge = true; 1214 int i = skb_shinfo(skb)->nr_frags; 1215 struct page_frag *pfrag = sk_page_frag(sk); 1216 1217 if (!sk_page_frag_refill(sk, pfrag)) 1218 goto wait_for_memory; 1219 1220 if (!skb_can_coalesce(skb, i, pfrag->page, 1221 pfrag->offset)) { 1222 if (i == sysctl_max_skb_frags || !sg) { 1223 tcp_mark_push(tp, skb); 1224 goto new_segment; 1225 } 1226 merge = false; 1227 } 1228 1229 copy = min_t(int, copy, pfrag->size - pfrag->offset); 1230 1231 if (!sk_wmem_schedule(sk, copy)) 1232 goto wait_for_memory; 1233 1234 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb, 1235 pfrag->page, 1236 pfrag->offset, 1237 copy); 1238 if (err) 1239 goto do_error; 1240 1241 /* Update the skb. */ 1242 if (merge) { 1243 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 1244 } else { 1245 skb_fill_page_desc(skb, i, pfrag->page, 1246 pfrag->offset, copy); 1247 get_page(pfrag->page); 1248 } 1249 pfrag->offset += copy; 1250 } 1251 1252 if (!copied) 1253 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 1254 1255 tp->write_seq += copy; 1256 TCP_SKB_CB(skb)->end_seq += copy; 1257 tcp_skb_pcount_set(skb, 0); 1258 1259 copied += copy; 1260 if (!msg_data_left(msg)) { 1261 tcp_tx_timestamp(sk, sockc.tsflags, skb); 1262 if (unlikely(flags & MSG_EOR)) 1263 TCP_SKB_CB(skb)->eor = 1; 1264 goto out; 1265 } 1266 1267 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair)) 1268 continue; 1269 1270 if (forced_push(tp)) { 1271 tcp_mark_push(tp, skb); 1272 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 1273 } else if (skb == tcp_send_head(sk)) 1274 tcp_push_one(sk, mss_now); 1275 continue; 1276 1277 wait_for_sndbuf: 1278 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1279 wait_for_memory: 1280 if (copied) 1281 tcp_push(sk, flags & ~MSG_MORE, mss_now, 1282 TCP_NAGLE_PUSH, size_goal); 1283 1284 err = sk_stream_wait_memory(sk, &timeo); 1285 if (err != 0) 1286 goto do_error; 1287 1288 mss_now = tcp_send_mss(sk, &size_goal, flags); 1289 } 1290 1291 out: 1292 if (copied) 1293 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal); 1294 out_nopush: 1295 release_sock(sk); 1296 return copied + copied_syn; 1297 1298 do_fault: 1299 if (!skb->len) { 1300 tcp_unlink_write_queue(skb, sk); 1301 /* It is the one place in all of TCP, except connection 1302 * reset, where we can be unlinking the send_head. 1303 */ 1304 tcp_check_send_head(sk, skb); 1305 sk_wmem_free_skb(sk, skb); 1306 } 1307 1308 do_error: 1309 if (copied + copied_syn) 1310 goto out; 1311 out_err: 1312 err = sk_stream_error(sk, flags, err); 1313 /* make sure we wake any epoll edge trigger waiter */ 1314 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN)) 1315 sk->sk_write_space(sk); 1316 release_sock(sk); 1317 return err; 1318 } 1319 EXPORT_SYMBOL(tcp_sendmsg); 1320 1321 /* 1322 * Handle reading urgent data. BSD has very simple semantics for 1323 * this, no blocking and very strange errors 8) 1324 */ 1325 1326 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags) 1327 { 1328 struct tcp_sock *tp = tcp_sk(sk); 1329 1330 /* No URG data to read. */ 1331 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data || 1332 tp->urg_data == TCP_URG_READ) 1333 return -EINVAL; /* Yes this is right ! */ 1334 1335 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE)) 1336 return -ENOTCONN; 1337 1338 if (tp->urg_data & TCP_URG_VALID) { 1339 int err = 0; 1340 char c = tp->urg_data; 1341 1342 if (!(flags & MSG_PEEK)) 1343 tp->urg_data = TCP_URG_READ; 1344 1345 /* Read urgent data. */ 1346 msg->msg_flags |= MSG_OOB; 1347 1348 if (len > 0) { 1349 if (!(flags & MSG_TRUNC)) 1350 err = memcpy_to_msg(msg, &c, 1); 1351 len = 1; 1352 } else 1353 msg->msg_flags |= MSG_TRUNC; 1354 1355 return err ? -EFAULT : len; 1356 } 1357 1358 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 1359 return 0; 1360 1361 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and 1362 * the available implementations agree in this case: 1363 * this call should never block, independent of the 1364 * blocking state of the socket. 1365 * Mike <pall@rz.uni-karlsruhe.de> 1366 */ 1367 return -EAGAIN; 1368 } 1369 1370 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len) 1371 { 1372 struct sk_buff *skb; 1373 int copied = 0, err = 0; 1374 1375 /* XXX -- need to support SO_PEEK_OFF */ 1376 1377 skb_queue_walk(&sk->sk_write_queue, skb) { 1378 err = skb_copy_datagram_msg(skb, 0, msg, skb->len); 1379 if (err) 1380 break; 1381 1382 copied += skb->len; 1383 } 1384 1385 return err ?: copied; 1386 } 1387 1388 /* Clean up the receive buffer for full frames taken by the user, 1389 * then send an ACK if necessary. COPIED is the number of bytes 1390 * tcp_recvmsg has given to the user so far, it speeds up the 1391 * calculation of whether or not we must ACK for the sake of 1392 * a window update. 1393 */ 1394 static void tcp_cleanup_rbuf(struct sock *sk, int copied) 1395 { 1396 struct tcp_sock *tp = tcp_sk(sk); 1397 bool time_to_ack = false; 1398 1399 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 1400 1401 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq), 1402 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n", 1403 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt); 1404 1405 if (inet_csk_ack_scheduled(sk)) { 1406 const struct inet_connection_sock *icsk = inet_csk(sk); 1407 /* Delayed ACKs frequently hit locked sockets during bulk 1408 * receive. */ 1409 if (icsk->icsk_ack.blocked || 1410 /* Once-per-two-segments ACK was not sent by tcp_input.c */ 1411 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss || 1412 /* 1413 * If this read emptied read buffer, we send ACK, if 1414 * connection is not bidirectional, user drained 1415 * receive buffer and there was a small segment 1416 * in queue. 1417 */ 1418 (copied > 0 && 1419 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) || 1420 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) && 1421 !icsk->icsk_ack.pingpong)) && 1422 !atomic_read(&sk->sk_rmem_alloc))) 1423 time_to_ack = true; 1424 } 1425 1426 /* We send an ACK if we can now advertise a non-zero window 1427 * which has been raised "significantly". 1428 * 1429 * Even if window raised up to infinity, do not send window open ACK 1430 * in states, where we will not receive more. It is useless. 1431 */ 1432 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) { 1433 __u32 rcv_window_now = tcp_receive_window(tp); 1434 1435 /* Optimize, __tcp_select_window() is not cheap. */ 1436 if (2*rcv_window_now <= tp->window_clamp) { 1437 __u32 new_window = __tcp_select_window(sk); 1438 1439 /* Send ACK now, if this read freed lots of space 1440 * in our buffer. Certainly, new_window is new window. 1441 * We can advertise it now, if it is not less than current one. 1442 * "Lots" means "at least twice" here. 1443 */ 1444 if (new_window && new_window >= 2 * rcv_window_now) 1445 time_to_ack = true; 1446 } 1447 } 1448 if (time_to_ack) 1449 tcp_send_ack(sk); 1450 } 1451 1452 static void tcp_prequeue_process(struct sock *sk) 1453 { 1454 struct sk_buff *skb; 1455 struct tcp_sock *tp = tcp_sk(sk); 1456 1457 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPREQUEUED); 1458 1459 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL) 1460 sk_backlog_rcv(sk, skb); 1461 1462 /* Clear memory counter. */ 1463 tp->ucopy.memory = 0; 1464 } 1465 1466 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off) 1467 { 1468 struct sk_buff *skb; 1469 u32 offset; 1470 1471 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) { 1472 offset = seq - TCP_SKB_CB(skb)->seq; 1473 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 1474 pr_err_once("%s: found a SYN, please report !