1 /* 2 * INET An implementation of the TCP/IP protocol suite for the LINUX 3 * operating system. INET is implemented using the BSD Socket 4 * interface as the means of communication with the user level. 5 * 6 * Implementation of the Transmission Control Protocol(TCP). 7 * 8 * Authors: Ross Biro 9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 10 * Mark Evans, <evansmp@uhura.aston.ac.uk> 11 * Corey Minyard <wf-rch!minyard@relay.EU.net> 12 * Florian La Roche, <flla@stud.uni-sb.de> 13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 14 * Linus Torvalds, <torvalds@cs.helsinki.fi> 15 * Alan Cox, <gw4pts@gw4pts.ampr.org> 16 * Matthew Dillon, <dillon@apollo.west.oic.com> 17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 18 * Jorge Cwik, <jorge@laser.satlink.net> 19 * 20 * Fixes: 21 * Alan Cox : Numerous verify_area() calls 22 * Alan Cox : Set the ACK bit on a reset 23 * Alan Cox : Stopped it crashing if it closed while 24 * sk->inuse=1 and was trying to connect 25 * (tcp_err()). 26 * Alan Cox : All icmp error handling was broken 27 * pointers passed where wrong and the 28 * socket was looked up backwards. Nobody 29 * tested any icmp error code obviously. 30 * Alan Cox : tcp_err() now handled properly. It 31 * wakes people on errors. poll 32 * behaves and the icmp error race 33 * has gone by moving it into sock.c 34 * Alan Cox : tcp_send_reset() fixed to work for 35 * everything not just packets for 36 * unknown sockets. 37 * Alan Cox : tcp option processing. 38 * Alan Cox : Reset tweaked (still not 100%) [Had 39 * syn rule wrong] 40 * Herp Rosmanith : More reset fixes 41 * Alan Cox : No longer acks invalid rst frames. 42 * Acking any kind of RST is right out. 43 * Alan Cox : Sets an ignore me flag on an rst 44 * receive otherwise odd bits of prattle 45 * escape still 46 * Alan Cox : Fixed another acking RST frame bug. 47 * Should stop LAN workplace lockups. 48 * Alan Cox : Some tidyups using the new skb list 49 * facilities 50 * Alan Cox : sk->keepopen now seems to work 51 * Alan Cox : Pulls options out correctly on accepts 52 * Alan Cox : Fixed assorted sk->rqueue->next errors 53 * Alan Cox : PSH doesn't end a TCP read. Switched a 54 * bit to skb ops. 55 * Alan Cox : Tidied tcp_data to avoid a potential 56 * nasty. 57 * Alan Cox : Added some better commenting, as the 58 * tcp is hard to follow 59 * Alan Cox : Removed incorrect check for 20 * psh 60 * Michael O'Reilly : ack < copied bug fix. 61 * Johannes Stille : Misc tcp fixes (not all in yet). 62 * Alan Cox : FIN with no memory -> CRASH 63 * Alan Cox : Added socket option proto entries. 64 * Also added awareness of them to accept. 65 * Alan Cox : Added TCP options (SOL_TCP) 66 * Alan Cox : Switched wakeup calls to callbacks, 67 * so the kernel can layer network 68 * sockets. 69 * Alan Cox : Use ip_tos/ip_ttl settings. 70 * Alan Cox : Handle FIN (more) properly (we hope). 71 * Alan Cox : RST frames sent on unsynchronised 72 * state ack error. 73 * Alan Cox : Put in missing check for SYN bit. 74 * Alan Cox : Added tcp_select_window() aka NET2E 75 * window non shrink trick. 76 * Alan Cox : Added a couple of small NET2E timer 77 * fixes 78 * Charles Hedrick : TCP fixes 79 * Toomas Tamm : TCP window fixes 80 * Alan Cox : Small URG fix to rlogin ^C ack fight 81 * Charles Hedrick : Rewrote most of it to actually work 82 * Linus : Rewrote tcp_read() and URG handling 83 * completely 84 * Gerhard Koerting: Fixed some missing timer handling 85 * Matthew Dillon : Reworked TCP machine states as per RFC 86 * Gerhard Koerting: PC/TCP workarounds 87 * Adam Caldwell : Assorted timer/timing errors 88 * Matthew Dillon : Fixed another RST bug 89 * Alan Cox : Move to kernel side addressing changes. 90 * Alan Cox : Beginning work on TCP fastpathing 91 * (not yet usable) 92 * Arnt Gulbrandsen: Turbocharged tcp_check() routine. 93 * Alan Cox : TCP fast path debugging 94 * Alan Cox : Window clamping 95 * Michael Riepe : Bug in tcp_check() 96 * Matt Dillon : More TCP improvements and RST bug fixes 97 * Matt Dillon : Yet more small nasties remove from the 98 * TCP code (Be very nice to this man if 99 * tcp finally works 100%) 8) 100 * Alan Cox : BSD accept semantics. 101 * Alan Cox : Reset on closedown bug. 102 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto(). 103 * Michael Pall : Handle poll() after URG properly in 104 * all cases. 105 * Michael Pall : Undo the last fix in tcp_read_urg() 106 * (multi URG PUSH broke rlogin). 107 * Michael Pall : Fix the multi URG PUSH problem in 108 * tcp_readable(), poll() after URG 109 * works now. 110 * Michael Pall : recv(...,MSG_OOB) never blocks in the 111 * BSD api. 112 * Alan Cox : Changed the semantics of sk->socket to 113 * fix a race and a signal problem with 114 * accept() and async I/O. 115 * Alan Cox : Relaxed the rules on tcp_sendto(). 116 * Yury Shevchuk : Really fixed accept() blocking problem. 117 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for 118 * clients/servers which listen in on 119 * fixed ports. 120 * Alan Cox : Cleaned the above up and shrank it to 121 * a sensible code size. 122 * Alan Cox : Self connect lockup fix. 123 * Alan Cox : No connect to multicast. 124 * Ross Biro : Close unaccepted children on master 125 * socket close. 126 * Alan Cox : Reset tracing code. 127 * Alan Cox : Spurious resets on shutdown. 128 * Alan Cox : Giant 15 minute/60 second timer error 129 * Alan Cox : Small whoops in polling before an 130 * accept. 131 * Alan Cox : Kept the state trace facility since 132 * it's handy for debugging. 133 * Alan Cox : More reset handler fixes. 134 * Alan Cox : Started rewriting the code based on 135 * the RFC's for other useful protocol 136 * references see: Comer, KA9Q NOS, and 137 * for a reference on the difference 138 * between specifications and how BSD 139 * works see the 4.4lite source. 140 * A.N.Kuznetsov : Don't time wait on completion of tidy 141 * close. 142 * Linus Torvalds : Fin/Shutdown & copied_seq changes. 143 * Linus Torvalds : Fixed BSD port reuse to work first syn 144 * Alan Cox : Reimplemented timers as per the RFC 145 * and using multiple timers for sanity. 146 * Alan Cox : Small bug fixes, and a lot of new 147 * comments. 148 * Alan Cox : Fixed dual reader crash by locking 149 * the buffers (much like datagram.c) 150 * Alan Cox : Fixed stuck sockets in probe. A probe 151 * now gets fed up of retrying without 152 * (even a no space) answer. 153 * Alan Cox : Extracted closing code better 154 * Alan Cox : Fixed the closing state machine to 155 * resemble the RFC. 156 * Alan Cox : More 'per spec' fixes. 157 * Jorge Cwik : Even faster checksumming. 158 * Alan Cox : tcp_data() doesn't ack illegal PSH 159 * only frames. At least one pc tcp stack 160 * generates them. 161 * Alan Cox : Cache last socket. 162 * Alan Cox : Per route irtt. 163 * Matt Day : poll()->select() match BSD precisely on error 164 * Alan Cox : New buffers 165 * Marc Tamsky : Various sk->prot->retransmits and 166 * sk->retransmits misupdating fixed. 167 * Fixed tcp_write_timeout: stuck close, 168 * and TCP syn retries gets used now. 169 * Mark Yarvis : In tcp_read_wakeup(), don't send an 170 * ack if state is TCP_CLOSED. 171 * Alan Cox : Look up device on a retransmit - routes may 172 * change. Doesn't yet cope with MSS shrink right 173 * but it's a start! 174 * Marc Tamsky : Closing in closing fixes. 175 * Mike Shaver : RFC1122 verifications. 176 * Alan Cox : rcv_saddr errors. 177 * Alan Cox : Block double connect(). 178 * Alan Cox : Small hooks for enSKIP. 179 * Alexey Kuznetsov: Path MTU discovery. 180 * Alan Cox : Support soft errors. 181 * Alan Cox : Fix MTU discovery pathological case 182 * when the remote claims no mtu! 183 * Marc Tamsky : TCP_CLOSE fix. 184 * Colin (G3TNE) : Send a reset on syn ack replies in 185 * window but wrong (fixes NT lpd problems) 186 * Pedro Roque : Better TCP window handling, delayed ack. 187 * Joerg Reuter : No modification of locked buffers in 188 * tcp_do_retransmit() 189 * Eric Schenk : Changed receiver side silly window 190 * avoidance algorithm to BSD style 191 * algorithm. This doubles throughput 192 * against machines running Solaris, 193 * and seems to result in general 194 * improvement. 195 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD 196 * Willy Konynenberg : Transparent proxying support. 197 * Mike McLagan : Routing by source 198 * Keith Owens : Do proper merging with partial SKB's in 199 * tcp_do_sendmsg to avoid burstiness. 200 * Eric Schenk : Fix fast close down bug with 201 * shutdown() followed by close(). 202 * Andi Kleen : Make poll agree with SIGIO 203 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and 204 * lingertime == 0 (RFC 793 ABORT Call) 205 * Hirokazu Takahashi : Use copy_from_user() instead of 206 * csum_and_copy_from_user() if possible. 207 * 208 * This program is free software; you can redistribute it and/or 209 * modify it under the terms of the GNU General Public License 210 * as published by the Free Software Foundation; either version 211 * 2 of the License, or(at your option) any later version. 212 * 213 * Description of States: 214 * 215 * TCP_SYN_SENT sent a connection request, waiting for ack 216 * 217 * TCP_SYN_RECV received a connection request, sent ack, 218 * waiting for final ack in three-way handshake. 219 * 220 * TCP_ESTABLISHED connection established 221 * 222 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete 223 * transmission of remaining buffered data 224 * 225 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote 226 * to shutdown 227 * 228 * TCP_CLOSING both sides have shutdown but we still have 229 * data we have to finish sending 230 * 231 * TCP_TIME_WAIT timeout to catch resent junk before entering 232 * closed, can only be entered from FIN_WAIT2 233 * or CLOSING. Required because the other end 234 * may not have gotten our last ACK causing it 235 * to retransmit the data packet (which we ignore) 236 * 237 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for 238 * us to finish writing our data and to shutdown 239 * (we have to close() to move on to LAST_ACK) 240 * 241 * TCP_LAST_ACK out side has shutdown after remote has 242 * shutdown. There may still be data in our 243 * buffer that we have to finish sending 244 * 245 * TCP_CLOSE socket is finished 246 */ 247 248 #define pr_fmt(fmt) "TCP: " fmt 249 250 #include <crypto/hash.h> 251 #include <linux/kernel.h> 252 #include <linux/module.h> 253 #include <linux/types.h> 254 #include <linux/fcntl.h> 255 #include <linux/poll.h> 256 #include <linux/inet_diag.h> 257 #include <linux/init.h> 258 #include <linux/fs.h> 259 #include <linux/skbuff.h> 260 #include <linux/scatterlist.h> 261 #include <linux/splice.h> 262 #include <linux/net.h> 263 #include <linux/socket.h> 264 #include <linux/random.h> 265 #include <linux/bootmem.h> 266 #include <linux/highmem.h> 267 #include <linux/swap.h> 268 #include <linux/cache.h> 269 #include <linux/err.h> 270 #include <linux/time.h> 271 #include <linux/slab.h> 272 #include <linux/errqueue.h> 273 #include <linux/static_key.h> 274 275 #include <net/icmp.h> 276 #include <net/inet_common.h> 277 #include <net/tcp.h> 278 #include <net/xfrm.h> 279 #include <net/ip.h> 280 #include <net/sock.h> 281 282 #include <linux/uaccess.h> 283 #include <asm/ioctls.h> 284 #include <net/busy_poll.h> 285 286 struct percpu_counter tcp_orphan_count; 287 EXPORT_SYMBOL_GPL(tcp_orphan_count); 288 289 long sysctl_tcp_mem[3] __read_mostly; 290 EXPORT_SYMBOL(sysctl_tcp_mem); 291 292 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */ 293 EXPORT_SYMBOL(tcp_memory_allocated); 294 295 #if IS_ENABLED(CONFIG_SMC) 296 DEFINE_STATIC_KEY_FALSE(tcp_have_smc); 297 EXPORT_SYMBOL(tcp_have_smc); 298 #endif 299 300 /* 301 * Current number of TCP sockets. 302 */ 303 struct percpu_counter tcp_sockets_allocated; 304 EXPORT_SYMBOL(tcp_sockets_allocated); 305 306 /* 307 * TCP splice context 308 */ 309 struct tcp_splice_state { 310 struct pipe_inode_info *pipe; 311 size_t len; 312 unsigned int flags; 313 }; 314 315 /* 316 * Pressure flag: try to collapse. 317 * Technical note: it is used by multiple contexts non atomically. 318 * All the __sk_mem_schedule() is of this nature: accounting 319 * is strict, actions are advisory and have some latency. 