1 /* 2 * INET An implementation of the TCP/IP protocol suite for the LINUX 3 * operating system. INET is implemented using the BSD Socket 4 * interface as the means of communication with the user level. 5 * 6 * Implementation of the Transmission Control Protocol(TCP). 7 * 8 * Authors: Ross Biro 9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 10 * Mark Evans, <evansmp@uhura.aston.ac.uk> 11 * Corey Minyard <wf-rch!minyard@relay.EU.net> 12 * Florian La Roche, <flla@stud.uni-sb.de> 13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 14 * Linus Torvalds, <torvalds@cs.helsinki.fi> 15 * Alan Cox, <gw4pts@gw4pts.ampr.org> 16 * Matthew Dillon, <dillon@apollo.west.oic.com> 17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 18 * Jorge Cwik, <jorge@laser.satlink.net> 19 * 20 * Fixes: 21 * Alan Cox : Numerous verify_area() calls 22 * Alan Cox : Set the ACK bit on a reset 23 * Alan Cox : Stopped it crashing if it closed while 24 * sk->inuse=1 and was trying to connect 25 * (tcp_err()). 26 * Alan Cox : All icmp error handling was broken 27 * pointers passed where wrong and the 28 * socket was looked up backwards. Nobody 29 * tested any icmp error code obviously. 30 * Alan Cox : tcp_err() now handled properly. It 31 * wakes people on errors. poll 32 * behaves and the icmp error race 33 * has gone by moving it into sock.c 34 * Alan Cox : tcp_send_reset() fixed to work for 35 * everything not just packets for 36 * unknown sockets. 37 * Alan Cox : tcp option processing. 38 * Alan Cox : Reset tweaked (still not 100%) [Had 39 * syn rule wrong] 40 * Herp Rosmanith : More reset fixes 41 * Alan Cox : No longer acks invalid rst frames. 42 * Acking any kind of RST is right out. 43 * Alan Cox : Sets an ignore me flag on an rst 44 * receive otherwise odd bits of prattle 45 * escape still 46 * Alan Cox : Fixed another acking RST frame bug. 47 * Should stop LAN workplace lockups. 48 * Alan Cox : Some tidyups using the new skb list 49 * facilities 50 * Alan Cox : sk->keepopen now seems to work 51 * Alan Cox : Pulls options out correctly on accepts 52 * Alan Cox : Fixed assorted sk->rqueue->next errors 53 * Alan Cox : PSH doesn't end a TCP read. Switched a 54 * bit to skb ops. 55 * Alan Cox : Tidied tcp_data to avoid a potential 56 * nasty. 57 * Alan Cox : Added some better commenting, as the 58 * tcp is hard to follow 59 * Alan Cox : Removed incorrect check for 20 * psh 60 * Michael O'Reilly : ack < copied bug fix. 61 * Johannes Stille : Misc tcp fixes (not all in yet). 62 * Alan Cox : FIN with no memory -> CRASH 63 * Alan Cox : Added socket option proto entries. 64 * Also added awareness of them to accept. 65 * Alan Cox : Added TCP options (SOL_TCP) 66 * Alan Cox : Switched wakeup calls to callbacks, 67 * so the kernel can layer network 68 * sockets. 69 * Alan Cox : Use ip_tos/ip_ttl settings. 70 * Alan Cox : Handle FIN (more) properly (we hope). 71 * Alan Cox : RST frames sent on unsynchronised 72 * state ack error. 73 * Alan Cox : Put in missing check for SYN bit. 74 * Alan Cox : Added tcp_select_window() aka NET2E 75 * window non shrink trick. 76 * Alan Cox : Added a couple of small NET2E timer 77 * fixes 78 * Charles Hedrick : TCP fixes 79 * Toomas Tamm : TCP window fixes 80 * Alan Cox : Small URG fix to rlogin ^C ack fight 81 * Charles Hedrick : Rewrote most of it to actually work 82 * Linus : Rewrote tcp_read() and URG handling 83 * completely 84 * Gerhard Koerting: Fixed some missing timer handling 85 * Matthew Dillon : Reworked TCP machine states as per RFC 86 * Gerhard Koerting: PC/TCP workarounds 87 * Adam Caldwell : Assorted timer/timing errors 88 * Matthew Dillon : Fixed another RST bug 89 * Alan Cox : Move to kernel side addressing changes. 90 * Alan Cox : Beginning work on TCP fastpathing 91 * (not yet usable) 92 * Arnt Gulbrandsen: Turbocharged tcp_check() routine. 93 * Alan Cox : TCP fast path debugging 94 * Alan Cox : Window clamping 95 * Michael Riepe : Bug in tcp_check() 96 * Matt Dillon : More TCP improvements and RST bug fixes 97 * Matt Dillon : Yet more small nasties remove from the 98 * TCP code (Be very nice to this man if 99 * tcp finally works 100%) 8) 100 * Alan Cox : BSD accept semantics. 101 * Alan Cox : Reset on closedown bug. 102 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto(). 103 * Michael Pall : Handle poll() after URG properly in 104 * all cases. 105 * Michael Pall : Undo the last fix in tcp_read_urg() 106 * (multi URG PUSH broke rlogin). 107 * Michael Pall : Fix the multi URG PUSH problem in 108 * tcp_readable(), poll() after URG 109 * works now. 110 * Michael Pall : recv(...,MSG_OOB) never blocks in the 111 * BSD api. 112 * Alan Cox : Changed the semantics of sk->socket to 113 * fix a race and a signal problem with 114 * accept() and async I/O. 115 * Alan Cox : Relaxed the rules on tcp_sendto(). 116 * Yury Shevchuk : Really fixed accept() blocking problem. 117 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for 118 * clients/servers which listen in on 119 * fixed ports. 120 * Alan Cox : Cleaned the above up and shrank it to 121 * a sensible code size. 122 * Alan Cox : Self connect lockup fix. 123 * Alan Cox : No connect to multicast. 124 * Ross Biro : Close unaccepted children on master 125 * socket close. 126 * Alan Cox : Reset tracing code. 127 * Alan Cox : Spurious resets on shutdown. 128 * Alan Cox : Giant 15 minute/60 second timer error 129 * Alan Cox : Small whoops in polling before an 130 * accept. 131 * Alan Cox : Kept the state trace facility since 132 * it's handy for debugging. 133 * Alan Cox : More reset handler fixes. 134 * Alan Cox : Started rewriting the code based on 135 * the RFC's for other useful protocol 136 * references see: Comer, KA9Q NOS, and 137 * for a reference on the difference 138 * between specifications and how BSD 139 * works see the 4.4lite source. 140 * A.N.Kuznetsov : Don't time wait on completion of tidy 141 * close. 142 * Linus Torvalds : Fin/Shutdown & copied_seq changes. 143 * Linus Torvalds : Fixed BSD port reuse to work first syn 144 * Alan Cox : Reimplemented timers as per the RFC 145 * and using multiple timers for sanity. 146 * Alan Cox : Small bug fixes, and a lot of new 147 * comments. 148 * Alan Cox : Fixed dual reader crash by locking 149 * the buffers (much like datagram.c) 150 * Alan Cox : Fixed stuck sockets in probe. A probe 151 * now gets fed up of retrying without 152 * (even a no space) answer. 153 * Alan Cox : Extracted closing code better 154 * Alan Cox : Fixed the closing state machine to 155 * resemble the RFC. 156 * Alan Cox : More 'per spec' fixes. 157 * Jorge Cwik : Even faster checksumming. 158 * Alan Cox : tcp_data() doesn't ack illegal PSH 159 * only frames. At least one pc tcp stack 160 * generates them. 161 * Alan Cox : Cache last socket. 162 * Alan Cox : Per route irtt. 163 * Matt Day : poll()->select() match BSD precisely on error 164 * Alan Cox : New buffers 165 * Marc Tamsky : Various sk->prot->retransmits and 166 * sk->retransmits misupdating fixed. 167 * Fixed tcp_write_timeout: stuck close, 168 * and TCP syn retries gets used now. 169 * Mark Yarvis : In tcp_read_wakeup(), don't send an 170 * ack if state is TCP_CLOSED. 171 * Alan Cox : Look up device on a retransmit - routes may 172 * change. Doesn't yet cope with MSS shrink right 173 * but it's a start! 174 * Marc Tamsky : Closing in closing fixes. 175 * Mike Shaver : RFC1122 verifications. 176 * Alan Cox : rcv_saddr errors. 177 * Alan Cox : Block double connect(). 178 * Alan Cox : Small hooks for enSKIP. 179 * Alexey Kuznetsov: Path MTU discovery. 180 * Alan Cox : Support soft errors. 181 * Alan Cox : Fix MTU discovery pathological case 182 * when the remote claims no mtu! 183 * Marc Tamsky : TCP_CLOSE fix. 184 * Colin (G3TNE) : Send a reset on syn ack replies in 185 * window but wrong (fixes NT lpd problems) 186 * Pedro Roque : Better TCP window handling, delayed ack. 187 * Joerg Reuter : No modification of locked buffers in 188 * tcp_do_retransmit() 189 * Eric Schenk : Changed receiver side silly window 190 * avoidance algorithm to BSD style 191 * algorithm. This doubles throughput 192 * against machines running Solaris, 193 * and seems to result in general 194 * improvement. 195 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD 196 * Willy Konynenberg : Transparent proxying support. 197 * Mike McLagan : Routing by source 198 * Keith Owens : Do proper merging with partial SKB's in 199 * tcp_do_sendmsg to avoid burstiness. 200 * Eric Schenk : Fix fast close down bug with 201 * shutdown() followed by close(). 202 * Andi Kleen : Make poll agree with SIGIO 203 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and 204 * lingertime == 0 (RFC 793 ABORT Call) 205 * Hirokazu Takahashi : Use copy_from_user() instead of 206 * csum_and_copy_from_user() if possible. 207 * 208 * This program is free software; you can redistribute it and/or 209 * modify it under the terms of the GNU General Public License 210 * as published by the Free Software Foundation; either version 211 * 2 of the License, or(at your option) any later version. 212 * 213 * Description of States: 214 * 215 * TCP_SYN_SENT sent a connection request, waiting for ack 216 * 217 * TCP_SYN_RECV received a connection request, sent ack, 218 * waiting for final ack in three-way handshake. 219 * 220 * TCP_ESTABLISHED connection established 221 * 222 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete 223 * transmission of remaining buffered data 224 * 225 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote 226 * to shutdown 227 * 228 * TCP_CLOSING both sides have shutdown but we still have 229 * data we have to finish sending 230 * 231 * TCP_TIME_WAIT timeout to catch resent junk before entering 232 * closed, can only be entered from FIN_WAIT2 233 * or CLOSING. Required because the other end 234 * may not have gotten our last ACK causing it 235 * to retransmit the data packet (which we ignore) 236 * 237 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for 238 * us to finish writing our data and to shutdown 239 * (we have to close() to move on to LAST_ACK) 240 * 241 * TCP_LAST_ACK out side has shutdown after remote has 242 * shutdown. There may still be data in our 243 * buffer that we have to finish sending 244 * 245 * TCP_CLOSE socket is finished 246 */ 247 248 #define pr_fmt(fmt) "TCP: " fmt 249 250 #include <linux/kernel.h> 251 #include <linux/module.h> 252 #include <linux/types.h> 253 #include <linux/fcntl.h> 254 #include <linux/poll.h> 255 #include <linux/init.h> 256 #include <linux/fs.h> 257 #include <linux/skbuff.h> 258 #include <linux/scatterlist.h> 259 #include <linux/splice.h> 260 #include <linux/net.h> 261 #include <linux/socket.h> 262 #include <linux/random.h> 263 #include <linux/bootmem.h> 264 #include <linux/highmem.h> 265 #include <linux/swap.h> 266 #include <linux/cache.h> 267 #include <linux/err.h> 268 #include <linux/crypto.h> 269 #include <linux/time.h> 270 #include <linux/slab.h> 271 272 #include <net/icmp.h> 273 #include <net/inet_common.h> 274 #include <net/tcp.h> 275 #include <net/xfrm.h> 276 #include <net/ip.h> 277 #include <net/netdma.h> 278 #include <net/sock.h> 279 280 #include <asm/uaccess.h> 281 #include <asm/ioctls.h> 282 283 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT; 284 285 struct percpu_counter tcp_orphan_count; 286 EXPORT_SYMBOL_GPL(tcp_orphan_count); 287 288 int sysctl_tcp_wmem[3] __read_mostly; 289 int sysctl_tcp_rmem[3] __read_mostly; 290 291 EXPORT_SYMBOL(sysctl_tcp_rmem); 292 EXPORT_SYMBOL(sysctl_tcp_wmem); 293 294 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */ 295 EXPORT_SYMBOL(tcp_memory_allocated); 296 297 /* 298 * Current number of TCP sockets. 299 */ 300 struct percpu_counter tcp_sockets_allocated; 301 EXPORT_SYMBOL(tcp_sockets_allocated); 302 303 /* 304 * TCP splice context 305 */ 306 struct tcp_splice_state { 307 struct pipe_inode_info *pipe; 308 size_t len; 309 unsigned int flags; 310 }; 311 312 /* 313 * Pressure flag: try to collapse. 314 * Technical note: it is used by multiple contexts non atomically. 315 * All the __sk_mem_schedule() is of this nature: accounting 316 * is strict, actions are advisory and have some latency. 317 */ 318 int tcp_memory_pressure __read_mostly; 319 EXPORT_SYMBOL(tcp_memory_pressure); 320 321 void tcp_enter_memory_pressure(struct sock *sk) 322 { 323 if (!tcp_memory_pressure) { 324 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES); 325 tcp_memory_pressure = 1; 326 } 327 } 328 EXPORT_SYMBOL(tcp_enter_memory_pressure); 329 330 /* Convert seconds to retransmits based on initial and max timeout */ 331 static u8 secs_to_retrans(int seconds, int timeout, int rto_max) 332 { 333 u8 res = 0; 334 335 if (seconds > 0) { 336 int period = timeout; 337 338 res = 1; 339 while (seconds > period && res < 255) { 340 res++; 341 timeout <<= 1; 342 if (timeout > rto_max) 343 timeout = rto_max; 344 period += timeout; 345 } 346 } 347 return res; 348 } 349 350 /* Convert retransmits to seconds based on initial and max timeout */ 351 static int retrans_to_secs(u8 retrans, int timeout, int rto_max) 352 { 353 int period = 0; 354 355 if (retrans > 0) { 356 period = timeout; 357 while (--retrans) { 358 timeout <<= 1; 359 if (timeout > rto_max) 360 timeout = rto_max; 361 period += timeout; 362 } 363 } 364 return period; 365 } 366 367 /* Address-family independent initialization for a tcp_sock. 368 * 369 * NOTE: A lot of things set to zero explicitly by call to 370 * sk_alloc() so need not be done here. 371 */ 372 void tcp_init_sock(struct sock *sk) 373 { 374 struct inet_connection_sock *icsk = inet_csk(sk); 375 struct tcp_sock *tp = tcp_sk(sk); 376 377 skb_queue_head_init(&tp->out_of_order_queue); 378 tcp_init_xmit_timers(sk); 379 tcp_prequeue_init(tp); 380 INIT_LIST_HEAD(&tp->tsq_node); 381 382 icsk->icsk_rto = TCP_TIMEOUT_INIT; 383 tp->mdev = TCP_TIMEOUT_INIT; 384 385 /* So many TCP implementations out there (incorrectly) count the 386 * initial SYN frame in their delayed-ACK and congestion control 387 * algorithms that we must have the following bandaid to talk 388 * efficiently to them. -DaveM 389 */ 390 tp->snd_cwnd = TCP_INIT_CWND; 391 392 /* See draft-stevens-tcpca-spec-01 for discussion of the 393 * initialization of these values. 