\n", __func__); 1475 offset--; 1476 } 1477 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) { 1478 *off = offset; 1479 return skb; 1480 } 1481 /* This looks weird, but this can happen if TCP collapsing 1482 * splitted a fat GRO packet, while we released socket lock 1483 * in skb_splice_bits() 1484 */ 1485 sk_eat_skb(sk, skb); 1486 } 1487 return NULL; 1488 } 1489 1490 /* 1491 * This routine provides an alternative to tcp_recvmsg() for routines 1492 * that would like to handle copying from skbuffs directly in 'sendfile' 1493 * fashion. 1494 * Note: 1495 * - It is assumed that the socket was locked by the caller. 1496 * - The routine does not block. 1497 * - At present, there is no support for reading OOB data 1498 * or for 'peeking' the socket using this routine 1499 * (although both would be easy to implement). 1500 */ 1501 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, 1502 sk_read_actor_t recv_actor) 1503 { 1504 struct sk_buff *skb; 1505 struct tcp_sock *tp = tcp_sk(sk); 1506 u32 seq = tp->copied_seq; 1507 u32 offset; 1508 int copied = 0; 1509 1510 if (sk->sk_state == TCP_LISTEN) 1511 return -ENOTCONN; 1512 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) { 1513 if (offset < skb->len) { 1514 int used; 1515 size_t len; 1516 1517 len = skb->len - offset; 1518 /* Stop reading if we hit a patch of urgent data */ 1519 if (tp->urg_data) { 1520 u32 urg_offset = tp->urg_seq - seq; 1521 if (urg_offset < len) 1522 len = urg_offset; 1523 if (!len) 1524 break; 1525 } 1526 used = recv_actor(desc, skb, offset, len); 1527 if (used <= 0) { 1528 if (!copied) 1529 copied = used; 1530 break; 1531 } else if (used <= len) { 1532 seq += used; 1533 copied += used; 1534 offset += used; 1535 } 1536 /* If recv_actor drops the lock (e.g. TCP splice 1537 * receive) the skb pointer might be invalid when 1538 * getting here: tcp_collapse might have deleted it 1539 * while aggregating skbs from the socket queue. 1540 */ 1541 skb = tcp_recv_skb(sk, seq - 1, &offset); 1542 if (!skb) 1543 break; 1544 /* TCP coalescing might have appended data to the skb. 1545 * Try to splice more frags 1546 */ 1547 if (offset + 1 != skb->len) 1548 continue; 1549 } 1550 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) { 1551 sk_eat_skb(sk, skb); 1552 ++seq; 1553 break; 1554 } 1555 sk_eat_skb(sk, skb); 1556 if (!desc->count) 1557 break; 1558 tp->copied_seq = seq; 1559 } 1560 tp->copied_seq = seq; 1561 1562 tcp_rcv_space_adjust(sk); 1563 1564 /* Clean up data we have read: This will do ACK frames. */ 1565 if (copied > 0) { 1566 tcp_recv_skb(sk, seq, &offset); 1567 tcp_cleanup_rbuf(sk, copied); 1568 } 1569 return copied; 1570 } 1571 EXPORT_SYMBOL(tcp_read_sock); 1572 1573 /* 1574 * This routine copies from a sock struct into the user buffer. 1575 * 1576 * Technical note: in 2.3 we work on _locked_ socket, so that 1577 * tricks with *seq access order and skb->users are not required. 1578 * Probably, code can be easily improved even more. 1579 */ 1580 1581 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock, 1582 int flags, int *addr_len) 1583 { 1584 struct tcp_sock *tp = tcp_sk(sk); 1585 int copied = 0; 1586 u32 peek_seq; 1587 u32 *seq; 1588 unsigned long used; 1589 int err; 1590 int target; /* Read at least this many bytes */ 1591 long timeo; 1592 struct task_struct *user_recv = NULL; 1593 struct sk_buff *skb, *last; 1594 u32 urg_hole = 0; 1595 1596 if (unlikely(flags & MSG_ERRQUEUE)) 1597 return inet_recv_error(sk, msg, len, addr_len); 1598 1599 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) && 1600 (sk->sk_state == TCP_ESTABLISHED)) 1601 sk_busy_loop(sk, nonblock); 1602 1603 lock_sock(sk); 1604 1605 err = -ENOTCONN; 1606 if (sk->sk_state == TCP_LISTEN) 1607 goto out; 1608 1609 timeo = sock_rcvtimeo(sk, nonblock); 1610 1611 /* Urgent data needs to be handled specially. */ 1612 if (flags & MSG_OOB) 1613 goto recv_urg; 1614 1615 if (unlikely(tp->repair)) { 1616 err = -EPERM; 1617 if (!(flags & MSG_PEEK)) 1618 goto out; 1619 1620 if (tp->repair_queue == TCP_SEND_QUEUE) 1621 goto recv_sndq; 1622 1623 err = -EINVAL; 1624 if (tp->repair_queue == TCP_NO_QUEUE) 1625 goto out; 1626 1627 /* 'common' recv queue MSG_PEEK-ing */ 1628 } 1629 1630 seq = &tp->copied_seq; 1631 if (flags & MSG_PEEK) { 1632 peek_seq = tp->copied_seq; 1633 seq = &peek_seq; 1634 } 1635 1636 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 1637 1638 do { 1639 u32 offset; 1640 1641 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */ 1642 if (tp->urg_data && tp->urg_seq == *seq) { 1643 if (copied) 1644 break; 1645 if (signal_pending(current)) { 1646 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN; 1647 break; 1648 } 1649 } 1650 1651 /* Next get a buffer. */ 1652 1653 last = skb_peek_tail(&sk->sk_receive_queue); 1654 skb_queue_walk(&sk->sk_receive_queue, skb) { 1655 last = skb; 1656 /* Now that we have two receive queues this 1657 * shouldn't happen. 1658 */ 1659 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq), 1660 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n", 1661 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, 1662 flags)) 1663 break; 1664 1665 offset = *seq - TCP_SKB_CB(skb)->seq; 1666 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 1667 pr_err_once("%s: found a SYN, please report !\n", __func__); 1668 offset--; 1669 } 1670 if (offset < skb->len) 1671 goto found_ok_skb; 1672 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1673 goto found_fin_ok; 1674 WARN(!(flags & MSG_PEEK), 1675 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n", 1676 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags); 1677 } 1678 1679 /* Well, if we have backlog, try to process it now yet. */ 1680 1681 if (copied >= target && !sk->sk_backlog.tail) 1682 break; 1683 1684 if (copied) { 1685 if (sk->sk_err || 1686 sk->sk_state == TCP_CLOSE || 1687 (sk->sk_shutdown & RCV_SHUTDOWN) || 1688 !timeo || 1689 signal_pending(current)) 1690 break; 1691 } else { 1692 if (sock_flag(sk, SOCK_DONE)) 1693 break; 1694 1695 if (sk->sk_err) { 1696 copied = sock_error(sk); 1697 break; 1698 } 1699 1700 if (sk->sk_shutdown & RCV_SHUTDOWN) 1701 break; 1702 1703 if (sk->sk_state == TCP_CLOSE) { 1704 if (!sock_flag(sk, SOCK_DONE)) { 1705 /* This occurs when user tries to read 1706 * from never connected socket. 1707 */ 1708 copied = -ENOTCONN; 1709 break; 1710 } 1711 break; 1712 } 1713 1714 if (!timeo) { 1715 copied = -EAGAIN; 1716 break; 1717 } 1718 1719 if (signal_pending(current)) { 1720 copied = sock_intr_errno(timeo); 1721 break; 1722 } 1723 } 1724 1725 tcp_cleanup_rbuf(sk, copied); 1726 1727 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) { 1728 /* Install new reader */ 1729 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) { 1730 user_recv = current; 1731 tp->ucopy.task = user_recv; 1732 tp->ucopy.msg = msg; 1733 } 1734 1735 tp->ucopy.len = len; 1736 1737 WARN_ON(tp->copied_seq != tp->rcv_nxt && 1738 !