320 */ 321 unsigned long tcp_memory_pressure __read_mostly; 322 EXPORT_SYMBOL_GPL(tcp_memory_pressure); 323 324 void tcp_enter_memory_pressure(struct sock *sk) 325 { 326 unsigned long val; 327 328 if (tcp_memory_pressure) 329 return; 330 val = jiffies; 331 332 if (!val) 333 val--; 334 if (!cmpxchg(&tcp_memory_pressure, 0, val)) 335 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES); 336 } 337 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure); 338 339 void tcp_leave_memory_pressure(struct sock *sk) 340 { 341 unsigned long val; 342 343 if (!tcp_memory_pressure) 344 return; 345 val = xchg(&tcp_memory_pressure, 0); 346 if (val) 347 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO, 348 jiffies_to_msecs(jiffies - val)); 349 } 350 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure); 351 352 /* Convert seconds to retransmits based on initial and max timeout */ 353 static u8 secs_to_retrans(int seconds, int timeout, int rto_max) 354 { 355 u8 res = 0; 356 357 if (seconds > 0) { 358 int period = timeout; 359 360 res = 1; 361 while (seconds > period && res < 255) { 362 res++; 363 timeout <<= 1; 364 if (timeout > rto_max) 365 timeout = rto_max; 366 period += timeout; 367 } 368 } 369 return res; 370 } 371 372 /* Convert retransmits to seconds based on initial and max timeout */ 373 static int retrans_to_secs(u8 retrans, int timeout, int rto_max) 374 { 375 int period = 0; 376 377 if (retrans > 0) { 378 period = timeout; 379 while (--retrans) { 380 timeout <<= 1; 381 if (timeout > rto_max) 382 timeout = rto_max; 383 period += timeout; 384 } 385 } 386 return period; 387 } 388 389 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp) 390 { 391 u32 rate = READ_ONCE(tp->rate_delivered); 392 u32 intv = READ_ONCE(tp->rate_interval_us); 393 u64 rate64 = 0; 394 395 if (rate && intv) { 396 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC; 397 do_div(rate64, intv); 398 } 399 return rate64; 400 } 401 402 /* Address-family independent initialization for a tcp_sock. 403 * 404 * NOTE: A lot of things set to zero explicitly by call to 405 * sk_alloc() so need not be done here. 406 */ 407 void tcp_init_sock(struct sock *sk) 408 { 409 struct inet_connection_sock *icsk = inet_csk(sk); 410 struct tcp_sock *tp = tcp_sk(sk); 411 412 tp->out_of_order_queue = RB_ROOT; 413 sk->tcp_rtx_queue = RB_ROOT; 414 tcp_init_xmit_timers(sk); 415 INIT_LIST_HEAD(&tp->tsq_node); 416 INIT_LIST_HEAD(&tp->tsorted_sent_queue); 417 418 icsk->icsk_rto = TCP_TIMEOUT_INIT; 419 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT); 420 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U); 421 422 /* So many TCP implementations out there (incorrectly) count the 423 * initial SYN frame in their delayed-ACK and congestion control 424 * algorithms that we must have the following bandaid to talk 425 * efficiently to them. -DaveM 426 */ 427 tp->snd_cwnd = TCP_INIT_CWND; 428 429 /* There's a bubble in the pipe until at least the first ACK. */ 430 tp->app_limited = ~0U; 431 432 /* See draft-stevens-tcpca-spec-01 for discussion of the 433 * initialization of these values. 434 */ 435 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 436 tp->snd_cwnd_clamp = ~0; 437 tp->mss_cache = TCP_MSS_DEFAULT; 438 439 tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering; 440 tcp_assign_congestion_control(sk); 441 442 tp->tsoffset = 0; 443 tp->rack.reo_wnd_steps = 1; 444 445 sk->sk_state = TCP_CLOSE; 446 447 sk->sk_write_space = sk_stream_write_space; 448 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE); 449 450 icsk->icsk_sync_mss = tcp_sync_mss; 451 452 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1]; 453 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1]; 454 455 sk_sockets_allocated_inc(sk); 456 sk->sk_route_forced_caps = NETIF_F_GSO; 457 } 458 EXPORT_SYMBOL(tcp_init_sock); 459 460 void tcp_init_transfer(struct sock *sk, int bpf_op) 461 { 462 struct inet_connection_sock *icsk = inet_csk(sk); 463 464 tcp_mtup_init(sk); 465 icsk->icsk_af_ops->rebuild_header(sk); 466 tcp_init_metrics(sk); 467 tcp_call_bpf(sk, bpf_op, 0, NULL); 468 tcp_init_congestion_control(sk); 469 tcp_init_buffer_space(sk); 470 } 471 472 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags) 473 { 474 struct sk_buff *skb = tcp_write_queue_tail(sk); 475 476 if (tsflags && skb) { 477 struct skb_shared_info *shinfo = skb_shinfo(skb); 478 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 479 480 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags); 481 if (tsflags & SOF_TIMESTAMPING_TX_ACK) 482 tcb->txstamp_ack = 1; 483 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK) 484 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1; 485 } 486 } 487 488 static inline bool tcp_stream_is_readable(const struct tcp_sock *tp, 489 int target, struct sock *sk) 490 { 491 return (tp->rcv_nxt - tp->copied_seq >= target) || 492 (sk->sk_prot->stream_memory_read ? 493 sk->sk_prot->stream_memory_read(sk) : false); 494 } 495 496 /* 497 * Socket is not locked. We are protected from async events by poll logic and 498 * correct handling of state changes made by other threads is impossible in 499 * any case. 500 */ 501 __poll_t tcp_poll_mask(struct socket *sock, __poll_t events) 502 { 503 struct sock *sk = sock->sk; 504 const struct tcp_sock *tp = tcp_sk(sk); 505 __poll_t mask = 0; 506 int state; 507 508 state = inet_sk_state_load(sk); 509 if (state == TCP_LISTEN) 510 return inet_csk_listen_poll(sk); 511 512 /* 513 * EPOLLHUP is certainly not done right. But poll() doesn't 514 * have a notion of HUP in just one direction, and for a 515 * socket the read side is more interesting. 516 * 517 * Some poll() documentation says that EPOLLHUP is incompatible 518 * with the EPOLLOUT/POLLWR flags, so somebody should check this 519 * all. But careful, it tends to be safer to return too many 520 * bits than too few, and you can easily break real applications 521 * if you don't tell them that something has hung up! 522 * 523 * Check-me. 524 * 525 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and 526 * our fs/select.c). It means that after we received EOF, 527 * poll always returns immediately, making impossible poll() on write() 528 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP 529 * if and only if shutdown has been made in both directions. 530 * Actually, it is interesting to look how Solaris and DUX 531 * solve this dilemma. I would prefer, if EPOLLHUP were maskable, 532 * then we could set it on SND_SHUTDOWN. BTW examples given 533 * in Stevens' books assume exactly this behaviour, it explains 534 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK 535 * 536 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent 537 * blocking on fresh not-connected or disconnected socket. --ANK 538 */ 539 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE) 540 mask |= EPOLLHUP; 541 if (sk->sk_shutdown & RCV_SHUTDOWN) 542 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP; 543 544 /* Connected or passive Fast Open socket? */ 545 if (state != TCP_SYN_SENT && 546 (state != TCP_SYN_RECV || tp->fastopen_rsk)) { 547 int target = sock_rcvlowat(sk, 0, INT_MAX); 548 549 if (tp->urg_seq == tp->copied_seq && 550 !sock_flag(sk, SOCK_URGINLINE) && 551 tp->urg_data) 552 target++; 553 554 if (tcp_stream_is_readable(tp, target, sk)) 555 mask |= EPOLLIN | EPOLLRDNORM; 556 557 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) { 558 if (sk_stream_is_writeable(sk)) { 559 mask |= EPOLLOUT | EPOLLWRNORM; 560 } else { /* send SIGIO later */ 561 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk); 562 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 563 564 /* Race breaker. If space is freed after 565 * wspace test but before the flags are set, 566 * IO signal will be lost. Memory barrier 567 * pairs with the input side. 568 */ 569 smp_mb__after_atomic(); 570 if (sk_stream_is_writeable(sk)) 571 mask |= EPOLLOUT | EPOLLWRNORM; 572 } 573 } else 574 mask |= EPOLLOUT | EPOLLWRNORM; 575 576 if (tp->urg_data & TCP_URG_VALID) 577 mask |= EPOLLPRI; 578 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) { 579 /* Active TCP fastopen socket with defer_connect 580 * Return EPOLLOUT so application can call write() 581 * in order for kernel to generate SYN+data 582 */ 583 mask |= EPOLLOUT | EPOLLWRNORM; 584 } 585 /* This barrier is coupled with smp_wmb() in tcp_reset() */ 586 smp_rmb(); 587 if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue)) 588 mask |= EPOLLERR; 589 590 return mask; 591 } 592 EXPORT_SYMBOL(tcp_poll_mask); 593 594 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg) 595 { 596 struct tcp_sock *tp = tcp_sk(sk); 597 int answ; 598 bool slow; 599 600 switch (cmd) { 601 case SIOCINQ: 602 if (sk->sk_state == TCP_LISTEN) 603 return -EINVAL; 604 605 slow = lock_sock_fast(sk); 606 answ = tcp_inq(sk); 607 unlock_sock_fast(sk, slow); 608 break; 609 case SIOCATMARK: 610 answ = tp->urg_data && tp->urg_seq == tp->copied_seq; 611 break; 612 case SIOCOUTQ: 613 if (sk->sk_state == TCP_LISTEN) 614 return -EINVAL; 615 616 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 617 answ = 0; 618 else 619 answ = tp->write_seq - tp->snd_una; 620 break; 621 case SIOCOUTQNSD: 622 if (sk->sk_state == TCP_LISTEN) 623 return -EINVAL; 624 625 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 626 answ = 0; 627 else 628 answ = tp->write_seq - tp->snd_nxt; 629 break; 630 default: 631 return -ENOIOCTLCMD; 632 } 633 634 return put_user(answ, (int __user *)arg); 635 } 636 EXPORT_SYMBOL(tcp_ioctl); 637 638 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb) 639 { 640 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 641 tp->pushed_seq = tp->write_seq; 642 } 643 644 static inline bool forced_push(const struct tcp_sock *tp) 645 { 646 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1)); 647 } 648 649 static void skb_entail(struct sock *sk, struct sk_buff *skb) 650 { 651 struct tcp_sock *tp = tcp_sk(sk); 652 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 653 654 skb->csum = 0; 655 tcb->seq = tcb->end_seq = tp->write_seq; 656 tcb->tcp_flags = TCPHDR_ACK; 657 tcb->sacked = 0; 658 __skb_header_release(skb); 659 tcp_add_write_queue_tail(sk, skb); 660 sk->sk_wmem_queued += skb->truesize; 661 sk_mem_charge(sk, skb->truesize); 662 if (tp->nonagle & TCP_NAGLE_PUSH) 663 tp->nonagle &= ~TCP_NAGLE_PUSH; 664 665 tcp_slow_start_after_idle_check(sk); 666 } 667 668 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags) 669 { 670 if (flags & MSG_OOB) 671 tp->snd_up = tp->write_seq; 672 } 673 674 /* If a not yet filled skb is pushed, do not send it if 675 * we have data packets in Qdisc or NIC queues : 676 * Because TX completion will happen shortly, it gives a chance 677 * to coalesce future sendmsg() payload into this skb, without 678 * need for a timer, and with no latency trade off. 679 * As packets containing data payload have a bigger truesize 680 * than pure acks (dataless) packets, the last checks prevent 681 * autocorking if we only have an ACK in Qdisc/NIC queues, 682 * or if TX completion was delayed after we processed ACK packet. 683 */ 684 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb, 685 int size_goal) 686 { 687 return skb->len < size_goal && 688 sock_net(sk)->ipv4.sysctl_tcp_autocorking && 689 skb != tcp_write_queue_head(sk) && 690 refcount_read(&sk->sk_wmem_alloc) > skb->truesize; 691 } 692 693 static void tcp_push(struct sock *sk, int flags, int mss_now, 694 int nonagle, int size_goal) 695 { 696 struct tcp_sock *tp = tcp_sk(sk); 697 struct sk_buff *skb; 698 699 skb = tcp_write_queue_tail(sk); 700 if (!skb) 701 return; 702 if (!(flags & MSG_MORE) || forced_push(tp)) 703 tcp_mark_push(tp, skb); 704 705 tcp_mark_urg(tp, flags); 706 707 if (tcp_should_autocork(sk, skb, size_goal)) { 708 709 /* avoid atomic op if TSQ_THROTTLED bit is already set */ 710 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) { 711 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING); 712 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 713 } 714 /* It is possible TX completion already happened 715 * before we set TSQ_THROTTLED. 716 */ 717 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize) 718 return; 719 } 720 721 if (flags & MSG_MORE) 722 nonagle = TCP_NAGLE_CORK; 723 724 __tcp_push_pending_frames(sk, mss_now, nonagle); 725 } 726 727 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, 728 unsigned int offset, size_t len) 729 { 730 struct tcp_splice_state *tss = rd_desc->arg.data; 731 int ret; 732 733 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe, 734 min(rd_desc->count, len), tss->flags); 735 if (ret > 0) 736 rd_desc->count -= ret; 737 return ret; 738 } 739 740 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss) 741 { 742 /* Store TCP splice context information in read_descriptor_t. */ 743 read_descriptor_t rd_desc = { 744 .arg.data = tss, 745 .count = tss->len, 746 }; 747 748 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv); 749 } 750 751 /** 752 * tcp_splice_read - splice data from TCP socket to a pipe 753 * @sock: socket to splice from 754 * @ppos: position (not valid) 755 * @pipe: pipe to splice to 756 * @len: number of bytes to splice 757 * @flags: splice modifier flags 758 * 759 * Description: 760 * Will read pages from given socket and fill them into a pipe. 761 * 762 **/ 763 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos, 764 struct pipe_inode_info *pipe, size_t len, 765 unsigned int flags) 766 { 767 struct sock *sk = sock->sk; 768 struct tcp_splice_state tss = { 769 .pipe = pipe, 770 .len = len, 771 .flags = flags, 772 }; 773 long timeo; 774 ssize_t spliced; 775 int ret; 776 777 sock_rps_record_flow(sk); 778 /* 779 * We can't seek on a socket input 780 */ 781 if (unlikely(*ppos)) 782 return -ESPIPE; 783 784 ret = spliced = 0; 785 786 lock_sock(sk); 787 788 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK); 789 while (tss.len) { 790 ret = __tcp_splice_read(sk, &tss); 791 if (ret < 0) 792 break; 793 else if (!ret) { 794 if (spliced) 795 break; 796 if (sock_flag(sk, SOCK_DONE)) 797 break; 798 if (sk->sk_err) { 799 ret = sock_error(sk); 800 break; 801 } 802 if (sk->sk_shutdown & RCV_SHUTDOWN) 803 break; 804 if (sk->sk_state == TCP_CLOSE) { 805 /* 806 * This occurs when user tries to read 807 * from never connected socket. 