394 */ 395 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 396 tp->snd_cwnd_clamp = ~0; 397 tp->mss_cache = TCP_MSS_DEFAULT; 398 399 tp->reordering = sysctl_tcp_reordering; 400 tcp_enable_early_retrans(tp); 401 icsk->icsk_ca_ops = &tcp_init_congestion_ops; 402 403 sk->sk_state = TCP_CLOSE; 404 405 sk->sk_write_space = sk_stream_write_space; 406 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE); 407 408 icsk->icsk_sync_mss = tcp_sync_mss; 409 410 /* TCP Cookie Transactions */ 411 if (sysctl_tcp_cookie_size > 0) { 412 /* Default, cookies without s_data_payload. */ 413 tp->cookie_values = 414 kzalloc(sizeof(*tp->cookie_values), 415 sk->sk_allocation); 416 if (tp->cookie_values != NULL) 417 kref_init(&tp->cookie_values->kref); 418 } 419 /* Presumed zeroed, in order of appearance: 420 * cookie_in_always, cookie_out_never, 421 * s_data_constant, s_data_in, s_data_out 422 */ 423 sk->sk_sndbuf = sysctl_tcp_wmem[1]; 424 sk->sk_rcvbuf = sysctl_tcp_rmem[1]; 425 426 local_bh_disable(); 427 sock_update_memcg(sk); 428 sk_sockets_allocated_inc(sk); 429 local_bh_enable(); 430 } 431 EXPORT_SYMBOL(tcp_init_sock); 432 433 /* 434 * Wait for a TCP event. 435 * 436 * Note that we don't need to lock the socket, as the upper poll layers 437 * take care of normal races (between the test and the event) and we don't 438 * go look at any of the socket buffers directly. 439 */ 440 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait) 441 { 442 unsigned int mask; 443 struct sock *sk = sock->sk; 444 const struct tcp_sock *tp = tcp_sk(sk); 445 446 sock_poll_wait(file, sk_sleep(sk), wait); 447 if (sk->sk_state == TCP_LISTEN) 448 return inet_csk_listen_poll(sk); 449 450 /* Socket is not locked. We are protected from async events 451 * by poll logic and correct handling of state changes 452 * made by other threads is impossible in any case. 453 */ 454 455 mask = 0; 456 457 /* 458 * POLLHUP is certainly not done right. But poll() doesn't 459 * have a notion of HUP in just one direction, and for a 460 * socket the read side is more interesting. 461 * 462 * Some poll() documentation says that POLLHUP is incompatible 463 * with the POLLOUT/POLLWR flags, so somebody should check this 464 * all. But careful, it tends to be safer to return too many 465 * bits than too few, and you can easily break real applications 466 * if you don't tell them that something has hung up! 467 * 468 * Check-me. 469 * 470 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and 471 * our fs/select.c). It means that after we received EOF, 472 * poll always returns immediately, making impossible poll() on write() 473 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP 474 * if and only if shutdown has been made in both directions. 475 * Actually, it is interesting to look how Solaris and DUX 476 * solve this dilemma. I would prefer, if POLLHUP were maskable, 477 * then we could set it on SND_SHUTDOWN. BTW examples given 478 * in Stevens' books assume exactly this behaviour, it explains 479 * why POLLHUP is incompatible with POLLOUT. --ANK 480 * 481 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent 482 * blocking on fresh not-connected or disconnected socket. --ANK 483 */ 484 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE) 485 mask |= POLLHUP; 486 if (sk->sk_shutdown & RCV_SHUTDOWN) 487 mask |= POLLIN | POLLRDNORM | POLLRDHUP; 488 489 /* Connected or passive Fast Open socket? */ 490 if (sk->sk_state != TCP_SYN_SENT && 491 (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk != NULL)) { 492 int target = sock_rcvlowat(sk, 0, INT_MAX); 493 494 if (tp->urg_seq == tp->copied_seq && 495 !sock_flag(sk, SOCK_URGINLINE) && 496 tp->urg_data) 497 target++; 498 499 /* Potential race condition. If read of tp below will 500 * escape above sk->sk_state, we can be illegally awaken 501 * in SYN_* states. */ 502 if (tp->rcv_nxt - tp->copied_seq >= target) 503 mask |= POLLIN | POLLRDNORM; 504 505 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) { 506 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) { 507 mask |= POLLOUT | POLLWRNORM; 508 } else { /* send SIGIO later */ 509 set_bit(SOCK_ASYNC_NOSPACE, 510 &sk->sk_socket->flags); 511 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 512 513 /* Race breaker. If space is freed after 514 * wspace test but before the flags are set, 515 * IO signal will be lost. 516 */ 517 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) 518 mask |= POLLOUT | POLLWRNORM; 519 } 520 } else 521 mask |= POLLOUT | POLLWRNORM; 522 523 if (tp->urg_data & TCP_URG_VALID) 524 mask |= POLLPRI; 525 } 526 /* This barrier is coupled with smp_wmb() in tcp_reset() */ 527 smp_rmb(); 528 if (sk->sk_err) 529 mask |= POLLERR; 530 531 return mask; 532 } 533 EXPORT_SYMBOL(tcp_poll); 534 535 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg) 536 { 537 struct tcp_sock *tp = tcp_sk(sk); 538 int answ; 539 bool slow; 540 541 switch (cmd) { 542 case SIOCINQ: 543 if (sk->sk_state == TCP_LISTEN) 544 return -EINVAL; 545 546 slow = lock_sock_fast(sk); 547 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 548 answ = 0; 549 else if (sock_flag(sk, SOCK_URGINLINE) || 550 !tp->urg_data || 551 before(tp->urg_seq, tp->copied_seq) || 552 !before(tp->urg_seq, tp->rcv_nxt)) { 553 554 answ = tp->rcv_nxt - tp->copied_seq; 555 556 /* Subtract 1, if FIN was received */ 557 if (answ && sock_flag(sk, SOCK_DONE)) 558 answ--; 559 } else 560 answ = tp->urg_seq - tp->copied_seq; 561 unlock_sock_fast(sk, slow); 562 break; 563 case SIOCATMARK: 564 answ = tp->urg_data && tp->urg_seq == tp->copied_seq; 565 break; 566 case SIOCOUTQ: 567 if (sk->sk_state == TCP_LISTEN) 568 return -EINVAL; 569 570 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 571 answ = 0; 572 else 573 answ = tp->write_seq - tp->snd_una; 574 break; 575 case SIOCOUTQNSD: 576 if (sk->sk_state == TCP_LISTEN) 577 return -EINVAL; 578 579 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 580 answ = 0; 581 else 582 answ = tp->write_seq - tp->snd_nxt; 583 break; 584 default: 585 return -ENOIOCTLCMD; 586 } 587 588 return put_user(answ, (int __user *)arg); 589 } 590 EXPORT_SYMBOL(tcp_ioctl); 591 592 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb) 593 { 594 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 595 tp->pushed_seq = tp->write_seq; 596 } 597 598 static inline bool forced_push(const struct tcp_sock *tp) 599 { 600 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1)); 601 } 602 603 static inline void skb_entail(struct sock *sk, struct sk_buff *skb) 604 { 605 struct tcp_sock *tp = tcp_sk(sk); 606 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 607 608 skb->csum = 0; 609 tcb->seq = tcb->end_seq = tp->write_seq; 610 tcb->tcp_flags = TCPHDR_ACK; 611 tcb->sacked = 0; 612 skb_header_release(skb); 613 tcp_add_write_queue_tail(sk, skb); 614 sk->sk_wmem_queued += skb->truesize; 615 sk_mem_charge(sk, skb->truesize); 616 if (tp->nonagle & TCP_NAGLE_PUSH) 617 tp->nonagle &= ~TCP_NAGLE_PUSH; 618 } 619 620 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags) 621 { 622 if (flags & MSG_OOB) 623 tp->snd_up = tp->write_seq; 624 } 625 626 static inline void tcp_push(struct sock *sk, int flags, int mss_now, 627 int nonagle) 628 { 629 if (tcp_send_head(sk)) { 630 struct tcp_sock *tp = tcp_sk(sk); 631 632 if (!(flags & MSG_MORE) || forced_push(tp)) 633 tcp_mark_push(tp, tcp_write_queue_tail(sk)); 634 635 tcp_mark_urg(tp, flags); 636 __tcp_push_pending_frames(sk, mss_now, 637 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle); 638 } 639 } 640 641 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, 642 unsigned int offset, size_t len) 643 { 644 struct tcp_splice_state *tss = rd_desc->arg.data; 645 int ret; 646 647 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len), 648 tss->flags); 649 if (ret > 0) 650 rd_desc->count -= ret; 651 return ret; 652 } 653 654 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss) 655 { 656 /* Store TCP splice context information in read_descriptor_t. */ 657 read_descriptor_t rd_desc = { 658 .arg.data = tss, 659 .count = tss->len, 660 }; 661 662 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv); 663 } 664 665 /** 666 * tcp_splice_read - splice data from TCP socket to a pipe 667 * @sock: socket to splice from 668 * @ppos: position (not valid) 669 * @pipe: pipe to splice to 670 * @len: number of bytes to splice 671 * @flags: splice modifier flags 672 * 673 * Description: 674 * Will read pages from given socket and fill them into a pipe. 675 * 676 **/ 677 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos, 678 struct pipe_inode_info *pipe, size_t len, 679 unsigned int flags) 680 { 681 struct sock *sk = sock->sk; 682 struct tcp_splice_state tss = { 683 .pipe = pipe, 684 .len = len, 685 .flags = flags, 686 }; 687 long timeo; 688 ssize_t spliced; 689 int ret; 690 691 sock_rps_record_flow(sk); 692 /* 693 * We can't seek on a socket input 694 */ 695 if (unlikely(*ppos)) 696 return -ESPIPE; 697 698 ret = spliced = 0; 699 700 lock_sock(sk); 701 702 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK); 703 while (tss.len) { 704 ret = __tcp_splice_read(sk, &tss); 705 if (ret < 0) 706 break; 707 else if (!ret) { 708 if (spliced) 709 break; 710 if (sock_flag(sk, SOCK_DONE)) 711 break; 712 if (sk->sk_err) { 713 ret = sock_error(sk); 714 break; 715 } 716 if (sk->sk_shutdown & RCV_SHUTDOWN) 717 break; 718 if (sk->sk_state == TCP_CLOSE) { 719 /* 720 * This occurs when user tries to read 721 * from never connected socket. 722 */ 723 if (!sock_flag(sk, SOCK_DONE)) 724 ret = -ENOTCONN; 725 break; 726 } 727 if (!timeo) { 728 ret = -EAGAIN; 729 break; 730 } 731 sk_wait_data(sk, &timeo); 732 if (signal_pending(current)) { 733 ret = sock_intr_errno(timeo); 734 break; 735 } 736 continue; 737 } 738 tss.len -= ret; 739 spliced += ret; 740 741 if (!timeo) 742 break; 743 release_sock(sk); 744 lock_sock(sk); 745 746 if (sk->sk_err || sk->sk_state == TCP_CLOSE || 747 (sk->sk_shutdown & RCV_SHUTDOWN) || 748 signal_pending(current)) 749 break; 750 } 751 752 release_sock(sk); 753 754 if (spliced) 755 return spliced; 756 757 return ret; 758 } 759 EXPORT_SYMBOL(tcp_splice_read); 760 761 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp) 762 { 763 struct sk_buff *skb; 764 765 /* The TCP header must be at least 32-bit aligned. */ 766 size = ALIGN(size, 4); 767 768 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp); 769 if (skb) { 770 if (sk_wmem_schedule(sk, skb->truesize)) { 771 skb_reserve(skb, sk->sk_prot->max_header); 772 /* 773 * Make sure that we have exactly size bytes 774 * available to the caller, no more, no less. 775 */ 776 skb->avail_size = size; 777 return skb; 778 } 779 __kfree_skb(skb); 780 } else { 781 sk->sk_prot->enter_memory_pressure(sk); 782 sk_stream_moderate_sndbuf(sk); 783 } 784 return NULL; 785 } 786 787 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now, 788 int large_allowed) 789 { 790 struct tcp_sock *tp = tcp_sk(sk); 791 u32 xmit_size_goal, old_size_goal; 792 793 xmit_size_goal = mss_now; 794 795 if (large_allowed && sk_can_gso(sk)) { 796 xmit_size_goal = ((sk->sk_gso_max_size - 1) - 797 inet_csk(sk)->icsk_af_ops->net_header_len - 798 inet_csk(sk)->icsk_ext_hdr_len - 799 tp->tcp_header_len); 800 801 /* TSQ : try to have two TSO segments in flight */ 802 xmit_size_goal = min_t(u32, xmit_size_goal, 803 sysctl_tcp_limit_output_bytes >> 1); 804 805 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal); 806 807 /* We try hard to avoid divides here */ 808 old_size_goal = tp->xmit_size_goal_segs * mss_now; 809 810 if (likely(old_size_goal <= xmit_size_goal && 811 old_size_goal + mss_now > xmit_size_goal)) { 812 xmit_size_goal = old_size_goal; 813 } else { 814 tp->xmit_size_goal_segs = 815 min_t(u16, xmit_size_goal / mss_now, 816 sk->sk_gso_max_segs); 817 xmit_size_goal = tp->xmit_size_goal_segs * mss_now; 818 } 819 } 820 821 return max(xmit_size_goal, mss_now); 822 } 823 824 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags) 825 { 826 int mss_now; 827 828 mss_now = tcp_current_mss(sk); 829 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB)); 830 831 return mss_now; 832 } 833 834 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset, 835 size_t size, int flags) 836 { 837 struct tcp_sock *tp = tcp_sk(sk); 838 int mss_now, size_goal; 839 int err; 840 ssize_t copied; 841 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 842 843 /* Wait for a connection to finish. One exception is TCP Fast Open 844 * (passive side) where data is allowed to be sent before a connection 845 * is fully established. 