(flags & (MSG_PEEK | MSG_TRUNC))); 1739 1740 /* Ugly... If prequeue is not empty, we have to 1741 * process it before releasing socket, otherwise 1742 * order will be broken at second iteration. 1743 * More elegant solution is required!!! 1744 * 1745 * Look: we have the following (pseudo)queues: 1746 * 1747 * 1. packets in flight 1748 * 2. backlog 1749 * 3. prequeue 1750 * 4. receive_queue 1751 * 1752 * Each queue can be processed only if the next ones 1753 * are empty. At this point we have empty receive_queue. 1754 * But prequeue _can_ be not empty after 2nd iteration, 1755 * when we jumped to start of loop because backlog 1756 * processing added something to receive_queue. 1757 * We cannot release_sock(), because backlog contains 1758 * packets arrived _after_ prequeued ones. 1759 * 1760 * Shortly, algorithm is clear --- to process all 1761 * the queues in order. We could make it more directly, 1762 * requeueing packets from backlog to prequeue, if 1763 * is not empty. It is more elegant, but eats cycles, 1764 * unfortunately. 1765 */ 1766 if (!skb_queue_empty(&tp->ucopy.prequeue)) 1767 goto do_prequeue; 1768 1769 /* __ Set realtime policy in scheduler __ */ 1770 } 1771 1772 if (copied >= target) { 1773 /* Do not sleep, just process backlog. */ 1774 release_sock(sk); 1775 lock_sock(sk); 1776 } else { 1777 sk_wait_data(sk, &timeo, last); 1778 } 1779 1780 if (user_recv) { 1781 int chunk; 1782 1783 /* __ Restore normal policy in scheduler __ */ 1784 1785 chunk = len - tp->ucopy.len; 1786 if (chunk != 0) { 1787 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk); 1788 len -= chunk; 1789 copied += chunk; 1790 } 1791 1792 if (tp->rcv_nxt == tp->copied_seq && 1793 !skb_queue_empty(&tp->ucopy.prequeue)) { 1794 do_prequeue: 1795 tcp_prequeue_process(sk); 1796 1797 chunk = len - tp->ucopy.len; 1798 if (chunk != 0) { 1799 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1800 len -= chunk; 1801 copied += chunk; 1802 } 1803 } 1804 } 1805 if ((flags & MSG_PEEK) && 1806 (peek_seq - copied - urg_hole != tp->copied_seq)) { 1807 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n", 1808 current->comm, 1809 task_pid_nr(current)); 1810 peek_seq = tp->copied_seq; 1811 } 1812 continue; 1813 1814 found_ok_skb: 1815 /* Ok so how much can we use? */ 1816 used = skb->len - offset; 1817 if (len < used) 1818 used = len; 1819 1820 /* Do we have urgent data here? */ 1821 if (tp->urg_data) { 1822 u32 urg_offset = tp->urg_seq - *seq; 1823 if (urg_offset < used) { 1824 if (!urg_offset) { 1825 if (!sock_flag(sk, SOCK_URGINLINE)) { 1826 ++*seq; 1827 urg_hole++; 1828 offset++; 1829 used--; 1830 if (!used) 1831 goto skip_copy; 1832 } 1833 } else 1834 used = urg_offset; 1835 } 1836 } 1837 1838 if (!(flags & MSG_TRUNC)) { 1839 err = skb_copy_datagram_msg(skb, offset, msg, used); 1840 if (err) { 1841 /* Exception. Bailout! */ 1842 if (!copied) 1843 copied = -EFAULT; 1844 break; 1845 } 1846 } 1847 1848 *seq += used; 1849 copied += used; 1850 len -= used; 1851 1852 tcp_rcv_space_adjust(sk); 1853 1854 skip_copy: 1855 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) { 1856 tp->urg_data = 0; 1857 tcp_fast_path_check(sk); 1858 } 1859 if (used + offset < skb->len) 1860 continue; 1861 1862 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1863 goto found_fin_ok; 1864 if (!(flags & MSG_PEEK)) 1865 sk_eat_skb(sk, skb); 1866 continue; 1867 1868 found_fin_ok: 1869 /* Process the FIN. */ 1870 ++*seq; 1871 if (!(flags & MSG_PEEK)) 1872 sk_eat_skb(sk, skb); 1873 break; 1874 } while (len > 0); 1875 1876 if (user_recv) { 1877 if (!skb_queue_empty(&tp->ucopy.prequeue)) { 1878 int chunk; 1879 1880 tp->ucopy.len = copied > 0 ? len : 0; 1881 1882 tcp_prequeue_process(sk); 1883 1884 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) { 1885 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1886 len -= chunk; 1887 copied += chunk; 1888 } 1889 } 1890 1891 tp->ucopy.task = NULL; 1892 tp->ucopy.len = 0; 1893 } 1894 1895 /* According to UNIX98, msg_name/msg_namelen are ignored 1896 * on connected socket. I was just happy when found this 8) --ANK 1897 */ 1898 1899 /* Clean up data we have read: This will do ACK frames. */ 1900 tcp_cleanup_rbuf(sk, copied); 1901 1902 release_sock(sk); 1903 return copied; 1904 1905 out: 1906 release_sock(sk); 1907 return err; 1908 1909 recv_urg: 1910 err = tcp_recv_urg(sk, msg, len, flags); 1911 goto out; 1912 1913 recv_sndq: 1914 err = tcp_peek_sndq(sk, msg, len); 1915 goto out; 1916 } 1917 EXPORT_SYMBOL(tcp_recvmsg); 1918 1919 void tcp_set_state(struct sock *sk, int state) 1920 { 1921 int oldstate = sk->sk_state; 1922 1923 switch (state) { 1924 case TCP_ESTABLISHED: 1925 if (oldstate != TCP_ESTABLISHED) 1926 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1927 break; 1928 1929 case TCP_CLOSE: 1930 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED) 1931 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS); 1932 1933 sk->sk_prot->unhash(sk); 1934 if (inet_csk(sk)->icsk_bind_hash && 1935 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK)) 1936 inet_put_port(sk); 1937 /* fall through */ 1938 default: 1939 if (oldstate == TCP_ESTABLISHED) 1940 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1941 } 1942 1943 /* Change state AFTER socket is unhashed to avoid closed 1944 * socket sitting in hash tables. 1945 */ 1946 sk_state_store(sk, state); 1947 1948 #ifdef STATE_TRACE 1949 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]); 1950 #endif 1951 } 1952 EXPORT_SYMBOL_GPL(tcp_set_state); 1953 1954 /* 1955 * State processing on a close. This implements the state shift for 1956 * sending our FIN frame. Note that we only send a FIN for some 1957 * states. A shutdown() may have already sent the FIN, or we may be 1958 * closed. 1959 */ 1960 1961 static const unsigned char new_state[16] = { 1962 /* current state: new state: action: */ 1963 [0 /* (Invalid) */] = TCP_CLOSE, 1964 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1965 [TCP_SYN_SENT] = TCP_CLOSE, 1966 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1967 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1, 1968 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2, 1969 [TCP_TIME_WAIT] = TCP_CLOSE, 1970 [TCP_CLOSE] = TCP_CLOSE, 1971 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN, 1972 [TCP_LAST_ACK] = TCP_LAST_ACK, 1973 [TCP_LISTEN] = TCP_CLOSE, 1974 [TCP_CLOSING] = TCP_CLOSING, 1975 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */ 1976 }; 1977 1978 static int tcp_close_state(struct sock *sk) 1979 { 1980 int next = (int)new_state[sk->sk_state]; 1981 int ns = next & TCP_STATE_MASK; 1982 1983 tcp_set_state(sk, ns); 1984 1985 return next & TCP_ACTION_FIN; 1986 } 1987 1988 /* 1989 * Shutdown the sending side of a connection. Much like close except 1990 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD). 1991 */ 1992 1993 void tcp_shutdown(struct sock *sk, int how) 1994 { 1995 /* We need to grab some memory, and put together a FIN, 1996 * and then put it into the queue to be sent. 1997 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92. 1998 */ 1999 if (!