808 */ 809 if (!sock_flag(sk, SOCK_DONE)) 810 ret = -ENOTCONN; 811 break; 812 } 813 if (!timeo) { 814 ret = -EAGAIN; 815 break; 816 } 817 /* if __tcp_splice_read() got nothing while we have 818 * an skb in receive queue, we do not want to loop. 819 * This might happen with URG data. 820 */ 821 if (!skb_queue_empty(&sk->sk_receive_queue)) 822 break; 823 sk_wait_data(sk, &timeo, NULL); 824 if (signal_pending(current)) { 825 ret = sock_intr_errno(timeo); 826 break; 827 } 828 continue; 829 } 830 tss.len -= ret; 831 spliced += ret; 832 833 if (!timeo) 834 break; 835 release_sock(sk); 836 lock_sock(sk); 837 838 if (sk->sk_err || sk->sk_state == TCP_CLOSE || 839 (sk->sk_shutdown & RCV_SHUTDOWN) || 840 signal_pending(current)) 841 break; 842 } 843 844 release_sock(sk); 845 846 if (spliced) 847 return spliced; 848 849 return ret; 850 } 851 EXPORT_SYMBOL(tcp_splice_read); 852 853 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp, 854 bool force_schedule) 855 { 856 struct sk_buff *skb; 857 858 /* The TCP header must be at least 32-bit aligned. */ 859 size = ALIGN(size, 4); 860 861 if (unlikely(tcp_under_memory_pressure(sk))) 862 sk_mem_reclaim_partial(sk); 863 864 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp); 865 if (likely(skb)) { 866 bool mem_scheduled; 867 868 if (force_schedule) { 869 mem_scheduled = true; 870 sk_forced_mem_schedule(sk, skb->truesize); 871 } else { 872 mem_scheduled = sk_wmem_schedule(sk, skb->truesize); 873 } 874 if (likely(mem_scheduled)) { 875 skb_reserve(skb, sk->sk_prot->max_header); 876 /* 877 * Make sure that we have exactly size bytes 878 * available to the caller, no more, no less. 879 */ 880 skb->reserved_tailroom = skb->end - skb->tail - size; 881 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 882 return skb; 883 } 884 __kfree_skb(skb); 885 } else { 886 sk->sk_prot->enter_memory_pressure(sk); 887 sk_stream_moderate_sndbuf(sk); 888 } 889 return NULL; 890 } 891 892 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now, 893 int large_allowed) 894 { 895 struct tcp_sock *tp = tcp_sk(sk); 896 u32 new_size_goal, size_goal; 897 898 if (!large_allowed) 899 return mss_now; 900 901 /* Note : tcp_tso_autosize() will eventually split this later */ 902 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER; 903 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal); 904 905 /* We try hard to avoid divides here */ 906 size_goal = tp->gso_segs * mss_now; 907 if (unlikely(new_size_goal < size_goal || 908 new_size_goal >= size_goal + mss_now)) { 909 tp->gso_segs = min_t(u16, new_size_goal / mss_now, 910 sk->sk_gso_max_segs); 911 size_goal = tp->gso_segs * mss_now; 912 } 913 914 return max(size_goal, mss_now); 915 } 916 917 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags) 918 { 919 int mss_now; 920 921 mss_now = tcp_current_mss(sk); 922 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB)); 923 924 return mss_now; 925 } 926 927 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset, 928 size_t size, int flags) 929 { 930 struct tcp_sock *tp = tcp_sk(sk); 931 int mss_now, size_goal; 932 int err; 933 ssize_t copied; 934 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 935 936 /* Wait for a connection to finish. One exception is TCP Fast Open 937 * (passive side) where data is allowed to be sent before a connection 938 * is fully established. 939 */ 940 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) && 941 !tcp_passive_fastopen(sk)) { 942 err = sk_stream_wait_connect(sk, &timeo); 943 if (err != 0) 944 goto out_err; 945 } 946 947 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk); 948 949 mss_now = tcp_send_mss(sk, &size_goal, flags); 950 copied = 0; 951 952 err = -EPIPE; 953 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 954 goto out_err; 955 956 while (size > 0) { 957 struct sk_buff *skb = tcp_write_queue_tail(sk); 958 int copy, i; 959 bool can_coalesce; 960 961 if (!skb || (copy = size_goal - skb->len) <= 0 || 962 !tcp_skb_can_collapse_to(skb)) { 963 new_segment: 964 if (!sk_stream_memory_free(sk)) 965 goto wait_for_sndbuf; 966 967 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, 968 tcp_rtx_and_write_queues_empty(sk)); 969 if (!skb) 970 goto wait_for_memory; 971 972 skb_entail(sk, skb); 973 copy = size_goal; 974 } 975 976 if (copy > size) 977 copy = size; 978 979 i = skb_shinfo(skb)->nr_frags; 980 can_coalesce = skb_can_coalesce(skb, i, page, offset); 981 if (!can_coalesce && i >= sysctl_max_skb_frags) { 982 tcp_mark_push(tp, skb); 983 goto new_segment; 984 } 985 if (!sk_wmem_schedule(sk, copy)) 986 goto wait_for_memory; 987 988 if (can_coalesce) { 989 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 990 } else { 991 get_page(page); 992 skb_fill_page_desc(skb, i, page, offset, copy); 993 } 994 995 if (!(flags & MSG_NO_SHARED_FRAGS)) 996 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG; 997 998 skb->len += copy; 999 skb->data_len += copy; 1000 skb->truesize += copy; 1001 sk->sk_wmem_queued += copy; 1002 sk_mem_charge(sk, copy); 1003 skb->ip_summed = CHECKSUM_PARTIAL; 1004 tp->write_seq += copy; 1005 TCP_SKB_CB(skb)->end_seq += copy; 1006 tcp_skb_pcount_set(skb, 0); 1007 1008 if (!copied) 1009 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 1010 1011 copied += copy; 1012 offset += copy; 1013 size -= copy; 1014 if (!size) 1015 goto out; 1016 1017 if (skb->len < size_goal || (flags & MSG_OOB)) 1018 continue; 1019 1020 if (forced_push(tp)) { 1021 tcp_mark_push(tp, skb); 1022 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 1023 } else if (skb == tcp_send_head(sk)) 1024 tcp_push_one(sk, mss_now); 1025 continue; 1026 1027 wait_for_sndbuf: 1028 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1029 wait_for_memory: 1030 tcp_push(sk, flags & ~MSG_MORE, mss_now, 1031 TCP_NAGLE_PUSH, size_goal); 1032 1033 err = sk_stream_wait_memory(sk, &timeo); 1034 if (err != 0) 1035 goto do_error; 1036 1037 mss_now = tcp_send_mss(sk, &size_goal, flags); 1038 } 1039 1040 out: 1041 if (copied) { 1042 tcp_tx_timestamp(sk, sk->sk_tsflags); 1043 if (!(flags & MSG_SENDPAGE_NOTLAST)) 1044 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal); 1045 } 1046 return copied; 1047 1048 do_error: 1049 if (copied) 1050 goto out; 1051 out_err: 1052 /* make sure we wake any epoll edge trigger waiter */ 1053 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && 1054 err == -EAGAIN)) { 1055 sk->sk_write_space(sk); 1056 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED); 1057 } 1058 return sk_stream_error(sk, flags, err); 1059 } 1060 EXPORT_SYMBOL_GPL(do_tcp_sendpages); 1061 1062 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset, 1063 size_t size, int flags) 1064 { 1065 if (!(sk->sk_route_caps & NETIF_F_SG)) 1066 return sock_no_sendpage_locked(sk, page, offset, size, flags); 1067 1068 tcp_rate_check_app_limited(sk); /* is sending application-limited? */ 1069 1070 return do_tcp_sendpages(sk, page, offset, size, flags); 1071 } 1072 EXPORT_SYMBOL_GPL(tcp_sendpage_locked); 1073 1074 int tcp_sendpage(struct sock *sk, struct page *page, int offset, 1075 size_t size, int flags) 1076 { 1077 int ret; 1078 1079 lock_sock(sk); 1080 ret = tcp_sendpage_locked(sk, page, offset, size, flags); 1081 release_sock(sk); 1082 1083 return ret; 1084 } 1085 EXPORT_SYMBOL(tcp_sendpage); 1086 1087 /* Do not bother using a page frag for very small frames. 1088 * But use this heuristic only for the first skb in write queue. 1089 * 1090 * Having no payload in skb->head allows better SACK shifting 1091 * in tcp_shift_skb_data(), reducing sack/rack overhead, because 1092 * write queue has less skbs. 1093 * Each skb can hold up to MAX_SKB_FRAGS * 32Kbytes, or ~0.5 MB. 1094 * This also speeds up tso_fragment(), since it wont fallback 1095 * to tcp_fragment(). 1096 */ 1097 static int linear_payload_sz(bool first_skb) 1098 { 1099 if (first_skb) 1100 return SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER); 1101 return 0; 1102 } 1103 1104 static int select_size(bool first_skb, bool zc) 1105 { 1106 if (zc) 1107 return 0; 1108 return linear_payload_sz(first_skb); 1109 } 1110 1111 void tcp_free_fastopen_req(struct tcp_sock *tp) 1112 { 1113 if (tp->fastopen_req) { 1114 kfree(tp->fastopen_req); 1115 tp->fastopen_req = NULL; 1116 } 1117 } 1118 1119 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, 1120 int *copied, size_t size) 1121 { 1122 struct tcp_sock *tp = tcp_sk(sk); 1123 struct inet_sock *inet = inet_sk(sk); 1124 struct sockaddr *uaddr = msg->msg_name; 1125 int err, flags; 1126 1127 if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) || 1128 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) && 1129 uaddr->sa_family == AF_UNSPEC)) 1130 return -EOPNOTSUPP; 1131 if (tp->fastopen_req) 1132 return -EALREADY; /* Another Fast Open is in progress */ 1133 1134 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request), 1135 sk->sk_allocation); 1136 if (unlikely(!tp->fastopen_req)) 1137 return -ENOBUFS; 1138 tp->fastopen_req->data = msg; 1139 tp->fastopen_req->size = size; 1140 1141 if (inet->defer_connect) { 1142 err = tcp_connect(sk); 1143 /* Same failure procedure as in tcp_v4/6_connect */ 1144 if (err) { 1145 tcp_set_state(sk, TCP_CLOSE); 1146 inet->inet_dport = 0; 1147 sk->sk_route_caps = 0; 1148 } 1149 } 1150 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0; 1151 err = __inet_stream_connect(sk->sk_socket, uaddr, 1152 msg->msg_namelen, flags, 1); 1153 /* fastopen_req could already be freed in __inet_stream_connect 1154 * if the connection times out or gets rst 1155 */ 1156 if (tp->fastopen_req) { 1157 *copied = tp->fastopen_req->copied; 1158 tcp_free_fastopen_req(tp); 1159 inet->defer_connect = 0; 1160 } 1161 return err; 1162 } 1163 1164 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size) 1165 { 1166 struct tcp_sock *tp = tcp_sk(sk); 1167 struct ubuf_info *uarg = NULL; 1168 struct sk_buff *skb; 1169 struct sockcm_cookie sockc; 1170 int flags, err, copied = 0; 1171 int mss_now = 0, size_goal, copied_syn = 0; 1172 bool process_backlog = false; 1173 bool zc = false; 1174 long timeo; 1175 1176 flags = msg->msg_flags; 1177 1178 if (flags & MSG_ZEROCOPY && size) { 1179 if (sk->sk_state != TCP_ESTABLISHED) { 1180 err = -EINVAL; 1181 goto out_err; 1182 } 1183 1184 skb = tcp_write_queue_tail(sk); 1185 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb)); 1186 if (!uarg) { 1187 err = -ENOBUFS; 1188 goto out_err; 1189 } 1190 1191 zc = sk->sk_route_caps & NETIF_F_SG; 1192 if (!zc) 1193 uarg->zerocopy = 0; 1194 } 1195 1196 if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect)) { 1197 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size); 1198 if (err == -EINPROGRESS && copied_syn > 0) 1199 goto out; 1200 else if (err) 1201 goto out_err; 1202 } 1203 1204 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 1205 1206 tcp_rate_check_app_limited(sk); /* is sending application-limited? */ 1207 1208 /* Wait for a connection to finish. One exception is TCP Fast Open 1209 * (passive side) where data is allowed to be sent before a connection 1210 * is fully established. 1211 */ 1212 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) && 1213 !tcp_passive_fastopen(sk)) { 1214 err = sk_stream_wait_connect(sk, &timeo); 1215 if (err != 0) 1216 goto do_error; 1217 } 1218 1219 if (unlikely(tp->repair)) { 1220 if (tp->repair_queue == TCP_RECV_QUEUE) { 1221 copied = tcp_send_rcvq(sk, msg, size); 1222 goto out_nopush; 1223 } 1224 1225 err = -EINVAL; 1226 if (tp->repair_queue == TCP_NO_QUEUE) 1227 goto out_err; 1228 1229 /* 'common' sending to sendq */ 1230 } 1231 1232 sockc.tsflags = sk->sk_tsflags; 1233 if (msg->msg_controllen) { 1234 err = sock_cmsg_send(sk, msg, &sockc); 1235 if (unlikely(err)) { 1236 err = -EINVAL; 1237 goto out_err; 1238 } 1239 } 1240 1241 /* This should be in poll */ 1242 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk); 1243 1244 /* Ok commence sending. */ 1245 copied = 0; 1246 1247 restart: 1248 mss_now = tcp_send_mss(sk, &size_goal, flags); 1249 1250 err = -EPIPE; 1251 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 1252 goto do_error; 1253 1254 while (msg_data_left(msg)) { 1255 int copy = 0; 1256 1257 skb = tcp_write_queue_tail(sk); 1258 if (skb) 1259 copy = size_goal - skb->len; 1260 1261 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) { 1262 bool first_skb; 1263 int linear; 1264 1265 new_segment: 1266 /* Allocate new segment. If the interface is SG, 1267 * allocate skb fitting to single page. 1268 */ 1269 if (!sk_stream_memory_free(sk)) 1270 goto wait_for_sndbuf; 1271 1272 if (process_backlog && sk_flush_backlog(sk)) { 1273 process_backlog = false; 1274 goto restart; 1275 } 1276 first_skb = tcp_rtx_and_write_queues_empty(sk); 1277 linear = select_size(first_skb, zc); 1278 skb = sk_stream_alloc_skb(sk, linear, sk->sk_allocation, 1279 first_skb); 1280 if (!skb) 1281 goto wait_for_memory; 1282 1283 process_backlog = true; 1284 skb->ip_summed = CHECKSUM_PARTIAL; 1285 1286 skb_entail(sk, skb); 1287 copy = size_goal; 1288 1289 /* All packets are restored as if they have 1290 * already been sent. skb_mstamp isn't set to 1291 * avoid wrong rtt estimation. 