846 */ 847 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) && 848 !tcp_passive_fastopen(sk)) { 849 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0) 850 goto out_err; 851 } 852 853 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags); 854 855 mss_now = tcp_send_mss(sk, &size_goal, flags); 856 copied = 0; 857 858 err = -EPIPE; 859 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 860 goto out_err; 861 862 while (size > 0) { 863 struct sk_buff *skb = tcp_write_queue_tail(sk); 864 int copy, i; 865 bool can_coalesce; 866 867 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) { 868 new_segment: 869 if (!sk_stream_memory_free(sk)) 870 goto wait_for_sndbuf; 871 872 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation); 873 if (!skb) 874 goto wait_for_memory; 875 876 skb_entail(sk, skb); 877 copy = size_goal; 878 } 879 880 if (copy > size) 881 copy = size; 882 883 i = skb_shinfo(skb)->nr_frags; 884 can_coalesce = skb_can_coalesce(skb, i, page, offset); 885 if (!can_coalesce && i >= MAX_SKB_FRAGS) { 886 tcp_mark_push(tp, skb); 887 goto new_segment; 888 } 889 if (!sk_wmem_schedule(sk, copy)) 890 goto wait_for_memory; 891 892 if (can_coalesce) { 893 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 894 } else { 895 get_page(page); 896 skb_fill_page_desc(skb, i, page, offset, copy); 897 } 898 899 skb->len += copy; 900 skb->data_len += copy; 901 skb->truesize += copy; 902 sk->sk_wmem_queued += copy; 903 sk_mem_charge(sk, copy); 904 skb->ip_summed = CHECKSUM_PARTIAL; 905 tp->write_seq += copy; 906 TCP_SKB_CB(skb)->end_seq += copy; 907 skb_shinfo(skb)->gso_segs = 0; 908 909 if (!copied) 910 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 911 912 copied += copy; 913 offset += copy; 914 if (!(size -= copy)) 915 goto out; 916 917 if (skb->len < size_goal || (flags & MSG_OOB)) 918 continue; 919 920 if (forced_push(tp)) { 921 tcp_mark_push(tp, skb); 922 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 923 } else if (skb == tcp_send_head(sk)) 924 tcp_push_one(sk, mss_now); 925 continue; 926 927 wait_for_sndbuf: 928 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 929 wait_for_memory: 930 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH); 931 932 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0) 933 goto do_error; 934 935 mss_now = tcp_send_mss(sk, &size_goal, flags); 936 } 937 938 out: 939 if (copied && !(flags & MSG_SENDPAGE_NOTLAST)) 940 tcp_push(sk, flags, mss_now, tp->nonagle); 941 return copied; 942 943 do_error: 944 if (copied) 945 goto out; 946 out_err: 947 return sk_stream_error(sk, flags, err); 948 } 949 950 int tcp_sendpage(struct sock *sk, struct page *page, int offset, 951 size_t size, int flags) 952 { 953 ssize_t res; 954 955 if (!(sk->sk_route_caps & NETIF_F_SG) || 956 !(sk->sk_route_caps & NETIF_F_ALL_CSUM)) 957 return sock_no_sendpage(sk->sk_socket, page, offset, size, 958 flags); 959 960 lock_sock(sk); 961 res = do_tcp_sendpages(sk, page, offset, size, flags); 962 release_sock(sk); 963 return res; 964 } 965 EXPORT_SYMBOL(tcp_sendpage); 966 967 static inline int select_size(const struct sock *sk, bool sg) 968 { 969 const struct tcp_sock *tp = tcp_sk(sk); 970 int tmp = tp->mss_cache; 971 972 if (sg) { 973 if (sk_can_gso(sk)) { 974 /* Small frames wont use a full page: 975 * Payload will immediately follow tcp header. 976 */ 977 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER); 978 } else { 979 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER); 980 981 if (tmp >= pgbreak && 982 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE) 983 tmp = pgbreak; 984 } 985 } 986 987 return tmp; 988 } 989 990 void tcp_free_fastopen_req(struct tcp_sock *tp) 991 { 992 if (tp->fastopen_req != NULL) { 993 kfree(tp->fastopen_req); 994 tp->fastopen_req = NULL; 995 } 996 } 997 998 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *size) 999 { 1000 struct tcp_sock *tp = tcp_sk(sk); 1001 int err, flags; 1002 1003 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE)) 1004 return -EOPNOTSUPP; 1005 if (tp->fastopen_req != NULL) 1006 return -EALREADY; /* Another Fast Open is in progress */ 1007 1008 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request), 1009 sk->sk_allocation); 1010 if (unlikely(tp->fastopen_req == NULL)) 1011 return -ENOBUFS; 1012 tp->fastopen_req->data = msg; 1013 1014 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0; 1015 err = __inet_stream_connect(sk->sk_socket, msg->msg_name, 1016 msg->msg_namelen, flags); 1017 *size = tp->fastopen_req->copied; 1018 tcp_free_fastopen_req(tp); 1019 return err; 1020 } 1021 1022 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg, 1023 size_t size) 1024 { 1025 struct iovec *iov; 1026 struct tcp_sock *tp = tcp_sk(sk); 1027 struct sk_buff *skb; 1028 int iovlen, flags, err, copied = 0; 1029 int mss_now = 0, size_goal, copied_syn = 0, offset = 0; 1030 bool sg; 1031 long timeo; 1032 1033 lock_sock(sk); 1034 1035 flags = msg->msg_flags; 1036 if (flags & MSG_FASTOPEN) { 1037 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn); 1038 if (err == -EINPROGRESS && copied_syn > 0) 1039 goto out; 1040 else if (err) 1041 goto out_err; 1042 offset = copied_syn; 1043 } 1044 1045 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 1046 1047 /* Wait for a connection to finish. One exception is TCP Fast Open 1048 * (passive side) where data is allowed to be sent before a connection 1049 * is fully established. 1050 */ 1051 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) && 1052 !tcp_passive_fastopen(sk)) { 1053 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0) 1054 goto do_error; 1055 } 1056 1057 if (unlikely(tp->repair)) { 1058 if (tp->repair_queue == TCP_RECV_QUEUE) { 1059 copied = tcp_send_rcvq(sk, msg, size); 1060 goto out; 1061 } 1062 1063 err = -EINVAL; 1064 if (tp->repair_queue == TCP_NO_QUEUE) 1065 goto out_err; 1066 1067 /* 'common' sending to sendq */ 1068 } 1069 1070 /* This should be in poll */ 1071 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags); 1072 1073 mss_now = tcp_send_mss(sk, &size_goal, flags); 1074 1075 /* Ok commence sending. */ 1076 iovlen = msg->msg_iovlen; 1077 iov = msg->msg_iov; 1078 copied = 0; 1079 1080 err = -EPIPE; 1081 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 1082 goto out_err; 1083 1084 sg = !!(sk->sk_route_caps & NETIF_F_SG); 1085 1086 while (--iovlen >= 0) { 1087 size_t seglen = iov->iov_len; 1088 unsigned char __user *from = iov->iov_base; 1089 1090 iov++; 1091 if (unlikely(offset > 0)) { /* Skip bytes copied in SYN */ 1092 if (offset >= seglen) { 1093 offset -= seglen; 1094 continue; 1095 } 1096 seglen -= offset; 1097 from += offset; 1098 offset = 0; 1099 } 1100 1101 while (seglen > 0) { 1102 int copy = 0; 1103 int max = size_goal; 1104 1105 skb = tcp_write_queue_tail(sk); 1106 if (tcp_send_head(sk)) { 1107 if (skb->ip_summed == CHECKSUM_NONE) 1108 max = mss_now; 1109 copy = max - skb->len; 1110 } 1111 1112 if (copy <= 0) { 1113 new_segment: 1114 /* Allocate new segment. If the interface is SG, 1115 * allocate skb fitting to single page. 1116 */ 1117 if (!sk_stream_memory_free(sk)) 1118 goto wait_for_sndbuf; 1119 1120 skb = sk_stream_alloc_skb(sk, 1121 select_size(sk, sg), 1122 sk->sk_allocation); 1123 if (!skb) 1124 goto wait_for_memory; 1125 1126 /* 1127 * Check whether we can use HW checksum. 1128 */ 1129 if (sk->sk_route_caps & NETIF_F_ALL_CSUM) 1130 skb->ip_summed = CHECKSUM_PARTIAL; 1131 1132 skb_entail(sk, skb); 1133 copy = size_goal; 1134 max = size_goal; 1135 } 1136 1137 /* Try to append data to the end of skb. */ 1138 if (copy > seglen) 1139 copy = seglen; 1140 1141 /* Where to copy to? */ 1142 if (skb_availroom(skb) > 0) { 1143 /* We have some space in skb head. Superb! */ 1144 copy = min_t(int, copy, skb_availroom(skb)); 1145 err = skb_add_data_nocache(sk, skb, from, copy); 1146 if (err) 1147 goto do_fault; 1148 } else { 1149 bool merge = true; 1150 int i = skb_shinfo(skb)->nr_frags; 1151 struct page_frag *pfrag = sk_page_frag(sk); 1152 1153 if (!sk_page_frag_refill(sk, pfrag)) 1154 goto wait_for_memory; 1155 1156 if (!skb_can_coalesce(skb, i, pfrag->page, 1157 pfrag->offset)) { 1158 if (i == MAX_SKB_FRAGS || !sg) { 1159 tcp_mark_push(tp, skb); 1160 goto new_segment; 1161 } 1162 merge = false; 1163 } 1164 1165 copy = min_t(int, copy, pfrag->size - pfrag->offset); 1166 1167 if (!sk_wmem_schedule(sk, copy)) 1168 goto wait_for_memory; 1169 1170 err = skb_copy_to_page_nocache(sk, from, skb, 1171 pfrag->page, 1172 pfrag->offset, 1173 copy); 1174 if (err) 1175 goto do_error; 1176 1177 /* Update the skb. */ 1178 if (merge) { 1179 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 1180 } else { 1181 skb_fill_page_desc(skb, i, pfrag->page, 1182 pfrag->offset, copy); 1183 get_page(pfrag->page); 1184 } 1185 pfrag->offset += copy; 1186 } 1187 1188 if (!copied) 1189 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 1190 1191 tp->write_seq += copy; 1192 TCP_SKB_CB(skb)->end_seq += copy; 1193 skb_shinfo(skb)->gso_segs = 0; 1194 1195 from += copy; 1196 copied += copy; 1197 if ((seglen -= copy) == 0 && iovlen == 0) 1198 goto out; 1199 1200 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair)) 1201 continue; 1202 1203 if (forced_push(tp)) { 1204 tcp_mark_push(tp, skb); 1205 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 1206 } else if (skb == tcp_send_head(sk)) 1207 tcp_push_one(sk, mss_now); 1208 continue; 1209 1210 wait_for_sndbuf: 1211 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1212 wait_for_memory: 1213 if (copied) 1214 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH); 1215 1216 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0) 1217 goto do_error; 1218 1219 mss_now = tcp_send_mss(sk, &size_goal, flags); 1220 } 1221 } 1222 1223 out: 1224 if (copied) 1225 tcp_push(sk, flags, mss_now, tp->nonagle); 1226 release_sock(sk); 1227 return copied + copied_syn; 1228 1229 do_fault: 1230 if (!skb->len) { 1231 tcp_unlink_write_queue(skb, sk); 1232 /* It is the one place in all of TCP, except connection 1233 * reset, where we can be unlinking the send_head. 1234 */ 1235 tcp_check_send_head(sk, skb); 1236 sk_wmem_free_skb(sk, skb); 1237 } 1238 1239 do_error: 1240 if (copied + copied_syn) 1241 goto out; 1242 out_err: 1243 err = sk_stream_error(sk, flags, err); 1244 release_sock(sk); 1245 return err; 1246 } 1247 EXPORT_SYMBOL(tcp_sendmsg); 1248 1249 /* 1250 * Handle reading urgent data. BSD has very simple semantics for 1251 * this, no blocking and very strange errors 8) 1252 */ 1253 1254 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags) 1255 { 1256 struct tcp_sock *tp = tcp_sk(sk); 1257 1258 /* No URG data to read. */ 1259 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data || 1260 tp->urg_data == TCP_URG_READ) 1261 return -EINVAL; /* Yes this is right ! */ 1262 1263 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE)) 1264 return -ENOTCONN; 1265 1266 if (tp->urg_data & TCP_URG_VALID) { 1267 int err = 0; 1268 char c = tp->urg_data; 1269 1270 if (!(flags & MSG_PEEK)) 1271 tp->urg_data = TCP_URG_READ; 1272 1273 /* Read urgent data. */ 1274 msg->msg_flags |= MSG_OOB; 1275 1276 if (len > 0) { 1277 if (!(flags & MSG_TRUNC)) 1278 err = memcpy_toiovec(msg->msg_iov, &c, 1); 1279 len = 1; 1280 } else 1281 msg->msg_flags |= MSG_TRUNC; 1282 1283 return err ? -EFAULT : len; 1284 } 1285 1286 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 1287 return 0; 1288 1289 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and 1290 * the available implementations agree in this case: 1291 * this call should never block, independent of the 1292 * blocking state of the socket. 1293 * Mike <pall@rz.uni-karlsruhe.de> 1294 */ 1295 return -EAGAIN; 1296 } 1297 1298 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len) 1299 { 1300 struct sk_buff *skb; 1301 int copied = 0, err = 0; 1302 1303 /* XXX -- need to support SO_PEEK_OFF */ 1304 1305 skb_queue_walk(&sk->sk_write_queue, skb) { 1306 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len); 1307 if (err) 1308 break; 1309 1310 copied += skb->len; 1311 } 1312 1313 return err ?: copied; 1314 } 1315 1316 /* Clean up the receive buffer for full frames taken by the user, 1317 * then send an ACK if necessary. COPIED is the number of bytes 1318 * tcp_recvmsg has given to the user so far, it speeds up the 1319 * calculation of whether or not we must ACK for the sake of 1320 * a window update. 1321 */ 1322 void tcp_cleanup_rbuf(struct sock *sk, int copied) 1323 { 1324 struct tcp_sock *tp = tcp_sk(sk); 1325 bool time_to_ack = false; 1326 1327 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 1328 1329 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq), 1330 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n", 1331 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt); 1332 1333 if (inet_csk_ack_scheduled(sk)) { 1334 const struct inet_connection_sock *icsk = inet_csk(sk); 1335 /* Delayed ACKs frequently hit locked sockets during bulk 1336 * receive. */ 1337 if (icsk->icsk_ack.blocked || 1338 /* Once-per-two-segments ACK was not sent by tcp_input.c */ 1339 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss || 1340 /* 1341 * If this read emptied read buffer, we send ACK, if 1342 * connection is not bidirectional, user drained 1343 * receive buffer and there was a small segment 1344 * in queue. 