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