1292 */ 1293 if (tp->repair) 1294 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED; 1295 } 1296 1297 /* Try to append data to the end of skb. */ 1298 if (copy > msg_data_left(msg)) 1299 copy = msg_data_left(msg); 1300 1301 /* Where to copy to? */ 1302 if (skb_availroom(skb) > 0 && !zc) { 1303 /* We have some space in skb head. Superb! */ 1304 copy = min_t(int, copy, skb_availroom(skb)); 1305 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy); 1306 if (err) 1307 goto do_fault; 1308 } else if (!zc) { 1309 bool merge = true; 1310 int i = skb_shinfo(skb)->nr_frags; 1311 struct page_frag *pfrag = sk_page_frag(sk); 1312 1313 if (!sk_page_frag_refill(sk, pfrag)) 1314 goto wait_for_memory; 1315 1316 if (!skb_can_coalesce(skb, i, pfrag->page, 1317 pfrag->offset)) { 1318 if (i >= sysctl_max_skb_frags) { 1319 tcp_mark_push(tp, skb); 1320 goto new_segment; 1321 } 1322 merge = false; 1323 } 1324 1325 copy = min_t(int, copy, pfrag->size - pfrag->offset); 1326 1327 if (!sk_wmem_schedule(sk, copy)) 1328 goto wait_for_memory; 1329 1330 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb, 1331 pfrag->page, 1332 pfrag->offset, 1333 copy); 1334 if (err) 1335 goto do_error; 1336 1337 /* Update the skb. */ 1338 if (merge) { 1339 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 1340 } else { 1341 skb_fill_page_desc(skb, i, pfrag->page, 1342 pfrag->offset, copy); 1343 page_ref_inc(pfrag->page); 1344 } 1345 pfrag->offset += copy; 1346 } else { 1347 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg); 1348 if (err == -EMSGSIZE || err == -EEXIST) { 1349 tcp_mark_push(tp, skb); 1350 goto new_segment; 1351 } 1352 if (err < 0) 1353 goto do_error; 1354 copy = err; 1355 } 1356 1357 if (!copied) 1358 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 1359 1360 tp->write_seq += copy; 1361 TCP_SKB_CB(skb)->end_seq += copy; 1362 tcp_skb_pcount_set(skb, 0); 1363 1364 copied += copy; 1365 if (!msg_data_left(msg)) { 1366 if (unlikely(flags & MSG_EOR)) 1367 TCP_SKB_CB(skb)->eor = 1; 1368 goto out; 1369 } 1370 1371 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair)) 1372 continue; 1373 1374 if (forced_push(tp)) { 1375 tcp_mark_push(tp, skb); 1376 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 1377 } else if (skb == tcp_send_head(sk)) 1378 tcp_push_one(sk, mss_now); 1379 continue; 1380 1381 wait_for_sndbuf: 1382 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1383 wait_for_memory: 1384 if (copied) 1385 tcp_push(sk, flags & ~MSG_MORE, mss_now, 1386 TCP_NAGLE_PUSH, size_goal); 1387 1388 err = sk_stream_wait_memory(sk, &timeo); 1389 if (err != 0) 1390 goto do_error; 1391 1392 mss_now = tcp_send_mss(sk, &size_goal, flags); 1393 } 1394 1395 out: 1396 if (copied) { 1397 tcp_tx_timestamp(sk, sockc.tsflags); 1398 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal); 1399 } 1400 out_nopush: 1401 sock_zerocopy_put(uarg); 1402 return copied + copied_syn; 1403 1404 do_fault: 1405 if (!skb->len) { 1406 tcp_unlink_write_queue(skb, sk); 1407 /* It is the one place in all of TCP, except connection 1408 * reset, where we can be unlinking the send_head. 1409 */ 1410 tcp_check_send_head(sk, skb); 1411 sk_wmem_free_skb(sk, skb); 1412 } 1413 1414 do_error: 1415 if (copied + copied_syn) 1416 goto out; 1417 out_err: 1418 sock_zerocopy_put_abort(uarg); 1419 err = sk_stream_error(sk, flags, err); 1420 /* make sure we wake any epoll edge trigger waiter */ 1421 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && 1422 err == -EAGAIN)) { 1423 sk->sk_write_space(sk); 1424 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED); 1425 } 1426 return err; 1427 } 1428 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked); 1429 1430 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size) 1431 { 1432 int ret; 1433 1434 lock_sock(sk); 1435 ret = tcp_sendmsg_locked(sk, msg, size); 1436 release_sock(sk); 1437 1438 return ret; 1439 } 1440 EXPORT_SYMBOL(tcp_sendmsg); 1441 1442 /* 1443 * Handle reading urgent data. BSD has very simple semantics for 1444 * this, no blocking and very strange errors 8) 1445 */ 1446 1447 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags) 1448 { 1449 struct tcp_sock *tp = tcp_sk(sk); 1450 1451 /* No URG data to read. */ 1452 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data || 1453 tp->urg_data == TCP_URG_READ) 1454 return -EINVAL; /* Yes this is right ! */ 1455 1456 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE)) 1457 return -ENOTCONN; 1458 1459 if (tp->urg_data & TCP_URG_VALID) { 1460 int err = 0; 1461 char c = tp->urg_data; 1462 1463 if (!(flags & MSG_PEEK)) 1464 tp->urg_data = TCP_URG_READ; 1465 1466 /* Read urgent data. */ 1467 msg->msg_flags |= MSG_OOB; 1468 1469 if (len > 0) { 1470 if (!(flags & MSG_TRUNC)) 1471 err = memcpy_to_msg(msg, &c, 1); 1472 len = 1; 1473 } else 1474 msg->msg_flags |= MSG_TRUNC; 1475 1476 return err ? -EFAULT : len; 1477 } 1478 1479 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 1480 return 0; 1481 1482 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and 1483 * the available implementations agree in this case: 1484 * this call should never block, independent of the 1485 * blocking state of the socket. 1486 * Mike <pall@rz.uni-karlsruhe.de> 1487 */ 1488 return -EAGAIN; 1489 } 1490 1491 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len) 1492 { 1493 struct sk_buff *skb; 1494 int copied = 0, err = 0; 1495 1496 /* XXX -- need to support SO_PEEK_OFF */ 1497 1498 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 1499 err = skb_copy_datagram_msg(skb, 0, msg, skb->len); 1500 if (err) 1501 return err; 1502 copied += skb->len; 1503 } 1504 1505 skb_queue_walk(&sk->sk_write_queue, skb) { 1506 err = skb_copy_datagram_msg(skb, 0, msg, skb->len); 1507 if (err) 1508 break; 1509 1510 copied += skb->len; 1511 } 1512 1513 return err ?: copied; 1514 } 1515 1516 /* Clean up the receive buffer for full frames taken by the user, 1517 * then send an ACK if necessary. COPIED is the number of bytes 1518 * tcp_recvmsg has given to the user so far, it speeds up the 1519 * calculation of whether or not we must ACK for the sake of 1520 * a window update. 1521 */ 1522 static void tcp_cleanup_rbuf(struct sock *sk, int copied) 1523 { 1524 struct tcp_sock *tp = tcp_sk(sk); 1525 bool time_to_ack = false; 1526 1527 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 1528 1529 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq), 1530 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n", 1531 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt); 1532 1533 if (inet_csk_ack_scheduled(sk)) { 1534 const struct inet_connection_sock *icsk = inet_csk(sk); 1535 /* Delayed ACKs frequently hit locked sockets during bulk 1536 * receive. */ 1537 if (icsk->icsk_ack.blocked || 1538 /* Once-per-two-segments ACK was not sent by tcp_input.c */ 1539 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss || 1540 /* 1541 * If this read emptied read buffer, we send ACK, if 1542 * connection is not bidirectional, user drained 1543 * receive buffer and there was a small segment 1544 * in queue. 1545 */ 1546 (copied > 0 && 1547 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) || 1548 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) && 1549 !icsk->icsk_ack.pingpong)) && 1550 !atomic_read(&sk->sk_rmem_alloc))) 1551 time_to_ack = true; 1552 } 1553 1554 /* We send an ACK if we can now advertise a non-zero window 1555 * which has been raised "significantly". 1556 * 1557 * Even if window raised up to infinity, do not send window open ACK 1558 * in states, where we will not receive more. It is useless. 1559 */ 1560 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) { 1561 __u32 rcv_window_now = tcp_receive_window(tp); 1562 1563 /* Optimize, __tcp_select_window() is not cheap. */ 1564 if (2*rcv_window_now <= tp->window_clamp) { 1565 __u32 new_window = __tcp_select_window(sk); 1566 1567 /* Send ACK now, if this read freed lots of space 1568 * in our buffer. Certainly, new_window is new window. 1569 * We can advertise it now, if it is not less than current one. 1570 * "Lots" means "at least twice" here. 1571 */ 1572 if (new_window && new_window >= 2 * rcv_window_now) 1573 time_to_ack = true; 1574 } 1575 } 1576 if (time_to_ack) 1577 tcp_send_ack(sk); 1578 } 1579 1580 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off) 1581 { 1582 struct sk_buff *skb; 1583 u32 offset; 1584 1585 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) { 1586 offset = seq - TCP_SKB_CB(skb)->seq; 1587 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 1588 pr_err_once("%s: found a SYN, please report !\n", __func__); 1589 offset--; 1590 } 1591 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) { 1592 *off = offset; 1593 return skb; 1594 } 1595 /* This looks weird, but this can happen if TCP collapsing 1596 * splitted a fat GRO packet, while we released socket lock 1597 * in skb_splice_bits() 1598 */ 1599 sk_eat_skb(sk, skb); 1600 } 1601 return NULL; 1602 } 1603 1604 /* 1605 * This routine provides an alternative to tcp_recvmsg() for routines 1606 * that would like to handle copying from skbuffs directly in 'sendfile' 1607 * fashion. 1608 * Note: 1609 * - It is assumed that the socket was locked by the caller. 1610 * - The routine does not block. 1611 * - At present, there is no support for reading OOB data 1612 * or for 'peeking' the socket using this routine 1613 * (although both would be easy to implement). 1614 */ 1615 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, 1616 sk_read_actor_t recv_actor) 1617 { 1618 struct sk_buff *skb; 1619 struct tcp_sock *tp = tcp_sk(sk); 1620 u32 seq = tp->copied_seq; 1621 u32 offset; 1622 int copied = 0; 1623 1624 if (sk->sk_state == TCP_LISTEN) 1625 return -ENOTCONN; 1626 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) { 1627 if (offset < skb->len) { 1628 int used; 1629 size_t len; 1630 1631 len = skb->len - offset; 1632 /* Stop reading if we hit a patch of urgent data */ 1633 if (tp->urg_data) { 1634 u32 urg_offset = tp->urg_seq - seq; 1635 if (urg_offset < len) 1636 len = urg_offset; 1637 if (!len) 1638 break; 1639 } 1640 used = recv_actor(desc, skb, offset, len); 1641 if (used <= 0) { 1642 if (!copied) 1643 copied = used; 1644 break; 1645 } else if (used <= len) { 1646 seq += used; 1647 copied += used; 1648 offset += used; 1649 } 1650 /* If recv_actor drops the lock (e.g. TCP splice 1651 * receive) the skb pointer might be invalid when 1652 * getting here: tcp_collapse might have deleted it 1653 * while aggregating skbs from the socket queue. 1654 */ 1655 skb = tcp_recv_skb(sk, seq - 1, &offset); 1656 if (!skb) 1657 break; 1658 /* TCP coalescing might have appended data to the skb. 1659 * Try to splice more frags 1660 */ 1661 if (offset + 1 != skb->len) 1662 continue; 1663 } 1664 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) { 1665 sk_eat_skb(sk, skb); 1666 ++seq; 1667 break; 1668 } 1669 sk_eat_skb(sk, skb); 1670 if (!desc->count) 1671 break; 1672 tp->copied_seq = seq; 1673 } 1674 tp->copied_seq = seq; 1675 1676 tcp_rcv_space_adjust(sk); 1677 1678 /* Clean up data we have read: This will do ACK frames. */ 1679 if (copied > 0) { 1680 tcp_recv_skb(sk, seq, &offset); 1681 tcp_cleanup_rbuf(sk, copied); 1682 } 1683 return copied; 1684 } 1685 EXPORT_SYMBOL(tcp_read_sock); 1686 1687 int tcp_peek_len(struct socket *sock) 1688 { 1689 return tcp_inq(sock->sk); 1690 } 1691 EXPORT_SYMBOL(tcp_peek_len); 1692 1693 static void tcp_update_recv_tstamps(struct sk_buff *skb, 1694 struct scm_timestamping *tss) 1695 { 1696 if (skb->tstamp) 1697 tss->ts[0] = ktime_to_timespec(skb->tstamp); 1698 else 1699 tss->ts[0] = (struct timespec) {0}; 1700 1701 if (skb_hwtstamps(skb)->hwtstamp) 1702 tss->ts[2] = ktime_to_timespec(skb_hwtstamps(skb)->hwtstamp); 1703 else 1704 tss->ts[2] = (struct timespec) {0}; 1705 } 1706 1707 /* Similar to __sock_recv_timestamp, but does not require an skb */ 1708 static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk, 1709 struct scm_timestamping *tss) 1710 { 1711 struct timeval tv; 1712 bool has_timestamping = false; 1713 1714 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) { 1715 if (sock_flag(sk, SOCK_RCVTSTAMP)) { 1716 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) { 1717 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPNS, 1718 sizeof(tss->ts[0]), &tss->ts[0]); 1719 } else { 1720 tv.tv_sec = tss->ts[0].tv_sec; 1721 tv.tv_usec = tss->ts[0].tv_nsec / 1000; 1722 1723 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMP, 1724 sizeof(tv), &tv); 1725 } 1726 } 1727 1728 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE) 1729 has_timestamping = true; 1730 else 1731 tss->ts[0] = (struct timespec) {0}; 1732 } 1733 1734 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) { 1735 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE) 1736 has_timestamping = true; 1737 else 1738 tss->ts[2] = (struct timespec) {0}; 1739 } 1740 1741 if (has_timestamping) { 1742 tss->ts[1] = (struct timespec) {0}; 1743 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPING, 1744 sizeof(*tss), tss); 1745 } 1746 } 1747 1748 /* 1749 * This routine copies from a sock struct into the user buffer. 1750 * 1751 * Technical note: in 2.3 we work on _locked_ socket, so that 1752 * tricks with *seq access order and skb->users are not required. 1753 * Probably, code can be easily improved even more. 1754 */ 1755 1756 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock, 1757 int flags, int *addr_len) 1758 { 1759 struct tcp_sock *tp = tcp_sk(sk); 1760 int copied = 0; 1761 u32 peek_seq; 1762 u32 *seq; 1763 unsigned long used; 1764 int err; 1765 int target; /* Read at least this many bytes */ 1766 long timeo; 1767 struct sk_buff *skb, *last; 1768 u32 urg_hole = 0; 1769 struct scm_timestamping tss; 1770 bool has_tss = false; 1771 1772 if (unlikely(flags & MSG_ERRQUEUE)) 1773 return inet_recv_error(sk, msg, len, addr_len); 1774 1775 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) && 1776 (sk->sk_state == TCP_ESTABLISHED)) 1777 sk_busy_loop(sk, nonblock); 1778 1779 lock_sock(sk); 1780 1781 err = -ENOTCONN; 1782 if (sk->sk_state == TCP_LISTEN) 1783 goto out; 1784 1785 timeo = sock_rcvtimeo(sk, nonblock); 1786 1787 /* Urgent data needs to be handled specially. */ 1788 if (flags & MSG_OOB) 1789 goto recv_urg; 1790 1791 if (unlikely(tp->repair)) { 1792 err = -EPERM; 1793 if (!