1345 */ 1346 (copied > 0 && 1347 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) || 1348 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) && 1349 !icsk->icsk_ack.pingpong)) && 1350 !atomic_read(&sk->sk_rmem_alloc))) 1351 time_to_ack = true; 1352 } 1353 1354 /* We send an ACK if we can now advertise a non-zero window 1355 * which has been raised "significantly". 1356 * 1357 * Even if window raised up to infinity, do not send window open ACK 1358 * in states, where we will not receive more. It is useless. 1359 */ 1360 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) { 1361 __u32 rcv_window_now = tcp_receive_window(tp); 1362 1363 /* Optimize, __tcp_select_window() is not cheap. */ 1364 if (2*rcv_window_now <= tp->window_clamp) { 1365 __u32 new_window = __tcp_select_window(sk); 1366 1367 /* Send ACK now, if this read freed lots of space 1368 * in our buffer. Certainly, new_window is new window. 1369 * We can advertise it now, if it is not less than current one. 1370 * "Lots" means "at least twice" here. 1371 */ 1372 if (new_window && new_window >= 2 * rcv_window_now) 1373 time_to_ack = true; 1374 } 1375 } 1376 if (time_to_ack) 1377 tcp_send_ack(sk); 1378 } 1379 1380 static void tcp_prequeue_process(struct sock *sk) 1381 { 1382 struct sk_buff *skb; 1383 struct tcp_sock *tp = tcp_sk(sk); 1384 1385 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED); 1386 1387 /* RX process wants to run with disabled BHs, though it is not 1388 * necessary */ 1389 local_bh_disable(); 1390 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL) 1391 sk_backlog_rcv(sk, skb); 1392 local_bh_enable(); 1393 1394 /* Clear memory counter. */ 1395 tp->ucopy.memory = 0; 1396 } 1397 1398 #ifdef CONFIG_NET_DMA 1399 static void tcp_service_net_dma(struct sock *sk, bool wait) 1400 { 1401 dma_cookie_t done, used; 1402 dma_cookie_t last_issued; 1403 struct tcp_sock *tp = tcp_sk(sk); 1404 1405 if (!tp->ucopy.dma_chan) 1406 return; 1407 1408 last_issued = tp->ucopy.dma_cookie; 1409 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1410 1411 do { 1412 if (dma_async_memcpy_complete(tp->ucopy.dma_chan, 1413 last_issued, &done, 1414 &used) == DMA_SUCCESS) { 1415 /* Safe to free early-copied skbs now */ 1416 __skb_queue_purge(&sk->sk_async_wait_queue); 1417 break; 1418 } else { 1419 struct sk_buff *skb; 1420 while ((skb = skb_peek(&sk->sk_async_wait_queue)) && 1421 (dma_async_is_complete(skb->dma_cookie, done, 1422 used) == DMA_SUCCESS)) { 1423 __skb_dequeue(&sk->sk_async_wait_queue); 1424 kfree_skb(skb); 1425 } 1426 } 1427 } while (wait); 1428 } 1429 #endif 1430 1431 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off) 1432 { 1433 struct sk_buff *skb; 1434 u32 offset; 1435 1436 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) { 1437 offset = seq - TCP_SKB_CB(skb)->seq; 1438 if (tcp_hdr(skb)->syn) 1439 offset--; 1440 if (offset < skb->len || tcp_hdr(skb)->fin) { 1441 *off = offset; 1442 return skb; 1443 } 1444 /* This looks weird, but this can happen if TCP collapsing 1445 * splitted a fat GRO packet, while we released socket lock 1446 * in skb_splice_bits() 1447 */ 1448 sk_eat_skb(sk, skb, false); 1449 } 1450 return NULL; 1451 } 1452 1453 /* 1454 * This routine provides an alternative to tcp_recvmsg() for routines 1455 * that would like to handle copying from skbuffs directly in 'sendfile' 1456 * fashion. 1457 * Note: 1458 * - It is assumed that the socket was locked by the caller. 1459 * - The routine does not block. 1460 * - At present, there is no support for reading OOB data 1461 * or for 'peeking' the socket using this routine 1462 * (although both would be easy to implement). 1463 */ 1464 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, 1465 sk_read_actor_t recv_actor) 1466 { 1467 struct sk_buff *skb; 1468 struct tcp_sock *tp = tcp_sk(sk); 1469 u32 seq = tp->copied_seq; 1470 u32 offset; 1471 int copied = 0; 1472 1473 if (sk->sk_state == TCP_LISTEN) 1474 return -ENOTCONN; 1475 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) { 1476 if (offset < skb->len) { 1477 int used; 1478 size_t len; 1479 1480 len = skb->len - offset; 1481 /* Stop reading if we hit a patch of urgent data */ 1482 if (tp->urg_data) { 1483 u32 urg_offset = tp->urg_seq - seq; 1484 if (urg_offset < len) 1485 len = urg_offset; 1486 if (!len) 1487 break; 1488 } 1489 used = recv_actor(desc, skb, offset, len); 1490 if (used <= 0) { 1491 if (!copied) 1492 copied = used; 1493 break; 1494 } else if (used <= len) { 1495 seq += used; 1496 copied += used; 1497 offset += used; 1498 } 1499 /* If recv_actor drops the lock (e.g. TCP splice 1500 * receive) the skb pointer might be invalid when 1501 * getting here: tcp_collapse might have deleted it 1502 * while aggregating skbs from the socket queue. 1503 */ 1504 skb = tcp_recv_skb(sk, seq - 1, &offset); 1505 if (!skb) 1506 break; 1507 /* TCP coalescing might have appended data to the skb. 1508 * Try to splice more frags 1509 */ 1510 if (offset + 1 != skb->len) 1511 continue; 1512 } 1513 if (tcp_hdr(skb)->fin) { 1514 sk_eat_skb(sk, skb, false); 1515 ++seq; 1516 break; 1517 } 1518 sk_eat_skb(sk, skb, false); 1519 if (!desc->count) 1520 break; 1521 tp->copied_seq = seq; 1522 } 1523 tp->copied_seq = seq; 1524 1525 tcp_rcv_space_adjust(sk); 1526 1527 /* Clean up data we have read: This will do ACK frames. */ 1528 if (copied > 0) { 1529 tcp_recv_skb(sk, seq, &offset); 1530 tcp_cleanup_rbuf(sk, copied); 1531 } 1532 return copied; 1533 } 1534 EXPORT_SYMBOL(tcp_read_sock); 1535 1536 /* 1537 * This routine copies from a sock struct into the user buffer. 1538 * 1539 * Technical note: in 2.3 we work on _locked_ socket, so that 1540 * tricks with *seq access order and skb->users are not required. 1541 * Probably, code can be easily improved even more. 1542 */ 1543 1544 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg, 1545 size_t len, int nonblock, int flags, int *addr_len) 1546 { 1547 struct tcp_sock *tp = tcp_sk(sk); 1548 int copied = 0; 1549 u32 peek_seq; 1550 u32 *seq; 1551 unsigned long used; 1552 int err; 1553 int target; /* Read at least this many bytes */ 1554 long timeo; 1555 struct task_struct *user_recv = NULL; 1556 bool copied_early = false; 1557 struct sk_buff *skb; 1558 u32 urg_hole = 0; 1559 1560 lock_sock(sk); 1561 1562 err = -ENOTCONN; 1563 if (sk->sk_state == TCP_LISTEN) 1564 goto out; 1565 1566 timeo = sock_rcvtimeo(sk, nonblock); 1567 1568 /* Urgent data needs to be handled specially. */ 1569 if (flags & MSG_OOB) 1570 goto recv_urg; 1571 1572 if (unlikely(tp->repair)) { 1573 err = -EPERM; 1574 if (!(flags & MSG_PEEK)) 1575 goto out; 1576 1577 if (tp->repair_queue == TCP_SEND_QUEUE) 1578 goto recv_sndq; 1579 1580 err = -EINVAL; 1581 if (tp->repair_queue == TCP_NO_QUEUE) 1582 goto out; 1583 1584 /* 'common' recv queue MSG_PEEK-ing */ 1585 } 1586 1587 seq = &tp->copied_seq; 1588 if (flags & MSG_PEEK) { 1589 peek_seq = tp->copied_seq; 1590 seq = &peek_seq; 1591 } 1592 1593 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 1594 1595 #ifdef CONFIG_NET_DMA 1596 tp->ucopy.dma_chan = NULL; 1597 preempt_disable(); 1598 skb = skb_peek_tail(&sk->sk_receive_queue); 1599 { 1600 int available = 0; 1601 1602 if (skb) 1603 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq); 1604 if ((available < target) && 1605 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) && 1606 !sysctl_tcp_low_latency && 1607 net_dma_find_channel()) { 1608 preempt_enable_no_resched(); 1609 tp->ucopy.pinned_list = 1610 dma_pin_iovec_pages(msg->msg_iov, len); 1611 } else { 1612 preempt_enable_no_resched(); 1613 } 1614 } 1615 #endif 1616 1617 do { 1618 u32 offset; 1619 1620 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */ 1621 if (tp->urg_data && tp->urg_seq == *seq) { 1622 if (copied) 1623 break; 1624 if (signal_pending(current)) { 1625 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN; 1626 break; 1627 } 1628 } 1629 1630 /* Next get a buffer. */ 1631 1632 skb_queue_walk(&sk->sk_receive_queue, skb) { 1633 /* Now that we have two receive queues this 1634 * shouldn't happen. 1635 */ 1636 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq), 1637 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n", 1638 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, 1639 flags)) 1640 break; 1641 1642 offset = *seq - TCP_SKB_CB(skb)->seq; 1643 if (tcp_hdr(skb)->syn) 1644 offset--; 1645 if (offset < skb->len) 1646 goto found_ok_skb; 1647 if (tcp_hdr(skb)->fin) 1648 goto found_fin_ok; 1649 WARN(!(flags & MSG_PEEK), 1650 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n", 1651 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags); 1652 } 1653 1654 /* Well, if we have backlog, try to process it now yet. */ 1655 1656 if (copied >= target && !sk->sk_backlog.tail) 1657 break; 1658 1659 if (copied) { 1660 if (sk->sk_err || 1661 sk->sk_state == TCP_CLOSE || 1662 (sk->sk_shutdown & RCV_SHUTDOWN) || 1663 !timeo || 1664 signal_pending(current)) 1665 break; 1666 } else { 1667 if (sock_flag(sk, SOCK_DONE)) 1668 break; 1669 1670 if (sk->sk_err) { 1671 copied = sock_error(sk); 1672 break; 1673 } 1674 1675 if (sk->sk_shutdown & RCV_SHUTDOWN) 1676 break; 1677 1678 if (sk->sk_state == TCP_CLOSE) { 1679 if (!sock_flag(sk, SOCK_DONE)) { 1680 /* This occurs when user tries to read 1681 * from never connected socket. 1682 */ 1683 copied = -ENOTCONN; 1684 break; 1685 } 1686 break; 1687 } 1688 1689 if (!timeo) { 1690 copied = -EAGAIN; 1691 break; 1692 } 1693 1694 if (signal_pending(current)) { 1695 copied = sock_intr_errno(timeo); 1696 break; 1697 } 1698 } 1699 1700 tcp_cleanup_rbuf(sk, copied); 1701 1702 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) { 1703 /* Install new reader */ 1704 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) { 1705 user_recv = current; 1706 tp->ucopy.task = user_recv; 1707 tp->ucopy.iov = msg->msg_iov; 1708 } 1709 1710 tp->ucopy.len = len; 1711 1712 WARN_ON(tp->copied_seq != tp->rcv_nxt && 1713 !(flags & (MSG_PEEK | MSG_TRUNC))); 1714 1715 /* Ugly... If prequeue is not empty, we have to 1716 * process it before releasing socket, otherwise 1717 * order will be broken at second iteration. 1718 * More elegant solution is required!!! 1719 * 1720 * Look: we have the following (pseudo)queues: 1721 * 1722 * 1. packets in flight 1723 * 2. backlog 1724 * 3. prequeue 1725 * 4. receive_queue 1726 * 1727 * Each queue can be processed only if the next ones 1728 * are empty. At this point we have empty receive_queue. 1729 * But prequeue _can_ be not empty after 2nd iteration, 1730 * when we jumped to start of loop because backlog 1731 * processing added something to receive_queue. 1732 * We cannot release_sock(), because backlog contains 1733 * packets arrived _after_ prequeued ones. 1734 * 1735 * Shortly, algorithm is clear --- to process all 1736 * the queues in order. We could make it more directly, 1737 * requeueing packets from backlog to prequeue, if 1738 * is not empty. It is more elegant, but eats cycles, 1739 * unfortunately. 1740 */ 1741 if (!skb_queue_empty(&tp->ucopy.prequeue)) 1742 goto do_prequeue; 1743 1744 /* __ Set realtime policy in scheduler __ */ 1745 } 1746 1747 #ifdef CONFIG_NET_DMA 1748 if (tp->ucopy.dma_chan) { 1749 if (tp->rcv_wnd == 0 && 1750 !skb_queue_empty(&sk->sk_async_wait_queue)) { 1751 tcp_service_net_dma(sk, true); 1752 tcp_cleanup_rbuf(sk, copied); 1753 } else 1754 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1755 } 1756 #endif 1757 if (copied >= target) { 1758 /* Do not sleep, just process backlog. */ 1759 release_sock(sk); 1760 lock_sock(sk); 1761 } else 1762 sk_wait_data(sk, &timeo); 1763 1764 #ifdef CONFIG_NET_DMA 1765 tcp_service_net_dma(sk, false); /* Don't block */ 1766 tp->ucopy.wakeup = 0; 1767 #endif 1768 1769 if (user_recv) { 1770 int chunk; 1771 1772 /* __ Restore normal policy in scheduler __ */ 1773 1774 if ((chunk = len - tp->ucopy.len) != 0) { 1775 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk); 1776 len -= chunk; 1777 copied += chunk; 1778 } 1779 1780 if (tp->rcv_nxt == tp->copied_seq && 1781 !skb_queue_empty(&tp->ucopy.prequeue)) { 1782 do_prequeue: 1783 tcp_prequeue_process(sk); 1784 1785 if ((chunk = len - tp->ucopy.len) != 0) { 1786 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1787 len -= chunk; 1788 copied += chunk; 1789 } 1790 } 1791 } 1792 if ((flags & MSG_PEEK) && 1793 (peek_seq - copied - urg_hole != tp->copied_seq)) { 1794 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n", 1795 current->comm, 1796 task_pid_nr(current)); 1797 peek_seq = tp->copied_seq; 1798 } 1799 continue; 1800 1801 found_ok_skb: 1802 /* Ok so how much can we use? */ 1803 used = skb->len - offset; 1804 if (len < used) 1805 used = len; 1806 1807 /* Do we have urgent data here? */ 1808 if (tp->urg_data) { 1809 u32 urg_offset = tp->urg_seq - *seq; 1810 if (urg_offset < used) { 1811 if (!urg_offset) { 1812 if (!sock_flag(sk, SOCK_URGINLINE)) { 1813 ++*seq; 1814 urg_hole++; 1815 offset++; 1816 used--; 1817 if (!used) 1818 goto skip_copy; 1819 } 1820 } else 1821 used = urg_offset; 1822 } 1823 } 1824 1825 if (!