(how & SEND_SHUTDOWN)) 2000 return; 2001 2002 /* If we've already sent a FIN, or it's a closed state, skip this. */ 2003 if ((1 << sk->sk_state) & 2004 (TCPF_ESTABLISHED | TCPF_SYN_SENT | 2005 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) { 2006 /* Clear out any half completed packets. FIN if needed. */ 2007 if (tcp_close_state(sk)) 2008 tcp_send_fin(sk); 2009 } 2010 } 2011 EXPORT_SYMBOL(tcp_shutdown); 2012 2013 bool tcp_check_oom(struct sock *sk, int shift) 2014 { 2015 bool too_many_orphans, out_of_socket_memory; 2016 2017 too_many_orphans = tcp_too_many_orphans(sk, shift); 2018 out_of_socket_memory = tcp_out_of_memory(sk); 2019 2020 if (too_many_orphans) 2021 net_info_ratelimited("too many orphaned sockets\n"); 2022 if (out_of_socket_memory) 2023 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n"); 2024 return too_many_orphans || out_of_socket_memory; 2025 } 2026 2027 void tcp_close(struct sock *sk, long timeout) 2028 { 2029 struct sk_buff *skb; 2030 int data_was_unread = 0; 2031 int state; 2032 2033 lock_sock(sk); 2034 sk->sk_shutdown = SHUTDOWN_MASK; 2035 2036 if (sk->sk_state == TCP_LISTEN) { 2037 tcp_set_state(sk, TCP_CLOSE); 2038 2039 /* Special case. */ 2040 inet_csk_listen_stop(sk); 2041 2042 goto adjudge_to_death; 2043 } 2044 2045 /* We need to flush the recv. buffs. We do this only on the 2046 * descriptor close, not protocol-sourced closes, because the 2047 * reader process may not have drained the data yet! 2048 */ 2049 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) { 2050 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq; 2051 2052 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 2053 len--; 2054 data_was_unread += len; 2055 __kfree_skb(skb); 2056 } 2057 2058 sk_mem_reclaim(sk); 2059 2060 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */ 2061 if (sk->sk_state == TCP_CLOSE) 2062 goto adjudge_to_death; 2063 2064 /* As outlined in RFC 2525, section 2.17, we send a RST here because 2065 * data was lost. To witness the awful effects of the old behavior of 2066 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk 2067 * GET in an FTP client, suspend the process, wait for the client to 2068 * advertise a zero window, then kill -9 the FTP client, wheee... 2069 * Note: timeout is always zero in such a case. 2070 */ 2071 if (unlikely(tcp_sk(sk)->repair)) { 2072 sk->sk_prot->disconnect(sk, 0); 2073 } else if (data_was_unread) { 2074 /* Unread data was tossed, zap the connection. */ 2075 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE); 2076 tcp_set_state(sk, TCP_CLOSE); 2077 tcp_send_active_reset(sk, sk->sk_allocation); 2078 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) { 2079 /* Check zero linger _after_ checking for unread data. */ 2080 sk->sk_prot->disconnect(sk, 0); 2081 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 2082 } else if (tcp_close_state(sk)) { 2083 /* We FIN if the application ate all the data before 2084 * zapping the connection. 2085 */ 2086 2087 /* RED-PEN. Formally speaking, we have broken TCP state 2088 * machine. State transitions: 2089 * 2090 * TCP_ESTABLISHED -> TCP_FIN_WAIT1 2091 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible) 2092 * TCP_CLOSE_WAIT -> TCP_LAST_ACK 2093 * 2094 * are legal only when FIN has been sent (i.e. in window), 2095 * rather than queued out of window. Purists blame. 2096 * 2097 * F.e. "RFC state" is ESTABLISHED, 2098 * if Linux state is FIN-WAIT-1, but FIN is still not sent. 2099 * 2100 * The visible declinations are that sometimes 2101 * we enter time-wait state, when it is not required really 2102 * (harmless), do not send active resets, when they are 2103 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when 2104 * they look as CLOSING or LAST_ACK for Linux) 2105 * Probably, I missed some more holelets. 2106 * --ANK 2107 * XXX (TFO) - To start off we don't support SYN+ACK+FIN 2108 * in a single packet! (May consider it later but will 2109 * probably need API support or TCP_CORK SYN-ACK until 2110 * data is written and socket is closed.) 2111 */ 2112 tcp_send_fin(sk); 2113 } 2114 2115 sk_stream_wait_close(sk, timeout); 2116 2117 adjudge_to_death: 2118 state = sk->sk_state; 2119 sock_hold(sk); 2120 sock_orphan(sk); 2121 2122 /* It is the last release_sock in its life. It will remove backlog. */ 2123 release_sock(sk); 2124 2125 2126 /* Now socket is owned by kernel and we acquire BH lock 2127 to finish close. No need to check for user refs. 2128 */ 2129 local_bh_disable(); 2130 bh_lock_sock(sk); 2131 WARN_ON(sock_owned_by_user(sk)); 2132 2133 percpu_counter_inc(sk->sk_prot->orphan_count); 2134 2135 /* Have we already been destroyed by a softirq or backlog? */ 2136 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE) 2137 goto out; 2138 2139 /* This is a (useful) BSD violating of the RFC. There is a 2140 * problem with TCP as specified in that the other end could 2141 * keep a socket open forever with no application left this end. 2142 * We use a 1 minute timeout (about the same as BSD) then kill 2143 * our end. If they send after that then tough - BUT: long enough 2144 * that we won't make the old 4*rto = almost no time - whoops 2145 * reset mistake. 2146 * 2147 * Nope, it was not mistake. It is really desired behaviour 2148 * f.e. on http servers, when such sockets are useless, but 2149 * consume significant resources. Let's do it with special 2150 * linger2 option. --ANK 2151 */ 2152 2153 if (sk->sk_state == TCP_FIN_WAIT2) { 2154 struct tcp_sock *tp = tcp_sk(sk); 2155 if (tp->linger2 < 0) { 2156 tcp_set_state(sk, TCP_CLOSE); 2157 tcp_send_active_reset(sk, GFP_ATOMIC); 2158 __NET_INC_STATS(sock_net(sk), 2159 LINUX_MIB_TCPABORTONLINGER); 2160 } else { 2161 const int tmo = tcp_fin_time(sk); 2162 2163 if (tmo > TCP_TIMEWAIT_LEN) { 2164 inet_csk_reset_keepalive_timer(sk, 2165 tmo - TCP_TIMEWAIT_LEN); 2166 } else { 2167 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 2168 goto out; 2169 } 2170 } 2171 } 2172 if (sk->sk_state != TCP_CLOSE) { 2173 sk_mem_reclaim(sk); 2174 if (tcp_check_oom(sk, 0)) { 2175 tcp_set_state(sk, TCP_CLOSE); 2176 tcp_send_active_reset(sk, GFP_ATOMIC); 2177 __NET_INC_STATS(sock_net(sk), 2178 LINUX_MIB_TCPABORTONMEMORY); 2179 } 2180 } 2181 2182 if (sk->sk_state == TCP_CLOSE) { 2183 struct request_sock *req = tcp_sk(sk)->fastopen_rsk; 2184 /* We could get here with a non-NULL req if the socket is 2185 * aborted (e.g., closed with unread data) before 3WHS 2186 * finishes. 