(flags & MSG_PEEK)) 1794 goto out; 1795 1796 if (tp->repair_queue == TCP_SEND_QUEUE) 1797 goto recv_sndq; 1798 1799 err = -EINVAL; 1800 if (tp->repair_queue == TCP_NO_QUEUE) 1801 goto out; 1802 1803 /* 'common' recv queue MSG_PEEK-ing */ 1804 } 1805 1806 seq = &tp->copied_seq; 1807 if (flags & MSG_PEEK) { 1808 peek_seq = tp->copied_seq; 1809 seq = &peek_seq; 1810 } 1811 1812 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 1813 1814 do { 1815 u32 offset; 1816 1817 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */ 1818 if (tp->urg_data && tp->urg_seq == *seq) { 1819 if (copied) 1820 break; 1821 if (signal_pending(current)) { 1822 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN; 1823 break; 1824 } 1825 } 1826 1827 /* Next get a buffer. */ 1828 1829 last = skb_peek_tail(&sk->sk_receive_queue); 1830 skb_queue_walk(&sk->sk_receive_queue, skb) { 1831 last = skb; 1832 /* Now that we have two receive queues this 1833 * shouldn't happen. 1834 */ 1835 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq), 1836 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n", 1837 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, 1838 flags)) 1839 break; 1840 1841 offset = *seq - TCP_SKB_CB(skb)->seq; 1842 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 1843 pr_err_once("%s: found a SYN, please report !\n", __func__); 1844 offset--; 1845 } 1846 if (offset < skb->len) 1847 goto found_ok_skb; 1848 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1849 goto found_fin_ok; 1850 WARN(!(flags & MSG_PEEK), 1851 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n", 1852 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags); 1853 } 1854 1855 /* Well, if we have backlog, try to process it now yet. */ 1856 1857 if (copied >= target && !sk->sk_backlog.tail) 1858 break; 1859 1860 if (copied) { 1861 if (sk->sk_err || 1862 sk->sk_state == TCP_CLOSE || 1863 (sk->sk_shutdown & RCV_SHUTDOWN) || 1864 !timeo || 1865 signal_pending(current)) 1866 break; 1867 } else { 1868 if (sock_flag(sk, SOCK_DONE)) 1869 break; 1870 1871 if (sk->sk_err) { 1872 copied = sock_error(sk); 1873 break; 1874 } 1875 1876 if (sk->sk_shutdown & RCV_SHUTDOWN) 1877 break; 1878 1879 if (sk->sk_state == TCP_CLOSE) { 1880 if (!sock_flag(sk, SOCK_DONE)) { 1881 /* This occurs when user tries to read 1882 * from never connected socket. 1883 */ 1884 copied = -ENOTCONN; 1885 break; 1886 } 1887 break; 1888 } 1889 1890 if (!timeo) { 1891 copied = -EAGAIN; 1892 break; 1893 } 1894 1895 if (signal_pending(current)) { 1896 copied = sock_intr_errno(timeo); 1897 break; 1898 } 1899 } 1900 1901 tcp_cleanup_rbuf(sk, copied); 1902 1903 if (copied >= target) { 1904 /* Do not sleep, just process backlog. */ 1905 release_sock(sk); 1906 lock_sock(sk); 1907 } else { 1908 sk_wait_data(sk, &timeo, last); 1909 } 1910 1911 if ((flags & MSG_PEEK) && 1912 (peek_seq - copied - urg_hole != tp->copied_seq)) { 1913 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n", 1914 current->comm, 1915 task_pid_nr(current)); 1916 peek_seq = tp->copied_seq; 1917 } 1918 continue; 1919 1920 found_ok_skb: 1921 /* Ok so how much can we use? */ 1922 used = skb->len - offset; 1923 if (len < used) 1924 used = len; 1925 1926 /* Do we have urgent data here? */ 1927 if (tp->urg_data) { 1928 u32 urg_offset = tp->urg_seq - *seq; 1929 if (urg_offset < used) { 1930 if (!urg_offset) { 1931 if (!sock_flag(sk, SOCK_URGINLINE)) { 1932 ++*seq; 1933 urg_hole++; 1934 offset++; 1935 used--; 1936 if (!used) 1937 goto skip_copy; 1938 } 1939 } else 1940 used = urg_offset; 1941 } 1942 } 1943 1944 if (!(flags & MSG_TRUNC)) { 1945 err = skb_copy_datagram_msg(skb, offset, msg, used); 1946 if (err) { 1947 /* Exception. Bailout! */ 1948 if (!copied) 1949 copied = -EFAULT; 1950 break; 1951 } 1952 } 1953 1954 *seq += used; 1955 copied += used; 1956 len -= used; 1957 1958 tcp_rcv_space_adjust(sk); 1959 1960 skip_copy: 1961 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) { 1962 tp->urg_data = 0; 1963 tcp_fast_path_check(sk); 1964 } 1965 if (used + offset < skb->len) 1966 continue; 1967 1968 if (TCP_SKB_CB(skb)->has_rxtstamp) { 1969 tcp_update_recv_tstamps(skb, &tss); 1970 has_tss = true; 1971 } 1972 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1973 goto found_fin_ok; 1974 if (!(flags & MSG_PEEK)) 1975 sk_eat_skb(sk, skb); 1976 continue; 1977 1978 found_fin_ok: 1979 /* Process the FIN. */ 1980 ++*seq; 1981 if (!(flags & MSG_PEEK)) 1982 sk_eat_skb(sk, skb); 1983 break; 1984 } while (len > 0); 1985 1986 /* According to UNIX98, msg_name/msg_namelen are ignored 1987 * on connected socket. I was just happy when found this 8) --ANK 1988 */ 1989 1990 if (has_tss) 1991 tcp_recv_timestamp(msg, sk, &tss); 1992 1993 /* Clean up data we have read: This will do ACK frames. */ 1994 tcp_cleanup_rbuf(sk, copied); 1995 1996 release_sock(sk); 1997 return copied; 1998 1999 out: 2000 release_sock(sk); 2001 return err; 2002 2003 recv_urg: 2004 err = tcp_recv_urg(sk, msg, len, flags); 2005 goto out; 2006 2007 recv_sndq: 2008 err = tcp_peek_sndq(sk, msg, len); 2009 goto out; 2010 } 2011 EXPORT_SYMBOL(tcp_recvmsg); 2012 2013 void tcp_set_state(struct sock *sk, int state) 2014 { 2015 int oldstate = sk->sk_state; 2016 2017 /* We defined a new enum for TCP states that are exported in BPF 2018 * so as not force the internal TCP states to be frozen. The 2019 * following checks will detect if an internal state value ever 2020 * differs from the BPF value. If this ever happens, then we will 2021 * need to remap the internal value to the BPF value before calling 2022 * tcp_call_bpf_2arg. 2023 */ 2024 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED); 2025 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT); 2026 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV); 2027 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1); 2028 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2); 2029 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT); 2030 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE); 2031 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT); 2032 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK); 2033 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN); 2034 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING); 2035 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV); 2036 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES); 2037 2038 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG)) 2039 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state); 2040 2041 switch (state) { 2042 case TCP_ESTABLISHED: 2043 if (oldstate != TCP_ESTABLISHED) 2044 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 2045 break; 2046 2047 case TCP_CLOSE: 2048 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED) 2049 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS); 2050 2051 sk->sk_prot->unhash(sk); 2052 if (inet_csk(sk)->icsk_bind_hash && 2053 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK)) 2054 inet_put_port(sk); 2055 /* fall through */ 2056 default: 2057 if (oldstate == TCP_ESTABLISHED) 2058 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 2059 } 2060 2061 /* Change state AFTER socket is unhashed to avoid closed 2062 * socket sitting in hash tables. 2063 */ 2064 inet_sk_state_store(sk, state); 2065 2066 #ifdef STATE_TRACE 2067 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]); 2068 #endif 2069 } 2070 EXPORT_SYMBOL_GPL(tcp_set_state); 2071 2072 /* 2073 * State processing on a close. This implements the state shift for 2074 * sending our FIN frame. Note that we only send a FIN for some 2075 * states. A shutdown() may have already sent the FIN, or we may be 2076 * closed. 2077 */ 2078 2079 static const unsigned char new_state[16] = { 2080 /* current state: new state: action: */ 2081 [0 /* (Invalid) */] = TCP_CLOSE, 2082 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 2083 [TCP_SYN_SENT] = TCP_CLOSE, 2084 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 2085 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1, 2086 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2, 2087 [TCP_TIME_WAIT] = TCP_CLOSE, 2088 [TCP_CLOSE] = TCP_CLOSE, 2089 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN, 2090 [TCP_LAST_ACK] = TCP_LAST_ACK, 2091 [TCP_LISTEN] = TCP_CLOSE, 2092 [TCP_CLOSING] = TCP_CLOSING, 2093 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */ 2094 }; 2095 2096 static int tcp_close_state(struct sock *sk) 2097 { 2098 int next = (int)new_state[sk->sk_state]; 2099 int ns = next & TCP_STATE_MASK; 2100 2101 tcp_set_state(sk, ns); 2102 2103 return next & TCP_ACTION_FIN; 2104 } 2105 2106 /* 2107 * Shutdown the sending side of a connection. Much like close except 2108 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD). 2109 */ 2110 2111 void tcp_shutdown(struct sock *sk, int how) 2112 { 2113 /* We need to grab some memory, and put together a FIN, 2114 * and then put it into the queue to be sent. 2115 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92. 2116 */ 2117 if (!(how & SEND_SHUTDOWN)) 2118 return; 2119 2120 /* If we've already sent a FIN, or it's a closed state, skip this. */ 2121 if ((1 << sk->sk_state) & 2122 (TCPF_ESTABLISHED | TCPF_SYN_SENT | 2123 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) { 2124 /* Clear out any half completed packets. FIN if needed. */ 2125 if (tcp_close_state(sk)) 2126 tcp_send_fin(sk); 2127 } 2128 } 2129 EXPORT_SYMBOL(tcp_shutdown); 2130 2131 bool tcp_check_oom(struct sock *sk, int shift) 2132 { 2133 bool too_many_orphans, out_of_socket_memory; 2134 2135 too_many_orphans = tcp_too_many_orphans(sk, shift); 2136 out_of_socket_memory = tcp_out_of_memory(sk); 2137 2138 if (too_many_orphans) 2139 net_info_ratelimited("too many orphaned sockets\n"); 2140 if (out_of_socket_memory) 2141 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n"); 2142 return too_many_orphans || out_of_socket_memory; 2143 } 2144 2145 void tcp_close(struct sock *sk, long timeout) 2146 { 2147 struct sk_buff *skb; 2148 int data_was_unread = 0; 2149 int state; 2150 2151 lock_sock(sk); 2152 sk->sk_shutdown = SHUTDOWN_MASK; 2153 2154 if (sk->sk_state == TCP_LISTEN) { 2155 tcp_set_state(sk, TCP_CLOSE); 2156 2157 /* Special case. */ 2158 inet_csk_listen_stop(sk); 2159 2160 goto adjudge_to_death; 2161 } 2162 2163 /* We need to flush the recv. buffs. We do this only on the 2164 * descriptor close, not protocol-sourced closes, because the 2165 * reader process may not have drained the data yet! 2166 */ 2167 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) { 2168 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq; 2169 2170 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 2171 len--; 2172 data_was_unread += len; 2173 __kfree_skb(skb); 2174 } 2175 2176 sk_mem_reclaim(sk); 2177 2178 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */ 2179 if (sk->sk_state == TCP_CLOSE) 2180 goto adjudge_to_death; 2181 2182 /* As outlined in RFC 2525, section 2.17, we send a RST here because 2183 * data was lost. To witness the awful effects of the old behavior of 2184 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk 2185 * GET in an FTP client, suspend the process, wait for the client to 2186 * advertise a zero window, then kill -9 the FTP client, wheee... 2187 * Note: timeout is always zero in such a case. 2188 */ 2189 if (unlikely(tcp_sk(sk)->repair)) { 2190 sk->sk_prot->disconnect(sk, 0); 2191 } else if (data_was_unread) { 2192 /* Unread data was tossed, zap the connection. */ 2193 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE); 2194 tcp_set_state(sk, TCP_CLOSE); 2195 tcp_send_active_reset(sk, sk->sk_allocation); 2196 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) { 2197 /* Check zero linger _after_ checking for unread data. */ 2198 sk->sk_prot->disconnect(sk, 0); 2199 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 2200 } else if (tcp_close_state(sk)) { 2201 /* We FIN if the application ate all the data before 2202 * zapping the connection. 2203 */ 2204 2205 /* RED-PEN. Formally speaking, we have broken TCP state 2206 * machine. State transitions: 2207 * 2208 * TCP_ESTABLISHED -> TCP_FIN_WAIT1 2209 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible) 2210 * TCP_CLOSE_WAIT -> TCP_LAST_ACK 2211 * 2212 * are legal only when FIN has been sent (i.e. in window), 2213 * rather than queued out of window. Purists blame. 2214 * 2215 * F.e. "RFC state" is ESTABLISHED, 2216 * if Linux state is FIN-WAIT-1, but FIN is still not sent. 2217 * 2218 * The visible declinations are that sometimes 2219 * we enter time-wait state, when it is not required really 2220 * (harmless), do not send active resets, when they are 2221 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when 2222 * they look as CLOSING or LAST_ACK for Linux) 2223 * Probably, I missed some more holelets. 2224 * --ANK 2225 * XXX (TFO) - To start off we don't support SYN+ACK+FIN 2226 * in a single packet! (May consider it later but will 2227 * probably need API support or TCP_CORK SYN-ACK until 2228 * data is written and socket is closed.) 