(flags & MSG_TRUNC)) { 1826 #ifdef CONFIG_NET_DMA 1827 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 1828 tp->ucopy.dma_chan = net_dma_find_channel(); 1829 1830 if (tp->ucopy.dma_chan) { 1831 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec( 1832 tp->ucopy.dma_chan, skb, offset, 1833 msg->msg_iov, used, 1834 tp->ucopy.pinned_list); 1835 1836 if (tp->ucopy.dma_cookie < 0) { 1837 1838 pr_alert("%s: dma_cookie < 0\n", 1839 __func__); 1840 1841 /* Exception. Bailout! */ 1842 if (!copied) 1843 copied = -EFAULT; 1844 break; 1845 } 1846 1847 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan); 1848 1849 if ((offset + used) == skb->len) 1850 copied_early = true; 1851 1852 } else 1853 #endif 1854 { 1855 err = skb_copy_datagram_iovec(skb, offset, 1856 msg->msg_iov, used); 1857 if (err) { 1858 /* Exception. Bailout! */ 1859 if (!copied) 1860 copied = -EFAULT; 1861 break; 1862 } 1863 } 1864 } 1865 1866 *seq += used; 1867 copied += used; 1868 len -= used; 1869 1870 tcp_rcv_space_adjust(sk); 1871 1872 skip_copy: 1873 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) { 1874 tp->urg_data = 0; 1875 tcp_fast_path_check(sk); 1876 } 1877 if (used + offset < skb->len) 1878 continue; 1879 1880 if (tcp_hdr(skb)->fin) 1881 goto found_fin_ok; 1882 if (!(flags & MSG_PEEK)) { 1883 sk_eat_skb(sk, skb, copied_early); 1884 copied_early = false; 1885 } 1886 continue; 1887 1888 found_fin_ok: 1889 /* Process the FIN. */ 1890 ++*seq; 1891 if (!(flags & MSG_PEEK)) { 1892 sk_eat_skb(sk, skb, copied_early); 1893 copied_early = false; 1894 } 1895 break; 1896 } while (len > 0); 1897 1898 if (user_recv) { 1899 if (!skb_queue_empty(&tp->ucopy.prequeue)) { 1900 int chunk; 1901 1902 tp->ucopy.len = copied > 0 ? len : 0; 1903 1904 tcp_prequeue_process(sk); 1905 1906 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) { 1907 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk); 1908 len -= chunk; 1909 copied += chunk; 1910 } 1911 } 1912 1913 tp->ucopy.task = NULL; 1914 tp->ucopy.len = 0; 1915 } 1916 1917 #ifdef CONFIG_NET_DMA 1918 tcp_service_net_dma(sk, true); /* Wait for queue to drain */ 1919 tp->ucopy.dma_chan = NULL; 1920 1921 if (tp->ucopy.pinned_list) { 1922 dma_unpin_iovec_pages(tp->ucopy.pinned_list); 1923 tp->ucopy.pinned_list = NULL; 1924 } 1925 #endif 1926 1927 /* According to UNIX98, msg_name/msg_namelen are ignored 1928 * on connected socket. I was just happy when found this 8) --ANK 1929 */ 1930 1931 /* Clean up data we have read: This will do ACK frames. */ 1932 tcp_cleanup_rbuf(sk, copied); 1933 1934 release_sock(sk); 1935 return copied; 1936 1937 out: 1938 release_sock(sk); 1939 return err; 1940 1941 recv_urg: 1942 err = tcp_recv_urg(sk, msg, len, flags); 1943 goto out; 1944 1945 recv_sndq: 1946 err = tcp_peek_sndq(sk, msg, len); 1947 goto out; 1948 } 1949 EXPORT_SYMBOL(tcp_recvmsg); 1950 1951 void tcp_set_state(struct sock *sk, int state) 1952 { 1953 int oldstate = sk->sk_state; 1954 1955 switch (state) { 1956 case TCP_ESTABLISHED: 1957 if (oldstate != TCP_ESTABLISHED) 1958 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1959 break; 1960 1961 case TCP_CLOSE: 1962 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED) 1963 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS); 1964 1965 sk->sk_prot->unhash(sk); 1966 if (inet_csk(sk)->icsk_bind_hash && 1967 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK)) 1968 inet_put_port(sk); 1969 /* fall through */ 1970 default: 1971 if (oldstate == TCP_ESTABLISHED) 1972 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 1973 } 1974 1975 /* Change state AFTER socket is unhashed to avoid closed 1976 * socket sitting in hash tables. 1977 */ 1978 sk->sk_state = state; 1979 1980 #ifdef STATE_TRACE 1981 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]); 1982 #endif 1983 } 1984 EXPORT_SYMBOL_GPL(tcp_set_state); 1985 1986 /* 1987 * State processing on a close. This implements the state shift for 1988 * sending our FIN frame. Note that we only send a FIN for some 1989 * states. A shutdown() may have already sent the FIN, or we may be 1990 * closed. 1991 */ 1992 1993 static const unsigned char new_state[16] = { 1994 /* current state: new state: action: */ 1995 /* (Invalid) */ TCP_CLOSE, 1996 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1997 /* TCP_SYN_SENT */ TCP_CLOSE, 1998 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN, 1999 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1, 2000 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2, 2001 /* TCP_TIME_WAIT */ TCP_CLOSE, 2002 /* TCP_CLOSE */ TCP_CLOSE, 2003 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN, 2004 /* TCP_LAST_ACK */ TCP_LAST_ACK, 2005 /* TCP_LISTEN */ TCP_CLOSE, 2006 /* TCP_CLOSING */ TCP_CLOSING, 2007 }; 2008 2009 static int tcp_close_state(struct sock *sk) 2010 { 2011 int next = (int)new_state[sk->sk_state]; 2012 int ns = next & TCP_STATE_MASK; 2013 2014 tcp_set_state(sk, ns); 2015 2016 return next & TCP_ACTION_FIN; 2017 } 2018 2019 /* 2020 * Shutdown the sending side of a connection. Much like close except 2021 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD). 2022 */ 2023 2024 void tcp_shutdown(struct sock *sk, int how) 2025 { 2026 /* We need to grab some memory, and put together a FIN, 2027 * and then put it into the queue to be sent. 2028 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92. 2029 */ 2030 if (!(how & SEND_SHUTDOWN)) 2031 return; 2032 2033 /* If we've already sent a FIN, or it's a closed state, skip this. */ 2034 if ((1 << sk->sk_state) & 2035 (TCPF_ESTABLISHED | TCPF_SYN_SENT | 2036 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) { 2037 /* Clear out any half completed packets. FIN if needed. */ 2038 if (tcp_close_state(sk)) 2039 tcp_send_fin(sk); 2040 } 2041 } 2042 EXPORT_SYMBOL(tcp_shutdown); 2043 2044 bool tcp_check_oom(struct sock *sk, int shift) 2045 { 2046 bool too_many_orphans, out_of_socket_memory; 2047 2048 too_many_orphans = tcp_too_many_orphans(sk, shift); 2049 out_of_socket_memory = tcp_out_of_memory(sk); 2050 2051 if (too_many_orphans) 2052 net_info_ratelimited("too many orphaned sockets\n"); 2053 if (out_of_socket_memory) 2054 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n"); 2055 return too_many_orphans || out_of_socket_memory; 2056 } 2057 2058 void tcp_close(struct sock *sk, long timeout) 2059 { 2060 struct sk_buff *skb; 2061 int data_was_unread = 0; 2062 int state; 2063 2064 lock_sock(sk); 2065 sk->sk_shutdown = SHUTDOWN_MASK; 2066 2067 if (sk->sk_state == TCP_LISTEN) { 2068 tcp_set_state(sk, TCP_CLOSE); 2069 2070 /* Special case. */ 2071 inet_csk_listen_stop(sk); 2072 2073 goto adjudge_to_death; 2074 } 2075 2076 /* We need to flush the recv. buffs. We do this only on the 2077 * descriptor close, not protocol-sourced closes, because the 2078 * reader process may not have drained the data yet! 2079 */ 2080 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) { 2081 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq - 2082 tcp_hdr(skb)->fin; 2083 data_was_unread += len; 2084 __kfree_skb(skb); 2085 } 2086 2087 sk_mem_reclaim(sk); 2088 2089 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */ 2090 if (sk->sk_state == TCP_CLOSE) 2091 goto adjudge_to_death; 2092 2093 /* As outlined in RFC 2525, section 2.17, we send a RST here because 2094 * data was lost. To witness the awful effects of the old behavior of 2095 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk 2096 * GET in an FTP client, suspend the process, wait for the client to 2097 * advertise a zero window, then kill -9 the FTP client, wheee... 2098 * Note: timeout is always zero in such a case. 2099 */ 2100 if (unlikely(tcp_sk(sk)->repair)) { 2101 sk->sk_prot->disconnect(sk, 0); 2102 } else if (data_was_unread) { 2103 /* Unread data was tossed, zap the connection. */ 2104 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE); 2105 tcp_set_state(sk, TCP_CLOSE); 2106 tcp_send_active_reset(sk, sk->sk_allocation); 2107 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) { 2108 /* Check zero linger _after_ checking for unread data. */ 2109 sk->sk_prot->disconnect(sk, 0); 2110 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 2111 } else if (tcp_close_state(sk)) { 2112 /* We FIN if the application ate all the data before 2113 * zapping the connection. 2114 */ 2115 2116 /* RED-PEN. Formally speaking, we have broken TCP state 2117 * machine. State transitions: 2118 * 2119 * TCP_ESTABLISHED -> TCP_FIN_WAIT1 2120 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible) 2121 * TCP_CLOSE_WAIT -> TCP_LAST_ACK 2122 * 2123 * are legal only when FIN has been sent (i.e. in window), 2124 * rather than queued out of window. Purists blame. 2125 * 2126 * F.e. "RFC state" is ESTABLISHED, 2127 * if Linux state is FIN-WAIT-1, but FIN is still not sent. 2128 * 2129 * The visible declinations are that sometimes 2130 * we enter time-wait state, when it is not required really 2131 * (harmless), do not send active resets, when they are 2132 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when 2133 * they look as CLOSING or LAST_ACK for Linux) 2134 * Probably, I missed some more holelets. 2135 * --ANK 2136 * XXX (TFO) - To start off we don't support SYN+ACK+FIN 2137 * in a single packet! (May consider it later but will 2138 * probably need API support or TCP_CORK SYN-ACK until 2139 * data is written and socket is closed.) 2140 */ 2141 tcp_send_fin(sk); 2142 } 2143 2144 sk_stream_wait_close(sk, timeout); 2145 2146 adjudge_to_death: 2147 state = sk->sk_state; 2148 sock_hold(sk); 2149 sock_orphan(sk); 2150 2151 /* It is the last release_sock in its life. It will remove backlog. */ 2152 release_sock(sk); 2153 2154 2155 /* Now socket is owned by kernel and we acquire BH lock 2156 to finish close. No need to check for user refs. 2157 */ 2158 local_bh_disable(); 2159 bh_lock_sock(sk); 2160 WARN_ON(sock_owned_by_user(sk)); 2161 2162 percpu_counter_inc(sk->sk_prot->orphan_count); 2163 2164 /* Have we already been destroyed by a softirq or backlog? */ 2165 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE) 2166 goto out; 2167 2168 /* This is a (useful) BSD violating of the RFC. There is a 2169 * problem with TCP as specified in that the other end could 2170 * keep a socket open forever with no application left this end. 2171 * We use a 3 minute timeout (about the same as BSD) then kill 2172 * our end. If they send after that then tough - BUT: long enough 2173 * that we won't make the old 4*rto = almost no time - whoops 2174 * reset mistake. 2175 * 2176 * Nope, it was not mistake. It is really desired behaviour 2177 * f.e. on http servers, when such sockets are useless, but 2178 * consume significant resources. Let's do it with special 2179 * linger2 option. --ANK 2180 */ 2181 2182 if (sk->sk_state == TCP_FIN_WAIT2) { 2183 struct tcp_sock *tp = tcp_sk(sk); 2184 if (tp->linger2 < 0) { 2185 tcp_set_state(sk, TCP_CLOSE); 2186 tcp_send_active_reset(sk, GFP_ATOMIC); 2187 NET_INC_STATS_BH(sock_net(sk), 2188 LINUX_MIB_TCPABORTONLINGER); 2189 } else { 2190 const int tmo = tcp_fin_time(sk); 2191 2192 if (tmo > TCP_TIMEWAIT_LEN) { 2193 inet_csk_reset_keepalive_timer(sk, 2194 tmo - TCP_TIMEWAIT_LEN); 2195 } else { 2196 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 2197 goto out; 2198 } 2199 } 2200 } 2201 if (sk->sk_state != TCP_CLOSE) { 2202 sk_mem_reclaim(sk); 2203 if (tcp_check_oom(sk, 0)) { 2204 tcp_set_state(sk, TCP_CLOSE); 2205 tcp_send_active_reset(sk, GFP_ATOMIC); 2206 NET_INC_STATS_BH(sock_net(sk), 2207 LINUX_MIB_TCPABORTONMEMORY); 2208 } 2209 } 2210 2211 if (sk->sk_state == TCP_CLOSE) { 2212 struct request_sock *req = tcp_sk(sk)->fastopen_rsk; 2213 /* We could get here with a non-NULL req if the socket is 2214 * aborted (e.g., closed with unread data) before 3WHS 2215 * finishes. 