2187 */ 2188 if (req) 2189 reqsk_fastopen_remove(sk, req, false); 2190 inet_csk_destroy_sock(sk); 2191 } 2192 /* Otherwise, socket is reprieved until protocol close. */ 2193 2194 out: 2195 bh_unlock_sock(sk); 2196 local_bh_enable(); 2197 sock_put(sk); 2198 } 2199 EXPORT_SYMBOL(tcp_close); 2200 2201 /* These states need RST on ABORT according to RFC793 */ 2202 2203 static inline bool tcp_need_reset(int state) 2204 { 2205 return (1 << state) & 2206 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 | 2207 TCPF_FIN_WAIT2 | TCPF_SYN_RECV); 2208 } 2209 2210 int tcp_disconnect(struct sock *sk, int flags) 2211 { 2212 struct inet_sock *inet = inet_sk(sk); 2213 struct inet_connection_sock *icsk = inet_csk(sk); 2214 struct tcp_sock *tp = tcp_sk(sk); 2215 int err = 0; 2216 int old_state = sk->sk_state; 2217 2218 if (old_state != TCP_CLOSE) 2219 tcp_set_state(sk, TCP_CLOSE); 2220 2221 /* ABORT function of RFC793 */ 2222 if (old_state == TCP_LISTEN) { 2223 inet_csk_listen_stop(sk); 2224 } else if (unlikely(tp->repair)) { 2225 sk->sk_err = ECONNABORTED; 2226 } else if (tcp_need_reset(old_state) || 2227 (tp->snd_nxt != tp->write_seq && 2228 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) { 2229 /* The last check adjusts for discrepancy of Linux wrt. RFC 2230 * states 2231 */ 2232 tcp_send_active_reset(sk, gfp_any()); 2233 sk->sk_err = ECONNRESET; 2234 } else if (old_state == TCP_SYN_SENT) 2235 sk->sk_err = ECONNRESET; 2236 2237 tcp_clear_xmit_timers(sk); 2238 __skb_queue_purge(&sk->sk_receive_queue); 2239 tcp_write_queue_purge(sk); 2240 __skb_queue_purge(&tp->out_of_order_queue); 2241 2242 inet->inet_dport = 0; 2243 2244 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK)) 2245 inet_reset_saddr(sk); 2246 2247 sk->sk_shutdown = 0; 2248 sock_reset_flag(sk, SOCK_DONE); 2249 tp->srtt_us = 0; 2250 tp->write_seq += tp->max_window + 2; 2251 if (tp->write_seq == 0) 2252 tp->write_seq = 1; 2253 icsk->icsk_backoff = 0; 2254 tp->snd_cwnd = 2; 2255 icsk->icsk_probes_out = 0; 2256 tp->packets_out = 0; 2257 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 2258 tp->snd_cwnd_cnt = 0; 2259 tp->window_clamp = 0; 2260 tcp_set_ca_state(sk, TCP_CA_Open); 2261 tcp_clear_retrans(tp); 2262 inet_csk_delack_init(sk); 2263 tcp_init_send_head(sk); 2264 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt)); 2265 __sk_dst_reset(sk); 2266 2267 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash); 2268 2269 sk->sk_error_report(sk); 2270 return err; 2271 } 2272 EXPORT_SYMBOL(tcp_disconnect); 2273 2274 static inline bool tcp_can_repair_sock(const struct sock *sk) 2275 { 2276 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) && 2277 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED)); 2278 } 2279 2280 static int tcp_repair_options_est(struct tcp_sock *tp, 2281 struct tcp_repair_opt __user *optbuf, unsigned int len) 2282 { 2283 struct tcp_repair_opt opt; 2284 2285 while (len >= sizeof(opt)) { 2286 if (copy_from_user(&opt, optbuf, sizeof(opt))) 2287 return -EFAULT; 2288 2289 optbuf++; 2290 len -= sizeof(opt); 2291 2292 switch (opt.opt_code) { 2293 case TCPOPT_MSS: 2294 tp->rx_opt.mss_clamp = opt.opt_val; 2295 break; 2296 case TCPOPT_WINDOW: 2297 { 2298 u16 snd_wscale = opt.opt_val & 0xFFFF; 2299 u16 rcv_wscale = opt.opt_val >> 16; 2300 2301 if (snd_wscale > 14 || rcv_wscale > 14) 2302 return -EFBIG; 2303 2304 tp->rx_opt.snd_wscale = snd_wscale; 2305 tp->rx_opt.rcv_wscale = rcv_wscale; 2306 tp->rx_opt.wscale_ok = 1; 2307 } 2308 break; 2309 case TCPOPT_SACK_PERM: 2310 if (opt.opt_val != 0) 2311 return -EINVAL; 2312 2313 tp->rx_opt.sack_ok |= TCP_SACK_SEEN; 2314 if (sysctl_tcp_fack) 2315 tcp_enable_fack(tp); 2316 break; 2317 case TCPOPT_TIMESTAMP: 2318 if (opt.opt_val != 0) 2319 return -EINVAL; 2320 2321 tp->rx_opt.tstamp_ok = 1; 2322 break; 2323 } 2324 } 2325 2326 return 0; 2327 } 2328 2329 /* 2330 * Socket option code for TCP. 2331 */ 2332 static int do_tcp_setsockopt(struct sock *sk, int level, 2333 int optname, char __user *optval, unsigned int optlen) 2334 { 2335 struct tcp_sock *tp = tcp_sk(sk); 2336 struct inet_connection_sock *icsk = inet_csk(sk); 2337 struct net *net = sock_net(sk); 2338 int val; 2339 int err = 0; 2340 2341 /* These are data/string values, all the others are ints */ 2342 switch (optname) { 2343 case TCP_CONGESTION: { 2344 char name[TCP_CA_NAME_MAX]; 2345 2346 if (optlen < 1) 2347 return -EINVAL; 2348 2349 val = strncpy_from_user(name, optval, 2350 min_t(long, TCP_CA_NAME_MAX-1, optlen)); 2351 if (val < 0) 2352 return -EFAULT; 2353 name[val] = 0; 2354 2355 lock_sock(sk); 2356 err = tcp_set_congestion_control(sk, name); 2357 release_sock(sk); 2358 return err; 2359 } 2360 default: 2361 /* fallthru */ 2362 break; 2363 } 2364 2365 if (optlen < sizeof(int)) 2366 return -EINVAL; 2367 2368 if (get_user(val, (int __user *)optval)) 2369 return -EFAULT; 2370 2371 lock_sock(sk); 2372 2373 switch (optname) { 2374 case TCP_MAXSEG: 2375 /* Values greater than interface MTU won't take effect. However 2376 * at the point when this call is done we typically don't yet 2377 * know which interface is going to be used */ 2378 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) { 2379 err = -EINVAL; 2380 break; 2381 } 2382 tp->rx_opt.user_mss = val; 2383 break; 2384 2385 case TCP_NODELAY: 2386 if (val) { 2387 /* TCP_NODELAY is weaker than TCP_CORK, so that 2388 * this option on corked socket is remembered, but 2389 * it is not activated until cork is cleared. 2390 * 2391 * However, when TCP_NODELAY is set we make 2392 * an explicit push, which overrides even TCP_CORK 2393 * for currently queued segments. 2394 */ 2395 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH; 2396 tcp_push_pending_frames(sk); 2397 } else { 2398 tp->nonagle &= ~TCP_NAGLE_OFF; 2399 } 2400 break; 2401 2402 case TCP_THIN_LINEAR_TIMEOUTS: 2403 if (val < 0 || val > 1) 2404 err = -EINVAL; 2405 else 2406 tp->thin_lto = val; 2407 break; 2408 2409 case TCP_THIN_DUPACK: 2410 if (val < 0 || val > 1) 2411 err = -EINVAL; 2412 else { 2413 tp->thin_dupack = val; 2414 if (tp->thin_dupack) 2415 tcp_disable_early_retrans(tp); 2416 } 2417 break; 2418 2419 case TCP_REPAIR: 2420 if (!tcp_can_repair_sock(sk)) 2421 err = -EPERM; 2422 else if (val == 1) { 2423 tp->repair = 1; 2424 sk->sk_reuse = SK_FORCE_REUSE; 2425 tp->repair_queue = TCP_NO_QUEUE; 2426 } else if (val == 0) { 2427 tp->repair = 0; 2428 sk->sk_reuse = SK_NO_REUSE; 2429 tcp_send_window_probe(sk); 2430 } else 2431 err = -EINVAL; 2432 2433 break; 2434 2435 case TCP_REPAIR_QUEUE: 2436 if (!tp->repair) 2437 err = -EPERM; 2438 else if (val < TCP_QUEUES_NR) 2439 tp->repair_queue = val; 2440 else 2441 err = -EINVAL; 2442 break; 2443 2444 case TCP_QUEUE_SEQ: 2445 if (sk->sk_state != TCP_CLOSE) 2446 err = -EPERM; 2447 else if (tp->repair_queue == TCP_SEND_QUEUE) 2448 tp->write_seq = val; 2449 else if (tp->repair_queue == TCP_RECV_QUEUE) 2450 tp->rcv_nxt = val; 2451 else 2452 err = -EINVAL; 2453 break; 2454 2455 case TCP_REPAIR_OPTIONS: 2456 if (!tp->repair) 2457 err = -EINVAL; 2458 else if (sk->sk_state == TCP_ESTABLISHED) 2459 err = tcp_repair_options_est(tp, 2460 (struct tcp_repair_opt __user *)optval, 2461 optlen); 2462 else 2463 err = -EPERM; 2464 break; 2465 2466 case TCP_CORK: 2467 /* When set indicates to always queue non-full frames. 2468 * Later the user clears this option and we transmit 2469 * any pending partial frames in the queue. This is 2470 * meant to be used alongside sendfile() to get properly 2471 * filled frames when the user (for example) must write 2472 * out headers with a write() call first and then use 2473 * sendfile to send out the data parts. 2474 * 2475 * TCP_CORK can be set together with TCP_NODELAY and it is 2476 * stronger than TCP_NODELAY. 2477 */ 2478 if (val) { 2479 tp->nonagle |= TCP_NAGLE_CORK; 2480 } else { 2481 tp->nonagle &= ~TCP_NAGLE_CORK; 2482 if (tp->nonagle&TCP_NAGLE_OFF) 2483 tp->nonagle |= TCP_NAGLE_PUSH; 2484 tcp_push_pending_frames(sk); 2485 } 2486 break; 2487 2488 case TCP_KEEPIDLE: 2489 if (val < 1 || val > MAX_TCP_KEEPIDLE) 2490 err = -EINVAL; 2491 else { 2492 tp->keepalive_time = val * HZ; 2493 if (sock_flag(sk, SOCK_KEEPOPEN) && 2494 !