2229 */ 2230 tcp_send_fin(sk); 2231 } 2232 2233 sk_stream_wait_close(sk, timeout); 2234 2235 adjudge_to_death: 2236 state = sk->sk_state; 2237 sock_hold(sk); 2238 sock_orphan(sk); 2239 2240 /* It is the last release_sock in its life. It will remove backlog. */ 2241 release_sock(sk); 2242 2243 2244 /* Now socket is owned by kernel and we acquire BH lock 2245 * to finish close. No need to check for user refs. 2246 */ 2247 local_bh_disable(); 2248 bh_lock_sock(sk); 2249 WARN_ON(sock_owned_by_user(sk)); 2250 2251 percpu_counter_inc(sk->sk_prot->orphan_count); 2252 2253 /* Have we already been destroyed by a softirq or backlog? */ 2254 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE) 2255 goto out; 2256 2257 /* This is a (useful) BSD violating of the RFC. There is a 2258 * problem with TCP as specified in that the other end could 2259 * keep a socket open forever with no application left this end. 2260 * We use a 1 minute timeout (about the same as BSD) then kill 2261 * our end. If they send after that then tough - BUT: long enough 2262 * that we won't make the old 4*rto = almost no time - whoops 2263 * reset mistake. 2264 * 2265 * Nope, it was not mistake. It is really desired behaviour 2266 * f.e. on http servers, when such sockets are useless, but 2267 * consume significant resources. Let's do it with special 2268 * linger2 option. --ANK 2269 */ 2270 2271 if (sk->sk_state == TCP_FIN_WAIT2) { 2272 struct tcp_sock *tp = tcp_sk(sk); 2273 if (tp->linger2 < 0) { 2274 tcp_set_state(sk, TCP_CLOSE); 2275 tcp_send_active_reset(sk, GFP_ATOMIC); 2276 __NET_INC_STATS(sock_net(sk), 2277 LINUX_MIB_TCPABORTONLINGER); 2278 } else { 2279 const int tmo = tcp_fin_time(sk); 2280 2281 if (tmo > TCP_TIMEWAIT_LEN) { 2282 inet_csk_reset_keepalive_timer(sk, 2283 tmo - TCP_TIMEWAIT_LEN); 2284 } else { 2285 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 2286 goto out; 2287 } 2288 } 2289 } 2290 if (sk->sk_state != TCP_CLOSE) { 2291 sk_mem_reclaim(sk); 2292 if (tcp_check_oom(sk, 0)) { 2293 tcp_set_state(sk, TCP_CLOSE); 2294 tcp_send_active_reset(sk, GFP_ATOMIC); 2295 __NET_INC_STATS(sock_net(sk), 2296 LINUX_MIB_TCPABORTONMEMORY); 2297 } else if (!check_net(sock_net(sk))) { 2298 /* Not possible to send reset; just close */ 2299 tcp_set_state(sk, TCP_CLOSE); 2300 } 2301 } 2302 2303 if (sk->sk_state == TCP_CLOSE) { 2304 struct request_sock *req = tcp_sk(sk)->fastopen_rsk; 2305 /* We could get here with a non-NULL req if the socket is 2306 * aborted (e.g., closed with unread data) before 3WHS 2307 * finishes. 2308 */ 2309 if (req) 2310 reqsk_fastopen_remove(sk, req, false); 2311 inet_csk_destroy_sock(sk); 2312 } 2313 /* Otherwise, socket is reprieved until protocol close. */ 2314 2315 out: 2316 bh_unlock_sock(sk); 2317 local_bh_enable(); 2318 sock_put(sk); 2319 } 2320 EXPORT_SYMBOL(tcp_close); 2321 2322 /* These states need RST on ABORT according to RFC793 */ 2323 2324 static inline bool tcp_need_reset(int state) 2325 { 2326 return (1 << state) & 2327 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 | 2328 TCPF_FIN_WAIT2 | TCPF_SYN_RECV); 2329 } 2330 2331 static void tcp_rtx_queue_purge(struct sock *sk) 2332 { 2333 struct rb_node *p = rb_first(&sk->tcp_rtx_queue); 2334 2335 while (p) { 2336 struct sk_buff *skb = rb_to_skb(p); 2337 2338 p = rb_next(p); 2339 /* Since we are deleting whole queue, no need to 2340 * list_del(&skb->tcp_tsorted_anchor) 2341 */ 2342 tcp_rtx_queue_unlink(skb, sk); 2343 sk_wmem_free_skb(sk, skb); 2344 } 2345 } 2346 2347 void tcp_write_queue_purge(struct sock *sk) 2348 { 2349 struct sk_buff *skb; 2350 2351 tcp_chrono_stop(sk, TCP_CHRONO_BUSY); 2352 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) { 2353 tcp_skb_tsorted_anchor_cleanup(skb); 2354 sk_wmem_free_skb(sk, skb); 2355 } 2356 tcp_rtx_queue_purge(sk); 2357 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue); 2358 sk_mem_reclaim(sk); 2359 tcp_clear_all_retrans_hints(tcp_sk(sk)); 2360 tcp_sk(sk)->packets_out = 0; 2361 } 2362 2363 int tcp_disconnect(struct sock *sk, int flags) 2364 { 2365 struct inet_sock *inet = inet_sk(sk); 2366 struct inet_connection_sock *icsk = inet_csk(sk); 2367 struct tcp_sock *tp = tcp_sk(sk); 2368 int err = 0; 2369 int old_state = sk->sk_state; 2370 2371 if (old_state != TCP_CLOSE) 2372 tcp_set_state(sk, TCP_CLOSE); 2373 2374 /* ABORT function of RFC793 */ 2375 if (old_state == TCP_LISTEN) { 2376 inet_csk_listen_stop(sk); 2377 } else if (unlikely(tp->repair)) { 2378 sk->sk_err = ECONNABORTED; 2379 } else if (tcp_need_reset(old_state) || 2380 (tp->snd_nxt != tp->write_seq && 2381 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) { 2382 /* The last check adjusts for discrepancy of Linux wrt. RFC 2383 * states 2384 */ 2385 tcp_send_active_reset(sk, gfp_any()); 2386 sk->sk_err = ECONNRESET; 2387 } else if (old_state == TCP_SYN_SENT) 2388 sk->sk_err = ECONNRESET; 2389 2390 tcp_clear_xmit_timers(sk); 2391 __skb_queue_purge(&sk->sk_receive_queue); 2392 tcp_write_queue_purge(sk); 2393 tcp_fastopen_active_disable_ofo_check(sk); 2394 skb_rbtree_purge(&tp->out_of_order_queue); 2395 2396 inet->inet_dport = 0; 2397 2398 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK)) 2399 inet_reset_saddr(sk); 2400 2401 sk->sk_shutdown = 0; 2402 sock_reset_flag(sk, SOCK_DONE); 2403 tp->srtt_us = 0; 2404 tp->write_seq += tp->max_window + 2; 2405 if (tp->write_seq == 0) 2406 tp->write_seq = 1; 2407 icsk->icsk_backoff = 0; 2408 tp->snd_cwnd = 2; 2409 icsk->icsk_probes_out = 0; 2410 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 2411 tp->snd_cwnd_cnt = 0; 2412 tp->window_clamp = 0; 2413 tcp_set_ca_state(sk, TCP_CA_Open); 2414 tp->is_sack_reneg = 0; 2415 tcp_clear_retrans(tp); 2416 inet_csk_delack_init(sk); 2417 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0 2418 * issue in __tcp_select_window() 2419 */ 2420 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS; 2421 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt)); 2422 __sk_dst_reset(sk); 2423 dst_release(sk->sk_rx_dst); 2424 sk->sk_rx_dst = NULL; 2425 tcp_saved_syn_free(tp); 2426 2427 /* Clean up fastopen related fields */ 2428 tcp_free_fastopen_req(tp); 2429 inet->defer_connect = 0; 2430 2431 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash); 2432 2433 if (sk->sk_frag.page) { 2434 put_page(sk->sk_frag.page); 2435 sk->sk_frag.page = NULL; 2436 sk->sk_frag.offset = 0; 2437 } 2438 2439 sk->sk_error_report(sk); 2440 return err; 2441 } 2442 EXPORT_SYMBOL(tcp_disconnect); 2443 2444 static inline bool tcp_can_repair_sock(const struct sock *sk) 2445 { 2446 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) && 2447 (sk->sk_state != TCP_LISTEN); 2448 } 2449 2450 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len) 2451 { 2452 struct tcp_repair_window opt; 2453 2454 if (!tp->repair) 2455 return -EPERM; 2456 2457 if (len != sizeof(opt)) 2458 return -EINVAL; 2459 2460 if (copy_from_user(&opt, optbuf, sizeof(opt))) 2461 return -EFAULT; 2462 2463 if (opt.max_window < opt.snd_wnd) 2464 return -EINVAL; 2465 2466 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd)) 2467 return -EINVAL; 2468 2469 if (after(opt.rcv_wup, tp->rcv_nxt)) 2470 return -EINVAL; 2471 2472 tp->snd_wl1 = opt.snd_wl1; 2473 tp->snd_wnd = opt.snd_wnd; 2474 tp->max_window = opt.max_window; 2475 2476 tp->rcv_wnd = opt.rcv_wnd; 2477 tp->rcv_wup = opt.rcv_wup; 2478 2479 return 0; 2480 } 2481 2482 static int tcp_repair_options_est(struct sock *sk, 2483 struct tcp_repair_opt __user *optbuf, unsigned int len) 2484 { 2485 struct tcp_sock *tp = tcp_sk(sk); 2486 struct tcp_repair_opt opt; 2487 2488 while (len >= sizeof(opt)) { 2489 if (copy_from_user(&opt, optbuf, sizeof(opt))) 2490 return -EFAULT; 2491 2492 optbuf++; 2493 len -= sizeof(opt); 2494 2495 switch (opt.opt_code) { 2496 case TCPOPT_MSS: 2497 tp->rx_opt.mss_clamp = opt.opt_val; 2498 tcp_mtup_init(sk); 2499 break; 2500 case TCPOPT_WINDOW: 2501 { 2502 u16 snd_wscale = opt.opt_val & 0xFFFF; 2503 u16 rcv_wscale = opt.opt_val >> 16; 2504 2505 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE) 2506 return -EFBIG; 2507 2508 tp->rx_opt.snd_wscale = snd_wscale; 2509 tp->rx_opt.rcv_wscale = rcv_wscale; 2510 tp->rx_opt.wscale_ok = 1; 2511 } 2512 break; 2513 case TCPOPT_SACK_PERM: 2514 if (opt.opt_val != 0) 2515 return -EINVAL; 2516 2517 tp->rx_opt.sack_ok |= TCP_SACK_SEEN; 2518 break; 2519 case TCPOPT_TIMESTAMP: 2520 if (opt.opt_val != 0) 2521 return -EINVAL; 2522 2523 tp->rx_opt.tstamp_ok = 1; 2524 break; 2525 } 2526 } 2527 2528 return 0; 2529 } 2530 2531 /* 2532 * Socket option code for TCP. 2533 */ 2534 static int do_tcp_setsockopt(struct sock *sk, int level, 2535 int optname, char __user *optval, unsigned int optlen) 2536 { 2537 struct tcp_sock *tp = tcp_sk(sk); 2538 struct inet_connection_sock *icsk = inet_csk(sk); 2539 struct net *net = sock_net(sk); 2540 int val; 2541 int err = 0; 2542 2543 /* These are data/string values, all the others are ints */ 2544 switch (optname) { 2545 case TCP_CONGESTION: { 2546 char name[TCP_CA_NAME_MAX]; 2547 2548 if (optlen < 1) 2549 return -EINVAL; 2550 2551 val = strncpy_from_user(name, optval, 2552 min_t(long, TCP_CA_NAME_MAX-1, optlen)); 2553 if (val < 0) 2554 return -EFAULT; 2555 name[val] = 0; 2556 2557 lock_sock(sk); 2558 err = tcp_set_congestion_control(sk, name, true, true); 2559 release_sock(sk); 2560 return err; 2561 } 2562 case TCP_ULP: { 2563 char name[TCP_ULP_NAME_MAX]; 2564 2565 if (optlen < 1) 2566 return -EINVAL; 2567 2568 val = strncpy_from_user(name, optval, 2569 min_t(long, TCP_ULP_NAME_MAX - 1, 2570 optlen)); 2571 if (val < 0) 2572 return -EFAULT; 2573 name[val] = 0; 2574 2575 lock_sock(sk); 2576 err = tcp_set_ulp(sk, name); 2577 release_sock(sk); 2578 return err; 2579 } 2580 case TCP_FASTOPEN_KEY: { 2581 __u8 key[TCP_FASTOPEN_KEY_LENGTH]; 2582 2583 if (optlen != sizeof(key)) 2584 return -EINVAL; 2585 2586 if (copy_from_user(key, optval, optlen)) 2587 return -EFAULT; 2588 2589 return tcp_fastopen_reset_cipher(net, sk, key, sizeof(key)); 2590 } 2591 default: 2592 /* fallthru */ 2593 break; 2594 } 2595 2596 if (optlen < sizeof(int)) 2597 return -EINVAL; 2598 2599 if (get_user(val, (int __user *)optval)) 2600 return -EFAULT; 2601 2602 lock_sock(sk); 2603 2604 switch (optname) { 2605 case TCP_MAXSEG: 2606 /* Values greater than interface MTU won't take effect. However 2607 * at the point when this call is done we typically don't yet 2608 * know which interface is going to be used 2609 */ 2610 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) { 2611 err = -EINVAL; 2612 break; 2613 } 2614 tp->rx_opt.user_mss = val; 2615 break; 2616 2617 case TCP_NODELAY: 2618 if (val) { 2619 /* TCP_NODELAY is weaker than TCP_CORK, so that 2620 * this option on corked socket is remembered, but 2621 * it is not activated until cork is cleared. 2622 * 2623 * However, when TCP_NODELAY is set we make 2624 * an explicit push, which overrides even TCP_CORK 2625 * for currently queued segments. 2626 */ 2627 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH; 2628 tcp_push_pending_frames(sk); 2629 } else { 2630 tp->nonagle &= ~TCP_NAGLE_OFF; 2631 } 2632 break; 2633 2634 case TCP_THIN_LINEAR_TIMEOUTS: 2635 if (val < 0 || val > 1) 2636 err = -EINVAL; 2637 else 2638 tp->thin_lto = val; 2639 break; 2640 2641 case TCP_THIN_DUPACK: 2642 if (val < 0 || val > 1) 2643 err = -EINVAL; 2644 break; 2645 2646 case TCP_REPAIR: 2647 if (!tcp_can_repair_sock(sk)) 2648 err = -EPERM; 2649 else if (val == 1) { 2650 tp->repair = 1; 2651 sk->sk_reuse = SK_FORCE_REUSE; 2652 tp->repair_queue = TCP_NO_QUEUE; 2653 } else if (val == 0) { 2654 tp->repair = 0; 2655 sk->sk_reuse = SK_NO_REUSE; 2656 tcp_send_window_probe(sk); 2657 } else 2658 err = -EINVAL; 2659 2660 break; 2661 2662 case TCP_REPAIR_QUEUE: 2663 if (!tp->repair) 2664 err = -EPERM; 2665 else if (val < TCP_QUEUES_NR) 2666 tp->repair_queue = val; 2667 else 2668 err = -EINVAL; 2669 break; 2670 2671 case TCP_QUEUE_SEQ: 2672 if (sk->sk_state != TCP_CLOSE) 2673 err = -EPERM; 2674 else if (tp->repair_queue == TCP_SEND_QUEUE) 2675 tp->write_seq = val; 2676 else if (tp->repair_queue == TCP_RECV_QUEUE) 2677 tp->rcv_nxt = val; 2678 else 2679 err = -EINVAL; 2680 break; 2681 2682 case TCP_REPAIR_OPTIONS: 2683 if (!tp->repair) 2684 err = -EINVAL; 2685 else if (sk->sk_state == TCP_ESTABLISHED) 2686 err = tcp_repair_options_est(sk, 2687 (struct tcp_repair_opt __user *)optval, 2688 optlen); 2689 else 2690 err = -EPERM; 2691 break; 2692 2693 case TCP_CORK: 2694 /* When set indicates to always queue non-full frames. 2695 * Later the user clears this option and we transmit 2696 * any pending partial frames in the queue. This is 2697 * meant to be used alongside sendfile() to get properly 2698 * filled frames when the user (for example) must write 2699 * out headers with a write() call first and then use 2700 * sendfile to send out the data parts. 2701 * 2702 * TCP_CORK can be set together with TCP_NODELAY and it is 2703 * stronger than TCP_NODELAY. 2704 */ 2705 if (val) { 2706 tp->nonagle |= TCP_NAGLE_CORK; 2707 } else { 2708 tp->nonagle &= ~TCP_NAGLE_CORK; 2709 if (tp->nonagle&TCP_NAGLE_OFF) 2710 tp->nonagle |= TCP_NAGLE_PUSH; 2711 tcp_push_pending_frames(sk); 2712 } 2713 break; 2714 2715 case TCP_KEEPIDLE: 2716 if (val < 1 || val > MAX_TCP_KEEPIDLE) 2717 err = -EINVAL; 2718 else { 2719 tp->keepalive_time = val * HZ; 2720 if (sock_flag(sk, SOCK_KEEPOPEN) && 2721 !