2216 */ 2217 if (req != NULL) 2218 reqsk_fastopen_remove(sk, req, false); 2219 inet_csk_destroy_sock(sk); 2220 } 2221 /* Otherwise, socket is reprieved until protocol close. */ 2222 2223 out: 2224 bh_unlock_sock(sk); 2225 local_bh_enable(); 2226 sock_put(sk); 2227 } 2228 EXPORT_SYMBOL(tcp_close); 2229 2230 /* These states need RST on ABORT according to RFC793 */ 2231 2232 static inline bool tcp_need_reset(int state) 2233 { 2234 return (1 << state) & 2235 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 | 2236 TCPF_FIN_WAIT2 | TCPF_SYN_RECV); 2237 } 2238 2239 int tcp_disconnect(struct sock *sk, int flags) 2240 { 2241 struct inet_sock *inet = inet_sk(sk); 2242 struct inet_connection_sock *icsk = inet_csk(sk); 2243 struct tcp_sock *tp = tcp_sk(sk); 2244 int err = 0; 2245 int old_state = sk->sk_state; 2246 2247 if (old_state != TCP_CLOSE) 2248 tcp_set_state(sk, TCP_CLOSE); 2249 2250 /* ABORT function of RFC793 */ 2251 if (old_state == TCP_LISTEN) { 2252 inet_csk_listen_stop(sk); 2253 } else if (unlikely(tp->repair)) { 2254 sk->sk_err = ECONNABORTED; 2255 } else if (tcp_need_reset(old_state) || 2256 (tp->snd_nxt != tp->write_seq && 2257 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) { 2258 /* The last check adjusts for discrepancy of Linux wrt. RFC 2259 * states 2260 */ 2261 tcp_send_active_reset(sk, gfp_any()); 2262 sk->sk_err = ECONNRESET; 2263 } else if (old_state == TCP_SYN_SENT) 2264 sk->sk_err = ECONNRESET; 2265 2266 tcp_clear_xmit_timers(sk); 2267 __skb_queue_purge(&sk->sk_receive_queue); 2268 tcp_write_queue_purge(sk); 2269 __skb_queue_purge(&tp->out_of_order_queue); 2270 #ifdef CONFIG_NET_DMA 2271 __skb_queue_purge(&sk->sk_async_wait_queue); 2272 #endif 2273 2274 inet->inet_dport = 0; 2275 2276 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK)) 2277 inet_reset_saddr(sk); 2278 2279 sk->sk_shutdown = 0; 2280 sock_reset_flag(sk, SOCK_DONE); 2281 tp->srtt = 0; 2282 if ((tp->write_seq += tp->max_window + 2) == 0) 2283 tp->write_seq = 1; 2284 icsk->icsk_backoff = 0; 2285 tp->snd_cwnd = 2; 2286 icsk->icsk_probes_out = 0; 2287 tp->packets_out = 0; 2288 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 2289 tp->snd_cwnd_cnt = 0; 2290 tp->bytes_acked = 0; 2291 tp->window_clamp = 0; 2292 tcp_set_ca_state(sk, TCP_CA_Open); 2293 tcp_clear_retrans(tp); 2294 inet_csk_delack_init(sk); 2295 tcp_init_send_head(sk); 2296 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt)); 2297 __sk_dst_reset(sk); 2298 2299 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash); 2300 2301 sk->sk_error_report(sk); 2302 return err; 2303 } 2304 EXPORT_SYMBOL(tcp_disconnect); 2305 2306 void tcp_sock_destruct(struct sock *sk) 2307 { 2308 inet_sock_destruct(sk); 2309 2310 kfree(inet_csk(sk)->icsk_accept_queue.fastopenq); 2311 } 2312 2313 static inline bool tcp_can_repair_sock(const struct sock *sk) 2314 { 2315 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) && 2316 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED)); 2317 } 2318 2319 static int tcp_repair_options_est(struct tcp_sock *tp, 2320 struct tcp_repair_opt __user *optbuf, unsigned int len) 2321 { 2322 struct tcp_repair_opt opt; 2323 2324 while (len >= sizeof(opt)) { 2325 if (copy_from_user(&opt, optbuf, sizeof(opt))) 2326 return -EFAULT; 2327 2328 optbuf++; 2329 len -= sizeof(opt); 2330 2331 switch (opt.opt_code) { 2332 case TCPOPT_MSS: 2333 tp->rx_opt.mss_clamp = opt.opt_val; 2334 break; 2335 case TCPOPT_WINDOW: 2336 { 2337 u16 snd_wscale = opt.opt_val & 0xFFFF; 2338 u16 rcv_wscale = opt.opt_val >> 16; 2339 2340 if (snd_wscale > 14 || rcv_wscale > 14) 2341 return -EFBIG; 2342 2343 tp->rx_opt.snd_wscale = snd_wscale; 2344 tp->rx_opt.rcv_wscale = rcv_wscale; 2345 tp->rx_opt.wscale_ok = 1; 2346 } 2347 break; 2348 case TCPOPT_SACK_PERM: 2349 if (opt.opt_val != 0) 2350 return -EINVAL; 2351 2352 tp->rx_opt.sack_ok |= TCP_SACK_SEEN; 2353 if (sysctl_tcp_fack) 2354 tcp_enable_fack(tp); 2355 break; 2356 case TCPOPT_TIMESTAMP: 2357 if (opt.opt_val != 0) 2358 return -EINVAL; 2359 2360 tp->rx_opt.tstamp_ok = 1; 2361 break; 2362 } 2363 } 2364 2365 return 0; 2366 } 2367 2368 /* 2369 * Socket option code for TCP. 2370 */ 2371 static int do_tcp_setsockopt(struct sock *sk, int level, 2372 int optname, char __user *optval, unsigned int optlen) 2373 { 2374 struct tcp_sock *tp = tcp_sk(sk); 2375 struct inet_connection_sock *icsk = inet_csk(sk); 2376 int val; 2377 int err = 0; 2378 2379 /* These are data/string values, all the others are ints */ 2380 switch (optname) { 2381 case TCP_CONGESTION: { 2382 char name[TCP_CA_NAME_MAX]; 2383 2384 if (optlen < 1) 2385 return -EINVAL; 2386 2387 val = strncpy_from_user(name, optval, 2388 min_t(long, TCP_CA_NAME_MAX-1, optlen)); 2389 if (val < 0) 2390 return -EFAULT; 2391 name[val] = 0; 2392 2393 lock_sock(sk); 2394 err = tcp_set_congestion_control(sk, name); 2395 release_sock(sk); 2396 return err; 2397 } 2398 case TCP_COOKIE_TRANSACTIONS: { 2399 struct tcp_cookie_transactions ctd; 2400 struct tcp_cookie_values *cvp = NULL; 2401 2402 if (sizeof(ctd) > optlen) 2403 return -EINVAL; 2404 if (copy_from_user(&ctd, optval, sizeof(ctd))) 2405 return -EFAULT; 2406 2407 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) || 2408 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED) 2409 return -EINVAL; 2410 2411 if (ctd.tcpct_cookie_desired == 0) { 2412 /* default to global value */ 2413 } else if ((0x1 & ctd.tcpct_cookie_desired) || 2414 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX || 2415 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) { 2416 return -EINVAL; 2417 } 2418 2419 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) { 2420 /* Supercedes all other values */ 2421 lock_sock(sk); 2422 if (tp->cookie_values != NULL) { 2423 kref_put(&tp->cookie_values->kref, 2424 tcp_cookie_values_release); 2425 tp->cookie_values = NULL; 2426 } 2427 tp->rx_opt.cookie_in_always = 0; /* false */ 2428 tp->rx_opt.cookie_out_never = 1; /* true */ 2429 release_sock(sk); 2430 return err; 2431 } 2432 2433 /* Allocate ancillary memory before locking. 2434 */ 2435 if (ctd.tcpct_used > 0 || 2436 (tp->cookie_values == NULL && 2437 (sysctl_tcp_cookie_size > 0 || 2438 ctd.tcpct_cookie_desired > 0 || 2439 ctd.tcpct_s_data_desired > 0))) { 2440 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used, 2441 GFP_KERNEL); 2442 if (cvp == NULL) 2443 return -ENOMEM; 2444 2445 kref_init(&cvp->kref); 2446 } 2447 lock_sock(sk); 2448 tp->rx_opt.cookie_in_always = 2449 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags); 2450 tp->rx_opt.cookie_out_never = 0; /* false */ 2451 2452 if (tp->cookie_values != NULL) { 2453 if (cvp != NULL) { 2454 /* Changed values are recorded by a changed 2455 * pointer, ensuring the cookie will differ, 2456 * without separately hashing each value later. 2457 */ 2458 kref_put(&tp->cookie_values->kref, 2459 tcp_cookie_values_release); 2460 } else { 2461 cvp = tp->cookie_values; 2462 } 2463 } 2464 2465 if (cvp != NULL) { 2466 cvp->cookie_desired = ctd.tcpct_cookie_desired; 2467 2468 if (ctd.tcpct_used > 0) { 2469 memcpy(cvp->s_data_payload, ctd.tcpct_value, 2470 ctd.tcpct_used); 2471 cvp->s_data_desired = ctd.tcpct_used; 2472 cvp->s_data_constant = 1; /* true */ 2473 } else { 2474 /* No constant payload data. */ 2475 cvp->s_data_desired = ctd.tcpct_s_data_desired; 2476 cvp->s_data_constant = 0; /* false */ 2477 } 2478 2479 tp->cookie_values = cvp; 2480 } 2481 release_sock(sk); 2482 return err; 2483 } 2484 default: 2485 /* fallthru */ 2486 break; 2487 } 2488 2489 if (optlen < sizeof(int)) 2490 return -EINVAL; 2491 2492 if (get_user(val, (int __user *)optval)) 2493 return -EFAULT; 2494 2495 lock_sock(sk); 2496 2497 switch (optname) { 2498 case TCP_MAXSEG: 2499 /* Values greater than interface MTU won't take effect. However 2500 * at the point when this call is done we typically don't yet 2501 * know which interface is going to be used */ 2502 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) { 2503 err = -EINVAL; 2504 break; 2505 } 2506 tp->rx_opt.user_mss = val; 2507 break; 2508 2509 case TCP_NODELAY: 2510 if (val) { 2511 /* TCP_NODELAY is weaker than TCP_CORK, so that 2512 * this option on corked socket is remembered, but 2513 * it is not activated until cork is cleared. 2514 * 2515 * However, when TCP_NODELAY is set we make 2516 * an explicit push, which overrides even TCP_CORK 2517 * for currently queued segments. 2518 */ 2519 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH; 2520 tcp_push_pending_frames(sk); 2521 } else { 2522 tp->nonagle &= ~TCP_NAGLE_OFF; 2523 } 2524 break; 2525 2526 case TCP_THIN_LINEAR_TIMEOUTS: 2527 if (val < 0 || val > 1) 2528 err = -EINVAL; 2529 else 2530 tp->thin_lto = val; 2531 break; 2532 2533 case TCP_THIN_DUPACK: 2534 if (val < 0 || val > 1) 2535 err = -EINVAL; 2536 else 2537 tp->thin_dupack = val; 2538 if (tp->thin_dupack) 2539 tcp_disable_early_retrans(tp); 2540 break; 2541 2542 case TCP_REPAIR: 2543 if (!tcp_can_repair_sock(sk)) 2544 err = -EPERM; 2545 else if (val == 1) { 2546 tp->repair = 1; 2547 sk->sk_reuse = SK_FORCE_REUSE; 2548 tp->repair_queue = TCP_NO_QUEUE; 2549 } else if (val == 0) { 2550 tp->repair = 0; 2551 sk->sk_reuse = SK_NO_REUSE; 2552 tcp_send_window_probe(sk); 2553 } else 2554 err = -EINVAL; 2555 2556 break; 2557 2558 case TCP_REPAIR_QUEUE: 2559 if (!tp->repair) 2560 err = -EPERM; 2561 else if (val < TCP_QUEUES_NR) 2562 tp->repair_queue = val; 2563 else 2564 err = -EINVAL; 2565 break; 2566 2567 case TCP_QUEUE_SEQ: 2568 if (sk->sk_state != TCP_CLOSE) 2569 err = -EPERM; 2570 else if (tp->repair_queue == TCP_SEND_QUEUE) 2571 tp->write_seq = val; 2572 else if (tp->repair_queue == TCP_RECV_QUEUE) 2573 tp->rcv_nxt = val; 2574 else 2575 err = -EINVAL; 2576 break; 2577 2578 case TCP_REPAIR_OPTIONS: 2579 if (!tp->repair) 2580 err = -EINVAL; 2581 else if (sk->sk_state == TCP_ESTABLISHED) 2582 err = tcp_repair_options_est(tp, 2583 (struct tcp_repair_opt __user *)optval, 2584 optlen); 2585 else 2586 err = -EPERM; 2587 break; 2588 2589 case TCP_CORK: 2590 /* When set indicates to always queue non-full frames. 2591 * Later the user clears this option and we transmit 2592 * any pending partial frames in the queue. This is 2593 * meant to be used alongside sendfile() to get properly 2594 * filled frames when the user (for example) must write 2595 * out headers with a write() call first and then use 2596 * sendfile to send out the data parts. 2597 * 2598 * TCP_CORK can be set together with TCP_NODELAY and it is 2599 * stronger than TCP_NODELAY. 2600 */ 2601 if (val) { 2602 tp->nonagle |= TCP_NAGLE_CORK; 2603 } else { 2604 tp->nonagle &= ~TCP_NAGLE_CORK; 2605 if (tp->nonagle&TCP_NAGLE_OFF) 2606 tp->nonagle |= TCP_NAGLE_PUSH; 2607 tcp_push_pending_frames(sk); 2608 } 2609 break; 2610 2611 case TCP_KEEPIDLE: 2612 if (val < 1 || val > MAX_TCP_KEEPIDLE) 2613 err = -EINVAL; 2614 else { 2615 tp->keepalive_time = val * HZ; 2616 if (sock_flag(sk, SOCK_KEEPOPEN) && 2617 !((1 << sk->sk_state) & 2618 (TCPF_CLOSE | TCPF_LISTEN))) { 2619 u32 elapsed = keepalive_time_elapsed(tp); 2620 if (tp->keepalive_time > elapsed) 2621 elapsed = tp->keepalive_time - elapsed; 2622 else 2623 elapsed = 0; 2624 inet_csk_reset_keepalive_timer(sk, elapsed); 2625 } 2626 } 2627 break; 2628 case TCP_KEEPINTVL: 2629 if (val < 1 || val > MAX_TCP_KEEPINTVL) 2630 err = -EINVAL; 2631 else 2632 tp->keepalive_intvl = val * HZ; 2633 break; 2634 case TCP_KEEPCNT: 2635 if (val < 1 || val > MAX_TCP_KEEPCNT) 2636 err = -EINVAL; 2637 else 2638 tp->keepalive_probes = val; 2639 break; 2640 case TCP_SYNCNT: 2641 if (val < 1 || val > MAX_TCP_SYNCNT) 2642 err = -EINVAL; 2643 else 2644 icsk->icsk_syn_retries = val; 2645 break; 2646 2647 case TCP_LINGER2: 2648 if (val < 0) 2649 tp->linger2 = -1; 2650 else if (val > sysctl_tcp_fin_timeout / HZ) 2651 tp->linger2 = 0; 2652 else 2653 tp->linger2 = val * HZ; 2654 break; 2655 2656 case TCP_DEFER_ACCEPT: 2657 /* Translate value in seconds to number of retransmits */ 2658 icsk->icsk_accept_queue.rskq_defer_accept = 2659 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ, 2660 TCP_RTO_MAX / HZ); 2661 break; 2662 2663 case TCP_WINDOW_CLAMP: 2664 if (!val) { 2665 if (sk->sk_state != TCP_CLOSE) { 2666 err = -EINVAL; 2667 break; 2668 } 2669 tp->window_clamp = 0; 2670 } else 2671 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ? 2672 SOCK_MIN_RCVBUF / 2 : val; 2673 break; 2674 2675 case TCP_QUICKACK: 2676 if (!val) { 2677 icsk->icsk_ack.pingpong = 1; 2678 } else { 2679 icsk->icsk_ack.pingpong = 0; 2680 if ((1 << sk->sk_state) & 2681 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) && 2682 inet_csk_ack_scheduled(sk)) { 2683 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 2684 tcp_cleanup_rbuf(sk, 1); 2685 if (!(val & 1)) 2686 icsk->icsk_ack.pingpong = 1; 2687 } 2688 } 2689 break; 2690 2691 #ifdef CONFIG_TCP_MD5SIG 2692 case TCP_MD5SIG: 2693 /* Read the IP->Key mappings from userspace */ 2694 err = tp->af_specific->md5_parse(sk, optval, optlen); 2695 break; 2696 #endif 2697 case TCP_USER_TIMEOUT: 2698 /* Cap the max timeout in ms TCP will retry/retrans 2699 * before giving up and aborting (ETIMEDOUT) a connection. 