((1 << sk->sk_state) & 2495 (TCPF_CLOSE | TCPF_LISTEN))) { 2496 u32 elapsed = keepalive_time_elapsed(tp); 2497 if (tp->keepalive_time > elapsed) 2498 elapsed = tp->keepalive_time - elapsed; 2499 else 2500 elapsed = 0; 2501 inet_csk_reset_keepalive_timer(sk, elapsed); 2502 } 2503 } 2504 break; 2505 case TCP_KEEPINTVL: 2506 if (val < 1 || val > MAX_TCP_KEEPINTVL) 2507 err = -EINVAL; 2508 else 2509 tp->keepalive_intvl = val * HZ; 2510 break; 2511 case TCP_KEEPCNT: 2512 if (val < 1 || val > MAX_TCP_KEEPCNT) 2513 err = -EINVAL; 2514 else 2515 tp->keepalive_probes = val; 2516 break; 2517 case TCP_SYNCNT: 2518 if (val < 1 || val > MAX_TCP_SYNCNT) 2519 err = -EINVAL; 2520 else 2521 icsk->icsk_syn_retries = val; 2522 break; 2523 2524 case TCP_SAVE_SYN: 2525 if (val < 0 || val > 1) 2526 err = -EINVAL; 2527 else 2528 tp->save_syn = val; 2529 break; 2530 2531 case TCP_LINGER2: 2532 if (val < 0) 2533 tp->linger2 = -1; 2534 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ) 2535 tp->linger2 = 0; 2536 else 2537 tp->linger2 = val * HZ; 2538 break; 2539 2540 case TCP_DEFER_ACCEPT: 2541 /* Translate value in seconds to number of retransmits */ 2542 icsk->icsk_accept_queue.rskq_defer_accept = 2543 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ, 2544 TCP_RTO_MAX / HZ); 2545 break; 2546 2547 case TCP_WINDOW_CLAMP: 2548 if (!val) { 2549 if (sk->sk_state != TCP_CLOSE) { 2550 err = -EINVAL; 2551 break; 2552 } 2553 tp->window_clamp = 0; 2554 } else 2555 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ? 2556 SOCK_MIN_RCVBUF / 2 : val; 2557 break; 2558 2559 case TCP_QUICKACK: 2560 if (!val) { 2561 icsk->icsk_ack.pingpong = 1; 2562 } else { 2563 icsk->icsk_ack.pingpong = 0; 2564 if ((1 << sk->sk_state) & 2565 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) && 2566 inet_csk_ack_scheduled(sk)) { 2567 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 2568 tcp_cleanup_rbuf(sk, 1); 2569 if (!(val & 1)) 2570 icsk->icsk_ack.pingpong = 1; 2571 } 2572 } 2573 break; 2574 2575 #ifdef CONFIG_TCP_MD5SIG 2576 case TCP_MD5SIG: 2577 /* Read the IP->Key mappings from userspace */ 2578 err = tp->af_specific->md5_parse(sk, optval, optlen); 2579 break; 2580 #endif 2581 case TCP_USER_TIMEOUT: 2582 /* Cap the max time in ms TCP will retry or probe the window 2583 * before giving up and aborting (ETIMEDOUT) a connection. 2584 */ 2585 if (val < 0) 2586 err = -EINVAL; 2587 else 2588 icsk->icsk_user_timeout = msecs_to_jiffies(val); 2589 break; 2590 2591 case TCP_FASTOPEN: 2592 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE | 2593 TCPF_LISTEN))) { 2594 tcp_fastopen_init_key_once(true); 2595 2596 fastopen_queue_tune(sk, val); 2597 } else { 2598 err = -EINVAL; 2599 } 2600 break; 2601 case TCP_TIMESTAMP: 2602 if (!tp->repair) 2603 err = -EPERM; 2604 else 2605 tp->tsoffset = val - tcp_time_stamp; 2606 break; 2607 case TCP_NOTSENT_LOWAT: 2608 tp->notsent_lowat = val; 2609 sk->sk_write_space(sk); 2610 break; 2611 default: 2612 err = -ENOPROTOOPT; 2613 break; 2614 } 2615 2616 release_sock(sk); 2617 return err; 2618 } 2619 2620 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval, 2621 unsigned int optlen) 2622 { 2623 const struct inet_connection_sock *icsk = inet_csk(sk); 2624 2625 if (level != SOL_TCP) 2626 return icsk->icsk_af_ops->setsockopt(sk, level, optname, 2627 optval, optlen); 2628 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2629 } 2630 EXPORT_SYMBOL(tcp_setsockopt); 2631 2632 #ifdef CONFIG_COMPAT 2633 int compat_tcp_setsockopt(struct sock *sk, int level, int optname, 2634 char __user *optval, unsigned int optlen) 2635 { 2636 if (level != SOL_TCP) 2637 return inet_csk_compat_setsockopt(sk, level, optname, 2638 optval, optlen); 2639 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2640 } 2641 EXPORT_SYMBOL(compat_tcp_setsockopt); 2642 #endif 2643 2644 /* Return information about state of tcp endpoint in API format. */ 2645 void tcp_get_info(struct sock *sk, struct tcp_info *info) 2646 { 2647 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */ 2648 const struct inet_connection_sock *icsk = inet_csk(sk); 2649 u32 now = tcp_time_stamp; 2650 unsigned int start; 2651 int notsent_bytes; 2652 u64 rate64; 2653 u32 rate; 2654 2655 memset(info, 0, sizeof(*info)); 2656 if (sk->sk_type != SOCK_STREAM) 2657 return; 2658 2659 info->tcpi_state = sk_state_load(sk); 2660 2661 info->tcpi_ca_state = icsk->icsk_ca_state; 2662 info->tcpi_retransmits = icsk->icsk_retransmits; 2663 info->tcpi_probes = icsk->icsk_probes_out; 2664 info->tcpi_backoff = icsk->icsk_backoff; 2665 2666 if (tp->rx_opt.tstamp_ok) 2667 info->tcpi_options |= TCPI_OPT_TIMESTAMPS; 2668 if (tcp_is_sack(tp)) 2669 info->tcpi_options |= TCPI_OPT_SACK; 2670 if (tp->rx_opt.wscale_ok) { 2671 info->tcpi_options |= TCPI_OPT_WSCALE; 2672 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale; 2673 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale; 2674 } 2675 2676 if (tp->ecn_flags & TCP_ECN_OK) 2677 info->tcpi_options |= TCPI_OPT_ECN; 2678 if (tp->ecn_flags & TCP_ECN_SEEN) 2679 info->tcpi_options |= TCPI_OPT_ECN_SEEN; 2680 if (tp->syn_data_acked) 2681 info->tcpi_options |= TCPI_OPT_SYN_DATA; 2682 2683 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto); 2684 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato); 2685 info->tcpi_snd_mss = tp->mss_cache; 2686 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss; 2687 2688 if (info->tcpi_state == TCP_LISTEN) { 2689 info->tcpi_unacked = sk->sk_ack_backlog; 2690 info->tcpi_sacked = sk->sk_max_ack_backlog; 2691 } else { 2692 info->tcpi_unacked = tp->packets_out; 2693 info->tcpi_sacked = tp->sacked_out; 2694 } 2695 info->tcpi_lost = tp->lost_out; 2696 info->tcpi_retrans = tp->retrans_out; 2697 info->tcpi_fackets = tp->fackets_out; 2698 2699 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime); 2700 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime); 2701 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp); 2702 2703 info->tcpi_pmtu = icsk->icsk_pmtu_cookie; 2704 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh; 2705 info->tcpi_rtt = tp->srtt_us >> 3; 2706 info->tcpi_rttvar = tp->mdev_us >> 2; 2707 info->tcpi_snd_ssthresh = tp->snd_ssthresh; 2708 info->tcpi_snd_cwnd = tp->snd_cwnd; 2709 info->tcpi_advmss = tp->advmss; 2710 info->tcpi_reordering = tp->reordering; 2711 2712 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3; 2713 info->tcpi_rcv_space = tp->rcvq_space.space; 2714 2715 info->tcpi_total_retrans = tp->total_retrans; 2716 2717 rate = READ_ONCE(sk->sk_pacing_rate); 2718 rate64 = rate != ~0U ? rate : ~0ULL; 2719 put_unaligned(rate64, &info->tcpi_pacing_rate); 2720 2721 rate = READ_ONCE(sk->sk_max_pacing_rate); 2722 rate64 = rate != ~0U ? rate : ~0ULL; 2723 put_unaligned(rate64, &info->tcpi_max_pacing_rate); 2724 2725 do { 2726 start = u64_stats_fetch_begin_irq(&tp->syncp); 2727 put_unaligned(tp->bytes_acked, &info->tcpi_bytes_acked); 2728 put_unaligned(tp->bytes_received, &info->tcpi_bytes_received); 2729 } while (u64_stats_fetch_retry_irq(&tp->syncp, start)); 2730 info->tcpi_segs_out = tp->segs_out; 2731 info->tcpi_segs_in = tp->segs_in; 2732 2733 notsent_bytes = READ_ONCE(tp->write_seq) - READ_ONCE(tp->snd_nxt); 2734 info->tcpi_notsent_bytes = max(0, notsent_bytes); 2735 2736 info->tcpi_min_rtt = tcp_min_rtt(tp); 2737 info->tcpi_data_segs_in = tp->data_segs_in; 2738 info->tcpi_data_segs_out = tp->data_segs_out; 2739 } 2740 EXPORT_SYMBOL_GPL(tcp_get_info); 2741 2742 static int do_tcp_getsockopt(struct sock *sk, int level, 2743 int optname, char __user *optval, int __user *optlen) 2744 { 2745 struct inet_connection_sock *icsk = inet_csk(sk); 2746 struct tcp_sock *tp = tcp_sk(sk); 2747 struct net *net = sock_net(sk); 2748 int val, len; 2749 2750 if (get_user(len, optlen)) 2751 return -EFAULT; 2752 2753 len = min_t(unsigned int, len, sizeof(int)); 2754 2755 if (len < 0) 2756 return -EINVAL; 2757 2758 switch (optname) { 2759 case TCP_MAXSEG: 2760 val = tp->mss_cache; 2761 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 2762 val = tp->rx_opt.user_mss; 2763 if (tp->repair) 2764 val = tp->rx_opt.mss_clamp; 2765 break; 2766 case TCP_NODELAY: 2767 val = !!