((1 << sk->sk_state) & 2722 (TCPF_CLOSE | TCPF_LISTEN))) { 2723 u32 elapsed = keepalive_time_elapsed(tp); 2724 if (tp->keepalive_time > elapsed) 2725 elapsed = tp->keepalive_time - elapsed; 2726 else 2727 elapsed = 0; 2728 inet_csk_reset_keepalive_timer(sk, elapsed); 2729 } 2730 } 2731 break; 2732 case TCP_KEEPINTVL: 2733 if (val < 1 || val > MAX_TCP_KEEPINTVL) 2734 err = -EINVAL; 2735 else 2736 tp->keepalive_intvl = val * HZ; 2737 break; 2738 case TCP_KEEPCNT: 2739 if (val < 1 || val > MAX_TCP_KEEPCNT) 2740 err = -EINVAL; 2741 else 2742 tp->keepalive_probes = val; 2743 break; 2744 case TCP_SYNCNT: 2745 if (val < 1 || val > MAX_TCP_SYNCNT) 2746 err = -EINVAL; 2747 else 2748 icsk->icsk_syn_retries = val; 2749 break; 2750 2751 case TCP_SAVE_SYN: 2752 if (val < 0 || val > 1) 2753 err = -EINVAL; 2754 else 2755 tp->save_syn = val; 2756 break; 2757 2758 case TCP_LINGER2: 2759 if (val < 0) 2760 tp->linger2 = -1; 2761 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ) 2762 tp->linger2 = 0; 2763 else 2764 tp->linger2 = val * HZ; 2765 break; 2766 2767 case TCP_DEFER_ACCEPT: 2768 /* Translate value in seconds to number of retransmits */ 2769 icsk->icsk_accept_queue.rskq_defer_accept = 2770 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ, 2771 TCP_RTO_MAX / HZ); 2772 break; 2773 2774 case TCP_WINDOW_CLAMP: 2775 if (!val) { 2776 if (sk->sk_state != TCP_CLOSE) { 2777 err = -EINVAL; 2778 break; 2779 } 2780 tp->window_clamp = 0; 2781 } else 2782 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ? 2783 SOCK_MIN_RCVBUF / 2 : val; 2784 break; 2785 2786 case TCP_QUICKACK: 2787 if (!val) { 2788 icsk->icsk_ack.pingpong = 1; 2789 } else { 2790 icsk->icsk_ack.pingpong = 0; 2791 if ((1 << sk->sk_state) & 2792 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) && 2793 inet_csk_ack_scheduled(sk)) { 2794 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 2795 tcp_cleanup_rbuf(sk, 1); 2796 if (!(val & 1)) 2797 icsk->icsk_ack.pingpong = 1; 2798 } 2799 } 2800 break; 2801 2802 #ifdef CONFIG_TCP_MD5SIG 2803 case TCP_MD5SIG: 2804 case TCP_MD5SIG_EXT: 2805 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) 2806 err = tp->af_specific->md5_parse(sk, optname, optval, optlen); 2807 else 2808 err = -EINVAL; 2809 break; 2810 #endif 2811 case TCP_USER_TIMEOUT: 2812 /* Cap the max time in ms TCP will retry or probe the window 2813 * before giving up and aborting (ETIMEDOUT) a connection. 2814 */ 2815 if (val < 0) 2816 err = -EINVAL; 2817 else 2818 icsk->icsk_user_timeout = msecs_to_jiffies(val); 2819 break; 2820 2821 case TCP_FASTOPEN: 2822 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE | 2823 TCPF_LISTEN))) { 2824 tcp_fastopen_init_key_once(net); 2825 2826 fastopen_queue_tune(sk, val); 2827 } else { 2828 err = -EINVAL; 2829 } 2830 break; 2831 case TCP_FASTOPEN_CONNECT: 2832 if (val > 1 || val < 0) { 2833 err = -EINVAL; 2834 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) { 2835 if (sk->sk_state == TCP_CLOSE) 2836 tp->fastopen_connect = val; 2837 else 2838 err = -EINVAL; 2839 } else { 2840 err = -EOPNOTSUPP; 2841 } 2842 break; 2843 case TCP_FASTOPEN_NO_COOKIE: 2844 if (val > 1 || val < 0) 2845 err = -EINVAL; 2846 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 2847 err = -EINVAL; 2848 else 2849 tp->fastopen_no_cookie = val; 2850 break; 2851 case TCP_TIMESTAMP: 2852 if (!tp->repair) 2853 err = -EPERM; 2854 else 2855 tp->tsoffset = val - tcp_time_stamp_raw(); 2856 break; 2857 case TCP_REPAIR_WINDOW: 2858 err = tcp_repair_set_window(tp, optval, optlen); 2859 break; 2860 case TCP_NOTSENT_LOWAT: 2861 tp->notsent_lowat = val; 2862 sk->sk_write_space(sk); 2863 break; 2864 default: 2865 err = -ENOPROTOOPT; 2866 break; 2867 } 2868 2869 release_sock(sk); 2870 return err; 2871 } 2872 2873 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval, 2874 unsigned int optlen) 2875 { 2876 const struct inet_connection_sock *icsk = inet_csk(sk); 2877 2878 if (level != SOL_TCP) 2879 return icsk->icsk_af_ops->setsockopt(sk, level, optname, 2880 optval, optlen); 2881 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2882 } 2883 EXPORT_SYMBOL(tcp_setsockopt); 2884 2885 #ifdef CONFIG_COMPAT 2886 int compat_tcp_setsockopt(struct sock *sk, int level, int optname, 2887 char __user *optval, unsigned int optlen) 2888 { 2889 if (level != SOL_TCP) 2890 return inet_csk_compat_setsockopt(sk, level, optname, 2891 optval, optlen); 2892 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 2893 } 2894 EXPORT_SYMBOL(compat_tcp_setsockopt); 2895 #endif 2896 2897 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp, 2898 struct tcp_info *info) 2899 { 2900 u64 stats[__TCP_CHRONO_MAX], total = 0; 2901 enum tcp_chrono i; 2902 2903 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) { 2904 stats[i] = tp->chrono_stat[i - 1]; 2905 if (i == tp->chrono_type) 2906 stats[i] += tcp_jiffies32 - tp->chrono_start; 2907 stats[i] *= USEC_PER_SEC / HZ; 2908 total += stats[i]; 2909 } 2910 2911 info->tcpi_busy_time = total; 2912 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED]; 2913 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED]; 2914 } 2915 2916 /* Return information about state of tcp endpoint in API format. */ 2917 void tcp_get_info(struct sock *sk, struct tcp_info *info) 2918 { 2919 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */ 2920 const struct inet_connection_sock *icsk = inet_csk(sk); 2921 u32 now; 2922 u64 rate64; 2923 bool slow; 2924 u32 rate; 2925 2926 memset(info, 0, sizeof(*info)); 2927 if (sk->sk_type != SOCK_STREAM) 2928 return; 2929 2930 info->tcpi_state = inet_sk_state_load(sk); 2931 2932 /* Report meaningful fields for all TCP states, including listeners */ 2933 rate = READ_ONCE(sk->sk_pacing_rate); 2934 rate64 = rate != ~0U ? rate : ~0ULL; 2935 info->tcpi_pacing_rate = rate64; 2936 2937 rate = READ_ONCE(sk->sk_max_pacing_rate); 2938 rate64 = rate != ~0U ? rate : ~0ULL; 2939 info->tcpi_max_pacing_rate = rate64; 2940 2941 info->tcpi_reordering = tp->reordering; 2942 info->tcpi_snd_cwnd = tp->snd_cwnd; 2943 2944 if (info->tcpi_state == TCP_LISTEN) { 2945 /* listeners aliased fields : 2946 * tcpi_unacked -> Number of children ready for accept() 2947 * tcpi_sacked -> max backlog 2948 */ 2949 info->tcpi_unacked = sk->sk_ack_backlog; 2950 info->tcpi_sacked = sk->sk_max_ack_backlog; 2951 return; 2952 } 2953 2954 slow = lock_sock_fast(sk); 2955 2956 info->tcpi_ca_state = icsk->icsk_ca_state; 2957 info->tcpi_retransmits = icsk->icsk_retransmits; 2958 info->tcpi_probes = icsk->icsk_probes_out; 2959 info->tcpi_backoff = icsk->icsk_backoff; 2960 2961 if (tp->rx_opt.tstamp_ok) 2962 info->tcpi_options |= TCPI_OPT_TIMESTAMPS; 2963 if (tcp_is_sack(tp)) 2964 info->tcpi_options |= TCPI_OPT_SACK; 2965 if (tp->rx_opt.wscale_ok) { 2966 info->tcpi_options |= TCPI_OPT_WSCALE; 2967 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale; 2968 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale; 2969 } 2970 2971 if (tp->ecn_flags & TCP_ECN_OK) 2972 info->tcpi_options |= TCPI_OPT_ECN; 2973 if (tp->ecn_flags & TCP_ECN_SEEN) 2974 info->tcpi_options |= TCPI_OPT_ECN_SEEN; 2975 if (tp->syn_data_acked) 2976 info->tcpi_options |= TCPI_OPT_SYN_DATA; 2977 2978 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto); 2979 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato); 2980 info->tcpi_snd_mss = tp->mss_cache; 2981 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss; 2982 2983 info->tcpi_unacked = tp->packets_out; 2984 info->tcpi_sacked = tp->sacked_out; 2985 2986 info->tcpi_lost = tp->lost_out; 2987 info->tcpi_retrans = tp->retrans_out; 2988 2989 now = tcp_jiffies32; 2990 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime); 2991 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime); 2992 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp); 2993 2994 info->tcpi_pmtu = icsk->icsk_pmtu_cookie; 2995 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh; 2996 info->tcpi_rtt = tp->srtt_us >> 3; 2997 info->tcpi_rttvar = tp->mdev_us >> 2; 2998 info->tcpi_snd_ssthresh = tp->snd_ssthresh; 2999 info->tcpi_advmss = tp->advmss; 3000 3001 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3; 3002 info->tcpi_rcv_space = tp->rcvq_space.space; 3003 3004 info->tcpi_total_retrans = tp->total_retrans; 3005 3006 info->tcpi_bytes_acked = tp->bytes_acked; 3007 info->tcpi_bytes_received = tp->bytes_received; 3008 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt); 3009 tcp_get_info_chrono_stats(tp, info); 3010 3011 info->tcpi_segs_out = tp->segs_out; 3012 info->tcpi_segs_in = tp->segs_in; 3013 3014 info->tcpi_min_rtt = tcp_min_rtt(tp); 3015 info->tcpi_data_segs_in = tp->data_segs_in; 3016 info->tcpi_data_segs_out = tp->data_segs_out; 3017 3018 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0; 3019 rate64 = tcp_compute_delivery_rate(tp); 3020 if (rate64) 3021 info->tcpi_delivery_rate = rate64; 3022 unlock_sock_fast(sk, slow); 3023 } 3024 EXPORT_SYMBOL_GPL(tcp_get_info); 3025 3026 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk) 3027 { 3028 const struct tcp_sock *tp = tcp_sk(sk); 3029 struct sk_buff *stats; 3030 struct tcp_info info; 3031 u64 rate64; 3032 u32 rate; 3033 3034 stats = alloc_skb(7 * nla_total_size_64bit(sizeof(u64)) + 3035 5 * nla_total_size(sizeof(u32)) + 3036 3 * nla_total_size(sizeof(u8)), GFP_ATOMIC); 3037 if (!stats) 3038 return NULL; 3039 3040 tcp_get_info_chrono_stats(tp, &info); 3041 nla_put_u64_64bit(stats, TCP_NLA_BUSY, 3042 info.tcpi_busy_time, TCP_NLA_PAD); 3043 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED, 3044 info.tcpi_rwnd_limited, TCP_NLA_PAD); 3045 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED, 3046 info.tcpi_sndbuf_limited, TCP_NLA_PAD); 3047 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT, 3048 tp->data_segs_out, TCP_NLA_PAD); 3049 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS, 3050 tp->total_retrans, TCP_NLA_PAD); 3051 3052 rate = READ_ONCE(sk->sk_pacing_rate); 3053 rate64 = rate != ~0U ? rate : ~0ULL; 3054 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD); 3055 3056 rate64 = tcp_compute_delivery_rate(tp); 3057 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD); 3058 3059 nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd); 3060 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering); 3061 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp)); 3062 3063 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits); 3064 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited); 3065 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh); 3066 3067 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una); 3068 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state); 3069 return stats; 3070 } 3071 3072 static int do_tcp_getsockopt(struct sock *sk, int level, 3073 int optname, char __user *optval, int __user *optlen) 3074 { 3075 struct inet_connection_sock *icsk = inet_csk(sk); 3076 struct tcp_sock *tp = tcp_sk(sk); 3077 struct net *net = sock_net(sk); 3078 int val, len; 3079 3080 if (get_user(len, optlen)) 3081 return -EFAULT; 3082 3083 len = min_t(unsigned int, len, sizeof(int)); 3084 3085 if (len < 0) 3086 return -EINVAL; 3087 3088 switch (optname) { 3089 case TCP_MAXSEG: 3090 val = tp->mss_cache; 3091 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 3092 val = tp->rx_opt.user_mss; 3093 if (tp->repair) 3094 val = tp->rx_opt.mss_clamp; 3095 break; 3096 case TCP_NODELAY: 3097 val = !!(tp->nonagle&TCP_NAGLE_OFF); 3098 break; 3099 case TCP_CORK: 3100 val = !!