2700 */ 2701 if (val < 0) 2702 err = -EINVAL; 2703 else 2704 icsk->icsk_user_timeout = msecs_to_jiffies(val); 2705 break; 2706 2707 case TCP_FASTOPEN: 2708 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE | 2709 TCPF_LISTEN))) 2710 err = fastopen_init_queue(sk, val); 2711 else 2712 err = -EINVAL; 2713 break; 2714 default: 2715 err = -ENOPROTOOPT; 2716 break; 2717 } 2718 2719 release_sock(sk); 2720 return err; 2721 } 2722 2723 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval, 2724 unsigned int optlen) 2725 { 2726 const struct inet_connection_sock *icsk = inet_csk(sk); 2727 2728 if (level != SOL_TCP) 2729 return icsk->icsk_af_ops->setsockopt(sk, level, optname, 2730 optval, optlen); 2731 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2732 } 2733 EXPORT_SYMBOL(tcp_setsockopt); 2734 2735 #ifdef CONFIG_COMPAT 2736 int compat_tcp_setsockopt(struct sock *sk, int level, int optname, 2737 char __user *optval, unsigned int optlen) 2738 { 2739 if (level != SOL_TCP) 2740 return inet_csk_compat_setsockopt(sk, level, optname, 2741 optval, optlen); 2742 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2743 } 2744 EXPORT_SYMBOL(compat_tcp_setsockopt); 2745 #endif 2746 2747 /* Return information about state of tcp endpoint in API format. */ 2748 void tcp_get_info(const struct sock *sk, struct tcp_info *info) 2749 { 2750 const struct tcp_sock *tp = tcp_sk(sk); 2751 const struct inet_connection_sock *icsk = inet_csk(sk); 2752 u32 now = tcp_time_stamp; 2753 2754 memset(info, 0, sizeof(*info)); 2755 2756 info->tcpi_state = sk->sk_state; 2757 info->tcpi_ca_state = icsk->icsk_ca_state; 2758 info->tcpi_retransmits = icsk->icsk_retransmits; 2759 info->tcpi_probes = icsk->icsk_probes_out; 2760 info->tcpi_backoff = icsk->icsk_backoff; 2761 2762 if (tp->rx_opt.tstamp_ok) 2763 info->tcpi_options |= TCPI_OPT_TIMESTAMPS; 2764 if (tcp_is_sack(tp)) 2765 info->tcpi_options |= TCPI_OPT_SACK; 2766 if (tp->rx_opt.wscale_ok) { 2767 info->tcpi_options |= TCPI_OPT_WSCALE; 2768 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale; 2769 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale; 2770 } 2771 2772 if (tp->ecn_flags & TCP_ECN_OK) 2773 info->tcpi_options |= TCPI_OPT_ECN; 2774 if (tp->ecn_flags & TCP_ECN_SEEN) 2775 info->tcpi_options |= TCPI_OPT_ECN_SEEN; 2776 if (tp->syn_data_acked) 2777 info->tcpi_options |= TCPI_OPT_SYN_DATA; 2778 2779 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto); 2780 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato); 2781 info->tcpi_snd_mss = tp->mss_cache; 2782 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss; 2783 2784 if (sk->sk_state == TCP_LISTEN) { 2785 info->tcpi_unacked = sk->sk_ack_backlog; 2786 info->tcpi_sacked = sk->sk_max_ack_backlog; 2787 } else { 2788 info->tcpi_unacked = tp->packets_out; 2789 info->tcpi_sacked = tp->sacked_out; 2790 } 2791 info->tcpi_lost = tp->lost_out; 2792 info->tcpi_retrans = tp->retrans_out; 2793 info->tcpi_fackets = tp->fackets_out; 2794 2795 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime); 2796 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime); 2797 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp); 2798 2799 info->tcpi_pmtu = icsk->icsk_pmtu_cookie; 2800 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh; 2801 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3; 2802 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2; 2803 info->tcpi_snd_ssthresh = tp->snd_ssthresh; 2804 info->tcpi_snd_cwnd = tp->snd_cwnd; 2805 info->tcpi_advmss = tp->advmss; 2806 info->tcpi_reordering = tp->reordering; 2807 2808 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3; 2809 info->tcpi_rcv_space = tp->rcvq_space.space; 2810 2811 info->tcpi_total_retrans = tp->total_retrans; 2812 } 2813 EXPORT_SYMBOL_GPL(tcp_get_info); 2814 2815 static int do_tcp_getsockopt(struct sock *sk, int level, 2816 int optname, char __user *optval, int __user *optlen) 2817 { 2818 struct inet_connection_sock *icsk = inet_csk(sk); 2819 struct tcp_sock *tp = tcp_sk(sk); 2820 int val, len; 2821 2822 if (get_user(len, optlen)) 2823 return -EFAULT; 2824 2825 len = min_t(unsigned int, len, sizeof(int)); 2826 2827 if (len < 0) 2828 return -EINVAL; 2829 2830 switch (optname) { 2831 case TCP_MAXSEG: 2832 val = tp->mss_cache; 2833 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 2834 val = tp->rx_opt.user_mss; 2835 if (tp->repair) 2836 val = tp->rx_opt.mss_clamp; 2837 break; 2838 case TCP_NODELAY: 2839 val = !!(tp->nonagle&TCP_NAGLE_OFF); 2840 break; 2841 case TCP_CORK: 2842 val = !!(tp->nonagle&TCP_NAGLE_CORK); 2843 break; 2844 case TCP_KEEPIDLE: 2845 val = keepalive_time_when(tp) / HZ; 2846 break; 2847 case TCP_KEEPINTVL: 2848 val = keepalive_intvl_when(tp) / HZ; 2849 break; 2850 case TCP_KEEPCNT: 2851 val = keepalive_probes(tp); 2852 break; 2853 case TCP_SYNCNT: 2854 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries; 2855 break; 2856 case TCP_LINGER2: 2857 val = tp->linger2; 2858 if (val >= 0) 2859 val = (val ? : sysctl_tcp_fin_timeout) / HZ; 2860 break; 2861 case TCP_DEFER_ACCEPT: 2862 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept, 2863 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ); 2864 break; 2865 case TCP_WINDOW_CLAMP: 2866 val = tp->window_clamp; 2867 break; 2868 case TCP_INFO: { 2869 struct tcp_info info; 2870 2871 if (get_user(len, optlen)) 2872 return -EFAULT; 2873 2874 tcp_get_info(sk, &info); 2875 2876 len = min_t(unsigned int, len, sizeof(info)); 2877 if (put_user(len, optlen)) 2878 return -EFAULT; 2879 if (copy_to_user(optval, &info, len)) 2880 return -EFAULT; 2881 return 0; 2882 } 2883 case TCP_QUICKACK: 2884 val = !icsk->icsk_ack.pingpong; 2885 break; 2886 2887 case TCP_CONGESTION: 2888 if (get_user(len, optlen)) 2889 return -EFAULT; 2890 len = min_t(unsigned int, len, TCP_CA_NAME_MAX); 2891 if (put_user(len, optlen)) 2892 return -EFAULT; 2893 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len)) 2894 return -EFAULT; 2895 return 0; 2896 2897 case TCP_COOKIE_TRANSACTIONS: { 2898 struct tcp_cookie_transactions ctd; 2899 struct tcp_cookie_values *cvp = tp->cookie_values; 2900 2901 if (get_user(len, optlen)) 2902 return -EFAULT; 2903 if (len < sizeof(ctd)) 2904 return -EINVAL; 2905 2906 memset(&ctd, 0, sizeof(ctd)); 2907 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ? 2908 TCP_COOKIE_IN_ALWAYS : 0) 2909 | (tp->rx_opt.cookie_out_never ? 2910 TCP_COOKIE_OUT_NEVER : 0); 2911 2912 if (cvp != NULL) { 2913 ctd.tcpct_flags |= (cvp->s_data_in ? 2914 TCP_S_DATA_IN : 0) 2915 | (cvp->s_data_out ? 2916 TCP_S_DATA_OUT : 0); 2917 2918 ctd.tcpct_cookie_desired = cvp->cookie_desired; 2919 ctd.tcpct_s_data_desired = cvp->s_data_desired; 2920 2921 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0], 2922 cvp->cookie_pair_size); 2923 ctd.tcpct_used = cvp->cookie_pair_size; 2924 } 2925 2926 if (put_user(sizeof(ctd), optlen)) 2927 return -EFAULT; 2928 if (copy_to_user(optval, &ctd, sizeof(ctd))) 2929 return -EFAULT; 2930 return 0; 2931 } 2932 case TCP_THIN_LINEAR_TIMEOUTS: 2933 val = tp->thin_lto; 2934 break; 2935 case TCP_THIN_DUPACK: 2936 val = tp->thin_dupack; 2937 break; 2938 2939 case TCP_REPAIR: 2940 val = tp->repair; 2941 break; 2942 2943 case TCP_REPAIR_QUEUE: 2944 if (tp->repair) 2945 val = tp->repair_queue; 2946 else 2947 return -EINVAL; 2948 break; 2949 2950 case TCP_QUEUE_SEQ: 2951 if (tp->repair_queue == TCP_SEND_QUEUE) 2952 val = tp->write_seq; 2953 else if (tp->repair_queue == TCP_RECV_QUEUE) 2954 val = tp->rcv_nxt; 2955 else 2956 return -EINVAL; 2957 break; 2958 2959 case TCP_USER_TIMEOUT: 2960 val = jiffies_to_msecs(icsk->icsk_user_timeout); 2961 break; 2962 default: 2963 return -ENOPROTOOPT; 2964 } 2965 2966 if (put_user(len, optlen)) 2967 return -EFAULT; 2968 if (copy_to_user(optval, &val, len)) 2969 return -EFAULT; 2970 return 0; 2971 } 2972 2973 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval, 2974 int __user *optlen) 2975 { 2976 struct inet_connection_sock *icsk = inet_csk(sk); 2977 2978 if (level != SOL_TCP) 2979 return icsk->icsk_af_ops->getsockopt(sk, level, optname, 2980 optval, optlen); 2981 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2982 } 2983 EXPORT_SYMBOL(tcp_getsockopt); 2984 2985 #ifdef CONFIG_COMPAT 2986 int compat_tcp_getsockopt(struct sock *sk, int level, int optname, 2987 char __user *optval, int __user *optlen) 2988 { 2989 if (level != SOL_TCP) 2990 return inet_csk_compat_getsockopt(sk, level, optname, 2991 optval, optlen); 2992 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 2993 } 2994 EXPORT_SYMBOL(compat_tcp_getsockopt); 2995 #endif 2996 2997 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, 2998 netdev_features_t features) 2999 { 3000 struct sk_buff *segs = ERR_PTR(-EINVAL); 3001 struct tcphdr *th; 3002 unsigned int thlen; 3003 unsigned int seq; 3004 __be32 delta; 3005 unsigned int oldlen; 3006 unsigned int mss; 3007 3008 if (!pskb_may_pull(skb, sizeof(*th))) 3009 goto out; 3010 3011 th = tcp_hdr(skb); 3012 thlen = th->doff * 4; 3013 if (thlen < sizeof(*th)) 3014 goto out; 3015 3016 if (!pskb_may_pull(skb, thlen)) 3017 goto out; 3018 3019 oldlen = (u16)~skb->len; 3020 __skb_pull(skb, thlen); 3021 3022 mss = skb_shinfo(skb)->gso_size; 3023 if (unlikely(skb->len <= mss)) 3024 goto out; 3025 3026 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) { 3027 /* Packet is from an untrusted source, reset gso_segs. */ 3028 int type = skb_shinfo(skb)->gso_type; 3029 3030 if (unlikely(type & 3031 ~(SKB_GSO_TCPV4 | 3032 SKB_GSO_DODGY | 3033 SKB_GSO_TCP_ECN | 3034 SKB_GSO_TCPV6 | 3035 0) || 3036 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))) 3037 goto out; 3038 3039 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss); 3040 3041 segs = NULL; 3042 goto out; 3043 } 3044 3045 segs = skb_segment(skb, features); 3046 if (IS_ERR(segs)) 3047 goto out; 3048 3049 delta = htonl(oldlen + (thlen + mss)); 3050 3051 skb = segs; 3052 th = tcp_hdr(skb); 3053 seq = ntohl(th->seq); 3054 3055 do { 3056 th->fin = th->psh = 0; 3057 3058 th->check = ~csum_fold((__force __wsum)((__force u32)th->check + 3059 (__force u32)delta)); 3060 if (skb->ip_summed != CHECKSUM_PARTIAL) 3061 th->check = 3062 csum_fold(csum_partial(skb_transport_header(skb), 3063 thlen, skb->csum)); 3064 3065 seq += mss; 3066 skb = skb->next; 3067 th = tcp_hdr(skb); 3068 3069 th->seq = htonl(seq); 3070 th->cwr = 0; 3071 } while (skb->next); 3072 3073 delta = htonl(oldlen + (skb->tail - skb->transport_header) + 3074 skb->data_len); 3075 th->check = ~csum_fold((__force __wsum)((__force u32)th->check + 3076 (__force u32)delta)); 3077 if (skb->ip_summed != CHECKSUM_PARTIAL) 3078 th->check = csum_fold(csum_partial(skb_transport_header(skb), 3079 thlen, skb->csum)); 3080 3081 out: 3082 return segs; 3083 } 3084 EXPORT_SYMBOL(tcp_tso_segment); 3085 3086 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb) 3087 { 3088 struct sk_buff **pp = NULL; 3089 struct sk_buff *p; 3090 struct tcphdr *th; 3091 struct tcphdr *th2; 3092 unsigned int len; 3093 unsigned int thlen; 3094 __be32 flags; 3095 unsigned int mss = 1; 3096 unsigned int hlen; 3097 unsigned int off; 3098 int flush = 1; 3099 int i; 3100 3101 off = skb_gro_offset(skb); 3102 hlen = off + sizeof(*th); 3103 th = skb_gro_header_fast(skb, off); 3104 if (skb_gro_header_hard(skb, hlen)) { 3105 th = skb_gro_header_slow(skb, hlen, off); 3106 if (unlikely(!th)) 3107 goto out; 3108 } 3109 3110 thlen = th->doff * 4; 3111 if (thlen < sizeof(*th)) 3112 goto out; 3113 3114 hlen = off + thlen; 3115 if (skb_gro_header_hard(skb, hlen)) { 3116 th = skb_gro_header_slow(skb, hlen, off); 3117 if (unlikely(!th)) 3118 goto out; 3119 } 3120 3121 skb_gro_pull(skb, thlen); 3122 3123 len = skb_gro_len(skb); 3124 flags = tcp_flag_word(th); 3125 3126 for (; (p = *head); head = &p->next) { 3127 if (!NAPI_GRO_CB(p)->same_flow) 3128 continue; 3129 3130 th2 = tcp_hdr(p); 3131 3132 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) { 3133 NAPI_GRO_CB(p)->same_flow = 0; 3134 continue; 3135 } 3136 3137 goto found; 3138 } 3139 3140 goto out_check_final; 3141 3142 found: 3143 flush = NAPI_GRO_CB(p)->flush; 3144 flush |= (__force int)(flags & TCP_FLAG_CWR); 3145 flush |= (__force int)((flags ^ tcp_flag_word(th2)) & 3146 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH)); 3147 flush |= (__force int)(th->ack_seq ^ th2->ack_seq); 3148 for (i = sizeof(*th); i < thlen; i += 4) 3149 flush |= *(u32 *)((u8 *)th + i) ^ 3150 *(u32 *)((u8 *)th2 + i); 3151 3152 mss = skb_shinfo(p)->gso_size; 3153 3154 flush |= (len - 1) >= mss; 3155 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq); 3156 3157 if (flush || skb_gro_receive(head, skb)) { 3158 mss = 1; 3159 goto out_check_final; 3160 } 3161 3162 p = *head; 3163 th2 = tcp_hdr(p); 3164 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH); 3165 3166 out_check_final: 3167 flush = len < mss; 3168 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH | 3169 TCP_FLAG_RST | TCP_FLAG_SYN | 3170 TCP_FLAG_FIN)); 3171 3172 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush)) 3173 pp = head; 3174 3175 out: 3176 NAPI_GRO_CB(skb)->flush |= flush; 3177 3178 return pp; 3179 } 3180 EXPORT_SYMBOL(tcp_gro_receive); 3181 3182 int tcp_gro_complete(struct sk_buff *skb) 3183 { 3184 struct tcphdr *th = tcp_hdr(skb); 3185 3186 skb->csum_start = skb_transport_header(skb) - skb->head; 3187 skb->csum_offset = offsetof(struct tcphdr, check); 3188 skb->ip_summed = CHECKSUM_PARTIAL; 3189 3190 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count; 3191 3192 if (th->cwr) 3193 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 3194 3195 return 0; 3196 } 3197 EXPORT_SYMBOL(tcp_gro_complete); 3198 3199 #ifdef CONFIG_TCP_MD5SIG 3200 static unsigned long tcp_md5sig_users; 3201 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool; 3202 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock); 3203 3204 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool) 3205 { 3206 int cpu; 3207 3208 for_each_possible_cpu(cpu) { 3209 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu); 3210 3211 if (p->md5_desc.tfm) 3212 crypto_free_hash(p->md5_desc.tfm); 3213 } 3214 free_percpu(pool); 3215 } 3216 3217 void tcp_free_md5sig_pool(void) 3218 { 3219 struct tcp_md5sig_pool __percpu *pool = NULL; 3220 3221 spin_lock_bh(&tcp_md5sig_pool_lock); 3222 if (--tcp_md5sig_users == 0) { 3223 pool = tcp_md5sig_pool; 3224 tcp_md5sig_pool = NULL; 3225 } 3226 spin_unlock_bh(&tcp_md5sig_pool_lock); 3227 if (pool) 3228 __tcp_free_md5sig_pool(pool); 3229 } 3230 EXPORT_SYMBOL(tcp_free_md5sig_pool); 3231 3232 static struct tcp_md5sig_pool __percpu * 3233 __tcp_alloc_md5sig_pool(struct sock *sk) 3234 { 3235 int cpu; 3236 struct tcp_md5sig_pool __percpu *pool; 3237 3238 pool = alloc_percpu(struct tcp_md5sig_pool); 3239 if (!pool) 3240 return NULL; 3241 3242 for_each_possible_cpu(cpu) { 3243 struct crypto_hash *hash; 3244 3245 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC); 3246 if (!hash || IS_ERR(hash)) 3247 goto out_free; 3248 3249 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash; 3250 } 3251 return pool; 3252 out_free: 3253 __tcp_free_md5sig_pool(pool); 3254 return NULL; 3255 } 3256 3257 struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk) 3258 { 3259 struct tcp_md5sig_pool __percpu *pool; 3260 bool alloc = false; 3261 3262 retry: 3263 spin_lock_bh(&tcp_md5sig_pool_lock); 3264 pool = tcp_md5sig_pool; 3265 if (tcp_md5sig_users++ == 0) { 3266 alloc = true; 3267 spin_unlock_bh(&tcp_md5sig_pool_lock); 3268 } else if (!pool) { 3269 tcp_md5sig_users--; 3270 spin_unlock_bh(&tcp_md5sig_pool_lock); 3271 cpu_relax(); 3272 goto retry; 3273 } else 3274 spin_unlock_bh(&tcp_md5sig_pool_lock); 3275 3276 if (alloc) { 3277 /* we cannot hold spinlock here because this may sleep. */ 3278 struct tcp_md5sig_pool __percpu *p; 3279 3280 p = __tcp_alloc_md5sig_pool(sk); 3281 spin_lock_bh(&tcp_md5sig_pool_lock); 3282 if (!p) { 3283 tcp_md5sig_users--; 3284 spin_unlock_bh(&tcp_md5sig_pool_lock); 3285 return NULL; 3286 } 3287 pool = tcp_md5sig_pool; 3288 if (pool) { 3289 /* oops, it has already been assigned. */ 3290 spin_unlock_bh(&tcp_md5sig_pool_lock); 3291 __tcp_free_md5sig_pool(p); 3292 } else { 3293 tcp_md5sig_pool = pool = p; 3294 spin_unlock_bh(&tcp_md5sig_pool_lock); 3295 } 3296 } 3297 return pool; 3298 } 3299 EXPORT_SYMBOL(tcp_alloc_md5sig_pool); 3300 3301 3302 /** 3303 * tcp_get_md5sig_pool - get md5sig_pool for this user 3304 * 3305 * We use percpu structure, so if we succeed, we exit with preemption 3306 * and BH disabled, to make sure another thread or softirq handling 3307 * wont try to get same context. 