(tp->nonagle&TCP_NAGLE_OFF); 2768 break; 2769 case TCP_CORK: 2770 val = !!(tp->nonagle&TCP_NAGLE_CORK); 2771 break; 2772 case TCP_KEEPIDLE: 2773 val = keepalive_time_when(tp) / HZ; 2774 break; 2775 case TCP_KEEPINTVL: 2776 val = keepalive_intvl_when(tp) / HZ; 2777 break; 2778 case TCP_KEEPCNT: 2779 val = keepalive_probes(tp); 2780 break; 2781 case TCP_SYNCNT: 2782 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries; 2783 break; 2784 case TCP_LINGER2: 2785 val = tp->linger2; 2786 if (val >= 0) 2787 val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ; 2788 break; 2789 case TCP_DEFER_ACCEPT: 2790 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept, 2791 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ); 2792 break; 2793 case TCP_WINDOW_CLAMP: 2794 val = tp->window_clamp; 2795 break; 2796 case TCP_INFO: { 2797 struct tcp_info info; 2798 2799 if (get_user(len, optlen)) 2800 return -EFAULT; 2801 2802 tcp_get_info(sk, &info); 2803 2804 len = min_t(unsigned int, len, sizeof(info)); 2805 if (put_user(len, optlen)) 2806 return -EFAULT; 2807 if (copy_to_user(optval, &info, len)) 2808 return -EFAULT; 2809 return 0; 2810 } 2811 case TCP_CC_INFO: { 2812 const struct tcp_congestion_ops *ca_ops; 2813 union tcp_cc_info info; 2814 size_t sz = 0; 2815 int attr; 2816 2817 if (get_user(len, optlen)) 2818 return -EFAULT; 2819 2820 ca_ops = icsk->icsk_ca_ops; 2821 if (ca_ops && ca_ops->get_info) 2822 sz = ca_ops->get_info(sk, ~0U, &attr, &info); 2823 2824 len = min_t(unsigned int, len, sz); 2825 if (put_user(len, optlen)) 2826 return -EFAULT; 2827 if (copy_to_user(optval, &info, len)) 2828 return -EFAULT; 2829 return 0; 2830 } 2831 case TCP_QUICKACK: 2832 val = !icsk->icsk_ack.pingpong; 2833 break; 2834 2835 case TCP_CONGESTION: 2836 if (get_user(len, optlen)) 2837 return -EFAULT; 2838 len = min_t(unsigned int, len, TCP_CA_NAME_MAX); 2839 if (put_user(len, optlen)) 2840 return -EFAULT; 2841 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len)) 2842 return -EFAULT; 2843 return 0; 2844 2845 case TCP_THIN_LINEAR_TIMEOUTS: 2846 val = tp->thin_lto; 2847 break; 2848 case TCP_THIN_DUPACK: 2849 val = tp->thin_dupack; 2850 break; 2851 2852 case TCP_REPAIR: 2853 val = tp->repair; 2854 break; 2855 2856 case TCP_REPAIR_QUEUE: 2857 if (tp->repair) 2858 val = tp->repair_queue; 2859 else 2860 return -EINVAL; 2861 break; 2862 2863 case TCP_QUEUE_SEQ: 2864 if (tp->repair_queue == TCP_SEND_QUEUE) 2865 val = tp->write_seq; 2866 else if (tp->repair_queue == TCP_RECV_QUEUE) 2867 val = tp->rcv_nxt; 2868 else 2869 return -EINVAL; 2870 break; 2871 2872 case TCP_USER_TIMEOUT: 2873 val = jiffies_to_msecs(icsk->icsk_user_timeout); 2874 break; 2875 2876 case TCP_FASTOPEN: 2877 val = icsk->icsk_accept_queue.fastopenq.max_qlen; 2878 break; 2879 2880 case TCP_TIMESTAMP: 2881 val = tcp_time_stamp + tp->tsoffset; 2882 break; 2883 case TCP_NOTSENT_LOWAT: 2884 val = tp->notsent_lowat; 2885 break; 2886 case TCP_SAVE_SYN: 2887 val = tp->save_syn; 2888 break; 2889 case TCP_SAVED_SYN: { 2890 if (get_user(len, optlen)) 2891 return -EFAULT; 2892 2893 lock_sock(sk); 2894 if (tp->saved_syn) { 2895 if (len < tp->saved_syn[0]) { 2896 if (put_user(tp->saved_syn[0], optlen)) { 2897 release_sock(sk); 2898 return -EFAULT; 2899 } 2900 release_sock(sk); 2901 return -EINVAL; 2902 } 2903 len = tp->saved_syn[0]; 2904 if (put_user(len, optlen)) { 2905 release_sock(sk); 2906 return -EFAULT; 2907 } 2908 if (copy_to_user(optval, tp->saved_syn + 1, len)) { 2909 release_sock(sk); 2910 return -EFAULT; 2911 } 2912 tcp_saved_syn_free(tp); 2913 release_sock(sk); 2914 } else { 2915 release_sock(sk); 2916 len = 0; 2917 if (put_user(len, optlen)) 2918 return -EFAULT; 2919 } 2920 return 0; 2921 } 2922 default: 2923 return -ENOPROTOOPT; 2924 } 2925 2926 if (put_user(len, optlen)) 2927 return -EFAULT; 2928 if (copy_to_user(optval, &val, len)) 2929 return -EFAULT; 2930 return 0; 2931 } 2932 2933 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval, 2934 int __user *optlen) 2935 { 2936 struct inet_connection_sock *icsk = inet_csk(sk); 2937 2938 if (level != SOL_TCP) 2939 return icsk->icsk_af_ops->getsockopt(sk, level, optname, 2940 optval, optlen); 2941 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2942 } 2943 EXPORT_SYMBOL(tcp_getsockopt); 2944 2945 #ifdef CONFIG_COMPAT 2946 int compat_tcp_getsockopt(struct sock *sk, int level, int optname, 2947 char __user *optval, int __user *optlen) 2948 { 2949 if (level != SOL_TCP) 2950 return inet_csk_compat_getsockopt(sk, level, optname, 2951 optval, optlen); 2952 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2953 } 2954 EXPORT_SYMBOL(compat_tcp_getsockopt); 2955 #endif 2956 2957 #ifdef CONFIG_TCP_MD5SIG 2958 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool); 2959 static DEFINE_MUTEX(tcp_md5sig_mutex); 2960 static bool tcp_md5sig_pool_populated = false; 2961 2962 static void __tcp_alloc_md5sig_pool(void) 2963 { 2964 struct crypto_ahash *hash; 2965 int cpu; 2966 2967 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC); 2968 if (IS_ERR(hash)) 2969 return; 2970 2971 for_each_possible_cpu(cpu) { 2972 struct ahash_request *req; 2973 2974 if (per_cpu(tcp_md5sig_pool, cpu).md5_req) 2975 continue; 2976 2977 req = ahash_request_alloc(hash, GFP_KERNEL); 2978 if (!req) 2979 return; 2980 2981 ahash_request_set_callback(req, 0, NULL, NULL); 2982 2983 per_cpu(tcp_md5sig_pool, cpu).md5_req = req; 2984 } 2985 /* before setting tcp_md5sig_pool_populated, we must commit all writes 2986 * to memory. See smp_rmb() in tcp_get_md5sig_pool() 2987 */ 2988 smp_wmb(); 2989 tcp_md5sig_pool_populated = true; 2990 } 2991 2992 bool tcp_alloc_md5sig_pool(void) 2993 { 2994 if (unlikely(!tcp_md5sig_pool_populated)) { 2995 mutex_lock(&tcp_md5sig_mutex); 2996 2997 if (!tcp_md5sig_pool_populated) 2998 __tcp_alloc_md5sig_pool(); 2999 3000 mutex_unlock(&tcp_md5sig_mutex); 3001 } 3002 return tcp_md5sig_pool_populated; 3003 } 3004 EXPORT_SYMBOL(tcp_alloc_md5sig_pool); 3005 3006 3007 /** 3008 * tcp_get_md5sig_pool - get md5sig_pool for this user 3009 * 3010 * We use percpu structure, so if we succeed, we exit with preemption 3011 * and BH disabled, to make sure another thread or softirq handling 3012 * wont try to get same context. 