(tp->nonagle&TCP_NAGLE_CORK); 3101 break; 3102 case TCP_KEEPIDLE: 3103 val = keepalive_time_when(tp) / HZ; 3104 break; 3105 case TCP_KEEPINTVL: 3106 val = keepalive_intvl_when(tp) / HZ; 3107 break; 3108 case TCP_KEEPCNT: 3109 val = keepalive_probes(tp); 3110 break; 3111 case TCP_SYNCNT: 3112 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries; 3113 break; 3114 case TCP_LINGER2: 3115 val = tp->linger2; 3116 if (val >= 0) 3117 val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ; 3118 break; 3119 case TCP_DEFER_ACCEPT: 3120 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept, 3121 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ); 3122 break; 3123 case TCP_WINDOW_CLAMP: 3124 val = tp->window_clamp; 3125 break; 3126 case TCP_INFO: { 3127 struct tcp_info info; 3128 3129 if (get_user(len, optlen)) 3130 return -EFAULT; 3131 3132 tcp_get_info(sk, &info); 3133 3134 len = min_t(unsigned int, len, sizeof(info)); 3135 if (put_user(len, optlen)) 3136 return -EFAULT; 3137 if (copy_to_user(optval, &info, len)) 3138 return -EFAULT; 3139 return 0; 3140 } 3141 case TCP_CC_INFO: { 3142 const struct tcp_congestion_ops *ca_ops; 3143 union tcp_cc_info info; 3144 size_t sz = 0; 3145 int attr; 3146 3147 if (get_user(len, optlen)) 3148 return -EFAULT; 3149 3150 ca_ops = icsk->icsk_ca_ops; 3151 if (ca_ops && ca_ops->get_info) 3152 sz = ca_ops->get_info(sk, ~0U, &attr, &info); 3153 3154 len = min_t(unsigned int, len, sz); 3155 if (put_user(len, optlen)) 3156 return -EFAULT; 3157 if (copy_to_user(optval, &info, len)) 3158 return -EFAULT; 3159 return 0; 3160 } 3161 case TCP_QUICKACK: 3162 val = !icsk->icsk_ack.pingpong; 3163 break; 3164 3165 case TCP_CONGESTION: 3166 if (get_user(len, optlen)) 3167 return -EFAULT; 3168 len = min_t(unsigned int, len, TCP_CA_NAME_MAX); 3169 if (put_user(len, optlen)) 3170 return -EFAULT; 3171 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len)) 3172 return -EFAULT; 3173 return 0; 3174 3175 case TCP_ULP: 3176 if (get_user(len, optlen)) 3177 return -EFAULT; 3178 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX); 3179 if (!icsk->icsk_ulp_ops) { 3180 if (put_user(0, optlen)) 3181 return -EFAULT; 3182 return 0; 3183 } 3184 if (put_user(len, optlen)) 3185 return -EFAULT; 3186 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len)) 3187 return -EFAULT; 3188 return 0; 3189 3190 case TCP_FASTOPEN_KEY: { 3191 __u8 key[TCP_FASTOPEN_KEY_LENGTH]; 3192 struct tcp_fastopen_context *ctx; 3193 3194 if (get_user(len, optlen)) 3195 return -EFAULT; 3196 3197 rcu_read_lock(); 3198 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx); 3199 if (ctx) 3200 memcpy(key, ctx->key, sizeof(key)); 3201 else 3202 len = 0; 3203 rcu_read_unlock(); 3204 3205 len = min_t(unsigned int, len, sizeof(key)); 3206 if (put_user(len, optlen)) 3207 return -EFAULT; 3208 if (copy_to_user(optval, key, len)) 3209 return -EFAULT; 3210 return 0; 3211 } 3212 case TCP_THIN_LINEAR_TIMEOUTS: 3213 val = tp->thin_lto; 3214 break; 3215 3216 case TCP_THIN_DUPACK: 3217 val = 0; 3218 break; 3219 3220 case TCP_REPAIR: 3221 val = tp->repair; 3222 break; 3223 3224 case TCP_REPAIR_QUEUE: 3225 if (tp->repair) 3226 val = tp->repair_queue; 3227 else 3228 return -EINVAL; 3229 break; 3230 3231 case TCP_REPAIR_WINDOW: { 3232 struct tcp_repair_window opt; 3233 3234 if (get_user(len, optlen)) 3235 return -EFAULT; 3236 3237 if (len != sizeof(opt)) 3238 return -EINVAL; 3239 3240 if (!tp->repair) 3241 return -EPERM; 3242 3243 opt.snd_wl1 = tp->snd_wl1; 3244 opt.snd_wnd = tp->snd_wnd; 3245 opt.max_window = tp->max_window; 3246 opt.rcv_wnd = tp->rcv_wnd; 3247 opt.rcv_wup = tp->rcv_wup; 3248 3249 if (copy_to_user(optval, &opt, len)) 3250 return -EFAULT; 3251 return 0; 3252 } 3253 case TCP_QUEUE_SEQ: 3254 if (tp->repair_queue == TCP_SEND_QUEUE) 3255 val = tp->write_seq; 3256 else if (tp->repair_queue == TCP_RECV_QUEUE) 3257 val = tp->rcv_nxt; 3258 else 3259 return -EINVAL; 3260 break; 3261 3262 case TCP_USER_TIMEOUT: 3263 val = jiffies_to_msecs(icsk->icsk_user_timeout); 3264 break; 3265 3266 case TCP_FASTOPEN: 3267 val = icsk->icsk_accept_queue.fastopenq.max_qlen; 3268 break; 3269 3270 case TCP_FASTOPEN_CONNECT: 3271 val = tp->fastopen_connect; 3272 break; 3273 3274 case TCP_FASTOPEN_NO_COOKIE: 3275 val = tp->fastopen_no_cookie; 3276 break; 3277 3278 case TCP_TIMESTAMP: 3279 val = tcp_time_stamp_raw() + tp->tsoffset; 3280 break; 3281 case TCP_NOTSENT_LOWAT: 3282 val = tp->notsent_lowat; 3283 break; 3284 case TCP_SAVE_SYN: 3285 val = tp->save_syn; 3286 break; 3287 case TCP_SAVED_SYN: { 3288 if (get_user(len, optlen)) 3289 return -EFAULT; 3290 3291 lock_sock(sk); 3292 if (tp->saved_syn) { 3293 if (len < tp->saved_syn[0]) { 3294 if (put_user(tp->saved_syn[0], optlen)) { 3295 release_sock(sk); 3296 return -EFAULT; 3297 } 3298 release_sock(sk); 3299 return -EINVAL; 3300 } 3301 len = tp->saved_syn[0]; 3302 if (put_user(len, optlen)) { 3303 release_sock(sk); 3304 return -EFAULT; 3305 } 3306 if (copy_to_user(optval, tp->saved_syn + 1, len)) { 3307 release_sock(sk); 3308 return -EFAULT; 3309 } 3310 tcp_saved_syn_free(tp); 3311 release_sock(sk); 3312 } else { 3313 release_sock(sk); 3314 len = 0; 3315 if (put_user(len, optlen)) 3316 return -EFAULT; 3317 } 3318 return 0; 3319 } 3320 default: 3321 return -ENOPROTOOPT; 3322 } 3323 3324 if (put_user(len, optlen)) 3325 return -EFAULT; 3326 if (copy_to_user(optval, &val, len)) 3327 return -EFAULT; 3328 return 0; 3329 } 3330 3331 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval, 3332 int __user *optlen) 3333 { 3334 struct inet_connection_sock *icsk = inet_csk(sk); 3335 3336 if (level != SOL_TCP) 3337 return icsk->icsk_af_ops->getsockopt(sk, level, optname, 3338 optval, optlen); 3339 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 3340 } 3341 EXPORT_SYMBOL(tcp_getsockopt); 3342 3343 #ifdef CONFIG_COMPAT 3344 int compat_tcp_getsockopt(struct sock *sk, int level, int optname, 3345 char __user *optval, int __user *optlen) 3346 { 3347 if (level != SOL_TCP) 3348 return inet_csk_compat_getsockopt(sk, level, optname, 3349 optval, optlen); 3350 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 3351 } 3352 EXPORT_SYMBOL(compat_tcp_getsockopt); 3353 #endif 3354 3355 #ifdef CONFIG_TCP_MD5SIG 3356 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool); 3357 static DEFINE_MUTEX(tcp_md5sig_mutex); 3358 static bool tcp_md5sig_pool_populated = false; 3359 3360 static void __tcp_alloc_md5sig_pool(void) 3361 { 3362 struct crypto_ahash *hash; 3363 int cpu; 3364 3365 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC); 3366 if (IS_ERR(hash)) 3367 return; 3368 3369 for_each_possible_cpu(cpu) { 3370 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch; 3371 struct ahash_request *req; 3372 3373 if (!scratch) { 3374 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) + 3375 sizeof(struct tcphdr), 3376 GFP_KERNEL, 3377 cpu_to_node(cpu)); 3378 if (!scratch) 3379 return; 3380 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch; 3381 } 3382 if (per_cpu(tcp_md5sig_pool, cpu).md5_req) 3383 continue; 3384 3385 req = ahash_request_alloc(hash, GFP_KERNEL); 3386 if (!req) 3387 return; 3388 3389 ahash_request_set_callback(req, 0, NULL, NULL); 3390 3391 per_cpu(tcp_md5sig_pool, cpu).md5_req = req; 3392 } 3393 /* before setting tcp_md5sig_pool_populated, we must commit all writes 3394 * to memory. See smp_rmb() in tcp_get_md5sig_pool() 3395 */ 3396 smp_wmb(); 3397 tcp_md5sig_pool_populated = true; 3398 } 3399 3400 bool tcp_alloc_md5sig_pool(void) 3401 { 3402 if (unlikely(!tcp_md5sig_pool_populated)) { 3403 mutex_lock(&tcp_md5sig_mutex); 3404 3405 if (!tcp_md5sig_pool_populated) 3406 __tcp_alloc_md5sig_pool(); 3407 3408 mutex_unlock(&tcp_md5sig_mutex); 3409 } 3410 return tcp_md5sig_pool_populated; 3411 } 3412 EXPORT_SYMBOL(tcp_alloc_md5sig_pool); 3413 3414 3415 /** 3416 * tcp_get_md5sig_pool - get md5sig_pool for this user 3417 * 3418 * We use percpu structure, so if we succeed, we exit with preemption 3419 * and BH disabled, to make sure another thread or softirq handling 3420 * wont try to get same context. 3421 */ 3422 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void) 3423 { 3424 local_bh_disable(); 3425 3426 if (tcp_md5sig_pool_populated) { 3427 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */ 3428 smp_rmb(); 3429 return this_cpu_ptr(&tcp_md5sig_pool); 3430 } 3431 local_bh_enable(); 3432 return NULL; 3433 } 3434 EXPORT_SYMBOL(tcp_get_md5sig_pool); 3435 3436 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp, 3437 const struct sk_buff *skb, unsigned int header_len) 3438 { 3439 struct scatterlist sg; 3440 const struct tcphdr *tp = tcp_hdr(skb); 3441 struct ahash_request *req = hp->md5_req; 3442 unsigned int i; 3443 const unsigned int head_data_len = skb_headlen(skb) > header_len ? 3444 skb_headlen(skb) - header_len : 0; 3445 const struct skb_shared_info *shi = skb_shinfo(skb); 3446 struct sk_buff *frag_iter; 3447 3448 sg_init_table(&sg, 1); 3449 3450 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len); 3451 ahash_request_set_crypt(req, &sg, NULL, head_data_len); 3452 if (crypto_ahash_update(req)) 3453 return 1; 3454 3455 for (i = 0; i < shi->nr_frags; ++i) { 3456 const struct skb_frag_struct *f = &shi->frags[i]; 3457 unsigned int offset = f->page_offset; 3458 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT); 3459 3460 sg_set_page(&sg, page, skb_frag_size(f), 3461 offset_in_page(offset)); 3462 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f)); 3463 if (crypto_ahash_update(req)) 3464 return 1; 3465 } 3466 3467 skb_walk_frags(skb, frag_iter) 3468 if (tcp_md5_hash_skb_data(hp, frag_iter, 0)) 3469 return 1; 3470 3471 return 0; 3472 } 3473 EXPORT_SYMBOL(tcp_md5_hash_skb_data); 3474 3475 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key) 3476 { 3477 struct scatterlist sg; 3478 3479 sg_init_one(&sg, key->key, key->keylen); 3480 ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen); 3481 return crypto_ahash_update(hp->md5_req); 3482 } 3483 EXPORT_SYMBOL(tcp_md5_hash_key); 3484 3485 #endif 3486 3487 void tcp_done(struct sock *sk) 3488 { 3489 struct request_sock *req = tcp_sk(sk)->fastopen_rsk; 3490 3491 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV) 3492 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 3493 3494 tcp_set_state(sk, TCP_CLOSE); 3495 tcp_clear_xmit_timers(sk); 3496 if (req) 3497 reqsk_fastopen_remove(sk, req, false); 3498 3499 sk->sk_shutdown = SHUTDOWN_MASK; 3500 3501 if (!sock_flag(sk, SOCK_DEAD)) 3502 sk->sk_state_change(sk); 3503 else 3504 inet_csk_destroy_sock(sk); 3505 } 3506 EXPORT_SYMBOL_GPL(tcp_done); 3507 3508 int tcp_abort(struct sock *sk, int err) 3509 { 3510 if (!sk_fullsock(sk)) { 3511 if (sk->sk_state == TCP_NEW_SYN_RECV) { 3512 struct request_sock *req = inet_reqsk(sk); 3513 3514 local_bh_disable(); 3515 inet_csk_reqsk_queue_drop_and_put(req->rsk_listener, 3516 req); 3517 local_bh_enable(); 3518 return 0; 3519 } 3520 return -EOPNOTSUPP; 3521 } 3522 3523 /* Don't race with userspace socket closes such as tcp_close. */ 3524 lock_sock(sk); 3525 3526 if (sk->sk_state == TCP_LISTEN) { 3527 tcp_set_state(sk, TCP_CLOSE); 3528 inet_csk_listen_stop(sk); 3529 } 3530 3531 /* Don't race with BH socket closes such as inet_csk_listen_stop. */ 3532 local_bh_disable(); 3533 bh_lock_sock(sk); 3534 3535 if (!sock_flag(sk, SOCK_DEAD)) { 3536 sk->sk_err = err; 3537 /* This barrier is coupled with smp_rmb() in tcp_poll() */ 3538 smp_wmb(); 3539 sk->sk_error_report(sk); 3540 if (tcp_need_reset(sk->sk_state)) 3541 tcp_send_active_reset(sk, GFP_ATOMIC); 3542 tcp_done(sk); 3543 } 3544 3545 bh_unlock_sock(sk); 3546 local_bh_enable(); 3547 tcp_write_queue_purge(sk); 3548 release_sock(sk); 3549 return 0; 3550 } 3551 EXPORT_SYMBOL_GPL(tcp_abort); 3552 3553 extern struct tcp_congestion_ops tcp_reno; 3554 3555 static __initdata unsigned long thash_entries; 3556 static int __init set_thash_entries(char *str) 3557 { 3558 ssize_t ret; 3559 3560 if (!str) 3561 return 0; 3562 3563 ret = kstrtoul(str, 0, &thash_entries); 3564 if (ret) 3565 return 0; 3566 3567 return 1; 3568 } 3569 __setup("thash_entries=", set_thash_entries); 3570 3571 static void __init tcp_init_mem(void) 3572 { 3573 unsigned long limit = nr_free_buffer_pages() / 16; 3574 3575 limit = max(limit, 128UL); 3576 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */ 3577 sysctl_tcp_mem[1] = limit; /* 6.25 % */ 3578 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */ 3579 } 3580 3581 void __init tcp_init(void) 3582 { 3583 int max_rshare, max_wshare, cnt; 3584 unsigned long limit; 3585 unsigned int i; 3586 3587 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > 3588 FIELD_SIZEOF(struct sk_buff, cb)); 3589 3590 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL); 3591 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL); 3592 inet_hashinfo_init(&tcp_hashinfo); 3593 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash", 3594 thash_entries, 21, /* one slot per 2 MB*/ 3595 0, 64 * 1024); 3596 tcp_hashinfo.bind_bucket_cachep = 3597 kmem_cache_create("tcp_bind_bucket", 3598 sizeof(struct inet_bind_bucket), 0, 3599 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL); 3600 3601 /* Size and allocate the main established and bind bucket 3602 * hash tables. 3603 * 3604 * The methodology is similar to that of the buffer cache. 3605 */ 3606 tcp_hashinfo.ehash = 3607 alloc_large_system_hash("TCP established", 3608 sizeof(struct inet_ehash_bucket), 3609 thash_entries, 3610 17, /* one slot per 128 KB of memory */ 3611 0, 3612 NULL, 3613 &tcp_hashinfo.ehash_mask, 3614 0, 3615 thash_entries ? 0 : 512 * 1024); 3616 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) 3617 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i); 3618 3619 if (inet_ehash_locks_alloc(&tcp_hashinfo)) 3620 panic("TCP: failed to alloc ehash_locks"); 3621 tcp_hashinfo.bhash = 3622 alloc_large_system_hash("TCP bind", 3623 sizeof(struct inet_bind_hashbucket), 3624 tcp_hashinfo.ehash_mask + 1, 3625 17, /* one slot per 128 KB of memory */ 3626 0, 3627 &tcp_hashinfo.bhash_size, 3628 NULL, 3629 0, 3630 64 * 1024); 3631 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size; 3632 for (i = 0; i < tcp_hashinfo.bhash_size; i++) { 3633 spin_lock_init(&tcp_hashinfo.bhash[i].lock); 3634 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain); 3635 } 3636 3637 3638 cnt = tcp_hashinfo.ehash_mask + 1; 3639 sysctl_tcp_max_orphans = cnt / 2; 3640 3641 tcp_init_mem(); 3642 /* Set per-socket limits to no more than 1/128 the pressure threshold */ 3643 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7); 3644 max_wshare = min(4UL*1024*1024, limit); 3645 max_rshare = min(6UL*1024*1024, limit); 3646 3647 init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM; 3648 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024; 3649 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare); 3650 3651 init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM; 3652 init_net.ipv4.sysctl_tcp_rmem[1] = 87380; 3653 init_net.ipv4.sysctl_tcp_rmem[2] = max(87380, max_rshare); 3654 3655 pr_info("Hash tables configured (established %u bind %u)\n", 3656 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size); 3657 3658 tcp_v4_init(); 3659 tcp_metrics_init(); 3660 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0); 3661 tcp_tasklet_init(); 3662 } 3663