3308 */ 3309 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void) 3310 { 3311 struct tcp_md5sig_pool __percpu *p; 3312 3313 local_bh_disable(); 3314 3315 spin_lock(&tcp_md5sig_pool_lock); 3316 p = tcp_md5sig_pool; 3317 if (p) 3318 tcp_md5sig_users++; 3319 spin_unlock(&tcp_md5sig_pool_lock); 3320 3321 if (p) 3322 return this_cpu_ptr(p); 3323 3324 local_bh_enable(); 3325 return NULL; 3326 } 3327 EXPORT_SYMBOL(tcp_get_md5sig_pool); 3328 3329 void tcp_put_md5sig_pool(void) 3330 { 3331 local_bh_enable(); 3332 tcp_free_md5sig_pool(); 3333 } 3334 EXPORT_SYMBOL(tcp_put_md5sig_pool); 3335 3336 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp, 3337 const struct tcphdr *th) 3338 { 3339 struct scatterlist sg; 3340 struct tcphdr hdr; 3341 int err; 3342 3343 /* We are not allowed to change tcphdr, make a local copy */ 3344 memcpy(&hdr, th, sizeof(hdr)); 3345 hdr.check = 0; 3346 3347 /* options aren't included in the hash */ 3348 sg_init_one(&sg, &hdr, sizeof(hdr)); 3349 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr)); 3350 return err; 3351 } 3352 EXPORT_SYMBOL(tcp_md5_hash_header); 3353 3354 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp, 3355 const struct sk_buff *skb, unsigned int header_len) 3356 { 3357 struct scatterlist sg; 3358 const struct tcphdr *tp = tcp_hdr(skb); 3359 struct hash_desc *desc = &hp->md5_desc; 3360 unsigned int i; 3361 const unsigned int head_data_len = skb_headlen(skb) > header_len ? 3362 skb_headlen(skb) - header_len : 0; 3363 const struct skb_shared_info *shi = skb_shinfo(skb); 3364 struct sk_buff *frag_iter; 3365 3366 sg_init_table(&sg, 1); 3367 3368 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len); 3369 if (crypto_hash_update(desc, &sg, head_data_len)) 3370 return 1; 3371 3372 for (i = 0; i < shi->nr_frags; ++i) { 3373 const struct skb_frag_struct *f = &shi->frags[i]; 3374 struct page *page = skb_frag_page(f); 3375 sg_set_page(&sg, page, skb_frag_size(f), f->page_offset); 3376 if (crypto_hash_update(desc, &sg, skb_frag_size(f))) 3377 return 1; 3378 } 3379 3380 skb_walk_frags(skb, frag_iter) 3381 if (tcp_md5_hash_skb_data(hp, frag_iter, 0)) 3382 return 1; 3383 3384 return 0; 3385 } 3386 EXPORT_SYMBOL(tcp_md5_hash_skb_data); 3387 3388 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key) 3389 { 3390 struct scatterlist sg; 3391 3392 sg_init_one(&sg, key->key, key->keylen); 3393 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen); 3394 } 3395 EXPORT_SYMBOL(tcp_md5_hash_key); 3396 3397 #endif 3398 3399 /* Each Responder maintains up to two secret values concurrently for 3400 * efficient secret rollover. Each secret value has 4 states: 3401 * 3402 * Generating. (tcp_secret_generating != tcp_secret_primary) 3403 * Generates new Responder-Cookies, but not yet used for primary 3404 * verification. This is a short-term state, typically lasting only 3405 * one round trip time (RTT). 3406 * 3407 * Primary. (tcp_secret_generating == tcp_secret_primary) 3408 * Used both for generation and primary verification. 3409 * 3410 * Retiring. (tcp_secret_retiring != tcp_secret_secondary) 3411 * Used for verification, until the first failure that can be 3412 * verified by the newer Generating secret. At that time, this 3413 * cookie's state is changed to Secondary, and the Generating 3414 * cookie's state is changed to Primary. This is a short-term state, 3415 * typically lasting only one round trip time (RTT). 3416 * 3417 * Secondary. (tcp_secret_retiring == tcp_secret_secondary) 3418 * Used for secondary verification, after primary verification 3419 * failures. This state lasts no more than twice the Maximum Segment 3420 * Lifetime (2MSL). Then, the secret is discarded. 3421 */ 3422 struct tcp_cookie_secret { 3423 /* The secret is divided into two parts. The digest part is the 3424 * equivalent of previously hashing a secret and saving the state, 3425 * and serves as an initialization vector (IV). The message part 3426 * serves as the trailing secret. 3427 */ 3428 u32 secrets[COOKIE_WORKSPACE_WORDS]; 3429 unsigned long expires; 3430 }; 3431 3432 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL) 3433 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2) 3434 #define TCP_SECRET_LIFE (HZ * 600) 3435 3436 static struct tcp_cookie_secret tcp_secret_one; 3437 static struct tcp_cookie_secret tcp_secret_two; 3438 3439 /* Essentially a circular list, without dynamic allocation. */ 3440 static struct tcp_cookie_secret *tcp_secret_generating; 3441 static struct tcp_cookie_secret *tcp_secret_primary; 3442 static struct tcp_cookie_secret *tcp_secret_retiring; 3443 static struct tcp_cookie_secret *tcp_secret_secondary; 3444 3445 static DEFINE_SPINLOCK(tcp_secret_locker); 3446 3447 /* Select a pseudo-random word in the cookie workspace. 3448 */ 3449 static inline u32 tcp_cookie_work(const u32 *ws, const int n) 3450 { 3451 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])]; 3452 } 3453 3454 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed. 3455 * Called in softirq context. 3456 * Returns: 0 for success. 3457 */ 3458 int tcp_cookie_generator(u32 *bakery) 3459 { 3460 unsigned long jiffy = jiffies; 3461 3462 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) { 3463 spin_lock_bh(&tcp_secret_locker); 3464 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) { 3465 /* refreshed by another */ 3466 memcpy(bakery, 3467 &tcp_secret_generating->secrets[0], 3468 COOKIE_WORKSPACE_WORDS); 3469 } else { 3470 /* still needs refreshing */ 3471 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS); 3472 3473 /* The first time, paranoia assumes that the 3474 * randomization function isn't as strong. But, 3475 * this secret initialization is delayed until 3476 * the last possible moment (packet arrival). 3477 * Although that time is observable, it is 3478 * unpredictably variable. Mash in the most 3479 * volatile clock bits available, and expire the 3480 * secret extra quickly. 3481 */ 3482 if (unlikely(tcp_secret_primary->expires == 3483 tcp_secret_secondary->expires)) { 3484 struct timespec tv; 3485 3486 getnstimeofday(&tv); 3487 bakery[COOKIE_DIGEST_WORDS+0] ^= 3488 (u32)tv.tv_nsec; 3489 3490 tcp_secret_secondary->expires = jiffy 3491 + TCP_SECRET_1MSL 3492 + (0x0f & tcp_cookie_work(bakery, 0)); 3493 } else { 3494 tcp_secret_secondary->expires = jiffy 3495 + TCP_SECRET_LIFE 3496 + (0xff & tcp_cookie_work(bakery, 1)); 3497 tcp_secret_primary->expires = jiffy 3498 + TCP_SECRET_2MSL 3499 + (0x1f & tcp_cookie_work(bakery, 2)); 3500 } 3501 memcpy(&tcp_secret_secondary->secrets[0], 3502 bakery, COOKIE_WORKSPACE_WORDS); 3503 3504 rcu_assign_pointer(tcp_secret_generating, 3505 tcp_secret_secondary); 3506 rcu_assign_pointer(tcp_secret_retiring, 3507 tcp_secret_primary); 3508 /* 3509 * Neither call_rcu() nor synchronize_rcu() needed. 3510 * Retiring data is not freed. It is replaced after 3511 * further (locked) pointer updates, and a quiet time 3512 * (minimum 1MSL, maximum LIFE - 2MSL). 3513 */ 3514 } 3515 spin_unlock_bh(&tcp_secret_locker); 3516 } else { 3517 rcu_read_lock_bh(); 3518 memcpy(bakery, 3519 &rcu_dereference(tcp_secret_generating)->secrets[0], 3520 COOKIE_WORKSPACE_WORDS); 3521 rcu_read_unlock_bh(); 3522 } 3523 return 0; 3524 } 3525 EXPORT_SYMBOL(tcp_cookie_generator); 3526 3527 void tcp_done(struct sock *sk) 3528 { 3529 struct request_sock *req = tcp_sk(sk)->fastopen_rsk; 3530 3531 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV) 3532 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 3533 3534 tcp_set_state(sk, TCP_CLOSE); 3535 tcp_clear_xmit_timers(sk); 3536 if (req != NULL) 3537 reqsk_fastopen_remove(sk, req, false); 3538 3539 sk->sk_shutdown = SHUTDOWN_MASK; 3540 3541 if (!sock_flag(sk, SOCK_DEAD)) 3542 sk->sk_state_change(sk); 3543 else 3544 inet_csk_destroy_sock(sk); 3545 } 3546 EXPORT_SYMBOL_GPL(tcp_done); 3547 3548 extern struct tcp_congestion_ops tcp_reno; 3549 3550 static __initdata unsigned long thash_entries; 3551 static int __init set_thash_entries(char *str) 3552 { 3553 ssize_t ret; 3554 3555 if (!str) 3556 return 0; 3557 3558 ret = kstrtoul(str, 0, &thash_entries); 3559 if (ret) 3560 return 0; 3561 3562 return 1; 3563 } 3564 __setup("thash_entries=", set_thash_entries); 3565 3566 void tcp_init_mem(struct net *net) 3567 { 3568 unsigned long limit = nr_free_buffer_pages() / 8; 3569 limit = max(limit, 128UL); 3570 net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3; 3571 net->ipv4.sysctl_tcp_mem[1] = limit; 3572 net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2; 3573 } 3574 3575 void __init tcp_init(void) 3576 { 3577 struct sk_buff *skb = NULL; 3578 unsigned long limit; 3579 int max_rshare, max_wshare, cnt; 3580 unsigned int i; 3581 unsigned long jiffy = jiffies; 3582 3583 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb)); 3584 3585 percpu_counter_init(&tcp_sockets_allocated, 0); 3586 percpu_counter_init(&tcp_orphan_count, 0); 3587 tcp_hashinfo.bind_bucket_cachep = 3588 kmem_cache_create("tcp_bind_bucket", 3589 sizeof(struct inet_bind_bucket), 0, 3590 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL); 3591 3592 /* Size and allocate the main established and bind bucket 3593 * hash tables. 3594 * 3595 * The methodology is similar to that of the buffer cache. 3596 */ 3597 tcp_hashinfo.ehash = 3598 alloc_large_system_hash("TCP established", 3599 sizeof(struct inet_ehash_bucket), 3600 thash_entries, 3601 17, /* one slot per 128 KB of memory */ 3602 0, 3603 NULL, 3604 &tcp_hashinfo.ehash_mask, 3605 0, 3606 thash_entries ? 0 : 512 * 1024); 3607 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) { 3608 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i); 3609 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i); 3610 } 3611 if (inet_ehash_locks_alloc(&tcp_hashinfo)) 3612 panic("TCP: failed to alloc ehash_locks"); 3613 tcp_hashinfo.bhash = 3614 alloc_large_system_hash("TCP bind", 3615 sizeof(struct inet_bind_hashbucket), 3616 tcp_hashinfo.ehash_mask + 1, 3617 17, /* one slot per 128 KB of memory */ 3618 0, 3619 &tcp_hashinfo.bhash_size, 3620 NULL, 3621 0, 3622 64 * 1024); 3623 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size; 3624 for (i = 0; i < tcp_hashinfo.bhash_size; i++) { 3625 spin_lock_init(&tcp_hashinfo.bhash[i].lock); 3626 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain); 3627 } 3628 3629 3630 cnt = tcp_hashinfo.ehash_mask + 1; 3631 3632 tcp_death_row.sysctl_max_tw_buckets = cnt / 2; 3633 sysctl_tcp_max_orphans = cnt / 2; 3634 sysctl_max_syn_backlog = max(128, cnt / 256); 3635 3636 tcp_init_mem(&init_net); 3637 /* Set per-socket limits to no more than 1/128 the pressure threshold */ 3638 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7); 3639 max_wshare = min(4UL*1024*1024, limit); 3640 max_rshare = min(6UL*1024*1024, limit); 3641 3642 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM; 3643 sysctl_tcp_wmem[1] = 16*1024; 3644 sysctl_tcp_wmem[2] = max(64*1024, max_wshare); 3645 3646 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM; 3647 sysctl_tcp_rmem[1] = 87380; 3648 sysctl_tcp_rmem[2] = max(87380, max_rshare); 3649 3650 pr_info("Hash tables configured (established %u bind %u)\n", 3651 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size); 3652 3653 tcp_metrics_init(); 3654 3655 tcp_register_congestion_control(&tcp_reno); 3656 3657 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets)); 3658 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets)); 3659 tcp_secret_one.expires = jiffy; /* past due */ 3660 tcp_secret_two.expires = jiffy; /* past due */ 3661 tcp_secret_generating = &tcp_secret_one; 3662 tcp_secret_primary = &tcp_secret_one; 3663 tcp_secret_retiring = &tcp_secret_two; 3664 tcp_secret_secondary = &tcp_secret_two; 3665 tcp_tasklet_init(); 3666 } 3667