3013 */ 3014 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void) 3015 { 3016 local_bh_disable(); 3017 3018 if (tcp_md5sig_pool_populated) { 3019 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */ 3020 smp_rmb(); 3021 return this_cpu_ptr(&tcp_md5sig_pool); 3022 } 3023 local_bh_enable(); 3024 return NULL; 3025 } 3026 EXPORT_SYMBOL(tcp_get_md5sig_pool); 3027 3028 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp, 3029 const struct tcphdr *th) 3030 { 3031 struct scatterlist sg; 3032 struct tcphdr hdr; 3033 3034 /* We are not allowed to change tcphdr, make a local copy */ 3035 memcpy(&hdr, th, sizeof(hdr)); 3036 hdr.check = 0; 3037 3038 /* options aren't included in the hash */ 3039 sg_init_one(&sg, &hdr, sizeof(hdr)); 3040 ahash_request_set_crypt(hp->md5_req, &sg, NULL, sizeof(hdr)); 3041 return crypto_ahash_update(hp->md5_req); 3042 } 3043 EXPORT_SYMBOL(tcp_md5_hash_header); 3044 3045 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp, 3046 const struct sk_buff *skb, unsigned int header_len) 3047 { 3048 struct scatterlist sg; 3049 const struct tcphdr *tp = tcp_hdr(skb); 3050 struct ahash_request *req = hp->md5_req; 3051 unsigned int i; 3052 const unsigned int head_data_len = skb_headlen(skb) > header_len ? 3053 skb_headlen(skb) - header_len : 0; 3054 const struct skb_shared_info *shi = skb_shinfo(skb); 3055 struct sk_buff *frag_iter; 3056 3057 sg_init_table(&sg, 1); 3058 3059 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len); 3060 ahash_request_set_crypt(req, &sg, NULL, head_data_len); 3061 if (crypto_ahash_update(req)) 3062 return 1; 3063 3064 for (i = 0; i < shi->nr_frags; ++i) { 3065 const struct skb_frag_struct *f = &shi->frags[i]; 3066 unsigned int offset = f->page_offset; 3067 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT); 3068 3069 sg_set_page(&sg, page, skb_frag_size(f), 3070 offset_in_page(offset)); 3071 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f)); 3072 if (crypto_ahash_update(req)) 3073 return 1; 3074 } 3075 3076 skb_walk_frags(skb, frag_iter) 3077 if (tcp_md5_hash_skb_data(hp, frag_iter, 0)) 3078 return 1; 3079 3080 return 0; 3081 } 3082 EXPORT_SYMBOL(tcp_md5_hash_skb_data); 3083 3084 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key) 3085 { 3086 struct scatterlist sg; 3087 3088 sg_init_one(&sg, key->key, key->keylen); 3089 ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen); 3090 return crypto_ahash_update(hp->md5_req); 3091 } 3092 EXPORT_SYMBOL(tcp_md5_hash_key); 3093 3094 #endif 3095 3096 void tcp_done(struct sock *sk) 3097 { 3098 struct request_sock *req = tcp_sk(sk)->fastopen_rsk; 3099 3100 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV) 3101 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 3102 3103 tcp_set_state(sk, TCP_CLOSE); 3104 tcp_clear_xmit_timers(sk); 3105 if (req) 3106 reqsk_fastopen_remove(sk, req, false); 3107 3108 sk->sk_shutdown = SHUTDOWN_MASK; 3109 3110 if (!sock_flag(sk, SOCK_DEAD)) 3111 sk->sk_state_change(sk); 3112 else 3113 inet_csk_destroy_sock(sk); 3114 } 3115 EXPORT_SYMBOL_GPL(tcp_done); 3116 3117 int tcp_abort(struct sock *sk, int err) 3118 { 3119 if (!sk_fullsock(sk)) { 3120 if (sk->sk_state == TCP_NEW_SYN_RECV) { 3121 struct request_sock *req = inet_reqsk(sk); 3122 3123 local_bh_disable(); 3124 inet_csk_reqsk_queue_drop_and_put(req->rsk_listener, 3125 req); 3126 local_bh_enable(); 3127 return 0; 3128 } 3129 sock_gen_put(sk); 3130 return -EOPNOTSUPP; 3131 } 3132 3133 /* Don't race with userspace socket closes such as tcp_close. */ 3134 lock_sock(sk); 3135 3136 if (sk->sk_state == TCP_LISTEN) { 3137 tcp_set_state(sk, TCP_CLOSE); 3138 inet_csk_listen_stop(sk); 3139 } 3140 3141 /* Don't race with BH socket closes such as inet_csk_listen_stop. */ 3142 local_bh_disable(); 3143 bh_lock_sock(sk); 3144 3145 if (!sock_flag(sk, SOCK_DEAD)) { 3146 sk->sk_err = err; 3147 /* This barrier is coupled with smp_rmb() in tcp_poll() */ 3148 smp_wmb(); 3149 sk->sk_error_report(sk); 3150 if (tcp_need_reset(sk->sk_state)) 3151 tcp_send_active_reset(sk, GFP_ATOMIC); 3152 tcp_done(sk); 3153 } 3154 3155 bh_unlock_sock(sk); 3156 local_bh_enable(); 3157 release_sock(sk); 3158 sock_put(sk); 3159 return 0; 3160 } 3161 EXPORT_SYMBOL_GPL(tcp_abort); 3162 3163 extern struct tcp_congestion_ops tcp_reno; 3164 3165 static __initdata unsigned long thash_entries; 3166 static int __init set_thash_entries(char *str) 3167 { 3168 ssize_t ret; 3169 3170 if (!str) 3171 return 0; 3172 3173 ret = kstrtoul(str, 0, &thash_entries); 3174 if (ret) 3175 return 0; 3176 3177 return 1; 3178 } 3179 __setup("thash_entries=", set_thash_entries); 3180 3181 static void __init tcp_init_mem(void) 3182 { 3183 unsigned long limit = nr_free_buffer_pages() / 16; 3184 3185 limit = max(limit, 128UL); 3186 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */ 3187 sysctl_tcp_mem[1] = limit; /* 6.25 % */ 3188 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */ 3189 } 3190 3191 void __init tcp_init(void) 3192 { 3193 unsigned long limit; 3194 int max_rshare, max_wshare, cnt; 3195 unsigned int i; 3196 3197 sock_skb_cb_check_size(sizeof(struct tcp_skb_cb)); 3198 3199 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL); 3200 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL); 3201 tcp_hashinfo.bind_bucket_cachep = 3202 kmem_cache_create("tcp_bind_bucket", 3203 sizeof(struct inet_bind_bucket), 0, 3204 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL); 3205 3206 /* Size and allocate the main established and bind bucket 3207 * hash tables. 3208 * 3209 * The methodology is similar to that of the buffer cache. 3210 */ 3211 tcp_hashinfo.ehash = 3212 alloc_large_system_hash("TCP established", 3213 sizeof(struct inet_ehash_bucket), 3214 thash_entries, 3215 17, /* one slot per 128 KB of memory */ 3216 0, 3217 NULL, 3218 &tcp_hashinfo.ehash_mask, 3219 0, 3220 thash_entries ? 0 : 512 * 1024); 3221 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) 3222 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i); 3223 3224 if (inet_ehash_locks_alloc(&tcp_hashinfo)) 3225 panic("TCP: failed to alloc ehash_locks"); 3226 tcp_hashinfo.bhash = 3227 alloc_large_system_hash("TCP bind", 3228 sizeof(struct inet_bind_hashbucket), 3229 tcp_hashinfo.ehash_mask + 1, 3230 17, /* one slot per 128 KB of memory */ 3231 0, 3232 &tcp_hashinfo.bhash_size, 3233 NULL, 3234 0, 3235 64 * 1024); 3236 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size; 3237 for (i = 0; i < tcp_hashinfo.bhash_size; i++) { 3238 spin_lock_init(&tcp_hashinfo.bhash[i].lock); 3239 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain); 3240 } 3241 3242 3243 cnt = tcp_hashinfo.ehash_mask + 1; 3244 3245 tcp_death_row.sysctl_max_tw_buckets = cnt / 2; 3246 sysctl_tcp_max_orphans = cnt / 2; 3247 sysctl_max_syn_backlog = max(128, cnt / 256); 3248 3249 tcp_init_mem(); 3250 /* Set per-socket limits to no more than 1/128 the pressure threshold */ 3251 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7); 3252 max_wshare = min(4UL*1024*1024, limit); 3253 max_rshare = min(6UL*1024*1024, limit); 3254 3255 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM; 3256 sysctl_tcp_wmem[1] = 16*1024; 3257 sysctl_tcp_wmem[2] = max(64*1024, max_wshare); 3258 3259 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM; 3260 sysctl_tcp_rmem[1] = 87380; 3261 sysctl_tcp_rmem[2] = max(87380, max_rshare); 3262 3263 pr_info("Hash tables configured (established %u bind %u)\n", 3264 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size); 3265 3266 tcp_metrics_init(); 3267 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0); 3268 tcp_tasklet_init(); 3269 } 3270